content
stringlengths 1
15.9M
|
|---|
\section{Intoduction}
The simplest form of classical Julia-Carath\'eodory theorem, given by Carath\'eodory\cite{car29} and Julia\cite{ju20}, follows.
\begin{thm}[Julia-Carath\'eodory theorem]
Let $f:\mathbb{D} \rightarrow \overline{\mathbb{D}}$ be an analytic function.
The limit
$\lim_{t\rightarrow 1^-} \frac{1 - |f(t)|}{1-|t|}$
exists if and only if
$\lim_{t\rightarrow 1^-} f(t)$
exists and has modulus $1$
and
the directional derivative at $1$ exists for all directions pointing into the disk.
\end{thm}
The Julia-Carath\'eodory theorem was extended to higher derivatives by Bolotnikov and Kheifets in \cite{bokh06,bokh08} and earlier, on the upper half plane, by Nevanlinna in his solution of the Hamburger moment problem\cite{nev22}.
There has been some effort to prove an analogue of the Julia-Carath\'eodory theorem in several variables in the works of Abate \cite{ab98,ab05}, Agler M\raise.45ex\hbox{c}Carthy, Young \cite{amy11a}, Jafari \cite{jaf93}, and \black Wlodarczyk \cite{wlo87}. We are interested in a fusion of the two approaches, that is, an analogue of
the Julia-Carath\'eodory theorem in several variables concerning higher derivatives.
Let $\Pi$ denote the upper half plane. On $\Pi$ analogues of the above program exist since $\Pi$ is conformally
equivalent to $\mathbb{D}.$
We will give an analogue of the Julia-Carath\'eodory theorem
on the domain $\Pi^2.$ We work in two variables since operator theoretic representation formulas exist
for analytic functions $f: \Pi^2 \rightarrow \overline{\Pi},$ but do not exist in general due to some classically
notorious obstruction\cite{par70, var71}.
On $\Pi,$ we have the luxury of the Nevanlinna representation.
\begin{thm}[R. Nevanlinna \cite{nev22}] \label{NevanlinnaRep}
Let $h: \Pi \to \mathbb{C}$. There exists a finite positive Borel measure $\mu$ on $\mathbb{R}$ such that
\begin{equation}\label{NevanlinnaRepFormula}
h(z) = \int\!\frac{1}{t-z}\, d\mu(t)
\end{equation}
if and only if $h$ is analytic, takes values in $\overline{\Pi},$ and
\begin{equation}\label{NevanlinnaRepFormulaAsy}
\liminf_{s\to\infty} s\abs{h(is)} < \infty.
\end{equation}
Moreover, for any Pick function $h$ satisfying Equation \eqref{NevanlinnaRepFormulaAsy}
the measure $\mu$ in Equation \eqref{NevanlinnaRepFormula} is uniquely determined.
\end{thm}
Notably, the condition \eqref{NevanlinnaRepFormulaAsy} is conformally related to the limit
in the classical Julia-Carath\'eodory theorem.
The Nevanlinna representation can be used to develop a theory of higher order regularity, since essentially
questions about regularity are equivalent to elementary questions in real analysis and measure theory.
Namely, since
$$\frac{1}{t-z} = -\sum^{\infty}_{n=0} \frac{t^n}{z^{n+1}}$$
questions about regularity at $\infty$ can be reduced to questions about the existence of moments
$\int t^n d\mu(t).$ The Nevanlinna representation in several variables is given in terms of operator
theory, and so questions there can be reduced to questions some operator theoretic analogue
of moments.
\subsection{The L\"owner class}
We denote the two variable Pick class, the set of holomorphic functions from $\Pi^2$ to $\Pi$, as $\mathcal{P}_2.$
In \cite{am11}, Agler and M\raise.45ex\hbox{c}Carthy defined the L\"owner class at infinity.
\begin{defin} \label{lndfn}
The L\"owner class at $\infty,$ denoted $\mathcal{L}^N,$ is the set of functions
$h\in\mathcal{P}_2$ such that $\lim_{s\rightarrow \infty} h(is,is) =0$ and there exists a multi-indexed sequence of real numbers $(\rho_n)_{|n|\leq 2N-1}$
(here, each $n = (n_1,n_2)$ for some non-negative integers
$n_1$ and $n_2$ and $|n| = n_1+n_2$)
such that
$$h(z)=\sum_{|n|\leq 2N-1} \frac{\rho_n}{z^n} + o\left(\frac{1}{\|z\|^{2N-1}}\right) \text{ nontangentially}.$$
\end{defin}
An asymptotic formula holds \emph{nontangentially} at $\infty$ if for each $c \in \mathbb{R}$
the formula holds for all $z$ large enough satisfying
$\norm{z} \le c\ \min{\impart{z_1}}{\impart{z_2}}.$
A weaker notion of regularity is given by the intermediate L\"owner class.
\begin{defin}
The intermediate L\"owner class at $\infty,$ denoted $\mathcal{L}^{N-},$ is the set of functions
$h\in\mathcal{P}_2$ such that $\lim_{s\rightarrow \infty} h(is,is) =0$ and there exists a multi-indexed sequence of real numbers $(\rho_n)_{|n|\leq 2N-2}$
such that
$$h(z)=\sum_{|n|\leq 2N-2} \frac{\rho_n}{z^n} + O\left(\frac{1}{\|z\|^{2N-1}}\right) \text{ nontangentially}.$$
\end{defin}
We show that $\mathcal{L}^N \neq \mathcal{L}^{N-}$ in Section $\ref{lnneqlnminus}.$
We examine an inductive relationship between $\mathcal{L}^{N-1}, \mathcal{L}^{N-},$ and $\mathcal{L}^N,$ which is given
in the following two theorems.
Our first main result describes when a function in $\mathcal{L}^{N-1}$ is in $\mathcal{L}^{N-}.$
\begin{thm} \label{thmoneprime}
Let $h\in \mathcal{P}_2.$
The following are equivalent:
\begin{enumerate}
\item $h \in \mathcal{L}^{N-}.$
\item $h\in \mathcal{L}^{N-1}$ and for each $b\in{(\mathbb{R}^+)}^2,$
$$s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]$$
is bounded for large $s$.
\end{enumerate}
\end{thm}
We prove Theorem \ref{thmoneprime} as Theorem \ref{thmone} in terms
of the language
of the Agler-M\raise.45ex\hbox{c}Carthy vector moment theory, which we will discuss later.
Our second main result describes when a function in $\mathcal{L}^{N-}$ is in $\mathcal{L}^N.$
\begin{thm} \label{thmtwoprime}
Let $h\in \mathcal{P}_2.$
The following are equivalent:
\begin{enumerate}
\item $h \in \mathcal{L}^N.$
\item $h \in \mathcal{L}^{N-}$ and there are residues, not necessarily real, $\{\rho_n\}_{n\leq 2N-1}$ such that
$$h(z) = \sum_{|n|\leq 2N-1} \frac{\rho_n}{z^n}+ o\left(\frac{1}{\|z\|^{2N-1}}\right)$$
nontangentially.
\end{enumerate}
\end{thm}
We prove Theorem \ref{thmtwoprime} as Theorem \ref{thmtwo} in terms
of the language of the Agler-M\raise.45ex\hbox{c}Carthy vector moment theory.
For $N=1$ our theorems are conformally equivalent the two-variable
Julia-Carath\'eodory theorem proven on $\mathbb{D}^2$ in
Agler M\raise.45ex\hbox{c}Carthy, Young \cite{amy11a} where the conformal
analogues of $\mathcal{L}^1$
and $\mathcal{L}^{1-}$ were called C-points and B-points.
An analysis for $N=1$ was given on $\Pi^2$ in Agler, Tully-Doyle, Young\cite{aty11,aty12}.
\section{The Agler-M\raise.45ex\hbox{c}Carthy vector moment theory}
A calculus was developed to calculate the residues of functions in $\mathcal{P}_2$ at $\infty$ in \cite{aty11, am11}.
\begin{thm}[Type I two variable Nevanlinna representation \cite{aty11}]
Let $h\in \mathcal{P}_2$ and suppose that
$sh(is,is)$
is bounded for real $s$ large enough.
Then, there is a separable \black Hilbert space $\mathcal{H},$ an unbounded self-adjoint operator $A$ on $\mathcal{H},$
a positive contraction $Y$ and a vector $\alpha \in \mathcal{H}$ such that
$$\langle(A-z_Y)^{-1}\alpha,\alpha\rangle $$
where $z_Y = Yz_1 + (1-Y)z_2.$
\end{thm}
In terms of the above representation, Agler and M\raise.45ex\hbox{c}Carthy defined \emph{vector moments,}
which occur in a way algebraically analogous
to the way classical moments occur in the theory of the Nevanlinna representation
in one variable.\cite{nev22}.
\begin{defin}
Given a separable \black Hilbert space $\mathcal{H},$ an unbounded self-adjoint operator $A$ on $\mathcal{H},$
a positive contraction $Y$ and a vector $\alpha \in \mathcal{H},$
we say \black $A$ has \emph{vector moments} to order $N$ denoted $(R_k)^N_{i=1}$ if
$$R_k(b)=(b_Y^{-1}A)^{k-1}b_Y^{-1}\alpha$$
exists for every $b\in {(\mathbb{R}^+)}^2.$
If $R_k$ is a vector-valued polynomial in $\frac{1}{b_1}$ and $\frac{1}{b_2},$ that is, there are vectors $(\alpha_n)_{|n|=k}$ such that
$$R_k(b) = \sum_{|n|=k} \frac{1}{b^n}\alpha_n ,\black$$
we extend $R_k$ to all of $\mathbb{C}^2$ via its formula.
\end{defin}
To prove Theorem \ref{thmoneprime} we prove the following equivalence in terms of the Agler-M\raise.45ex\hbox{c}Carthy
vector moment theory.
\begin{thm} \label{thmone}
Let $h\in \mathcal{P}_2.$
The following are equivalent:
\begin{enumerate}
\item $h \in \mathcal{L}^{N-}.$
\item $h \in \mathcal{L}^{N-1}$ and for any type I representation of $h,$
$$h(z) = \langle(A-z_Y)^{-1}\alpha,\alpha\rangle ,$$
$A$ has real vector $(Y,\alpha)$-moments to order $N-1.$
\item For each $b\in{(\mathbb{R}^+)}^2,$
$$s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]$$
is bounded for large $s$.
\end{enumerate}
\end{thm}
We prove Theorem \ref{thmone} in several parts. The implication
$(1)\Rightarrow (2)$ is given in Proposition \ref{prop3.100}. The implication
$(2)\Rightarrow (1)$ is given in Proposition \ref{prop3.200}. The implication
$(2) \Leftrightarrow (3)$ is given in Proposition \ref{thm3.400}.
Our second result, Theorem \ref{thmtwoprime}, becomes the following in the language of the Agler-M\raise.45ex\hbox{c}Carthy moment theory.
\begin{thm} \label{thmtwo}
Let $h\in \mathcal{P}_2.$
The following are equivalent:
\begin{enumerate}
\item $h \in \mathcal{L}^N.$
\item $h \in \mathcal{L}^{N-}$ and for any type I representation of $h,$
$$h(z) = \langle(A-z_Y)^{-1}\alpha,\alpha\rangle ,$$
$A$ has vector $(Y,\alpha)$-moments to order $N-1$ and $R_{N-1}$ is a vector valued polynomial.
\item $h \in \mathcal{L}^{N-}$ and there are residues, not necessarily real, $\{\rho_n\}_{n\leq 2N-1}$ such that
$$h(z) = \sum_{|n|\leq 2N-1} \frac{\rho_n}{z^n}+ o\left(\frac{1}{\|z\|^{2N-1}}\right)$$
nontangentially.
\end{enumerate}
\end{thm}
Theorem \ref{thmtwo} is also proven in several parts. $(1)\Leftrightarrow (2)$
follows directly from the Agler-M\raise.45ex\hbox{c}Carthy moment theory, specifically
their theorem given here as Theorem \ref{hvmstores}, in the light of
Theorem \ref{thmone}. The implication $(1)\Leftrightarrow (3)$ is proven as
Proposition \ref{thm3.700}.
\subsection{Some facts about moments}
In \cite{am11}, Agler and M\raise.45ex\hbox{c}Carthy proved the following:
\begin{thm}[Agler, M\raise.45ex\hbox{c}Carthy \cite{am11}] \label{hmvsrce}
Let $\mathcal{H}$ be a Hilbert space, let $\alpha \in \mathcal{H}$ and assume that $A$ and $Y$ are operators acting on $\mathcal{H}$, with $A$ an unbounded self-adjoint and $Y$ a positive contraction. The following conditions are equivalent.\\ \\
(i)\ \ \ $A$ has finite complex vector $(Y,\alpha)$-moments to order $N-1$ and \\
\hspace*{10mm} for each $l=1,\ldots,N$ there exist vectors $\alpha_n$, $\abs{n}=l$ such that
\begin{equation}
R_l(z)=\sum_{\mid n\mid=l}\frac{1}{z^n}\alpha_n \notag
\end{equation}
\ \ \ \ whenever $z \in \mathbb{C}^2 \setminus \set{z}{z_2 \ne 0, z_1/z_2\notin (-\infty, 0]}$.
\\\\
(ii)\ \ \ $A$ has finite real vector $(Y,\alpha)$-moments to order $N-1$ and\\
\hspace*{10mm} for each $l=1,\ldots,N$ there exist vectors $\alpha_n$, $\abs{n}=l$ such that
\begin{equation}\label{eq2.4.6}
R_l(b)=\sum_{\mid n\mid=l}\frac{1}{b^n}\alpha_n
\end{equation}
\ \ \ \ whenever $b \in {\mathbb{R}^+}^2$.
\end{thm}
We also define scalar moments.
\begin{defin}
The $k$th real scalar moment is
$$r_k(b) = \langle R_{\lceil{k/2}\rceil}(b),AR_{\lfloor{k/2}\rfloor}(b)\rangle . $$
\end{defin}
Notably, the $r_k$ are always real valued when they are defined and furthermore if the $r_k$
are given by polynomials in $\frac{1}{b_1}, \frac{1}{b_2}$ then they must have real coefficients.
We will use the following key result about scalar moments.
\begin{thm}[Agler, M\raise.45ex\hbox{c}Carthy \cite{am11}]\label{hvmstores} \label{hmvsmtr}
A function $h\in \mathcal{L}^N$ if and only if $h$ has a type I representation
$$h(z)= \langle{(A-z_Y)}^{-1} \alpha,\alpha\rangle ,$$
such that $A$ has polynomial vector $(Y,\alpha)$-moments to order $N-1.$
Moreover,
$$r_{k}(z)=-\sum_{|n|=k} \frac{\rho_n}{z^n}$$
where $\rho_n$ are as in Definition \ref{lndfn}.
\end{thm}
The following telescoping lemma gives a formula that will let us prove the main results.
\begin{lem}\label{lem3.100}
Let $h\in\mathcal{P}_2$ with type I representation $$h(z) = \langle{(A-z_Y)}^{-1} \alpha,\alpha\rangle ,$$
be such that $A$ has vector $(Y,\alpha)$-moments to order $N-1$ and
scalar moments up to order $2N-1.$
Let $b\in {(\mathbb{R}^+)}^2.$ Let $X_b = b_Y^{-1/2}Ab_Y^{-1/2}.$ Let $\beta_k = X_b^{k}b_Y^{-1/2}\alpha.$
Then,
$$h(isb) + \sum^{2N-1}_{k=1} r_{k}(isb) =
\cofrac{(is)^{2(N-1)}}{\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle }.$$
\end{lem}
\begin{proof}
Note $b_Y^{-1/2}$ exists since $b_Y$ is strictly positive.
Note that the expressions for $r_{2k-1}$ and $r_{2k}$ in the notation of the lemma become:
$$r_{2k-1}(isb) = (is)^{-(2k-1)}\langle\beta_k,\beta_k\rangle $$
$$r_{2k}(isb) = (is)^{-2k}\langle\beta_{k-1},\beta_k\rangle .$$
The proof will go by induction.
When $N=1,$
$$h(isb) + r_{1}(isb) = \langle(A-isb_Y)^{-1}\alpha,\alpha\rangle + \langle(isb_Y)^{-1}\alpha,\alpha\rangle $$
$$= \langle(X_b-is)^{-1}\beta_0,\beta_0\rangle + \langle(is)^{-1} \beta_0,\beta_0\rangle $$
$$= \langle[(X_b-is)^{-1} + (is)^{-1}]\beta_0,\beta_0\rangle .$$
So we are done.
Now suppose, by induction,
$$h(isb) + \sum^{2N-1}_{k=1} r_{2k-1}(isb) = \cofrac{(is)^{2(N-1)}}{\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle} $$
and additionally we have vector $(Y,\alpha)$-moments to order $N$ and scalar $(Y,\alpha)$-moments to order $2N+1.$
So,
\begin{align*}
& h(isb) + \sum^{2N+1}_{k=1} r_{k}(isb)\\
&=
\cofrac{(is)^{2(N-1)}}{\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle}
+ \cofrac{(is)^{2N}}{\langle\beta_{N-1},\beta_N\rangle}
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle} \\
& = \cofrac{(is)^{2(N-1)}}
{\langle[(X_b - is)^{-1} + (is)^{-1} + (is)^{-2}X_b] \beta_{N-1},\beta_{N-1}\rangle }
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle} \\
& = \cofrac{(is)^{2(N-1)}}{\langle[(X_b - is)^{-1} + (is)^{-2}(X_b+is)] \beta_{N-1},\beta_{N-1}\rangle }
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle} \\
& = \cofrac{(is)^{2(N-1)}}{\langle[(X_b - is)^{-1} + (is)^{-2}(X_b+is)(X_b-is)(X_b-is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle }
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle} \\
& = \cofrac{(is)^{2(N-1)}}
{\langle[(X_b - is)^{-1} + ((is)^{-2}X_b^2 - 1)(X_b-is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle}
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle} \\
& = \cofrac{(is)^{2N}}{\langle[X_b^2(X_b-is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle }
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle} \\
& = \cofrac{(is)^{2N}}{\langle[(X_b-is)^{-1}] X_b\beta_{N-1},X_b\beta_{N-1}\rangle}
+ \cofrac{(is)^{2N+1}}{\langle\beta_{N},\beta_N\rangle } \\
& = \cofrac{(is)^{2N}}{\langle[(X_b-is)^{-1} +(is)^{-1}] \beta_{N},X_b\beta_{N}\rangle}.
\end{align*}
This concludes the proof.
\end{proof}
\section{Proofs of results}
\subsection{Proof of operator theoretic results}
First we endeavor to prove the equivalence of $(1)$ and $(2)$ in Theorem \ref{thmone}. We separate the proof \black into two parts.
\begin{prop}\label{prop3.100}
Let $h\in\mathcal{P}_2$ with type I representation $$h(z) = \langle{(A-z_Y)}^{-1} \alpha,\alpha\rangle .$$ If
$h\in\mathcal{L}^{N-}$, then $h\in\mathcal{L}^{N-1}$ and $A$ has vector $(Y,\alpha)$-moments to order $N-1.$
\end{prop}
\begin{proof}
Suppose $h\in\mathcal{L}^{N-}.$ Then, $h\in\mathcal{L}^{N-1}.$
We will show $A$ has real vector $(Y,\alpha)$-moments to order $N-1.$
This is sufficient by Theorem \ref{hmvsrce}.
By Theorem \ref{hmvsmtr},
$$h(isb) + \sum^{2N-3}_{k=1} r_{k}(isb)= h(isb) - \sum_{|n|\leq 2N-3} \frac{\rho_n}{(isb)^n},$$
and $A$ has $(Y,\alpha)$-moments to order $N-2.$
So by Lemma \ref{lem3.100}, adopting its notation,
$$(is)^{2N-1}\left[
h(isb) - \sum_{|n|\leq 2N-2} \frac{\rho_n}{(isb)^n}\right] = (is)^3\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle - is\sum_{|n|= 2N-2} \frac{\rho_n}{b^n}.$$
Since $h\in \mathcal{L}^{N-},$ for some $C> 0$,
$$\abs{(is)^{2N-1}\left[
h(isb) - \sum_{|n|\leq 2N-2} \frac{\rho_n}{(isb)^n}\right]}\leq C$$
So,
$$\abs{(is)^3\langle[{(X_b-is)}^{-1} - (is)^{-1}] \beta_{N-2},\beta_{N-2}\rangle - (is)\sum_{|n|=2N-2} \frac{\rho_n}{b^n}} \leq C.$$
Taking the \black real part preserves this inequality. Thus,
$$\abs{\text{Re } (is)^3\langle[{(X_b-is)}^{-1} + (is)^{-1}] \beta_{N-2},\beta_{N-2}\rangle } \text{ }\leq C.$$
Simplifying,
$$\abs{\text{Re } (is)^2\langle\frac{X_b}{X_b-is} \beta_{N-2},\beta_{N-2}\rangle } \text{ }\leq C$$
$$\abs{\text{Re } s^2\langle\frac{X_b^2+isX_b}{X_b^2+s^2} \beta_{N-2},\beta_{N-2}\rangle } \text{ }\leq C$$
$$\left\langle\frac{s^2X_b^2}{X_b^2+s^2} \beta_{N-2},\beta_{N-2}\right\rangle \text{ }\leq C$$
By the spectral theorem, there is a measure $\mu$ so that,
$$\left\langle\frac{s^2X_b^2}{X_b^2+s^2} \beta_{N-2},\beta_{N-2}\right\rangle =\int \frac{s^2x^2}{x^2+s^2} \abs{\beta_{N-2}(x)}^2 d\mu(x).$$
Note the integrand is monotone increasing in $s,$ so apply monotone convergence theorem to get
$$\int |x\beta_{N-2}(x)|^2 d\mu(x) = \int |\beta_{N-1}(x)|^2 d\mu(x)$$
exists and is finite.
So $X_b\beta_{N-2}\in\text{Dom } X_b.$ That is, $(b_Y^{-1/2}Ab_Y^{-1/2})^{N-1}b_Y^{-1/2}\alpha\in\text{Dom }b_Y^{-1/2}Ab_Y^{-1/2}.$
So, $(Ab_Y^{-1})^{N-1} \in \text{Dom }A.$
Thus, $A$ has $(Y,\alpha)$-moments to order $N-1.$
\end{proof}
The other direction goes as follows.
\begin{prop} \label{prop3.200}
Let $h\in\mathcal{P}_2$ with type I representation $$h(z) = \langle{(A-z_Y)}^{-1} \alpha,\alpha\rangle .$$ Then, if
$h\in\mathcal{L}^{N-1}$ and $A$ has vector $(Y,\alpha)$-moments to order $N-1,$ then
$h\in\mathcal{L}^{N-}.$
\end{prop}
\begin{proof}
Suppose $h\in \mathcal{L}^{N-1}$ and $h$ has $(Y,\alpha)$-moments to order $N-1.$
By Theorem \ref{hmvsrce}, \black
$$z_{Y}^{-1}(Az_Y^{-1})^{N-2}\alpha=R_{N-1}(z) = \sum_{|n|=N-1} \frac{1}{z^n}\alpha_n.\black $$
Since we have $(Y,\alpha)$-moments to order $N-1,$
$$(Az_Y^{-1})^{N-1}\alpha=AR_{N-1}(z) = A\sum_{|n|=N-1} \frac{1}{z^n}\alpha_n$$
is well defined. Note, by linear independence of monomials, each $\alpha_n\in \text{Dom} A.$ Thus,
$$(Az_Y^{-1})^{N-1}\alpha=AR_{N-1}(z) = \sum_{|n|=N-1} \frac{1}{z^n}A\alpha_n.$$
So,
\begin{align*}
r_{2N-2}(z) &= \langle AR_{N-1}(z),R_{N-1}(\overline{z})\rangle \\
&= \sum_{|m|=N-1}\sum_{|n|=N-1} \frac{1}{z^nz^m}\langle A\alpha_n,\alpha_m\rangle =\sum_{|n|=2N-2} \frac{\rho_n}{z^{n}},
\end{align*}
where $\rho_{n} = \sum_{n+m=2N-2} \langle A\alpha_n,\alpha_m\rangle .$
Note that if $b\in {(\mathbb{R}^+)}^2$,
$$r_{2N-2}(isb)=\frac{1}{(is)^{2N-2}}=\sum_{|n|=2N-2} \frac{\rho_n}{z^{n}}$$
is real valued. Thus, by linear independence \black of monomials, each $\rho_n$ is real valued. So
$$h(z)-\sum^{2N-2}_{l=1} r_l(z) = \langle(A-z_Y)^{-1}AR_{N-1}(z),R_{N-1}(\overline{z})\rangle $$
$$\|z\|^{2N-1}(h(z)-\sum^{2N-2}_{l=1} r_l(z)) = \|z\|^{2N-1}\langle(A-z_Y)^{-1}AR_{N-1}(z),R_{N-1}(\overline{z})\rangle .$$
Now notice
$$\|z\|^{2N-1}|(h(z)-\sum^{2N-2}_{l=1} r_l(z))| \leq \|z\|\|(A-z_Y)^{-1}\|\|z\|^{N-1}\|AR_{N-1}(z)\|\|z\|^{N-1}\|R_{N-1}(z)\|$$
is nontangentially bounded.
So, $h\in \mathcal{L}^{N-}.$
\end{proof}
This concludes the proof of the equivalence of $(1)$ and $(2)$ in Theorem \ref{thmone}.
\subsection{Proof of function theoretic results}
We now seek to prove the implication $(1) \Leftrightarrow (3)$ in Theorem \ref{thmone}
and Theorem \ref{thmtwo}.
We begin with the following lemma which will allow us to prove $(1) \Leftrightarrow (3)$
for Theorem \ref{thmone}.
\begin{lem}\label{thm3.500}
Let $h\in\mathcal{P}_2.$ Suppose $h\in\mathcal{L}^{N-1}$ and for each $b\in {(\mathbb{R}^+)}^2,$ for large $s,$
$$(is)^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]$$
is bounded.
Then \black
$$r_{2N-1}(b) = \lim_{s\rightarrow\infty} (is)^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right].$$
\end{lem}
\begin{proof}
Suppose
$h\in\mathcal{L}^{N-1}$ and for each $b\in {(\mathbb{R}^+)}^2,$
$$J_b(s) := s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]$$
is bounded.
Let $h$ have a \black type I representation $h(z) = \langle{(A-z_Y)}^{-1} \alpha,\alpha\rangle .$
We will show $A$ has vector $(Y,\alpha)$-moments to order $N-1$ and apply the equivalence of 1 and 2 in Theorem \ref{thmone}.
By Lemma \ref{lem3.100}, adopting its notation,
$$h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}} = (is)^{-2(N-2)}\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-2},\beta_{N-2}\rangle .$$
With this substitution,
$$J_b(s)
=(is)^{2N-1}\text{Im }
(is)^{-2(N-2)}\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-2},\beta_{N-2}\rangle .$$
Simplified, we obtain
$$J_b(s)=\langle\frac{-s^2X_b^2}{X_b^2+s^2}\beta_{N-2},\beta_{N-2}\rangle .$$
Applying the spectral theorem and monotone convergence theorem as in the proof of the equivalence on $(1)$ and $(2)$ in Theorem \ref{thmone}, we get
$$\lim_{s\rightarrow\infty} (is)^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right] = -\|\beta_{N-1}\|^2= r_{2N-1}(b).$$
\end{proof}
We now prove $(1) \Leftrightarrow (3)$
for Theorem \ref{thmone}.
\begin{prop}\label{thm3.400}
Let $h\in\mathcal{P}_2.$ Then,
$h\in\mathcal{L}^{N-}$ if and only if $h\in\mathcal{L}^{N-1}$ and for each $b\in {(\mathbb{R}^+)}^2,$ for large $s,$
$$s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]$$
is bounded.
\end{prop}
\begin{proof}
Suppose $h\in\mathcal{L}^{N-}.$
The term $\sum_{|n|={2N-2}} \frac{\rho_n}{(isb)^{n}}$ is real. So,
$$s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]
=s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-2}} \frac{\rho_n}{(isb)^{n}}\right],$$
which is bounded since $h\in \mathcal{L}^{N-}.$
On the other hand, suppose
$h\in\mathcal{L}^{N-1}$ and for each $b\in {(\mathbb{R}^+)}^2,$
$$s^{2N-1}\text{Im }\left[h(isb)-\sum_{|n|\leq{2N-3}} \frac{\rho_n}{(isb)^{n}}\right]$$
is bounded.
By Theorem \ref{thm3.500}, we have scalar moments to order $2N-1$ and thus vector $(Y,\alpha)$-moments.
So by the equivalence of $(1)$ and $(2)$ in Theorem \ref{thmone}, we are done.
\end{proof}
The following finishes the proof of Theorem \ref{thmtwo} by showing that
We now prove $(1) \Leftrightarrow (3).$
\begin{prop} \label{thm3.700}
Let $h\in\mathcal{P}_2.$ Then, $h\in \mathcal{L}^N$ if and only if $h \in \mathcal{L}^{N-}$ and there are residues, not necessarily real, $\{\rho_n\}_{n\leq 2N-1}$ such that
$$h(z) = \sum_{|n|\leq 2N-1} \frac{\rho_n}{z^n}+ o\left(\frac{1}{\|z\|^{2N-1}}\right)$$
nontangentially.
\end{prop}
\begin{proof}
The forward direction is true by definition.
On the other hand, suppose $h \in \mathcal{L}^{N-}$ and the residues exist.
Let $b\in {(\mathbb{R}^+)}^2.$ Let $h$ have a type I representation
$$h(z) = \langle{(A-z_Y)}^{-1} \alpha,\alpha\rangle .$$
By Lemma \ref{lem3.100}, adopting its notation,
$$h(isb) + \sum^{2N-1}_{k=1} r_{k}(isb) = (is)^{-2(N-1)}\langle[(X_b - is)^{-1} + (is)^{-1}] \beta_{N-1},\beta_{N-1}\rangle .$$
Multiply through by $(is)^{2N-1}.$
$$(is)^{2N-1}\left[ h(isb) + \sum^{2N-1}_{k=1} r_{k}(isb)\right] = \langle[is(X_b - is)^{-1} + 1] \beta_{N-1},\beta_{N-1}\rangle .$$
So,
$$(is)^{2N-1}\left[ h(isb) + \sum^{2N-1}_{k=1} r_{k}(isb)\right] = \langle[\frac{-s^2}{X_b^2 + s^2} + 1] \beta_{N-1},\beta_{N-1}\rangle .$$
Applying the spectral theorem and taking limits gives
$$\lim_{s\rightarrow \infty}(is)^{2N-1}\left[ h(isb) + \sum^{2N-1}_{k=1} r_{k}(isb)\right] = - \|\beta_{N-1}\|^2 + \|\beta_{N-1}\|^2 = 0.$$
Now
$$\lim_{s\rightarrow \infty}(is)^{2N-1}\left[ h(isb) + \sum^{2N-1}_{k=1} r_{k}(isb) - h(isb) + \sum_{|n|\leq 2N-1} \frac{\rho_n}{(isb)^n} \right] = 0.$$
Applying Theorem \ref{hmvsmtr},
$$\lim_{s\rightarrow \infty}(is)^{2N-1}\left[ r_{2N-1}(isb) + \sum_{|n|= 2N-1} \frac{\rho_n}{(isb)^n} \right] = 0.$$
Simplifying,
$$\lim_{s\rightarrow \infty}r_{2N-1}(b) + \sum_{|n|= 2N-1} \frac{\rho_n}{b^n} = 0,$$
that is, $$r_{2N-1}(b) = -\sum_{|n|= 2N-1} \frac{\rho_n}{b^n}.$$
\end{proof}
\section{$\mathcal{L}^N\neq\mathcal{L}^{N-}$} \label{lnneqlnminus}
Now we give an example that shows the hierarchy of L\"owner classes in two variables at infinity does not collapse, that is, $\mathcal{L}^N\neq\mathcal{L}^{N-},$ which was shown for the case $N=1$ in \cite{amy11a}.
That $\mathcal{L}^N\neq\mathcal{L}^{N-}$ is in stark contrast to the theory in one variable where the
classes are identical\cite{ju20,car29}.
Let
$\mathcal{H} = l^2(Z_{2(n-1)}),$
and $\pi: Z_{2(n-1)}\rightarrow B(l^2(Z_{2(n-1)}))$ a left regular representation i.e. $\pi(j)e_i = e_{j+i}.$
Let $A = [\pi(1) + \pi(-1)]$ and $Y$ be a diagonal matrix satisfying
$Ye_i = e_i$, for $i\neq n$, and $Ye_{n-1} = te_{n-1}.$
Let $\alpha = e_0.$
Let $f$ be the Pick function defined by
$$f(z) = \langle(A-z_Y)^{-1}\alpha ,\alpha \rangle .$$
Recall
$R_k(z) = (z_Y)^{-1}(Az_Y^{-1})^{k-1}e_0.$
If $k<n,$ it can be shown inductively that
$$R_k(z)=z_1^{-k}\sum^{k-1}_{l=0} \begin{pmatrix}
k-1 \\ l
\end{pmatrix}e_{-(k-1)+2l}.$$
Furthermore,
$$AR_{n-1}(z)=z_1^{-(n-1)}\sum^{n-1}_{l=0} \begin{pmatrix}
n-1 \\ l
\end{pmatrix}e_{-(n-1)+2l}$$
and
$$R_n(z)=\frac{1}{tz_1+(1-t)z_2}z_1^{-(n-1)}e_{n-1} + z_1^{-n}\sum^{n-2}_{l=0} \begin{pmatrix}
n-1 \\ l
\end{pmatrix}e_{-(n-1)+2l}.$$
So, $r_{2n-1}(z) = \langle R_n(z),AR_{n-1}(\overline{z})\rangle $ is not a polynomial,
but for $k< 2n-1,$ $r_k$ is a polynomial. That is, $f \in \mathcal{L}^{N-},$ but $f \notin \mathcal{L}^N.$
|
\section{Introduction}
\label{sec:intro}
\setcounter{equation}{0}
Solitons and integrable systems play an important role in the study of non-linear phenomena because often they appear in the description of some physical systems. The soliton properties are intimately related to the integrability of the relevant mathematical models in which they arise \cite{babelonbook, faddeev}. Some deformations of these theories have been shown to possess solitary waves that behave in a scattering process in a similar way to true solitons \cite{jhep1, jhep2, jhep3, jhep4, jhep5, jhep7, jhep6, arxiv1}.
The deformations of the relativistic integrable $SU(N)$ Toda \cite{jhep6} (the $N=2$ case is the sine-Gordon (SG) model in disguise \cite{jhep1,jhep3, arxiv1}) and Bullough-Dodd (BD) \cite{jhep4} models have been shown to posses an infinite number of quantities which are not exactly time-independent but are, however,
asymptotically conserved. Similar phenomena have been observed in the deformations of the non-relativistic focusing and defocusing non-linear Schr\"odinger (NLS) model possessing bright and dark solitons \cite{jhep2, jhep5, jhep7}. For earlier observations on related phenomena, such as the elastic scattering of solitons in some non-integrable theories, see e.g. \cite{hietarinta}.
Some calculations in solitary wave collisions have been done in the regimes which
are close to either integrable or related to a particular relationship between the parameters of
the colliding solitons, e.g. high relative velocity, one of the solitons is significantly larger than
the other one, fast thin solitons and slow broad solitons are among the cases considered in the
literature (see e.g. \cite{shift} and references therein). An analytical approach on inelastic solitary wave interactions for a quartic gKdV equation has been considered showing the absence of a pure 2-soliton solution in a special regime \cite{martel}.
Recently, certain modified defocusing and focusing NLS models \cite{jhep5, jhep7}, with dark and bright soliton solutions, respectively, have been shown to exhibit the new feature of an infinite tower of exactly conserved charges. For the special case of
two dark (bright) soliton solutions, where the field components are eigenstates of a space-reflection
symmetry, they exhibit an alternating sequence of exactly conserved and asymptotically conserved charges for
the scattering process of the solitons. These were the distinguishing new features associated to the non-relativistic deformed defocusing (focusing)
NLS with dark (bright) soliton solutions, as compared to the previous relativistic quasi-integrable models \cite{jhep1, jhep2, jhep3, jhep4, jhep6, arxiv1}.
This work is a continuation of \cite{jhep1}, in which a deformed sine-Gordon model has been used to introduce the quasi-integrability concept. The main result of our paper is that half of the infinite set of quasi-conserved quantities of the deformed sine-Gordon model of ref. \cite{jhep1} are in fact exactly
conserved, provided that some two-soliton configurations are eigenstates (even or odd) of the space-reflection operator. By considering linear combinations of the asymptotically conserved charges of \cite{jhep1} we show, through analytical and numerical methods, that one tower of them becomes {\sl exactly} conserved and the other
one remains quasi-conserved after the combination; in this way, we have strengthened the arguments of \cite{jhep1} for two-soliton configurations with definite parity. Notice that we have dealt with deformations of the sine-Gordon integrable relativistic field theory and reproduced analogous results to the ones in the non-relativistic non-linear Schr\"odinger model, as presented in the recent contributions for defocusing NLS with dark solitons \cite{jhep5} and the focusing NLS with bright solitons \cite{jhep7}. In addition, in these last references the authors have found, through numerical simulations, that the first non-trivial anomaly vanishes even in the cases where the space-time parity and space-parity arguments do not indicate they should vanish. We believe that the results in \cite{jhep5, jhep7} and the ones in the present paper open the way for new investigations on the nature of the quasi-integrability phenomena.
We consider the both anomalous Lax representations of the deformed sine-Gordon model, and show that the composition of the space-reflection parity with a special internal symmetry turns out to be a symmetry relating the both anomalous Lax representations. Our analytical and numerical results indicate that the charges associated to two-solitons with a definite parity under space-reflection symmetry, i.e. kink-kink (odd parity) and kink-antikink (even parity) scatterings with equal and opposite velocities, split into two infinite towers of conserved and asymptotically conserved charges. In the case of the positive parity breather-like (kink-antikink bound state) solution one has a tower of exactly conserved charges and a sequence of charges which oscillate around a fixed value. We also show, through numerical simulations, the existence of long lived breather-like states in these models \cite{arxiv1}, which in our formulation exhibit a subset of exactly conserved charges.
However, it seems to be that such parity property is a necessary condition in order to have the sequence of the exactly conserved
charges in the kink-kink and kink-antikink systems. In fact, as we will show by numerical simulations, there are some soliton-like configurations without this symmetry in laboratory coordinates (two-solitons without definite parity under space-reflection symmetry: kink-kink and kink-antikink scatterings with unequal and opposite velocities) which exhibit asymptotically conserved charges only. However, we show that in the center-of-mass reference frame of the two solitons the parity symmetries are restored, and then their associated set of exactly conserved charges would be constructed.
In addition, to simulate the time dependence
of field configurations for computing soliton collisions we used the 4th order Runge-Kutta method supplied with non-reflecting boundary conditions suitable to allow the radiation generated as outgoing waves cross the boundary points $x= \pm L$ freely \cite{nonreflec}.
Our simulations show that some radiation is produced by the soliton systems and the rate of loss of the energy depends on the initial conditions of the system.
The paper is organized as follows: in section \ref{qzc0} we discuss the dual set of quasi-zero curvature representations introduced in \cite{jhep1}, based on the $sl(2)$
loop algebra, for a real scalar field
theory subjected to a generic potential, and the dual sets of infinite
number of quasi-conservation laws. We discuss the relationship between the space-time
parity and asymptotically conserved charges. In section \ref{sec:symcharges} we introduce the space-reflection parity symmetry and an order two automorphism of the $sl(2)$ loop algebra relating the both dual sets of quasi-conserved quantities. We discuss the relationships between the space-reflection parity and the exactly conserved charges. A tower of new exactly conserved charges is constructed for each field configuration possessing a definite space-reflection parity. In section \ref{sec:expansion}
we perform the expansion of the theory (\ref{eq1}) around the
sine-Gordon model in power series on the deformation parameter $\epsilon$, and discuss the interplay between the parity of the solution and its dynamics. In section \ref{kink} we study the space-time and space-reflection symmetries of the kink-antikink, kink-kink and breather solitons of the standard sine-Gordon model. In section \ref{lorentz} the Lorentz transformation is considered in order to study the Lorentz boost transformation of the anomalies and charges. It is shown the vanishing of the anomalies associated to solitary waves. In section \ref{sec:numerical} we present the results of our numerical simulations which allowed us to compute and study various properties of the kink-kink, kink-antikink and a system involving a kink and an antikink bound state (breather). In \ref{conclu} we present some conclusions and discussions. The appendix presents useful $\epsilon$-expansions.
\section{The model}
\label{qzc0}
\setcounter{equation}{0}
We consider Lorentz invariant field theories in
$(1+1)$-dimensions with equation of
motion
\begin{equation}
\partial^2 \varphi + \frac{\partial\, V\(\varphi\)}{\partial\, \varphi}=0,
\label{eq1}
\end{equation}
where $\varphi$ is a real scalar field $\varphi$ and $V( \varphi )$ is the scalar potential and the operator $\partial^2$ stands for $\( \frac{\partial^2}{\partial t^2}-\frac{\partial^2}{\partial x^2}\)$.
Let us consider a deformed potential \cite{Bazeia, jhep1}
\begin{eqnarray}
\label{dpot}
V(\varphi, q) = \frac{2}{q^2} \tan^2{\varphi} [1- |\sin{\varphi}|^q]^2,
\end{eqnarray}
where $q$ is a real parameter such that for $q=2$ the potential reduces to the SG potential, i.e. $V(\varphi, 2) = \frac{1}{16} [1- \cos{(4 \varphi)}]$. So, we introduce the deformation parameter $\epsilon$ as $q= 2+ \epsilon$, such that in the limit $\epsilon = 0$ one reproduces the SG model.
In our numerical simulations we will consider the deformed potential (\ref{dpot}), however most of the analytical and numerical discussions are valid for any positive parity potential under the space-time and space-reflection symmetries of the field $\varphi$ defined below.
Next, we discuss a pair of dual sets of anomalous zero curvature representations of the equations of motion (\ref{eq1}) in order to construct the new set of exactly conserved charges. So, in the following subsections, we summarize the relevant steps regarding the construction of the charges and anomalies in the both dual representations \cite{jhep1}, which will be relevant to our discussions in the section 3.
\subsection{A first set of quasi-conserved charges}
\label{fsc}
Consider the first representation through the Lax potentials
\begin{eqnarray}
A_{+}&=& \frac{1}{2}\left[ \(\omega^2 \, V -m\)\, b_{1}
-i\,\omega\, \frac{\partial\,V}{\partial\,\varphi}\,F_1\right],
\label{pot11}\\
A_{-}&=& \frac{1}{2}\, b_{-1} - \frac{i}{2}\,
\omega\, \partial_{-}\varphi\, F_0.
\label{pot1}
\end{eqnarray}
The Lax potentials $A_{\pm}$ are $sl(2)$ loop algebra valued functions of two variables $x_{\pm}= \frac{1}{2} (t \pm x)$. We follow the notations of the loop algebra commutation relations provided in the appendix A of \cite{jhep1}.
The curvature of the connection (\ref{pot11})-(\ref{pot1}) can be written as
\begin{equation}
F_{+-}\equiv \partial_{+}A_{-}-\partial_{-}A_{+}+\sbr{A_{+}}{A_{-}}= X
\, F_1 -\frac{i\,\omega}{2}\left[\partial^2 \varphi + \frac{\partial\,
V}{\partial\, \varphi} \right]\,F_0
\label{zc1}
\end{equation}
with
\begin{equation}
X = \frac{i\,\omega}{2}\, \partial_{-}\varphi\,
\left[\frac{\partial^2\,V}{\partial\,\varphi^2}+\omega^2\, V-m\right].
\label{x1}
\end{equation}
When the equation of motion (\ref{eq1}) is satisfied one has that the term in (\ref{zc1}) which lies on the direction of the Lie algebra generator $F_0$ vanishes.
The construction follows through the so-called abelianization procedure \cite{jhep1, mpla}. So, let us consider the gauge transformation
\begin{equation}
A_{\mu}\rightarrow a_{\mu}=g\, A_{\mu}\,g^{-1}-\partial_{\mu}g\,
g^{-1},\,\,\,\mbox{with} \,\,\,\,g={\rm exp}\left[\sum_{n=1}^{\infty} \zeta_n\, F_n\right].
\label{g11}
\end{equation}
Note that the group element $g$ is an exponentiation of generators lying in
the positive grade subspace generated by the $F_n$'s, with $\zeta_n$ being parameters to be determined. So, when the anomaly does not vanish the new curvature becomes
\begin{equation}
F_{+-}\rightarrow
g\,F_{+-}\,g^{-1}=\partial_{+}a_{-}-\partial_{-}a_{+}+\sbr{a_{+}}{a_{-}}=
X \, g\, F_1\,g^{-1},
\label{newc}
\end{equation}
where the equations of motion (\ref{eq1}) has been used.
The connection $a_{-}$ can be decomposed into its graded components; so one has \cite{jhep1}
\begin{eqnarray}
a_{-}&=&\frac{1}{2}\,b_{-1}
\label{zz1}\\
&-&\frac{1}{2}\,\zeta_1\,\sbr{b_{-1}}{F_1}- \frac{i}{2}\,
\omega\, \partial_{-}\varphi\, F_0\nonumber\\
&-&\frac{1}{2}\,\zeta_2\,\sbr{b_{-1}}{F_2}
+\frac{1}{4}\,\zeta_1^2\,\sbr{\sbr{b_{-1}}{F_1}}{F_1}
- \frac{i}{2}\,
\omega\, \partial_{-}\varphi\, \zeta_1\,\sbr{F_1}{F_0}
-\partial_{-}\zeta_1\,F_1
\nonumber\\
&\vdots&\nonumber\\
&-&\frac{1}{2}\,\zeta_n\,\sbr{b_{-1}}{F_n}+ \ldots. \nonumber
\end{eqnarray}
See appendix B in \cite{jhep1} for the first $\zeta_n$'s obtained by requiring that the component in the direction of $F_{n-1}$ cancels out in $a_{-}$. For example, we can set $\zeta_1= \frac{i}{2}\,
\omega\, \partial_{-}\varphi$, and so on. Therefore, the component $a_{-}$ gets rotated into the abelian sub-algebra generated by the set $\{b_{2n+1}\}$. So, one has
\begin{equation}
\label{am33}
a_{-}=\frac{1}{2}\,b_{-1}+\sum_{n=0}^{\infty}a_{-}^{(2n+1)}\, b_{2n+1}.
\end{equation}
Since the anomaly term $X$ in (\ref{zc1}) is non-vanishing it is not possible to transform $a_{+}$ into
the abelian sub-algebra generated by the $b_{2n+1}$. So, $a_{+}$ takes the form
\begin{equation}
\label{apco}
a_{+}=\sum_{n=0}^{\infty}a_{+}^{(2n+1)}\, b_{2n+1}
+\sum_{n=2}^{\infty} c_{+}^{(n)}\,F_n.
\end{equation}
The first few terms of the coefficients $a_{\pm}^{(2n+1)}$ and $c_{+}^{(n)}$ are provided in appendix B of ref. \cite{jhep1}. The process follows by defining the components of $g F_{1} g^{-1}$ in the r.h.s. of eq. (\ref{newc}) as
\begin{eqnarray}
\label{ff1}
g F_{1} g^{-1} = \sum_{n=0}^{\infty} \gamma^{(2n+1)} b_{2n+1} + \mbox{terms lying on the}\, F_{n}\mbox{'s}.
\end{eqnarray}
The terms proportional to the $F_n$'s exactly cancel out and the curvature (\ref{newc}) is left with the terms in the direction of the $b_{2n+1}$ only. So, the transformed curvature (\ref{newc}) provides the equation
\begin{eqnarray}
\partial_{+}a_{-}^{(2n+1)}-\partial_{-}a_{+}^{(2n+1)}& = & X \gamma^{(2n+1)} \label{qsc0} \\
&\equiv & \beta^{(2n+1)}\qquad\qquad n=0,1,2,\ldots ;
\label{qsc}
\end{eqnarray}
where the quantities $\beta^{(2n+1)}$ are linear in the anomaly $X$ given in
(\ref{x1}).
The eq. (\ref{qsc}) can be written in the $x$ and $t$ coordinates as
\begin{equation}
\frac{d\,Q^{(2n+1)}}{d\,t}=- \,\alpha^{(2n+1)}+
a_{t}^{(2n+1)}\mid_{x=-\infty}^{x=\infty}
\end{equation}
with
\begin{eqnarray}
Q^{(2n+1)}\equiv \int_{-\infty}^{\infty}dx\,a_{x}^{(2n+1)},\qquad\qquad\qquad
\alpha^{(2n+1)}\equiv \int_{-\infty}^{\infty}dx\,\beta^{(2n+1)}.
\label{cad1}
\end{eqnarray}
Since we are interested in finite energy solutions of the theory
(\ref{eq1}) the field
configurations must satisfy the boundary conditions
\begin{eqnarray}
\partial_{\mu}\varphi \rightarrow 0\; ; \qquad\qquad V\(\varphi\)\rightarrow \mbox{\rm
global minimum}\qquad\qquad {\rm as} \qquad x\rightarrow \pm \infty.
\label{bc}
\end{eqnarray}
Therefore, for finite energy solutions satisfying the boundary condition (\ref{bc}), one has
\begin{eqnarray}
\frac{d\,Q^{(1)}}{d\,t}= 0\; , \qquad\qquad\qquad
\frac{d\,Q^{(2n+1)}}{d\,t}=- \,\alpha^{(2n+1)}\qquad
n=1,2,\ldots
\label{qsc1}
\end{eqnarray}
For $n=0$ one has $\beta^{(1)}=0$, and $Q^{(1)}$, as discussed in section \ref{sec:symcharges}, is related to the energy and momentum.
The charges $Q^{(2n+1)}, n\geq 1 $ are not conserved due to the non-trivial anomaly $\alpha^{(2n+1)}$.
Next we summarize the main result of \cite{jhep1}. If the field $\varphi$ is an eigenstate of the space-time reflection around a given point $(x_\Delta , t_\Delta )$ given by
\begin{eqnarray}
\label{stsym1}
P:\,\,\,\,\,\,\,\,\,\,\, \varphi & \rightarrow & -\varphi + \mbox{const.}\\
\label{stsym2}
P : (\widetilde{x}\, ,\, \widetilde{t}) & \rightarrow & (-\widetilde{x}\, ,\, -\widetilde{t}),\,\,\,\,\mbox{with} \,\,\,\,\widetilde{x}=x-x_\Delta\, ,\,\mbox{and},\,\, \widetilde{t}=t-t_{\Delta},
\end{eqnarray}
then, the charges $Q^{(2n+1)}$ in (\ref{qsc1}) satisfy the relationship
\begin{eqnarray}
\label{mirror1}
Q^{(2n+1)}(t=\widetilde{t}_o + t_{\Delta} ) = Q^{(2n+1)}(t = -\widetilde{t}_o + t_{\Delta} ).
\end{eqnarray}
For the two-soliton solutions the charges are asymptotically conserved, i.e. they have the same charges before ($\widetilde{t}_o=-\infty$) and after ($\widetilde{t}_o=+ \infty$) the collision $
Q^{(2n+1)}(+ \infty) = Q^{(2n+1)}(-\infty)$. Whereas, for the breather-like solution, the charges oscillate around a fixed value.
\subsection{A second set of quasi-conserved charges}
\label{ssc}
Next we consider another quasi-zero curvature representation of the
equation of motion (\ref{eq1}) and a second set of quasi-conserved charges for the theory. The new Lax potentials are obtained from
(\ref{pot11})-(\ref{pot1}) by interchanging $x_{+}$ with $x_{-}$, and by reverting the signs of
the grades of the relevant generators. Thus we consider the dual Lax potential \cite{jhep1}
\begin{eqnarray}
\widetilde{A}_{-}&=& \frac{1}{2}\left[ \(\omega^2\,V-m\)\, b_{-1}
-i\,\omega\, \frac{\partial\,V}{\partial\,\varphi}\,F_{-1}\right],
\label{pot22}\\
\widetilde{A}_{+}&=& \frac{1}{2}\, b_{1} - \frac{i}{2}\,
\omega\, \partial_{+}\varphi\, F_0.
\label{pot2}
\end{eqnarray}
The curvature of such a connection becomes
\begin{eqnarray}
\label{cur22}
{\widetilde F}_{+-}\equiv \partial_{+}\widetilde{A}_{-}-\partial_{-} \widetilde{A}_{+}+\sbr{\widetilde{ A}_{+}}{\widetilde{A}_{-}}= \widetilde{X} \, F_{-1}
+\frac{i}{2}\,\omega\,\left[\partial^2\varphi+\frac{\partial V}{\partial
\varphi}\right]\, F_0,
\end{eqnarray}
with
\begin{equation}
\widetilde{X} = -\frac{i}{2}\,
\omega\, \partial_{+}\varphi\,\left[\frac{\partial^2\,V}{\partial\,\varphi^2}+\omega^2\,
V-m\right].
\label{x2}
\end{equation}
The construction of the corresponding charges follows the same procedure as in
section \ref{fsc}. One performs the gauge transformation
\begin{equation}
\widetilde{A}_{\mu}\rightarrow \widetilde{a}_{\mu}=\widetilde{g}\, \widetilde{A}_{\mu}\,\widetilde{g}^{-1}-\partial_{\mu}\widetilde{g}\, \widetilde{g}^{-1} ,
\label{gp1}
\end{equation}
with the group element being
\begin{equation}
\widetilde{g}={\rm exp}\left[\sum_{n=1}^{\infty} \zeta_{-n}\,
F_{-n}\right]
\label{gp2}.
\end{equation}
So, one has
\begin{equation}
\partial_{+} \widetilde{a}_{-}-\partial_{-} \widetilde{a}_{+}+\sbr{\widetilde{a}_{+}}{\widetilde{a}_{-}}= \widetilde{X} \, \widetilde{g}\,
F_{-1}\,\widetilde{g}^{-1},
\label{nct}
\end{equation}
where the equation of motion (\ref{eq1}) has been used to cancel the
component of $\widetilde{F}_{+-}$ in the direction of $F_0$. The new connection takes the form
\begin{eqnarray}
\widetilde{a}_{+}&=&\frac{1}{2}\,b_{1}
+\sum_{n=0}^{\infty}\widetilde{a}_{+}^{(-2n-1)}\, b_{-2n-1},
\label{wap1}\\
\widetilde{a}_{-}&=&\sum_{n=0}^{\infty}\widetilde{a}_{-}^{(-2n-1)}\, b_{-2n-1}
+\sum_{n=2}^{\infty} \widetilde{c}_{+}^{(-n)}\,F_{-n}.
\label{wan1}
\end{eqnarray}
We refer to the appendix C of ref. \cite{jhep1} for the first few terms of the coefficients ${\widetilde a}_{\pm}^{(-2n-1)}$ and $\widetilde{c}_{+}^{(n)}$. Let us define the components of $\widetilde{g} F_{-1} \widetilde{g}^{-1}$ in the r.h.s. of eq. (\ref{nct}) as
\begin{eqnarray}
\label{f221}
\widetilde{g} F_{-1} \widetilde{g}^{-1} = \sum_{n=0}^{\infty} \widetilde{\gamma}^{(-2n-1)} b_{-2n-1} + \mbox{terms lying on the}\, F_{-n}\mbox{'s}.
\end{eqnarray}
A similar construction as in the section (\ref{fsc}) allows us to write the curvature (\ref{nct}) as
\begin{eqnarray}
\label{qsct1}
\partial_{+}{\widetilde a}_{-}^{(-2n-1)}-\partial_{-}{\widetilde a}_{+}^{(-2n-1)} &=& \widetilde{X}\, \widetilde{\gamma}^{(-2n-1)} \\
& \equiv & {\widetilde \beta}^{(-2n-1)}\qquad\qquad n=0,1,2,\ldots
\label{qsct}
\end{eqnarray}
with the quantity ${\widetilde \beta}^{(2n+1)}$ being linear in the anomaly ${\widetilde X}$, given in
(\ref{x2}).
Applying the same boundary conditions as in section \ref{fsc}, for finite energy solutions the quasi conservation laws become
\begin{equation}
\frac{d\,\widetilde{Q}^{(-1)}}{d\,t}= 0\; , \qquad\qquad\qquad
\frac{d\,\widetilde{Q}^{(-2n-1)}}{d\,t}=- \,{\widetilde
\alpha}^{(-2n-1)}\qquad
n=1,2,\ldots
\label{qsc2}
\end{equation}
with
\begin{eqnarray}
\widetilde{Q}^{(-2n-1)}\equiv \int_{-\infty}^{\infty}dx\,{\widetilde
a}_{x}^{(-2n-1)},\qquad\qquad\qquad
\widetilde{\alpha}^{(-2n-1)}\equiv \int_{-\infty}^{\infty}dx\,\widetilde{\beta}^{(-2n-1)}.
\label{cad2}
\end{eqnarray}
Following similar arguments as in the last subsection, one can conclude that the above construction also admit a set of charges which obey a mirror time-symmetry (\ref{mirror1}) and, consequently, a set of asymptotically conserved charges $
\widetilde{Q}^{(-2n-1)}(+ \infty) = \widetilde{Q}^{(-2n-1)}(-\infty). $
In the next section we will define a new set of infinitely many quasi-conserved charges and relevant anomalies associated to the pair of charges $\(Q^{(2n+1)} ,\, \widetilde{Q}^{(-2n-1)}\)$ and anomalies $\(\alpha^{(2n+1)},\,\widetilde{\alpha}^{(-2n-1)} \)$ by means of their linear combinations. Similar construction has been performed in the usual sine-Gordon model \cite{chodos} in order to obtain an infinite number of conservation laws written in the space-time coordinates $\(x,t\)$. The new representations of the charges will turn out to be convenient ones in order to analyse their properties in relation to the space-reflection symmetries of the soliton field configurations.
\section{Space-reflection parity and conserved charges}
\label{sec:symcharges}
We will consider the above dual sets of anomalous Lax representations of the deformed sine-Gordon model and their associated quasi-conservation laws in order to construct a sequence of conserved charges and vanishing anomalies. The construction below follows by relating the dual sets through an operator which is a composition of the space-reflection and an order two automorphism of the $sl(2)$ affine loop algebra. The space-reflection symmetry of some soliton solutions of the deformed SG model will imply the existence of an infinite tower of conserved charges. In order to see this it is convenient to examine a linear combination, at each order $n=1,2,...$, of the above two sets of quasi-conserved charges $Q^{(2n+1)}$ (\ref{cad1}) and $\widetilde{Q}^{(-2n-1)}$ (\ref{cad2}). The discussion for the particular case $n=0$ will be presented below. Notice that each set of the quasi-conserved charges was constructed considering a particular quasi-zero curvature representation of the same deformed sine-Gordon model. So, let us consider the new quasi-conservation laws
\begin{eqnarray}
\label{lcom}
\frac{dQ_{\pm}^{(2n+1)}}{dt} = - \alpha_{\pm}^{(2n+1)}, \,\,\,\,n=1,2,...,
\end{eqnarray}
with the charges $Q_{\pm}^{(2n+1)}$ and anomalies $\alpha_{\pm}^{(2n+1)}$ being defined as the following linear combinations
\begin{eqnarray}
\label{charlc}
Q_{\pm}^{(2n+1)} & \equiv & \mp \frac{1}{w^2} \(Q^{(2n+1)} \pm \widetilde{Q}^{(-2n-1)}\),\\
\alpha_{\pm}^{(2n+1)} &\equiv & \mp \frac{1}{w^2} \(\alpha^{(2n+1)} \pm \widetilde{\alpha}^{(2n+1)} \)\label{anolc0}\\
&=& \mp \frac{1}{w^2} \int_{-\infty}^{+ \infty} dx\, \( \beta^{(2n+1)} \pm \widetilde{\beta}^{(-2n-1)}\),\label{anolc}
\end{eqnarray}
in which the quantities $Q^{(2n+1)}$ and $ \alpha^{(2n+1)}$ defined in eqs. (\ref{cad1}) and (\ref{qsc1}) and the quantities $\widetilde{Q}^{(-2n-1)}$ and $\widetilde{\alpha}^{(2n+1)} $ in eqs. (\ref{qsc2})-(\ref{cad2}) have been used.
Since the theory (\ref{eq1}) is invariant under space-time translations one has that the energy
momentum tensor is conserved. This conservation law is related to the vanishing of the anomalies $\alpha_{\pm}^{(1)}=0$ (i.e. the case $n=0$), which follow from the vanishing of the integrand functions $\beta^{(1)}$ and $\widetilde{\beta}^{(-1)}$. In fact, the linear combinations of the charges $Q^{(1)}$ and $\widetilde{Q}^{(-1)}$, lead to the
energy and momentum of the field configuration, respectively, i.e.
\begin{eqnarray}
\label{ener}
Q_{+}^{(1)} &=& \int_{-\infty}^{+ \infty} dx\,\,\Big[\frac{1}{2}(\partial_t \varphi)^2 + \frac{1}{2}(\partial_x \varphi)^2 + V \Big], \\
Q_{-}^{(1)} &=& \int_{-\infty}^{+ \infty} dx\,\, \partial_x \varphi \partial_t \varphi,
\label{mom}\end{eqnarray}
where $E=Q_{+}^{(1)} $ is the energy and $P=Q_{-}^{(1)}$ is the momentum.
The first non-trivial anomalies from (\ref{anolc}) become
\begin{eqnarray}
\alpha_{\pm}^{(3)} &=& \pm \frac{1}{2}\int_{-\infty}^{+ \infty} dx\, [V'' + w^2 V -m ] \{ \partial_{-} [ (\partial_{-} \varphi)^2] \mp \partial_{+} [ (\partial_{+} \varphi)^2] \}\\
\alpha_{\pm}^{(5)} &=& \pm \frac{1}{2}\int_{-\infty}^{+ \infty} dx \,[V'' + w^2 V -m ] \times\\
&& \Big[\(\frac{3}{2} w^2 (\partial_{-}\varphi)^2 \partial_{-}^2\varphi + \partial_{-}^4\varphi\)\partial_{-}\varphi \pm \(\frac{3}{2} w^2 (\partial_{+}\varphi)^2 \partial_{+}^2\varphi + \partial_{+}^4\varphi\)\partial_{+}\varphi \Big].
\end{eqnarray}
The behaviour of the charges $Q_{\pm}^{(2n+1)}$ and anomalies $\alpha_{\pm}^{(2n+1)}$ in the quasi-conservation laws (\ref{lcom}) for soliton collisions will depend on the symmetry properties of the associated field configurations, in particular on the space-reflection symmetry of each anomaly density $ ( \beta^{(2n+1)} \pm \widetilde{\beta}^{(-2n-1)} )$ of the corresponding anomaly $\alpha_{\pm}^{(2n+1)}$ in (\ref{anolc}), as we will see below. So, let us examine the space-reflection symmetry of those densities.
Consider the space-reflection transformation
\begin{eqnarray}
\label{px}
{\cal P}_x: x_{+} \leftrightarrow x_{-},
\end{eqnarray}
and assume that the scalar field is an eigenstate of the operator ${\cal P}_x$
\begin{eqnarray}
\label{pxvp}
{\cal P}_x: \varphi \rightarrow \varrho\,\, \varphi,\,\,\,\,\,\varrho = \pm 1.
\end{eqnarray}
In addition, consider an even potential $V$ under ${\cal P}_x$
\begin{eqnarray}
\label{evenpot}
{\cal P}_x (V) = V.
\end{eqnarray}
So, the Lax potentials $A_{\pm}$ in (\ref{pot11})-(\ref{pot1}) and $\widetilde{A}_{\pm}$ in (\ref{pot22})-(\ref{pot2}) are related by
\begin{eqnarray}
\widetilde{\Omega}(A_{\pm}) = \widetilde{A}_{\mp}, \\
\widetilde{\Omega} \equiv \widetilde{\Sigma} {\cal P}_x,
\end{eqnarray}
where $\widetilde{\Sigma}$ is an order two automorphism of the $sl(2)$ loop algebra
\begin{eqnarray}
\nonumber
\widetilde{\Sigma}(
b_{2n+1}) &=& b_{-2n-1} \\
\widetilde{\Sigma}(
F_{2n+1}) &=& \varrho\, F_{-2n-1}\label{aut22}\\
\widetilde{\Sigma}(
F_{2n}) &=& \varrho\, F_{-2n}.\nonumber
\end{eqnarray}
Next, let us examine the transformation properties of the anomalies integrand functions $\beta^{(2n+1)}$ and $\widetilde{\beta}^{(-2n-1)}$. From (\ref{x1}) and (\ref{x2}) one can see that the functions $X$ and $\widetilde{X}$ are related by the transformation
\begin{eqnarray}
\label{pxxx}
{\cal P}_x (X) = - \varrho \, \widetilde{X}.
\end{eqnarray}
Moreover, the quantities $\gamma^{(2n+1)}$ (\ref{qsc0}) and $\widetilde{\gamma}^{(-2n-1)}$ (\ref{qsct1}) are related through the transformation
\begin{eqnarray}
\label{pxg}
{\cal P}_x \( \gamma^{(2n+1)} \) = \varrho \widetilde{\gamma}^{(-2n-1)}.
\end{eqnarray}
The above relationship can be obtained by applying $\widetilde{\Omega}$ on the both sides of the eq. (\ref{ff1}) such that
\begin{eqnarray}
\label{omf1}
\widetilde{\Omega} \( g F_{1} g^{-1}\) &=& \sum_{n=0}^{\infty} {\cal P}_x \(\gamma^{(2n+1)}\) \widetilde{\Sigma} ( b_{2n+1}) + \mbox{terms lying on the}\, \varrho F_{-n}\mbox{'s}\\
\varrho \widetilde{g} F_{-1} \widetilde{g}^{-1} & = & \sum_{n=0}^{\infty} {\cal P}_x\(\gamma^{(2n+1)}\) b_{-2n-1} + \mbox{terms lying on the}\, \varrho F_{-n}\mbox{'s} , \label{omf2}
\end{eqnarray}
and comparing the equation (\ref{omf2}) with the eq. (\ref{f221}) one gets (\ref{pxg}). In order to relate the group elements $g$ in (\ref{g11}) and $\widetilde{g}$ in (\ref{gp2}) as $\widetilde{\Omega} (g) = \widetilde{g}$ we have used the automorphism transformation $\widetilde{\Sigma}$ (\ref{aut22}) of the generators $F_n$ and the transformation ${\cal P}_x(\zeta_n) =\varrho \zeta_{-n}$ in (\ref{g11}). This transformation property can also be seen by inspecting the explicit expressions of the first few $\zeta_{\pm n}'s$ presented in the appendix B and C of \cite{jhep1}.
From the relationships (\ref{pxxx}) and (\ref{pxg}) one can get
\begin{eqnarray}
{\cal P}_x \( X \gamma^{(2n+1)}\) = - \widetilde{X} \widetilde{\gamma}^{(-2n-1)} \,\,\, \rightarrow\,\,\, {\cal P}_x \( \beta^{(2n+1)}\) = - \widetilde{\beta}^{(-2n-1)},\,\,\,\,n= 1,2,...
\end{eqnarray}
where the eqs. (\ref{qsc0})-(\ref{qsc}) for $\beta^{(2n+1)}$ and the eqs. (\ref{qsct1})-(\ref{qsct}) for $\widetilde{\beta}^{(-2n-1)}$ have been used.
Therefore, for the linear combinations we have
\begin{eqnarray}
\label{odd1}
{\cal P}_x \(\beta^{(2n+1)} + \widetilde{\beta}^{(-2n-1)}\) &=& - \(\beta^{(2n+1)} + \widetilde{\beta}^{(-2n-1)}\), \,\,\,\,n=1,2,...\\
\label{even1}
{\cal P}_x \(\beta^{(2n+1)} - \widetilde{\beta}^{(-2n-1)}\) &=& + \(\beta^{(2n+1)} - \widetilde{\beta}^{(-2n-1)}\), \,\,\,\,n=1,2,...
\end{eqnarray}
Therefore, for these type of field configurations satisfying (\ref{pxvp}) we have that the anomaly densities possess a definite space-reflection parity, i.e. the anomaly density of $\alpha_{+}^{(2n+1)}$ ($\alpha_{-}^{(2n+1)}$) is an odd (even) function of $x$. So, from the eq. (\ref{anolc}) and using the symmetry property (\ref{odd1}) one has that the set of anomalies $\alpha_{+}^{(2n+1)}$ vanish, i.e.
\begin{eqnarray}
\label{ano123}
\alpha_{+}^{(2n+1)} = 0, \,\,\,\,n=1,2,...
\end{eqnarray}
This implies that a subset of the quasi-conservation laws (\ref{lcom}) turn out to be truly conservation laws
\begin{eqnarray}
\label{cons1}
\frac{dQ_{+}^{(2n+1)}}{dt} = 0, \,\,\,\,n=1,2,...,
\end{eqnarray}
with the charges $Q_{+}^{(2n+1)}$ defined as
\begin{eqnarray}
\label{charge11}
Q_{+}^{(2n+1)} & \equiv & - \frac{1}{w^2} \(Q^{(2n+1)} + \widetilde{Q}^{(-2n-1)}\),
\end{eqnarray}
in which the charges $Q^{(2n+1)}$ and $\widetilde{Q}^{(-2n-1)}$ have been defined in (\ref{cad1}) and (\ref{cad2}), respectively.
Some comments are in order here:
First, if the field $\varphi$ is an eigenstate (\ref{pxvp}) of the space reflection symmetry (\ref{px}) for a fixed time one can construct a tower of conserved charges $Q_{+}^{(2n+1)}$ in (\ref{cons1}). Since the energy $E=Q_{+}^{(1)} $ in (\ref{ener}) is a conserved quantity for any solution, it can be included into the sequence of exactly conserved charges (\ref{cons1}), i. e. we will have the infinite sequence of exactly conserved charges
\begin{eqnarray}
\label{cons11}
\frac{dQ_{+}^{(2n+1)}}{dt} = 0, \,\,\,\,n=0, 1, 2,...
\end{eqnarray}
Second, notice that for an even or odd parity field $\varphi$ the momentum $P=Q_{-}^{(1)}$ in (\ref{mom}) vanishes, since the momentum density is an odd function under (\ref{px}). As we will see below, in their center-of-mass reference frames the kink-kink, kink-antikink and breather (kink-antikink bound state) solutions of the usual sine-Gordon model posses definite parities under (\ref{px}), and so their total momenta will vanish.
Third, since the charges $Q_{\pm}^{(2n+1)}$ in (\ref{lcom}) are linear combinations of asymptotically conserved charges, provided that the field configurations $\varphi$ are eigenstates of the space-time symmetry (\ref{stsym1})-(\ref{stsym2}) as discussed in the last section, we may conclude that the charges $Q_{\pm}^{(2n+1)}$ will also be asymptotically conserved.
Let us write the integrands of the anomalies $\alpha_{\pm}^{(3)}$ in terms of the $\partial_x$ and $\partial_t$ derivatives. So, once the eq. of motion (\ref{eq1}) is used to substitute $\partial_t^2 \varphi \rightarrow [\partial_x^2\varphi -V'(\varphi)]$ in the integrand of $\alpha_{+}^{(3)}$, one has
\begin{eqnarray}
\label{alf1}
\alpha_{+}^{(3)}=-2 \int \, dx \,[V'' + w^2 V -m ] \Big\{ \partial_{x} [ (\partial_{t} \varphi)^2] + \partial_{x} [ (\partial_{x} \varphi)^2] - \partial_x V(\varphi)\Big\}.
\end{eqnarray}
Notice that the `surface' terms of the form $\partial_x\{ \star \}$ in the integrand of (\ref{alf1}) can be discarded by taking into account the boundary condition (\ref{bc}). Therefore, one is left with
\begin{eqnarray}
\label{alf11}
\alpha_{+}^{(3)} & =&-2 \int \, dx \,[V'' + w^2 V] \Big\{ \partial_{x} [ (\partial_{t} \varphi)^2] + \partial_{x} [ (\partial_{x} \varphi)^2]\Big\},\\
& \equiv & \int \, dx \, f_{+}^{(3)}(x,t),\label{alf112}
\end{eqnarray}
where we have defined the anomaly density $f_{+}^{(3)}$. Notice that for even parity potentials (\ref{evenpot}) and for definite parity (even or odd eigenstate) fields $\varphi$ (\ref{pxvp}) the density $f_{+}^{(3)}$ is an odd function, and thus the anomaly $\alpha_{+}^{(3)}$ vanishes.
Similarly, once the eq. of motion (\ref{eq1}) is used to substitute $\partial_x^2 \varphi \rightarrow [\partial_t^2\varphi + V'(\varphi)]$ in the integrand of $\alpha_{-}^{(3)}$, one has
\begin{eqnarray}
\label{alf2}
\alpha_{-}^{(3)}=-2 \int \, dx \,[V'' + w^2 V -m ] \Big\{ \partial_{t} [ (\partial_{t} \varphi)^2] + \partial_{t} [ (\partial_{x} \varphi)^2] + \partial_t V(\varphi)\Big\}.
\end{eqnarray}
Notice that the terms which can be written in the form $\partial_t [ \int \, dx \, ( \star ) ]$ in (\ref{alf2}) can be carried to the l.h.s. of the quasi-conservation law (\ref{lcom}) such that the charge $Q^{(3)}_{-}$ gets redefined as $Q^{(3)}_{-} \rightarrow Q^{(3)}_{-} + \int \, dx \, ( \star )$. Therefore, discarding the terms $[V'' + w^2 V -m ] \partial_t V(\varphi)$ and $-m \{\partial_{t}[ (\partial_{t} \varphi)^2] + \partial_{t} [ (\partial_{x} \varphi)^2] \}$ in the anomaly density above, one is left with the redefined anomaly
\begin{eqnarray}
\label{alf22}
\alpha_{-}^{(3)} = -2 \int \, dx \,[V'' + w^2 V] \Big\{ \partial_{t} [ (\partial_{t} \varphi)^2] + \partial_{t} [ (\partial_{x} \varphi)^2]\Big\}.
\end{eqnarray}
Furthermore, once we substitute $\partial_t^2 \varphi \rightarrow [\partial_x^2\varphi - V'(\varphi)]$ into this expression and follow similar reasoning as above it acquires the following form
\begin{eqnarray}
\label{alf223}
\alpha_{-}^{(3)} &=& -4 \int \, dx \,[V'' + w^2 V] \Big\{ \partial_{t} \varphi \partial_{x}^2 \varphi + \partial_{x} \varphi \partial_{x} \partial_t \varphi \Big\},\\
&\equiv& \int \, dx \, f_{-}^{(3)}(x,t),\label{alf2233}
\end{eqnarray}
where we have defined the anomaly density $f_{-}^{(3)}$. Notice that for even parity potentials (\ref{evenpot}) and for definite parity (even or odd eigenstate) fields $\varphi$ (\ref{pxvp}) the density $f_{-}^{(3)}$ is an even function, and thus the anomaly $\alpha_{-}^{(3)}$ will not vanish solely by a space-reflection parity reason.
Notice that the redefined effective anomalies provided by the expressions (\ref{alf11}) and (\ref{alf223}) will vanish when evaluated on static solutions; in particular, they will vanish for static kink-type solutions. In sec. \ref{vanano} we will show that the anomalies $\alpha_{\pm}^{(3)}$ vanish for traveling solutions, and in particular for traveling kinks, in all Lorentz frames. Moreover, the effective expressions (\ref{alf11}) and (\ref{alf223}), being more amenable to numerical simulations, and the relevant time integrated expressions $\int dt \, \alpha_{\pm}^{(3)}$, will be computed numerically for various two-soliton and breather-like configurations below.
\section{Perturbation theory and space-reflection symmetry}
\label{sec:expansion}
\setcounter{equation}{0}
Next let us analyse the solutions which admit the equation (\ref{eq1}) such that $\varphi$ satisfies the space-reflection symmetry (\ref{px})-(\ref{pxvp}). In sec. \ref{spref} we will examine the space-reflection symmetry of the two-soliton (kink-kink, kink-antikink and breather) solutions of the usual SG model. We perform this construction in perturbation theory around solutions of the SG model, so let us expand the solutions of (\ref{eq1}) into power series in $\epsilon$, as
\begin{equation}
\varphi= \varphi_0+\varphi_1\,\varepsilon +\varphi_2\,\varepsilon^2+\ldots.
\end{equation}
The expansion of the potential (\ref{dpot}) is presented in Appendix \ref{expansion}.
Therefore, expanding the equation of motion (\ref{eq1}) in powers of $\varepsilon$ we find that the
order zero field $\varphi_0$ must satisfy the sine-Gordon equation,
{\it i.e.}
\begin{equation}
\partial^2\varphi_0+\frac{1}{4}\,\sin\(4\,\varphi_0\)=0.
\label{vp0}
\end{equation}
The higher order component of the field $\varphi_n$ satisfies the
equation
\begin{equation}
\partial^2\varphi_n+ \cos{(4 \varphi_0)}\, \varphi_n = f_n
\label{fn},
\end{equation}
where the result $\frac{\partial^2 V}{\partial \varphi^2}|_{\varepsilon=0} = \cos{(4 \varphi_0)}$ for the deformed potential (\ref{dpot}) has been used. The first few $f_n$'s are provided in the appendix \ref{expansion}.
Next, we split (\ref{fn}) into the even and odd parts under the space-reflection parity ${\cal P}_x$ such that
\begin{eqnarray}
\label{fnpm}
\partial^2 \varphi_{n}^{(\pm)} + \Big[\cos{( 4\, \varphi_0)}\, \varphi_n\Big]^{(\pm)} = f_n^{(\pm)},\,\,\,\,\mbox{where}\,\,\,\, \star^{(\pm)} \equiv \frac{1}{2} \(1\pm {\cal P}_x \) \star,\,\,\,\,\,n=1,2,...
\end{eqnarray}
In the next section we consider the two sectors separately and label them by the parameter $\varrho$ in (\ref{pxvp}). It is assumed that $\varphi_0$ possesses a definite parity under the transformation (\ref{px}).
\subsection{The case $\varrho = 1$ }
In this case one has from (\ref{pxvp}) that $\varphi_0^{+} \neq 0$ and $\varphi_0^{-}=0$, and so
\begin{eqnarray}
\label{order0}
{\cal P}_x : \varphi_0 \rightarrow + \varphi_0.
\end{eqnarray}
Therefore, $\varphi_0$ assumes an even parity and (\ref{f11}) implies that $f_1^{(+)} \neq 0$ and $f_1^{(-)}=0$. So, the first order equations become
\begin{eqnarray}
\label{f1p}
\partial^2 \varphi_{1}^{(+)} + \cos{( 4\, \varphi_0^{(+)})}\,\, \varphi_1^{(+)} &=& f_1^{(+)}(\varphi_0^{(+)}),\\
\partial^2 \varphi_{1}^{(-)} + \cos{( 4\, \varphi_0^{(+)})}\,\, \varphi_1^{(-)} &=& 0.\label{f11p}
\end{eqnarray}
Some comments are in order here. First, the pair of fields $\varphi^{(+)}_1$ and $\varphi_{1}^{(-)}$ satisfies the uncoupled linear system of equations (\ref{f1p})-(\ref{f11p}) with variable coefficients. Second, the equation of motion for the even component $\varphi_{1}^{(+)}$ (\ref{f1p}) satisfies a non-homogeneous equation and so it can never vanish. However, the odd part component $\varphi_{1}^{(-)}$ (\ref{f11p}) satisfies a homogeneous equation and so it can vanish. Third, if the field without definite parity $\varphi_{1}$ is a solution, so is the field combination $\varphi_{1}-\varphi_{1}^{(-)}=\varphi_{1}^{(+)}$. Then, one can always choose a first order solution which is even under the space-reflection parity, i.e.
\begin{eqnarray}
\label{order1}
{\cal P}_x : \varphi_1 \rightarrow \varphi_1.
\end{eqnarray}
Next, choosing the zeroth (\ref{order0}) and first order (\ref{order1}) solutions one has that $f_2$ in (\ref{f22}) splits into $f_2^{(+)} \neq 0$ and
$f_2^{(-)} = 0$. So, the second order terms ${\cal O}(\epsilon^2)$ in (\ref{fnpm}) become
\begin{eqnarray}
\label{f2p}
\partial^2 \varphi_{2}^{(+)} + \cos{( 4\, \varphi_0^{(+)})}\,\, \varphi_2^{(+)} &=& f_2^{(+)}(\varphi_0^{(+)}, \varphi_1^{(+)} ),\\
\partial^2 \varphi_{2}^{(-)} + \cos{( 4\, \varphi_0^{(+)})}\,\, \varphi_2^{(-)} &=& 0.\label{f22p}
\end{eqnarray}
Similarly, as in the first order case, from (\ref{f2p})-(\ref{f22p}) one can always choose a second order solution which is even under the space-reflection parity, i.e.
\begin{eqnarray}
\label{order2}
{\cal P}_x : \varphi_2 \rightarrow \varphi_2.
\end{eqnarray}
Following similar reasoning as above one notices that choosing (\ref{order0}), (\ref{order1}) and (\ref{order2}) solutions the function $f_3$ in (\ref{f33}) splits into $f_3^{(+)} \neq 0$ and
$f_3^{(-)} = 0$, and then one can write for the third order solution
\begin{eqnarray}
\label{order3}
{\cal P}_x : \varphi_3 \rightarrow \varphi_3.
\end{eqnarray}
So, one can choose again the third order solution to be even, and this construction can be repeated in all orders ${\cal O}(\epsilon^n)$ in order to construct a perturbative solution which satisfies (\ref{px})-(\ref{pxvp}) with $\varrho=1$ , and so has
charges satisfying (\ref{cons1}). Therefore, the theory (\ref{eq1}) possesses a subset of solutions such that the charges $Q^{(2n+1)}_{+},\,n=1,2,3,...$ are exactly conserved. It must be included into this series the charges $ Q^{(1)}_{\pm}$ presented in (\ref{ener})-(\ref{mom}).
\begin{figure}
\centering
\label{fig1}
\includegraphics[width=12cm,scale=3, angle=0,height=4.5cm]{sas.eps}
\includegraphics[width=12cm,scale=3, angle=0,height=4.5cm]{breather.eps}
\parbox{6in}{\caption{(color online) a) $\varphi_0$ with even space-reflection parity for an antikink/kink pair sent at $v_1=-v_2=0.7$, for initial $t_i$ (green), collision $t_c$ (horizontal black line, i.e. $\varphi(x, t_c)=0$) and final $t_f$ (red) times, respectively. At $t_c$ the solitons have merged and cancel to each other. After collision, $t_f$, each soliton emerges as kink or antikink, respectively. b) $\varphi_0$ with even space-reflection parity for the SG breather oscillating with period $T=14.0496$ plotted for successive times $t_0< t_2....< t_6$, respectively. At $t_0 = T$ the soliton/antisoliton pair has merged and they cancel to each other.}}
\end{figure}
\subsection{The case $\varrho = -1$ }
In this case one has from (\ref{pxvp}) that $\varphi_0^{+} = 0$ and $\varphi_0^{-} \neq 0$, and so
\begin{eqnarray}
\label{order00}
{\cal P}_x : \varphi_0 \rightarrow -\varphi_0.
\end{eqnarray}
Therefore, $\varphi_0$ assumes an odd parity and (\ref{f11}) implies that $f_1^{(+)} = 0$ and $f_1^{(-)} \neq 0$. So, the first order equations become
\begin{eqnarray}
\label{f1m}
\partial^2 \varphi_{1}^{(+)} + \cos{( 4\, \varphi_0^{(-)})}\,\, \varphi_1^{(+)} &=& 0,\\
\partial^2 \varphi_{1}^{(-)} + \cos{( 4\, \varphi_0^{(-)})}\,\, \varphi_1^{(-)} &=& f_1^{(-)}(\varphi_0^{(-)}).\label{f11m}.
\end{eqnarray}
It follows similar arguments as above. First, the pair of fields $\varphi^{(-)}_1$ and $\varphi_{1}^{(+)}$ satisfies the uncoupled linear system of equations (\ref{f1m})-(\ref{f11m}) with variable coefficients. Second, the equation of motion for the odd component $\varphi_{1}^{(-)}$ (\ref{f11m}) satisfies a non-homogeneous equation and so it can never vanish. However, the even part component $\varphi_{1}^{(+)}$ (\ref{f1m}) satisfies a homogeneous equation and so it can vanish. Third, if the field without definite parity $\varphi_{1}$ is a solution, so is the field combination $\varphi_{1}-\varphi_{1}^{(+)}=\varphi_{1}^{(-)}$. Then, one can always choose a first order solution which is odd under the space-reflection parity, i.e.
\begin{eqnarray}
\label{order11}
{\cal P}_x : \varphi_1 \rightarrow - \varphi_1 .
\end{eqnarray}
Next, choosing the zeroth (\ref{order00}) and first order (\ref{order11}) solutions one has that $f_2$ in (\ref{f22}) splits into $f_2^{(+)} = 0$ and
$f_2^{(-)} \neq 0$. So, the second order terms ${\cal O}(\epsilon^2)$ in (\ref{fnpm}) become
\begin{eqnarray}
\label{f2pp}
\partial^2 \varphi_{2}^{(+)} + \cos{( 4\, \varphi_0^{(+)})}\,\, \varphi_2^{(+)} &=& 0,\\
\partial^2 \varphi_{2}^{(-)} + \cos{( 4\, \varphi_0^{(+)})}\,\, \varphi_2^{(-)} &=& f_2^{(-)}(\varphi_0^{(-)}, \varphi_1^{(-)} ).\label{f22pp}
\end{eqnarray}
Similarly, as in the first order case, from (\ref{f2pp})-(\ref{f22pp}) one can always choose a second order solution which is odd under the space-reflection parity, i.e.
\begin{eqnarray}
\label{order22}
{\cal P}_x : \varphi_2 \rightarrow -\varphi_2.
\end{eqnarray}
Next, choosing (\ref{order00}), (\ref{order11}) and (\ref{order22}) solutions the function $f_3$ in (\ref{f33}) splits into $f_3^{(+)} = 0$ and
$f_3^{(-)} \neq 0$, and then one can write for the third order solution
\begin{eqnarray}
\label{order33}
{\cal P}_x : \varphi_3 \rightarrow -\varphi_3.
\end{eqnarray}
One can choose again the third order solution to be odd, and this process can be repeated order by order with $\varphi_n$ satisfying (\ref{px})-(\ref{pxvp}) with $\varrho=-1$, and so one has charges satisfying (\ref{cons1}). Therefore, the theory (\ref{eq1}) possesses a subset of solutions such that the charges $Q^{(2n+1)}_{+},\,n=1,2,3,...$ are exactly conserved. It must also be included into this series the charges $ Q^{(1)}_{\pm}$ provided in (\ref{ener})-(\ref{mom}).
Let us summarize the main results so far. The main result of \cite{jhep1} is that the deformed SG model presents an infinite number of asymptotically conserved charge, for each anomalous zero-curvature representation, and for solitons satisfying the space-time symmetry (\ref{stsym1})-(\ref{stsym2}). Next, for solitons satisfying the same space-time symmetry, as well as the space-reflection symmetry (\ref{px})-(\ref{pxvp}) one can say more. In this case, the sequence of the charges $\{Q^{(1)}_{\pm},\, Q^{(2n+1)}_{+},\,n=1,2,...\}$ become indeed exactly conserved (\ref{cons1}), and so, it constitutes a new result for deformed models of the SG type. So, the model supports an infinite number of conserved charges $\{Q^{(1)}_{\pm},\, Q^{(2n+1)}_{+},\,n=1,2,...\}$ and asymptotically conserved charges $\{Q^{(2n+1)}_{-},\,n=1,2,...\}$ associated to soliton solutions satisfying the both space-time and space-reflection symmetries.
\begin{figure}
\centering
\label{fig2}
\includegraphics[width=12cm,scale=3, angle=0,height=4.5cm]{ss.eps}
\parbox{6in}{\caption{(color online) $\varphi_0$ with odd space-reflection parity for a kink/kink pair sent at $v_1=-v_2=0.7$, for initial $t_i$ (green), collision $t_c$ (blue) and final $t_f$ (red) times, respectively. After collision, $t_f$, each soliton emerges as a kink.}}
\end{figure}
\begin{figure}
\centering
\label{fig3}
\includegraphics[width=12cm,scale=3, angle=0,height=4.5cm]{sasasi.eps}
\includegraphics[width=12cm,scale=3, angle=0,height=4.5cm]{ssasi.eps}
\parbox{6in}{\caption{(color online) a) $\varphi_0$ without definite space-reflection parity for an antikink/kink pair sent at $v_1=0.017, v_2=-0.044$, for initial $t_i$ (green), collision $t_c$ (blue) and final $t_f$ (red) times, respectively. After collision, $t_f$, each soliton emerges as kink or antikink, respectively. b) $\varphi_0$ without definite space-reflection parity for a kink/kink pair sent at $v_1=0.017, v_2=-0.049$, for initial $t_i$ (green), collision $t_c$ (blue) and final $t_f$ (red) times, respectively. After collision, $t_f$, each soliton emerges as a kink.}}
\end{figure}
\section{Space-time symmetries of sine-Gordon solitons}
\label{kink}
Next we discuss the both space-time and space-reflection symmetries in the general two-soliton and breather solutions of the integrable sine-Gordon model. The solutions of the SG model (\ref{vp0}) can be found in terms of the tau functions as \cite{babelon, jhep00}
\begin{eqnarray}
\label{vp0tau}
\varphi_0 = - \frac{i}{2} \log{\frac{\tau_0}{\tau_1}}.
\end{eqnarray}
Since $\varphi_0$ is real we must have $\tau_1 =\tau_0^{\star}$ and so
\begin{eqnarray}
\label{vp0arctan}
\varphi_0 = \arctan{\frac{Im(\tau_0)}{Re(\tau_0)}},
\end{eqnarray}
where the $\tau_0$ function is given by\footnote{Since $\varphi_0 \rightarrow \varphi_0 + \pi/2$ leaves invariant the SG eq. (\ref{vp0}), the relevant expression of \cite{jhep1} can be rewritten as $e^{2 i \varphi_0} = \frac{-i\, \tau_0}{i \, \tau_1}$, so the form of the tau function $\tau_0$ in (\ref{tau0}) follows from that given in \cite{jhep1} provided that the transformation $\tau_0 \rightarrow -i \tau_0$, $\tau_1 \rightarrow i \tau_1$ is performed.}
\begin{eqnarray}
\label{tau0}
\tau_0 = -i + e^{\Gamma_1} + e^{\Gamma_2} + i \gamma e^{\Gamma_1 + \Gamma_2} ,
\end{eqnarray}
with
\begin{eqnarray}
\Gamma_i &=& \epsilon_i \frac{x- v_i t -x_0^{(i)}}{\sqrt{1-v_i ^2}},\,\,\,\,\,i=1,2 \\
\gamma &=& \Big[\tanh{(\frac{\alpha_2-\alpha_1}{2})}\Big]^{2 \epsilon_1 \epsilon_2},
\end{eqnarray}
where $\epsilon_i = \pm 1$, for kink and anti-kink, and $v_i = \tanh{\alpha_i}$, is the velocity (since $c=1$, one must have $v_i < 1 $) of the kink (antikink) $i$.
The solution (\ref{vp0arctan}) can be written as
\begin{eqnarray}
\label{vp0sch}
\varphi_0 = \arctan{\frac{\sqrt{\gamma} \sinh{z_+}}{\cosh{z_{-}}}},
\end{eqnarray}
where
\begin{eqnarray}
z_{+} = \frac{\Gamma_1 + \Gamma_2}{2}+\log{\sqrt{\gamma}},\,\,\,\,\,z_{-} = \frac{\Gamma_1 - \Gamma_2}{2}.
\end{eqnarray}
Then, the space-time parity transformation (\ref{stsym2}) in the new variables $z_{\pm}$ becomes
\begin{eqnarray}
P\,:\,\,\,\,\, (z_{+}\,,\,z_{-}) \,\,\rightarrow \,\, (-z_{+} \,,\, -z_{-}) ,
\end{eqnarray}
and the solution $\varphi_0$ in (\ref{vp0sch}), taking the domain of $\arctan$ to be $(-\pi/2,\pi/2)$, transforms as
\begin{eqnarray}
P(\varphi_0) = - \varphi_0.
\end{eqnarray}
This space-time symmetry has been presented in \cite{jhep1}.
\subsection{Space-reflection symmetry of sine-Gordon solitons}
\label{spref}
Next, let us examine the space-reflection symmetry (\ref{px})-(\ref{pxvp}) of the two-soliton solution (\ref{vp0sch}). From (\ref{vp0sch}) we expect to recover the well known kink-kink, kink-antikink and breather type solutions (see e.g. \cite{rajaraman}). So, we will be interested on some set of parameters $\{v_1,v_2,\epsilon_1, \epsilon_2 \}$ such that the two-soliton solution provides the field $\varphi_0$ with a definite parity under the space-reflection symmetry for any given shifted time $\widetilde{t}=t-t_{\Delta}$, i.e
\begin{eqnarray}
\label{pxxy}
{\cal P}_x &:& x \rightarrow -x,\\
\label{pxx1}
{\cal P}_x &: & \varphi_0 (x, t)\rightarrow \varrho\,\, \varphi_0(x,t),\,\,\,\,\,\varrho = \pm 1.
\end{eqnarray}
For $\epsilon_1=-\epsilon_2 = +1,\, v_2=-v_1$ one has the kink-antikink solution
\begin{eqnarray}
\label{sa1}
\varphi_0^{SA} = \arctan{\Big[\frac{1}{v_1} \frac{\sinh{(\frac{v_1 t}{\sqrt{1-v_1^2}})}}{\cosh{(\frac{x}{\sqrt{1-v_1^2}})}}\Big]}.
\end{eqnarray}
Notice that the kink-antikink solution possesses an even parity ($\varrho =+1$) under the space-reflection transformation (\ref{pxxy})-(\ref{pxx1}), and it can be regarded as the zeroth order solution of deformed SG model in perturbation theory, as elaborated in sec. \ref{sec:expansion}. This configuration is plotted in Fig. 1a for three successive times.
For $\epsilon_1=\epsilon_2 = +1,\, v_2=-v_1$ one has the kink-kink solution
\begin{eqnarray}
\label{ss1}
\varphi_0^{SS} = \arctan{\Big[v_2 \frac{\sinh{(\frac{x}{\sqrt{1-v_2^2}})}}{\cosh{(\frac{v_ 2 t}{\sqrt{1-v_2^2}})}}\Big]}.
\end{eqnarray}
Notice that the kink-kink solution possesses an odd parity ($\varrho =-1$) under the space-reflection transformation (\ref{pxxy})-(\ref{pxx1}), and constitutes a good candidate to be the zeroth order solution of deformed SG model in perturbation theory, as elaborated in sec. \ref{sec:expansion}. This configuration is plotted in Fig. 2 for three successive times.
Making the transformation $v_1 \rightarrow i \nu/\sqrt{1-\nu^2}$ in (\ref{sa1}), i.e. convert the real parameter $v_1$ into an imaginary one, one has
\begin{eqnarray}
\label{br1}
\varphi_0^{b} = \arctan{\Big[\frac{\sqrt{1-\nu^2}}{\nu} \frac{\sin{( \nu\, t )}}{\cosh{(\sqrt{1-\nu^2}\, x)}}\Big]},
\end{eqnarray}
where $|\nu|< 1$ is its frequency. This is another real solution for $\varphi_0$, and it is interpreted as a ``bound" solution of a kink-antikink pair at rest. This solution is the well known breather solution of the SG model and it is a periodic in time solution with period $T_0 = 2\pi/\nu$.
Notice that the breather solution at rest satisfies $\varphi_0^{b} (-x, t) = \varphi_0^{b}(x, t)$ and therefore it possesses an even parity ($\varrho =+1$) under the space-reflection transformation (\ref{pxxy})-(\ref{pxx1}), and it could also be used as the zeroth order solution of the deformed SG model in perturbation theory as considered in sec. \ref{sec:expansion}. This configuration is plotted in Fig. 1b for various times $0<t<T_0$.
Moreover, it is possible to have two-soliton solutions without definite space-reflection parity
\begin{eqnarray}
\label{2kasym1}
\varphi_0^{SA} &=& \arctan{\Big\{\frac{1}{\tanh{(\frac{\alpha_1-\alpha_2}{2})}} \frac{\sinh{\Big[\frac{1}{2} \( \frac{t\, v_1 - x}{\sqrt{1-v_1^2}}- \frac{t\, v_2 -x}{\sqrt{1-v_2^2}} \)\Big]}}{\cosh{\Big[\frac{1}{2} \( \frac{t\, v_1 - x}{\sqrt{1-v_1^2}}+ \frac{t\, v_2 -x}{\sqrt{1-v_2^2}} \)\Big]}} \Big\}},\,\,\,\,\,\\
\label{2kasym2}
\varphi_0^{SS}&=& \arctan{\Big\{\tanh{(\frac{\alpha_2-\alpha_1}{2})} \frac{\sinh{\Big[\frac{1}{2} \( \frac{t\, v_1 - x}{\sqrt{1-v_1^2}}+ \frac{t\, v_2 -x}{\sqrt{1-v_2^2}} \)\Big]}}{\cosh{\Big[\frac{1}{2} \( \frac{t\, v_1 - x}{\sqrt{1-v_1^2}} - \frac{t\, v_2 -x}{\sqrt{1-v_2^2}} \)\Big]}} \Big\}},
\end{eqnarray}
where $v_1 = \tanh{\alpha_1},\,v_2 = \tanh{\alpha_2}$. In Fig. 3 we plot these two-solitons. In Fig. 3a the case $\epsilon_1=-\epsilon_2 = +1,\, v_2 \neq -v_1$ is plotted for three successive times, which is an asymmetric kink-antikink pair. In Fig 3b the case $\epsilon_1=\epsilon_2 = +1,\, v_2 \neq -v_1$ is plotted as an asymmetric kink-kink solution for three successive times. However, one can transform the above solutions through the Lorentz boosts $(x,t)\rightarrow (x', t')$ in order to write them in the forms (\ref{sa1}) and (\ref{ss1}) for the kink-antikink and kink-kink, respectively. The relevant Lorentz boosts $(x,t)\rightarrow (x', t')$ become
\begin{eqnarray}
\label{lor1}\mbox{kink-antikink}: x' &=& -x\cosh{\chi} + t \sinh{\chi},\,\, t' = -x \sinh{\chi} + t \cosh{\chi},\\
\mbox {kink-kink}: x' &=& -x \cosh{\chi} + t \sinh{\chi}\, ,\,\, t' = x \sinh{\chi} - t \cosh{\chi} ,\label{lor2}
\end{eqnarray}
where $\chi \equiv (\alpha_1+\alpha_2)/2$. In the new coordinates $(x', t')$ the two-solitons will be parity eigenstates under the space-reflection: $x' \rightarrow - x';\, t' \rightarrow t' $ transformation.
The last explicit examples show that the space-reflection parity symmetries of the kink-antikink and kink-kink solutions are not, in general, Lorentz invariant, i.e. ${\cal P}_{x'} \neq \Lambda {\cal P}_x \Lambda^{-1}$ with $\Lambda$ being a Lorentz transformation. In fact, in the above examples we have shown that the kink-kink and kink-antikink solutions are eigenstates of ${\cal P}_x$ only in their centre of mass reference frames, respectively.
\section{Lorentz transformations of charges and anomalies}
\label{lorentz}
Let us consider the two dimensional Lorentz transformation
\begin{eqnarray}
\Lambda : \,\,\,\,x_{\pm} \rightarrow e^{\mp \kappa} x_{\pm},
\end{eqnarray}
where the rapidity $\kappa$ and the velocity $v$ are related by $ v= \tanh{\kappa}$. And, in addition let $\sigma$ be an automorphism of the loop algebra $sl_{2}$
\begin{eqnarray}
\sigma(T) = e^{\kappa d}\, T\, e^{-\kappa d},
\end{eqnarray}
where $d$ is the grading operator. So, it follows that under the composition of the above transformations, the Lax operators of the both representations (\ref{pot11})-(\ref{pot1}) and (\ref{pot22})-(\ref{pot2}), respectively, transform as vectors
\begin{eqnarray}
\Theta (A_{\pm}) = e^{\pm \kappa} A_{\pm},\,\,\,\,\,\,\, \Theta ({\widetilde A}_{\pm}) = e^{\pm \kappa} {\widetilde A}_{\pm},\,\,\,\,\, \Theta \equiv \Lambda \sigma.
\end{eqnarray}
Therefore, the curvatures (\ref{zc1}) and (\ref{cur22}) are invariant under the composed transformation $\Theta$, and so are the anomalous terms $X F_1$ and $\widetilde{X} F_{-1}$, respectively. In order to see the Lorentz transformation laws of the charges and anomalies in (\ref{qsc1}) and (\ref{qsc2}) we need to examine the properties of the quantities $\gamma^{(2n+1)}$ introduced in (\ref{ff1}) and $\widetilde{\gamma}^{(-2n-1)}$ in (\ref{f221}), respectively.
Let us examine the relevant quantities of the first gauge transformation performed in section \ref{fsc}. So, the gauge transformation (\ref{g11}) of the connection $A_{-}$ in (\ref{pot1}) gives rise to the new connection $a_{-}$, which has been decomposed in (\ref{zz1}). The term $\frac{i}{2} w \partial_{-}\varphi F_{0}$ in the r.h.s. of the second equation of (\ref{zz1}) undergoes a transformation
\begin{eqnarray}
\Theta(\frac{i}{2} w \partial_{-}\varphi F_{0}) = e^{-\kappa} \,\, \frac{i}{2} w \partial_{-}\varphi F_{0}.
\end{eqnarray}
Since we have chosen the parameter $\zeta_1$ such that the both terms in the second eq. of (\ref{zz1}) cancel to each other, it follows that
\begin{eqnarray}
\Theta(\zeta_1) = \Lambda(\zeta_1) = e^{-\kappa} \zeta_1.
\end{eqnarray}
This implies that each one of the last three terms on the r.h.s. of the third eq. of (\ref{zz1}) gets multiplied by $e^{-\kappa}$ under the action of $\Theta$. Therefore, in order to cancel the $F_1$ component in that eq. one must have $\Theta(\zeta_2) = e^{-2 \kappa} \zeta_2$. Continuing an analogous reasoning, order by order, one can show that
\begin{eqnarray}
\label{lzetan}
\Theta(\zeta_n) = \Lambda(\zeta_n) = e^{-n \kappa} \zeta_n,
\end{eqnarray} which implies that $\Theta(\zeta_n F_{n} ) = \zeta_n F_{n}$. So, the group element $g$ related to the gauge transformation (\ref{g11}) becomes invariant
\begin{eqnarray}
\label{glt}
\Theta(g) = g.
\end{eqnarray}
Therefore, as the connection $A_{\pm}$, the new connection $a_{\pm}$ transforms as a vector
\begin{eqnarray}
\Theta \( a_{\pm}\) = e^{\pm \kappa} a_{\pm}.
\end{eqnarray}
From (\ref{glt}) one notices that $\Theta(g F_{1} g^{-1} ) = e^{\kappa}\, g F_{1} g^{-1}$, and then each term on the r.h.s. of (\ref{ff1}) turns out to be multiplied by $ e^{\kappa}$ when it is applied with $\Theta$. Then, since $\Theta(b_{2n+1})= e^{(2n+1)\kappa} b_{2n+1}$, it follows that
\begin{eqnarray}
\Theta(\gamma^{(2n+1)}) = \Lambda(\gamma^{(2n+1)}) = e^{-2 n \kappa} \gamma^{(2n+1)}.
\end{eqnarray}
Next, using the relationship $\Theta(X) = \Lambda(X) = e^{-\kappa} X$ in the eqs. (\ref{qsc0})-(\ref{qsc}) and the defining equation for $\alpha^{(2n+1)}$ (\ref{cad1}), one can write the transformation rule for the expression $\alpha^{(2n+1)} dt$ as
\begin{eqnarray}
\label{tensor1}
\Theta\(\alpha^{(2n+1)} dt\) = e^{(-2n-1) \kappa} \(\alpha^{(2n+1)} dt \).
\end{eqnarray}
Therefore, from the quasi-conservation law (\ref{qsc1}) one has that $-\alpha^{(2n+1)} dt$, and so $d Q^{(2n+1)}$, is a tensor under the two-dimensional Lorentz group.
A quite analogous procedure as above can be implemented for the second anomalous Lax representation leading to the following transformation rule for the expression $\widetilde{\alpha}^{(-2n-1)} dt$
\begin{eqnarray}
\label{tensor2}
\Theta\(\widetilde{\alpha}^{(-2n-1)} dt\) = e^{(2n+1) \kappa} \(\widetilde{\alpha}^{(-2n-1)} dt\).
\end{eqnarray}
Therefore, from the quasi-conservation law (\ref{qsc2}) one has that $-\widetilde{\alpha}^{(-2n-1)} dt$, and so $d \widetilde{Q}^{(-2n-1)}$, is a tensor under the two-dimensional Lorentz group.
\subsection{Solitary waves and vanishing anomalies}
\label{vanano}
For static solutions, i.e. field configurations which are $x-$dependent only, the charges $Q^{(2n+1)}$ and $\widetilde{Q}^{(2n+1)}$ are obviously $t-$independent, therefore their associated anomalies $\alpha^{(2n+1)} $ and $\widetilde{\alpha}^{(-2n-1)} $ vanish, respectively. In addition, traveling solutions can be obtained through a Lorentz boost transformation from the static ones. Next we will show that the solitary waves, being special traveling waves, provide vanishing anomalies even for non-integrable theories and, consequently, the charges $Q^{(2n+1)}_{\pm}$ in (\ref{lcom}) are exactly conserved for traveling solutions. In order to achieve this, we will show that $\alpha^{(2n+1)}_{\pm} dt $ and $d Q^{(2n+1)}_{\pm}$ vanish for traveling waves in all $(1+1)-$dimensional Lorentz frames. Therefore, if $d Q^{(2n+1)}_{\pm}=0$ in the rest frame of the static solution, they must vanish in all Lorentz frames.
In fact, the expressions $\alpha^{(2n+1)} dt$ in (\ref{tensor1}) and $\widetilde{\alpha}^{(-2n-1)} dt$ in (\ref{tensor2}) transform as tensors under the $1+1$ Lorentz group, consequently if they vanish in the rest frame of the solution, they must vanish in all Lorentz reference frames. Therefore, if $d Q^{(2n+1)}$ and $d \widetilde{Q}^{(-2n-1)}$ vanish on the rest frame of the solution, they should vanish in
all Lorentz frames. One then concludes that the charges $Q^{(2n+1)}$ and $\widetilde{Q}^{(2n+1)}$ are exactly conserved for
traveling wave solutions (like one-soliton solutions) of (\ref{eq1}). In fact, such conclusion holds
for any functional of the scalar field $\varphi$ and its derivatives, which is a tensor under the
Lorentz group. Consequently, from (\ref{lcom}) one has that the composed anomalies also vanish in all Lorentz frames, i.e. $\alpha^{(2n+1)}_{\pm}=0$. Consequently, if $d Q^{(2n+1)}_{\pm}$ vanish on the rest frame of the solution, it should vanish in
all Lorentz frames. One then concludes that the composed charges $Q^{(2n+1)}_{\pm}$ are exactly conserved for
traveling wave solutions, such as the solitary waves, of the model (\ref{eq1}).
We must emphasize that this property holds despite the fact that the expressions $\alpha_{\pm}^{(2n+1)} dt $ involving the composed anomalies (\ref{anolc0}) do not transform as tensors of the Lorentz group, as it is clear from the transformations rules (\ref{tensor1}) and (\ref{tensor2}) for each term of their linear combinations, i.e. $\alpha_{\pm}^{(2n+1)} dt = (\alpha^{(2n+1)} dt \pm \widetilde{\alpha}^{(2n+1)} dt )$ in (\ref{lcom}). The vanishing of the anomalies $\alpha_{\pm}^{(2n+1)}$ merely reflects the fact that each term of their linear combinations vanish in all Lorentz frames.
\section{Numerical support}
\label{sec:numerical}
\setcounter{equation}{0}
In order to check our results on the anomalies $\alpha_{\pm}^{(3)}$ we have performed various numerical
simulations of the Bazeia at. al. model, studying kink-antikink, kink-kink and a system involving
a kink and an antikink bound state (breather). We will study the behaviour of the quasi-conservation laws in (\ref{lcom}) through numerical simulations of soliton collisions. After integration they can be written as
\begin{eqnarray}
\label{timeint}
Q^{(3)}_{\pm}(t) - Q^{(3)}_{\pm}(t_0) =- \int^t_{t_0} dt' \alpha^{(3)}_{\pm}(t'),
\end{eqnarray}
where $t_0$ is the initial time of the simulation, taking to be zero. We will consider the effective anomaly expressions (the ones with ``surface" terms and total time derivatives discarded) as presented in (\ref{alf11})-(\ref{alf112}) and (\ref{alf223})-(\ref{alf2233}), respectively. We used various grid sizes and number of points. The two-soliton (kink-antikink and kink-kink) simulations were performed on a lattice of $2000$ lattice points with lattice spacing of $\Delta x =0.03$ (in the interval $[-L,L]=[-30,30]$). The time step for all of our simulations was $\Delta t =0.001$. The breather-like simulations were performed on a lattice of $10000$ lattice points with lattice spacing of $\Delta x =0.01$ (in the interval $[-L,L]=[-50,50]$). The time evolution was simulated by the fourth order Runge - Kuta method provided that the so-called non-reflecting (transparent) boundary conditions (see e.g. \cite{nonreflec} and references therein) are assumed at the both ends of the lattice grids
\begin{eqnarray}
\label{nonreflec}
\frac{\partial }{\partial t} \varphi (\pm L, t) \pm \frac{\partial }{\partial x} \varphi (\pm L, t) =0.
\end{eqnarray}
So, the total energy is not conserved in the interval $[-L,L]$, but the only energy which flows to the outer regions $|x| > L$ is the energy of radiation waves. The b.c.'s (\ref{nonreflec}) assure that the radiation generated as outgoing waves cross the boundary points $x= \pm L$ freely (it is expected that the radiation leaves the domain $[-L,L]$ without being reflected back), then the total remaining energy is effectively the energy of the field configurations related to the interacting solitons.
Our simulations show that some radiation is produced by the soliton systems and the rate of loss of the energy depends on the initial conditions and the parameter values for each system. The results of our extensive numerical simulations confirm the usefulness of the transparent boundary condition (\ref{nonreflec}).
\subsection{kink-antikink}
The simulations of the kink-antikink system of the deformed SG model will consider, as the initial condition, two analytical solitary wave solutions. In fact, in order to have a kink-antikink system for $t=0$ we consider a kink ($\eta_1= 1,\eta_2=1, l=0$) and an antikink ($\eta_1=-1,\eta_2=1, l=0$), according to the solution in eq. (1.2) of \cite{jhep1}, located some distance
apart and stitched together at the middle point $x=0$.
These kink-antikink simulations are presented in the Figs. 4-5. In the Fig. 4 we show the results for the
collision of equal and opposite velocity solitons with parameters $v_2=-v_1=0.5$ and $\epsilon = 0.06$. The relevant anomaly densities $f^{(3)}_{\pm}(x,t)$, anomalies $\alpha^{(3)}_{\pm}(t)$, as presented in (\ref{alf11})-(\ref{alf112}) and (\ref{alf223})-(\ref{alf2233}), respectively, and the time integrated anomalies $\int^t dt' \alpha^{(3)}_{\pm}(t')$ as functions of time, are plotted. The relevant anomaly densities $f^{(3)}_{\pm}(x,t)$ in (\ref{alf112}) and (\ref{alf2233}) have been plotted as functions of $x$ for three successive times (top figures of each column). These plots show qualitatively the behaviour of these functions such that their integration in the whole space furnish vanishing $\alpha^{(3)}_{+}(t)$ and non-vanishing $\alpha^{(3)}_{-}(t)$ anomalies (middle figures in the both columns). The bottom figures of the left columns show the vanishing expressions $\int^t dt' \alpha^{(3)}_{+}(t')$, whereas the bottom figures of the right columns show the asymptotically vanishing of the expressions $\int^t dt' \alpha^{(3)}_{-}(t')$. Therefore, according to (\ref{timeint}) our numerical simulations show the asymptotically conservation of the charge $Q^{(3)}_{-}$ and the conservation of the charge $Q^{(3)}_{+}$, within numerical accuracy.
In the Fig. 5 we show the results for the
collision of unequal and opposite velocity kinks with parameters $v_2=0.7, v_1=-0.5$ and $\epsilon = 0.03$. Similarly, these simulations of the deformed SG model are performed considering as the initial condition two analytical solitary wave solutions. These results show that the anomalies $\alpha^{(3)}_{\pm}(t)$ do not vanish and the time integrated anomalies $\int^t dt' \alpha^{(3)}_{\pm}(t')$ vanish only asymptotically. According to (\ref{timeint}), the Fig. 5 shows numerically that the both charges $Q^{(3)}_{\pm}$ are conserved only asymptotically, provided that the soliton configuration breaks the even parity associated to the space-reflection symmetry (i.e. it describes a kink-antikink collision with different velocities).
\begin{figure}
\centering
\label{fig4}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{saape06v5.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{saane06v5.eps}
\parbox{6in}{\caption{(color online) Left column figures for (\ref{alf11})-(\ref{alf112}), from top to bottom, anomaly density $f^{(3)}_{+}$, anomaly $\alpha_{+}^{(3)}$ and time integrated anomaly of kink-antikink collision with equal and opposite velocities $v_2=-v_1=0.5$ and $\epsilon = 0.06$. The right column figures show the relevant results for (\ref{alf223})-(\ref{alf2233}) related to anomaly $\alpha_{-}^{(3)}$. The density figures in the both columns correspond to initial (green), collision (blue) and final (red) configurations of the kink-antikink scattering.}}
\end{figure}
\begin{figure}
\centering
\label{fig8}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{saape03v7vn5.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{saane03v7vn5.eps}
\parbox{6in}{\caption{(color online) Left column corresponds to $\alpha_{+}^{(3)}$ and right column to $\alpha_{-}^{(3)}$ for kink-antikink collision with different velocities $v_2=0.7$ (left soliton),\,$v_1=-0.5$ (right soliton) and $\epsilon = 0.03$.}}
\end{figure}
\[
\]
\subsection{kink-kink}
The simulations of the kink-kink system of the deformed SG model will consider, as the initial condition, two analytical solitary wave solutions. In fact, in order to have a kink-kink system for $t=0$ we consider two kinks ($\eta_1= 1,\eta_2=1, l=0$), according to the solution in eq. (1.2) of \cite{jhep1}, located some distance apart and stitched together at the middle point $x=0$.
These kink-kink simulations are presented in the Figs. 6-7. In the Fig. 6 we show the results for the
collision of equal and opposite velocity solitons with parameters $v_2=-v_1=0.5$ and $\epsilon = 0.06$. The relevant anomaly densities $f^{(3)}_{\pm}(x,t)$, anomalies $\alpha^{(3)}_{\pm}(t)$, as presented in (\ref{alf11})-(\ref{alf112}) and (\ref{alf223})-(\ref{alf2233}), respectively, and the time integrated anomalies $\int^t dt' \alpha^{(3)}_{\pm}(t')$ as functions of time, are plotted. The relevant anomaly densities $f^{(3)}_{\pm}(x,t)$ in (\ref{alf112}) and (\ref{alf2233}) have been plotted as functions of $x$ for three successive times (top figures of each column). These plots show qualitatively the behaviour of these functions such that their integration in the whole space furnishes vanishing $\alpha^{(3)}_{+}(t)$ and non-vanishing $\alpha^{(3)}_{-}(t)$ anomalies (middle figures in the both columns). The bottom figures of the left columns show the vanishing expressions $\int^t dt' \alpha^{(3)}_{+}(t')$, whereas the bottom figures of the right columns show the asymptotically vanishing of the expressions $\int^t dt' \alpha^{(3)}_{-}(t')$. Therefore, according to (\ref{timeint}) our numerical results show the asymptotically conservation of the charge $Q^{(3)}_{-}$ and the conservation of the charge $Q^{(3)}_{+}$, within numerical accuracy. Thus, these results provide qualitatively similar behaviour to the kink-antikink collision anomalies simulated in the previous subsection.
In the Fig. 7 we show the results for the
collision of unequal and opposite velocity kink-kink collisions with parameters $v_2=0.9, v_1=-0.4$ and $\epsilon = 0.06$. Similarly, these simulations of the deformed SG model are performed considering as the initial condition two analytical solitary wave solutions. These results show that the anomalies $\alpha^{(3)}_{\pm}(t)$ do not vanish and the time integrated anomalies $\int^t dt' \alpha^{(3)}_{\pm}(t')$ vanish only asymptotically. Therefore, in the Fig. 7 we have shown numerically that the charges $Q^{(3)}_{\pm}$ are only asymptotically conserved when
the configuration does not possess a space-reflection symmetry (kink-kink collision with different velocities).
Therefore, according to the above numerical results, we may conclude that the even(odd) parity associated to the space-reflection symmetry of the kink-antikink (kink-kink) configuration is a necessary condition in order to have a conserved $Q^{(3)}_{+}$ charge, within the numerical accuracy of our simulations. In fact, the Fig. 5 shows the non-vanishing of the anomaly $\alpha^{(3)}_{\pm}(t)$ computed for kink-antikink configuration without even space-reflection parity. In addition, the Fig. 7 shows the non-vanishing of these anomalies computed for kink-kink configuration without odd space-reflection parity.
\begin{figure}
\centering
\label{fig10}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{ssape06v5.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{ssane06v5.eps}
\parbox{6in}{\caption{(color online) Left column figures for (\ref{alf11})-(\ref{alf112}), from top to bottom, anomaly density $f^{(3)}_{+}$, anomaly $\alpha_{+}^{(3)}$ and time integrated anomaly of kink-kink collision with equal and opposite velocities $v_2=-v_1=0.5$ and $\epsilon = 0.06$. The right column figures show the relevant results for (\ref{alf223})-(\ref{alf2233}) related to anomaly $\alpha_{-}^{(3)}$. The density figures in the both columns correspond to initial (green), collision (blue) and final (red) configurations of the kink-kink scattering.}}
\end{figure}
\begin{figure}
\centering
\label{fig14}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{ssape06v9vn4.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{ssane06v9vn4.eps}
\parbox{6in}{\caption{(color online) Left column corresponds to $\alpha_{+}^{(3)}$ and right column to $\alpha_{-}^{(3)}$ for kink-kink collision with different velocities $v_2=0.9$ (left soliton),\,$v_1=-0.4$ (right soliton) and $\epsilon = 0.06$.}}
\end{figure}
However, for the both kink-antikink and kink-kink solitons of the SG model with opposite and different velocities (\ref{2kasym1})-(\ref{2kasym2}) we have shown that in the center-of-mass reference frame ($x', t'$) provided in (\ref{lor1})-(\ref{lor2}) the parity symmetries are recovered, as discussed in the last paragraph of sec. \ref{spref}. So, the simulations performed in these reference frames, in the both kink-antikink and kink-kink cases, will provide vanishing $\alpha_{+}^{(3)}$ anomalies as they have been shown above.
\subsection{Breather: kink-antikink bound state}
Some properties of the breather-like configurations of the deformed sine-Gordon models have been studied in \cite{jhep1,jhep3, arxiv1} through numerical simulations. In \cite{jhep1} the breather-like configurations were obtained by numerically evolving a kink/anti-kink pair initially located some distance apart, as they tend to attract to each other, then this pair progressively becomes a bound state once the radiation emitted by the configuration is absorbed at the extreme regions of the grid. The second approach, adopted in \cite{jhep3, arxiv1}, considers the analytical sine-Gordon breather solution and its time derivative evaluated at $t=0$ as the initial condition for the breather simulation of the deformed model. This procedure also produced some radiation which has been absorbed at the edges of the grid. Here we adopt the second approach which seems to be more appropriate to our purposes; for example, the initial condition will be an analytical function of the SG breather at rest and it already possesses the positive parity of the breather-like configuration which we are looking for. In addition, the initial energy expression in terms of the frequency turns out to be equal to the one associated to the usual SG model, as we will see below. Nevertheless, the first approach deserves a careful examination in the context of our constructions and we will postpone it for a future work.
In the Figs. 8-11 the simulations for the breather-like solution of the deformed SG model are presented. As mentioned above, the input for our program will be the SG breather solution at rest (\ref{br1}), and so the initial condition for our simulations becomes
\begin{eqnarray}
\label{br2}
\varphi_0^{b}|_{t=0} = 0,\,\,\,\,\frac{d}{dt}\varphi_0^{b}|_{t=0} = \frac{\sqrt{1-\nu^2_0}}{\cosh{(\sqrt{1-\nu^2_0}\, x)}}.
\end{eqnarray}
For this initial configuration we have $\frac{d}{dx}\varphi_0^{b}|_{t=0} =0 $, and $V(\varphi=0) = 0 $, and then from (\ref{ener}) one has that the initial energy $E_0 = Q^{(1)}_{+}$ becomes
\begin{eqnarray}
\label{eneri}
E_0 = \sqrt{1-\nu^2_0}.
\end{eqnarray}
In fact, this is the initial energy of the sine-Gordon breather. So, our simulations of the deformed SG model consider as the initial condition the analytical breather solution presented in Fig. 1 b. We will consider the parameters $\nu_0 = 0.8944$ and $\epsilon = \pm 0.06, \pm 0.03$. The initial energy corresponding to this frequency becomes $E_0 = 0.4473 $ and the amplitude of the breather configuration is $A_0 =\varphi_0^{b}(0, \frac{\pi}{2 \nu_0}) = 0.46371$. The relevant anomaly densities $f^{(3)}_{\pm}(x,t)$, anomalies $\alpha^{(3)}_{\pm}(t)$, as presented in (\ref{alf11})-(\ref{alf112}) and (\ref{alf223})-(\ref{alf2233}), respectively, and the time integrated anomalies $\int^t dt' \alpha^{(3)}_{\pm}(t')$ as functions of time, are plotted. The relevant anomaly densities $f^{(3)}_{\pm}(x,t)$ in (\ref{alf112}) and (\ref{alf2233}), respectively, have been plotted as functions of $x$ for three successive times (top figures of each column). These plots show qualitatively the behaviour of these functions such that their integration in the whole space furnishes vanishing $\alpha^{(3)}_{+}(t)$ and non-vanishing (periodic in time) $\alpha^{(3)}_{-}(t)$ anomalies (middle figures in the both columns). The bottom figures of the left columns show the vanishing expressions $\int^t dt' \alpha^{(3)}_{+}(t')$, whereas the bottom figures of the right columns show that the expressions $\int^t dt' \alpha^{(3)}_{-}(t')$ are periodic in time. Therefore, according to (\ref{timeint}) our numerical results show the oscillation of the charges $Q^{(3)}_{-}$ around a fixed value and the exact conservation of the charges $Q^{(3)}_{+}$, within numerical accuracy.
A sensible definition of the ``lifetime" of a breather in deformed SG models is not yet available; however, a working definition has been put forward recently in relation to a particular deformation of the SG model ( see \cite{arxiv1} and references therein). They consider ``short-lived" and ``long-lived" breathers according to the energy loss of the field configuration in $4 \times 10^4$ units of time. The ``short-lived" system was characterized by the energy loss of more than $10 \%$ during this time, and the ``long-lived" system if it was less than $2 \%$. So, following this terminology our simulations for $|\epsilon|=0.03, 0.06$ and $\nu_0 =0.8944 $ can be termed as ``long-lived" breathers, since for the time of the order of $10^5$ one has an energy loss of less than $2\%$. In fact, from the data in Fig. 9 for $|\epsilon| = 0.03$ one has $\frac{(E_0-E_f)}{E_0} \times 100 \approx 0.05\% $, where $E_f \approx 0.4471$. Similarly, the data presented in Fig. 9 for $|\epsilon|= 0.06$ and $\nu_0 =0.8944 $ show the energy loss of the order of $\sim 0.1\%$, thus showing that in this case we also have ``long-lived" breathers. So, we are confident on our extensive numerical simulations showing the existence of long-lived breathers in the region $|\epsilon| < 0.1$. Moreover, we noticed that for some set of parameters $\epsilon$ and $\nu_0$, and after about $ 5\times 10^4$ units of time, the quasi-breathers start moving slowly, and thus break their even parity symmetry. This phenomenon lies beyond our present scope and deserves future investigations.
Notice that the time dependence of the energies of the breather systems presented in Fig. 9 resemble qualitatively to the ones studied in \cite{arxiv1}, when a relevant deformation of SG was considered such that the space-time parity symmetry holds (see e.g. their Figs. 7 and 9). The authors in \cite{arxiv1} analysed the time oscillating component $\alpha^{(2n+1)}$ as presented in (\ref{qsc1}) of the present paper, instead our numerical results consider the both composed anomalies $\alpha_{\pm}^{(3)}$. In fact, our numerical simulations show that the vanishing of the composed anomaly $\alpha_{+}^{(3)}$ leads to the exactly conserved charge $Q^{(3)}_{+}$, whereas the other composed anomaly $\alpha_{-}^{(3)}$ oscillates in time (see Fig. 8), and then the relevant charge $Q^{(3)}_{-}$ will oscillate around a fixed value, according to the equations (\ref{lcom}) and (\ref{timeint}).
\begin{figure}
\centering
\label{fig18}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{brape06u2.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{brane06u2.eps}
\parbox{6in}{\caption{(color online) Left column figures for (\ref{alf11})-(\ref{alf112}), from top to bottom, anomaly density $f^{(3)}_{+}$, anomaly $\alpha_{+}^{(3)}$ and time integrated anomaly for the breather (kink-antikink bound state) oscillation with $\epsilon = 0.06$ and initial condition frequency $\nu_0=0.8944$ in (\ref{br2}). The right column figures show the relevant results for (\ref{alf223})-(\ref{alf2233}) related to anomaly $\alpha_{-}^{(3)}$. The density figures in the both columns correspond to three successive times ($t_i < t< t_f$) in the interval $[t_f -T_0, t_f]$,\, with $T_0=7.025$. The long-lived breather is characterized by the final time of the order $t_f \approx 10^5$.}}
\end{figure}
\begin{figure}
\centering
\label{fig20}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{energy1.eps}
\parbox{6in}{\caption{(color online) Breather simulations of the deformed SG theory (\ref{eq1})-(\ref{dpot}) with initial configuration (\ref{br2}), deformation and initial condition parameters given by $\epsilon = \pm 0.06,\, \pm 0.03$ and $\nu_0=0.8944,\,T_0=7.025$, respectively. The plots show the time dependence of the energy $E=Q_{+}^{(1)}$ (\ref{ener}) (in adimensional units). The final energies $E_f$ of the stabilized breathers are lower than the initial energy $E_0 = 0.4473$ of the SG breather. Notice that the final time of the order of $t_f \approx 10^5$ characterizes the ``long-lived" breathers.}}
\end{figure}
\begin{figure}
\centering
\label{fig21}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{fielde03.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{fielden03.eps}
\parbox{6in}{\caption{(color online) Breather oscillations at $x=0$ during the initial $i=[0, 10^2]$ (top figures) and final $f=[10^5, 1.0010 \times 10^5]$ (bottom figures) intervals of time with parameters $\epsilon = 0.03$ (left column) and $\epsilon = -0.03$ (right column), respectively. The frequency increased $(\nu_i < \nu_f)$ and the amplitude decreased ( $A_i > A_f $ ) in each case. The $\nu_i\,'s$ are averaged over several periods of time and the $\nu_f\,'s$ are the final stabilized frequencies.}}
\end{figure}
\begin{figure}
\centering
\label{fig22}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{fielde06.eps}
\includegraphics[width=2cm,scale=6, angle=0,height=6cm]{fielden06.eps}
\parbox{6in}{\caption{(color online) Breather oscillations at $x=0$ during the initial $i=[0, 10^2]$ (top figures) and final $f=[10^5, 1.0010 \times 10^5]$ (bottom figures) intervals of time with parameters $\epsilon = 0.06$ (left column) and $\epsilon = -0.06$ (right column), respectively. The frequency increased $(\nu_i < \nu_f)$ and the amplitude decreased ( $A_i > A_f $ ) in each case. The $\nu_i\,'s$ are averaged over several periods of time and the $\nu_f\,'s$ are the final stabilized frequencies.}}
\end{figure}
Finally, let us see what would happen if the energy dependence of the stabilized breather on the frequency is assumed to be similar to the usual SG breather as in eq. (\ref{eneri}). So, for the stabilized breather we would have $E_{f} \sim \sqrt{1-\nu^2_{f}}$. Next, consider the relationship
\begin{eqnarray}
\label{rat}
\(\frac{E_{0}}{E_f}\)^2 = \frac{1-\nu_{0}^2}{1-\nu_{f}^2},
\end{eqnarray}
and compute the validity of this ratio for the various simulations. In fact, in Fig. 10 one notices that the breather simulations for $\epsilon=0.03,\,E_0 = 0.4473$ show that the final frequency becomes $\nu_{f}=0.88617$ and the initial energy decreases to $E_f=0.4471$. For these numerical values the l.h.s. of relationship (\ref{rat}) becomes $\sim 1.0009$, whereas the r.h.s. $\sim 1.0736$, thus showing that the ratio is a good approximation to within $7\%$. Similar computations, using the data from Fig. 11 for the case $\epsilon = 0.06$, show the validity of (\ref{rat}) to within $6\%$. Whereas, the l.h.s and r.h.s of (\ref{rat}) computed for the relevant values associated to each of the negative parameters $\epsilon= -0.03$ and $-0.06$ differ to within $15\%$ and $25\%$, respectively. These results, and some other ones, such as the existence of moving quasi-breathers briefly mentioned above, show that the problems of stability, energy dependence on frequency $\nu$ and the deformation parameter $\epsilon$ for breathers deserve careful analytical and numerical studies. We hope to consider such systems in more detail in the near future.
\section{Discussions and some conclusions}
\label{conclu}
The well known results of \cite{jhep1} for soliton solutions can be summarized as follow: a) the model possesses an infinite number of exactly conserved quantities for one-soliton type solutions, i.e. for kink-type solitary waves traveling with a constant speed. b) For two-soliton (kink-kink, kink-antikink) type solutions possessing a special space-time parity symmetry (\ref{stsym1})-(\ref{stsym2}), the charges are asymptotically conserved. This means that these quantities vary in time during the collision process (and sometimes can vary quite a lot) of two one-solitons but return, in the distant future (after the collision), to the values they had in the distant past (before the collision). c) For breather-type solutions, with the same space-time parity symmetry, those charges oscillate around a fixed value.
We have shown that the quasi-conserved charges of the deformed sine-Gordon model (\ref{eq1}) studied in \cite{jhep1} split into two subsets, with different conservation properties. Through linear combinations, at each level $n$ in (\ref{charlc}), of the dual set of asymptotically conserved charges we have obtained a subset comprising a new infinite tower of exactly conserved charges (\ref{cons1}), and a second subset containing the remaining asymptotically conserved ones. The two-soliton (kink-kink and kink-antikink) and breather-like (kink-antikink bound state) solutions in their center-of-mass frame possessing a special space-reflection parity symmetry (\ref{px})-(\ref{pxvp}) give rise to the new tower of exactly conserved charges. In order to examine the parity symmetry and the vanishing of the anomalies (\ref{ano123}) we have used a technique involving the space-reflection and an order two $\rlx\hbox{\sf Z\kern-.4em Z}_2$ automorphism $\widetilde{\Sigma}$ (\ref{aut22}) of the $sl(2)$ loop algebra, in the context of dual anomalous zero-curvature (Lax) representations of the deformed model.
In sec. \ref{sec:expansion} we have implemented a perturbation theory, by a power series expansion on $\epsilon$, in order to study the interplay between the space-reflection symmetry and dynamics of the solutions. In this context, the zeroth order solution in $\epsilon$ can be chosen to be the two-soliton solutions (kink-antikink, kink-kink and breather) of the integrable sine-Gordon model, which
satisfy the symmetries (\ref{pxxy})-(\ref{pxx1}). We have provided an order by order prescription to construct a solution of the deformed model satisfying the parity property, i.e. even or odd parity solution, which implies the existence of a tower of exactly conserved charges associated to a field configuration with definite parity.
The vanishing of the anomalies $\alpha^{(2 n+1)}_{\pm}$ (\ref{anolc0}) for a general solitary wave solution has been verified in section \ref{lorentz} by showing that their components $\alpha^{(2n+1)} dt $ and $\widetilde{\alpha}^{(-2n-1)} dt$ transform as tensors under the Lorentz group. Since for static solitary wave solutions the anomalies vanish, and the moving waves can be obtained through a Lorentz boost, then we may conclude that the anomalies must vanish in all Lorentz frames.
We have checked the predictions of our analytical calculations through numerical simulations. For the special deformed potential (\ref{dpot}) we have computed the first non-trivial anomalies $\alpha^{(3)}_{\pm}$ of the quasi-conservation laws involving the charges $Q^{(3)}_{\pm}$. We have verified that the anomaly $\alpha^{(3)}_{+}$ vanishes, and consequently the exact conservation of the charge $Q^{(3)}_{+}$ holds for various two-soliton configurations, within numerical accuracy. Moreover, the anomaly $\alpha^{(3)}_{-}$ does not vanish, and then the charge $Q^{(3)}_{-}$ is asymptotically conserved for the various two-soliton configurations.
We have verified numerically for the collision of kink-antikink and kink-kink systems traveling in opposite directions with equal velocities, as shown in the Figs. 4-5 and Figs. 6-7, respectively. Moreover, for the breather-type solutions with even parity under space-reflection symmetry (breather at rest) we have constructed a tower of exactly conserved charges and a subset of charges which oscillate around a fixed value (see the Fig. 8). In addition, our extensive numerical simulations show the existence of long-lived breathers in the region $| \epsilon | < 0.1$ (the time dependence of the energy is shown in Fig. 9 for various $\epsilon$ parameters).
For those two-soliton (kink-antikink and kink-kink) solutions in laboratory coordinates and without the space-reflection parity symmetry we have checked, through numerical simulations, that the both set of charges are only asymptotically conserved (see Fig.5 and Fig.7, respectively). However, in the case of solutions of the usual sine-Gordon model we have shown that in their center-of-mass reference frames (\ref{lor1})-(\ref{lor2}), the both systems recover their space-reflection symmetries (in this frame they would be suitable solutions in perturbation theory as the zeroth order in powers of $\epsilon$), and thus the existence of the tower of exactly conserved charges would be guaranteed by the parity symmetry for each pair of solitons in that frame, as shown in the Fig. 4 and Fig. 6, respectively.
The mechanism responsible for the exact conservation of the charges is not well understood yet. As in all of the examples in relativistic field theories where the asymptotically conserved charges have been observed so far, associated to a space-time parity symmetry \cite{jhep3, jhep4, jhep6}, the two-soliton type (kink-kink, kink-antikink and breather) solutions associated to a sequence of exactly conserved charges present special properties under a space-reflection parity transformation. The only explanation we have found, so far, for this sector of the charges, is that those special soliton-like solutions are eigenstates under a space-reflection parity transformation for a fixed time, where the point in $x-$coordinate around which space is reversed depends upon the parameters of the particular solution under consideration (in particular, we have set $x_{\Delta}=0$). In the absence of these symmetries in the kink-kink and kink-antikink collisions we have checked numerically the existence of the asymptotically conserved charges only, which allow us to argue that the parity symmetries of the systems of two-solitons in their center-of-mass reference frame is a necessary condition in order to get the set of exactly conserved charges.
Further research work is necessary to settle such
questions as the decay of the breathers and the parameters $\epsilon$ and $\nu$ dependence of the energy of long-lived breathers, which involve the non-linear dynamics of the scattering. The symmetries involved in the quasi-integrability phenomenon deserve further investigation; in particular, the transition between the both phases: broken (laboratory coordinate frame) and unbroken (center-of-mass frame) parity symmetric phases of the two-soliton solutions and their bound state. In particular, the even parity breather-type system (at rest) for $\epsilon \approx 0.1 $ start moving slowly after $\sim 5\times 10^4$ units of time, and so the parity symmetry is broken and its effect on the tower of exactly conserved charges deserves a careful examination. Finally, the space-time and internal symmetries involved in the quasi-integrability phenomenon deserve further investigation, since they have potential applications in many areas of non-linear sciences.
\noindent {\bf Acknowledgements:}
HB thanks Prof. L.A. Ferreira for enlightening discussions on quasi-integrability and FAPEMAT for partial financial support in the initial stage of the work.
|
\section{Introduction}
Resonance peaks are observed in many natural phenomena. They are treated in
numerous textbooks (see, e.g., \cite{fls}). The traditional way
of their description is to compare their shapes with the relativistic
Breit-Wigner formula \cite{bw, china}
\begin{equation}
f=\frac {k}{(m_{r }^2-M^2)^2+M^2\Gamma _r^2},
\label{bw}
\end{equation}
where $f$ denotes a signal strength, $k$ is the normalization constant,
$M$ is the scanning energy, $m_r$ is the maximum position (the resonance
mass), $\Gamma _r$ is the resonance width. One measures the signal intensity
$f$ at different energies $M$. For example, the atoms considered as oscillators emit
as Breit-Wigner resonances. The resonance shape is a purely statistical
phenomenon. It depends on many details of interactions and need not to be
necessarily symmetric.
The asymmetric resonance peaks were experimentally observed in various
fields of physics even before the formula (\ref{bw}) was proposed.
E.g., the resonance of He atoms observed in the inelastic scattering of
electrons is strongly asymmetric (see \cite{rice, fano}). Many spectroscopic
studies of atoms were devoted to this phenomenon. The name of Fano
resonances was attached to observed asymmetric resonances. In nuclear physics,
they are known as Feshbach resonances \cite{fesh}.
Both types of resonances (we can call them as FF-resonances) are the same thing
when it comes to their mathematical essence \cite{fano, fano1, fesh}. The
asymmetry is explained as a byproduct of the quantum-mechanical interference
between two separate channels of the reaction. Namely, interference between
a background (continuum of states) and a resonant (excitation of the discrete
states) scattering process produces the asymmetric line-shapes. The two
separate channels (closed and open) differ but they couple to each other. The
resonance has an energy width that depends on the coupling between the channels.
In particle physics, such peaks are identified with unstable particles.
Let us mention, however, that sometimes even the "ordinary" hadrons are
treated as FF-resonances \cite{rlj}. It is usually claimed that the symmetric
shape is observed for resonances directly produced in particle collisions.
Their characteristics are compiled by the PDG (Particle Data Group)
\cite{china}. However,
some asymmetry was recently experimentally noticed for narrow resonances (with
c- and b-quarks). It was explained as a consequence of FF-effects
\cite{len1, len2} with interference of bound states and continuum induced
by vacuum excitation of light quark-antiquark pairs leading to creation of
D-mesons.
The situation has strongly changed after the high energy nucleus-nucleus
collisions became available. The created particles have to leak somehow from
the nuclear medium. Medium interactions may lead to some modification of their
characteristics. Really, there are numerous experimental data
\cite{agak, adam, arna, damj, trnk, naru, muto, kozl, kotu, tser} about the
in-medium modification of widths and positions of prominent wide vector meson
resonances. Some of them even contradict each other. They are mainly obtained
from the shapes of dilepton (decay products) mass and transverse momentum
spectra in nucleus-nucleus collisions. Some excess over the expected
symmetric shape of the distribution was observed. Dilepton spectroscopy
directly probes the vector component of the spectral function of the hadronic
medium. The dilepton mass spectra decrease approximately exponentially
with increase of masses but show peaks over this trend at some masses which
can be identified with prominent resonances. The $\rho $-meson peak is usually
the strongest one \cite{agak, adam, arna, damj} in the ratio
$\rho :\omega :\phi =10:1:2$. Below, we concentrate on properties of
in-medium $\rho $-mesons with the special attention to be paid to the
asymmetry of their shape.
Several approaches have been advocated for explanation of the observed excess
and properties of in-medium resonances \cite{pisa, hara, brow, bore, dusl, leup,
rupp, chiu, elet, elio, mart, rapp, hees}. Most of them use effective hadronic
Lagrangians to compute loop corrections and/or just hydrodynamics ideas
of the expanding fireball. See also the review \cite{haya} on cold nuclear
matter effects. However, either positions, widths or heights presented some
problems. Therefore we will not review these attempts in detail.
Here we concentrate not as much on their particular values which we just fit by the
corresponding parameters as on the asymmetry of the resonance shape.
\section{Experiment, Fano-Feshbach-effect and \\ Cherenkov gluons}
The dilepton mass spectrum in semi-central In-In collisions at 158 AGeV
measured by NA60-Collaboration \cite{arna} is shown in Fig.\ref{fitFF_PDG}
by dots with error bars in the region of $\rho$ and $\phi , \; \omega$-mesons.
Its asymmetry is easily seen with some excess in the low-mass wing. The shape
is quite distinct from the familiar $\rho$-peak with PDG-parameters
$m_{\rho}=775$ MeV, $\Gamma _{\rho}=149$ MeV \cite{china}
shown in Fig.1 by the dashed line.
The in-medium modification of $\rho$-meson parameters can not be accounted by
simple variation of them within the Breit-Wigner formula (\ref{bw}). It is
demonstrated by the dash-dotted line in Fig. 1 with fit parameters $m_r=775$
MeV, $\; \Gamma _r=336$ MeV (much larger width!) which does not reproduce the
observed asymmetry.
\begin{figure}[ht]
\includegraphics[width=\textwidth]{fitFF_ident-eps-converted-to.pdf}
\caption{The spectrum of dileptons in semi-central collisions
In(158 A GeV)-In measured by NA60-Collaboration \cite{damj} (points with error
bars). The solid line shows that the fits by Eq. (\ref{sigma1}) with FF-effect
and by Eq. (\ref{ll}) with gluon Cherenkov effect taken into account coincide.
The dashed line corresponds to the Breit-Wigner shape of $\rho $-meson
with PDG-parameters \cite{china}. The dash-dotted line shows this shape
with the modified width.}
\label{fitFF_PDG}
\end{figure}
As described above, the interference of the continuum states and discrete
levels of the reaction leads in quantum mechanics to the well known FF-effect
\cite{fano, fano1, fesh}. They interfere with opposite phase on the two sides
of the resonance as shown in \cite{fano1}. The resonance asymmetry is described
by the following formula derived in Ref. \cite{fano1}:
\begin{equation}
\sigma = \frac {(q+\epsilon )^2}{1+\epsilon ^2}=1+\frac {q^2-1+2q\epsilon }
{\epsilon ^2 +1},
\label{sigma}
\end{equation}
where at the relativistic notation
\begin{equation}
\epsilon = \frac {M^2-m_r^2}{M\Gamma _r}.
\end{equation}
This expression replaces its non-relativistic form in the original paper
\cite{fano} to reproduce properly the contribution due to
the Breit-Wigner resonance shape (\ref{bw}).
The parameter $q$ describes the relative strength of discrete states and
unperturbed continuum. The term linear in $\epsilon $ is in charge of asymmetry.
After subtracting the constant background and normalizing the Breit - Wigner
part, one gets the following expression which can be fitted to experimental
data of Fig.~\ref{fitFF_PDG} when multiplied by the overall normalization
factor $C$:
\begin{equation}
C\frac {1-\sigma }{1-q^2} = C\frac {1-2q\epsilon /(1-q^2)}{\epsilon ^2 +1}.
\label{sigma1}
\end{equation}
Unfortunately, one can not use for the fitting procedure the experimental
data in the whole region of masses $M$ shown in Fig.~\ref{fitFF_PDG} because
the admixture of higher mass resonances is large in the right wing of the
Figure. Therefore we had to use the data in the smaller interval of masses
below the peak, i.e., in the left wing only.
The fit shown by the solid line in Fig.~\ref{fitFF_PDG} resulted in the
value of $q\approx 0.363$ ($\Gamma_r=184$ MeV, $C=29$). Let us note that
the width is larger than the ordinary one.
One concludes that the interference parameter $q$ is quite noticeable to
explain the shape of the resonance with excess of mesons in the
left wing (smaller masses) of the resonance profile.
The admixture of the contribution of direct states to this effect compared
to the influence of the continuum was estimated in \cite{fano1} equal
$\pi q^2/2$ which amounts to about 0.21 in our case. Thus we conclude that
the interference of the continuum with quasibound states is quite important.
The above treatment is based on general quantum-mechanical principles and does
not reveal what particular mechanisms are in charge of the interfering open and
closed channels. One can propose emission of Cherenkov gluons in the nuclear
media \cite{drem} as one of these channels that explains a possible source of
the left-wing asymmetry.
The necessary condition for Cherenkov effects to be observable within some
energy interval is an excess of the refractivity index of the medium $n$ over 1.
It is well known for ordinary media that such an excess happens
due to electromagnetic interactions just in
the left wing of any resonance (e.g., see Fig. 31.5 in \cite{fls}).
According to general formulas (e.g., see \cite{akpo}) this excess is
proportional to the real part of the forward (depicted by 0 below)
scattering amplitude.
\begin{equation}
\Delta n ={\rm Re }n -1 \propto {\rm Re }F(M,0)>0.
\label{delta}
\end{equation}
For the nuclear quark-gluon medium this requirement should be fulfilled for
the chromopermittivity of gluons \cite{drle}. The real part of the
Breit-Wigner amplitude leading to Eq. (\ref{bw}) is positive just within the
low-energy (left) wing of any resonance described by this equation (see, e.g.,
\cite{elet, elio}). Herefrom one gets the general prediction that the shape of
{\it any} resonance formed in high energy nuclei collisions must become
asymmetric with some excess within its left wing compared to the usual
Breit-Wigner shape. One could expect that some collective excitations of the
quark-gluon medium may contribute in these energy intervals in addition to the
traditional effects. Since the probability of Cherenkov radiation is
proportional to $\Delta n$ the asymmetry must be proportional to it. Then the
dilepton mass distribution must get the shape (the formula in \cite{drem} is
slightly corrected):
\begin{equation}
\frac{dN_{ll}}{dM}=\frac {A}{(m_{r }^2-M^2)^2+M^2\Gamma _r^2}
\left(1+w_r\frac{m_{r}^2-M^2}{M\Gamma _r}\Theta (m_{r}-M)\right). \label{ll}
\end{equation}
This formula was also used to fit experimental data in Fig.~\ref{fitFF_PDG}.
The second term is due to the coherent Cherenkov gluon response of the medium
to the penetrating quark
proportional to the real part of the amplitude. It is in charge of the
observed asymmetry. It vanishes at energies above the resonance peak $M>m_r$
because only positive $\Delta n$ lead to Cherenkov effects. Here, we take into
account that the ratio of real to imaginary parts of Breit-Wigner amplitudes is
\begin{equation}
\frac {{\rm Re}F(M,0)}{{\rm Im}F(M,0)}=\frac {m_r^2-M^2}{M\Gamma _r}.
\label{ratio}
\end{equation}
The weight of the second term is described by the only adjustable parameter
$w_r$ for a given resonance $r$. As we see, the general structure of
Eqs (\ref{sigma1}), (\ref{ll}) is the same with
\begin{equation}
w_r=2q/(1-q^2).
\label{wq}
\end{equation}
The adjusted parameters were obtained from the independent fit to
experimental points in Fig. 1.
They are $A=25, \; \Gamma _r=184$ MeV,$\; w_r=0.838$. The relation (\ref{wq})
is well fulfilled. Therefore both fits according to (\ref{sigma1}) and
(\ref{ll}) practically coincide and are shown in Fig. 1 by a single solid line.
The quantum interference of continuum and quasibound states is at the origin
both of asymmetric resonances and of the classical phenomenological prescription
of $\Delta n>0$, which is required for Cherenkov effect. Therefore, FF-effect
can serve as the quantum-mechanics foundation of classical Cherenkov effect
in general. The overlap of both fits demonstrated by the solid line
in Fig.~\ref{fitFF_PDG} supports this conclusion.
\section{QCD perspective}
At that stage one is tempted to speculate about the QCD interpretation of such
a statement. High energy nuclei collisions give rise to a state of the boiling
quark-gluon matter (plasma?). The numerous quark-antiquark pairs with different
colors and masses (energy in the center of mass system of the pair) are
produced there. The color-neutral pairs whose mass fits the Breit-Wigner
shape form the resonance.
However, most pairs in the plasma are in a color-octet state and may not create
resonances. They are considered as a continuum. After interaction with a gluon
or collective excitation in the medium (shown by the dashed line in Fig. 2)
some quarks can change the color and
get excited. If such a quark finds a partner to form a color-neutral pair in the
left wing of the resonance it can emit Cherenkov gluon as allowed by the
chromopermittivity (argued above in classical terms). This gluon transforms to
the quark-antiquark pair. That is how the color-neutral ${\bar q}qg$-component
(or four-quark component) of the resonance can be formed from the initial
colored two-quark state. Thus beside the common color-neutral
${\bar q}q$-component in the low-mass branch of the $\rho$-meson peak there
appears new ${\bar q}qg$-component (as well as others, possibly). Using the
wave-line notation for the Cherenkov gluon shown in Fig.~\ref{diagram} one
gets a tetraquark state which contributes to the left-wing cross section.
\begin{figure}[ht]
\centerline{\includegraphics[width=0.5\textwidth]{diagram1-eps-converted-to.pdf}}
\caption{Origin of the tetraquark state.
One of the quarks of the initially colored pair interacts within the medium
(shown by the dashed line attached to it),
changes its color and gets excited, finds its partner for the color-free bound
state within the low-mass wing of a resonance, gets de-excited by emission of
Cherenkov gluon that leads to the tetraquark states in this wing.}
\label{diagram}
\end{figure}
The distributions of the decay products of the left-wing and right-wing states
of asymmetric heavy resonances created in the process of nucleus-nucleus
collisions can slightly differ. If tetraquarks have additional decay modes
compared to dimers, these modes can serve for identification of tetraquarks.
Then these special modes should reveal themselves only in the low-mass wing of
an asymmetric resonance. It is a distinctive feature of ${\bar q}qg$-Cherenkov
states. That is more probable for heavy resonances because the presence of a
heavy quark in heavy flavor hadrons provides an additional variety of decay
channels with new energy scales. Even though the admixture of new decay
channels is small, their search deserves special attention.
Surely, the emission of Cherenkov gluons asks for interaction of "$\rho $-meson
quarks" within the medium (gluon, collective modes...?). Such an
interaction can result either in continuum or quasistable states. In our case,
the nature of the primary interaction with $\rho $-meson quarks initiating
emission of Cherenkov gluons is however left unknown yet. The main conclusion
is that some nuclear forces initiate sometimes the additional binding
to tetraquarks by emitted Cherenkov gluons of two otherwise independent
unbound color-octet quarks, created in the nuclear medium.
The formation of the weakly bound triple-states (${\bar q}qg$) in the
collisions of three particles when two-particle forces are too weak to produce
bound dimers is known as Efimov effect \cite{efim, ferl}. In all the cases
the role of the third component is crucial for experimental observation of
this effect. It is important to reveal the physics nature of the component.
In simplest models, it was ascribed to light quark pairs produced in vacuum
in case of narrow resonances \cite{len1, len2} and to Cherenkov gluons for
wide resonances \cite{drem}.
\section{Conclusion}
To conclude, we have shown that asymmetry of vector mesons produced in
nuclear collisions can be successfully described as Fano-Feshbach effect
and further interpreted in terms of emission of Cherenkov gluons as a
particular detalization of the quantum interference pattern.
From the theoretical side, models of (collective?) excitations in the nuclear
medium which help to get an insight to this problem are welcome.
From the experimental side, the error bars in experiments with $\rho $-mesons
are quite large. There are some plans to improve experimental accuracy up to
two orders of magnitude (private communication). Very little is still known
about other resonances but the low-mass asymmetry seems universal and gives
some hope for further progress.
The dilepton spectra and, especially, the asymmetry of vector mesons produced
in nuclei collisions deserve further precise experimental studies at
RHIC and LHC (see discussion in \cite{rapp1}).
\medskip
{\bf Acknowledgments}
\medskip
We are grateful for support by the RFBR-grant
14-02-00099 and the RAS-CERN program.
|
\section{Introduction}\label{section_intro}
In recent years, there has been significant progress in the study of the
charmed baryon spectrum, mainly from the Belle and BaBar experiments \citep{Mizuk:2004yu,Chistov:2006zj,
Aubert:2006je,Aubert:2006sp,Abe:2006rz,Aubert:2007dt,Kato:2013ynr,Solovieva:2008fw,Lesiak:2008wz}.
In the charmed strange baryon sector, a number of excited states ($\Xi_c^{\ast}$) have been observed.
Belle reported evidence for two
excited states, $\Xi_c(2980)$ and $\Xi_c(3080)$, in the $\Lambda_{c}^{+}K^{-}\pi^{+}$
and $\Lambda_{c}^{+}K_{S}^{0}\pi^{-}$ final states \cite{Chistov:2006zj}.
These states have been confirmed by BaBar \cite{Aubert:2007dt}. In the same paper,
BaBar also claimed evidence for two resonances, the $\Xi_c(3055)^{+}$ and the $\Xi_c(3123)^{+}$, observed
in the $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$
final states. Recently, Belle confirmed the existence of the $\Xi_c(3055)^{+}$,
but no evidence was found for the $\Xi_c(3123)^{+}$ \cite{Kato:2013ynr}.
As discussed in Refs.~\cite{Cheng:2006dk,Liu:2012sj}, the decay pattern of
charmed baryons provides an important contribution to our understanding of the nature of the states.
To date, all measurements of $\Xi_{c}^{\ast}$ baryons were performed using decays in which
the charm quark is contained in the final-state baryon. Measurements of final states
in which the charm quark is part of the final state meson provide complementary information.
In this paper, we report studies of $\Xi_{c}^{\ast}$ baryons decaying to the $\Lambda D^{+}$ and $\Lambda D^{0}$
final states using a data sample with an integrated luminosity of 980 fb$^{-1}$ collected with the
Belle detector at the KEKB asymmetric-energy $e^{+}e^{-}$ collider. We find significant signals for
$\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$ decays into $\Lambda D^{+}$.
In the $\Lambda D^{0}$ final state, we report observation of the $\Xi_{c}(3055)^{0}$.
These measurements constitute the first observation and evidence for the $\Xi_{c}(3055)$
and $\Xi_{c}(3080)$ into the $\Lambda D$ final states, and the first-ever observation of the $\Xi_{c}(3055)^{0}$.
We also perform a combined analysis of the $\Lambda D^{+}$ , $\Sigma_{c}^{++}K^{-}$, and $\Sigma_{c}^{\ast ++}K^{-}$ final states to measure the ratios of branching fractions and to
improve the accuracy of the mass and width measurements.
The remaining sections of the paper are organized as follows.
In Sections \ref{section_data} and \ref{section_selection}, we describe the data sample and event selections.
In Section \ref{section_signal}, observations and measurements of $\Xi_{c}^{\ast}$ baryons
in the $\Lambda D^{+}$ and $\Lambda D^{0}$ final states are presented.
In Section \ref{section_combine}, the combined analysis with the $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ final states
is presented. Finally, the summary and conclusion are given.
\section{Data samples and the Belle detector}\label{section_data}
We use a data sample with a total integrated luminosity of 980 fb$^{-1}$ recorded with
the Belle detector at the KEKB asymmetric-beam-energy $e^{+}e^{-}$ collider \cite{KEKB}.
The Belle detector is a large-solid-angle magnetic
spectrometer that consists of a silicon vertex detector (SVD),
a 50-layer central drift chamber (CDC), an array of
aerogel threshold Cherenkov counters (ACC),
a barrel-like arrangement of time-of-flight
scintillation counters (TOF), and an electromagnetic calorimeter
comprised of CsI(Tl) crystals (ECL) located inside
a superconducting solenoid coil that provides a 1.5~T
magnetic field. An iron flux-return located outside of
the coil is instrumented to detect $K_L^0$ mesons and to identify
muons. The detector is described in detail elsewhere~\cite{Belle}.
Two inner detector configurations were used. A 2.0-cm radius beampipe
and a 3-layer silicon vertex detector was used for the first sample
of 156 fb$^{-1}$, while a 1.5-cm radius beampipe, a 4-layer
silicon detector and a small-cell inner drift chamber were used to record
the remaining 824 fb$^{-1}$ \cite{svd2}.\par
We use a GEANT-based Monte Carlo (MC) simulation \cite{GEANT} to model the detector response
and its acceptance to obtain the reconstruction efficiency and the mass resolution for the signal.
We re-weight the signal MC sample according to the scaled-momentum $x_{p}=p^{\ast}/p_{\rm max}$ distributions,
based on the measurements in real data, to obtain the correct reconstruction efficiency.
Here, $p^{\ast}$ is the momentum of the $\Xi_{c}^{\ast}$ system
in the center-of-mass frame and $p_{\rm max} = \sqrt{s/4-M^{2}c^{4}}/c$, where $s$ is the total center-of-mass
energy squared, $M$ is the invariant mass of the $\Xi_{c}^{\ast}$ system, and $c$ is the speed of light.
We also use MC events generated with EVTGEN \cite{evtgen} and JETSET \cite{jetset}
to study the mass distribution in the background process $e^{+}e^{-} \to q\bar{q}$ process ($q=u,d,s,c$ and $b$).
\section{Event selection}\label{section_selection}
Our analysis is optimized to search for decays of $\Xi_{c}^{\ast}$ baryons into the $\Lambda D^{+}$ and $\Lambda D^{0}$ final states.
Throughout this paper, the inclusion of the charge-conjugate decay mode is implied
unless otherwise stated.
A $\Lambda$ candidate is reconstructed via its decay into $p \pi^{-}$.
A $D^{+}$ candidate is reconstructed via its decay into $K^{-}\pi^{+}\pi^{+}$.
A $D^{0}$ candidate is reconstructed via its decay into $K^{-}\pi^{+}$, $K^{-}\pi^{+}\pi^{+}\pi^{-}$, and $K^{-}\pi^{+}\pi^{0}$.
The selection of charged hadrons is based on information from the tracking system (SVD and CDC) and
hadron identification system (CDC, ACC, and TOF). The charged hadrons that are not
associated with the $\Lambda$ candidate are required to have a point of closest approach to the interaction point
that is within 2 cm along the $z$ axis and within 0.2 cm in the transverse ($r$-$\phi$) plane.
The $z$ axis is opposite the positron beam direction. For each track, likelihood values $\mathcal{L}_{p}$, $\mathcal{L}_{K}$, and $\mathcal{L}_{\pi}$
are provided by the hadron identification system, based on the ionization losses in the CDC,
the number of detected Cherenkov photons in the ACC, and the time-of-flight measured by the TOF.
The likelihood ratio is defined as $\mathcal{L}(i:j)=\mathcal{L}_{i}/(\mathcal{L}_{i}+\mathcal{L}_{j})$.
A track is identified as a proton if the likelihood ratios $\mathcal{L}(p:\pi)$ and $\mathcal{L}(p:K)$ are greater than 0.6,
as a kaon if the likelihood ratios $\mathcal{L}(K:\pi)$ and $\mathcal{L}(K:p)$ are greater than 0.6, or as a
pion if the likelihood ratios $\mathcal{L}(\pi:K)$ and $\mathcal{L}(\pi:p)$ are greater than 0.6.
In addition, an electron likelihood is provided based on information from the ECL, ACC, and CDC \cite{eid}.
A track with an electron likelihood greater than 0.95 is rejected. \par
The momentum-averaged efficiencies of hadron identification are about 90$\%$, 90$\%$, and 93$\%$ for pions,
kaons, and protons, respectively.
The momentum-averaged probability to misidentify a pion as a kaon is about 9$\%$,
to misidentify a kaon as a pion about 10$\%$, and to
misidentify a pion or kaon as a proton about 5$\%$.
The $\pi^{0}$ candidates are reconstructed from pairs of photons whose invariant mass ($M_{\gamma \gamma}$) satisfies
$120$ MeV/$c^{2}<M_{\gamma \gamma}<150$ MeV/${\it c}$$^{2}$, which corresponds to $\pm 2.5 \sigma$ (where $\sigma$ is the one-standard-deviation of the resolution).
The energy of each photon in the laboratory frame is required to be greater than $50$ MeV
and the energy of the $\pi^{0}$ candidate in the laboratory frame is required to be greater than 500 MeV.
The $D^{+}$ candidates are selected by requiring $|M(K^{-}\pi^{+}\pi^{+})-m_{D^{+}}|<$ 12 MeV/${\it c}$$^{2}$,
where $m_{D^{+}}$ is the $D^{+}$ mass \cite{PDG}.
The $D^{0}$ candidates for each decay mode of the $D^{0}$ are selected by requiring
$|M(K^{-}\pi^{+})-m_{D^{0}}|<$ 14 MeV/${\it c}$$^{2}$,
$|M(K^{-}\pi^{+}\pi^{+}\pi^{-})-m_{D^{0}}|<$ 11 MeV/${\it c}$$^{2}$,
and $|M(K^{-}\pi^{+}\pi^{0})-m_{D^{0}}|<$ 27 MeV/${\it c}$$^{2}$, where $m_{D^{0}}$ is the $D^{0}$ mass.
These mass ranges correspond to $\pm 2.5\sigma$.
To improve the momentum resolution, the daughter particles are fitted to a
common vertex together with an invariant mass constrained to the $D^{+}$ or $D^{0}$ mass.
The $\Lambda$ candidates are selected using cuts on four parameters:
the angular difference between the $\Lambda$ flight direction and
the direction pointing from IP to the decay vertex in the transverse plane; the distance between each track and
the IP in the transverse plane; the distance between the decay vertex and the IP in the transverse plane; and
the displacement along $z$ of the closest-approach points of the two tracks to the beam axis.
Also, the invariant mass of a $\Lambda$ candidate is required to be within 3 MeV/${\it c}$$^{2}$ of the
$\Lambda$ mass, which corresponds to $\pm 3 \sigma$.
Excited charmed baryons are known to be produced with much higher average momenta than the
combinatorial background. We thus require that $x_{p}$ be greater than
0.7 for the $\Lambda D^{+}$ and 0.8 for the $\Lambda D^{0}$ modes.
\section{Observation of $\Xi_{c}t^{\ast}\to\Lambda D$ decays}\label{section_signal}
Figure \ref{mlamd} shows the $\Lambda D$ invariant-mass ($M(\Lambda D)$) distributions
for data after the application of all the selection criteria; signals near 3055
and 3080 MeV/${\it c}$$^{2}$ are seen. We do not observe any such peaks in the
distributions in wrong-sign $\overline{\Lambda} D$ combinations, in data from the $D$ meson mass sideband, nor in MC events that do not include these resonances.
Hereinafter, $\Xi_{c}^{\ast}$ baryons corresponding to these peaks are referred to as $\Xi_c(3055)$ and $\Xi_{c}(3080)$.
In order to evaluate the masses, widths, and statistical significances of the $\Xi_{c}^{\ast}$ states,
we apply an unbinned extended maximum likelihood (UML) fit to the mass spectra in the invariant mass range of 3.0$-$3.2 GeV/$c^{2}$.
For the $\Lambda D^{0}$ mode, the fit is performed simultaneously for the three different $D^{0}$ decay modes,
with their relative yields fixed using the product of their known branching fractions \cite{PDG} and detection efficiencies.
The masses and widths of the $\Xi_{c}^{\ast}$ states are constrained to be the same for all modes.
The detection efficiencies for the $\Xi_{c}(3055)^{0}$ and $\Xi_{c}(3080)^{0}$ are found to exhibit no difference within the statistical
precision of the MC sample, which is smaller than $1\%$. Therefore, we use common efficiency values for these states.
The relative yields are fixed to $K^{-}\pi^{+}$:$K^{-}\pi^{+}\pi^{+}\pi^{-}$:$K^{-}\pi^{+}\pi^{0}$= 1.00:1.30:1.15.
The probability density functions (PDF) for the $\Xi_{c}^{\ast}$ components are represented by convolutions of Breit-Wigner shapes
with Gaussian distributions to take the intrinsic invariant mass resolution, $\sigma_{\rm res}$, into account.
Using the signal MC events, we determine $\sigma_{\rm res}$ for the $\Lambda D^{+}$ mode to be
1.1 MeV/${\it c}$$^{2}$ for the $\Xi_c(3055)^{+}$ and 1.3 MeV/${\it c}$$^{2}$ for the $\Xi_c(3080)^{+}$.
In the $\Lambda D^{0}$ mode, we determine $\sigma_{\rm res}$ to be 1.1 and 2.0 MeV/${\it c}$$^{2}$
for the $D^{0}$ decay mode without and with $\pi^{0}$, respectively for the $\Xi_{c}(3055)^{0}$
and 1.3 and 2.2 MeV/${\it c}$$^{2}$ for the $\Xi_{c}(3080)^{0}$.
The masses, widths and yields of the $\Xi_{c}^{\ast}$ states are treated as free parameters.
A third-order Chebyshev polynomial is used to model the PDF for the combinatorial background.
The statistical significance is evaluated from $-2\ln{(\mathcal{L}_{0}/\mathcal{L})}$,
where $\mathcal{L}_{0}$ ($\mathcal{L}$) is the likelihood for the fit without (with) the signal component. When we evaluate $\mathcal{L}_{0}$ for one of the $\Xi_c^{\ast}$ states, the other $\Xi_c^{\ast}$
state is included in the fit. The $-2\ln{(\mathcal{L}_{0}/\mathcal{L})}$ values are 144.6 for the $\Xi_c(3055)^{+}$, 30.0 for the $\Xi_c(3080)^{+}$,
83.1 for the $\Xi_c(3055)^{0}$, and 6.6 for the $\Xi_c(3080)^{0}$.
By taking into account the change by 3 of the number of degrees of freedom in the UML fit
associated with the inclusion of the $\Xi_c^{\ast}$ states, the statistical significances are
11.7$\sigma$, 4.8$\sigma$, 8.6$\sigma$, and 1.7$\sigma$
for the $\Xi_c(3055)^{+}$, $\Xi_c(3080)^{+}$, $\Xi_c(3055)^{0}$, and $\Xi_c(3080)^{0}$, respectively.
The peak for the $\Xi_c(3080)^{0}$ is not statistically significant.
We estimate the systematic uncertainty of the masses and widths of $\Xi_{c}(3055)^{0}$, $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$
in the $\Lambda D^{+}$ decay mode as the changes produced by giving reasonable variations to the fitting technique.
The stability of the background shape is checked by changing the fit region and background PDF.
The maximum deviation from the nominal fit is taken as the systematic uncertainty.
To check the uncertainty due to $\sigma_{\rm res}$, the ratio $r_{\sigma}=\sigma_{D}^{\rm MC}/ \sigma_{D}^{\rm data}$
is evaluated, where $\sigma_{D}^{\rm MC}$ and $\sigma_{D}^{\rm data}$ are the $D^{0}$ mass resolution for MC and data.
For the $\Lambda D^{0}$ mode, $r_{\sigma}$ is 1.16, 1.16 and 1.08 for the $D^{0}$ final state of
$K^{-}\pi^{+}$, $K^{-}\pi^{+}\pi^{+}\pi^{-}$ and $K^{-}\pi^{+}\pi^{0}$, respectively.
We evaluate the uncertainty by fitting data with $\sigma_{\rm res}$ scaled by 16$\%$ for all the decay modes.
To check the uncertainty on the mass due to a possible mis-calibration of the momentum and energy measurements,
we check the reconstructed $D^{0}$ masses for both data and signal MC.
In each mode, the peak position is observed to have a distinct but small deviation from the world average.
However, these deviations are well reproduced by the MC and, because of the mass-constrained fit,
have little effect on the determination of the masses of the $\Xi_c^{\ast}$ baryons. In the signal MC, the
differences between the input and output masses of the $\Xi_{c}^{\ast}$ baryons is less than 0.1 MeV/$c^{2}$
for all $D^{0}$ decay modes. We assign a systematic uncertainty of 0.1 MeV$/c^{2}$
on the mass measurements. We perform fits that include the interference of the $\Xi_{c}(3055)$ and $\Xi_{c}(3080)$
by introducing the phase between two Breit-Wigner amplitudes.
The systematic uncertainties are summarized in Table \ref{summary_sys_lamd0}.
The fit result for the $\Xi_{c}(3080)^{+}$ width is ($1.4\pm1.8$) MeV, which is consistent with zero.
Therefore, we set a 90$\%$ confidence level upper limit on the width.
We redo the fit by changing the width; the width for which
the likelihood ratio $-2\ln{(\mathcal{L}/\mathcal{L}_{0 \Gamma})}$ is 2.7, where $L_{0 \Gamma}$
is the likelihood with the zero width for $\Xi_{c}(3080)^{+}$, is assigned as the 90$\%$
confidence level upper limit. We obtain the upper limit $\Gamma_{\Xi_{c}(3080)^{+}}<6.3$ MeV.
The measurements of the masses and widths are summarized in Table \ref{summary_mass_lambdad}.
Note that the final values for the $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$ masses and widths
in this paper are those combined with $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ decay modes.
Values in the $\Lambda D$ mode only are shown to compare with other decay modes.
We find that the mass of the $\Xi_{c}(3055)^{+}$ and widths of $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$
are consistent with our previous measurements with the $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ decay modes
\cite{Kato:2013ynr}. However, we find a small inconsistency for the mass of the $\Xi_{c}(3080)^{+}$, which may indicate the
possible underestimation of the systematic uncertainty for the determination of the masses.
We determine the combined value for the masses of the $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$ by taking the weighted average.
The uncertainty is scaled by $\sqrt{\chi^{2}/(N-1)}$, where $N$ is the number of different decay modes, which is 2 for $\Xi_{c}(3055)^{+}$ and 3 for $\Xi_{c}(3080)^{+}$,
if the $\chi^{2}/(N-1)$ is greater than one; this is the recipe used in Ref.~\cite{PDG}.
The scale factor for the $\Xi_c(3055)^{+}$ is 1.0 and that for the $\Xi_c(3080)^{+}$ is 3.3.
The measured mass of the $\Xi_{c}(3055)^{+}$ is $(3055.9\pm0.4)$ MeV/$c^{2}$ and that for $\Xi_{c}(3080)^{+}$ is $(3077.9\pm0.9)$ MeV/$c^{2}$. The combined values for the widths are determined by simultaneous fit
with $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ decay modes as described in the next section.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[scale=0.3]{fig/roofit_mlamdplus_all.eps}
\includegraphics[scale=0.3]{fig/roofit_mlamd0_1.eps}
\includegraphics[scale=0.3]{fig/roofit_mlamd0_2.eps}
\includegraphics[scale=0.3]{fig/roofit_mlamd0_3.eps}
\caption{$M(\Lambda D)$ distributions. Points with statistical error bars are data.
Blue solid lines show the fit results. The red dashed, magenta dotted, and black dashed-dotted lines show
the $\Xi_{c}(3055)$ signal, the $\Xi_{c}(3080)$ signal, and the background components, respectively.
(a) $M(\Lambda D^{+})$ distribution; $M(\Lambda D^{0})$ distributions for the (b) $K^{-}\pi^{+}$,
(c) $K^{-}\pi^{+}\pi^{+}\pi^{-}$, and (d) $K^{-}\pi^{+}\pi^{0}$ $D^{0}$ decay modes.}
\label{mlamd}
\end{center}
\end{figure*}
\begin{center}
\begin{table*}[htbp]
\caption{Systematic uncertainties for the mass (MeV/$c^{2}$) and width (MeV) of the $\Xi_{c}^{\ast}$.}
\begin{tabular}{c|c@{\hspace{0.2cm}}c@{\hspace{0.2cm}}c@{\hspace{0.2cm}}c@{\hspace{0.2cm}}c@{\hspace{0.2cm}}c@{\hspace{0.2cm}}c@{\hspace{0.2cm}}c} \hline \hline
Source & M$_{\Xi_c(3055)^{0}}$ & $\Gamma_{\Xi_c(3055)^{0}}$ & M$_{\Xi_c(3055)^{+}}$ & $\Gamma_{\Xi_c(3055)^{+}}$ & M$_{\Xi_c(3080)^{+}}$\\ \hline
Background shape & 0.6 & 1.0 & 0.1 & 1.5 & 0.0 \\
Resolution & 0.0 & 0.2 & 0.0 & 0.2 & 0.0 \\
Mass scale & 0.1 & 0.0 & 0.1 & 0.0 & 0.1 \\
Interference & 0.1 & 0.3 & 0.1 & 0.1 & 0.1 \\ \hline
Total & 0.6 & 1.1 & 0.2 & 1.5 & 0.1 \\ \hline \hline
\end{tabular}
\label{summary_sys_lamd0}
\end{table*}
\end{center}
\begin{center}
\begin{table*}[htbp]
\caption{Summary of the masses, widths and significances of the $\Xi_{c}^{\ast}$ baryons measured in the $\Lambda D$ modes.
The first error is statistical and the second is systematic.
We set a 90$\%$ confidence level upper limit for the width of $\Xi_c(3080)^{+}$.}
\begin{tabular}{c|ccc} \hline \hline
Resonance & Mass (MeV/$c^{2}$) & Width (MeV) & Significance ($\sigma$) \\ \hline
$\Xi_c(3055)^{0}$ & $3059.0\pm0.5\pm0.6$ & $6.4\pm2.1\pm1.1$ & 8.6 \\
$\Xi_c(3055)^{+}$ & $3055.8\pm0.4\pm0.2$ & $7.0\pm1.2\pm1.5$ & 11.7 \\
$\Xi_c(3080)^{+}$ & $3079.6\pm0.4\pm0.1$ & $<$ 6.3 & 4.8 \\ \hline \hline
\end{tabular}
\label{summary_mass_lambdad}
\end{table*}
\end{center}
\section{Combined analysis with the $\Sigma_{c}^{++} K^{-}$ and $\Sigma_{c}^{\ast ++} K^{-}$ decay modes}\label{section_combine}
We measure the ratio of branching fractions,
${{\cal B}(\Xi_{c}^{\ast +} \to \Lambda D^{+})}/{{\cal B}(\Xi_{c}^{\ast +} \to \Sigma_{c}^{++}K^{-})} \equiv R_{{\cal B} (\Lambda D)}$,
using the following equations:
\begin{eqnarray}
\lefteqn{R_{{\cal B} (\Lambda D)}=R_{{\rm yield} (\Lambda D)}\times({\cal B} \times \epsilon)_{\Sigma_{c}K}/({\cal B} \times \epsilon)_{\Lambda D^{+}}}, \\
\lefteqn{({\cal B} \times \epsilon)_{\Lambda D^{+}}={\cal B}(D^{+} \to K^{-}\pi^{+}\pi^{+}) } \nonumber \\
\lefteqn{\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\: \:\:\:\:\:\:\: \times {\cal B}(\Lambda \to p \pi^{-}) \times \epsilon(\Lambda D^{+}) } \\
\lefteqn{({\cal B} \times \epsilon)_{\Sigma_{c}K}={\cal B}(\Lambda_{c}^{+}\to pK^{-}\pi^{+}) \times[ \epsilon_{p K^{-}\pi^{+}} + R_{pK^{0}_{S}} } \nonumber \\
\lefteqn{\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\: \:\:\:\:\: \:\:\:\:\: \times {\cal B}(K^{0}_{S} \to \pi^{+}\pi^{-}) \times \epsilon_{p K^{0}_{S}} ], }
\end{eqnarray}
where $\epsilon(\Lambda D^{+})$ is the reconstruction efficiency for the $\Lambda D^{+}$ mode, $\epsilon_{i}$ is the
reconstruction efficiency for the $\Sigma_{c}^{++}K^{-}$ mode with the $i^{\rm th}$ sub-decay of the $\Lambda_{c}^{+}$,
$R_{pK^{0}_{S}}$ is the ratio of branching fraction ${\cal B}( \Lambda_{c}^{+} \to p K^{0}_{S})/{\cal B}( \Lambda_{c}^{+} \to p K^{-} \pi^{+})$
and $R_{{\rm yield} (\Lambda D)}$ is the ratio of the yields of $\Xi_{c}^{\ast}$ baryons in the $\Lambda D^{+}$ and the $\Sigma_{c}^{++}K^{-}$ modes.
For ${\cal B}(\Lambda_{c}^{+}\to pK^{-}\pi^{+})$, we use the latest Belle measurement \cite{Zupanc:2013iki}.
Other branching fraction values are taken from Ref.~\cite{PDG}.
We also measure the ratio of branching fractions,
${{\cal B}(\Xi_{c}(3080)^{+} \to \Sigma_c^{\ast ++} K^{-})}/{{\cal B}(\Xi_{c}(3080)^{+} \to \Sigma_{c}^{++}K^{-})}$ $\equiv R_{{\cal B}\Sigma_{c}^{\ast} K}$
using the equation
\begin{eqnarray}
\lefteqn{R_{{\cal B} (\Sigma_{c}^{\ast} K)}=R_{{\rm yield} (\Sigma_{c}^{\ast} K)}\times({\cal B} \times \epsilon)_{\Sigma_{c}K}/({\cal B} \times \epsilon)_{\Sigma_{c}^{\ast} K}},
\end{eqnarray}
where $R_{{\rm yield} (\Sigma_{c}^{\ast} K)}$ is the ratio of yields of $\Xi_{c}(3080)^{+}$ in the $\Sigma_c^{\ast ++} K^{-}$ decay mode and $\Sigma_{c}^{++}K^{-}$ decay modes.
$({\cal B} \times \epsilon)_{\Sigma_{c}^{\ast} K}$ shares the form of Eq. (3) for
$({\cal B} \times \epsilon)_{\Sigma_{c}K}$ after replacing the reconstruction efficiency
for $\Sigma_{c}^{\ast ++}K^{-}$ with that for $\Sigma_{c}^{++}K^{-}$.
The data set used for the $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ decay modes is the same as that for the $\Lambda D^{+}$ mode.
Event selections are the same as those in Ref.~\cite{Kato:2013ynr}.
A $\Sigma_{c}^{++}$ or $\Sigma_{c}^{\ast ++}$ candidate is reconstructed via its decay into $\Lambda_{c}^{+} \pi^{+}$; the $\Lambda_{c}^{+}$ candidate here is reconstructed via its decay into
$p K^{-} \pi^{+}$ and $p K^{0}_{S}$. Note that the requirement $x_{p}>0.7$ is the same as that for the $\Lambda D^{+}$ mode
and so it is possible to directly compare the three decay modes.
To obtain $R_{\rm yield}$ and to measure the width of the $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$ with greater
accuracy than is possible using a single decay mode, we perform a simultaneous UML fit with the widths of the
$\Xi_{c}^{\ast}$ states constrained to be the same among the three decay modes, as discussed in the previous section.
The masses are not constrained because we find inconsistency for the mass of the $\Xi_{c}(3080)^{+}$ among the three decay modes.
We also fit the mass distribution of the $\Sigma_{c}^{++}$ sideband region, defined as
$|M(\Lambda_c^{+}\pi^{+})-(m_{\Sigma_c^{++}} \pm 15 $ MeV/${\it c}$$^{2}$$)| \: <$ 5 MeV/${\it c}$$^{2}$,
where $m_{\Sigma_c^{++}}$ is the $\Sigma_c^{++}$ mass, to subtract the contribution from non-resonant $\Lambda_{c}^{+} K^{-} \pi^{+}$ decays
in the signal region.
We subtract half of the yield found in the sideband regions because the mass range of the sideband region is double
the width of the $\Sigma_{c}^{++}$ signal region.
It is difficult to define the $\Sigma_{c}^{\ast ++}$ sideband regions because the maximum
mass that is possible for combinations to contribute to the $\Xi_{c}(3080)^{+}$ is only slightly higher
than the $\Sigma_{c}^{\ast ++}$ mass, and a low mass sideband would overlap with the $\Sigma_{c}^{++}$ region.
Thus, we estimate the contribution under the $\Sigma_{c}^{\ast ++}$ by scaling the yield in the
$\Sigma_{c}^{++}$ sideband regions by 2.9, a factor estimated using signal MC.
We assume no interference between $\Sigma_{c}^{++} K^{-}$ or $\Sigma_{c}^{\ast ++} K^{-}$ with non-resonant $\Lambda_{c}^{+}K^{-}\pi^{+}$.
The PDFs and fit region for the $\Lambda D^{+}$ are the same as those described in Section \ref{section_signal}.
The fit conditions for the $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ modes are the same as in Ref.~\cite{Kato:2013ynr}.
For the fit to the events from the $\Sigma_{c}^{++}$ sideband region, we use the $3.0-3.2$ ${\rm GeV}/c^{2}$ $\Sigma_{c}^{++}K^{-}$ mass range.
The $\Xi_{c}^{\ast}$ contributions are represented by a Gaussian-convolved Breit-Wigner
with the same mass resolution of the $\Xi_{c}^{\ast}$ states as that used for the $\Sigma_{c}^{++}$ signal region.
The combinatorial background is represented by a second-order Chebyshev polynomial.
Figure \ref{simultaneousfit} shows the results of the simultaneous fit.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[scale=0.3]{fig/roofit_simu_mlamdplus.eps}
\includegraphics[scale=0.3]{fig/roofit_simu_msk.eps}
\includegraphics[scale=0.3]{fig/roofit_simu_msk_side.eps}
\includegraphics[scale=0.3]{fig/roofit_simu_msks.eps}
\caption{The simultaneous fit results. Points with error bars are data.
The blue solid lines show the fit result. The red dashed, magenta dotted, green dotted, and black dash-dotted lines
show the contributions from the $\Xi_{c}(3055)^{+}$, $\Xi_{c}(3080)^{+}$, $\Xi_{c}(2980)^{+}$, and background, respectively.
(a) $M(\Lambda D^{+}$), (b) $M(\Sigma_{c}^{++}K^{-}$), (c) $M(\Lambda_{c}^{+}K^{-} \pi^{+}$)
for the $\Sigma_{c}^{++}$ sideband region, and (d) $M(\Sigma_{c}^{\ast ++}K^{-})$.}
\label{simultaneousfit}
\end{center}
\end{figure*}
The following systematic uncertainties are taken into account for the combined analysis
for the measurements of the ratios of branching fractions and width.
The systematic uncertainty due to the pion- and kaon-identification efficiency
is estimated from the ratio of the yields of the
$D^{\ast +} \to D^{0}\pi^{+}$, $D^{0}\to K^{-}\pi^{+}$ with and without
the pion- and kaon-identification requirements for data and MC.
The difference of the ratio between data and MC is used to correct the efficiency
and the statistical error of this correction is treated as the systematic uncertainty.
We conservatively assume no correlation in the systematic uncertainty for pion and kaon identification between
$\Lambda D^{+}$ and $\Sigma_{c}^{++}K^{-}$ decay modes as the momentum ranges for these
decay modes are distinct; the systematic uncertainty for
$\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$ cancel.
The systematic uncertainty due to the efficiency of proton identification
is determined using the ratio of the yields of the
$\Lambda \to p \pi^{-}$ with and without the proton identification requirement.
The difference of the ratio between data and MC is used to correct the efficiency
and the statistical uncertainty of this correction is regarded as the systematic uncertainty.
The systematic uncertainty due to the reconstruction efficiency of the $\Lambda$ is determined using the yield ratio of
$B \to \Lambda \bar{\Lambda} K^{+}$ with and without the $\Lambda$ selection cut as a function of momenta of $\Lambda$.
By taking the weighted average of the momentum, it is estimated to be 3$\%$.
The uncertainties of the branching fractions \cite{PDG,Zupanc:2013iki} are included as systematic uncertainties.
The stability of the background shape is checked by changing the fit region and background PDF.
The maximum deviation from the nominal fit among the various changes is regarded as the systematic uncertainty.
To assess the uncertainty due to $\sigma_{\rm res}$, $r_{\sigma}$ is evaluated as $\sigma_{D}^{MC}/ \sigma_{D}^{data}=1.15$
for the $\Lambda D^{+}$ mode and $\sigma_{\Lambda_{c}^{+}}^{MC}/\sigma_{\Lambda_{c}^{+}}^{data}=1.08$
for the $\Sigma_{c}^{++}K^{-}$; we perform a fit with $\sigma_{\rm res}$ scaled by a factor of $r_{\sigma}$
and use the difference of the result from the nominal fit as the systematic uncertainty.
To check the uncertainty due to a possible mis-calibration of momentum and energy measurements,
we evaluate the difference between the reconstructed and nominal $D^{+}$ and $\Lambda_{c}^{+}$ masses for both data and MC.
In data, the reconstructed $D^{+}$ mass differs from the world average \cite{PDG} by 0.1 MeV/$c^{2}$
whereas, in the MC, the $D^{+}$ mass differs by 0.2 MeV/$c^2$. No deviation is observed for $\Lambda_{c}^{+}$
for both data and MC. In the signal MC, the difference of the input and output $\Xi_{c}^{\ast}$ masses
in the $\Lambda D^{+}$ mode is 0.1 MeV/$c^{2}$, which is smaller than the deviation observed in the $D^{+}$ mass because of the
mass-constrained fit. We conservatively assign the systematic uncertainty of 0.1 MeV/$c^{2}$ on the mass measurement.
Table \ref{summary_sys_rel} summarizes the systematic uncertainties.
Table \ref{summary_simfit} summarizes the measurement of yields and widths of the $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$ and
Table \ref{summary_branch_ratio} summarizes the values related to the ratio of branching fractions measurements.
\begin{center}
\begin{table*}[htbp]
\caption{Summary of the systematic uncertainties for the width (MeV) and ratio of branching fraction ratios ($\%$) measurements from the combined analysis.}
\begin{tabular}{c|ccccc} \hline \hline
Source & $\Gamma_{\Xi_c(3055)^{+}}$ & $R_{{\cal B} (\Lambda D)}$ for $\Xi_c(3055)^{+}$ & $\Gamma_{\Xi_c(3080)^{+}}$ & $R_{{\cal B} (\Lambda D)}$ for $\Xi_c(3080)^{+}$ & $R_{{\cal B} (\Sigma_c^{\ast} K)}$\\ \hline
$\pi K p$ identification &- &1.4 &- &1.4 & - \\
$\Lambda$ identification &- &3.0 &- &3.0 & - \\
Branching fractions &- &5.7 &- &5.7 & - \\
Background shape &1.5 &13.1 &0.4 &9.7 & 1.0 \\
Resolution &0.2 &2.1 &0.2 &1.6 & 0.5 \\
Mass scale &0.0 &0.0 &0.0 &0.0 & 0.0 \\ \hline
Total &1.5 &14.9 &0.4 &12.0 & 1.1 \\ \hline \hline
\end{tabular}
\label{summary_sys_rel}
\end{table*}
\end{center}
\begin{center}
\begin{table*}[htbp]
\caption{Summary of results from the simultaneous fits to the $\Lambda D^{+}$ and $\Sigma_{c}^{++}K^{-}$ modes.}
\begin{tabular}{c|ccccc} \hline \hline
Resonance & Width (MeV) & Yield for $\Lambda D^{+}$ & Yield for $\Sigma_{c}^{++}K^{-}$ & Yield for sideband & Yield for $\Sigma_{c}^{\ast ++}$ \\ \hline
$\Xi_{c}(3055)^{+}$ & $7.8\pm1.2\pm1.5$ & $721\pm90$ & $173\pm30$ & $21\pm18$ & -\\
$\Xi_{c}(3080)^{+}$ & $3.0\pm0.7\pm0.4$ & $186\pm40$ & $176\pm23$ & $20\pm12$ & $234\pm30$\\ \hline \hline
\end{tabular}
\label{summary_simfit}
\end{table*}
\end{center}
\begin{center}
\begin{table*}[htbp]
\caption{Summary of the values related to the measurements of the ratio of branching fractions.
The branching fraction values are taken from Ref.~\cite{PDG,Zupanc:2013iki}.
For the ratios of branching fractions, the first error is statistical and second is systematic.}
\begin{tabular}{c|c} \hline \hline
Variable & Value \\ \\ \hline
${\cal B}(D^{+} \to K^{-}\pi^{+}\pi^{+})$ & $0.0913\pm0.0019$ \\
${\cal B}(\Lambda \to p \pi^{-})$ & $0.639\pm0.005$ \\
${\cal B}(\Lambda_{c}^{+}\to pK^{-}\pi^{+})$ & $0.0684\pm0.036$ \\
${\cal B}(K^{0}_{S}\to\pi^{+}\pi^{-})$ & $0.6920\pm0.0005$ \\
${\cal B}(\Lambda_{c}^{+}\to p K^{0}_{S})/{\cal B}(\Lambda_{c}^{+}\to pK^{-}\pi^{+})$ & $0.24\pm0.02$ \\
$\epsilon(\Lambda D^{+})$ & $0.1771$ \\
$\epsilon_{p K^{-}\pi^{+}}$ ($\Sigma_{c}^{++}K^{-}$) & $0.149$ \\
$\epsilon_{p K^{0}_{S}}$ ($\Sigma_{c}^{++}K^{-}$) & $0.155$ \\
$\epsilon_{p K^{-}\pi^{+}}$ ($\Sigma_{c}^{\ast ++}K^{-}$) & $0.146$ \\
$\epsilon_{p K^{0}_{S}}$ ($\Sigma_{c}^{\ast ++}K^{-}$) & $0.153$ \\
$({\cal B} \times \epsilon)_{\Lambda D^{+}}$ & $0.0103$ \\
$({\cal B} \times \epsilon)_{\Sigma_{c}K}$ & $0.0119$ \\
$({\cal B} \times \epsilon)_{\Sigma_{c}^{\ast}K}$ & $0.0117$ \\
$R_{{\rm yield} (\Lambda D)}$ for $\Xi_c(3055)^{+}$ & $4.41\pm0.87$ \\
$R_{{\rm yield} (\Lambda D)}$ for $\Xi_c(3080)^{+}$ & $1.12\pm0.26$ \\
$R_{{\rm yield} (\Sigma_{c}^{\ast} K)}$ & $1.05\pm0.27$ \\
$R_{{\cal B} (\Lambda D)}$ for $\Xi_c(3055)^{+}$ & $5.09\pm1.01\pm0.76$ \\
$R_{{\cal B} (\Lambda D)}$ for $\Xi_c(3080)^{+}$ & $1.29\pm0.30\pm0.15$ \\
$R_{{\cal B} (\Sigma_{c}^{\ast}K)}$ & $1.07\pm0.27\pm0.01$ \\ \hline \hline
\end{tabular}
\label{summary_branch_ratio}
\end{table*}
\end{center}
\section{Summary and conclusions}\label{section_conclusion}
We present studies of $\Xi_{c}^{\ast}$ baryons decaying into the $\Lambda D^{+}$ and $\Lambda D^{0}$ final states.
We report the first observation of the $\Xi_{c}(3055)^{0}$ in the $\Lambda D^{0}$ mode with a significance of $8.6 \sigma$.
The mass and width of the $\Xi_{c}(3055)^{0}$ are measured to be (3059.0$\pm$0.5$\pm$0.6) MeV/$c^{2}$ and (6.4$\pm$2.1$\pm$1.1) MeV,
respectively. We report the first observation of the $\Xi_{c}(3055)^{+}$ decay and evidence for the $\Xi_{c}(3080)^{+}$
in the $\Lambda D^{+}$ final state.
The mass and width of the $\Xi_{c}(3055)^{+}$ obtained from the $\Lambda D$ final states only are
($3055.8\pm0.4\pm0.2$) MeV/$c^{2}$ and $(7.0\pm1.2\pm1.5)$ MeV, respectively, and those for $\Xi_c(3080)^{+}$ are
($3079.6\pm0.4\pm0.1$) MeV/$c^{2}$ and $<$ 6.3 MeV, respectively.
The measured values for $\Xi_{c}(3055)^{+}$ are more accurate than the world average thanks to the high statistics in this decay mode.
We perform a combined analysis of these particles by comparing their decays into $\Lambda D^{+}$ with those into $\Sigma_{c}^{++}K^{-}$ and $\Sigma_{c}^{\ast ++}K^{-}$.
We measure the ratios of branching fractions ${\cal B}(\Xi_{c}(3055)^{+} \to \Lambda D^{+})/{\cal B}(\Xi_{c}(3055)^{+} \to \Sigma_{c}^{++}K^{-})=5.09\pm1.01\pm0.76$,
${\cal B}(\Xi_{c}(3080)^{+} \to \Lambda D^{+})/{\cal B}(\Xi_{c}(3080)^{+} \to \Sigma_{c}^{++}K^{-})=1.29\pm0.30\pm0.15$,
and ${\cal B}(\Xi_{c}(3080)^{+} \to \Sigma_{c}^{\ast ++}K^{-} )/{\cal B}(\Xi_{c}(3080)^{+} \to \Sigma_{c}^{++}K^{-})=1.07\pm0.27\pm0.01$.
The width of the $\Xi_{c}(3055)^{+}$ is $(7.8\pm1.2\pm1.5)$ MeV and that of the $\Xi_{c}(3080)^{+}$ is $(3.0\pm0.7\pm0.4)$ MeV.
We take the weighted average of the measurements in the different decay modes to find the masses of the $\Xi_{c}(3055)^{+}$ and $\Xi_{c}(3080)^{+}$
to be $(3055.9\pm0.4)$ MeV/$c^{2}$ and $(3077.9\pm0.9)$ MeV/$c^{2}$, respectively,
where the uncertainties are scaled by $\sqrt{\chi^{2}/(N-1)}$ to account for small inconsistencies in the $N$ individual measurements.
The uncertainties on the masses incorporate the statistical and systematic values.
The masses and widths of $\Xi_{c}(3055)^{+}$ and $\Xi_{c }(3080)^{+}$, after combining other decay modes,
supersede our previous measurements \cite{Kato:2013ynr}.
Our measurements provide information on the nature of these baryons.
For instance, the chiral quark model has been used to identify the $\Xi_c(3055)$ as the D-wave excitation in the
N=2 shell, and predicts ${\cal B} (\Xi_{c}(3055) \to \Sigma_c \bar{K})$:${\cal B} (\Xi_{c}(3055) \to \Lambda D)$ to be 2.3:0.1 or 5.6:0.0,
depending on the possible excitation modes \cite{Liu:2012sj}.
It further identifies the $\Xi_c(3080)$ as an S-wave excitation mode of the $\Xi_{c}$ in N=2 shell
and predicts that its decay into $\Lambda D$ is forbidden.
Both of these predictions are in contradiction with our measurements.
Further experimental and theoretical work is needed to understand these baryons.
\acknowledgments
We thank the KEKB group for the excellent operation of the
accelerator; the KEK cryogenics group for the efficient
operation of the solenoid; and the KEK computer group,
the National Institute of Informatics, and the
PNNL/EMSL computing group for valuable computing
and SINET4 network support. We acknowledge support from
the Ministry of Education, Culture, Sports, Science, and
Technology (MEXT) of Japan, the Japan Society for the
Promotion of Science (JSPS), and the Tau-Lepton Physics
Research Center of Nagoya University;
the Australian Research Council;
Austrian Science Fund under Grant No.~P 22742-N16 and P 26794-N20;
the National Natural Science Foundation of China under Contracts
No.~10575109, No.~10775142, No.~10875115, No.~11175187, No.~11475187
and No.~11575017;
the Chinese Academy of Science Center for Excellence in Particle Physics;
the Ministry of Education, Youth and Sports of the Czech
Republic under Contract No.~LG14034;
the Carl Zeiss Foundation, the Deutsche Forschungsgemeinschaft, the
Excellence Cluster Universe, and the VolkswagenStiftung;
the Department of Science and Technology of India;
the Istituto Nazionale di Fisica Nucleare of Italy;
the WCU program of the Ministry of Education, National Research Foundation (NRF)
of Korea Grants No.~2011-0029457, No.~2012-0008143,
No.~2012R1A1A2008330, No.~2013R1A1A3007772, No.~2014R1A2A2A01005286,
No.~2014R1A2A2A01002734, No.~2015R1A2A2A01003280 , No. 2015H1A2A1033649;
the Basic Research Lab program under NRF Grant No.~KRF-2011-0020333,
Center for Korean J-PARC Users, No.~NRF-2013K1A3A7A06056592;
the Brain Korea 21-Plus program and Radiation Science Research Institute;
the Polish Ministry of Science and Higher Education and
the National Science Center;
the Ministry of Education and Science of the Russian Federation and
the Russian Foundation for Basic Research;
the Slovenian Research Agency;
Ikerbasque, Basque Foundation for Science and
the Euskal Herriko Unibertsitatea (UPV/EHU) under program UFI 11/55 (Spain);
the Swiss National Science Foundation;
the Ministry of Education and the Ministry of Science and Technology of Taiwan;
and the U.S.\ Department of Energy and the National Science Foundation.
This work is supported by a Grant-in-Aid for Scientific Research (S) ''Probing New Physics with Tau-Lepton'' (No.26220706),
Grant-in-Aid for Scientific Research on Innovative Areas ''Elucidation of New Hadrons with a Variety of Flavors'',
Grant-in-Aid from MEXT for Science Research in a Priority Area (``New Development of Flavor Physics'')
and from JSPS for Creative Scientific Research (``Evolution of Tau-lepton Physics'').
|
\chapter{Introduction}\label{chap:intro}
These notes form the second part of a detailed account of the theory of nilspaces developed by Camarena and Szegedy in the paper \cite{CamSzeg}. The material in these notes expands on the third chapter of their paper, providing more detailed proofs of the main results. To that end we also include several additional results that are implicit in \cite{CamSzeg}. This material relies strongly on the first part of our exposition, given in \cite{Cand:Notes1}. We shall have to assume some familiarity with the basic theory of nilspaces and some of their algebraic properties, but we shall always refer to the relevant results in \cite{Cand:Notes1}.
\smallskip
Recall the notion of a nilspace from \cite[Definition 1.2.1]{Cand:Notes1}. The principal objects treated in these notes are \emph{compact nilspaces}. These are nilspaces equipped with a compact topology that is compatible with the cube structure. To give the precise definition, we adopt the following terminology.
\begin{defn}\label{def:compspace}
Throughout the sequel, to be concise we shall use the term \emph{compact space} to mean a compact, Hausdorff, second-countable topological space.
\end{defn}
\noindent In particular, compact spaces in these notes are Polish spaces \cite[Theorem 5.3]{Ke}. More generally, topological spaces will usually be assumed to be Hausdorff and second-countable.
Recall from \cite[Definition 1.2.3]{Cand:Notes1} the notion of a cubespace.
\begin{defn}\label{def:compnils}
A cubespace $\ns$ is called a \emph{compact cubespace} if $\ns$ is a compact space and $\cu^n(\ns)$ is a closed subset of $\ns^{\{0,1\}^n}$ for every $n\in \mathbb{N}$. A \emph{compact nilspace} is a compact cubespace satisfying the ergodicity and corner-completion axioms from \cite[Definition 1.2.1]{Cand:Notes1}.
\end{defn}
\noindent The main goal in the sequel is to characterize compact nilspaces.
A central result in this direction is a description of a general compact nilspace as an inverse limit of simpler spaces, namely compact nilspaces of finite rank (these are defined in Section \ref{sec:CFRdef}). This result is treated in Section \ref{sec:invlim} (see Theorem \ref{thm:invlim}).
Another central result concerns finite-rank compact nilspaces, and involves \emph{nilmanifolds}, i.e. compact homogeneous spaces of nilpotent Lie groups, the study of which goes back to \cite{Malcev}. To situate the result, let us note that nilmanifolds provide important examples of compact nilspaces. Indeed, a nilmanifold becomes a compact nilspace when, given a filtration on the Lie group (with certain topological properties), the nilmanifold is equipped with the natural cubes associated with the filtration, namely the cubes introduced by Host and Kra \cite{HK,HKparas}. This construction is treated from a purely algebraic viewpoint in \cite[Section 2.3]{Cand:Notes1}, and its basic topological aspects are detailed in Section \ref{sec:filnilmcompns} below. The central result in question here goes in the converse direction, and states that if a compact nilspace of finite rank has connected structure groups,\footnote{For the notion of the structure groups of a nilspace, recall \cite[Definition 3.2.17]{Cand:Notes1} and \cite[Theorem 3.2.19]{Cand:Notes1}.} then it is isomorphic to a nilmanifold with a cube structure of the kind mentioned above. This is treated in Section \ref{sec:CFRnilsnilm} (see Theorem \ref{thm:toralnilspace}).
The work toward these main theorems yields several other results of inherent interest, including the following: endowing every compact nilspace with a Borel probability measure that generalizes the Haar measure on compact abelian groups (Proposition \ref{prop:nilspaceHaar}); an automatic continuity result for Borel morphisms between compact nilspaces (Theorem \ref{thm:Klepgen}); a rigidity result for morphisms into compact nilspaces of finite rank (Theorem \ref{thm:rigidity}). Several of the main tools used to obtain these results rely on the theory of \emph{continuous systems of measures}, which provides a natural framework for measure theoretic aspects of compact nilspaces. This is detailed in Section \ref{sec:measprel}.
An alternative treatment of compact nilspaces is given by Gutman, Manners, and Varj\'u in the series of papers \cite{GMV1,GMV2,GMV3}.
Let us end this introduction by evoking one of the central motivations for the study of compact nilspaces. This motivation concerns the analysis of \emph{uniformity norms}. These norms were introduced in arithmetic combinatorics by Gowers \cite{GSz}, and independently in ergodic theory (as the analogous \emph{uniformity seminorms}) by Host and Kra \cite{HK}. The uniformity norms, or $U^d$ norms (one such norm for each integer $d\geq 2$), are defined on the space of complex-valued functions on a finite (or more generally a compact) abelian group. These norms provide useful tools to control averages of bounded functions over certain linear configurations in abelian groups, configurations such as arithmetic progressions (such control can be obtained via the so-called \emph{generalized Von Neumann theorems}); see \cite{GTarith} and \cite{GTlin}. A major topic concerning these norms is the analysis, for each $d\geq 2$, of the harmonics that determine whether the $U^d$ norm of a function is small or large. For the $U^2$ norm these characteristic harmonics are simply the characters from Fourier analysis. For $d>2$, this topic leads to the theory of \emph{higher order Fourier analysis}. Central to this theory are the results known as \emph{inverse theorems} for the $U^d$ norms, proved by Green, Tao and Ziegler \cite{GTZ}, and independently by Szegedy \cite{Szegedy:HFA}. Essentially, these theorems tell us that, for the $U^d$ norm of functions on $\Zmod{N}$, one can use as characteristic harmonics the functions known as \emph{$(d-1)$-step nilsequences}. In the approach to these theorems developed by Szegedy (consisting principally in \cite{Szegedy:HFA} and his work with Camarena \cite{CamSzeg}), compact nilspaces play a fundamental role. Indeed, roughly speaking, in this approach a nilsequence on $\Zmod{N}$ is obtained as a composition $F\co\varphi$, where $\varphi$ is a nilspace morphism from $\Zmod{N}$ to a compact nilspace $\ns$, and $F$ is some continuous function $\ns\to\mathbb{C}$. In this way, compact nilspaces play a role in higher order Fourier analysis that generalizes the role played by the circle group $\mathbb{R}/\mathbb{Z}$ in the definition of characters in Fourier analysis. In these notes we shall concentrate on the theory of compact nilspaces in itself. For further introduction to higher order Fourier analysis and its applications, we refer the reader to the survey \cite{GHFA}.
\vspace{3cm}
\noindent \textbf{Acknowledgements.} I am very grateful to Bal\'azs Szegedy for discussions that were crucial for my understanding of \cite{CamSzeg}, and to anonymous referees for advice that helped to improve this paper. I also thank Yonatan Gutman, Frederick Manners and P\'eter Varj\'u for informing me of their work prior to publication, and for useful comments. The present work was supported by the ERC Consolidator Grant No. 617747.
\newpage
\section{Filtered nilmanifolds as compact nilspaces}\label{sec:filnilmcompns}
\medskip
Recall from \cite[Section 2.2]{Cand:Notes1} the notion of a filtration $G_\bullet$ on a group $G$ and the basic properties of cube sets of the form $\cu^n(G_\bullet)$.
\begin{defn}\label{def:nilmani}
A \emph{nilmanifold} is a quotient space $G/\Gamma$ where $G$ is a connected nilpotent Lie group and $\Gamma$ is a discrete, cocompact subgroup. If $G_\bullet$ is a filtration on $G$ of degree at most $k$, with terms $G_i$ being closed subgroups of $G$, and with each subgroup $\Gamma \cap G_i$ being cocompact in $G_i$, then we call $(G/\Gamma, G_\bullet)$ a \emph{filtered nilmanifold} of degree at most $k$.
\end{defn}
\noindent The groups in the filtration $G_\bullet$ are sometimes required to be connected (in particular in applications of nilmanifolds in arithmetic combinatorics; see for instance \cite[Definition 1.3]{GTarith}), but we shall not need this assumption here.
A compact nilspace $\ns$ is \emph{connected} if the underlying topological space is connected. We now show that every filtered nilmanifold of degree $k$ can be viewed as a $k$-step connected compact nilspace, in the following sense.
\begin{proposition}\label{prop:nilmanilspace}
Let $G$ be a connected nilpotent Lie group with a degree-$k$ filtration $G_\bullet$ of closed subgroups, let $\Gamma\leq G$ be discrete, and for each $i\in [k]=\{1,2,\dots,k\}$ let $\Gamma_i=\Gamma \cap G_i$. Let $\ns$ denote the coset space $G/\Gamma$ equipped with the sets $\cu^n(\ns)=\{\pi_\Gamma\co \q: \q\in \cu^n(G_\bullet)\}$, where $\pi_\Gamma$ is the quotient map $G\to G/\Gamma$. Then $\ns$ is a connected compact nilspace if and only if $\Gamma_i$ is cocompact in $G_i$ for each $i\in [k]$.
\end{proposition}
\begin{proof}
The nilspace axioms are proved for $\ns$ in \cite[Proposition 2.3.1]{Cand:Notes1}. To complete the proof, we show that the compactness of each $G_i/\Gamma_i$ is necessary and sufficient for each set $\cu^n(\ns)$ to be closed in $\ns^{\{0,1\}^n}$.\\
\indent To see the sufficiency, it is enough to deduce that $\cu^n(G_\bullet)\cap \Gamma^{\{0,1\}^n}$ is cocompact in $\cu^n(G_\bullet)$, as then $\cu^n(\ns)$ is homeomorphic to the compact set $\cu^n(G_\bullet)/ \big(\cu^n(G_\bullet)\cap \Gamma^{\{0,1\}^n}\big)$ (see \cite[Proposition 2.7]{Leib}). This cocompactness is proved elsewhere in the literature but we include a proof here for completeness, following the argument from \cite[Lemma E.10]{GTlin}. By Definition \ref{def:nilmani} there is a compact set $K_j \subseteq G_j$ such that the product set $K_j \,\Gamma_j$ equals $G_j$. Recall from \cite[Lemma 2.2.5]{Cand:Notes1} that every cube $\q\in \cu^n(G_\bullet)$ has a unique factorization $\q= g_0^{F_0}g_1^{F_1}\cdots g_{2^n-1}^{F_{2^n-1}}$
with $g_i\in G_{\codim(F_i)}$ for each $i$, where the $F_i$ are the upper faces of $\{0,1\}^n$, of the form $F_i=F(v_i)=\{v:\supp(v)\supset \supp(v_i)\}$ and the $v_i$ are ordered in the colex order. We then have the following inductive formula for the coefficients $g_i$:
\begin{equation}\label{eq:compcubecoeffs}
g_0=\q(v_0),\text{ and for each }i>0,\; g_i= \Big( \prod_{ j < i : \; \supp(v_j)\subset\supp(v_i)} g_j^{F_j}\Big)^{-1}\cdot \q \,(v_i).
\end{equation}
Let $H_0=\cu^n(G_\bullet)$ and for each $i\in [2^n]$ let $H_i=\{\q\in\cu^n(G_{\bullet}): g_0= \dots = g_{i-1} = \mathrm{id}_G\}$. It follows from \eqref{eq:compcubecoeffs} that $\q\in H_i$ if and only if $\q(v_j)=\mathrm{id}_G$ for all $j<i$. We deduce that $H_i$ is a normal subgroup of $\cu^n(G_\bullet)$. We shall now use these subgroups to show that every $\q\in \cu^n(G_\bullet)$ is a product of an element of $\cu^n(G_\bullet)\cap \Gamma^{\{0,1\}^n}$ with an element of $K:=\{ k_0^{F_0} \dots k_{2^n-1}^{F_{2^n-1}} : k_i \in K_i\}$. This will prove the claimed sufficiency, since the set $K$ is compact. Let $i \in [2^n]$ and suppose that $\q \in H_{i-1}$. Then $\q = g_i^{F_i}\,\q'$, where $\q' \in H_i$. We then have $g_i^{F_i}=(k_i \gamma_i)^{F_i}$ where $k_i \in K_{\codim(F_i)}$ and $\gamma_i \in \Gamma_{\codim(F_i)}$. Since $H_i$ is normal, we have
$\q = k_i^{F_i}\; \q'' \; \gamma_i^{F_i}$ with $\q''\in H_i$. Starting with any cube in $H_0$ and repeating this procedure, we eventually obtain the desired factorization showing that $\q\in K\cdot\big(\cu^n(G_\bullet)\cap \Gamma^{\{0,1\}^n}\big)$.\\
\indent To see the necessity of each set $G_i/\Gamma_i$ being compact, let $Q_i$ denote the set of cubes $\q\in \cu^i(\ns)$ such that $\q(v)$ is the trivial coset $\Gamma$ for all $v\neq 0^i$. Since $\cu^i(\ns)$ is compact, so is $Q_i$. We have $\q=\pi_\Gamma\co \q'$ for some $\q'\in \cu^i(G_\bullet)$ with $\q'(v)\in \Gamma$ for $v\neq 0^i$. The fact that $\q'\in \cu^i(G_\bullet)$ implies that the Gray-code product $\sigma_i(\q')$ lies in $G_i$ (recall \cite[Definition 2.2.22 and Proposition 2.2.25]{Cand:Notes1}). By the formula for the Gray-code product, we deduce that $\sigma_i(\q')\q'(0^i)^{-1} \in \Gamma$. It follows that $\q'(0^i)\in G_i\,\Gamma$. Thus the projection $\pi_{0^i}: \q\to \q(0^i)$ maps $Q_i$ into $(G_i\,\Gamma)/\Gamma$. Conversely, for every coset $g \Gamma$ with $g\in G_i$ there is a cube $\q\in Q_i$ with $\pi_{0^i}(\q)=g\Gamma$, namely $\q=\pi_\Gamma\co \q'$ where $\q'(0^i)=g$ and $\q'(v)=\mathrm{id}_G$ otherwise. It follows that $\pi_{0^i}(Q_i)=(G_i\,\Gamma)/\Gamma$, so the compactness of $Q_i$ implies that $(G_i\,\Gamma)/\Gamma$ is compact, whence $G_i/\Gamma_i$ is compact.
\end{proof}
\chapter{Characterization of compact nilspaces}\label{chap:compnils}
\section{Topological preliminaries}\label{sec:toprelims}
\medskip
In order to characterize a general compact nilspace, we first have to examine how the basic algebraic structures treated in \cite{Cand:Notes1} behave under the additional topological assumptions in Definition \ref{def:compnils}.\\
\noindent We begin with a result showing that the closure condition on $\cu^n(\ns)$ needs to be checked only for $n=k+1$.
\begin{lemma}\label{lem:compclosure}
A $k$-step nilspace $\ns$ is a compact nilspace if and only if $\ns$ is a compact space and $\cu^{k+1}(\ns)$ is a closed subset of $\ns^{\{0,1\}^{k+1}}$.
\end{lemma}
\begin{proof}
Note that for all $n\leq m$ we can view $\cu^n(\ns)$ as the projection of $\cu^m(\ns)$ to $\ns^{\{0,1\}^n}$ (by the composition axiom, embedding $\{0,1\}^n$ as a face in $\{0,1\}^m$). In particular, if we assume that $\cu^{k+1}(\ns)$ is closed then for every $n<k+1$ the projection $\cu^n(\ns)$ is also closed. If $n> k+1$, then for any $(k+1)$-dimensional face $F$ let
\[
Q_F=\big\{f:\{0,1\}^n\to \ns ~|~ f\co \phi_F \in \cu^{k+1}(\ns)\big\},
\]
where $f\co \phi_F$ is the restriction of $f$ to $F$ (recall \cite[Definition 1.1.5]{Cand:Notes1}). Thus $Q_F$ is the preimage of a closed set under the restriction map $f\mapsto f\co \phi_F $, and is therefore closed. Then, by \cite[Lemma 3.2.13]{Cand:Notes1}, we have $\cu^n(\ns)=\bigcap_F Q_F$ where $F$ runs through $(k+1)$-dimensional faces of $\{0,1\}^n$.
\end{proof}
\noindent In the remainder of this section we establish topological properties of the structures in \cite[Chapter 3]{Cand:Notes1} under the compactness assumption. The main result that we shall obtain along these lines is that if a finite-step nilspace is compact, then the associated abelian bundle obtained in \cite[Theorem 3.2.19]{Cand:Notes1} is endowed with a topology making it an iterated \emph{continuous} abelian bundle (in a sense that we shall formalize in Definition \ref{def:CpctAbBund}), with structure groups being compact abelian groups.
We begin by looking at the behaviour of some basic constructions. Recall from \cite[Definitions 3.1.19 and 3.1.22]{Cand:Notes1} the notions of arrow spaces and the spaces $\partial_x\ns$.
\begin{lemma}
Let $\ns$ be a compact nilspace, let $\ab$ be a compact abelian group, and fix any $x\in \ns$. Then the arrow spaces $\ns \Join_i \ns$, $i\geq 1$, the nilspaces $\cD_k(\ab)$, $k\geq 1$, and the space $\partial_x \ns$, are all compact nilspaces.
\end{lemma}
\begin{proof}
The space $\ns \times \ns$ underlying $\ns \Join_i \ns$ is compact and $\cu^{k+1}(\ns \Join_i \ns)$ is the image under a continuous map of a closed set of cubes in $\cu^{k+1+i}(\ns)$ (see \cite[(3.4)]{Cand:Notes1}), so it is a compact set as required. The cube set $\cu^{k+1}(\cD_k(\ab))$ is also compact, since it is the set of solutions of a linear equation (recall \cite[(2.9)]{Cand:Notes1}). Finally, the cube set $\cu^{k+1}(\partial_x \ns)$ is a closed subset of $\cu^{k+1}(\ns\Join_1 \ns)$ (recall \cite[Definition 3.1.22]{Cand:Notes1}).
\end{proof}
\noindent We now look at the more involved constructions, especially those using quotients by equivalence relations.
\begin{remark}\label{rem:stfacts} We shall use the following standard facts.
\begin{enumerate}\vspace{-0.2cm}
\item Let $X,Y$ be compact spaces (as in Definition \ref{def:compspace}). Then a function $f:X\rightarrow Y$ is continuous if and only if its graph is closed. (This is a special case of the closed graph theorem \cite[Ex. 8, p. 171]{Munkres}.)
\item Compact spaces are metrizable. (This follows from Urysohn's metrization theorem \cite[Thm. 34.1, p. 215]{Munkres}, using that a compact Hausdorff space is regular \cite[Ex. 3, p. 205]{Munkres}.)
\item Let $X$ be a compact space and let $\sim$ be a closed equivalence relation on $X$, that is $\{(x,y): x\sim y\}$ is a closed subset of $X\times X$. Then $X/\sim$ is compact in the quotient topology \cite[Prop. 8, p. 105]{Bourb1}.
\end{enumerate}
\end{remark}
\noindent Recall from \cite[Definition 1.2.3]{Cand:Notes1} that $\ns$ is said to be $k$\emph{-fold ergodic} if $\cu^k(\ns)=\ns^{\{0,1\}^k}$.
\begin{lemma}\label{lem:topkfolderg} Let $\ns$ be a $k$-step $k$-fold ergodic compact nilspace. Then $\ns$ is isomorphic as a compact nilspace to $\cD_k(\ab)$ for some compact abelian group $\ab$.
\end{lemma}
\begin{proof} As seen in \cite[Proposition 2.4.1]{Cand:Notes1}, in the case $k=1$, if we fix any $e\in \ns$ as a distinguished point, then the cube structure on $\ns$ yields a commutative group operation on $\ns$ with identity element $e$, making $\ns$ a principal homogeneous space of the resulting abelian group $\ab$. Let us check the continuity of addition and inverse on $\ab$. The graph of addition can be written
\[
\{(x_{00},x_{10},x_{01},x_{11})\in \cu^2(\ns) : x_{00}=e\}.
\]
The graph of the inverse is similarly defined by the equation $x_{00}=x_{11}=e$. These are closed sets, so the operations are continuous.
For $k>1$, as in \cite[Proposition 3.2.14]{Cand:Notes1} we consider the 1-step nilspace $\partial_e^{k-1}(\ns)$, which is here a compact nilspace, and therefore isomorphic to $\cD_1(\ab)$ for some compact abelian group $\ab$. We then argue just as in the proof of \cite[Proposition 3.2.14]{Cand:Notes1} to show that $\cu^n(\ns)=\cu^n(\cD_k(\ab))$ for all $n$.
\end{proof}
\noindent Recall from \cite[Definition 3.2.3]{Cand:Notes1} the relation $\sim_k$ on a nilspace $\ns$ and the corresponding nilspace factor $\cF_k(\ns)$ from \cite[Lemma 3.2.10]{Cand:Notes1}.
\begin{lemma}\label{lem:compfactor} Let $\ns$ be a compact nilspace and let $k\in\mathbb{N}$. Then $\cF_k(\ns)$ with the quotient topology is a compact nilspace.
\end{lemma}
\begin{proof} As a subset of $\ns\times \ns$, the relation $\sim_k$ consists of restrictions to the 1-face $\{0^{k+1},(1,0,\dots,0)\}$ of the cubes satisfying the condition in \cite[Lemma 3.2.4]{Cand:Notes1}. The set of these cubes is
\[
\cu^{k+1}(\ns)\;\cap\;\big\{f\in \ns^{\{0,1\}^{k+1}}~:~\forall\, v,v'\in \{0,1\}^{k+1}\setminus\{0^{k+1}\},\;f(v)=f(v')\big\},
\]
and is therefore closed.
\end{proof}
\noindent We now move toward the main result of this section, namely the addition of suitable topological properties to the bundle decomposition of a nilspace.
\subsection{Continuous abelian bundles}\label{sec:contbundec}
Recall from \cite[Definition 3.2.17]{Cand:Notes1} the notion of a $k$-fold abelian bundle.
\begin{defn}\label{def:CpctAbBund}
Let $\bnd$ be an abelian bundle with base $S$, structure group $\ab$ and projection $\pi$. We say that $\bnd$ is \emph{continuous} if the following conditions hold:
\begin{enumerate}
\item $\bnd$ and $S$ are topological spaces.
\item $\ab$ is an abelian topological group.
\item The action $\alpha:\ab\times \bnd\to \bnd$ is continuous.
\item A set $U\subset S$ is open if and only if $\pi^{-1}(U)$ is open in $\bnd$.
\end{enumerate}
We say that $\bnd$ is a \emph{compact} abelian bundle if, in addition to these conditions, $\bnd,S,\ab$ are compact spaces.
\end{defn}
\begin{remark}\label{rem:open-and-closed}
Recall that by \cite[Definition 3.2.17]{Cand:Notes1} the projection $\pi$ yields a bijection between $S$ and the orbits of $\alpha$, namely the bijection $s\mapsto \pi^{-1}(s)$. It follows that we obtain a definition equivalent to Definition \ref{def:CpctAbBund} if we replace condition (iv) with the following:\vspace{0.1cm}\\
\hspace*{0.15cm} (iv')\; Denoting by $\bnd/\alpha$ the set of $\ab$-orbits equipped with the quotient topology, we have that the map\\
\hspace*{1cm} $S\to \bnd/\alpha$, $\;s\mapsto \pi^{-1}(s)$ is a homeomorphism. \vspace{0.1cm}\\
\noindent Thus the base $S$ is topologically identified with the orbit space for the $\ab$-action on $\bnd$. In particular, since the orbit map $\bnd\to \bnd/\alpha$, $x\mapsto x+\ab$ is an open map, we have that the projection $\pi$ is an open map. Note also that if $\ab$ is compact, then the orbit map is also a closed map, and hence so is $\pi$.
\end{remark}
\begin{defn}\label{def:compdegkbund}
Let $\ns=\bnd_k$ be a $k$-fold abelian bundle, with factors $\bnd_i$, $i=0,\ldots,k$. We say that $\ns$ is a \emph{$k$-fold compact abelian bundle} if for each $i\in [k]$ the factor $\bnd_i$ is a compact $\ab_i$-bundle with base $\bnd_{i-1}$. A \emph{compact degree-$k$ bundle} is a $k$-fold compact abelian bundle that is also a compact cubespace with respect to the same topology, and such that condition (3.5) from \cite[Definition 3.2.18]{Cand:Notes1} is satisfied, that is, for every $i \in [0,k-1]$ and every $n\in \mathbb{N}$, we have $\cu^n(\bnd_i) = \{\pi_i\co \q: \q\in \cu^n(\ns)\}$, and for every $\q\in \cu^n(\bnd_{i+1})$ we have
\begin{equation}\label{eq:comp-k-deg-bund}
\{\q_2 \in \cu^n(\bnd_{i+1}): \pi_i\co \q = \pi_i\co \q_2\} = \{\q+\q_3: \q_3\in \cu^n(\cD_{i+1}(\ab_{i+1}))\}.
\end{equation}
\end{defn}
We can now establish the main result of this subsection.
\begin{proposition}\label{prop:topbundec} A compact cubespace $\ns$ is a compact degree-$k$ bundle if and only if $\ns$ is a $k$-step compact nilspace.
\end{proposition}
\begin{proof} The algebraic part of the statement is given by \cite[Theorem 3.2.19]{Cand:Notes1}, so we only need to check the topological properties in each direction.\\
\indent If $\ns$ is a compact degree-$k$ bundle then in particular it is a compact cubespace satisfying the nilspace axioms, so it is a $k$-step compact nilspace. Conversely, if $\ns$ is a $k$-step compact nilspace then, to show that the degree-$k$ bundle structure given by \cite[Theorem 3.2.19]{Cand:Notes1} is compact, it suffices to show that the abelian structure group $\ab_k$ can be equipped with a topology making it a compact abelian group such that $\ns$ is a compact $\ab_k$-bundle with base $\cF_{k-1}(\ns)$ (if we show this then the result follows by repeating the argument for $\ab_{k-1}$ and so on by induction). For this it suffices to show that $\ab_k$ can be equipped with a compact topology such that the action of $\ab_k$ on $\ns$ is continuous (since Lemma \ref{lem:compfactor} then implies the remaining condition (iv') in Definition \ref{def:CpctAbBund}).
Let $F$ be a class of $\sim_{k-1}$ in $\ns$. From the proof of \cite[Corollary 3.2.16]{Cand:Notes1} we know that $F$ together with the $F$-valued cubes in each $\cu^n(\ns)$ is a $k$-fold ergodic $k$-step nilspace. Using \cite[Lemma 3.2.4]{Cand:Notes1} we deduce that $F$ is closed and it follows that $F$ is a compact nilspace. Then by Lemma \ref{lem:topkfolderg} there is a compact topology on $\ab_k$ such that $F$ is isomorphic to $\cD_k(\ab_k)$ as a compact nilspace. We now check that the action of $\ab_k$ is continuous on $F$ by showing that its graph $\{(a,y_0,y_0+a): a\in \ab,y_0\in \ns\}$ is closed in $\ab_k\times \ns^2$. This graph can be described using the arrow space $\nss=\ns\Join_1 \ns$. Indeed, recall from \cite[Lemma 3.2.15]{Cand:Notes1} that for $x=(x_0,x_1)$, $y=(y_0,y_1)$ in $F^2\subset \nss$ we have $x\sim_{k-1} y$ if and only if $x_0-x_1=y_0-y_1$ in $\ab_k$. It follows that, fixing some $e\in F$, the graph above is homeomorphic to the closed set $\{(x,y) \in F^2\times F^2: x \sim_{k-1} y,\; x_0 =e\}$. We have thus shown that from any given fibre $F$ the group $\ab_k$ acquires a compatible compact topology making its action on $F$ continuous. To see that this topology is the same for every fibre, recall that by \cite[Lemma 3.2.24]{Cand:Notes1} for any two fibres $F_0,F_1$ and fixed points $x_i\in F_i$, we have an isomorphism $\vartheta:\ab_{F_0}\to \ab_{F_1}$, defined by $\vartheta(a)=b$ if and only if $(x_0,x_1)\sim_{k-1}(x_0+a,x_1+b)$. We then see that $\vartheta$ is continuous by identifying its graph with the closed set $\{y \in F_0\times F_1: y \sim_{k-1} (x_0,x_1)\}$. It follows that $\vartheta$ is a homeomorphism.
\end{proof}
\noindent Recall the notions of a bundle morphism \cite[Definition 3.3.1]{Cand:Notes1} and of a sub-bundle \cite[Definition 3.3.3]{Cand:Notes1}. These notions are transferred to the setting of compact abelian bundles in a straightforward manner, by specifying in their definitions that all the maps involved are continuous. The same holds for the kernel of a bundle morphism $\psi:\bnd\to\bnd'$ (recall \cite[Definition 3.3.4]{Cand:Notes1}). Note that such a kernel $K$ is indeed a compact abelian bundle, since each set $K_i$ in \cite[Definition 3.3.4]{Cand:Notes1} is a closed subset of $\bnd_i\times \bnd'$ and therefore compact, and every structure group $\ker(\alpha_i)$ is compact.
Recall also that in the purely algebraic setting we had a description of a set of restricted morphisms $\hom_f(P,\ns)$ as a sub-bundle of $\ns^P$, namely \cite[Lemma 3.3.11]{Cand:Notes1}. This has the following version for compact nilspaces.
\begin{lemma}\label{lem:top-restrmorph=subbund}
Let $P$ be a subcubespace of $\{0,1\}^n$ with the extension property, let $S$ be a subcubespace of $P$ with the extension property in $P$, let $\ns$ be a $k$-step compact nilspace and let $f:S\to \ns$ be a morphism. Then $\hom_f(P,\ns)$ is a compact $k$-fold abelian bundle that is a sub-bundle of $\ns^P$, with factors $\hom_{\pi_i\co f}(P,\ns_i)$ and structure groups $\hom_{S\to 0}(P,\cD_i(\ab_i))$, where $\ab_i$ is the $i$-th structure group of $\ns$.
\end{lemma}
\begin{proof}
We just have to check the topological properties. The set $\hom_f(P,\ns)$ is closed in $\ns^P$ and therefore compact. For each $i\in [k]$ the abelian group $\hom_{S\to 0}(P,\cD_i(\ab_i))$ is also compact, as a closed subgroup of $\ab_i^P$, and the action of $\hom_{S\to 0}(P,\cD_i(\ab_i))$ inherits continuity from that of $\ab_i$. Condition (iv) from Definition \ref{def:CpctAbBund} is also inherited from $\ns^P$ since the topology on each factor $\hom_{\pi_i\co f}(P,\ns_i)$ is the subspace topology from $\ns_i^P$.
\end{proof}
\subsection{Metrics}
\noindent By Remark \ref{rem:stfacts} (ii), we know that a compact nilspace can always be equipped with a compatible metric. We shall often use the fact that this metric can be assumed to be invariant under the action of the last structure group. We record this fact as follows.
\begin{lemma}\label{lem:d-invariance}
Let $\ns$ be a $k$-step compact nilspace and let $d_0$ be a metric on $\ns$ generating its topology. Define $d(x,y)=\int_{z\in \ab_k} d_0(x+z,y+z) \,\mathrm{d}\mu_{\ab_k}(z)$, for any $x,y\in \ns$. Then $d$ is a $\ab_k$-invariant metric that generates the same topology on $\ns$.
\end{lemma}
\noindent This invariant metric is a basic construction in the theory of $G$-spaces (see \cite[Proposition 1.1.12]{Pal-G}).
\begin{proof}
It is clear that $d$ is a $\ab_k$-invariant metric. To see that $d$ generates the same topology as $d_0$, note first that the function $\ns\times \ns\to \mathbb{R}_{\geq 0}$, $(x,y)\mapsto d(x,y)$ is continuous and in particular for a fixed $x\in \ns$ we have that $y\mapsto d(x,y)$ is continuous (relative to $d_0$). It follows that the topology generated by $d$ is a subtopology of the given topology on $\ns$. In particular the former topology is compact. However, that topology is also Hausdorff, so it is maximal among the compact topologies on $\ns$, and so it must be the original topology on $\ns$ (see \cite[\S 9.4, Corollary 3]{Bourb1}).
\end{proof}
\noindent From now on we always assume that a metric on $\ns$ is $\ab_k$-invariant in this sense. Given such a metric $d$ on $\ns$, we may define the following quotient metric on $\cF_{k-1}(\ns)$ (see \cite[Proposition 1.1.12]{Pal-G}):
\begin{equation}\label{eq:quotientmetric}
d'(x',y')=\inf\big\{d(x,y): x,y\in \ns,\; \pi_{k-1}(x)=x',\; \pi_{k-1}(y)=y'\big\}.
\end{equation}
\noindent In the sequel we shall use the following fact several times.
\begin{lemma}\label{lem:contcomp}
Let $\ns$ be a $k$-step compact nilspace, and let $\cor^{k+1}(\ns)$ denote the set of $(k+1)$-corners on $\ns$ with the subspace topology obtained from the product topology on $\ns^{\{0,1\}^{k+1}\setminus \{1^{k+1}\}}$. Let $\comp:\cor^{k+1}(\ns)\to \ns$ denote the map sending a $(k+1)$-corner $\q'$ to $\q(1^{k+1})$, where $\q$ is the unique completion of $\q'$. Then $\comp$ is continuous.
\end{lemma}
\begin{proof}
We argue by contradiction. If $\comp$ were not continuous then there would exist some corner $\q'$, some open set $U\ni \comp(\q')$ and a sequence of corners $\q_n'$ converging to $\q'$ such that for all $n$ we have $\comp(\q_n')\not\in U$. Let $\q_n$ denote the completion in $\cu^{k+1}(\ns)$ of $\q'_n$. By the compactness of $\cu^{k+1}(\ns)$ there exists a subsequence $(\q_m)$ of $(\q_n)$ such that $\q_m$ converges to some cube $\q\in \cu^{k+1}(\ns)$. This convergence combined with the assumption that $\q_n'\to \q'$ implies that for each $v\neq 1^{k+1}$ we have $\q(v)=\q'(v)$, so by uniqueness of completion we have $\q(1^{k+1})=\comp(\q')\in U$. On the other hand, the convergence $\q_m\to \q$ also implies that $\q_m(1^{k+1})\to \q(1^{k+1})$. Since by assumption $\q_m(1^{k+1})=\comp(\q_m')\not\in U$ for all $m$, we have $\q(1^{k+1})\not\in U$, a contradiction.
\end{proof}
\noindent Let us record also the following consequence of Lemma \ref{lem:contcomp}, to the effect that proximate cubes on $\cF_{k-1}(\ns)$ can always be lifted to proximate cubes on $\ns$.
\begin{lemma}\label{lem:closelifts}
Let $\ns$ be a $k$-step compact nilspace and let $d,d'$ be the metrics on $\ns$, $\cF_{k-1}(\ns)$ given by Lemma \ref{lem:d-invariance} and \eqref{eq:quotientmetric} respectively. Then for every $\epsilon>0$ there exists $\delta>0$ such that the following holds. If $\q_1,\q_2\in \cu^{k+1}(\cF_{k-1}(\ns))$ satisfy $d'(\q_1(v),\q_2(v))\leq \delta$ for all $v$, then there exist cubes $\tilde\q_1,\tilde\q_2\in \cu^{k+1}(\ns)$ such that $\pi_{k-1}\co\tilde\q_i=\q_i$ for $i=1,2$ and $d\big(\tilde\q_1(v),\tilde\q_2(v)\big)\leq \epsilon$ for all $v$.
\end{lemma}
\begin{proof}
Let $P=\{0,1\}^{k+1}\setminus\{1^{k+1}\}$ and note that, by definition of $d'$ and \cite[Remark 3.2.12]{Cand:Notes1}, the corners $\q_i|_P$ can be lifted to corners $\q_i'$, $i=1,2$ such that $d\big(\q_1'(v),\q_2'(v)\big)\leq \delta$ for every $v\in P$. By Lemma \ref{lem:contcomp}, if $\delta\leq \epsilon$ is sufficiently small then the unique completions $\tilde\q_i$ of $\q_i'$ satisfy $d\big(\tilde\q_1(1^{k+1}),\tilde\q_2(1^{k+1})\big)\leq \epsilon$.
\end{proof}
\medskip
\section{Measure-theoretic preliminaries}\label{sec:measprel}
\medskip
In this section we collect the main tools from measure theory that we shall use in the sequel.\\
\indent A notion that unifies most of these tools is that of a \emph{continuous system of measures}, discussed in Subsection \ref{subsec:csm}.\\
\indent Using continuous systems of measures, we shall then define in Subsection \ref{subsec:Haar} an invariant measure on a continuous abelian bundle, and use this to construct a generalization of the Haar measure for compact nilspaces.\\
\indent In Subsection \ref{subsec:morph-prob-spaces} we collect several constructions of probability spaces that will be used repeatedly later in the chapter.
\subsection{Continuous systems of measures}\label{subsec:csm}
Recall that a measure $\mu$ on a measure space $X$ is said to be \emph{concentrated on} $S\subset X$ if $\mu(X\setminus S)=0$.
The study of continuous systems of measures goes back at least to Bourbaki (see \cite[Ch. V, \S 3]{BourInt}). A recent treatment is given in \cite{C&Gra}. We shall use the following definition.
\begin{defn}\label{def:CSM}
Let $X,Y$ be compact spaces and let $\pi : X \to Y$ be continuous. A \emph{continuous system of measures} (CSM) on the map $\pi$ is a family of Borel measures $\{ \mu_y :y \in Y\}$ on $X$ satisfying the following conditions:
\begin{enumerate}
\item For every $y\in Y$ the measure $\mu_y$ is concentrated on $\pi^{-1}(y)$.
\item\label{contprop} For every continuous function $f:X\to\mathbb{C}$, the function $Y\to \mathbb{C}$, $y \mapsto \int_{\pi^{-1}(y)} f \; \,\mathrm{d}\mu_y$ is continuous.
\end{enumerate}
\end{defn}
\noindent As explained in \cite{C&Gra}, replacing $\mathbb{C}$ by $\mathbb{R}$ in condition \eqref{contprop} yields an equivalent definition.\\
\noindent Note that all the CSMs that we shall consider in these notes consist of strictly positive measures $\mu_y$, i.e. Borel measures taking positive values on non-empty open sets.
A simple example of a CSM is obtained by taking a product space, as follows.
\begin{lemma}\label{lem:GenCSM}
Let $V$ and $Y$ be compact spaces, let $\mu$ be a Borel probability measure on $V$, let $X=V\times Y$ and let $\pi : X\to Y,\, (v,y)\mapsto y$. For each $y\in Y$ let $\mu_y$ denote the measure on $X$ concentrated on $\pi^{-1}(y)$, where it equals the pushforward of $\mu$ under the map $V\to \pi^{-1}(y)$, $v\mapsto (v,y)$. Then $\{\mu_y:y\in Y\}$ is a CSM on $\pi$.
\end{lemma}
\begin{proof}
Let $f : V\times Y \to \mathbb{R}$ be a continuous function. Then $f$ is a uniform limit of functions of the form $F:(v,y) \mapsto \sum_{i=1}^n f_{i,1}(v)f_{i,2}(y)$ where $f_{i,1}:V\to \mathbb{R}$, $f_{i,2}:Y\to \mathbb{R}$ are continuous. Indeed, the algebra of such functions $F$ satisfies the assumptions of the Stone-Weierstrass theorem \cite[Appendix A, \S A14]{Rudin}. (The fact that this algebra separates points follows from Urysohn's lemma \cite[Theorem 33.1]{Munkres}.) Now, for a function $f_{i,1}(v)f_{i,2}(y)$ as above, it is clear that integrating over $v$ yields a continuous function of $y$. It follows that $y\mapsto \int_{\pi^{-1}(y)} f\; \,\mathrm{d}\mu_y$ is a uniform limit of continuous functions and is therefore continuous.
\end{proof}
\noindent One of our main uses of CSMs below is to equip any compact abelian bundle with a probability measure that is invariant under the action of the structure group. This measure will be obtained as a simple application of the following result concerning general CSMs.
\begin{lemma}\label{lem:CSM-measure}
Let $\{\mu_y:y\in Y\}$ be a family of Borel probability measures on $X$ forming a CSM on $\pi : X \to Y$, and let $\nu$ be a Borel probability measure on $Y$. Then the following function on Borel sets $E\subset X$ is a Borel probability measure:
\[
\mu(E) = \int_Y \mu_y\big(E \cap \pi^{-1}(y)\big)\, \,\mathrm{d}\nu(y).
\]
\end{lemma}
\begin{proof}
This is a special case of \cite[Corollary 3.7]{C&Gra}, using \cite[Proposition 2.23]{C&Gra}.
\end{proof}
\subsection{Haar measure on compact abelian bundles and nilspaces}\label{subsec:Haar}
Recall that every compact abelian group can be equipped with a regular Borel probability measure that is invariant under translation, called the (normalized) Haar measure \cite{Rudin}. In this subsection we shall construct a generalization of this measure for compact nilspaces. As we shall see, there is a natural inductive construction of a probability measure on a $k$-fold compact abelian bundle, starting with the Haar measure on a compact abelian group. This construction will yield the desired measure on a $k$-step compact nilspace, thanks to \cite[Theorem 3.2.19]{Cand:Notes1}.
Our first step, then, is to define a Haar measure on a compact abelian bundle.
\begin{lemma}\label{lem:bundHaar}
Let $\bnd$ be a compact abelian bundle with base $S$, structure group $\ab$ and projection $\pi$. Let $\mu_S$ be a regular Borel probability measure on $S$. Then there is a unique regular Borel probability measure $\mu$ on $\bnd$ that is invariant under the action of $\ab$ and satisfies $\mu_S=\mu\co\pi^{-1}$.
\end{lemma}
We call $\mu$ the \emph{Haar measure} on $\bnd$ (given $\mu_S$).
\begin{proof}
For each $s\in S$ the set $\pi^{-1}(s)$ is a principal homogeneous space of $\ab$, so there is a $\ab$-equivariant homeomorphism $f:\ab\to \pi^{-1}(s)$. Let $\mu_s$ be the pushfoward measure $\mu_{\ab}\co f^{-1}$ on $\pi^{-1}(s)$, where $\mu_{\ab}$ is the Haar measure on $\ab$. If we show that $\{\mu_s:s\in S\}$ is a CSM on $\pi$, then by Lemma \ref{lem:CSM-measure} we can define a Borel measure $\mu$ on $\bnd$ by the following formula:
\begin{equation}\label{eq:HaarMeas}
\mu(E)=\int_S \; \mu_s\big(\pi^{-1}(s)\cap E\big)\; \,\mathrm{d}\mu_S(s),
\end{equation}
for every Borel set $E\subset \bnd$.
To show that $\{\mu_s:s\in S\}$ is a CSM on $\pi$ we check condition \eqref{contprop} from Definition \ref{def:CSM}. Given any continuous function $f : \bnd \to \mathbb{R}$, we need to show that $g : S \to \mathbb{R}$,
$s\mapsto \int_{\pi^{-1}(s)} f \, \,\mathrm{d}\mu_s$
is continuous. Since $S$ has the quotient topology (Definition \ref{def:CpctAbBund}), it suffices to show that $g \circ \pi : \bnd \to \mathbb{R}$ is continuous.
Now $g(\pi(b)) = \int_{\ab} f(a+b) \; \,\mathrm{d}\mu_{\ab}(a) = \int_{\ab} f'(a,b) \; \,\mathrm{d}\mu_{\ab}(a)$, where the function $f':\ab\times \bnd \to \mathbb{R}, \; (a,b) \mapsto f(a+b)$ is continuous. Therefore, the continuity of $g \circ \pi$ follows from (the proof of) Lemma \ref{lem:GenCSM} applied with $V=\ab$ and $Y=\bnd$.\\
\indent We thus obtain the Borel measure $\mu$ given by \eqref{eq:HaarMeas}, and $\mu$ is regular by the standard fact that any Borel probability measure on a metric space is regular \cite[Theorem 1.1]{Bill2}. The $\ab$-invariance of $\mu$ clearly follows from that of each measure $\mu_s$. The uniqueness can be deduced from uniqueness of Haar measure on $\ab$ combined with a disintegration result for invariant measures, such as \cite[Proposition 2]{Rip}.
\end{proof}
As mentioned above, this lemma yields the following result.
\begin{proposition}\label{prop:nilspaceHaar}
Let $\bnd$ be a compact $k$-fold abelian bundle, with factors $\bnd_0,\bnd_1,\dots,\bnd_k=\bnd$ and structure groups $\ab_1,\dots,\ab_k$. Then there is a unique regular Borel probability measure $\mu$ on $\bnd$ with the following property: for each $i\in [k]$, letting $\pi_i$ be the projection $\bnd\to \bnd_i$, we have that the Borel probability measure $\mu\co\pi_i^{-1}$ on $\bnd_i$ is invariant under the action of $\ab_i$.
\end{proposition}
\noindent In particular, for every $k$-step compact nilspace $\ns$, applying the proposition to the bundle structure we obtain the measure that we shall call the \emph{Haar measure on} $\ns$.
\begin{proof}
By induction on $i\in [k]$, starting with the Haar probability on $\ab_1$ and applying Lemma \ref{lem:bundHaar} for $i=2,\dots,k$.
\end{proof}
\noindent Recall from \cite[Definition 3.3.1]{Cand:Notes1} that a bundle morphism $\phi:\bnd\to \bnd'$ between $k$-fold abelian bundles is said to be \emph{totally surjective} if the structure morphism $\alpha_i:\ab_i\to\ab_i'$ is surjective for each $i\in [k]$.
The following basic result is the analogue for nilspaces of the fact that continuous surjective homomorphisms between compact abelian groups preserve the Haar measures.
\begin{lemma}\label{lem:MeasPres}
Let $\bnd,\bnd'$ be compact $k$-fold abelian bundles, and let $\phi:\bnd\to \bnd'$ be a totally surjective continuous bundle morphism. Then $\phi$ preserves the Haar measures, that is, for every Borel set $E\subset \bnd'$ we have $\mu_{\bnd}\big(\phi^{-1}(E)\big)=\mu_{\bnd'}(E)$.
\end{lemma}
\begin{proof}
We argue by induction on $k$. The case $k=1$ follows from the fact recalled above. For $k>1$, recall that by \cite[Definition 3.3.1]{Cand:Notes1} the map $\phi$ induces a totally surjective continuous bundle morphism $\phi_{k-1}:\bnd_{k-1}\to \bnd'_{k-1}$. Let $\{\mu_s:s\in \bnd_{k-1}\}$ and $\{\mu'_{s'}:s'\in \bnd'_{k-1}\}$ be the CSMs on $\pi_{k-1}:\bnd\to\bnd_{k-1}$, $\pi_{k-1}':\bnd'\to\bnd_{k-1}'$ respectively (as in the proof of Lemma \ref{lem:bundHaar}). For $s\in \bnd_{k-1}$ and any Borel set $E \subset \bnd'$, let $f_{\bnd}(s)=\mu_s\big(\pi_{k-1}^{-1}(s)\cap \phi^{-1}(E)\big)$, and for $s'\in \bnd_{k-1}'$ let $f_{\bnd'}(s')=\mu'_{s'}\big({\pi'}^{-1}_{k-1}(s')\cap E\big)$. The restriction of $\phi$ to a fibre $\pi_{k-1}^{-1}(s)$, being a continuous affine homomorphism onto some fibre ${\pi'}^{-1}_{k-1}(s')$, preserves the measures $\mu_s,\mu'_{s'}$. Hence we have $f_{\bnd}(s)= f_{\bnd'}(\phi_{k-1}(s))$ for each $s \in \bnd_{k-1}$. Since $\phi_{k-1}$ is measure-preserving (by the induction hypothesis), we have
\[
\mu_{\bnd}\big(\phi^{-1}(E)\big) = \int_{\bnd_{k-1}} f_{\bnd}\,\mathrm{d} \mu_{\bnd_{k-1}}= \int_{\bnd_{k-1}} f_{\bnd'}\co\phi_{k-1}\,\mathrm{d} \mu_{\bnd_{k-1}}
= \int_{\bnd_{k-1}'} f_{\bnd'}\,\mathrm{d} \mu_{\bnd_{k-1}'}= \mu_{\bnd'}(E). \qedhere
\]
\end{proof}
\noindent Recall from \cite[Definition 3.3.7]{Cand:Notes1} the notion of a fibre-surjective morphism.
\begin{corollary}\label{cor:ctsfibsurmorph}
Let $\ns,\ns'$ be $k$-step compact nilspaces and let $\phi:\ns\to\ns'$ be a continuous fibre-surjective morphism. Then $\phi$ preserves the Haar measures.
\end{corollary}
\begin{proof}
We combine Lemma \ref{lem:MeasPres} with \cite[Lemma 3.3.8]{Cand:Notes1}.
\end{proof}
\noindent In the sequel we shall have to use CSMs on certain maps from cube sets $\cu^n(\ns)$ to $\ns$, such as the evaluation map $\q\mapsto \q(0^n)$. These CSMs will be obtained as special cases of the construction in the next lemma.\\
\indent Recall from \cite[Definition 3.2.17]{Cand:Notes1} the notion of a \emph{relative} $k$-fold abelian bundle. We may similarly define a \emph{compact} relative $k$-fold abelian bundle $\bnd$ with ground set $\bnd_0$. Note that each fibre of the projection $\pi_0:\bnd\to\bnd_0$ is then itself a compact $k$-fold abelian bundle and so it has a Haar measure by Lemma \ref{lem:bundHaar}.
\begin{lemma}\label{lem:relcsm} Let $\bnd$ be a compact relative $k$-fold abelian bundle with factors $\bnd_0, \bnd_1, \ldots,\bnd_{k-1}$. Then the Haar measures on the fibres of the projection $\pi_0:\bnd\to \bnd_0$ form a CSM on $\pi_0$.
\end{lemma}
To prove this we shall have to compose several CSMs, in the following sense.
\begin{defn}
Given CSMs $\{\mu_y: y \in Y\}$ on $\pi : X \to Y$ and $\{\mu_z: z \in Z\}$ on $\tau : Y \to Z$, their \emph{composition} is the CSM $\{\nu_z: z \in Z\}$ on $\tau \co \pi : X \to Z$ consisting of measures defined for Borel sets $E\subset X$ by
\begin{equation}\label{eq:CSMcomp}
\nu_z(E) = \int_{\tau^{-1}(z)} \mu_y\big( \pi^{-1}(y) \cap E\big) \; \,\mathrm{d}\mu_z(y).
\end{equation}
\end{defn}
The fact that this composition is indeed a CSM is established in \cite[Proposition 3.3]{C&Gra}.
\begin{proof}[Proof of Lemma \ref{lem:relcsm}]
By the proof of Lemma \ref{lem:bundHaar}, for each $i\in [k]$ there is a CSM on the bundle map $\pi_{i-1,i}:\bnd_i\to \bnd_{i-1}$ consisting of the pushforward of the Haar measure on $\ab_i$ to each fibre of $\pi_{i-1,i}$. Let us compose these CSMs for $i\in [k]$, and let $\{\mu_x: x\in \bnd_0\}$ be the resulting CSM on $\pi_0$. For each $x\in \bnd_0$, a straightforward inductive argument using \eqref{eq:CSMcomp} shows that the measure $\mu_x$ on the compact $k$-fold abelian bundle $\pi^{-1}(x)$ satisfies the properties of the measure in Proposition \ref{prop:nilspaceHaar}, so it must be the Haar measure on $\pi^{-1}(x)$.
\end{proof}
\noindent The following result is an analogue for compact abelian bundles of the quotient integral formula.
\begin{lemma}\label{lem:quotint}
Let $\bnd,\bnd'$ be compact $k$-fold abelian bundles, with Haar measures $\mu,\mu'$ respectively, and let $\phi:\bnd\to \bnd'$ be a totally surjective continuous bundle morphism. For each $t\in \bnd'$ let $\mu_t$ denote the Haar measure on $\phi^{-1}(t)$. Then the measures $\mu_t$, $t\in \bnd'$ disintegrate $\mu$ relative to $\mu'$, that is for every Borel set $E\subset \bnd$ we have $\mu(E)=\int_{\bnd'}\mu_t\big(\phi^{-1}(t)\cap E\big)\,\,\mathrm{d}\mu'(t)$.
\end{lemma}
Recall from \cite[Definition 3.3.4]{Cand:Notes1} the notion of the kernel of a bundle morphism.
\begin{proof}
The kernel of $\phi$ is a compact relative $k$-fold abelian bundle, with ground set $\bnd'$, and the fibres $\phi^{-1}(t)$ are compact $k$-fold abelian bundles (recall \cite[Lemma 3.3.6]{Cand:Notes1}). By Lemma \ref{lem:relcsm} the Haar measures $\mu_t$ form a CSM on $\phi$. By Lemma \ref{lem:CSM-measure}, with this CSM and $\mu'$ we can define a Borel probability $\nu$ on $\bnd$ by
\begin{equation}\label{eq:nukey}
\nu(E)=\int_{\bnd'}\mu_t\big(\phi^{-1}(t)\cap E\big)\,\,\mathrm{d}\mu'(t).
\end{equation}
We have to show that $\nu=\mu$. For this it suffices to show that $\nu$ has the invariance properties that characterize the Haar measure $\mu$, namely that for each projection $\pi_i:\bnd\to\bnd_i$ the measure $\nu\co\pi_i^{-1}$ on $\bnd_i$ is $\ab_i$-invariant (recall Proposition \ref{prop:nilspaceHaar}). We prove this by induction on $k$.\\
\indent For $k=1$, we have that $\bnd,\bnd'$ are principal homogeneous spaces of compact abelian groups $\ab,\ab'$ respectively, and we just have to show that $\nu$ is $\ab$-invariant. But from the quotient integral formula \cite[Theorem 1.5.2]{D&E} it actually follows that $\nu=\mu$ in this case, so $\nu$ is indeed $\ab$-invariant.\\
\indent Supposing now that the claim holds for $k\geq 1$, we prove it for $k+1$. First we show that we can suppose by induction that $\nu\co\pi_i^{-1}$ is already $\ab_i$-invariant on $\bnd_i$ for each $i\in [k]$. For this it suffices to show that $\nu\co \pi_k^{-1}$ is of the same form as $\nu$ (as in \eqref{eq:nukey}), but relative to the totally surjective morphism $\phi_k:\bnd_k\to \bnd_k'$ induced by $\phi:\bnd_{k+1}\to \bnd_{k+1}'$. By construction, the Haar measure $\mu'$ has a disintegration relative to the Haar measure $\mu'_k$ on $\bnd'_k$, into measures $\mu'_s$, $s\in \bnd_k'$, each of which is the Haar measure on $\ab_{k+1}'$ pushed forward to the fibre ${\pi'_k}^{-1}(s)$. Thus for any Borel set $E\subset \bnd_k$ we have
\[
\nu\co\pi_k^{-1}(E)\; = \;\int_{\bnd'}\mu_t\big(\pi_k^{-1}(E)\cap \phi^{-1}(t)\big) \,\mathrm{d}\mu'(t) \; = \; \int_{\bnd'_k}\Big(\int_{{\pi'_k}^{-1}(s)} \mu_t\big(\pi_k^{-1}(E)\cap \phi^{-1}(t)\big) \,\mathrm{d}\mu'_s(t)\Big) \,\mathrm{d}\mu'_k(s).
\]
For each $s\in \bnd'_k$ and each $t\in \bnd'$ with $\pi'_k(t)=s$, note that $\phi^{-1}(t)$ is a $(k+1)$-fold compact abelian bundle, with $k$-th factor $\phi_k^{-1}(s)$ having Haar measure denoted $\mu_s$. Then $\pi_k$ restricted to $\phi^{-1}(t)$ pushes $\mu_t$ forward to $\mu_s$. It follows that for any such $s,t$ we have $\mu_t\big(\pi_k^{-1}(E)\cap \phi^{-1}(t)\big)=\mu_s\big(E\cap \phi_k^{-1}(s)\big)$. Hence
\[
\nu\co\pi_k^{-1}(E) = \int_{\bnd'_k}\Big(\int_{{\pi'_k}^{-1}(s)} \mu_s\big(E\cap \phi_k^{-1}(s)\big)\,\mathrm{d}\mu_s'(t)\Big) \,\mathrm{d}\mu_k'(s) = \int_{\bnd'_k} \mu_s\big(E\cap \phi_k^{-1}(s)\big) \,\mathrm{d}\mu_k'(s).
\]
The right side here is indeed the measure on $\bnd_k$ constructed in the same way as $\nu$ in \eqref{eq:nukey}. Therefore, as mentioned above, by induction it now suffices to show that $\nu$ is $\ab_{k+1}$-invariant.\\
\indent Again we use the disintegration of $\mu'$ relative to $\mu'_{k-1}$, so that for any Borel set $E\subset \bnd'$ we have
\[
\nu(E) = \int_{\bnd'_k}\Big(\int_{{\pi'_k}^{-1}(s)} \mu_t\big(E\cap \phi^{-1}(t)\big) \,\mathrm{d}\mu_s'(t)\Big) \,\mathrm{d}\mu_k'(s).
\]
For each $s\in \bnd_k'$, we claim that the inner integral on the right side here integrates $1_E$ over a union of orbits of $\ab_{k+1}$. Indeed, we have $\mu_t\big( E\cap \phi^{-1}(t)\big)=\int_{\phi^{-1}(t)} 1_E(x)\,\mathrm{d}\mu_t(x) $, and by disintegrating $\mu_t$ into the fibres of $\pi_k:\phi^{-1}(t)\to \phi_k^{-1}(s)$, the inner integral in question is written
\[
\int_{{\pi'_k}^{-1}(s)} \; \int_{\phi_k^{-1}(s)} \;\int_{\pi_k^{-1}(y)\cap \phi^{-1}(t)} 1_E(x) \;\,\mathrm{d}\mu_{\pi_k^{-1}(y)}(x) \; \,\mathrm{d}\mu_s(y) \; \,\mathrm{d}\mu_s'(t).
\]
By Fubini's theorem we can interchange the two outer integrals, so this equals
\[
\int_{\phi_k^{-1}(s)} \; \int_{{\pi'_k}^{-1}(s)} \;\int_{\pi_k^{-1}(y)\cap \phi^{-1}(t)} 1_E(x) \;\,\mathrm{d}\mu_{\pi_k^{-1}(y)}(x) \; \,\mathrm{d}\mu_s'(t)\; \,\mathrm{d}\mu_s(y) .
\]
For each $y\in \phi_k^{-1}(s)$, note that ${\pi'_k}^{-1}(s)$ is a $\ab_{k+1}'$-torsor (or principal homogeneous space of $\ab_{k+1}'$) and for every $t$ in this torsor we have that $\pi_k^{-1}(y)\cap \phi^{-1}(t)$ is a $\ker(\alpha_{k+1})$-torsor, where $\alpha_{k+1}$ is the $(k+1)$-th structure morphism of $\phi$ (recall \cite[Definition 3.3.1 and Lemma 3.3.6]{Cand:Notes1}). Fixing any $t_0$ in the former torsor, and then for each $r\in \ab_{k+1}'$ fixing some $x_r\in \pi_k^{-1}(y)\cap \phi^{-1}(t_0+r)$, the last inner double-integral equals
\[
\int_{\ab_{k+1}'} \int_{\pi_k^{-1}(y)\cap \phi^{-1}(t_0+r)} 1_E(x)\, \,\mathrm{d}\mu_{\pi_k^{-1}(y)}(x)\, \,\mathrm{d}\mu_{\ab_{k+1}'}(r)
= \int_{\ab_{k+1}'} \int_{\ker(\alpha_{k+1})} 1_E(x_r+u) \,\,\mathrm{d}\mu_{\ker(\alpha_{k+1})}(u)\, \,\mathrm{d}\mu_{\ab_{k+1}'}(r).
\]
Note that the sets $\{x_r+u:u\in \ker(\alpha_{k+1})\}$, $r\in \ab_{k+1}'$ form a partition of $\pi_k^{-1}(y)$. It follows from the quotient integral formula \cite[Theorem 1.5.2]{D&E} that for any $x_0\in \pi_k^{-1}(y)$ this integral equals
\[
\int_{\ab_{k+1}} 1_E(x_0+z) \,\mathrm{d}\mu_{\ab_{k+1}}(z)= \int_{\pi_k^{-1}(y)} 1_E(x) \,\mathrm{d}\mu_{\pi_k^{-1}(y)}(x).
\]
This is clearly invariant under shifting $E$ by any fixed $z\in \ab_{k+1}$. The $\ab_{k+1}$-invariance of $\nu$ follows.
\end{proof}
\noindent Recall the fact that the Haar measure on a compact abelian group is strictly positive. This generalizes to compact nilspaces.
\begin{proposition}\label{prop:posmeasopen}
Let $\ns$ be a $k$-step compact nilspace, with Haar measure $\mu$. Then for every open set $U\subset \ns$ we have $\mu(U)>0$.
\end{proposition}
\begin{proof}
We argue by induction on $k$. The case $k=1$ follows from the fact recalled above. Let $k>1$ and let $U$ be open in $\ns$. By Remark \ref{rem:open-and-closed} we have that $\pi_{k-1}$ is an open map, so $B=\pi_{k-1}(U)$ is open in $\cF_{k-1}(\ns)$. Hence, letting $\mu'$ denote the Haar measure on $\cF_{k-1}(\ns)$, we have by induction that $\mu'(B)>0$. By regularity of $\mu'$ there is a compact subset $C\subset B$ such that $\mu'(C)>0$, and then since $\pi_{k-1}$ preserves Haar measures we have $\mu(\pi_{k-1}^{-1}(C))>0$. Now by Remark \ref{rem:open-and-closed} we also have that $\pi_{k-1}$ is a proper map, so $\pi_{k-1}^{-1}(C)$ is a compact set. This set is covered by the union of open sets $\bigcup_{z\in \ab_k} U+z$, so there is a finite set of translates of $U$ by elements of $\ab_k$ that covers $\pi_{k-1}^{-1}(C)$, and it follows that $\mu(U)>0$.
\end{proof}
\subsection{Probability spaces of morphisms}\label{subsec:morph-prob-spaces}
Recall from \cite[Definition 3.1.3]{Cand:Notes1} that given a cubespace $P$ and a subcubespace $S$ of $P$, we say that $S$ has the \emph{extension property} in $P$ if for every non-empty nilspace $\ns$ and every morphism $g': S \to \ns$ there is a morphism $g: P\to \ns$ with $g|_S=g'$.
In this subsection we collect several constructions of probability spaces of morphisms between nilspaces, which will be used repeatedly in the sequel. These constructions rely on two basic ideas.\\
\indent Firstly, every set of restricted morphisms $\hom_f(P,\ns)$ has a compact abelian-bundle structure, as described in Lemma \ref{lem:top-restrmorph=subbund}, which we restate here for convenience.
\begin{lemma}\label{lem:comp-restrmorph=subbund}
Let $P$ be a subcubespace of $\{0,1\}^n$ with the extension property, let $S$ be a subcubespace of $P$ with the extension property in $P$, let $\ns$ be a $k$-step compact nilspace and let $f:S\to \ns$ be a morphism. Then $\hom_f(P,\ns)$ is a compact $k$-fold abelian bundle that is a sub-bundle of $\ns^P$, with factors $\hom_{\pi_i\co f}(P,\ns_i)$ and structure groups $\hom_{S\to 0}(P,\cD_i(\ab_i))$, where $\ab_i$ is the $i$-th structure group of $\ns$.
\end{lemma}
\noindent As a consequence of this structure we have a Haar measure on $\hom_f(P,\ns)$, by Proposition \ref{prop:nilspaceHaar}.
The second idea consists in giving a simple description of the measure-theoretic properties of a restriction map from one set of such morphisms to another. In particular, we want a convenient criterion to decide whether, for a subcubespace $S$ of $P$, the restriction map $\hom(P,\ns)\to \hom(S,\ns)$ preserves the Haar measures. The following notion provides a general criterion of this type.
\begin{defn}\label{defn:GoodPair}
Let $P$ be a cubespace, and let $P_1,P_2$ be subcubespaces of $P$. We call the pair of sets $P_1,P_2$ a \emph{good pair} if $P_1$ and $P_1\cap P_2$ both have the extension property in $P$ and, for every abelian group $\ab$ and every $k\in \mathbb{N}$, every morphism $f':P_2\to \cD_k(\ab)$ satisfying $f'|_{P_1\cap P_2}=0$ extends to a morphism $f:P\to \cD_k(\ab)$ satisfying $f|_{P_1}=0$.
\end{defn}
\noindent Note that in particular if $S$ has the extension property in $P\subset \{0,1\}^n$ and $P_1=\emptyset$ then $P_1,S$ is a good pair in $P$.
The main purpose of this definition is the following result.
\begin{lemma}\label{lem:GoodPairHoms}
Let $P\subset\{0,1\}^n$ be a subcubespace with the extension property, and let $P_1,P_2$ be a good pair in $P$. Let $\ns$ be a $k$-step nilspace and let $f:P_1\to \ns$ be a morphism. Then the restriction map
\begin{equation}\label{eq:restricthom}
\phi:\hom_f(P,\ns)\to \hom_{f|_{P_1\cap P_2}} (P_2,\ns)
\end{equation}
is a totally-surjective bundle morphism. In particular, if $\ns$ is a $k$-step compact nilspace, then $\phi$ preserves the Haar measures. Moreover, the Haar measures on the fibres $\phi^{-1}(t)$ form a CSM on $\phi$ that disintegrates the Haar measure on $\hom_f(P,\ns)$ relative to the Haar measure on $\hom_{f|_{P_1\cap P_2}} (P_2,\ns)$.
\end{lemma}
\begin{proof}
We prove that $\phi$ is a totally-surjective bundle morphism by induction on $k$. For $k=0$ the claim holds trivially, so let $k\geq 1$ and suppose that $\ns$ is a $k$-step nilspace and that the lemma holds for $(k-1)$-step nilspaces. To see that condition (i) from \cite[Definition 3.3.1]{Cand:Notes1} holds, note that by a straightforward calculation we have that $\phi$ induces a map $\phi_{k-1}:\hom_{\pi_{k-1}\co f}(P,\ns_{k-1})\to \hom_{\pi_{k-1}\co f|_{P_1\cap P_2}} (P_2,\ns_{k-1})$ well-defined by $\pi_{k-1}\co g\mapsto \pi_{k-1}\co \phi(g)$, so the condition holds for $i=k-1$. We also have that $\phi_{k-1}$ is precisely the restriction map on $\hom_{\pi_{k-1}\co f}(P,\ns_{k-1})$, so by induction the condition holds for all $i\leq k-1$. Condition (ii) from the same definition is clearly satisfied, with the structure morphism $\alpha_i$ being the restriction $\hom_{P_1\to 0}(P,\cD_i(\ab_i))\to \hom_{P_1\cap P_2\to 0}(P_2,\cD_i(\ab_i))$. To check that $\phi$ is totally surjective, by induction it suffices to check that $\alpha_k$ is surjective, which holds by Definition \ref{defn:GoodPair}. Lemma \ref{lem:MeasPres} then implies that $\phi$ preserves the Haar measures. The last sentence of the lemma follows from Lemma \ref{lem:quotint}.
\end{proof}
Another useful feature of good pairs is the following extension result for morphisms.
\begin{lemma}\label{lem:Gpair-union-ext}
Let $P$ be a subcubespace of $\{0,1\}^n$ with the extension property, and let $P_1,P_2$ be a good pair in $P$. Then $P_1\cup P_2$ equipped with the union of the cube structures on $P_1,P_2$ has the following extension property: any morphism $f:P_1\cup P_2\to \ns$ into a non-empty $k$-step nilspace $\ns$ extends to a morphism $f':P\to \ns$.
\end{lemma}
\begin{proof}
We argue again by induction on $k$. The statement is trivial for $k=0$, as $\ns$ is then a 1-point space, so suppose that the lemma holds for $(k-1)$-step nilspaces and let $\ns$ be a $k$-step nilspace.
Let $f:P_1\cup P_2\to \ns$ be a morphism and let $f_2:P\to\cF_{k-1}(\ns)$ be an extension of $\pi_{k-1}\co f$. By \cite[Lemma 3.2.11]{Cand:Notes1} there is a morphism $f_2':P\to \ns$ such that $\pi_{k-1}\co f_2'=f_2$. Let $g=f'_2|_{P_1}-f|_{P_1}$. By Definition \ref{defn:GoodPair}, there is an extension $f_3:P\to\cD_k(\ab_k)$ of $g$. Let $g_2=f_2'-f_3$. We have that $g_2|_{P_1}=f|_{P_1}$. Now let $g_3$ be an extension of $g_2|_{P_2}-f|_{P_2}$ to $P$ with $g_3|_{P_1}=0$. Then $f'=g_2-g_3$ is an extension of $f$ to $P$.
\end{proof}
\noindent For the remainder of this section we apply the results above to define various probability spaces of morphisms. We begin with the cube sets $\cu^n(\ns)$ on a compact nilspace. Note that $\cu^n(\ns)=\hom(\{0,1\}^n,\ns)$ is a compact $k$-fold abelian bundle and therefore has a Haar measure (by Proposition \ref{prop:nilspaceHaar}).
\begin{lemma}\label{lem:surjmorphcubemeas}
Let $\ns,\nss$ be $k$-step compact nilspaces and let $n\in \mathbb{N}$. Then for every continuous fibre-surjective morphism $\psi:\ns \to \nss$, the induced map $\cu^n(\ns)\to \cu^n(\nss)$, $\q\mapsto \psi\co\q$ preserves the Haar measures. Moreover, the fibres of this map can all be equipped with the Haar measure.
\end{lemma}
\begin{proof}
By \cite[Lemma 3.3.12 (ii)]{Cand:Notes1} we have that $\q\mapsto \psi\co\q$ is a totally surjective bundle morphism, so it preserves the Haar measures by Lemma \ref{lem:MeasPres}. Moreover, by \cite[Lemma 3.3.12 (iii)]{Cand:Notes1}, the fibres of this map are compact $k$-fold abelian bundles, so the last claim in the lemma follows from Proposition \ref{prop:nilspaceHaar}.
\end{proof}
\noindent We shall often want to use a disintegration of the Haar measure on $\cu^n(\ns)$ into a CSM on the projection $\q\mapsto \q(0^n)$. For this purpose we record the following special case of Lemma \ref{lem:GoodPairHoms}.
\begin{lemma}\label{lem:cube-set-CSM}
Let $\ns$ be a $k$-step compact nilspace with Haar measure $\mu$, let $\pi:\cu^n(\ns)\to \ns$, $\q\mapsto \q(0^n)$, and let $\mu_n$ denote the Haar measure on $\cu^n(\ns)$. Then for every $x\in \ns$, the space $\pi^{-1}(x)=\cu^n_x(\ns)$ is a compact $k$-fold abelian bundle and therefore has a Haar measure $\mu_x$. The family $\{\mu_x:x\in \ns\}$ is a CSM on $\pi$ disintegrating $\mu_n$ relative to $\mu$.
\end{lemma}
\begin{proof}
By Lemma \ref{lem:GoodPairHoms} with $P=\{0,1\}^n$, with $P_1=\emptyset$, $P_2=\{0^n\}$, identifying $\ns$ with $\hom(P_2,\ns)$.
\end{proof}
\noindent We move on to some results concerning the tricube $T_n$. Recall from \cite[Subsection 3.1.3]{Cand:Notes1} the definitions of tricubes, of morphisms $\trem_v$, and of the outer-point map $\omega_n:\{0,1\}^n\to T_n$.
\begin{lemma}\label{lem:ope-to-tricube-ext}
The set $\omega_n(\{0,1\}^n)\subset T_n$ has the extension property in $T_n$.
\end{lemma}
\begin{proof}
Let $\ns$ be a nilspace, let $P=\omega_n(\{0,1\}^n)$ and let $f:P\to \ns$ be a morphism. Let $h:\{-1,0,1\}\to \{0,1\}$ be the map $-1\mapsto 0$, $0\mapsto 1$, $1\mapsto 1$. Then the map $f'=f\co h^n$ is an extension of $f$ to $T_n$.
\end{proof}
\begin{lemma}\label{lem:subcube-in-Tn}
For every $v\in \{0,1\}^n$, the subcube $\Psi_v(\{0,1\}^n)$ of $T_n$ has the extension property in $T_n$.
\end{lemma}
\begin{proof}
The claim is checked for $v=0^n$ using the map $h:\{-1,0,1\}\to \{0,1\}$, $-1\mapsto 1$, $0\mapsto 0$, $1\mapsto 1$. (Indeed then $f\co h^n$ extends $f$ to $T_n$.) The lemma then follows by the composition axiom.
\end{proof}
\begin{corollary}\label{cor:tricubeprobspaces}
Let $P\subset T_n$ be $\Psi_v(\{0,1\}^n)$ or $\omega_n(\{0,1\}^n)$, let $\ns$ be a $k$-step compact nilspace, and let $f:P \to \ns$ be a morphism. Then $\hom_f(T_n,\ns)$ has a Haar measure.
\end{corollary}
\begin{proof}
This follows again from Lemma \ref{lem:comp-restrmorph=subbund} and Proposition \ref{prop:nilspaceHaar}.
\end{proof}
\begin{lemma}\label{lem:Gpair-in-Tn}
The sets $P_1=\omega_n(\{0,1\}^n)=\{-1,1\}^n$ and $P_2=\{0,1\}^n$ form a good pair in $T_n$.
\end{lemma}
\begin{proof}
Note first that we may assume that $\omega_n\in \cu^n(T_n)$, since by \cite[Lemma 3.1.16]{Cand:Notes1} this assumption does not affect any set of morphisms from $T_n$ into any nilspace. Now since $P_1\cap P_2=\{1^n\}$ is a singleton, it has the extension property. The set $P_1$ itself has the extension property, by Lemma \ref{lem:ope-to-tricube-ext}.
Let $f:\{0,1\}^n \to \cD_k(\ab)$ be a morphism with $f(1^n)=0$. Let $h(-1)=1$, $h(0)=0$, $h(1)=1$. Then $f'=f \circ h^n$ is an extension of $f$ with $f|_{P_1}=0$.
\end{proof}
\noindent Combining this with Lemma \ref{lem:GoodPairHoms}, we obtain that a set of morphisms $T_n\to \ns$ with a constraint at the outer points of $T_n$ is still large enough to cover a whole set of rooted cubes $\cu^n_x(\ns)$, in the following sense.
\begin{corollary}\label{cor:ext-in-Tn}
Let $\ns$ be a $k$-step compact nilspace, let $P_1$ and $P_2$ be as in Lemma \ref{lem:Gpair-in-Tn}, and let $f:P_1\to \ns$ be a morphism. Then the restriction map $\hom_f(T_n,\ns)\,\to\, \hom_{0^n\to f(1^n)}(P_2,\ns)=\cu^n_{f(1^n)}(\ns)$, $\;t\mapsto t\co \trem_{0^n}$ preserves the Haar measures.
\end{corollary}
\noindent As a complement to this corollary, we also have that a set of morphisms $T_n\to \ns$ with a given constraint at just one outer point is always large enough to cover a full cube set $\cu^n(\ns)$, as follows.
\begin{lemma}\label{lem:key-tricube-mp}
Let $\ns$ be a $k$-step compact nilspace, let $x\in \ns$, and let $Q_x$ denote the compact abelian bundle $\hom_{1^n\mapsto x}(T_n,\ns)$. Then for every $v\in\{0,1\}^n\setminus \{0^n\}$, the map $Q_x\to \cu^n(\ns)$, $t\mapsto t\circ\Psi_v$ preserves the Haar measures.
\end{lemma}
\begin{proof}
There is a Haar measure on $Q_x$ by Lemma \ref{lem:comp-restrmorph=subbund}. Let $P_1=\{1^n\}$ and $P_2=\Psi_v(\{0,1\}^n)$. The lemma will follow if we show that $P_1,P_2$ form a good pair in $T_n$, by Lemma \ref{lem:subcube-in-Tn}. Since the singleton $P_1$ has the extension property in $T_n$ and so does $P_1\cap P_2=\emptyset$, it suffices to show that any given morphism $f':P_2\to \cD_k(\ab)$ can be extended to a morphism $f:T_n\to\cD_k(\ab)$ satisfying $f|_{P_1}=0$. We can certainly extend $f'$ to a morphism $f_2:T_n\to\cD_k(\ab)$, by Lemma \ref{lem:subcube-in-Tn}. Now $v\neq 0^n$ implies that $\trem_v(0^n)\neq 1^n$, so this has some coordinate equal to $-1$, say without loss of generality $\trem_v(0^n)\sbr{1}=-1$. Then the set $F=\{1\}\times\{-1,0,1\}^{n-1}$ is disjoint from $P_2$. Let $g:T_n\to\cD_k(\ab)$ equal $f_2(1^n)$ on $F$ and $0$ otherwise. A straightforward calculation shows that $g$ is a morphism, using \cite[(2.9)]{Cand:Notes1}. Then the function $f=f_2-g$ is also a morphism $T_n\to\cD_k(\ab)$, it extends $f'$, and satisfies $f|_{P_1}=0$.
\end{proof}
\section{Topological spaces associated with continuous systems of measures}\label{subsec:CSM-bundle}
\medskip
\begin{defn}
Let $X,Y$ be compact spaces and let $\mu$ be a Borel measure on $X$. We denote by $L(X,Y)$ the quotient of the set of Borel measurable functions $f:X\to Y$ by the equivalence relation $\sim$ defined by $f\sim g\;$ $\Leftrightarrow$ $\;\mu\big(\{x\in X: f(x)\neq g(x)\}\big)=0$.
\end{defn}
\noindent In this section, given a continuous system of measures, we construct an associated topological space that will be crucial in the sequel. In order to motivate the construction, let us outline how it will be used.
The central application will be given in the next section, which is one of the main sections of this chapter. The aim there is to show that given a compact nilspace $\ns$ and a compact abelian group $\ab$, given a cocycle $\rho:\cu^k(\ns)\to \ab$ (recall \cite[Definition 3.3.14]{Cand:Notes1}), if $\rho$ is Borel measurable then the extension $M(\rho)$ constructed in \cite[Proposition 3.3.26]{Cand:Notes1} can be equipped with a compact nilspace structure compatible with that of $\ns$. The purpose of the construction in this section is to provide an ambient continuous abelian bundle from which $M(\rho)$ will inherit the desired topological structure. The construction will be completed with Proposition \ref{prop:CSM-dif-bund}.\\
\indent Let us now turn to the definition of the topological space in the construction. To motivate this, consider the following general theme of the application mentioned above: given a Borel measurable function on a space (e.g. the cocycle $\rho$), we have to build a topological space that is associated with the function in a useful way. A very basic example of such a construction is the following: let $\ab$ be a compact abelian group, suppose that we only know the Borel $\sigma$-algebra on $\ab$ (not the underlying topology), and that we are given a Borel function $f:\ab \to \mathbb{R}$. Then we can build a topology on $\ab$ naturally associated with $f$ using the $L^1$ seminorm on the Borel functions on $\ab$ (relative to the Haar measure), by restricting this seminorm to the translates of $f$, i.e. the functions $f_z:x\mapsto f(x+z)$, $z\in \ab$. More precisely, we can define a pseudometric $d$ on $\ab$ by $d(z,z')=\|f_z-f_{z'}\|_{L^1(\ab)}$, and then take the topology generated by the open balls with respect to $d$. The central application mentioned above will use a more elaborate version of this idea. Indeed, recall from \cite[(3.11)]{Cand:Notes1} that the extension generated by the cocycle $\rho$ is $M(\rho)=\bigcup_{x\in \ns} \{\rho_x+z:z\in \ab\}$, where for each $x$ we denote by $\rho_x$ the restriction of $\rho$ to the set of rooted cubes $\cu^k_x(\ns)=\{\q\in\cu^k(\ns):\q(0^k)=x\}$. To put a useful topology on $M(\rho)$, we shall first put an analogue of the $L^1$-topology on each family of shifts $\rho_x +z,\,z\in \ab$, for each $x\in \ns$, and we shall then want to tie together adequately these different topologies over different points $x$ into a global topology. Now the Haar measures on sets $\cu^k_x(\ns)$ form a CSM, by Lemma \ref{lem:cube-set-CSM}. This leads to the following definition, which concerns CSMs more generally, and which provides a topology that will be adequate for our purposes.
\begin{defn}\label{def:extspace}
Let $V,W,Z$ be compact spaces, and let $\{\mu_w: w\in W\}$ be a family of strictly positive Borel probability measures forming a CSM on a continuous map $\pi:V\to W$. For each function $f$ in $\bigcup_{w\in W} L\big(\pi^{-1}(w),Z\big)$, let $\tilde\pi(f)$ be the element $w\in W$ such that $f\in L(\pi^{-1}(w),Z)$. We define the topological space $\cL(V,Z)$ to be the set $\bigcup_{w\in W} L\big(\pi^{-1}(w),Z\big)$ equipped with the coarsest topology making the following functions continuous:
\begin{equation}\label{eq:initopfns}
\varphi_{F_1,F_2}:f\mapsto \int_{\pi^{-1}(\tilde{\pi}(f))}F_1(f(v))\;F_2(v)~\,\mathrm{d}\mu_{\tilde{\pi}(f)}(v),
\end{equation}
for every pair of continuous functions $F_1:Z\to\mathbb{C}$ and $F_2:V\to\mathbb{C}$.
\end{defn}
\noindent We shall discuss below how this topology relates to the $L^1$ topology, but for now let us record some of its basic properties.
\begin{proposition}\label{prop:CSM-top}
The topological space $\cL(V,Z)$ is regular Hausdorff and second-countable, and the map $\tilde\pi$ is continuous.
\end{proposition}
In particular $\cL(V,Z)$ is metrizable, by Urysohn's theorem \cite[Theorem 34.1]{Munkres}.
\begin{proof}
We show first that $\tilde\pi$ is continuous. Fix a metric on $W$, and let $U$ be any open subset of $W$. Then for any $w\in U$ there is an open ball $B_r(w)\subset U$ of radius $r>0$. By Urysohn's lemma \cite[Theorem 33.1]{Munkres} there exists a continuous function $f_w$ with value 1 on the closure of $B_{r/2}(w)$ and value 0 outside $B_r(w)$. Let $B_w'$ denote the open set $\{x\in W: f_w(x)>0 \}$, and note that $U=\bigcup_{w\in U} B_w'$. By definition of $\cL(V,Z)$, for every $w\in U$ the function $\varphi_{1,f_w\co\pi}$ is continuous. Since $\varphi_{1,f_w\co\pi}(f)=f_w(\tilde\pi(f))$, this function is positive precisely on $\tilde\pi^{-1}(B_w')$, so this set is open. Therefore $\tilde\pi^{-1}(U)=\bigcup_{w\in U} \tilde\pi^{-1}(B_w')$ is an open set.\\
\indent Next we show that $\cL(V,Z)$ is Hausdorff. Let $f,f'$ be distinct points in $\cL(V,Z)$. We distinguish two cases. In the first case $f,f'$ lie in different fibres of $\tilde\pi$. Then since $W$ is Hausdorff there exist disjoint open sets $U,U'$ containing $\tilde\pi(f)$, $\tilde\pi(f')$ respectively. The preimages of these sets under $\tilde\pi$ are open sets separating $f,f'$. In the second case we have $f,f'$ in the same fibre $L(\pi^{-1}(w),Z)$. Then there are Borel functions $f_0,f_0'$ in the classes $f,f'$ respectively such that the set $D=\{v:f_0(v)\neq f_0'(v)\}$ has $\mu_w(D)>0$. For some $\epsilon<\mu_w(D)$ to be fixed later, by Lusin's theorem \cite[Appendix E, \S E8]{Rudin} there exist continuous functions $g,g'$ and a set $C\subset \pi^{-1}(w)$ of probability at most $\epsilon$ such that $f_0=g$ and $f_0'=g'$ outside $C$. Then there is a point $v_0\in D\setminus C$ where $g(v_0)\neq g'(v_0)$. By continuity there exist $r_1,r_2,r_3>0$ such that for all $v\in B_{r_1}(v_0)\subset V$ we have $g(v)\in B_{r_2}(g(v_0))\subset Z$, $g'(v)\in B_{r_3}(g'(v_0))$, and $B_{r_3}(g'(v_0))\cap B_{2r_2}(g(v_0))=\emptyset$. Urysohn's lemma gives us continuous functions $F_1:Z\to [0,1]$, $F_2:V\to [0,1]$, such that $1_{B_{r_2}(g(v_0))}\leq F_1\leq 1_{B_{2r_2}(g(v_0))}$ and $1_{B_{r_1/2}(v_0)}\leq F_2 \leq 1_{B_{r_1}(v_0)}$. We then have $\varphi_{F_1,F_2}(g)\geq \mu_w(B_{r_1/2}(v_0))>0$, whereas $\varphi_{F_1,F_2}(g')=0$. We also have $|\varphi_{F_1,F_2}(g)-\varphi_{F_1,F_2}(f_0)|\leq 2 \mu_w(C)$ and similarly for $g',f_0'$. Hence, choosing $\epsilon$ small enough, we deduce that $\varphi_{F_1,F_2}(f_0)>\varphi_{F_1,F_2}(f_0')$.\\
\indent To see that $\cL(V,Z)$ is regular, first note that since by definition the finite intersections of sets $\varphi_{F_1,F_2}^{-1}(U)$ form a base for the topology, we have that for every closed set $S$ and $f\not\in S$, there exist $\varphi_{F_{1,i},F_{2,i}}^{-1}(U_i)$, $i\in [N]$ such that $f\in \bigcap_i \varphi_{F_{1,i},F_{2,i}}^{-1}(U_i)$ and $S\subset \bigcup_i \varphi_{F_{1,i},F_{2,i}}^{-1}(\mathbb{C}\setminus U_i)$. Since $\mathbb{C}$ is regular, for each $i$ there exist disjoint open sets $U_i',U_i''$ such that $\varphi_{F_{1,i},F_{2,i}}(f)\in U_i'\subset U_i$ and $\mathbb{C}\setminus U_i \subset U_i''$. Then the open sets $\bigcap_i \varphi_{F_{1,i},F_{2,i}}^{-1}(U_i')$ and $\bigcup_i \varphi_{F_{1,i},F_{2,i}}^{-1}(U_i'')$ are disjoint, the former contains $f$, and the latter includes $S$.\\
\indent Finally, we show that $\cL(V,Z)$ is second-countable. It suffices to find sequences of continuous functions $F_{1,i}:Z\to \mathbb{C}$, $F_{2,j}:V\to \mathbb{C}$ and open sets $U_k\subset \mathbb{C}$, $i,j,k\in \mathbb{N}$, such that for every $\varphi_{F_1,F_2}$, every open set $U\subset \mathbb{C}$, and every $f\in \cL(V,Z)$ with $\varphi_{F_1,F_2}(f)\in U$, there exist $i,j,k$ such that $f\in \varphi_{F_{1,i},F_{2,j}}^{-1}(U_k)\subset \varphi_{F_1,F_2}^{-1}(U)$. Indeed, if this holds then the finite intersections of sets $\varphi_{F_{1,i},F_{2,j}}^{-1}(U_k)$ form a countable base for the topology on $\cL(V,Z)$.
Let $\{U_k:k\in \mathbb{N}\}$ be a base of open discs in $\mathbb{C}$. The spaces of continuous functions $C(V,\mathbb{C})$ and $C(Z,\mathbb{C})$, equipped with the uniform metric, are separable \cite[Theorem (4.19)]{Ke}. Let $(F_{1,i})_{i\in \mathbb{N}}$ and $(F_{2,j})_{j\in \mathbb{N}}$ be separating sequences for $C(V,\mathbb{C})$ and $C(Z,\mathbb{C})$ respectively. Given any $\varphi_{F_1,F_2}$, open set $U\subset \mathbb{C}$, and $f\in \varphi_{F_1,F_2}^{-1}(U)$, for some $\epsilon>0$ there is a disc $U_k\subset U$ of radius at most $\epsilon$ with center at most $\epsilon/2$ away from $\varphi_{F_1,F_2}(f)$ and such that the distance from $U_k$ to the complement of $U$ is at least $\epsilon$. There exist also $i,j$ such that $\|F_1- F_{1,i}\|_\infty<\epsilon/(4(1+\|F_2\|_\infty))$ and $\|F_2- F_{2,j}\|_\infty< \epsilon/(4(1+\|F_{1,i}\|_\infty))$. Then for every $g\in \cL(V,Z)$, letting $w=\tilde\pi(g)$, we have
\[
|\varphi_{F_1,F_2}(g)-\varphi_{F_{1,i},F_{2,j}}(g)| \leq \int_{\pi^{-1}(w)} |F_1(g(v)) F_2(v)- F_{1,i}(g(v)) F_{2,j}(v)|\;\,\mathrm{d}\mu_w(v).
\]
By the triangle inequality, this is at most $\|F_2\|_\infty \, \|F_1- F_{1,i}\|_\infty + \|F_{1,i}\|_\infty \, \|F_2- F_{2,j}\|_\infty < \epsilon/2$.
This implies that $f\in \varphi_{F_{1,i},F_{2,j}}^{-1}(U_k)\subset \varphi_{F_1,F_2}^{-1}(U)$.
\end{proof}
\begin{remark}
There are spaces $\cL(V,Z)$ that do not admit any complete metric. For example, let $V$ be the one-point compactification of the half-strip $[0,\infty)\times [0,1] \subset \mathbb{R}^2$, let $W$ be the one-point compactification of $[0,\infty)$, let $\pi:V\to W$ be the projection to the first coordinate, let $Z=\{0,1\}$, and for each $w\in W\setminus \{\infty\}$ let the interval $\pi^{-1}(w)$ be equipped with the Lebesgue probability measure. For each $n\in \mathbb{N}$ let $f_n$ be the function on $\pi^{-1}(\{n\})$ that takes value $0$ on $\{n\}\times [0,1/2)$ and $1$ on $\{n\}\times [1/2,1]$. The sequence $(f_n)$ is Cauchy for any compatible metric on $\cL(V,Z)$ but it does not converge to a function $\{\infty\}\to Z$. To see this, we can view $\cL(V,Z)$ as a subspace of a larger space $D(V,Z)=\bigcup_{w\in W} L\big(\pi^{-1}(w),P(Z)\big)$, where each fibre $L\big(\pi^{-1}(w),P(Z)\big)$ consists of probability-valued functions $\tilde f$ on $\pi^{-1}(w)$, i.e. functions mapping each $v\in \pi^{-1}(w)$ to a probability measure $p$ on $Z$. (We can identify $L\big(\pi^{-1}(w),P(Z)\big)$ with $L\big(\pi^{-1}(w),[0,1]\big)$ by identifying $\tilde f$ with the map $v\mapsto p_v(\{0\})$.) The topology on $D(V,Z)$ is the initial topology generated by the functions $\tilde f\mapsto \int_{\pi^{-1}(w)} \mathbb{E}\big(F_1\,|\,\tilde f(v)\big)\;F_2(v)~\,\mathrm{d}\mu_w(v)$. This restricts to the original topology on $\cL(V,Z)$. A straightforward calculation shows that $f_n$ converges in $D(V,Z)$ to the function $\tilde f$ mapping $\infty$ to the uniform distribution on $\{0,1\}$. (I am grateful to Szegedy for providing this example.)
\end{remark}
\noindent The topology on $\cL(V,Z)$ can be viewed as a generalization of the $L^1$ topology, as illustrated by the following result.
\begin{lemma}\label{lem:Ltop-restrict}
Let $V,W$ be compact spaces, let $Z$ be the closed unit disc in $\mathbb{C}$, and let $w\in W$. Then a sequence $(f_n)$ in $L(\pi^{-1}(w),Z)$ converges to $f$ in $\cL(V,Z)$ if and only if it converges to $f$ in $L^1(\pi^{-1}(w),Z)$.
\end{lemma}
\noindent This is a special case of a more general result that we shall use in the sequel, namely Lemma \ref{lem:Ltop-gen-restrict} below. This result involves the following generalization of the $L^1$ distance for functions taking values in a compact metric space.
\begin{defn}
Given a compact space $Y$ with a Borel probability measure, and a compact space $Z$ with a compatible metric $d$, we define the metric $d_1$ on $L(Y,Z)$ by
\begin{equation}\label{eq:gen-L1}
d_1(f_1,f_2)=\mathbb{E}_{v\in Y}\; d\big(f_1(v),f_2(v)\big).
\end{equation}
\end{defn}
\begin{lemma}\label{lem:Ltop-gen-restrict}
Let $V,W,Z$ be compact spaces, let $\{\mu_w:w\in W\}$ be a CSM on $\pi:V\to W$, let $d$ be a compatible complete metric on $Z$, and let $w\in W$. Then a sequence $(f_n)$ in $L(\pi^{-1}(w),Z)$ converges to $f$ in $\cL(V,Z)$ if and only if $d_1(f,f_n)\to 0$ as $n\to \infty$.
\end{lemma}
\begin{proof}
Let $\varphi_{F_1,F_2}$ be any of the functions generating the topology on $\cL(V,Z)$. We then have
\[
|\varphi_{F_1,F_2}(f_n)- \varphi_{F_1,F_2}(f)|\leq \|F_2\|_\infty \int_{\pi^{-1}(w)} |F_1(f_n(v))-F_1(f(v)) | \,\mathrm{d}\mu_w(v).
\]
We claim that if $d_1(f,f_n)\to 0$ then the last integral tends to 0 as $n\to\infty$. To see this fix any $\epsilon>0$ and note that by uniform continuity of $F_1$ there is $\delta_0>0$ such that if $d(f_n(v),f(v))<\delta_0$ then $|F_1(f_n(v))-F_1(f(v))|<\epsilon/2$. Fix a positive $\delta<\min(\delta_0,\epsilon/4 \|F_1\|_\infty )$. For $n$ sufficiently large we have $d_1(f_n,f)<\delta^2$, whence, by Markov's inequality, the set $D_\delta=\{v\in \pi^{-1}(w):d(f_n(v),f(v))\geq \delta\}$ has $\mu_w(D_\delta)\leq \delta$. It follows that
\[
\int_{\pi^{-1}(w)} |F_1(f_n(v))-F_1(f(v)) | \,\,\mathrm{d}\mu_w\; \leq\; \epsilon/2+ 2 \|F_1\|_\infty\, \mu_w(D_\delta) \;\leq\; \epsilon.
\]
The claim follows, and so $d_1(f_n,f)\to 0$ implies that $f_n\to f$ in $\cL(V,Z)$.\\
\indent For the converse, suppose that $f_n\to f$ in
$\mathcal{L}(V,Z)$. By \cite[Theorem (4.17)]{Ke} and \cite[Theorem 1]{Ander}, there exists a compact subset $B$ of the real Hilbert space $\ell^2$ such that there is a homeomorphism $F:Z\to B$. We have that $F\co f_n$, $F\co f$ are Bochner measurable functions in $L^2(V,\ell^2)$. We claim that if $F\co f_n\to F\co f$ in $L^2(V,\ell^2)$ (that is $\int_{\pi^{-1}(w)}\|F\co f_n(v) - F\co f(v)\|_{\ell^2}\,\mathrm{d}\mu_w \to 0$) then $d_1(f_n,f)\to 0$ as required. Indeed, for any $\epsilon>0$, the set $E=\{(z,z'):d(z,z')\geq \epsilon\}$ is closed in $Z\times Z$ so compact, and therefore the continuous function $\frac{d(z,z')}{\|F(z) - F(z')\|_{\ell^2}}$ on $E$ reaches a maximum $C$, whence for $n$ sufficiently large we have
\begin{eqnarray*}
d_1(f_n,f) & = & \int_{\pi^{-1}(w)} d\big(f_n(v),f(v)\big)\,\mathrm{d}\mu_w \;\;\leq\;\; \epsilon+C \int_{\pi^{-1}(w)}1_E\big(f_n(v),f(v)\big)\,\|F\co f_n(v) - F\co f(v)\|_{\ell^2}\,\mathrm{d}\mu_w\\
& \leq & 2 \epsilon.
\end{eqnarray*}
This proves our claim. We now show that $F\co f_n\to F\co f$ in $L^2(V,\ell^2)$.\\
\indent Firstly, for every $F_0\in L^2(V,\ell^2)$, we show that as $n\to \infty$ we have $\int_{\pi^{-1}(w)} \langle F\co f_n(v),F_0(v)\rangle_{\ell^2}\,\mathrm{d}\mu_w\to \int_{\pi^{-1}(w)} \langle F\co f(v),F_0(v)\rangle_{\ell^2}\,\mathrm{d}\mu_w$. By definition of the Bochner integral, it suffices to prove this for simple functions $F_0$. By linearity, it then suffices to prove it for $F_0$ of the form $\alpha\,1_A$ where $A\subset \pi^{-1}(w)$ is measurable and $\alpha\in \ell^2(\mathbb{R})$. But in this case the function $F_1:Z\to \mathbb{R}$, $z\mapsto \langle F(z),\alpha\rangle$ is continuous, and the function $1_A:\pi^{-1}(w)\to \mathbb{R}$ can be approximated in $L^1(\mu_w)$ by continuous functions, hence in this case the convergence follows from the definition of the topology on $\cL(V,Z)$.\\
\indent Secondly, we show that $\int_{\pi^{-1}(w)} \| F\co f_n(v) \|_{\ell^2}^2\,\mathrm{d}\mu_w$ converges to $\int_{\pi^{-1}(w)} \| F\co f(v) \|_{\ell^2}^2\,\mathrm{d}\mu_w$. This follows from the fact that $F_1:Z\to \mathbb{R}$, $z\to \| F(z) \|_{\ell^2}^2$ is continuous, so that letting $F_2$ be the function with value 1 on all of $V$ we have $\int_{\pi^{-1}(w)} \| F\co f_n(v) \|_{\ell^2}^2\,\mathrm{d}\mu_w=\varphi_{F_1,F_2}(f_n)$, which converges to $\varphi_{F_1,F_2}(f)$ as required.\\
\indent From the last two paragraphs we deduce the desired convergence by a standard argument, namely
\begin{eqnarray*}
&& \int_{\pi^{-1}(w)}\|F\co f_n(v) - F\co f(v)\|_{\ell^2}^2\,\,\mathrm{d}\mu_w \; = \; \int_{\pi^{-1}(w)} \langle F\co f_n - F\co f,F\co f_n - F\co f\rangle\, \,\mathrm{d}\mu_w\\
& = & \int \|F\co f_n\|_{\ell^2}^2 -2 \langle F\co f_n ,F\co f\rangle +\|F\co f\|_{\ell^2}^2\,\,\mathrm{d}\mu_w
\; \to \; \int 2\|F\co f\|_{\ell^2}^2 -2 \langle F\co f ,F\co f\rangle\, \,\mathrm{d}\mu_w\;= \;0.\qedhere
\end{eqnarray*}
\end{proof}
\noindent It is a basic fact that if $(f_n)$ is a sequence of real-valued functions on a standard probability space such that $f_n\to f$ in $L^1$ and each $f_n$ is constant almost surely, then the limit $f$ is also constant almost surely. We shall use the following analogous fact for $\cL(V,Z)$.
\begin{lemma}\label{lem:constconv}
Let $V,W,Z$ be compact spaces, let $\{\mu_w:w\in W\}$ be a CSM on $\pi:V\to W$, and let $(f_n)$ be a convergent sequence in $\cL(V,Z)$ such that each $f_n$ is constant almost surely on its fibre $\pi^{-1}(w_n)$. Then the limit is also constant almost surely.
\end{lemma}
\begin{proof}
Let $f$ be the limit of the sequence $(f_n)$. By Urysohn's lemma on $Z$, it suffices to show that for every continuous function $F:Z\to\mathbb{R}$ the function $F\co f$ is constant. Now since $F$ and $F^2$ are both continuous, by definition of the topology on $\cL(V,Z)$ we have $\int_{\pi^{-1}(w_n)} F\co f_n \,\mathrm{d}\mu_{w_n} \to \int_{\pi^{-1}(w)} F\co f \,\mathrm{d}\mu_w$ and $\int_{\pi^{-1}(w_n)} F^2\co f_n \,\mathrm{d}\mu_{w_n} \to \int_{\pi^{-1}(w)} F^2\co f \,\mathrm{d}\mu_w$. On the other hand, since each $f_n$ is constant almost everywhere, we have $\Big(\int_{\pi^{-1}(w_n)} F\co f_n \,\mathrm{d}\mu_{w_n}\Big)^2 - \int_{\pi^{-1}(w_n)} F^2\co f_n \,\mathrm{d}\mu_{w_n} =0$. Taking limits as $n\to\infty$, we deduce that $\Big(\int_{\pi^{-1}(w)} F\co f \,\mathrm{d}\mu_w\Big)^2 - \int_{\pi^{-1}(w)} F^2\co f \,\mathrm{d}\mu_w =0$. By the equality case of the Cauchy-Schwarz inequality, this implies that $F\co f$ is constant almost everywhere.
\end{proof}
The following result will be used several times in the next section.
\begin{lemma}\label{lem:prodCSMconv}
Let $V,W,Z$ be compact spaces, and let $\{\mu_w : w\in W\}$ be a CSM on $\pi:V\to W$. Suppose that $(f_n)$ and $(g_n)$ are sequences in $\cL(V,Z)$ converging to $f$ and $g$ respectively, where $f_n,g_n$ are defined on $\pi^{-1}(w_n)$ and $f,g$ are defined on $\pi^{-1}(w)$. Then the functions $v\mapsto (f_n(v),g_n(v))$ converge to $v\mapsto (f(v),g(v))$ in $\cL(V,Z\times Z)$.
\end{lemma}
\begin{proof}
It suffices to show that for any continuous functions $F_1:Z\times Z\to\mathbb{C}$ and $F_2:V\to\mathbb{C}$ we have $\lim_{n\to\infty}\int_{\pi^{-1}(w_n)}F_1(f_n(v),g_n(v))\, F_2(v)\,\mathrm{d}\mu_{w_n}(v)=\int_{\pi^{-1}(w)}F_1(f(v),g(v))\,F_2(v)\,\mathrm{d}\mu_w(v)$.
By the Stone-Weierstrass theorem the function $F_1$ can be approximated arbitrarily closely in the uniform metric by finite linear combinations of functions of the form $(z_1,z_2)\mapsto F_{1,1}(z_1)F_{1,2}(z_2)$ where $F_{1,1},F_{1,2}:Z\to\mathbb{C}$ are continuous, so we may assume that $F_1$ is of this form. Fix any $\epsilon>0$. By Lusin's theorem there exists a continuous function $q':\pi^{-1}(w)\to \mathbb{C}$ such that $\|F_{1,2}\co g-q'\|_{L^2}<\epsilon$ and $\|q'\|_\infty\leq \|F_{1,2}\|_\infty$, and then by the Tietze extension theorem \cite[Theorem 35.1]{Munkres} there exists a continuous function $q:V\to \mathbb{C}$ such that the restriction $q|_{\pi^{-1}(w)}$ equals $q'$ and $\|q\|_\infty\leq \|F_{1,2}\|_\infty$. For each $n$ let $q_n=q|_{\pi^{-1}(w_n)}$. Now
\begin{eqnarray}\label{eq:L2-arg}
\|F_{1,2}\co g_n - q_n\|_{L^2}^2 & = & \int_{\pi^{-1}(w_n)}|F_{1,2}\co g_n - q_n|^2\,\mathrm{d}\mu_{w_n} \nonumber \\
& = & \int_{\pi^{-1}(w_n)}|F_{1,2}\co g_n|^2 - 2\tRe\,(F_{1,2}\co g_n\,\overline{ q_n})+|q_n|^2\,\mathrm{d}\mu_{w_n}.
\end{eqnarray}
Using that $g_n\to g$ in $\cL(V,Z)$, that $w_n\to w$, and the uniform continuity of $q$, we deduce that the integral above converges to $\|F_{1,2}\co g -q' \|_{L^2}^2$. Setting $h_n=(F_{1,1}\co f_n) (F_{1,2}\co g_n-q_n)$, we then have
\[
F_1\co(f_n,g_n)=(F_{1,1}\co f_n) \big(q_n+(F_{1,2}\co g_n-q_n)\big)=(F_{1,1}\co f_n)\,q_n+h_n
\]
and similarly $F_1\co(f,g)=(F_{1,1}\co f)\, q'+h$, where for $n$ sufficiently large we have $\|h_n\|_{L^2}$ and $\|h\|_{L^2}$ both at most $2\epsilon \|F_{1,1}\|_\infty$. The convergence $f_n\to f$ also implies that
\[
\int_{\pi^{-1}(w_n)} (F_{1,1}\co f_n)\,q_n\, F_2\, \,\mathrm{d}\mu_{w_n}\;\to\; \int_{\pi^{-1}(w)}(F_{1,1}\co f)\, q'\, F_2\, \,\mathrm{d}\mu_w.
\]
Combining the above estimates we deduce that for all $n$ sufficiently large we have
\[
\Big|\int_{\pi^{-1}(w_n)}F_1\co (f_n,g_n)\, F_2\,\,\mathrm{d}\mu_{w_n}- \int_{\pi^{-1}(w)}F_1\co(f,g)\,F_2\,\mathrm{d}\mu_w\Big| \leq 5 \epsilon \|F_{1,1}\|_\infty\|F_2\|_\infty.
\]
Since $\epsilon> 0$ was arbitrary, the result follows.
\end{proof}
\noindent For the main applications of the space $\cL(V,Z)$ in the sequel, we will suppose that $Z$ is a compact abelian group. We now show that in this case the group has a continuous action on the space.
\begin{lemma}\label{lem:contact} Let $V,W$ be compact spaces, let $\ab$ be a compact abelian group, and let $\{\mu_w: w\in W\}$ be a CSM on $\pi:V\to W$. Then we have the following free continuous action of $\ab$ on $\cL(V,\ab)$:
\[
\alpha\,:\;\ab\times\cL(V,\ab)\;\to\;\cL(V,\ab),\quad \alpha(z, f): v\mapsto f(v) + z.
\]
\end{lemma}
\begin{proof}
The topology on $\cL(V,\ab)$ is generated by the functions $\varphi_{F_1,F_2}$, so to check continuity of $\alpha$ it suffices to show that for every $\varphi_{F_1,F_2}$ the map $\varphi_{F_1,F_2}\circ \alpha$ is continuous. Thus, it suffices to show that for any continuous functions $F_1 : \ab \to \mathbb{C}$ and $F_2 : V \to \mathbb{C}$, the following function is continuous:
\begin{equation}\label{eq:contact}
q: \;(z, f) \;\mapsto\; \int_{\pi^{-1}(\tilde{\pi}(f))} F_1(f(v)+z)\; F_2(v) \; \,\mathrm{d}\mu_{\tilde{\pi}(f)}(v).
\end{equation}
Applying the Stone-Weierstrass theorem as in previous proofs, the continuous function $\ab \times \ab \to \mathbb{C}$ defined by $(z_1,z_2) \mapsto F_1(z_1 + z_2)$ can be approximated uniformly by linear combinations of functions of the form $(z_1, z_2) \mapsto h_1(z_1)\,h_2(z_2)$ where $h_1$ and $h_2$ are continuous. Substituting any of these into \eqref{eq:contact}, we obtain a function of the form
\[
q':\;(z, f) \;\mapsto \;h_2(z) \int_{\pi^{-1}(\tilde{\pi}(f))} h_1(f(v))\; F_2(v) \; \,\mathrm{d}\mu_{\tilde{\pi}(f)}(v) \;=\;h_2(z)\, \varphi_{h_1,F_2}(f).
\]
This is continuous by definition of $\cL(V,\ab)$. Hence $q$ is a uniform limit of continuous functions and is therefore continuous.
\end{proof}
\noindent In our uses of $\cL(V,\ab)$, we shall want to view this space as a continuous $\ab$-bundle, for some appropriate base $S$ and projection $\cE$. We would like $S,\cE$ to have the following features. Firstly, the projection should send two functions $f_1,f_2$ in $\cL(V,\ab)$ to the same point of $S$ if and only if $f_1$ and $f_2$ are in the same $\ab$-orbit (i.e. $f_2=f_1+z$ for some $z\in \ab$). Secondly, we would like $S$ to be also of the form $\cL(Y,\ab)$, for some other similar space $Y$ having a projection to $W$, as this would help to relate the topologies on $S$ and $\cL(V,\ab)$.\\
\indent A natural way to obtain these features is to define $\cE$ as a difference map, sending $f$ to the function $\cE(f):V\times V\to \ab$, $(v_0,v_1)\mapsto f(v_0)-f(v_1)$. We are thus led to the following construction.
\begin{defn}
Let $V,W$ be compact spaces and let $\{\mu_w:w\in W\}$ be a CSM on $\pi:V\to W$. We define the compact space
\[
V\times_W V = \{(v_0,v_1): v_0,v_1\in V,\;\pi(v_0)=\pi(v_1)\},
\]
and define the projection $\pi':V\times_W V\to W$ by $\pi'(v_0,v_1)=\pi(v_0)$.
\end{defn}
\begin{lemma}
The family of measures $\{\mu_w\times\mu_w:w\in W\}$ on $V\times_W V$ is a CSM on $\pi'$.
\end{lemma}
\begin{proof}
For every $w\in W$ the measure $\mu_w$ is concentrated on $\pi^{-1}(w)$, so in $V\times V$ the measure $\mu_w\times \mu_w$ is concentrated on $\pi^{-1}(w) \times \pi^{-1}(w)={\pi'}^{-1}(w)$. For any continuous function $f:V\times_W V\to \mathbb{C}$, the continuity of $w\mapsto \int_{{\pi'}^{-1}(w)} f \,\mathrm{d} (\mu_w\times \mu_w)$ is proved by an argument similar to previous ones (e.g. in the proof of Lemma \ref{lem:contact}), showing that this function is a uniform limit of continuous functions $F_n$, where $F_n$ is obtained by approximating $f$ to within $1/n$ in the supremum norm by a finite linear combination of functions of the form $(v_0,v_1)\mapsto h_0(v_0)\,h_1(v_1)$ with $h_i:V\to \mathbb{C}$ continuous, and then using the continuity of $w\mapsto \prod_{i=0,1}\int_{\pi^{-1}(w)} h_i \,\mathrm{d} \mu_w$ guaranteed by the CSM on $\pi$.
\end{proof}
\noindent We are now able to complete the construction of the continuous abelian bundle announced at the beginning of this section, by showing that the map $\cE$ has the required properties to be a suitable projection.
\begin{proposition}\label{prop:CSM-dif-bund}
Let $V,W$ be compact spaces, let $\{\mu_w :w\in W\}$ be a CSM on $\pi:V\to W$, and let $\ab$ be a compact abelian group. Let $\cE:\cL(V,\ab)\to\cL(V\times_W V,\ab)$ be defined by $\cE(g): (v_0,v_1) \mapsto g(v_0)-g(v_1)$. Then $\cL(V,\ab)$ is a continuous $\ab$-bundle over $\cE(\cL(V,\ab))$ with projection $\cE$.
\end{proposition}
\begin{proof} Recall that the topology on $\cL(V\times_W V,\ab)$ is generated by functions $\varphi_{F_1,F_2}$ sending $f:V\times_W V\to \ab$ with $\tilde\pi(f)=w$ to $\int_{{\pi'}^{-1}(w)} F_1(f(v_0,v_1))\,F_2(v_0,v_1) \,\mathrm{d}\mu_w(v_0,v_1)$. To see that $\cE$ is continuous, fix any such function and suppose that $g_n\to g$ in $\cL(V,\ab)$. Then, approximating the continuous functions $(z_0,z_1)\mapsto F_1(z_0-z_1)$ on $\ab^2$ and $F_2$ on ${\pi'}^{-1}(w)=\pi^{-1}(w)\times \pi^{-1}(w)$ by combinations of continuous functions $h_0\,h_1$ as in previous proofs, we approximate $\varphi_{F_1,F_2}(\cE(g_n))$ by finite combinations of products $\varphi_0(g_n)\varphi_1(g_n)$, where $\varphi_0,\varphi_1$ are functions among those generating the $\cL(V,\ab)$ topology. By assumption, each of these products satisfies $\varphi_0(g_n)\varphi_1(g_n)\to \varphi_0(g)\varphi_1(g)$, and it follows that $\varphi_{F_1,F_2}(\cE(g_n))\to \varphi_{F_1,F_2}(\cE(g))$. This proves the continuity of $\cE$.\\
\indent To ensure that we have a continuous $\ab$-bundle, it only remains to check the `if' part of condition (iv) in Definition \ref{def:CpctAbBund}. Equivalently, it now suffices to prove that if $\cE^{-1}(U)$ is closed then $U$ is closed. For that, we claim it is enough to show that if for a sequence of functions $f_n\in\cL(V,\ab)$ the images $\cE(f_n)$ converge, then there is a convergent subsequence $(f_m)$ of $(f_n)$. Indeed, suppose that this holds and suppose for a contradiction that $U$ is not closed. Then (using the surjectivity of $\cE$) there is a sequence of functions $\cE(f_n)\in U$ converging to a function $\cE(f)\not\in U$. But then letting $f'$ be the limit of the guaranteed subsequence $(f_m)$, the closure of $\cE^{-1}(U)$ implies that $f'\in \cE^{-1}(U)$, and so by continuity of $\cE$ we have $U\ni \cE(f')=\lim_{m\to \infty} \cE(f_m) =\cE(f) \not\in U$, a contradiction. This proves our claim.\\
\indent So let us suppose that $\cE(f_n)$ converges to $\cE(f)$ for some $f\in\cL(V,\ab)$. Let $w_n=\tilde\pi(f_n)$ and $w=\tilde\pi(f)$, and note that $w_n\to w$ in $W$ by continuity of $\tilde{\pi}$.\\
\indent Let $\cC$ denote the circle group, viewed as the unit circle in $\mathbb{C}$. We shall prove the following fact:
\begin{eqnarray}\label{eq:cE-keyfact}
&& \textrm{There exists a subsequence }(f_m)\textrm{ and some }t\in \ab\textrm{ such that, for every character }\\
&& \chi:\ab\to \cC\textrm{ in }\widehat{\ab},\textrm{ we have }\chi\co f_m\to \chi\co\,(f+t)\textrm{ in }\cL(V,\cC)\textrm{ as }m\to \infty. \nonumber
\end{eqnarray}
We prove this below, but before that let us show how it implies that $f_m$ converges to $f+t$ in $\cL(V,\ab)$.\\
\indent Let $\varphi_{F_1,F_2}$ be any of the functions generating the $\cL(V,\ab)$ topology, and note that by the Stone-Weierstrass theorem for every $\epsilon>0$ there exists a trigonometric polynomial $F_1'=\sum_{j\in [K]} \lambda_j \chi_j$ on $\ab$ such that $\|F_1-F_1'\|_\infty \leq\epsilon$. Therefore it suffices to show that $\varphi_{F_1',F_2}(f_m)\to \varphi_{F_1',F_2}(f+t)$ for any such polynomial $F_1'$. We have
\[
\varphi_{F_1',F_2}(f_m)=\sum_{j\in [K]} \lambda_j \int_{\pi^{-1}(w_m)} (\chi_j\co f_m)\, F_2 \;\,\mathrm{d}\mu_{w_m}=\sum_{j\in [K]} \lambda_j\, \varphi_{1,F_2}(\chi_j\co f_m).
\]
Since $\varphi_{1,F_2}$ is one of the functions generating the topology on $\cL(V,\cC)$, we have $\varphi_{1,F_2}(\chi_j\co f_m)\to \varphi_{1,F_2}\big(\chi_j\co\,(f+t)\big)$ by assumption, and the desired convergence follows.\\
\indent We now prove \eqref{eq:cE-keyfact}. Given $\chi\in\widehat{\ab}$, the function $\chi\circ f$ is in $L^2(\mu_w)$. Continuous functions are dense in this space, and any such function can be extended to a continuous function on $V$ (by the Tietze extension theorem). Hence, for every $\delta>0$ there exists a continuous function $\chi_\delta:V\to\cC$ whose restriction to $\pi^{-1}(w)$ satisfies $\|\chi_\delta-\chi\co f\|_{L^2(\mu_w)}\leq \delta$. We claim that for every $\chi\in\widehat{\ab}$ and $\epsilon>0$, there exist $\delta>0$, $N\in\mathbb{N}$, and $\lambda\in \cC$, such that for all $n>N$ we have $\|(\chi\co f_n)\,\overline{\chi_\delta}-\lambda\|_{L^2(\mu_{w_n})} \leq \epsilon$.
To prove this claim, let $\cE':\cL(V,\cC)\to\cL(V\times_W V,\cC)$ be defined by $\cE'(g):(v_0,v_1)\mapsto g(v_0)\overline{g(v_1)}$. We have that $\cE'$ is continuous, by an argument similar to the one used above for $\cE$. It is also clear that for any $g_1,g_2\in \cL(V,\cC)$ lying in the same fibre of $\tilde\pi$ we have $\cE'(g_1g_2)=\cE'(g_1)\cE'(g_2)$, and that $\cE'(\chi\circ g)=\chi\circ\cE(g)$ for every $\chi\in\widehat{\ab}$, $g\in \cL(V,\ab)$. Now, from the definition of $\chi_\delta$, the continuity of $\cE'$, and Lemma \ref{lem:Ltop-gen-restrict}, it follows that for $\delta$ sufficiently small we have $\big\| \cE'\big((\chi\circ f)\,\overline{\chi_\delta}\big)-1\big\|_{L^2(\mu_w\times\mu_w)}\leq \epsilon/2$. On the other hand, we have $\cE'\big((\chi\circ f_n)\overline{\chi_\delta}\big)=\cE'(\chi\circ f_n)\,\cE'(\overline{\chi_\delta})=\chi\co \cE(f_n)\,\cE'(\overline{\chi_\delta})$ and similarly for $f$. This together with the assumption $\cE(f_n)\to\cE(f)$ implies, by an argument similar to \eqref{eq:L2-arg}, that for $n$ sufficiently large we have $\big\| \cE'\big((\chi\circ f_n)\,\overline{\chi_\delta}\big)-1\big\|_{L^2(\mu_{w_n}\times \mu_{w_n})}\leq \epsilon$. Taking a minimum over $v_1$ in this integral, we deduce that there exists $v_1\in \pi^{-1}(w_n)$ such that, for $\lambda= (\chi\circ f_n)(v_1)\overline{\chi_\delta}(v_1)$, we have indeed $\|(\chi\circ f_n)\overline{\chi_\delta}-\lambda\|_{L^2(\mu_{w_n})}\leq \epsilon$.
Since there are at most countably many elements in $\widehat{\ab}$, it follows by a standard diagonalization argument (using the above claim with smaller and smaller $\epsilon$ for each $\chi$) that for some sequence $(m_n)_{n\in \mathbb{N}}$ of positive integers, for every $\chi\in\widehat{\ab}$ there is a constant $\lambda_\chi\in\cC$ such that $\chi\co f_{m_n}\to \lambda_\chi (\chi\co f)$ in $\cL(V,\cC)$ as $n\to\infty$. A straightforward calculation shows that the function $\chi\mapsto \lambda_\chi$ is a homomorphism, hence a character on $\widehat{\ab}$, which by duality must be of the form $\chi\mapsto \chi(t)$ for some $t\in \ab$. This confirms \eqref{eq:cE-keyfact} and completes the proof.
\end{proof}
We close this section with two technical lemmas that we shall use several times.
\begin{lemma}\label{lem:csmtech} Let $\{\mu_w:w\in W\}$ be a CSM on $\pi:V\to W$. Let $K$ be a compact space with a Borel measure $\nu$. Suppose that $f_1:V \to K$ is continuous and that, for every $w\in W$, the restriction of $f_1$ to $\pi^{-1}(w)$ satisfies $\mu_w\co f_1^{-1}=\nu$. Let $Z$ be a compact space and let $f_2:K\to Z$ be a Borel function. Let $h: W\to\cL(V,Z)$ be the map $w\;\mapsto \;f_2\co f_1|_{\pi^{-1}(w)}$. Then $h$ is continuous.
\end{lemma}
\begin{proof} By definition of the topology on $\cL(V,Z)$ it suffices to show that for every $\varphi_{F_1,F_2}$ the function $\varphi_{F_1,F_2}\circ h:W\to \mathbb{C}$ is continuous. So let $F_1:Z\to\mathbb{C}$ and $F_2:V \to \mathbb{C}$ be continuous and let $q$ denote $\varphi_{F_1,F_2}\circ h$, that is $q: \; w\;\mapsto\;\int_{\pi^{-1}(w)}\;F_1(f_2\circ f_1(v))\;F_2(v)~\,\mathrm{d}\mu_w(v)$. To prove that $q$ is continuous we show that it is a uniform limit of continuous functions.
Fix an arbitrary $\epsilon>0$. By Lusin's theorem there exists a continuous function $F_3:K\to \mathbb{C}$ such that $\|F_1\co f_2-F_3\|_{L^1(\nu)}\leq\epsilon$. Letting $q': w\mapsto\int_{\pi^{-1}(w)}\;F_3(f_1(v))\;F_2(v)~\,\mathrm{d}\mu_w(v)$, the triangle inequality and the assumption that $f_1$ is measure-preserving imply that $\sup_{w\in W}|q(w)-q'(w)|\leq \epsilon \|F_2\|_\infty$. The function $q'$ is continuous by condition \eqref{contprop} from Definition \ref{def:CSM}. The result follows.
\end{proof}
\begin{lemma}\label{lem:CSMlincombin}
Let $\{\mu_w:w\in W\}$ be a CSM on $\pi:V\to W$, let $\ab$ be a compact abelian group, let $P$ be a finite set, and for each $r\in P$ let $\lambda_r$ be an integer. Let $\Sigma_P:\cL(V,\ab^P)\to \cL(V,\ab)$, $f\mapsto \sum_{r\in P} \lambda_r\, \pi_r\co f$, where $\pi_r$ is the projection to the $r$-component on $\ab^P$. Then $\Sigma_P$ is continuous.
\end{lemma}
\begin{proof}
For any of the functions $\varphi_{F_1,F_2}$ generating the topology on $\cL(V,\ab)$, letting $w=\tilde\pi(f)$, we have
$\varphi_{F_1,F_2}(\Sigma_P(f))=\int_{\pi^{-1}(w)} F_1\co\Sigma_P(f)\, F_2\, \,\mathrm{d}\mu_w$. But $F_1\co \Sigma_P(f)=F_1'\co f$, where $F_1':\ab^P\to \mathbb{C}$, $(z_r)_{r\in P}\mapsto F_1\big(\sum_r \lambda_r z_r\big)$ is continuous. Hence, by definition of $\cL(V,\ab^P)$, if $f_n\to f$ in this space then $\varphi_{F_1,F_2}( \Sigma_P(f_n))\to\varphi_{F_1,F_2}( \Sigma_P(f))$, and since this holds for every $\varphi_{F_1,F_2}$, we have $\Sigma_P(f_n)\to \Sigma_P(f)$ in $\cL(V,\ab)$.
\end{proof}
\medskip
\section{Borel measurable cocycles generate compact extensions}
\medskip
Let $\ns$ be a compact nilspace and let $\ab$ be a compact abelian group. The purpose of this section is to show that any measurable cocycle $\rho:\cu^k(\ns)\to\ab$ yields a compact nilspace that is a $\ab$-bundle over $\ns$.
Recall that for each $x\in \ns$ we denote by $\cu^k_x(\ns)$ the set of $k$-cubes $\q$ satisfying $\pi(\q):=\q(0^k)=x$. By Lemma \ref{lem:cube-set-CSM} each space $\cu^k_x(\ns)$ has a Haar measure, denoted $\mu_x$, and the family $\{\mu_x: x\in \ns\}$ is a CSM on $\pi$ of strictly positive measures.
\begin{defn}
We denote by $\cL_k(\ns,\ab)$ the space $\cL(V,\ab)$ with $V=\cu^n(\ns)$, $W=\ns$, and $\pi:\q\mapsto \q(0^k)$.
\end{defn}
\noindent Thus $\cL_k(\ns,\ab)=\bigcup_{x\in \ns} L\big(\cu_x^k(\ns),\ab\big)$, its topology is second-countable regular Hausdorff, and we have a projection $\tilde\pi:\mathcal{L}_k(\ns,\ab)\to \ns$ defined by $\tilde\pi(f)=x$ for $f\in L(\cu_x^k(\ns),\ab)$. For such a function $f$, we denote by $f+\ab$ the set of functions $\{f+z: z\in \ab\}$.
The desired nilspace extending $\ns$ is found inside $\cL_k(\ns,\ab)$ by taking, for each $x\in \ns$, the orbit of the restricted cocycle $\rho_x=\rho|_{\cu_x^k(\ns)}$.
\begin{proposition}\label{prop:cocyclext}
Let $\rho:\cu^k(\ns)\to \ab$ be a Borel measurable cocycle and let $M=\bigcup_{x\in \ns} \,(\rho_x+\ab)$. Then, as a subspace of $\cL_k(\ns,\ab)$ with the inherited $\ab$-action, the space $M$ is a compact $\ab$-bundle over $\ns$.
\end{proposition}
\begin{proof}
Let $Y=\cu^k(\ns)\times_{\ns} \cu^k(\ns)$, that is (recalling the notation from Proposition \ref{prop:CSM-dif-bund}),
\[
Y\,=\,\big\{ (\q_0,\q_1):\,\q_0,\q_1 \in \cu^k(\ns),\, \q_0(0^k)=\q_1(0^k)\big\}\,=\,\bigcup_{x\in \ns} \cu^k_x(\ns)\times \cu^k_x(\ns).
\]
Recall that $\cE:\cL(\cu^k(\ns),\ab)\to \cL(Y,\ab)$ is defined by $\cE(f)\;:\; (\q_0,\q_1)\mapsto f(\q_0)-f(\q_1)$. We now compose the function $x\mapsto \rho_x$ with $\cE$, obtaining the function
\begin{eqnarray}
g':\;\ns & \to & L\big(\,\cu^k_x(\ns)\times \cu^k_x(\ns)\,,\,\ab\,\big)\subset \cL(Y,\ab)\\
x & \mapsto & \cE(\rho_x).\nonumber
\end{eqnarray}
Note that $g'$ is injective, and that its inverse is the projection $\tilde\pi$ on $\cL(Y,\ab)$, a continuous map. We have
\begin{equation}\label{eq:MoverX}
M=\cE^{-1}(g'(\ns)).
\end{equation}
Indeed, if $f\in \cL(\cu^k(\ns),\ab)$ satisfies $\cE(f)=\cE(\rho_x)$, then for every $(\q_0,\q_1)\in \cu^k_x(\ns)^2\subset Y$ we have $f(\q_0)-\rho_x(\q_0)= f(\q_1)-\rho_x(\q_1)$, so $f-\rho_x$ is a constant $\ab$-valued function on $\cu^k_x(\ns)$.
Now, by Proposition \ref{prop:CSM-dif-bund} the preimage under $\cE$ of every point is a compact set (homeomorphic to $\ab$), and $\cE$ is also a closed map (using Remark \ref{rem:open-and-closed}). Hence $\cE$ is a proper map (see \cite[\S 10.2, Theorem 1]{Bourb1}). In particular, the preimage under $\cE$ of any compact set is compact (\cite[\S 10.2, Proposition 6]{Bourb1}). Therefore, if we show that $g'$ is continuous, then $g'(\ns)$ is compact and then $M$ must be compact, by \eqref{eq:MoverX}. (Note that $M$ is then also a continuous $\ab$-bundle over $\ns$, since by \eqref{eq:MoverX} it is such a bundle over $g'(\ns)$, and $g'$ is a homeomorphism $\ns\to g'(\ns)$.)
To prove the continuity of $g'$, the key idea is to express this function in terms of compositions of $\rho$ with certain tricube morphisms $T_k\to \ns$ (recall \cite[Definition 3.1.13]{Cand:Notes1}), and thereby to use these morphisms in a smoothening operation that turns the measurability of $\rho$ into the continuity of $g'$.\\
\indent Let $Q=\hom(T_k,\ns)$ and $Q_x=\hom_{1^k\to x}(T_k,\ns)$. For each $v\in \{0,1\}^k\setminus\{0^k\}$, let
\[
g_v:\ns\to\cL(Q,\ab),\;\; x\;\mapsto \;\big(g_v(x):t\mapsto \rho( t\co\Psi_v)\,\in L(Q_x,\ab)\big).
\]
We claim that $g_v$ is continuous. Indeed, by Lemma \ref{lem:key-tricube-mp} if $v\neq 0^k$ then $t\mapsto t\circ\Psi_v$ is measure-preserving from $Q_x$ to $\cu^k(\ns)$, so we can deduce the continuity of $g_v$ by applying Lemma \ref{lem:csmtech} (setting $V=Q$, $K=\cu^k(\ns)$, $W=\ns$, $f_2=\rho$, $h=g_v$).\\
\indent Now consider the following function:
\[
g\;:\; \ns \to \cL(Q,\ab),\quad x\;\mapsto\; \sum_{v\in\{0,1\}^k\setminus \{0^k\}}(-1)^{|v|}\,g_v.
\]
The continuity of each function $g_v$ implies that $g$ is continuous. Indeed, letting $P=\{0,1\}^k\setminus \{0^k\}$, by Lemma \ref{lem:prodCSMconv} the function $\ns \to \cL(Q,\ab^P)$, $x\,\mapsto\, \prod_{v\in P}g_v(x)$ is continuous, and then Lemma \ref{lem:CSMlincombin} gives the continuity of $g$. Now, by \cite[Lemma 3.3.31]{Cand:Notes1}, for each $x\in \ns$ we have that $g(x)$ is the function $t\mapsto \rho_x(t\circ\omega_k)-\rho_x(t\circ\Psi_{0^k})$ (defined on $Q_x$). In particular, given any $t\in Q_x$, we have that $\q_0:=t\co\omega_k$ and $\q_1:=t\co\Psi_{0^k}$ are both cubes in $\cu^k_x(\ns)$ ($\q_0$ is a cube by \cite[Lemma 3.1.16]{Cand:Notes1}), and $g(x)(t)=\rho_x(\q_0)-\rho_x(\q_1) = g'(x)(\q_0,\q_1)$. Given this equality, we now claim that continuity of $g'$ follows from continuity of $g$. Indeed, we know that $g'$ is continuous if for any of the functions $\varphi_{F_1,F_2}$ generating the topology on $\cL(Y,\ab)$ we have that $\ns \to \mathbb{C}$, $x\mapsto \varphi_{F_1,F_2}(g'(x))$ is continuous. Letting $Y_x=\cu^k_x(\ns)\times \cu^k_x(\ns)$, we have
\[
\varphi_{F_1,F_2}(g'(x)) = \int_{Y_x} F_1(g'(x)(\q_0,\q_1))\; F_2((\q_0,\q_1)) \;\,\mathrm{d}\mu((\q_0,\q_1) ),
\]
where $F_1:\ab\to \mathbb{C}$ and $F_2:Y_x\to \mathbb{C}$ are both continuous.
Now let $\theta : Q_x \to Y_x $, $t\mapsto (t\co \omega_k, t\co\Psi_{0^k})$. If we show that this map is surjective and measure-preserving, then since $g'(x)\co\theta(t)=g(x)(t)$, we would conclude that
\[
\varphi_{F_1,F_2}(g'(x)) = \int_{Q_x} F_1(g(x)(t))\; F_2(\theta(t))\; \,\mathrm{d}\mu(t)
\]
and continuity of $g'$ would then indeed follow from that of $g$ (via approximating the measurable function $F_2\co \theta$ by continuous functions as in previous proofs).\\
\indent The surjectivity of $\theta$ follows from combining Lemma \ref{lem:Gpair-in-Tn} and \ref{lem:Gpair-union-ext}. (To apply the latter result recall that we are implicitly embedding $T_k$ as a subcubespace in $\{0,1\}^{2k}$, using \cite[Lemma 3.1.17]{Cand:Notes1}.)\\
\indent To see that $\theta$ is measure-preserving, first note that by Lemma \ref{lem:GoodPairHoms} there is a CSM on the map $Q_x\to \cu^k_x(\ns),\; t\mapsto t\circ \omega_k$ consisting of the Haar probabilities on the fibres, giving a disintegration of the Haar probability on $Q_x$. On each of these fibres, of the form $\hom_{\q\co \omega_k^{-1}}(T_k,\ns)$, $\q\in \cu^k_x(\ns)$, it follows from Corollary \ref{cor:ext-in-Tn} that the map $\hom_{\q\co \omega_k^{-1}}(T_k,\ns) \to \cu^k_x(\ns)$, $t\mapsto t\co\Psi_{0^k}$ preserves the Haar measures. Using this, a straightforward calculation shows that for product sets $A_1\times A_2\subset \cu^k_x(\ns)\times \cu^k_x(\ns)$ we have $\mu_{Q_x}\co\theta^{-1}(A_1\times A_2)=\mu_{\cu^k_x(\ns)\times \cu^k_x(\ns)}(A_1\times A_2)$, and it follows that $\theta$ is measure-preserving as claimed.
\end{proof}
\noindent Let us now recall the definition of cubes on $M$ from \cite[Definition 3.3.25]{Cand:Notes1}. Note that, by definition of $M$, for every function $f:\{0,1\}^k\to M$ there is some function $a=a_f:\{0,1\}^k\to \ab$ such that $\rho_x(v)=f(v)+a(v)$, where $x=\tilde\pi(f(v))\in \ns$.
\begin{defn}\label{def:compextcubes}
We define $\cu^k(M)$ to be the set of functions $f:\{0,1\}^k\to M$ such that $\tilde\pi\co f\in \cu^k(\ns)$ and
\begin{equation}\label{eq:compextcubes}
\rho(\tilde\pi\co f) = \sigma_k(a).
\end{equation}
For $n\neq k$, a function $f:\{0,1\}^n\to M$ is declared to be in $\cu^n(M)$ if $\tilde\pi\co f\in \cu^n(\ns)$ and every $k$-dimensional face-restriction of $f$ is in $\cu^k(M)$.
\end{defn}
\indent We can now complete the main task of this section.
\begin{proposition}\label{prop:comp-extcompleted}
The space $M$ together with the cube sets $\cu^n(M)$ is a compact nilspace.
\end{proposition}
\begin{proof}
By \cite[Proposition 3.3.26]{Cand:Notes1} we have that $M$ is a nilspace, and by Proposition \ref{prop:cocyclext} we have that $M$ is a compact space. Arguing as in the proof of Lemma \ref{lem:compclosure}, we see that it suffices to show that $\cu^k(M)$ is a closed subset of $M^{\{0,1\}^k}$.
We use the probability space $Q=\hom(T_k,\ns)$ again, but this time we let $\pi_o:Q\to \cu^k(\ns)$ be the map $t\mapsto t\co \omega_k$, and apply Lemma \ref{lem:GoodPairHoms} (with $P=T_k$ embedded in $\{0,1\}^{2k}$, $P_1=\emptyset$, and $P_2=\omega_k(\{0,1\}^k)$ also embedded in $\{0,1\}^{2k}$) to obtain a CSM on $\pi_o$ consisting of the Haar measures on the fibres $\pi_o^{-1}(\q)=\hom_{\q\co\omega_k^{-1}}(T_k,\ns)$. Let $\tilde\pi$ be the projection $M\to \ns$, $\rho_x+z\mapsto x$, and let
\[
R_k(M)=\{f:\{0,1\}^k\to M : \tilde\pi\co f \in \cu^k(\ns)\}.
\]
Since $\tilde\pi$ is continuous and $\cu^k(\ns)$ is closed, we have that $R_k(M)$ is closed in $M^{\{0,1\}^k}$. We define the following map:
\begin{eqnarray*}
\phi\;:\;\;R_k(M)& \quad\to\quad & L\big(\pi_o^{-1}(\tilde\pi\co f),\ab\big) \; \subset \; \mathcal{L}(Q,\ab)\\
f &\quad \mapsto\quad & \phi_f(t)=\sum_{v\in\{0,1\}^k}(-1)^{|v|}\,f(v)(t\co\trem_v).
\end{eqnarray*}
Arguing as in the proof of \cite[Lemma 3.3.32]{Cand:Notes1}, we see that $\phi_f$ is actually a constant function for every $f\in R_k(M)$, namely the constant $\rho(\tilde\pi\co f)-\sigma_k(a)\in \ab$. By \cite[(3.16)]{Cand:Notes1} we have that $\cu^k(M)=\phi^{-1}(0)$. Hence it now suffices to show that $\phi$ is continuous, as then $\cu^k(M)$ is closed as required.\\
\indent To see the continuity of $\phi$, we first express it as the composition of simpler functions. For each $v\in \{0,1\}^k$, let
\begin{eqnarray*}
\phi_v\;:\;\;R_k(M)& \quad\to\quad & L\big(\pi_o^{-1}(\tilde\pi\co f),\ab\big) \; \subset \; \mathcal{L}(Q,\ab)\\
f &\quad \mapsto\quad & \phi_{v,f}(t)= f(v)(t\co\trem_v).
\end{eqnarray*}
We claim that $\phi_v$ is continuous. To see this, we suppose that $f_n\to f$ in $R_k(M)$, which implies that the component $f_n(v)$ converges to $f(v)$ in $M$, and we want to show that then $\phi_{v,f_n}\to \phi_{v,f}$ in $\cL(Q,\ab)$. For this it suffices to show that $\varphi_{F_1',F_2'}(\phi_{v,f_n})\to \varphi_{F_1',F_2'}(\phi_{v,f})$ for any continuous functions $F_1':\ab\to \mathbb{C}$, $F_2':Q\to \mathbb{C}$, where
\[
\varphi_{F_1',F_2'}(\phi_{v,f})=\int_{\pi_o^{-1}(\tilde\pi\co f)} F_1'(\phi_{v,f}(t))\; F_2'(t) \,\mathrm{d}\mu_{\tilde\pi \co f}(t)=\int_{\pi_o^{-1}(\tilde\pi\co f)} F_1'\big(f(v)(t\co\trem_v)\big)\; F_2'(t) \,\mathrm{d}\mu_{\tilde\pi\co f}(t).
\]
We shall now show that, for one of the functions $\varphi_{F_1,F_2}$ generating the topology on $M$, we have $\varphi_{F_1',F_2'}(\phi_{v,f})=\varphi_{F_1,F_2}(f(v))$. Recall that for any $g\in M$ we have
\begin{equation}\label{eq:testfn1}
\varphi_{F_1,F_2}(g)=\int_{\pi^{-1}(\tilde\pi(g))} F_1(g(\q))\; F_2(\q) \,\mathrm{d}\mu_{\tilde\pi(g)}(\q).
\end{equation}
Now for each $t\in Q$ let $Q_t$ denote the space of morphisms $t'\in Q$ such that $\trem_v\co t'=\trem_v\co t$, and let $\mu_{Q_t}$ denote the Haar measure on this space given by Corollary \ref{cor:tricubeprobspaces}. These measures form a CSM by Lemma \ref{lem:GoodPairHoms}. We then have
\begin{eqnarray*}
\varphi_{F_1',F_2'}(\phi_{v,f}) & = & \int_{\pi_o^{-1}(\tilde\pi\co f)}\;\int_{Q_t} F_1'\big(f(v)(t'\co\trem_v)\big)\; F_2'(t')\,\mathrm{d}\mu_{Q_t}(t') \;\,\mathrm{d}\mu_{\tilde\pi\co f}(t)\\
& = & \int_{\pi_o^{-1}(\tilde\pi\co f)}F_1'\big(f(v)(t\co\trem_v)\big)\; \Big(\int_{Q_t}F_2'(t')\,\mathrm{d}\mu_{Q_t}(t')\Big)\,\mathrm{d}\mu_{\tilde\pi\co f}(t)\\
& = & \int_{\pi_o^{-1}(\tilde\pi\co f)}F_1'\big(f(v)(t\co\trem_v)\big)\; F_2''(t\co \trem_v)\,\mathrm{d}\mu_{\tilde\pi\co f}(t),
\end{eqnarray*}
where $F_2''(t\co \trem_v)=\int_{t'\in Q_t}F_2'(t')\,\mathrm{d}\mu_{Q_t}(t')$ is a continuous function of $\q=t\co \trem_v\in \cu^k(\ns)$, by Definition \ref{def:CSM}. We can now use the fact that $t\mapsto t\co \trem_v$ is measure-preserving from $\pi_o^{-1}(\tilde\pi\co f)$ to $\pi^{-1}(\tilde\pi\co f(v))$, to conclude that for $f(v)\in M$ we have
\[
\varphi_{F_1',F_2'}(\phi_{v,f}) = \int_{\pi^{-1}(\tilde\pi\co f(v))}F_1'\big(f(v)(\q)\big)\; F_2''(\q)\,\mathrm{d}\mu_{\tilde\pi\co f}(\q).
\]
This is of the form $\varphi_{F_1',F_2''}(f(v))$ as in \eqref{eq:testfn1}, and the continuity of $\phi_v$ follows.\\
\indent The continuity of each function $\phi_v$ implies that the following function is continuous:
\begin{eqnarray*}
\phi_1\;:\;\;R_k(M)& \quad\to\quad & L\big(\pi_o^{-1}(\tilde\pi\co f),\ab^{\{0,1\}^k}\big) \; \subset \; \mathcal{L}(Q,\ab^{\{0,1\}^k})\\
f &\quad \mapsto\quad & \phi_{1,f}(t)= (f(v)(t\co\trem_v))_{v\in \{0,1\}^k}.
\end{eqnarray*}
Indeed, if $f_n\to f$ in $R_k(M)$ then for each $v$ by continuity of $\phi_v$ we have $\phi_{v,f_n}\to \phi_{v,f}$ in $\cL(Q,\ab)$. This then implies that $\phi_{1,f_n}\to\phi_{1,f}$ in $\cL(Q,\ab^{\{0,1\}^k})$, by Lemma \ref{lem:prodCSMconv}.\\
\indent The proof can now be completed using Lemma \ref{lem:CSMlincombin}. Indeed, for $P=\{0,1\}^k$ and appropriate choices of coefficients $\lambda_r$ in that lemma, the original function $\phi: R_k(M)\to \cL(Q,\ab)$ is the composition $\Sigma_P\co \phi_1$. The lemma tells us that $\Sigma_P$ is continuous, and so $\phi$ is continuous.
\end{proof}
\noindent Recall from \cite[Lemma 3.3.28]{Cand:Notes1} that, up to isomorphisms of nilspaces, every extension of $\ns$ by $\ab$ is of the form $M(\rho)$ above for some cocycle $\rho$. We now want an analogue of this result for compact nilspaces, which requires that the isomorphism be also a homeomorphism. To ensure this, recall first from \cite[Lemma 3.3.21]{Cand:Notes1} that given a degree-$k$ extension of a nilspace, every cross section $\cs$ for this extension generates a cocycle, denoted $\rho_{\cs}$. We now show that for compact nilspaces one can take $\rho_{\cs}$ to be Borel measurable.
\begin{lemma}\label{lem:meascross}
Let $\ns$ be a compact nilspace, and let $\nss$ be a compact nilspace that is a degree-$k$ extension of $\ns$ by a compact abelian group $\ab$. Then there is a Borel cross section $\cs$ for this extension and therefore a Borel cocycle $\rho_{\cs}:\cu^{k+1}(\ns)\to \ab$.
\end{lemma}
\begin{proof}
Let $\pi : \nss \to \ns$ be the projection of the extension. Let $P = \{(x,y) \in \ns \times \nss \,:\, \pi(y) = x \}$. A cross-section for $\pi$ is a subset of $P$ which happens to be the graph of a function $\ns \to \nss$. Let $\mathcal{P}(\nss)$ denote the set of Borel probability measures on $\nss$. By \cite[Corollary 18.7]{Ke}, a sufficient condition for a Borel cross section to exist is that for some Borel function $\mu : \ns \to \mathcal{P}(\nss)$ we have $\mu_x(P \cap (\{x\} \times \nss)) >0$. The measures $\mu_x$ from the CSM structure on $\pi$ satisfy this, and the function $x \to \mu_x$ is continuous, by Definition \ref{def:CSM}. There is therefore a Borel cross section $\cs:\ns\to \nss$. We can then define a Borel function $\rho:\cu^{k+1}(\nss)\to \ab$ by $\rho(\q')=\sigma_{k+1}(\cs\co\pi\co\q' -\q')$. Moreover, as is shown in the proof of \cite[Lemma 3.3.21]{Cand:Notes1}, the function $\rho$ is constant on each fibre of the map $\cu^{k+1}(\nss)\to \cu^{k+1}(\ns)$, $\q'\mapsto \pi\co\q'$, thus yielding the well-defined cocycle $\rho_{\cs}:\cu^{k+1}(\ns)\to \ab$. It can then be seen by general results that $\rho_{\cs}$ is also Borel measurable. For instance, one can use that by \cite[Definition 3.3.13]{Cand:Notes1} the space $\cu^{k+1}(\nss)$ is a continuous abelian bundle over $\cu^{k+1}(\ns)$ with Polish structure group $\cu^{k+1}(\cD_k(\ab))$, so that by \cite[Theorem (12.16) and Corollary (15.2)]{Ke} there is a Borel cross section $\q\mapsto\q'$ for this bundle.
\end{proof}
\noindent We shall also need the following useful generalization of a classical automatic-continuity result. The classical result in question is a theorem of Kleppner stating that a Borel measurable homomorphism between two locally compact groups must be continuous \cite[Theorem 1]{Klep}. We extend the case of compact abelian groups as follows.
\begin{theorem}\label{thm:Klepgen}
Let $\ns,\nss$ be compact nilspaces of finite step, and let $\phi:\ns\to \nss$ be a Borel measurable morphism. Then $\phi$ is continuous.
\end{theorem}
\begin{proof} Suppose that $\nss$ is a $(k-1)$-step nilspace, and consider the CSM on $\pi:\cu^k(\ns)\to \ns$, $\q\mapsto \q(0^k)$, given by Lemma \ref{lem:cube-set-CSM}. For each $v\in \{0,1\}^k\setminus \{0^k\}$ let $\pi_v:\cu^k(\ns)\to \ns$, $\q\mapsto \q(v)$. We have that $\pi_v$ is continuous and each of its restrictions to a space $\cu^k_x(\ns)$ is measure-preserving, by Lemma \ref{lem:GoodPairHoms} (applied with $P_1=\{0^k\}$, $P_2=\{v\}$ and $f:0^k\mapsto x$, identifying $\hom_{f|_{P_1\cap P_2}}(P_2,\ns)$ with $\ns$). Now for each $v$ let $f_v:\ns\to\cL(\cu^k(\ns),\nss)$ be the function sending $x$ to the restriction $\phi\co\pi_v|_{\cu_x^k(\ns)}$. Then $f_v$ is continuous, by Lemma \ref{lem:csmtech} (applied with $V=\cu^k(\ns)$, $W=K=\ns$, $f_1=\pi_v$ and $f_2=\phi$).\\
\indent Let $f:\ns\to\cL\big(\cu^k(\ns),\nss\big)$ be the function mapping $x\in \ns$ to the constant function with value $\phi(x)$ on $\cu_x^k(\ns)$. From the definition of $\cL\big(\cu^k(\ns),\nss\big)$ it follows that if $f$ is continuous then so is $\phi$.\\
\indent To see that $f$ is continuous, recall that $\cor^k(\nss)$ denotes the space of $k$-corners on $\nss$ and consider the following function:
\[
\xi : \ns \to \cL\big(\cu^k(\ns),\cor^k(\nss)\big),\;\;\; x\mapsto\;\; \Big( \xi_x:\q\in\cu^k_x(\ns)\; \mapsto\; \big(f_v(x)(\q)\big)_{v\neq 0^k}=\big(\phi\co\q(v)\big)_{v\neq 0^k}\Big).
\]
Combining the continuity of each function $f_v$ with Lemma \ref{lem:prodCSMconv}, we deduce that $\xi$ is continuous. Now let $\cK$ denote the function $\cL\big(\cu^k(\ns),\cor^k(\nss)\big)\to \cL\big(\cu^k(\ns),\nss\big)$ that sends a function $g$ to $\comp\co g$, where $\comp:\cor^k(\nss)\to\nss$ is the unique completion of $k$-corners on $\nss$. By Lemma \ref{lem:contcomp} we know that $\comp$ is continuous, and we claim that $\cK$ is therefore continuous. Indeed, suppose we have a sequence $g_n$ in $\cL\big(\cu^k(\ns),\cor^k(\nss)\big)$ converging to $g$, and let $F_1:\nss\to \mathbb{C}$ and $F_2:\cu^k(\ns)\to \mathbb{C}$ be continuous. Then
\[
\varphi_{F_1,F_2}(\cK(g_n))\;=\;\int_{\cu^k_{x_n}(\ns)} F_1(\comp \co g_n(\q))\;F_2(\q)\; \,\mathrm{d}\mu(\q)\;=\;\varphi_{F_1\co\comp, F_2}(g_n),
\]
where $\varphi_{F_1\co\comp, F_2}$ is one of the functions generating the topology on $\cL(\cu^k(\ns),\cor^k(\nss))$. Hence by assumption this converges to $\varphi_{F_1\co\comp, F_2}(g)=\varphi_{F_1,F_2}(\cK(g))$, so $\cK$ is continuous.\\
\indent Finally, note that $f=\cK\co\, \xi$, so $f$ is continuous, which completes the proof.
\end{proof}
We are now able to prove the analogue of \cite[Lemma 3.3.28]{Cand:Notes1} for compact nilspaces.
\begin{lemma}\label{lem:compextisotoM} Let $\ns$ be a compact nilspace of finite step. Let $\nss$ be a degree-$k$ extension of $\ns$ by a compact abelian group $\ab$. Let $\cs:\ns\to \nss$ be a Borel cross-section and let $\rho=\rho_{\cs}$ be the associated Borel cocycle. Then $\nss$ is isomorphic as a compact nilspace to the extension $M(\rho)$ from Proposition \ref{prop:cocyclext}.
\end{lemma}
\begin{proof} Let $\pi:\nss\to \ns$ be the projection for the extension. The isomorphism is given by the following map:
\begin{equation}
\theta:\nss\to M,\;\; x\mapsto \rho_{\pi(x)}+(x-\cs\co\pi(x)).
\end{equation}
It is checked in the proof of \cite[Lemma 3.3.28]{Cand:Notes1} that $\theta$ is an isomorphism of nilspaces. Since $\nss$ and $M$ are both compact Hausdorff spaces, if we show that $\theta$ is continuous then it is a homeomorphism \cite[Theorem 26.6]{Munkres}. Since $\theta$ is Borel measurable, it is continuous by Theorem \ref{thm:Klepgen}.
\end{proof}
\noindent Recall from \cite[Subsection 3.3.3]{Cand:Notes1} that $H_k(\ns,\ab)$ denoted the quotient of the abelian group of degree-$k$ cocycles by the subgroup of coboundaries. In the category of compact nilspaces we restrict these groups to the Borel measurable cocycles. Combining the results from this section, we now deduce the following analogue of \cite[Corollary 3.3.29]{Cand:Notes1} for compact nilspaces.
\begin{corollary}\label{cor:compclassrep} Let $\Phi$ denote the map which sends each class $C\in H_k(\ns,\ab)$ to the isomorphism class of $M(\rho)$, for any choice of $\rho\in C$. Then $\Phi$ is a surjection from $H_k(\ns,\ab)$ to the set of isomorphism classes of compact degree-$k$ extensions of $\ns$ by $\ab$.
\end{corollary}
\noindent The proof is similar to that of \cite[Corollary 3.3.29]{Cand:Notes1}.
\medskip
\section{Compact abelian bundles of finite rank, and averaging}\label{sec:CFRdef}
\medskip
In this section we define a notion of rank for compact abelian bundles and nilspaces, and we give a first topological description of nilspaces of finite rank (Lemma \ref{lem:finrankloctriv}). We then define an averaging operation for certain functions on these spaces, which enables us to prove a useful rigidity result concerning cocycles (Lemma \ref{lem:smallco}).
\begin{defn}
Let $\bnd$ be a $k$-fold compact abelian bundle, with structure groups $\ab_1,\dots,\ab_k$. We define the \emph{rank} of $\bnd$ by
\[
\mathrm{rk}(\bnd)=\sum_{i=1}^k \mathrm{rk}(\widehat{\ab_i})
\]
where $\widehat{\ab_i}$ is the Pontryagin dual of $\ab_i$ and $\mathrm{rk}(\widehat{\ab_i})$ is the minimal number of generators of $\widehat{\ab_i}$. If $\ns$ is a $k$-step compact nilspace then we define its rank $\mathrm{rk}(\ns)$ to be $\mathrm{rk}(\bnd)$ for the associated bundle $\bnd$ given by Proposition \ref{prop:topbundec}.
\end{defn}
\noindent When working with a compact $\ab$-bundle $\bnd$ over $S$, we shall often want to ensure that the bundle is locally trivial, meaning that for every $x\in \bnd$ there is an open set $U\subset S$ containing the projection $\pi(x)$, and a homeomorphism $\pi^{-1}(U)\to U\times \ab$ of the form $x\mapsto (\pi(x),\varphi(x))$ where $\varphi:\pi^{-1}(U)\to \ab$ is $\ab$-equivariant (that is $\varphi(x+z)=\varphi(x)+z$ for every $x\in \pi^{-1}(U)$, $z\in \ab$). A theorem of Gleason \cite[Theorem 3.3]{Gl} tells us that this local triviality holds provided that $\ab$ is a Lie group (see also \cite[Theorem 1.1]{A&D}). We record the special case of this result that we shall use.
\begin{proposition}[Consequence of Gleason's slice theorem]\label{prop:Gleason}
Let $\ab$ be a compact abelian Lie group, and let $\bnd$ be a continuous $\ab$-bundle. Then $\bnd$ is locally trivial.
\end{proposition}
\noindent Recalling that a compact abelian group is a Lie group if and only if it has finite rank, we deduce the following result.
\begin{lemma}\label{lem:finrankloctriv}
Let $\ns$ be a $k$-step compact nilspace of finite rank with structure groups $\ab_1,\dots,\ab_k$ and factors $\ns_0,\dots,\ns_k$. Then $\ns_i$ is a locally trivial $\ab_i$-bundle over $\ns_{i-1}$ for each $i\in [k]$.
\end{lemma}
Thus, topologically a $k$-step compact nilspace of finite rank is a finite dimensional manifold.
\begin{defn}\label{def:calgmetric}
Recall that every compact abelian Lie group $\ab$ is isomorphic to a direct product $F\times \mathbb{T}^n$ for some finite abelian group $F$ and $n\geq 0$. We can then define a natural metric $d_2$ on $\ab$ as follows. For two elements $x,y\in \mathbb{T}^n=\mathbb{R}^n/\mathbb{Z}^n$ we define $d_2(x,y)$ as the minimum of the Euclidean distances between preimages of $x$ and $y$ under the map $\mathbb{R}^n\to\mathbb{R}^n/\mathbb{Z}^n$. If $\ab$ is not connected, then for points $x,y$ in different connected components we declare that $d_2(x,y)=\infty$.
\end{defn}
\noindent We now turn to the definition of averaging for a function on a nilspace taking values in a compact abelian group $\ab$. Here we face the problem that in general the average of a $\ab$-valued function is not well-defined. This is clear already in the simple case of the circle group $\ab=\mathbb{T}$, where there is no suitable notion of multiplication by a real number. However, if we assume that the function takes all its values in a short interval, then we can lift this $\mathbb{T}$-valued function to a real-valued one, then take the usual average, and then project the obtained value back to $\mathbb{T}$. We capture this idea more generally as follows.
\begin{defn}\label{def:calgaver}
Let $\ab\cong F\times \mathbb{T}^n$ be a compact abelian Lie group. Let $f$ be a Borel measurable function from a probability space $X$ to $\ab$, and suppose there exists $z=(g,\theta)\in \ab$, ($\theta\in \mathbb{T}^n$) such that every value of $f$ lies in the ball $B_{1/4}(z)$ in the metric $d_2$. Then, letting $\pi$ denote the projection $\mathbb{R}^n\to \mathbb{T}^n$, for any fixed lift $\theta'\in \mathbb{R}^n$ there is a unique function $f':X\to \mathbb{R}^n$ such that $f(x)=(g,\pi\co f'(x))$ and $f'$ takes all its values in $B_{z'}(1/4)$ in the euclidean distance. We then define the average value $\mathbb{E} f$ to be $(g,\pi(\mathbb{E} f'))$.
\end{defn}
\noindent Note that $\mathbb{E}(f)$ does not depend on the choice of the lift of $z$. Let us say that a function $f:X\to \ab$ is $\delta$\emph{-concentrated} if there exists $z\in \ab$ such that $f(x)\in B_\delta(z)$ for all $x\in X$.
\begin{lemma}\label{lem:avadd}
Let $\ab$ be a compact abelian Lie group, and let $m\in \mathbb{N}$. Then for any $(1/4m)$-concentrated measurable functions $f_1,\dots,f_m$ from a probability space $X$ to $\ab$, the average of $f_1+\cdots+f_m$ is well defined and satisfies $\mathbb{E}(f_1+\cdots+f_m)=\mathbb{E} f_1+\cdots+\mathbb{E} f_m$.
\end{lemma}
\begin{proof}
The assumption implies that $f_1+\cdots+f_m$ is $1/4$-concentrated, so its average is well-defined. The additivity then follows in a straightforward way from the additivity of the usual average for real-valued functions.
\end{proof}
\noindent As a first application of this notion of averaging, we obtain the following rigidity result for cocycles. Stronger versions of this result will be very useful in later sections. Recall from \cite[Definition 3.3.18]{Cand:Notes1} that a function $\cu^k(\ns)\to\ab$ is called a coboundary if it is of the form $\q\mapsto \sigma_k(f\co \q)$ for some function $f:\ns\to \ab$.
\begin{lemma}\label{lem:smallco} Let $\ns$ be a compact $\ell$-step nilspace and let $\ab$ be a compact abelian group of finite rank. There exists $\epsilon>0$ such that, for every Borel measurable cocycle $\rho:\cu^k(\ns)\to \ab$, if $d_2(\rho(\q),0)\leq\epsilon$ for every $\q\in \cu^k(\ns)$ then there exists a Borel measurable function $g:\ns\to \ab$ such that $\rho$ is the coboundary $\q\mapsto \sigma_k(g\co\q)$.
\end{lemma}
\begin{proof}
For any $x\in\ns$, consider the space of cubes $\cu^k_x(\ns)$ with the Haar probability given by Lemma \ref{lem:cube-set-CSM}. By assumption $\rho$ is $\epsilon$-concentrated on $\cu^k_x(\ns)$, so we may define
\[
g:\ns\to \ab,\; x\mapsto \mathbb{E}_{\cu^k_x(\ns)}\,\rho(\q).
\]
Fix any $\q\in \cu^k(\ns)$ and let $\Omega=\hom_{\q\co\omega_k^{-1}}(T_k,\ns)$. For any $t\in \Omega$, by \cite[Lemma 3.3.31]{Cand:Notes1} we have $\rho(\q)=\sum_{v\in\{0,1\}^k} (-1)^{|v|} \rho(t\co\Psi_v)$. By Corollary \ref{cor:ext-in-Tn}, we have that for each $v$ the map $t\mapsto t\co\Psi_v$, $\Omega\to \cu^k_{\q(v)}(\ns)$ preserves the Haar measures (where the Haar measure on $\Omega$ is given by Corollary \ref{cor:tricubeprobspaces}). Hence, by averaging over $t$ in the last equation, we obtain $\rho(\q) = \sum_{v\in\{0,1\}^k} (-1)^{|v|} g(\q(v))=\sigma_k(g\co\q)$.
\end{proof}
\medskip
\section{Counting isomorphism classes of compact finite-rank nilspaces}
\medskip
In this section we use the parametrization of compact extensions by cocycles, given in Corollary \ref{cor:compclassrep}, to prove the following result.
\begin{theorem}\label{thm:countably} There are countably many isomorphism classes of compact nilspaces of finite rank.
\end{theorem}
\noindent To prove this we shall use the following lemma, that strengthens Lemma \ref{lem:smallco} by allowing its premise to fail on a null set.
\begin{lemma} \label{lem:smallco2} Let $\ns$ be a compact $\ell$-step nilspace and $\ab$ be a compact abelian group of finite rank. Then there exists $\epsilon>0$ such that every Borel measurable cocycle $\rho:\cu^k(\ns)\to \ab$ satisfying $d_2(\rho(\q),0)\leq\epsilon$ for almost every $\q\in \cu^k(\ns)$ is a coboundary.
\end{lemma}
This relies on the following fact.
\begin{lemma}\label{lem:almostzero} Let $\ns$ be a compact $\ell$-step nilspace and let $\ab$ be a compact abelian group of finite rank. Let $\rho:\cu^k(\ns)\to \ab$ be a Borel measurable cocycle such that $\rho=0$ for almost every element in $\cu^k(\ns)$. Then $\rho$ is a coboundary.
\end{lemma}
\begin{proof}
Let $S=\{x\in \ns : \rho_x=0\textrm{ almost surely on }\cu_x^k(\ns)\}$. It follows from the properties of the CSM on $\cu^k(\ns)\to \ns$, $\q\mapsto \q(0^k)$ (given by Lemma \ref{lem:cube-set-CSM}) that $S$ has probability $1$. For every $x\in S$ the set $\rho_x+\ab\,\subset \cL_k(\ns,\ab)$ consists of functions that are constant almost everywhere. Since by Proposition \ref{prop:posmeasopen} every open set in $\ns$ has positive measure, we have that $S$ is dense in $\ns$. We claim that for every $x\in \ns$ the function $\rho_x$ equals a constant almost everywhere. Indeed, by density of $S$ there exists a sequence of points $x_n\in S$ converging to $x$. For each $n$ choose an element $\rho_{x_n}+a_n \in M(\rho)$ equal to a constant almost surely. By Proposition \ref{prop:cocyclext} the extension $M$ is compact and so there is a subsequence of such functions $\rho_{x_m}+a_m$ converging in $M$. By Lemma \ref{lem:constconv}, the limit is constant almost everywhere, and (as a point of $M$) is $\rho_x+a$ for some $a\in \ab$, which proves our claim. From this, we deduce that the function $g:\ns\to \ab$, $x\mapsto \mathbb{E}_{\q\in \cu^k_x(\ns)} \rho(\q)$ is well-defined and that for each $x\in \ns$ we have $\rho(\q)=g(x)$ for almost every $\q\in \cu^k_x(\ns)$. We now claim that for every $x\in \ns$, we have $\rho(\q)=\sigma_k(g\co \q)$ for almost every $\q\in \cu^k(\ns)$. This follows from an argument similar to the second paragraph in the proof of Lemma \ref{lem:smallco}. Thus, denoting by $f$ the coboundary $\q\mapsto \sigma_k(g\co \q)$, the cocycle $\rho'=\rho-f$ has the property that for every $x\in \ns$ we have $\rho'_x=0$ almost surely on $\cu_x^k(\ns)$. This implies that $\rho'=0$ everywhere. Indeed, for an arbitrary fixed $\q\in \cu^k(\ns)$, and any $t\in \hom_{\q\co\omega_k^{-1}} (T_k,\ns)$, by \cite[Lemma 3.3.31]{Cand:Notes1} we have $\rho'(\q)=\beta(t,\rho')$, and then, averaging both sides of this equation over $t$, the right side vanishes since $\rho_x'$ is 0 almost everywhere (using again that each map $t\mapsto t\circ \Psi_v$ is probability-preserving, by Corollary \ref{cor:ext-in-Tn}).
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:smallco2}] We argue as in the proof of Lemma \ref{lem:smallco}. Let $S$ be the set of elements $x\in \ns$ for which $\rho_x$ is almost surely within $\epsilon$ of $0$ in $d_2$. By the assumption, we have that $S$ has probability $1$. We define $g:\ns\to \ab$ for $x\in S$ by $g(x)=\mathbb{E}_{\cu^k_x(\ns)}\,\rho(\q)$, and for $x\in \ns\setminus S$ by setting $g(x)=0$. Letting $\rho_2(\q)=\sigma_k(g\co \q)$, the same averaging argument shows that $\rho_2=\rho$ almost surely. By Lemma \ref{lem:almostzero} the difference $\rho-\rho_2$ is a coboundary, whence $\rho$ is a coboundary.
\end{proof}
We can now prove the main result of this section.
\begin{proof}[Proof of Theorem \ref{thm:countably}]
We claim that for every $k\in \mathbb{N}$, for every compact $(k-1)$-step nilspace $\ns$ of finite rank and every compact abelian Lie group $\ab$, there are at most countably many non-isomorphic degree-$k$ extensions of $\ns$ by $\ab$. This implies the theorem by induction on $k$.\\
\indent To prove the claim, we first associate with each $\ab$-extension of $\ns$ a measurable cocycle $\rho_{\cs}$ generated by a piecewise-continuous cross section $\cs$ for the extension. We construct $\cs$ as follows.
Since $\ns$ is a finite-dimensional compact manifold, there exist disjoint open sets $U_1,U_2,\dots,U_r$, each homeomorphic to the open unit Euclidean ball in $\mathbb{R}^n$ for some $n\geq 0$, such that $U=\bigcup_{i\in [r]} U_i$ has measure $1$ in $\ns$. Let $M$ denote the given degree-$k$ extension of $\ns$ by $\ab$, with projection $\pi:M\to \ns$. Every $\ab$-bundle over a contractible space is trivial \cite[Ch. 4, Corollary 10.3]{Husem}. Therefore there exists a continuous cross section $\cs_i$ locally on each $U_i$, and combining these we obtain a piecewise continuous (hence Borel-measurable) cross section $\cs:\ns\to M$, with $\cs|_{U_i}=\cs_i$. Let $T\subseteq \cu^{k+1}(\ns)$ denote the set of cubes whose vertices are all in $U$. We can partition the cubes $\q\in T$ according to which set $U_i$ contains each value $\q(v)$, that is $T=\bigsqcup_{i\in [r]^{\{0,1\}^{k+1}}} T_i$, where $T_i\,=\,T\;\cap \prod_{v\in \{0,1\}^{k+1}} U_{i_v}$. The cocycle $\rho_{\cs}:\cu^{k+1}(\ns)\to \ab$ is defined by $\rho_{\cs}(\q)=\sigma_{k+1}(\cs\co \q-\q')$ for any $\q'\in\cu^{k+1}(M)$ with $\pi\co\q'=\q$ (recall \cite[Lemma 3.3.21]{Cand:Notes1}). Note that $\rho_{\cs}$ is continuous on each part $T_i$.\\
\indent Now let $\epsilon>0$ be as given by Lemma \ref{lem:smallco2}. Then for any other degree-$k$ extension $M'$, we can generate a cocycle $\rho_{\cs'}$ as above, and if $d_2\big(\rho_{\cs}(\q),\rho_{\cs'}(\q)\big)<\epsilon$ for almost every $\q$ then by the lemma $\rho_{\cs'}$ must equal $\rho_{\cs}$ plus a coboundary, which implies, by Corollary \ref{cor:compclassrep}, that $M'$ is isomorphic to $M$.\\
\indent Let $d_\infty$ denote the uniform metric on the set of continuous functions $\cu^{k+1}(\ns)\to\ab$, defined by $d_\infty(f_1,f_2)=\sup_{\q} d_2\big(f_1(\q),f_2(\q)\big)$. For each $i$, the space of continuous functions $T_i\to \ab$ with $d_\infty$ is separable (essentially by separability of the space $C(Y,\ab)$ for any compact subset $Y$ of a Euclidean space \cite[Theorem (4.19)]{Ke}). Therefore there exists a countable $\frac{\epsilon}{2}$-net $(f_{i,j})_{j\in \mathbb{N}}$ of such continuous functions on each $T_i$. Then, by the previous paragraph, for each choice of functions $\{f_{i,j_i}: i\in [r]^{\{0,1\}^{k+1}}\}$ from these nets, there can only be at most one isomorphism class of extensions having an associated cocycle $\rho_{\cs}$ satisfying $d_\infty(\rho_{\cs}|_{T_i},f_{i,j_i})<\epsilon/2$ for each $i$.
\end{proof}
\medskip
\section{Compact nilspaces as inverse limits of finite-rank nilspaces}\label{sec:invlim}
\medskip
In this section we treat one of the central results from \cite{CamSzeg}, namely that every compact nilspace can be expressed as an inverse limit of compact nilspaces of finite rank. Before we give the formal statement, let us detail the inverse limit construction in this category.
\begin{defn}
An \emph{inverse system} (or projective system) of compact nilspaces over $\mathbb{N}$ is a family of continuous nilspace morphisms $\{\varphi_{ij}: \ns_j\to \ns_i~|~ i,j\in \mathbb{N}, i\leq j\}$, where $\ns_j$, $j\in \mathbb{N}$ are compact nilspaces, such that $\varphi_{jj}$ is the identity morphism for all $j\in \mathbb{N}$, and for all $i,j,k\in \mathbb{N}$ with $i\leq j\leq k$ we have $\varphi_{ij}\co\varphi_{jk}=\varphi_{ik}$. The inverse system is said to be \emph{strict} if every morphism $\varphi_{ij}$ is fibre-surjective.
\end{defn}
\begin{lemma}
Let $S=(\varphi_{ij}: \ns_j\to \ns_i)_{i\leq j}$ be a strict inverse system of compact nilspaces. Let
\[\ns=\Big\{(x_i)\in \prod_{i\in \mathbb{N}} \ns_i ~|~ \varphi_{ij}(x_j)=x_i\;\; \forall\,i,j\in \mathbb{N},\;i\leq j\Big\},
\]
and for each $n\geq 0$ let $\cu^n(\ns)=\big\{(\q_i)\in \prod_{i\in \mathbb{N}} \cu^n(\ns_i)~|~ \varphi_{ij}\co \q_j= \q_i\;\;\forall\,i,j\in \mathbb{N},\;i\leq j\big\}$. The space $\ns$ together with the cube sets $\cu^n(\ns)$ is a compact nilspace, called the \emph{inverse limit} of $S$, and denoted $\varprojlim_i \ns_i$. We have that $\ns$ is $k$-step if and only if every $\ns_i$ is $k$-step.
\end{lemma}
\noindent We call the maps $\varphi_{ij}$ the \emph{transition morphisms}. The \emph{projections} on $\ns$, denoted by $\varphi_j$, are the coordinate projections on $\prod_{i\in \mathbb{N}} \ns_i$ restricted to $\ns$. If the inverse system is strict then the projections are also fibre-surjective morphisms. The $\ell$-th structure group of $\varprojlim_i \ns_i$ is the inverse limit of the system $(\alpha_{\ell,ij}:\ab_{\ell,j}\to \ab_{\ell,i})_{i\leq j}$, where $\ab_{\ell,j}$ is the $\ell$-th structure group of $\ns_j$ and $\alpha_{\ell,ij}$ is the $\ell$-th structure morphism of $\varphi_{ij}$.
\begin{proof}
A straightforward argument with convergent sequences shows that $\ns$ is a closed subspace of the compact space $\prod_{i\in \mathbb{N}} \ns_i$, and similarly for each set $\cu^n(\ns)$. The ergodicity and composition axioms are clear. Let us prove that $\ns$ satisfies the completion axiom. Let $\q'$ be an $n$-corner on $\ns$. Then for each $i\in \mathbb{N}$ the projection $\q_i'=\varphi_i\co \q'$ is an $n$-corner on $\ns_i$. Now $\q_1'$ has a completion $\q_1$ and by fibre-surjectivity there exists $\tilde \q_1\in \cu^n(\ns_2)$ such that $\varphi_{12}\co \tilde\q_1=\q_1$ (recall \cite[Lemma 3.3.9]{Cand:Notes1}).
The restriction of $\tilde\q_1$ to $P=\{0,1\}^n\setminus\{1^n\}$ is in the same fibre of the map $\q'\mapsto \varphi_{12}\co \q'$ as $\q_2'$, so by \cite[Lemma 3.3.12 (ii)]{Cand:Notes1} these two corners differ by an $n$-corner on $\cD_k(\ker(\alpha_k))$, where $\alpha_k$ is the $k$-th structure morphism of $\varphi_{12}$. Now by modifying $\tilde\q_1$, we can obtain a cube $\q_2$ that is still in this fibre (so that $\varphi_{12}\co \q_2=\q_1$) and such that $\q_2|_P=\q_2'|_P$. Indeed, we can obtain $\q_2$ by adding first an appropriate element of $\ker(\alpha_k)$ to the values of $\tilde\q_1$ at the vertices of the 1-face $\{0^n,(1,0,\dots,0)\}$ so that it agrees with $\q_2'$ at $0^n$, then repeat this for each subsequent 1-face along a Hamiltonian path in the graph of 1-faces on $\{0,1\}^n$ (similar arguments were used several times in \cite{Cand:Notes1}, for instance in the proof of \cite[Lemma 3.2.25]{Cand:Notes1}). We have thus obtained a cube $\q_2$ completing $\q_2'$ and such that $\varphi_{12}\co \q_2=\q_1$. Now we repeat this operation to obtain $\q_3$ completing $\q_3'$ such that $\varphi_{23}\co \q_3=\q_2$, and so on. This yields a sequence of cubes $\q_i\in \cu^n(\ns)$ such that for each $i$ we have $\varphi_j\co \q_i$ is a completion of $\q_j'$ for all $j\leq i$. The limit of a convergent subsequence of $(\q_i)$ is a completion of $\q'$. To see the last claim in the lemma, recall from \cite[Definition 3.3.7]{Cand:Notes1} that a fibre-surjective morphism preserves the property of being $k$-step.
\end{proof}
We can now state the main result of this section.
\begin{theorem}\label{thm:invlim} Every $k$-step compact nilspace is an inverse limit of compact nilspaces of finite rank.
\end{theorem}
\noindent The basic case of this theorem, concerning 1-step nilspaces, is given by the standard result that every compact abelian group is an inverse limit of compact abelian Lie groups (see \cite[Corollary 2.43]{H&M}). This standard result also yields the following example of an inverse limit of nilspaces, which will be used in the proof of Theorem \ref{thm:invlim}.
\begin{proposition}\label{prop:k-level-invlim}
Let $\ns$ be a $k$-step compact nilspace. Then $\ns$ is an inverse limit of $k$-step nilspaces $\ns_i$, each with $k$-th structure group a compact abelian Lie group, and with $\cF_{k-1}(\ns_i)$ isomorphic to $\cF_{k-1}(\ns)$. \end{proposition}
\begin{proof}
By the standard result for compact abelian groups recalled above, for the $k$-th structure group $\ab_k$ of $\ns$ we have for each $i\in \mathbb{N}$ a compact abelian Lie group $\ab_{k,i}$ and a surjective continuous homomorphism $\alpha_i:\ab_k\to \ab_{k,i}$ such that $\ab_k=\varprojlim \ab_{k,i}$. (For $i\leq j$ the transition morphism $\alpha_{i,j}$ is $x\mapsto x+\ker \alpha_i$ for $x\in \ab_{k,j}\cong \ab_k/\ker \alpha_j$.) Define for each $i$ the $k$-step compact nilspace $\ns_i$ as the image of $\ns$ under the map $\varphi_i$ that sends $x\in \ns$ to the orbit $x+\ker \alpha_i$. A straightforward calculation shows that $\ns_i$, with the quotient cube-structure of $\ns$ by $\ker \alpha_i$ and the quotient topology, is a $k$-step compact nilspace with $k$-th structure group $\ab_{k,i}$. The projections $\ns\to \cF_{k-1}(\ns)$ and $\ns_i\to \cF_{k-1}(\ns_i)$ are both given by the orbit map for the $\ab_k$-action, and it follows that $\cF_{k-1}(\ns)\cong \cF_{k-1}(\ns_i)$ as compact nilspaces. The transition morphisms $\varphi_{ij}:\ns_j\to \ns_i$, defined by $x+\ker \alpha_j\mapsto x+\ker \alpha_i$, are all clearly fibre-surjective.
\end{proof}
\noindent Recall from \cite[Lemma 3.3.21]{Cand:Notes1} that given a degree-$k$ extension $\nss$ of $\ns$ by $\ab$, and given a cross section $\cs:\ns\to \nss$, letting $f:\nss\to \ab$, $y\mapsto \cs\co \pi(y)-y$, the cocycle generated by $\cs$ is the degree-$k$ cocycle $\rho_{\cs}:\cu^{k+1}(\ns)\to \ab$ defined by $\rho_{\cs}(\q)=\sigma_{k+1}(f\co \q')$, for any $\q'\in \cu^{k+1}(\nss)$ such that $\pi\co \q'=\q$.
For the proof of Theorem \ref{thm:invlim}, it will be useful to be able to tell in a simple way whether a given cocycle on a nilspace $\ns$ induces a well-defined cocycle on another nilspace $\ns'$ via some given fibre-surjective morphism $\ns\to \ns'$. The following definition provides such a criterion.
\begin{defn}\label{def:crossfactdef}
Let $\ns$ be a $k$-step compact nilspace, and for some compact nilspace $\ns'$ let $\psi:\cF_{k-1}(\ns)\to \ns'$ be a continuous fibre-surjective morphism. We say that a measurable cross section $\cs:\cF_{k-1}(\ns)\to \ns$ is \emph{consistent with the factor} $\ns'$ if the cocycle $\rho_{\cs}:\cu^{k+1}(\cF_{k-1}(\ns))\to \ab_k$ satisfies $\rho_{\cs}(\q_1)=\rho_{\cs}(\q_2)$ for every $\q_1,\q_2$ such that $\psi\co \q_1=\psi\co \q_2$.
\end{defn}
\noindent Thus if $\cs$ is consistent with $\ns'$ then $\rho_{\cs}$ induces a cocycle $\rho':\cu^{k+1}(\ns')\to \ab_k$, well-defined by $\rho'(\q')=\rho_{\cs}(\q)$ for any $\q$ such that $\psi\co\q=\q'$. Note that we shall often refer to the image of a fibre-surjective morphism on $\ns$ as a \emph{fibre-surjective factor} of $\ns$.
\begin{lemma}\label{lem:crossfactor}
Let $\psi:\cF_{k-1}(\ns)\to \ns'$ be a fibre-surjective morphism and let $\cs:\cF_{k-1}(\ns)\to \ns$ be a cross section consistent with $\ns'$. Let $\sim$ be the equivalence relation on $\ns$ defined by
\[
x\sim y\;\;\Leftrightarrow\;\; \big(\;\psi\co \pi_{k-1}(x)=\psi\co \pi_{k-1}(y)\textrm{ and }x-\cs\co\pi_{k-1}(x)=y-\cs\co\pi_{k-1}(y)\;\big).
\]
Then the quotient nilspace $\ns/_{\sim}$ is a fibre-surjective factor of $\ns$ which is an extension of $\ns'$ by $\ab_k$.
\end{lemma}
\begin{proof} Let $\rho':\cu^{k+1}(\ns')\to \ab_k$ be the cocycle induced by $\rho_{\cs}$. Let $M(\rho')$ be the compact nilspace extending $\ns'$ by $\ab_k$ given by Proposition \ref{prop:cocyclext}. Let $f:\ns\to M(\rho')$ be defined by $f(x)=\rho'_{\psi(\pi_{k-1}(x))}+x-\cs\co\pi_{k-1}(x)$. Using Definition \ref{def:compextcubes} we check that $f$ is a morphism (a similar calculation appears in the proof of \cite[Lemma 3.3.28]{Cand:Notes1}). It is also checked easily that $f$ is fibre-surjective. Moreover, clearly $f(x_1)=f(x_2)$ if and only if $x_1\sim x_2$. It follows that $f$ factors through $\sim$ giving a bijective measurable morphism $f':\ns/_{\sim}\to\nss$. Then $f'$ is a continuous isomorphism of compact nilspaces, by Theorem \ref{thm:Klepgen}.
\end{proof}
\noindent The following lemma will be used in the proof of Theorem \ref{thm:invlim} and also in Section \ref{sec:CFRnilsnilm}. Recall the notation $d_2$ for the metric from Definition \ref{def:calgmetric}.
\begin{lemma}\label{lem:vertiparathin}
Let $\ns$ be a $k$-step compact nilspace such that the $k$-th structure group $\ab_k$ has finite rank, and let $d$ be a compatible metric on $\ns$. Then for any $\epsilon>0$ there exists $\delta>0$ such that the following holds. Let $x_0,x_1,y_0,y_1\in \ns$ be such that $\pi_{k-1}(x_i)=\pi_{k-1}(y_i)$ for $i=0,1$ and $d(x_0,x_1),d(y_0,y_1)$ are both at most $\delta$. Then in $\ab_k$ we have $d_2(y_0-x_0,y_1-x_1)\leq \epsilon$.
\end{lemma}
\begin{proof}
Consider the local translation $\phi=\phi_{x_0,x_1}$ (recall \cite[Definition 3.2.26]{Cand:Notes1}). By definition of $\phi$ we have $y_0-x_0=\phi(y_0)-x_1$ in $\ab_k$ (using implicitly the isomorphism from \cite[Lemma 3.2.24]{Cand:Notes1}), so it suffices to show that $d_2(\phi(y_0)-x_1,y_1-x_1)\leq \epsilon$ for $\delta$ sufficiently small.
Since $d(x_0,x_1)$ is unchanged by adding $y_0-x_0=\phi(y_0)-x_1$ to $x_0$ and $x_1$, we have $d(y_0,\phi(y_0))\leq \delta$. This together with $d(y_0,y_1)\leq \delta$ implies that $d(\phi(y_0),y_1)\leq 2\delta$. Thus, the lemma will follow if we prove the following statement: for every $\epsilon>0$ there exists $\delta>0$ such that whenever $x_1,x_2,x_3\in \ns$ are in the same fibre of $\pi_{k-1}$ and $d(x_2,x_3)\leq \delta$, we have $d_2(x_2-x_1,x_3-x_1)\leq \epsilon$. This can be shown by a straightforward argument using the local triviality of the $\ab_k$-bundle $\ns$ and the compactness of $\ns_{k-1}$.
\end{proof}
\subsection{Proof of the inverse limit theorem}
To prove Theorem \ref{thm:invlim}, we argue by induction on $k$, starting with the trivial case $k=0$. For $k>0$ we suppose that $\ns$ is a $k$-step compact nilspace such that the nilspace $\nss=\cF_{k-1}(\ns)$ is the inverse limit of a strict inverse system $(\tau_{ij}:\nss_j\to\nss_i)_{i,j\in \mathbb{N}, i\leq j}$. We denote by $\pi$ the projection $\ns\to \nss$, and for each $i$ we denote by $\tau_i$ the projection $\nss\to\nss_i$ and let $\mathcal{Q}_i=\{\tau_i^{-1}(Q):Q\subset \nss_i,\; Q\textrm{ open}\}$. Since by assumption the topology on $\nss$ is the initial topology generated by the maps $\tau_i$, we have that the finite intersections of sets in $\bigcup_{i\in \mathbb{N}}\mathcal{Q}_i$ form a base for this topology.
\subsubsection{The inductive step: construction of the sequence $S$}
We now apply Proposition \ref{prop:k-level-invlim} to $\ns$. Let $B_0=\ab_k$ be the $k$-th structure group of $\ns$, and let $B_i=\ker \alpha_i\leq \ab_k$ for each $i\in \mathbb{N}$ (for the homomorphisms $\alpha_i$ from the proof of the proposition). Then $B_0 \geq B_1 \geq \cdots$, with $\bigcap_{i\in \mathbb{N}} B_i=\{0\}$, and $\ns = \varprojlim_i \ns/_{B_i}$. Let $q_{i-1,i}$ denote the transition morphism $\ns/_{B_i}\to \ns/_{B_{i-1}}$ and $q_i$ the projection $\ns\to \ns/_{B_i}$. Note that $\ns/_{B_i}$ has $k$-th structure group $\ab_k/B_i$, of finite rank.
The main goal of the inductive step is to construct, starting with the 1-point nilspace $\ns_0$, a sequence $S=\{(\ns_i,\psi_i,\psi_i',h_i):i\in \mathbb{N}\}$, where for each $i$ we have a compact nilspace $\ns_i$ of finite rank, and fibre-surjective morphisms $\psi_i:\ns/_{B_i}\to \ns_i$ and $\psi_i':\ns_i\to \ns_{i-1}$, with the following properties:
\begin{enumerate}
\item For every $i$, the restriction of $\psi_i$ to each class of $\sim_{k-1}$ in $\ns/_{B_i}$ is injective. In other words, the $k$-th structure group of $\ns_i$ is the same as that of $\ns/_{B_i}$, namely $\ab_k/B_i$.
\item For every $i$, we have $\psi'_i\circ\psi_i=\psi_{i-1}\co q_{i-1,i}$.
\item Letting $\pi_i$ denote the projection $\ns_i\mapsto \cF_{k-1}(\ns_i)$, and $\kappa_i$ denote the projection $\ns/_{B_i}\to \nss$, we have $\cF_{k-1}(\ns_i)\cong \nss_{h_i}$ and $\pi_i\co \psi_i= \tau_{h_i}\co \kappa_i$. \end{enumerate}
Properties (i) and (iii) imply that any two distinct points in $\ns$ are separated by a map $\psi_i\co q_i$ for some $i$, so the initial topology on $\ns$ generated by these maps is Hausdorff. Since the original topology on $\ns$ is compact, it follows that this initial topology equals the original topology (see \cite[\S 9.4, Corollary 3]{Bourb1}). The maps $\varphi_{ij}=\psi'_{i+1}\co\cdots\co \psi'_j$ and $\varphi_i:=\psi_i\co q_i$ satisfy the relations $\varphi_{ij}\co \varphi_j= \varphi_i$ thanks to property (ii). Hence $\ns$ is the inverse limit of the system $(\varphi_{ij}:\ns_j\to\ns_i)_{i\leq j}$. \medskip \\
\includegraphics{InvLim1}\\
\noindent We shall now construct the sequence $S$ inductively starting from $\ns_0$. Suppose that $(\ns_i,\psi_i,\psi'_i,h_i)$ has been constructed for $i\leq m$. Then, to begin with, we construct a Borel cocycle $\cu^{k+1}(\nss)\to \ab_k/B_{m+1}$ generated by a cross section $\cs$ that is almost consistent with some appropriate factor of $\nss$.
\noindent \textbf{Constructing a measurable cross section $\cs$:} since $\ns_m$ is a compact locally-trivial $\ab_k/B_m$-bundle over $\nss_{h_m}$, there is a finite cover of $\nss_{h_m}$ by closed subsets $W_1,\ldots,W_r$ such that every preimage $\pi_m^{-1}(W_i)\subset \ns_m$ is a trivial $\ab_k/B_m$-bundle over $W_i$. For each $i\in [r]$, let $\theta_i:W_i\to \ns_m$ be a continuous cross section (thus $\pi_m\co\theta_i$ is the identity map on $W_i$).\\
\indent For each $a\in [r]$ let $W'_a$ denote the preimage of the set $\theta_i(W_a)$ under $\psi_m\co q_{m,m+1}:\ns/_{B_{m+1}}\to \ns_m$. Note that $W'_a$ is a $\ab_k/B_{m+1}$-bundle over $\tau_{h_m}^{-1}(W_a)\subset \nss$, so it is locally trivial by Proposition \ref{prop:Gleason}. Let $d$ be a metric on $\ns/_{B_{m+1}}$ generating its topology.\\
\begin{center}\includegraphics{InvLim2}\end{center}
\noindent Now fix an arbitrary $\epsilon>0$. We claim that for every $p\in \tau_{h_m}^{-1}(W_a)$ there exists an open set $U_p\subset \tau_{h_m}^{-1}(W_a)$ containing $p$ with the following properties:\\ \vspace{-0.7cm}
\begin{enumerate}
\setcounter{enumi}{3}
\item There exists a continuous cross section $\cs_p:U_p\to \ns/_{B_{m+1}}$.\\ \vspace{-0.7cm}
\item The diameter of $\cs_p(U_p)$ in the metric $d$ is at most $\epsilon$.\\ \vspace{-0.7cm}
\item For some $t(p)\in\mathbb{N}$, the set $U_p$ is in the collection $\mathcal{Q}_{t(p)}$ defined above.\\ \vspace{-0.7cm}
\end{enumerate}
\noindent Property (iv) is ensured by the local triviality of the bundle, and it can be satisfied together with (v) by letting $U_p$ be a ball of sufficiently small radius, say. Property (vi) can also be satisfied since $\bigcup_i \cQ_i$ generates the topology on $\nss$ and for every $i\leq j$ and open $Q\subset \nss_i$ we have $\tau_i^{-1}Q=\tau_j^{-1}(\tau_{ij}^{-1}Q)\in \cQ_j$. Since $\nss$ is compact there exists a finite cover by such sets $U_p$, thus for some points $p_1,p_2,\dots,p_n$ we have $\bigcup_{i=1}^n U_{p_i}\,\supset\,\tau_{h_m}^{-1}(W_a)$. Let $t_a=\max\{t(p_i):i \in [n]\}$, so that for every $i\in [n]$ we have $U_{p_i}\subset \mathcal{Q}_{t_a}$. Then, by dividing $\tau_{h_m}^{-1}(W_a)$ into the atoms of the Boolean algebra generated by $U_{p_1},\dots,U_{p_n}$, and then using one of the cross sections $\cs_p$ for each atom, we can construct a Borel measurable cross section $\cs_a:\tau_{h_m}^{-1}(W_a)\to \ns/_{B_{m+1}}$ with the following properties:\\ \vspace{-0.7cm}
\begin{enumerate}
\setcounter{enumi}{6}
\item The cross section $\cs_a$ is continuous on $\tau_{t_a}^{-1}(v)$ for every $v\in \nss_{t_a}$.\\ \vspace{-0.7cm}
\item The diameter of $\cs_a(\tau_{t_a}^{-1}(v))$ in $d$ is at most $\epsilon$ for every $v\in \nss_{t_a}$.\\ \vspace{-0.7cm}
\end{enumerate}
Note that we obtain (viii) from (v) because each $\tau_{t_a}^{-1}(v)$ is entirely contained in an atom of the Boolean algebra, which contains $\tau_{t_a}^{-1}(Q)$ for some open $Q\supset v$. Let $t=\max\{t_a: a\in [r]\}\cup \{h_m+1\}$. Using these partial cross sections, we now claim that we can construct a global Borel cross section $\cs:\nss\to \ns/_{B_{m+1}}$ with the following properties:\\ \vspace{-0.7cm}
\begin{enumerate}
\setcounter{enumi}{8}
\item The cross section $\cs$ is continuous on every preimage $\tau_t^{-1}(v)$ where $v\in \nss_t$.\\ \vspace{-0.7cm}
\item The diameter of $\cs(\tau_t^{-1}(v))$ in $d$ is at most $\epsilon$ for every $v\in \nss_t$.\\ \vspace{-0.7cm}
\item If $v_1,v_2\in \nss$ satisfy $\tau_{h_m}(v_1)=\tau_{h_m}(v_2)$, then $\psi_m\co q_{m,m+1} (\cs(v_1))=\psi_m\co q_{m,m+1}(\cs(v_2))$.
\end{enumerate}
Indeed, we can construct $\cs$ by again dividing $\nss$ into the atoms of the Boolean algebra generated by the sets $\tau_{h_m}^{-1}(W_a)$ and then using one of the cross sections $\cs_a$ on each atom. This immediately gives (ix) and (x), and then (xi) follows from (iii) above. \\
\indent Let $\rho_{\cs}:\cu^{k+1}(\nss)\to \ab_k/B_{m+1}$ be the cocycle generated by $\cs$.\\
\indent Now for any fixed $\epsilon_2>0$, we note that if $\epsilon>0$ is small enough, then property (x) implies the following fact, which tells us that $\cs$ is almost consistent with the factor $\nss_t$:
\begin{equation}\label{eq:rhoclose}
\forall\,\q_1,\q_2\in \cu^{k+1}(\nss)\textrm{ with } \tau_t\co\q_1= \tau_t\co\q_2,\textrm{ we have } d_2\big(\rho_{\cs}(\q_1)-\rho_{\cs}(\q_2)\big)\leq \epsilon_2.
\end{equation}
Indeed $\tau_t\co\q_1= \tau_t\co\q_2$ tells us that for each $v$ the values $\q_1(v),\q_2(v)$ are in the same fibre of $\tau_t$ and so by (x) we have that $d\big(\cs\co \q_1(v),\cs\co\q_1(v)\big)\leq \epsilon$ for all $v$. Then by continuity of $\kappa_{m+1}$ the values $\q_i(v)$, $i=1,2$ are also close in $\nss$ for each $v$. Therefore by Lemma \ref{lem:closelifts} there exist $\tilde \q_1,\tilde \q_2\in \cu^{k+1}(\ns/_{B_{m+1}})$ such that $d\big(\tilde\q_1(v),\tilde\q_2(v)\big)\leq \epsilon_2$ for every $v$ and such that $\kappa_{m+1}\co \tilde\q_i=\q_i$, $i=1,2$. Finally, from the expression $\rho_{\cs}(\q_i)=\sigma_{k+1}(\cs\co \q_i-\tilde\q_i)$ we obtain the inequality in \eqref{eq:rhoclose} by applying Lemma \ref{lem:vertiparathin}.\\
\indent Note that property (xi) also implies that
\begin{equation}\label{eq:modBm}
\forall\,\q_1,\q_2\in \cu^{k+1}(\nss)\textrm{ with } \tau_{h_m} \co\q_1 = \tau_{h_m}\co \q_2,\textrm{ we have } \rho_{\cs}(\q_1)-\rho_{\cs}(\q_2)\in B_m/B_{m+1}.
\end{equation}
Indeed, since $\psi_m$ is injective by (i), we have by (xi) that $\tau_{h_m} \co\q_1 = \tau_{h_m}\co \q_2$ implies $q_{m,m+1}\co \cs\co\q_1(v)=q_{m,m+1}\co \cs\co \q_2(v)$, so $\cs\co \q_1(v)$, $\cs\co \q_2(v)$ are in the same orbit of $B_m/B_{m+1}$ for every $v$. In particular we have $\rho_{\cs}(\q_1)-\rho_{\cs}(\q_2)=\sigma_{k+1}(\cs\co \q_1-\cs\co \q_2)$, whence \eqref{eq:modBm} follows.
\noindent To complete the proof we shall now use $\rho_{\cs}$ to find a new cross section $\cs': \nss\to \ns/B_{m+1}$ that is consistent with the factor $\nss_t$, and that is also congruent to $\cs$ modulo $B_m$, that is $q_{m,m+1}\co\cs'=q_{m,m+1}\co\cs$. If we find $\cs'$ then, firstly, thanks to the consistency, we can apply Lemma \ref{lem:crossfactor} with map $\psi=\tau_t$, obtaining a new fibre-surjective factor $\ns_{m+1}$ of $\ns/_{B_{m+1}}$ that is an extension of $\nss_t$ by $\ab_k/B_{m+1}$, and so (i) holds. Secondly $\psi_m\co q_{m,m+1}$ factors through $\ns_{m+1}$, which yields $\psi_{m+1}'$ satisfying (ii). Indeed, by definition of the relation $\sim$ in Lemma \ref{lem:crossfactor} we have that $x,y\in \ns/_{B_{m+1}}$ are in the same fibre of $\psi_{m+1}$ if and only if $\tau_t\co\kappa_{m+1}(x)=\tau_t\co\kappa_{m+1}(y)$ and $x-\cs'\co\kappa_{m+1}(x)=y-\cs'\co\kappa_{m+1}(y)$. The first equality here implies, thanks to the congruence mod $B_m$ and (xi), that $\psi_m\co q_{m,m+1}(\cs'\co \kappa_{m+1}(x))=\psi_m\co q_{m,m+1}(\cs'\co\kappa_{m+1}(y))$, and this together with the second equality gives $\psi_m\co q_{m,m+1}(x)=\psi_m\co q_{m,m+1}(y)$. Finally, property (iii) holds, since $x\sim y\Rightarrow \tau_t\co\kappa_{m+1}(x)=\tau_t\co\kappa_{m+1}(y)$ implies that $\tau_t\co\kappa_{m+1}$ factors through $\psi_{m+1}$, and it must then factor as $\pi_{m+1}$, by definition of $\kappa_{m+1}$. We will have thus found the next term in the sequence $S$.\\
\noindent \textbf{Averaging $\rho_{\cs}$ to obtain a fibre-surjective factor $\ns_{m+1}$ of $\ns/_{B_{m+1}}$:} to find $\cs'$, the idea is firstly to define a new function $\rho'$ by averaging $\rho_{\cs}$, in such a way that the approximate equality in \eqref{eq:rhoclose} becomes an equality for $\rho'$, and secondly then to show that $\rho'$ is itself a cocycle generated by some cross section, which will be the desired $\cs'$.\\
\indent Recall that, by \cite[Lemma 3.3.12 (ii) \& (iii)]{Cand:Notes1}, for each $i\in \mathbb{N}$ the map
$\beta_i:\cu^i(\nss)\to\cu^i(\nss_t)$, $\q\mapsto \tau_t\co \q$ is a totally surjective bundle morphism, and for every $\q\in \cu^i(\nss_t)$ the preimage $\beta_i^{-1}(\q)$ is a $(k-1)$-fold sub-bundle of $\cu^i(\nss)$. (Recall also Lemmas \ref{lem:surjmorphcubemeas} and \ref{lem:cube-set-CSM}.)
We define $\rho':\cu^{k+1}(\nss)\to \ab_k/B_{m+1}$ by
\begin{equation}\label{eq:avcocycle}
\rho'(\q)=\mathbb{E}_{\q'\in\beta_{k+1}^{-1}(\beta_{k+1}(\q))}\;\rho_{\cs}(\q').
\end{equation}
The averaging here is relative to the Haar measure on $\beta_{k+1}^{-1}(\beta_{k+1}(\q))$, and is well-defined for $\epsilon_2$ sufficiently small, thanks to \eqref{eq:rhoclose} (recall Definition \ref{def:calgaver}). By \eqref{eq:modBm} and the fact that $t>h_m$, we have that $\rho'(\q)-\rho_{\cs}(\q)\in B_m/B_{m+1}$ for every $\q\in \cu^{k+1}(\nss)$. Note also that it follows clearly from \eqref{eq:avcocycle} that
\begin{equation}\label{eq:rhoexact}
\forall\,\q_1,\q_2\in \cu^{k+1}(\nss)\textrm{ with } \tau_t\co \q_1 = \tau_t\co \q_2,\textrm{ we have }\rho'(\q_1)=\rho'(\q_2).
\end{equation}
\noindent We shall now prove that $\rho'$ is a cocycle by showing that it inherits the required properties from $\rho_{\cs}$ thanks to the linearity of the averaging in \eqref{eq:avcocycle}. The required properties are the two axioms from \cite[Definition 3.3.14]{Cand:Notes1}. Axiom (i) follows straight from the same property for $\rho_{\cs}$ and linearity using the fact that any set $\beta_{k+1}^{-1}(\beta_{k+1}(\q))$ is globally invariant under composition with automorphisms in $\aut(\{0,1\}^{k+1})$. To see the second axiom we take two adjacent cubes $\q_1,\q_2\in \cu^{k+1}(\nss)$ with concatenation $\q_3$ and embed them into a $(k+2)$-dimensional cube $\q\in \cu^{k+2}(\nss)$ as restrictions to two adjacent $(k+1)$-dimensional faces $F_1$ and $F_2$ in $\{0,1\}^{k+2}$. The concatenation of $F_1$ and $F_2$ is a diagonal subcube $F_3$. We have that the concatenation of $\q_1$ and $\q_2$ is the restriction of $\q$ to $F_3$. Note that the existence of $\q$ is guaranteed by simplicial completion \cite[Lemma 3.1.5]{Cand:Notes1}. By \cite[Lemma 3.3.12 (iii)]{Cand:Notes1}, we have a Haar probability on $\Omega=\beta_{k+2}^{-1}(\beta_{k+2}(\q))$. Then, by \cite[Lemma 3.3.12 (iv)]{Cand:Notes1}, the probability spaces $\beta_{k+1}^{-1}(\beta_{k+1}(\q_i))$, $i=1,2,3$ are faithfully embedded as factors and coupled in the big probability space $\Omega$. Using the concatenation property for $\rho$ in $\Omega$ (when the random cube is restricted to $F_1,F_2,F_3$) and linearity of expectation we obtain that $\rho'(c_3)=\rho'(c_1)+\rho'(c_2)$.
It remains to prove that the cocycle $\rho'$ is generated by a cross section. Let $\rho''$ denote the cocycle $\rho'-\rho_{\cs}$, which takes values in $B_m/B_{m+1}$ (as noted just before \eqref{eq:rhoexact}). It follows from \eqref{eq:rhoclose} that $d_2(\rho''(\q),0)\leq\epsilon_2$ for every $\q\in \cu^{k+1}(\nss)$, so for $\epsilon_2$ sufficiently small we have that $\rho''$ is a coboundary, by Lemma \ref{lem:smallco}. Thus for some Borel function $g:\nss\to B_m/B_{m+1}$ we have $\rho''(\q)=\sigma_{k+1}(g\co \q)$. Hence, defining the Borel cross section $\cs':\nss\to \ns/_{B_{m+1}}$ by $\cs'(y)= \cs(y)+g(y)$, we have $\rho'=\rho_{\cs'}$.
By \eqref{eq:rhoexact} we have that $\cs'$ is consistent with the factor $\nss_t$. Applying Lemma \ref{lem:crossfactor} as described above, we find the new term in the sequence $S$.
This completes the proof of Theorem \ref{thm:invlim}.
\medskip
\section{Rigidity of morphisms}
\medskip
Given a set $X$, a metric space $(Y,d)$, and maps $\phi,\phi':X\to Y$, we say that $\phi'$ is an \emph{$\epsilon$-modification} of $\phi$ if for every $x\in X$ we have $d(\phi(x),\phi'(x))\leq \epsilon$.
\begin{defn}
Let $\ns$ be a compact nilspace, and let $\nss$ be a $k$-step compact nilspace with a metric $d$ generating its topology. A map $\phi:\ns\to\nss$ is a $\delta$\emph{-quasimorphism} if for every $\q\in \cu^{k+1}(\ns)$ there exists $\q'\in \cu^{k+1}(\nss)$ such that $d\big(\phi\co\q(v),\q'(v)\big)\leq\delta$ for every $v\in \{0,1\}^{k+1}$.
\end{defn}
In this section we establish the following result.
\begin{theorem}\label{thm:rigidity}
Let $\nss$ be a $k$-step compact nilspace of finite rank, with a metric $d$ generating its topology. Then for every $\epsilon>0$ there exists $\delta>0$ such that the following holds. If $\ns$ is a compact nilspace and $\phi:\ns\to \nss$ is a Borel $\delta$-quasimorphism, then $\phi$ has an $\epsilon$-modification that is a continuous morphism.
\end{theorem}
\noindent In the proof we shall use the fact that the metric $d$ can be assumed to be invariant under the action of the structure group $\ab_k$ (Lemma \ref{lem:d-invariance}). We shall also use the following `rectification result' for cubes.
\begin{lemma}\label{lem:verticubapprox}
Let $\ns$ be a $k$-step compact nilspace with metric $d$. For every $\epsilon>0$ there exists $\delta>0$ such that the following holds. If $\q\in \cu^{k+1}(\ns)$ satisfies $d'\big(\pi_{k-1}\co \q(v,0),\pi_{k-1}\co\q(v,1)\big)\leq \delta$ for every $v\in \{0,1\}^k$, then $\q$ has an $\epsilon$-modification $\q'\in \cu^{k+1}(\ns)$ such that $\pi_{k-1}\co \q'(v,0) = \pi_{k-1}\co\q'(v,1)$ for every $v\in \{0,1\}^k$.
\end{lemma}
\begin{proof}
Suppose for a contradiction that for some $\epsilon>0$ there is a sequence of cubes $\q_n\in \cu^{k+1}(\ns)$ such that $d'\big(\pi_{k-1}\co \q_n(v,0),\pi_{k-1}\co\q_n(v,1)\big)\leq 1/n$ for every $v\in \{0,1\}^k$ and yet, for every cube $\q'\in \cu^{k+1}(\ns)$ such that $\pi_{k-1}\co \q'(v,0) = \pi_{k-1}\co\q'(v,1)$ for all $v\in\{0,1\}^k$, there is some $v\in \{0,1\}^{k+1}$ such that $d(\q_n(v),\q'(v))> \epsilon$. By compactness of $\cu^{k+1}(\ns)$, there is a subsequence $(\q_m)$ of $(\q_n)$ such that $\q_m\to \q^*\in \cu^{k+1}(\ns)$. The topology on $\cu^{k+1}(\ns)$ can be metrized by $d_\infty(\q_1,\q_2):=\max_{v\in \{0,1\}^{k+1}} d(\q_1(v),\q_2(v))$. It then follows from $d_\infty(\q_m, \q^*)\to 0$ and the triangle inequality for $d'$ that $\pi_{k-1}\co \q^*(v,0) = \pi_{k-1}\co\q^*(v,1)$ for every $v\in \{0,1\}^k$. However, it also follows from $\q_m\to \q^*$ that $d(\q_m(v),\q^*(v))\leq \epsilon$ for all $v\in \{0,1\}^{k+1}$, for sufficiently large $m$, and this contradicts our initial assumption since $\pi_{k-1}\co \q^*(\cdot,0) = \pi_{k-1}\co\q^*(\cdot,1)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:rigidity}]
We argue by induction on $k$, starting from the trivial case $k=0$. For $k>0$, suppose that the result holds for $k-1$, and fix any $\epsilon>0$. Let $\delta>0$ be a parameter to be determined later, suppose that $\phi:\ns\to\nss$ is a $\delta$-quasimorphism, and let $\phi_1=\pi_{k-1}\co\phi$. Then $\phi_1$ is a $\delta$-quasimorphism into $\cF_{k-1}(\nss)$. By induction, for some $\delta_1(\epsilon)>0$ to be fixed later, we have that if $\delta$ is small enough then $\phi_1$ has a $\delta_1$-modification $\phi_2:\ns\to \cF_{k-1}(\nss)$ that is a continuous morphism.
We shall now obtain a Borel measurable map $\phi_3:\ns\to \nss$ that lifts $\phi_2$, i.e. such that $\pi_{k-1}\co \phi_3=\phi_2$, and which is a $(\delta+\delta_1)$-modification of $\phi$, so that $\phi_3$ is a $\delta_2$-quasimorphism with $\delta_2=\delta_1+2\delta$. To obtain this, the idea is first to obtain such a lift locally for each neighbourhood in some appropriate cover of $\ns$, and then combine the local lifts into a global one. To this end we use the following set:
\[
G=\{(x,y):\pi_{k-1}(y)=\phi_2(x)\}\subset \ns\times \nss.
\]
We claim that $G$ is compact. Indeed, on one hand the map $f_1\times f_2:\ns\times \nss\to \cF_{k-1}(\nss)\times \cF_{k-1}(\nss)$, $(x,y)\mapsto (\phi_2(x),\pi_{k-1}(y))$ is continuous, and on the other hand $(x,y)\in G$ if and only if $f_1\times f_2(x,y)$ is in the diagonal of $\cF_{k-1}(\nss)\times \cF_{k-1}(\nss)$, which is a closed set in the product topology since $\cF_{k-1}(\nss)$ is Hausdorff. Hence $G$ is closed in $\ns\times \nss$ and therefore compact. We also claim that $G$ is a $\ab_k$-bundle over $\ns$, where $\ab_k$ is the $k$-th structure group of $\nss$. Indeed, the action of $\ab_k$ is given by $a\cdot (x,y)=(x,y+a)$, and the projection map is just $(x,y)\mapsto x\in \ns$.\\
\indent By Proposition \ref{prop:Gleason}, the bundle $G$ is locally trivial and so for every $p\in \ns$ we can find an open set $U_p\subset \ns$ containing $p$ with a continuous cross section $U_p\to G$, which must then be of the form $x\mapsto (x,\tau(x))$ for some continuous map $\tau:U_p\to \nss$. Thus $\phi_2(x)=\pi_{k-1}(\tau(x))$ for every $x\in U_p$. Then, by the definition of $d'$, since for all $x\in U_p$ we have
\[
d'\big(\pi_{k-1}(\tau(x)),\pi_{k-1}(\phi(x))\big)=d'(\phi_2(x),\phi_1(x))\leq \delta_1,
\]
in particular for $p$ there exists $z_p\in \ab_k$ such that $d\big(\tau(p)+z_p,\phi(p)\big)\leq \delta_1$. But then, since the function $f_p:x\mapsto \tau(x)+z_p$ is continuous on $U_p$, there is an open set $U_p'\subset U_p$ containing $p$ such that for all $x\in U_p'$ we have $d(\phi(x),\tau(x)+z_p)<\delta_1+\delta$. The function $f_p$ on $U'_p$ is the desired local lift of $\phi_2$.\\
\indent Now by compactness there is a finite covering of $\ns$ by such open sets $U_p'$. Let $U_1,\dots,U_s$ be the atoms of the Boolean algebra generated by the sets $U_p'$. For each $i\in [s]$ we fix one of the functions $f_p$ defined on $U_s$ and rename it $f_i$. We then define the Borel function $\phi_3$ by $\phi_3(x)=\sum_{i\in [s]}1_{U_i}(x) f_i(x)$.
We shall now apply an averaging argument to $\phi_3$ that will produce
a continuous morphism $\phi_4$ which is close to $\phi$ as desired.
Let $P_2=\{0,1\}^{k+1}\setminus\{0^{k+1}\}$. Note that since $\phi_2$ is a morphism $\ns\to \cF_{k-1}(\nss)$, it preserves $k$-cubes,
and since $\phi_3$ is a lift of $\phi_2$, by \cite[Remark 3.2.12]{Cand:Notes1} we have that $\phi_3$ also preserves $k$-cubes on $\nss$. This implies that for every $\q\in \cu^{k+1}(\ns)$ the restriction of $\phi_3\co \q$ to $P_2$ is a $(k+1)$-corner on $\nss$. For any given cube $\q\in \cu^{k+1}(\ns)$ let us denote by $\comp_0(\q)$ the value at $0^{k+1}$ of the unique completion in $\cu^{k+1}(\nss)$ of $(\phi_3\co \q)|_{P_2}$. Let
\begin{equation}\label{eq:fi4}
\phi_4(x)=\mathbb{E}_{\q\in \cu^{k+1}_x(\ns)}\, \comp_0(\q).
\end{equation}
Let us show that this averaging is well-defined. Note that for every $\q\in\cu^{k+1}_x(\ns)$ we have $\comp_0(\q)\sim_{k-1} \phi_3(x)$, indeed since the morphism $\phi_2$ preserves $(k+1)$-cubes, and $\phi_3$ is a lift of $\phi_2$, we must have $\comp_0(\q)$ and $\phi_3(x)$ both in the fibre $\pi_{k-1}^{-1}(\phi_2(x))$. We have to show that if $\delta_2>0$ is small enough then the set of values $\comp_0(\q)$, $\q\in \cu^{k+1}_x(\ns)$ has small diameter in the fibre. We will show that $\comp_0(\q)$ is actually close to $\phi_3(x)$ for every $\q\in \cu^{k+1}_x(\ns)$, which will also be useful later. For any $\q\in \cu^{k+1}_x(\ns)$, since $\phi_3$ is a $\delta_2$-quasimorphism, there is $\q_0\in \cu^{k+1}(\nss)$ such that for all $v$ we have $d\big(\phi_3\co\q(v),\q_0(v)\big)\leq\delta_2$. In particular, this inequality holds for all $v\in P_2$. By continuity of the corner-completion (Lemma \ref{lem:contcomp}) it follows that for $\lambda>0$, if $\delta_2$ is small enough then we have $d\big(\comp_0(\q),\q_0(0^{k+1})\big)\leq\lambda$. Hence $d\big(\phi_3(x),\comp_0(\q)\big)=d\big(\phi_3\co \q(0^{k+1}),\comp_0(\q)\big)\leq \delta_2+\lambda$. We can now fix $\lambda$ so that the averaging in \eqref{eq:fi4} is indeed well-defined.\\
\indent Let us now prove that $\phi_4$ is a morphism. By \cite[Lemma 3.2.13]{Cand:Notes1}, it suffices to show that for every $\q\in \cu^{k+1}(\ns)$ we have $\phi_4\co \q\in \cu^{k+1}(\nss)$. We shall do this using another averaging argument, but working this time with the tricube $T_{k+1}$. Let $B=T_{k+1}\setminus \{-1,1\}^{k+1}$. For every $t\in\hom_{\q\co \omega_{k+1}^{-1}}(T_{k+1},\ns)$, there is a unique completion of $\phi_3 \co t|_B$ to a morphism $t':T_{k+1}\to \nss$. Let $\q_t=t'\co \omega_{k+1}$, which is in $\cu^{k+1}(\nss)$ by \cite[Lemma 3.1.16]{Cand:Notes1}. If $\delta_2$ is small enough then the following function is well defined:
\[
\q_2: \{0,1\}^{k+1}\to \nss,\; v\mapsto \mathbb{E}_{t\in\hom_{\q\co \omega_{k+1}^{-1}}(T_{k+1},\ns)}\,\q_t(v).
\]
Indeed, by Corollary \ref{cor:ext-in-Tn}, for each $v$ this average is equal to $\mathbb{E}_{\q'\in \cu^{k+1}_{\q(v)}(\ns)} \comp_0(\q')=\phi_4(\q(v))$. Thus, we have to show that $\q_2\in \cu^{k+1}(\ns)$. To see this, recall that for every $\q'\in \cu^{k+1}_{\q(v)}(\ns)$ we have $\comp_0(\q')\sim_{k-1}\phi_3(\q(v))$, and so $\q_2(v)\sim_{k-1}\phi_3\co \q(v)$, which implies that $\pi_{k-1}\co\q_2= \phi_2\co \q\in \cu^{k+1}(\cF_{k-1}(\nss))$. Therefore, by \cite[Theorem 3.2.19]{Cand:Notes1} and the definition of degree-$k$ extensions, we just have to check that for some (any) $t_0$ we have $\sigma_{k+1}(\q_2-\q_{t_0})=0$. This follows from linearity of averaging (Lemma \ref{lem:avadd}) and the fact that for $\q_t,\q_{t_0}$ are cubes with the same projection to $\cF_{k-1}(\nss)$ for every $t$, so
\[
\sigma_{k+1}(\q_2-\q_{t_0})\;=\;\sum_v (-1)^{|v|}\; \mathbb{E}_{t\in\hom_{\q\co \omega_{k+1}^{-1}}(T_{k+1},\ns)}\,\q_t(v) - \q_{t_0}(v)\;=\;\mathbb{E}_t\, \sigma_{k+1}\big(\q_t - \q_{t_0}\big) = 0.
\]
Finally, let us prove that $\phi_4$ is continuous. The argument is similar to the proof of Theorem \ref{thm:Klepgen}. Consider the CSM structure on $\pi_0:\cu^{k+1}(\ns)\to \ns$, $\q\mapsto \q(0^{k+1})$, given by Lemma \ref{lem:cube-set-CSM}. For each $v\in \{0,1\}^{k+1}\setminus \{0^{k+1}\}$ let $\pi_v:\cu^{k+1}(\ns)\to \ns$, $\q\mapsto \q(v)$. We have that $\pi_v$ is continuous and each of its restrictions to a space $\cu^{k+1}_x(\ns)$ is measure-preserving, by Lemma \ref{lem:GoodPairHoms}. For each $v$ we let $f_v:\ns\to\cL(\cu^{k+1}(\ns),\nss)$ be the function sending $x$ to the restriction of $\phi_3\co\pi_v$ to $\cu_x^{k+1}(\ns)$. Then, just as in the proof of Theorem \ref{thm:Klepgen}, we have that $f_v$ is continuous, by Lemma \ref{lem:csmtech}. We now want to express $\phi_4$ in terms of the functions $f_v$. To that end, consider again the map $\xi : \ns \to \cL(\cu^{k+1}(\ns),\cor^{k+1}(\nss))$, sending $x$ to the function $\xi_x:\q\in\cu^{k+1}_x(\ns) \mapsto (\q(v)=f_v(x)(\q))_{v\neq 0^{k+1}}$. Just as in the proof of Theorem \ref{thm:Klepgen}, the continuity of each function $f_v$ implies that $\xi$ is continuous. We then use again the continuous map $\cK: \cL(\cu^{k+1}(\ns),\cor^{k+1}(\nss))\to \cL(\cu^{k+1}(\ns),\nss)$ that sends a function $g$ to $g\co \comp$, where $\comp$ is the unique completion of $(k+1)$-corners on $\nss$. Now note that $\cK\co\,\xi: \ns\to \cL(\cu^{k+1}(\ns),\nss)$ is precisely the map that sends each $x\in \ns$ to the function $\q\mapsto \comp_0(\q)$ on $\cu^{k+1}_x(\ns)$. This map is continuous, and by the discussion justifying the averaging \eqref{eq:fi4}, we have that in fact $\cK\co\,\xi$ takes values in the following subset of $\cL\big(\cu^{k+1}(\ns),\nss\big)$:
\[
\cU= \bigcup_{x\in \ns} \Big\{g\in L\big(\cu^{k+1}_x(\ns),\pi_{k-1}^{-1}(\phi_2(x))\big)~:~\forall \q\in \cu^{k+1}_x(\ns),\;\; d\big(g(\q),\phi_3(x)\big)\leq \delta_2+\lambda\Big\}.
\]
Now $\phi_4$ is the composition of $\cK\co\,\xi$ with the averaging operator $\cA:\,\cU \to \pi_{k-1}^{-1}(\phi_2(x))$ that sends a function $g$ to $\mathbb{E}_{\q\in \cu^{k+1}_x(\ns)} g(\q)$. Thus it now suffices to show that $\cA$ is continuous. Note that since $ \pi_{k-1}^{-1}(\phi_2(x))$ is homeomorphic to $\ab_k$, it follows from the definition of the metric $d_2$ on $\ab_k$ that if $\delta_2+\lambda$ is sufficiently small then every function $g\in \cU$ can be lifted continuously to a function $g':\cu^{k+1}_x(\ns)\to \mathbb{R}^n$, where $\ab_k=F\times \mathbb{T}^n$, in such a way that two functions $g_1,g_2\in \cU$ are close in the metric $d_1$ from \eqref{eq:gen-L1} if and only if their lifts $g_1',g_2'$ are close in the same metric but with $d$ now being the Euclidean distance on $\mathbb{R}^n$. By Lemma \ref{lem:Ltop-gen-restrict}, if $g_n\to g$ in the topology on $\cL(\cu^{k+1}(\ns),\nss)$ restricted to $L(\cu^{k+1}_x(\ns),\nss)$ , then $d_1(g_n,g)\to 0$, whence by the triangle inequality $\cA(g_n)\to \cA(g)$, so $\cA$ is continuous as required.
\end{proof}
\medskip
\section{Characterizing compact connected nilspaces of finite rank}\label{sec:CFRnilsnilm}
\medskip
Recall from \cite[Subsection 3.2.4]{Cand:Notes1} that given a $k$-step nilspace $\ns$ we denote by $\tran_i(\ns)$ the group of translations of height $i$ on $\ns$ (or $i$-translations). By a slight abuse of notation, when $\ns$ is a compact nilspace we shall write $\tran_i(\ns)$ for the group of $i$-translations that are also \emph{continuous} functions. A central goal in this section is to show that if $\ns$ has finite rank then $\tran_i(\ns)$ is a Lie group for each $i$ and $\tran_1(\ns)$ acts transitively on the connected components of $\ns$. This will then enable us to show that if $\ns$ has connected structure groups then it can be identified with a filtered nilmanifold (Theorem \ref{thm:toralnilspace}).
Throughout this section we shall abbreviate `compact and finite-rank' by writing `\textsc{cfr}'.
\begin{lemma}
Let $\ns$ be a \textsc{cfr} $k$-step nilspace. Then every element of $\tran(\ns)$ is a homeomorphism $\ns\to \ns$ preserving the Haar measure.
\end{lemma}
\begin{proof}
Each translation is an invertible continuous map from the compact Hausdorff space $\ns$ to itself and so its inverse is also continuous \cite[Theorem 26.6]{Munkres}. Moreover, the translation is a fibre-surjective automorphism of $\ns$, so it is measure preserving by Corollary \ref{cor:ctsfibsurmorph}.
\end{proof}
\noindent For a compact space $X$ and a space $Y$ with metric $d$, recall that $C(X,Y)$ denotes the space of continuous functions $f:X\to Y$, with the topology induced by the uniform metric $d_\infty(f_1,f_2)=\sup_{x\in X} d(f_1(x),f_2(x))$. We record the following basic fact.
\begin{lemma}\label{lem:transpolish}
Let $\ns$ be a \textsc{cfr} $k$-step nilspace and let $d$ be a metric generating the topology on $\ns$. Then for every $i\in [k]$, the group $\tran_i(\ns)$ equipped with the restriction of the uniform metric on $C(\ns,\ns)$ is a Polish group.
\end{lemma}
\begin{proof}
The group $H(\ns)$ of homeomorphisms on $\ns$, with the relative topology from $C(X,X)$, is a Polish group \cite[\S 1.3, Example (v)]{Be&Ke}. It follows from the definition of translations \cite[Definition 3.2.27]{Cand:Notes1} and the closure of cube sets that $\tran_i(\ns)$ is a closed subgroup of $H(\ns)$, hence it is also a Polish group. By definition the metric $d_\infty$ generates the topology on $\tran_i(\ns)$.
\end{proof}
\noindent Recall that by \cite[Lemma 3.2.31]{Cand:Notes1} a translation maps every class of the relation $\sim_{k-1}$ onto another such class. This enables us to define a translation on the factor $\ns_{k-1}=\cF_{k-1}(\ns)$, as in the following lemma. Recall also that the topology on $\ns_{k-1}$ is the quotient topology from $\ns$, which can be metrized by the quotient metric $d'$ defined in \eqref{eq:quotientmetric}. We then denote by $d'_\infty$ the uniform metric on $\tran(\ns_{k-1})$ relative to $d'$.
\begin{lemma}\label{lem:hdefn}
For each $i\in \mathbb{N}$ let $h$ be the map sending each $\alpha\in \tran_i(\ns)$ to the map $h(\alpha)$ on $\ns_{k-1}$ defined by $h(\alpha)(y)=\pi_{k-1}(\alpha(x))$, for any $x\in \ns$ such that $\pi_{k-1}(x)=y$. Then $h$ is a continuous homomorphism $\tran_i(\ns)\to \tran_i(\ns_{k-1})$.
\end{lemma}
\begin{proof}
\noindent That $h$ is a homomorphism $\tran_i(\ns)\to \tran_i(\ns_{k-1})$ follows from the definitions. The continuity follows similarly, thus
\begin{eqnarray*}
d'_\infty\big(h(\alpha_1),h(\alpha_2)\big)& =& \sup_{y\in \ns_{k-1}} d'\big(h(\alpha_1)(y),h(\alpha_2)(y)\big)\;\;=\;\;\sup_{x\in \ns} d'\big(\pi_{k-1}(\alpha_1(x)),\pi_{k-1}(\alpha_2(x))\big)\\
& = & \sup_{x\in \ns}\;\;\; \inf_{\substack{x_1\sim_{k-1} \alpha_1(x)\\ x_2\sim_{k-1} \alpha_2(x)}} d(x_1,x_2)\;\;\leq\;\;\sup_{x\in \ns}d\big(\alpha_1(x),\alpha_2(x)\big)= d_\infty(\alpha_1,\alpha_2).
\end{eqnarray*}
\end{proof}
\noindent An important result toward our goal in this section is that if an $i$-translation on $\ns_{k-1}$ is sufficiently close to the identity then it can be lifted to an $i$-translation on $\ns$, in the following sense.
\begin{lemma}\label{lem:small-trans-lift}
Let $\ns$ be a \textsc{cfr} $k$-step nilspace and let $i\in \mathbb{N}$. There exists $\epsilon>0$ such that if $\alpha\in\tran_i(\ns_{k-1})$ satisfies $d'_\infty(\alpha,\mathrm{id})< \epsilon$ then there exists $\beta\in\tran_i(\ns)$ such that $h(\beta)=\alpha$.
\end{lemma}
\begin{proof}
We shall use notation from \cite[Lemma 3.3.38]{Cand:Notes1}. Recall that by that lemma the translation bundle $\mathcal{T}^*(\alpha,\ns,i)=\cF_{k-1}(\mathcal{T}(\alpha,\ns,i))$ is, from the purely algebraic viewpoint, a degree-$(k-i)$ extension of $\ns_{k-1}$ by $\ab_k$, and note that by the results from Section \ref{sec:toprelims} we have that this extension is a continuous $\ab_k$-bundle. We show first that if $\epsilon$ is sufficiently small then this extension has a measurable cross section $\cs$ such that the corresponding cocycle $\rho_{\cs}$ is a coboundary.\\
\indent Let $\gamma$ be the projection $\mathcal{T}^*(\alpha,\ns,i)\to \ns_{k-1}$. We claim that if $\epsilon>0$ is sufficiently small, then we can choose a Borel cross section $\cs:\ns_{k-1}\to \cT^*$ with the property that for each $\cs(x)$, for every $(x_0,x_1)$ in the equivalence class $\pi_{k-1,\cT}^{-1}(\cs(x))\subset \cT$, we have $d(x_0,x_1)\leq \epsilon$. (Recall that we have $\alpha(\pi_{k-1}(x_0))=\pi_{k-1}(x_1)$ for $(x_0,x_1)\in \cT$, by \cite[Definition 3.3.34]{Cand:Notes1}.) Indeed, by definition $d'_\infty(\alpha,\mathrm{id})< \epsilon$ implies that, for each $x\in \ns_{k-1}$, there exist $x_0,x_1\in \ns$ with $\pi_{k-1}(x_0)=x$, $\pi_{k-1}(x_1)=\alpha(x)$, and such that $d(x_0,x_1)< \epsilon$. Then, by $\ab_k$-invariance of $d$, we have that every pair $(y_0,y_1)$ in the same equivalence class as $(x_0,x_1)$ in $\cT^*$ satisfies $d(y_0,y_1)\leq \epsilon$. We choose $\cs(x)$ to be one of these pairs, say $(x_0,x_1)$. Note that this choice can be made in a Borel measurable way. Indeed, using the fact that $\ns$ is a locally trivial $\ab_k$-bundle over $\ns_{k-1}$ (Proposition \ref{prop:Gleason}), we can construct $\cs$ as a piecewise continuous function.
Now let $\rho_{\cs}$ be the cocycle generated by $\cs$, defined for $\q\in \cu^{k-i+1}(\ns_{k-1})$ by $\rho_{\cs}(\q)= \sigma_{k-i+1}(\cs\co\q-\q')$ for any lift $\q'\in \cu^{k-i+1}(\cT^*)$ of $\q$. We have that $\rho_{\cs}$ is measurable (recall the end of the proof of Lemma \ref{lem:meascross}). We claim that, for some $\epsilon_2>0$ to be fixed later, if $\epsilon$ is sufficiently small then $\rho_{\cs}$ is also small in the sense that $d_2(\rho_{\cs}(\q),0_{\ab_k})\leq \epsilon_2$ for every $\q\in \cu^{k-i+1}(\ns_{k-1})$. To show this we first give an alternative expression of the function $\cs\co\q-\q'$. The lift $\q'$, being a cube on $\cT^*$, has itself a lift $\tilde \q\in \cu^{k-i+1}(\cT)$. By definition we have $\tilde \q=\tilde\q_0\times \tilde\q_1$ where $\pi_{k-1}\co\tilde\q_1(v) = \alpha (\pi_{k-1}\co\tilde\q_0(v))$ for all $v$. By the claim in the previous paragraph, for each $v$ we have that $\cs\co \q(v)$ is a class of pairs $(x_0,x_1)\in \pi_{k-1}^{-1}(\q(v))\times \pi_{k-1}^{-1}(\alpha\co\q(v))$ satisfying $d(x_0,x_1)\leq \epsilon$ and corresponding to a local translation $\phi$, thus $x_1=\phi(x_0)$ for every such pair. The difference $\cs\co\q(v)-\q'(v)$ is then the element $a_v\in \ab_k$ which has to be added to $\tilde\q_0(v)$ in order to have $\phi(\tilde\q_0(v)+a_v)=\tilde\q_1(v)$, namely $a_v=\phi^{-1}(\tilde\q_1(v))-\tilde\q_0(v)$. Thus, letting $a$ denote the function $\{0,1\}^{k-i+1}\to \ab_k$, $v\mapsto a_v$, we have $\rho_{\cs}(\cs\co\q-\q')=\sigma_{k-i+1}(a)$. To show that this must be a small element of $\ab_k$, we argue as follows. By assumption on $\alpha$, the cube $\tilde\q$ satisfies the premise of Lemma \ref{lem:verticubapprox} with $\delta=\epsilon$. Thus, given $\epsilon_1>0$ to be fixed later, if $\epsilon<\epsilon_1$ is sufficiently small then by that lemma there must exist $\q^*=\q^*_0\times \q^*_1\in \cu^{k-i+1}(\cT)$ such that for all $v$ we have $d(\q^*_i(v),\tilde\q_i(v))\leq \epsilon_1$ for $i=0,1$, and with $\pi_{k-1}\co \q^*_1=\pi_{k-1}\co \q^*_0$. For each $v$ let $a^*(v)=\q^*_1(v)-\q^*_0(v)\in \ab_k$. It now suffices to show that $d_2(a^*(v),a(v))\leq \epsilon_2/2^{k-i+1}$ for every $v$, for then, by virtue of $\arr{\q^*_0,\q^*_1}_1$ being a cube, we must have $\sigma_{k-i+1}(a^*)=0$ and so $\sigma_{k-i+1}(a)\leq \epsilon_2$ as claimed. To show that $d_2(a^*(v),a(v))$ is small we can apply Lemma \ref{lem:vertiparathin} with $\delta=2\epsilon_1$, since we know on one hand that $d(\q_0^*(v),\tilde \q_0(v))\leq \epsilon_1$, and on the other hand that $d(\phi^{-1}(\tilde\q_1(v)),\tilde\q_1(v))\leq \epsilon$ and $d(\q_1^*(v),\tilde \q_1(v))\leq \epsilon_1$, so that $d(\phi^{-1}(\tilde\q_1(v)),\q_1^*(v))\leq \epsilon+\epsilon_1\leq 2\epsilon_1$.
We can now apply Lemma \ref{lem:smallco}. Thus if $\epsilon_2$ is sufficiently small then there exists a Borel function $g:\cF_{k-1}(\ns)\to \ab_k$ such that $\rho_{\cs}(\q)=\sigma_{k-1+i}(g\co \q)$ for all $\q\in \cu^{k-i+1}\big(\cF_{k-1}(\ns)\big)$. Let $m:\cF_{k-1}(\ns)\to \cT^*$ be the measurable function $x\mapsto \cs(x)-g(x)$. We have that $m$ is a morphism, indeed for any $\q\in \cu^n(\cF_{k-1}(\ns))$, on one hand we clearly have that the $\cT^*$-valued map $m\co \q$ is a lift of $\q$, and on the other hand for some (any) lift $\q'\in \cu^n(\cT^*)$ of $\q$ we have $\sigma_{k-i+1}(m\co \q - \q')=\sigma_{k-i+1}(\cs\co \q - \q')-\sigma_{k-i+1}(g\co \q)=\rho_{\cs}(\q)-\rho_{\cs}(\q)=0$, so by \cite[Definition 3.3.13 (ii)]{Cand:Notes1} we have indeed $m\co\q\in \cu^n(\cT^*)$. Now \cite[Proposition 3.3.39]{Cand:Notes1} gives us an $i$-translation $\beta$ on $\ns$ that is a lift of $\alpha$, and is defined as follows: for each $x\in \ns$ we have $\beta(x)=\phi(x)$, where $\phi$ is the local translation corresponding to the equivalence class $m\co\pi_{k-1}(x)\in \cT^*$. It only remains to check that $\beta$ is Borel measurable, as then, being a morphism, it must be continuous by Theorem \ref{thm:Klepgen}. By \cite[Definition 3.3.34 and Proposition 3.3.36]{Cand:Notes1}, the nilspace $\cT$ is a continuous $\ab$-bundle over $\cT^*$, for the polish group $\ab=\{(z,z):z\in \ab_k\}\leq \ab_k\times \ab_k$. Therefore, arguing as in the end of the proof of Lemma \ref{lem:meascross}, we obtain a Borel cross section $\lambda:\cT^*\to\cT$, and so the map $\lambda\co m\co\pi_{k-1}:\ns\to \cT$ is Borel. Moreover, letting $\lambda\co m\co\pi_{k-1}(x)=(y_0,y_1)$, we have $\beta(x)=y_1+(x-y_0)$, so $\beta$ is indeed Borel.
\end{proof}
\noindent We shall need a useful criterion for a translation to lie in the kernel of $h$. Note that if $\alpha\in \ker(h)$ then, since by \cite[Lemma 3.2.31]{Cand:Notes1} the restriction of $\alpha$ to each fibre of $\pi_{k-1}$ is a local translation, we deduce that the map $x\mapsto \alpha(x)-x$ is a constant element of $\ab_k$ for every $x$ in a given fibre. Thus $\alpha$ induces a well-defined map $\alpha':\ns_{k-1}\to \ab_k$, $\pi_{k-1}(x) \mapsto \alpha(x)-x$. Conversely, given $\alpha':\ns_{k-1}\to \ab_k$ we can define $\alpha:\ns \to \ns$, $x\mapsto x+\alpha'(\pi_{k-1}(x))$, and we can ask when is $\alpha$ in $\tran(\ns)$.
\begin{lemma}\label{lem:h-kernel}
Let $i<k$. Then $\alpha\in \ker (h)\cap \tran_i(\ns)$ if and only if $\alpha'\in \hom(\ns_{k-1},\cD_{k-i}(\ab_k))$.
\end{lemma}
\begin{proof}
By \cite[Lemma 3.2.32]{Cand:Notes1} we know that $\alpha\in \tran_i(\ns)$ if and only if for every $\q\in \cu^n(\ns)$ we have $\arr{\q,\alpha\co \q}_i\in \cu^{n+i}(\ns)$. If $\alpha\in \ker(h)\cap \tran_i(\ns)$ then we can take the difference $\alpha\co \q-\q=\alpha'\co\q'$ for $\q'=\pi_{k-1}\co\q$, and by definition of degree-$k$ extensions this difference must be a cube in $\cD_k(\ab_k)$. But we also have $\arr{\q,\alpha\co \q}_i-\arr{\q,\q}_i\in \cu^{n+i}(\cD_k(\ab_k))$, and this equals $\arr{0,\alpha'\co\q'}_i$. By \cite[Lemma 3.3.37]{Cand:Notes1}, we have $\arr{0,\alpha'\co\q'}_i\in \cu^{n+i}(\cD_k(\ab_k))$ if and only if $\alpha'\co\q'\in \cu^n(\cD_{k-i}(\ab_k)$. The converse follows similarly.
\end{proof}
\noindent The next main tool that we need, Lemma \ref{lem:hom-oscillation} below, is a type of rigidity result for morphisms between abelian torsors of the form $\cD_k(\ab)$. The proof will use the following basic fact.
\begin{lemma}\label{lem:abhomoscil}
Let $\ab'$ be a \textsc{cfr} abelian group, and let $d_2$ be the metric from Definition \ref{def:calgmetric}. Then there exists $\eta>0$ such that for every compact abelian group $\ab$ and every non-constant continuous affine homomorphism $\phi:\ab\to \ab'$, there exist $x,y\in \ab$ such that $d_2(\phi(x),\phi(y))\geq \eta$.
\end{lemma}
\begin{proof}
We can assume without loss that $\phi$ is a homomorphism (not just an affine one). If $\ab'$ is a finite group then the claim is verified by any two points $x,y$ such that $\phi(x)\neq \phi(y)$. We may therefore suppose that $\ab'=F\times \mathbb{T}^n$ with $n>0$ and $F$ a finite abelian group, and that for some $i\in [n]$ the projection $\pi_i:\mathbb{T}^n\to \mathbb{T}$, $x\mapsto x_i$ satisfies that $\pi_i\co\phi(\ab)$ is a nontrivial subgroup of $\mathbb{T}$. Identifying $\mathbb{T}^n$ with $[0,1)^n$, there is then $y\in \ab$ such that $|\pi_i\co\phi(y)-0|_{\mathbb{T}}\geq 1/4$ and so with $x=0_{\ab}$ we have $d_2(\phi(x),\phi(y))\geq 1/4$.
\end{proof}
\begin{lemma}\label{lem:hom-oscillation}
Let $k,\ell\in\mathbb{N}$, let $\ab,\ab'$ be compact abelian groups, and suppose that $\ab'$ has finite rank. Then there exists $\epsilon=\epsilon(\ell,\ab')>0$ such that if $\phi\in \hom(\cD_k(\ab),\cD_\ell(\ab'))$ satisfies $d_2(\phi(x),\phi(y))\leq \epsilon$ for every $x,y\in \ab$ then $\phi$ is a constant function.
\end{lemma}
\begin{proof}
Since $\hom(\cD_k(\ab),\cD_\ell(\ab'))\subseteq\hom(\cD_1(\ab),\cD_\ell(\ab'))$ we can assume that $k=1$. Let $\phi$ be an arbitrary non-constant morphism from $\cD_1(\ab)$ to $\cD_\ell(\ab')$. As mentioned in \cite[Example 2.2.13]{Cand:Notes1} (recall also \cite[Definition 2.2.30]{Cand:Notes1}), the map $\phi$ is a polynomial map of degree at most $\ell$. Thus, if for every $t\in \ab$ and every $f:\ab\to \ab'$ we denote by $\Delta_t f$ the function $x\mapsto f(x+t)-f(x)$, then there is $i<\ell$ and elements $t_1,t_2,\dots,t_i\in \ab$ such that $\phi'=\Delta_{t_1}\co \Delta_{t_2}\co\cdots\co \Delta_{t_i}\phi$ is non-constant but $\Delta_t \phi'$ is constant for every $t\in \ab$. It follows that $\phi'$ is a non-constant continuous affine homomorphism from $\ab$ to $\ab'$, so by Lemma \ref{lem:abhomoscil} there are $x,y\in \ab$ with $d_2(\phi'(x),\phi'(y))\geq \eta(\ab')>0$. This cannot hold if $\max_{x,y}d_2(\phi(x),\phi(y))$ is too small, so this quantity must have a positive lower bound (depending only on $\ab'$ and $\ell$).
\end{proof}
Lemma \ref{lem:hom-oscillation} has the following consequence.
\begin{corollary}\label{cor:hom-oscillation}
Let $\ell\in \mathbb{N}$, let $\ns$ be a $k$-step compact nilspace, and let $\ab$ be a \textsc{cfr} abelian group. Then there exists $\epsilon=\epsilon(\ell,\ab)>0$ such that if $\phi\in \hom(\ns,\cD_\ell(\ab))$ satisfies $d_2(\phi(x),\phi(y))\leq \epsilon$ for every $x,y\in \ns$ then $\phi$ is constant. In particular, for any $x_0\in \ns$, $z_0\in \ab$, the set $Y=\{\phi\in \hom(\ns,\cD_\ell(\ab)): \phi(x_0)=z_0\}$, equipped with the metric $d_\infty(\phi,\phi')=\max_{x\in \ns}d_2(\phi(x),\phi'(x))$, is discrete.
\end{corollary}
\begin{proof}
Suppose that $d_2(\phi(x),\phi(y))<\epsilon$ for every $x,y\in \ns$, where $\epsilon=\epsilon(\ell,\ab)$ is the constant from Lemma \ref{lem:hom-oscillation}. We prove by induction on $k$ that $\phi$ is constant.\\
\indent If $k=1$ then $\ns$ is abelian and the result follows from Lemma \ref{lem:hom-oscillation}. Suppose that the statement holds for $k-1$. Then Lemma \ref{lem:hom-oscillation} tells us that $\phi$ is constant on the $\sim_{k-1}$ classes of $\ns$. This means that $\phi$ can be regarded as a function on $\cF_{k-1}(\ns)$, and so by our assumption $\phi$ is constant.\\
\indent To see the last sentence in the corollary, note that if $\phi,\phi'\in Y$ satisfy $d_\infty(\phi,\phi')=d_\infty(\phi-\phi',0)\leq \epsilon/2$, then by the triangle inequality $\phi-\phi'$ satisfies the premise of the previous sentence in the corollary, and then by that sentence and the definition of $Y$ we deduce that $\phi-\phi'$ must be the constant $0$.
\end{proof}
With these results we can now give a useful description of the kernel of $h$.
\begin{lemma}\label{lem:h-kernel-Lie}
Let $\ns$ be a \textsc{cfr} $k$-step nilspace, and let $h:\tran(\ns)\to \tran(\ns_{k-1})$ be the homomorphism from Lemma \ref{lem:hdefn}. Then $\ker(h)$ is isomorphic as a topological group to $F\times \ab_k$ for some discrete group $F$. In particular $\ker(h)$ is a Lie group.
\end{lemma}
\begin{proof}
Let $x\in \ns$ be arbitrary and let $F$ be the stabilizer $\stab_{\ker(h)}(x)=\{\alpha\in \ker(h):\alpha(x)=x\}$. Let $\tau:\ab_k\to \ker(h)$ be the map sending $z\in \ab_k$ to the translation $\tau_z: x\mapsto x+z$, and note that $\tau$ is an isomorphism of topological groups between $\ab_k$ and $\tau(\ab_k)\leq \ker(h)$. Now given any $\alpha \in \ker(h)$, letting $z=\alpha(x)-x$ we have that $\alpha-\tau_z\in F$, so $\ker(h)=F\cdot \tau(\ab_k)$. Moreover, since every $\alpha\in F$ must in fact stabilize every point in the fibre of $x$ (as the trivial local translation from this fibre to itself), it follows that $F\lhd\,\ker(h)$. We also have $\tau(\ab_k)\lhd\,\ker(h)$. (In fact $\tau(\ab_k)\subset Z(\ker(h))$.) Using Lemma \ref{lem:h-kernel} and Corollary \ref{cor:hom-oscillation} we see that $F$ is discrete. It follows that $\ker(h)\cong F\times \ab_k$ as claimed. Since $\ab_k$ is a Lie group, the proof is complete.
\end{proof}
\noindent If $G$ is a topological group then we denote the connected component of $\mathrm{id}_G$ by $G^0$. This is a closed normal subgroup of $G$. It is also standard that $G$ is a Lie group if and only if $G^0$ is a Lie group.
\begin{theorem}\label{thm:transurj} Let $\ns$ be a \textsc{cfr} $k$-step nilspace and let $i\in [k]$. Then the following statements hold.\\ \vspace{-0.7cm}
\begin{enumerate}
\item $\tran_i(\ns)$ is a Lie group,
\item $h\big(\tran_i(\ns)^0\big)=\tran_i(\ns_{k-1})^0$.
\end{enumerate}
\end{theorem}
\begin{proof} We prove the statements by induction on $k$. The case $k=0$ being trivial, let $k>0$ and suppose that the statements hold for $k-1$.
To see statement $(i)$, note first that since $h$ is a continuous homomorphism and $\tran_i(\ns)$ is closed, we have that $h(\tran_i(\ns))$ is a closed subgroup of $\tran_i(\ns_{k-1})$. By induction the latter is a Lie group, so by Cartan's theorem $h(\tran_i(\ns))$ is a Lie group \cite[Theorem 2.12.6]{Var}. By Lemma \ref{lem:h-kernel-Lie} we have that $\ker(h)$ is a Lie group, and so $H:= \tran_i(\ns)\cap\ker (h)$ is a Lie group (being a closed subgroup of $\ker(h)$). On the other hand $h(\tran_i(\ns))$ is isomorphic to $\tran_i(\ns)/H$ as a topological group (see \cite[Theorem 1.2.6]{Be&Ke}). Therefore $\tran_i(\ns)/H$ must be a Lie group (this can be seen using \cite[Theorem 3.1]{GleasonSmall}). We can now deduce that $\tran_i(\ns)$ is itself a Lie group, by \cite[Lemma A.3]{HK}.
Next we show that
\begin{equation}\label{eq:inclutrans}
\tran_i(\ns_{k-1})^0\subseteq h(\tran_i(\ns)).
\end{equation}
To see this we use that $\tran_i(\ns_{k-1})$ is a Lie group and so every element $\alpha\in\tran_i(\ns_{k-1})^0$ is connected to the identity by a continuous path $p:[0,1]\to\tran_i(\ns_{k-1})$ with $p(0)=\mathrm{id}$ and $p(1)=\alpha$. For $n\in\mathbb{N}$ let $\alpha_i=p((i-1)/n)^{-1}p(i/n)$ for $i\in [n]$ and let $\alpha=\prod_{i=1}^n\alpha_i$.
Lemma \ref{lem:small-trans-lift} implies that if $n$ is sufficiently large then for every $\alpha_i$ there is $\beta_i\in\tran_i(\ns)$ with $h(\beta_i)=\alpha_i$. Let $\beta=\prod_{i=1}^n\beta_i$. Since $h$ is a homomorphism we have $h(\beta)=\alpha$.
To obtain statement $(ii)$, we first show that
\begin{equation}\label{eq:inclutrans2}
h(\tran_i(\ns)^0)= h(\tran_i(\ns))^0.
\end{equation}
On one hand we have $h(\tran_i(\ns)^0)\subset h(\tran_i(\ns))^0$ since $h$ is continuous. On the other hand $h(\tran_i(\ns))^0$ is open in $h(\tran_i(\ns))$
and $h$ is an open map $\tran_i(\ns)\to h(\tran_i(\ns))$ (again by \cite[Theorem 1.2.6]{Be&Ke}), so $h(\tran_i(\ns)^0)$ is open in $h(\tran_i(\ns))$. Since it is also closed, we have that both $h(\tran_i(\ns)^0)$ and its complement in $h(\tran_i(\ns))^0$ are open, so this complement must be empty, since $h(\tran_i(\ns))^0$ is connected, and \eqref{eq:inclutrans2} follows. Now in statement $(ii)$, the inclusion $h(\tran_i(\ns)^0)\subset \tran_i(\ns_{k-1})^0$ follows by continuity of $h$, and for the opposite inclusion, note that by \eqref{eq:inclutrans} we have $\tran_i(\ns_{k-1})^0\subseteq h(\tran_i(\ns))^0$, so by \eqref{eq:inclutrans2} we have $\tran_i(\ns_{k-1})^0\subseteq h(\tran_i(\ns)^0)$.
\end{proof}
\begin{remark}
It follows from statement $(ii)$ above that $\tran_i(\ns)^0$ is a topological extension of $\tran_i(\ns_{k-1})^0$ by $\ker(h)\cap \tran_i(\ns)^0$, that is we have the short exact sequence
\[
0\;\;\xrightarrow{\makebox[0.5cm]{}}\;\; \ker(h)\cap \tran_i(\ns)^0\;\; \xrightarrow{\makebox[0.5cm]{$\iota$}}\;\; \tran_i(\ns)^0 \;\;\xrightarrow{\makebox[0.5cm]{$h$}}\;\; \tran_i(\ns_{k-1})^0 \;\;\xrightarrow{\makebox[0.5cm]{}}\;\; 0.
\]
This gives another way to obtain statement $(i)$ by induction given $(ii)$, using that a topological extension of a Lie group by a Lie group is again a Lie group (see \cite[\S 2.6]{Tao:Hilbert}).
\end{remark}
\begin{corollary}\label{cor:actrans} The Lie group $\tran(\ns)^0$ acts transitively on the connected components of $\ns$.
\end{corollary}
\begin{proof}
We argue by induction on $k$. For $k>0$ suppose that $\tran(\ns_{k-1})^0$ acts transitively on the connected components of $\ns_{k-1}$. Let $\tau$ be the isomorphism $\ab_k\to \tau(\ab_k)\leq \tran(\ns)$ in the proof of Lemma \ref{lem:h-kernel-Lie}. Consider the group $\tau(\ab_k)\cdot \tran(\ns)^0 = \{\tau(z)\,\alpha:z\in \ab_k,\,\alpha\in\tran(\ns)^0\} \leq \tran(\ns)$, and let $M$ be any connected component of $\ns$. We claim that $\tau(\ab_k)\cdot \tran(\ns)^0$ acts transitively on $M$. Indeed, for any $x,y\in M$, by Theorem \ref{thm:transurj} (ii) and the induction hypothesis, there exists $\alpha\in \tran(\ns)^0$ such that $\pi_{k-1}\co\alpha(x)=\pi_{k-1}(y)$. Combining this with the transitivity of $\ab_k$ on the fibres of $\pi_{k-1}$ we deduce that there exists $\alpha'\in \tau(\ab_k)\cdot \tran(\ns)^0$ such that $\alpha'(x)=y$, which proves our claim. It follows that $\tran(\ns)$ acts transitively on $M$, and we can then deduce that $\tran(\ns)^0$ also acts transitively on $M$ by standard results (e.g. \cite[Chapter II, Theorem 3.2 and Proposition 4.3]{Helga}).
\end{proof}
\begin{example}\label{ex:heisentoral}
The main results of this section so far can be illustrated with the Heisenberg nilmanifold $H/\Gamma$, where $H=\begin{psmallmatrix} 1 & \mathbb{R} & \mathbb{R}\\[0.1em] & 1 & \mathbb{R} \\[0.1em] & & 1 \end{psmallmatrix}$, $\Gamma=\begin{psmallmatrix} 1 & \mathbb{Z} & \mathbb{Z}\\[0.1em] & 1 & \mathbb{Z} \\[0.1em] & & 1 \end{psmallmatrix}$ (recall \cite[Example 2.3.2]{Cand:Notes1}). Let $\ns$ denote $H/\Gamma$ equipped with the cubes associated with the lower central series on $H$. We then have that $\ns$ is a 2-step compact nilspace, with 1-step factor $\ns_1=\mathbb{T}^2$ equipped with the standard degree-1 cubes, and we compute that
\[
\tran_1(\ns)=\begin{psmallmatrix} 1 & \mathbb{R} & \mathbb{T}\\[0.1em] & 1 & \mathbb{R} \\[0.1em] & & 1 \end{psmallmatrix}, \;\;\tran_2(\ns)=\begin{psmallmatrix} 1 & 0 & \mathbb{T}\\[0.1em] & 1 & 0 \\[0.1em] & & 1 \end{psmallmatrix} \cong \mathbb{T},\;\; \tran(\ns_1)=\mathbb{T}^2.
\]
We also check that $\ker(h)=\begin{psmallmatrix} 1 & \mathbb{Z} & \mathbb{T}\\[0.1em] & 1 & \mathbb{Z} \\[0.1em] & & 1 \end{psmallmatrix}\;\cong\; \mathbb{Z}^2 \times \mathbb{T}$. Note that the structure groups of $\ns$ are tori ($\ab_2=\mathbb{T}$, $\ab_1=\mathbb{T}^2$). Thus $\ns$ is also an example of the type of connected \textsc{cfr} nilspace to which we turn next.
\end{example}
\subsection{Toral nilspaces}
\indent Recall that a \textsc{cfr} abelian group $\ab$ has torsion-free dual group if and only if $\ab\cong \mathbb{T}^n$ for some $n\geq 0$.
\begin{defn} We say that a $k$-step nilspace $\ns$ is \emph{torsion-free} if every structure group of $\ns$ has torsion-free dual group. We call $\ns$ a \emph{toral nilspace} if it is \textsc{cfr} and torsion-free.
\end{defn}
\noindent Thus a $k$-step compact nilspace is toral if and only if its structure groups are all tori. One of the reasons for treating these nilspaces in particular is that they occur naturally in the study of the regularizations of ultralimits of functions on abelian groups, which can be used to prove inverse theorems for the Gowers uniformity norms (see the characteristic-0 case of \cite[Definition 1.3]{Szegedy:HFA}).
\begin{lemma}\label{lem:torcon}
Every toral nilspace is connected.
\end{lemma}
\begin{proof}
Recall the general fact that a fibre bundle is connected if the base space and the fibre are connected (see \cite[Part I, \S2.12]{Steenrod}). Thus if $\ab$ has finite rank and is connected then a continuous $\ab$-bundle over a connected space is connected. Now given a $k$-step toral nilspace $\ns$, with factors $\ns_i$, $i\in [k]$, a simple induction on $i$ using this fact shows that $\ns$ is connected.
\end{proof}
\begin{remark}
The converse of Lemma \ref{lem:torcon} is false. Consider for example the connected 2-step nilspace $\ns$ consisting of the circle group $\mathbb{T}$ with cube sets $\cu^n(G_\bullet)$ for the degree-2 filtration $G_\bullet$ with $G_0=G_1=\mathbb{T}$ and $G_2=\{0,1/2\}\cong \mathbb{Z}/2\mathbb{Z}$ (viewing $\mathbb{T}$ as $[0,1)$ with addition mod 1). A straightforward calculation shows that $x\sim_1 y$ in $\ns$ if and only if $x-y\in G_2$, whence the structure group $\ab_2$ is the disconnected group $G_2$.
\end{remark}
We can now establish the main result of this section.
\begin{theorem}\label{thm:toralnilspace}
Let $\ns$ be a $k$-step toral nilspace, let $G=\tran(\ns)^0$, let $G_\bullet$ denote the degree-$k$ filtration $(\tran_i(\ns)^0)_{i\geq 0}$, and for an arbitrary fixed $x\in \ns$ let $\Gamma=\stab_G(x)$. Then $\ns$ is isomorphic as a compact nilspace to the nilmanifold $G/\Gamma$ with cube sets $\cu^n(\ns)=(\cu^n(G_\bullet)\cdot \Gamma^{\{0,1\}^n})/\Gamma^{\{0,1\}^n}$, $n\geq 0$.
\end{theorem}
\noindent Note that for the nilmanifold obtained here, every group in the filtration $G_\bullet$ is connected. (The fact that $G_\bullet$ here is indeed a filtration follows from the fact that if $H_1,H_2$ are closed subgroups of a topological group and one of these subgroups is connected, then $[H_1,H_2]$ is connected.)
\begin{proof} We argue by induction on $k$. For $k=1$, by Lemma \ref{lem:topkfolderg} we have that $\ns$ is the principal homogeneous space of a torus with cubes being the projections of the standard cubes on the torus, so the statement holds. Supposing then that the statement holds for $k-1$, we fix $x\in \ns$ and let $\Gamma=\stab_G(x)$.
We first claim that $\Gamma$ is discrete. To see this, note that $h(\Gamma)$ is a subgroup of the stabilizer of $\pi_{k-1}(x)$ in $\tran(\ns_{k-1})$, so by induction it is discrete. Then using that $h^{-1}(h(\Gamma))$ is a union of cosets of $\ker(h)$, we have that it suffices to show that $\Gamma\cap\ker(h)$ is discrete. But this follows from Lemma \ref{lem:h-kernel-Lie}, since no non-trivial element of $\tau(\ab_k)$ stabilizes $x$.\\
\indent By Corollary \ref{cor:actrans} the Lie group $\tran(\ns)^0$ acts transitively on the connected space $\ns$, and it follows that $\ns$ is homeomorphic to the coset space $G/\Gamma$ (see \cite[Chapter II, Theorem 3.2]{Helga}). Therefore $\Gamma$ is cocompact.
It now only remains to determine the cubes on $\ns$. Recall from \cite[Definition 3.2.38]{Cand:Notes1} that two cubes $\q_1,\q_2\in \cu^n(\ns)$ are said to be translation equivalent if there is an element $\q\in \cu^n(G_\bullet)$ such that $\q_2(v)=\q(v)\cdot \q_1(v)$. We claim that for every cube $\q\in \cu^n(\ns)$ there is a cube $\q'\in \cu^n(\ns)$ that is translation equivalent to the constant $x$ cube and such that $\pi_{k-1}\co \q=\pi_{k-1}\co \q'$. Indeed, given $\q\in \cu^n(\ns)$, we have $\pi_{k-1}\co\q\in \cu^n(\ns_{k-1})$, and by induction the latter cube is translation equivalent to the cube with constant value $x'=\pi_{k-1}(x)$, i.e. it is of the form $\tilde\q\cdot x'$ for some cube $\tilde \q$ on the group $\tran(\ns_{k-1})^0$ with the filtration $\big(\tran_i(\ns_{k-1})^0\big)_{i\geq 0}$. Now by the unique factorization result for these cubes \cite[Lemma 2.2.5]{Cand:Notes1}, we have $\tilde\q={\tilde g_0}^{F_0}\cdots {\tilde g_{2^n-1}}^{F_{2^n-1}}$ where $\tilde g_j\in \tran_{\codim(F_j)}(\ns_{k-1})^0$. Theorem \ref{thm:transurj} (ii) then tells us that for each $j\in [0,2^n)$ there is $g_j\in \tran_{\codim(F_j)}(\ns)^0$ such that $h(g_j)=\tilde g_j$.
Let $\q^*$ be the cube in $\cu^n(\tran(\ns)^0)$ defined by $\q^*={g_0}^{F_0}\cdots {g_{2^n-1}}^{F_{2^n-1}}$. Let $\q'=\q^*\cdot x$. This is in $\cu^n(\ns)$, and is translation equivalent to the constant $x$ cube. Moreover, by construction we have
\[
\pi_{k-1}\co \q'= \pi_{k-1}(\q^*\cdot x)= \Big(\prod_j h(g_j)^{F_j}\Big)\cdot x'= \Big(\prod_j \tilde g_j^{F_j}\Big)\cdot x'=\tilde \q \cdot x'=\pi_{k-1}\co \q.
\]
This proves our claim.
It follows from \cite[Theorem 3.2.19]{Cand:Notes1} and the definition of degree-$k$ bundles (in particular \cite[(3.5)]{Cand:Notes1}) that $\q-\q'\in \cu^n(\cD_k(\ab_k))$. But then using translations from $\tau(\ab_k)$ we can correct $\q'$ further to obtain $\q$, thus showing that $\q$ is itself a translation cube with translations from $\tran(\ns)^0$. (Such a correction procedure has been used in previous arguments, see for instance the proof of \cite[Lemma 3.2.25]{Cand:Notes1}.)
We have thus shown that $\cu^n(\ns)\subset (\cu^n(G_\bullet)\cdot \Gamma^{\{0,1\}^n})/\Gamma^{\{0,1\}^n}$. The opposite inclusion is clear, by definition of the groups $\tran_i(\ns)$.
\end{proof}
\begin{remark}\label{rem:simpconn} In Theorem \ref{thm:toralnilspace} we can replace the groups $G_i$ by \emph{simply connected} Lie groups $\tilde G_i$ while conserving the nilspace isomorphism with $\ns$. More precisely, let $\tilde G$ be the universal covering group of the connected Lie group $G$, with covering homomorphism $p:\tilde G\to G$ (see \cite[\S 47, Theorem 61]{Pont}). Let $\tilde\Gamma=p^{-1}(\Gamma)$, and for each $i\geq 0$, let $\tilde G_i$ be the identity component of the closed (hence Lie) subgroup $p^{-1}(G_i)$ of $\tilde G$. Then $\tilde G_\bullet=(\tilde G_i)_{i\geq 0}$ is a degree-$k$ filtration of closed connected (hence simply-connected \cite[Corollary 1.2.2]{C&G}) Lie subgroups of $\tilde G$. It can be checked in a straightforward way that $\tilde G/\tilde \Gamma$ with the cubes determined by $\tilde G_\bullet$ is isomorphic to $\ns$ as a compact nilspace.
\end{remark}
\noindent We know by a theorem of Palais and Stewart that $\ns$ is the total space of a principal torus bundle over a torus if and only if it is a 2-step nilmanifold (see \cite[Theorem 1]{P&S}). Theorem \ref{thm:toralnilspace} can be viewed as a generalization of this result for arbitrary step, in the category of compact nilspaces (in particular, making use of the assumed cube structure).
One may wonder whether, by adding some simple assumption in Theorem \ref{thm:toralnilspace}, the description of the filtered nilmanifold can then be made even more precise. For instance, having seen Example \ref{ex:heisentoral}, we may seek some simple additional condition under which Theorem \ref{thm:toralnilspace} yields the Heisenberg nilmanifold $H/\Gamma$. In this direction we have the following result.
\begin{proposition}
Let $\ns$ be a 2-step toral nilspace of dimension 3. Then $\ns$ is isomorphic to one of the following compact nilspaces: \vspace{-0.15cm}
\begin{enumerate}
\item The torus $\mathbb{R}^3/\mathbb{Z}^3$ with cube structure determined by a filtration of degree at most 2 on $\mathbb{R}^3$. \vspace{-0.15cm}
\item The Heisenberg nilmanifold $H/\Gamma$ with cube structure determined by the lower central series on $H$.
\end{enumerate}
\end{proposition}
\noindent Here the dimension is that of $\ns$ when viewed as a manifold (a view which we noted after Lemma \ref{lem:finrankloctriv}).
\begin{proof}
Applying Theorem \ref{thm:toralnilspace} and Remark \ref{rem:simpconn}, we have that $\ns$ is isomorphic as a compact nilspace to $G/\Gamma$ for a simply-connected Lie group $G$ with a filtration $G_\bullet$ of degree at most 2 consisting of closed (therefore Lie) simply-connected groups $G_i\leq G$, and a discrete cocompact subgroup $\Gamma\leq G$. The assumption that $\ns$ has dimension 3 implies that $G$ is 3-dimensional.
If $\ns$ is a 1-step nilspace then $\ns$ is isomorphic to the 3-torus $\mathbb{R}^3/\mathbb{Z}^3$ in case $(i)$, with the standard degree-1 cube structure.
If $\ns$ is not 1-step then, if $G$ is abelian, we have again that $\ns$ is isomorphic to $\mathbb{R}^3/\mathbb{Z}^3$ but now the cube structure is given by a degree-2 filtration on $\mathbb{R}^3$ (for example it could be the filtration determining the degree-2 structure on $\mathbb{R}^3/\mathbb{Z}^3$ defined in \cite[Definition 2.2.30]{Cand:Notes1}). The remaining possibility is that $G$ is non-abelian 2-step nilpotent. The simply-connected Lie group $G$ is uniquely determined by its Lie algebra, which is a real non-abelian, 2-step nilpotent Lie algebra of dimension 3. It follows from the Bianchi classification \cite{Bianchi} that the only such Lie algebra is the 3-dimensional Heisenberg Lie algebra, and so $G$ is the Heisenberg group. Finally, we have that the only degree-2 filtration of closed connected subgroups of $G$ is the lower central series.
\end{proof}
\noindent To close this section, we prove additional pleasant properties of toral nilspaces. In particular, property $(iii)$ below gives a simple relation between the translation groups and the structure groups.
\begin{proposition}\label{prop:abstranstrucrel}
Let $\ns$ be a $k$-step toral nilspace. For each $j\in [k]$ let $h_{j-1,j}$ denote the continuous surjective homomorphism $\tran(\ns_j)^0\to \tran(\ns_{j-1})^0$ defined as in Lemma \ref{lem:hdefn}, and for each $i\in [k]$ let $h_i=h_{i,i+1}\co\cdots\co h_{k-1,k}:\tran(\ns)^0\to \tran(\ns_i)^0$. Fix an arbitrary element $x\in \ns$, and for each $i\in [k]$ let $\Gamma_i=\tran_i(\ns)^0\cap \stab_{\tran(\ns)^0}(x)$. Then for each $i$, the following statements hold:
\begin{enumerate}
\item $\tran_i(\ns)^0$ acts transitively on each fibre of $\pi_{i-1}:\ns\to \ns_{i-1}$.
\item $\ker(h_i)= \Gamma_i\cdot \tran_{i+1}(\ns)^0$.
\item $\tran_i(\ns)^0\;/\; \big(\Gamma_i\cdot \tran_{i+1}(\ns)^0\big)\;\cong\; \ab_i$.
\end{enumerate}
\end{proposition}
\begin{proof}
We prove $(i)$ by downward induction on $i$. For the case $i=k$, since each fibre of $\pi_{k-1}$ is a homogeneous space of $\ab_k$, it suffices to show that $\tran_k(\ns)^0=\tau(\ab_k)$. This is immediate from the following equality, which actually holds for a general nilspace and is established in \cite[Lemma 3.2.37]{Cand:Notes1}:
\begin{equation}\label{eq:abstktk}
\tran_k(\ns)=\tau(\ab_k).
\end{equation}
Suppose now that $i\leq k-1$ and that the statement holds for $i+1$, and note that any fibre of $\pi_{i-1}=\pi_{i-1,i}\co \pi_i$ has the form $\pi_i^{-1}(\pi_{i-1,i}^{-1}(y))$ for some $y\in \ns_{i-1}$. Given $w,w'$ in such a fibre, we have $\pi_i(w),\pi_i(w')\in \pi_{i-1,i}^{-1}(y)$. Since $h_i(\tran_i(\ns)^0)= \tran_i(\ns_i)^0=\tau(\ab_i)$, we have that this acts transitively on $\pi_{i-1,i}^{-1}(y)$ and so there is $\alpha\in \tran_i(\ns)^0$ such that $\pi_i(\alpha(w))=\pi_i(w')$. Then, since $\tran_i(\ns)^0$ contains $\tran_{i+1}(\ns)^0$, by induction it acts transitively on each $\pi_i$-fibre and so there exists $\alpha'\in \tran_i(\ns)^0$ such that $\alpha'\alpha(w)=w'$. This proves statement $(i)$.\\
\indent To see statement $(ii)$, note that $\tran_{i+1}(\ns)^0\leq \ker(h_i)$, since $h_i$ sends $\tran_{i+1}(\ns)^0$ into the trivial group $\tran_{i+1}(\ns_i)^0$. We also have that $\Gamma_i\subset \ker(h_i)$, because by \eqref{eq:abstktk} we have $h_i(\Gamma_i)\subset \tau(\ab_i)$ but also each element of $h_i(\Gamma_i)$ fixes $\pi_i(x)$, so must be the trivial translation in $\tau(\ab_i)$. It follows that $\Gamma_i\cdot \tran_{i+1}(\ns)^0 \subset \ker(h_i)$. To see the opposite inclusion, note that for $\alpha\in \ker(h_i)$ we have from the definitions that $\pi_i(\alpha(x))=h_i(\alpha)(\pi_i(x))=\pi_i(x)$. Therefore, by statement $(i)$ there exists $\alpha'\in \tran_{i+1}(\ns)^0$ such that $\alpha'\alpha(x)=x$, so $\alpha'\alpha\in \Gamma_i$, and so $\alpha\in \tran_{i+1}(\ns)^0\cdot \Gamma_i$.\\
\indent Statement $(iii)$ now follows from $(ii)$, the first isomorphism theorem for $h_i$, and \eqref{eq:abstktk} for $k=i$.
\end{proof}
|
\section{Introduction}
\setlength{\epigraphwidth}{0.8\textwidth}
\epigraph{\textit{We cannot hope to prove that every definition, every
symbol, every abbreviation that we introduce is free from
potential ambiguities, that it does not bring about the
possibility of a contradiction that might not otherwise have been
present.}}{N. Bourbaki~\cite{Bourbaki1949}}
\epigraph{\textit{There is an error, I can confess now. Some 40 years
after the paper was published, the logician Robert M. Solovay from
the University of California sent me a communication pointing out
the error. I thought: “How could it be?” I started to look at it
and finally I realized} [$\ldots$]}{John F. Nash Jr.~\cite{nash}}
\epigraph{\textit{Mathematics arises from all sorts of application or
insights but in the end must always consist of proofs}, [but]
\textit{although a real proof is not simply a formalized document
but a sequence of ideas and insights}, [a] \textit{real proof is
not something just probably correct.}}{Saunders Mac Lane
~\cite{maclane}}
What constitutes a \emph{real proof} is a question at the origin of
mathematical logic. In effect, a real proof is one that can be reduced
to the use of only a few accepted `ideal' principles such as the
axioms for a set theory like ZFC. And yet certain ideal principles
are far from self-evident as Euclid's axiomatic method would require
them to be. Proof theory was conceived by Hilbert with the program to
further ``recognize the non-contradictory character of all the usual
[ideal] mathematical methods without exceptions''. Around 1960, these
concerns were addressed for the theory of arithmetic and analysis in
the so called \emph{modified} Hilbert program using the early models
of \emph{computation}---proof theory was also pivotal for the
development of computer science (Hilbert's Entscheidungsproblem).
Applying mathematical rigor to formal proofs as the object of study
brought an answer to the question of what a real proof is: a formal
proof can be given semantics in terms of Gödel's system T and
Spector's bar recursion, thereby eliminating logic in favor of pure
computation. However, these early models of computation that were used
to provide the answer to the consistency question, although satisfying
in terms of precision, are cumbersome to use in practice.
Firstly, it is far from clear why the old computational
interpretations are the right ones, for it is often hard to
distinguish them (bar recursion) from brute force search
procedures. We would like to understand the computational answer to
the main consistency questions in terms of modern and more finely
grained computing abstractions, such as the ones developed over the
course of the past four decades in the theory of programming
languages---for research on (natural) models of computation surely did
not end with the invention of the Turing machine and recursive
function theory.
Secondly, the cumbersome machinery, although ingenious, makes it
difficult to address the next level of research questions. Once that
we have the answer to what a real proof \emph{is}, we need to know
what constitutes the \emph{essential data} of a proof---curiously,
this question of finding criteria of greatest simplicity for proofs
was already listed as \nth{24} in Hilbert's famous list of open
problems, but being premature was not included among the ones finally
published~\cite{Thiele2003}.
The title of this chapter refers to \emph{proof unwinding}, the
pioneering research program from the 1950's of Georg Kreisel
\cite{Luckhardt1996}, who started to use the computational approach,
not for foundational purposes, but to extract numerical content from
actual mathematical arguments. We aim at the \emph{working
mathematician}, a term used by Bourbaki~\cite{Bourbaki1949} who
meant by it a researcher with a pragmatic attitude toward
foundations. The time is ripe for a
leap forward, both in foundations and unwinding applications. The
present chapter has as goal to propose bringing proof unwinding on a
par with the latest computing abstractions from the theory of
programming languages, with the ambition to turn such
streamlined proof theoretic methods into a toolbox readily used by the
working mathematician, rather than the rare specialist in proof theory
as it has been the case up to now.
\section{New Unwinding Toolbox}
\label{subsec:toolbox}
Conducted with the goal:
\begin{quote}
\textit{``To determine the constructive (recursive) content or the
constructive equivalent of the non-constructive concepts and
theorems used in mathematics''}~\cite{Kreisel1958},
\end{quote}
Kreisel's research program applied the proof theory of the day, namely
Hilbert's $\epsilon$-substitution method, Herbrand's theorem, and the
no-counterexample interpretation, combined with then brand new theory
of recursive functions, to extract new bounds and algorithms from
prima facie ineffective proofs. But, even in the hands of masters,
the early unwinding methodology was apparently difficult to apply, if
one is to judge from the lapses of time in between applications:
Littlewood's theorem by Kreisel in 1951~\cite{Kreisel1951}, Artin's
proof of Hilbert's \nth{17} problem by Kreisel first in 1957 and again
in 1978~\cite{Delzell1996}), the Thue-Siegel-Roth theorem by Kreisel
and Macintyre in 1982 and Luckhardt in 1989
\cite{KreiselM1982,Luckhardt1989}, Van der Waerden's theorem by Girard
in 1987~\cite{Girard1987}. The unwinding methods are so complex that
there are even doubts cast on some of the results by authorities in
proof theory~\cite{Feferman1996}.
However, there is a more recent application of unwinding to functional
analysis in the \emph{proof mining} program of Kohlenbach
\cite{Kohlenbach2005}. This very successful unwinding program has at
its methodological core the classic unwinding approach using
Kolmogorov's double-negation translation (1929) and Gödel's functional
`Dialectica' interpretation (1941).
In parallel, in constructive mathematics, there have been equally
significant results in the program of \emph{constructive analysis}
\cite{bishopbridges} and \emph{constructive algebra}
\cite{mines88,Lombardi}, although these are primarily theory
reconstruction programs and rely little on direct application of proof
theoretic methods to unwind ineffective proofs.
But, both mining and research in constructive mathematics have not
sought to reap the benefits of notable proof theoretic advances
directly inspired by the theory of programming languages. This theory,
a continuation of the work on the early models of computation, has
arrived at highly abstract notions for structuring programs. We shall
now describe the proof theoretic state-of-the-art for three such proof
unwinding techniques. This new methodology will be referred to as
the \emph{unwinding toolbox}.
\subsection{Computational Side-Effects}
\label{methodology:control}
The first of these methods concerns \emph{computational
side-effect}. Namely, since the work of Griffin~\cite{Griffin1990},
it has been known that the principle of proof by contradiction can be
interpreted by a programming language mechanism (a computational
side-effect) for \emph{control operators}. Although, in absence of
mathematical axioms additional to the reductio-ad-absurdum principle,
control operators provide not much more than a very elegant way to
obtain Herbrand's theorem (an important very early result on classical
first-order logic from 1930), in presence of additional axioms like
induction or choice, when Herbrand's theorem no longer holds, one
begins to get new results. For instance, by the use of computational
side-effects, in set theory, Krivine has managed to extend Cohen's
forcing method from the usual sets of conditions to realizability
algebras~\cite{Krivine}.
However, the promise that control operators can turn every proof by
contradiction into an effective one is a mirage: there are classically
provable formulas whose effective proof would allow to decide the
Halting problem. Whether an ineffectively proved formula can be
unwound, in general needs to be considered on a case-by-case
basis. Nevertheless, there \emph{are} whole \emph{classes} of formulas
which we know can be unwound upfront, like the class of
$\Pi^0_2$-formulas. \emph{Delimiting} control operators only to
formulas in these classes allows to get a new constructive logic. This
logic still respects the existence property, characteristic of
intuitionistic logic that is at the bases of current constructive
mathematics, but the obtained new constructive logic manages to prove
intuitionistically non-provable principles.
For instance, Herbelin~\cite{HerbelinMP} showed that Markov's
principle (MP), \[\neg\neg\exists x A_0(x)\to \exists x A_0(x),\] where
$x\in\mathbb{N}$ and $A_0$ is quantifier-free, an axiom crucial for
constructive proofs of completeness of first-order logic
\cite{IlikThesis}, can be interpreted with the help of a computational
side-effect known as \emph{(delimited) exceptions}. The author further
showed~\cite{Ilik2010} that the double negation shift principle (DNS),
\[\forall x \neg\neg A(x) \to \neg\neg\forall x A(x),\] where
$x\in\mathbb{N}$, a principle crucial for the interpretation of the
classical axiom of choice---and that has only been interpreted before
by the \emph{generally}-recursive schema of bar recursion---can be
interpreted computationally by a generalization of the exceptions
effect to so called delimited continuations, or \emph{delimited
control operators}. The key
observation from these results is that---when
delimited---computational side-effects like control operators can be
used to unwind ineffective proofs and at the same time not run into
non-decidability problems. The newly obtained logics are intermediate
logics, in between classical and intuitionistic logic, and take the
best of both worlds.
These results are controversial from the point of view of the orthodox
constructive mathematician who is used to intuitionistic logic as
first codified by Heyting's analysis of Brouwer's work in
intuitionistic mathematics. Namely, the only previous computational
interpretation of MP were either trivial (as given by Gödel's
functional interpretation) or proceeded by unbounded search. As for
DNS, the computational interpretation was only given by the
generally-recursive bar recursion schema, whose termination must be
ensured by Brouwer's bar induction or continuity principle. As
unbounded/general recursion can lead to an inconsistent formal system,
intuitionists have been understandably wary of accepting these
principle. By replacing the mentioned computational interpretations by
computationally meaningful realizers, we not only propose to
intuitionists to reconsider the constructivity of principles like MP
and DNS, but we are re-establishing the link between modern proof
theory and one of the offspring of Hilbert's proof theory, the theory
of programming languages.
A further principle interpreted in this way, in a joint work of Nakata
and the author~\cite{IlikN2014}, was the open induction principle,
\[\forall \alpha(\forall \beta<\alpha(\beta\in U) \to \alpha\in U) \to
\forall \alpha(\alpha\in U),\]
where $\alpha,\beta$ range over Cantor space and $U$ is open. This
principle is the only known equivalent form of the axiom of choice
that is stable under double-negation translation (even if we replace
Cantor space by Baire space). The principle is also interesting for
combinatorics, where it leads to a direct version of Nash-Williams'
proof of Kruskal's tree theorem~\cite{Veldman2001}, as well as in
algebra where it is used to replace Zorn's lemma~\cite{Schuster}.
We finally mention a last result~\cite{Ilik2014}, still in review, on
the nature of the programming-language inspired proof rules. It
concerns higher-type primitive recursion---Gödel's system T---versus
general recursion---Spector's bar recursion. Namely, already in 1979,
Schwichtenberg has shown that bar recursion of type 0 and 1 does not
allow to define functions beyond system T
\cite{Schwichtenberg1979}. Since a previous analysis of Kreisel
\cite{Kreisel1959} shows that these types are sufficient for all
practical purposes (realizing $\Sigma^0_2$-theorems), it follows that
we know for a long time that we should not need bar recursion for the
computational interpretation of ideal proof principles. What we
proposed is how to circumvent bar recursion and generate System T
terms directly, using delimited control operators as an intermediary
step. This also shows that the extensions of system T with
computational side-effects are in fact conservative extensions. In
order to establish this fact, we relied on \emph{partial evaluation},
the second set of techniques of our unwinding toolbox that we explain
in the following subsection.
\subsection{Partial Evaluation and Formalization}
\label{methodology:interpret}
The second programming languages method that we intend to employ for
proof unwinding concerns formalization of proofs in proof assistant
software and, more specifically, the use of formalized
\emph{(type-directed) partial evaluators}.
The topic of partial evaluators came up in our previous research on
constructive versions of completeness theorems
\cite{IlikThesis}. These logical theorems establish the adequacy of a
given formal system to encode actual mathematical proofs. As it turns
out, and thanks to initial work on the link between normalization proofs
and completeness of intuitionistic logic for Kripke models
\cite{CoquandD1997}, the computational content of proofs of
intuitionistic completeness can be expressed by type-directed partial
evaluation algorithms~\cite{Danvy1996}. Having a rich theory of such
algorithms in the theory of programming languages, allowed to cover
cases of constructive completeness proofs that were beyond the
previous state-of-the art in proof theory. More precisely, we now know
how to partially evaluate (i.e. show constructively completeness for)
not only classical logic~\cite{Ilik2010}, intuitionistic logic with
disjunction~\cite{Ilik2013}, but also simultaneous presence of
delimited control operators and higher-type primitive recursion
\cite{IlikTDPE,Ilik2014} (the second citation is still in review).
The development of these logical meta-theorems was conducted formally,
in the Coq and Agda proof assistants. Since the formalized proofs are
constructive, they can be used to \emph{compute a proof
transformation} for every actual formalized argument. What this
allows is to perform unwinding of actual mathematical proofs more
directly, by pushing the complexity of doing a manual double-negation
transformation (like done in the classic unwinding approach and used,
for instance, by Kohlenbach in his program of proof mining in
analysis) into the realizability model, that is, into the reduction
mechanism of the proof assistant used.
Proof assistants are most well known for their use in the full
formalization of complex proofs, such as the four-color theorem
\cite{Gonthier2008}, the Kepler conjecture~\cite{Hales2012}, or the
Feit-Thompson theorem~\cite{Gonthier2013}. However, as far as proof
unwinding is concerned, one can in general \emph{avoid} needing a
\emph{fully} formalized version of an actual mathematical proof. It
suffices to notice that lemmas that have a computationally irrelevant
form, such as $\Pi^0_1$, can be simply assumed without proof. A more
refined analysis of computational relevance of formulas can be found
in~\cite{SchwichtenbergW2012} where the classes of so called definite
and goal formulas are isolated. This allows to greatly decrease the
burden of formalizing i.e. we are only dealing with \emph{partial
formalization} which nonetheless contains as much of algorithmic
content as a full formalization.
The important lesson that we learned from partial evaluation is that
proofs need not be interpreted uniformly, by `oracles' such as bar
recursion that work uniformly (for example the realization of DNS by
bar recursion is agnostic of the concrete formula $A$ in the instance
of DNS). Rather, it is possible to specialize (i.e. partially
evaluate) proofs, even if they are highly ineffective, and, when one
in addition uses a proof assistant like Coq, the specialization of the
(partially) formalized theorem can become automatic. This is one of
the principal advantages of our toolbox over the old toolbox built on
Herbrand's theorem, $\epsilon$-substitution, double-negation- and
A-translation, and functional interpretation: while unwinding, the
mathematician can concentrate on the essential parts of a proof rather
than get lost in manual proof transformations.
\subsection{Type Isomorphisms}\label{methodology:typeiso}
The final third method of our unwinding toolbox concerns \emph{type
isomorphisms}. Mathematically, this notion is the same as the one of
\emph{constructive cardinality} of sets~\cite{mines88}, saying not
only that sets are of the same size, but moreover that they have
indistinguishable structure. In programming languages theory, the
notion allows to generalize the notion of type assigned to a program,
which allows to test more easily if a programs conforms to a formal
specification~\cite{Rittri}.
The link that brings us to the study of type isomorphisms is
Tarski's high-school algebra problem~\cite{burris04}. This basic
question, asking whether the system of eleven arithmetic equation
taught in high-school suffices to derive all the true equations
between \emph{exponential} polynomials, had taken some time to be
answered in mathematical logic. It turned out that the high-school
system is not complete, as shown by a counter-example of Wilkie in
1981~\cite{wilkie00}, a true statement which is not derivable by only
using the eleven equations. Gurevič further showed that the system
cannot be completed by any additional finite set of axioms
\cite{gurevic90}.
Now, by the Curry-Howard correspondence, formulas of intuitionistic
logic can be seen as types (conjunction correspond to products,
disjunction to coproducts, and implication to exponentials) and proofs
of formulas can be seen as computer programs of the corresponding
type. Following the correspondence, one gets a notion of strong
equivalence, or formula isomorphism, from isomorphism of types. A new
correspondence is thus obtained: the language of formulas is the same
as the language of exponential polynomials---and, moreover---formula
isomorphism generalizes equality of exponential polynomials in the
standard model of positive natural numbers, that is
\[
A \cong B \text{ implies } \mathbb{N}^+ \vDash A=B.
\]
The link that one establishes in this way allows to use the rich
theory on exponential polynomials to obtain proof theoretic
results. For instance, Fiore, Di Cosmo, and Balat, showed that the
non-finite-axiomatizability result of Gurevič also hold for the theory
of type isomorphism~\cite{fiore06}.
Using results of Richardson, Martin, Levitz, Wilkie, Macintyre, Henson
and Gurevič, the PI proved that although not finitely axiomatizable,
type i.e. formula isomorphism is recursively axiomatizable and
moreover decidable~\cite{sumaxioms}. The value of this unexpected
positive result is still somewhat limited because the existence of a
\emph{practical} decision algorithm for type isomorphism is open.
Nevertheless, even if the meta-theory of type isomorphism has
remaining open questions to be resolved, applications to proof theory
are well under way. Recently, the PI proposed a pseudo-normal form of
types~\cite{explog}, inspired by the decomposition of the exponential
function in exponential fields~\cite{hardy}, called the exp-log normal
form, that allows to decompose the axioms of the notoriously non-local
theory of $\beta\eta$-equality for the lambda calculus with coproduct
type. This equality can be seen as the essence of \emph{identity of
proofs} for intuitionistic propositional logic with disjunction. An
extension to the first-order case has also been proposed in a joint
work with Brock-Nannestad~\cite{highschool}, where the normal form
appears to produce the first arithmetical hierarchy for formulas of
intuitionistic logic that copes with both quantifiers equally well;
the only previously known hierarchy, the one of Burr~\cite{Burr2000},
covers well only the universal quantifier. This has been a long
standing open problem for constructive logic, although for classical
logic an arithmetical hierarchy exists since the 1930s.
\section{Perspectives}
Today, a paradigm change in proof unwinding is possible, thanks to the
notions from contemporary programming languages theory comprising our
\nameref{subsec:toolbox}. These long-evolved techniques provide
proof-theoretic simplifications of the order that makes them more
accessible even to non-specialists in proof theory.
The overall goal of this undertaking would be to exploit the full potential of
the novel toolbox and apply it, beyond logic itself, to proofs of
landmark results in number theory, combinatorics, and homotopy
theory. In parallel, it would be necessary to address the foundations
of unwinding i.e. tackle long-standing open questions in the
foundations of constructive mathematics such as identity of proofs and
simplified computational interpretations of semi-intuitionistic
principles. We have thus two sets of objectives, work on applications
and work on foundations.
\subsection*{Objective I --- Applications of Proof Unwinding}
\label{obj:unwinding} The first set of objectives concerns
applications to areas that are important for the `working
mathematician', that is, analysis, number theory, and combinatorics,
as well as an application to unwinding incompleteness theorems in
logic. Objective I would be achieved by tackling three more specific
objectives, called \emph{perspectives}: \nameref{task:mining},
\nameref{task:unwinding}, and \nameref{task:incompleteness}.
\subsection*{Objective II --- Foundations of Proof Unwinding}
\label{obj:nextgen} The second set of objectives concerns work on
foundations of constructive mathematics that are both necessary to
guarantee the soundness of applying unwinding and as an update to the
current foundational theories. The two more specific objectives, or
perspectives, to be tackled are: \nameref{task:hott} and
\nameref{task:metatheory}.
\vfill
The immediate effects of the project would be, on the one hand, to show
that our new proof theoretic methods can be used by the working
mathematician to extract numerical bounds and algorithms from prima
facie ineffective proofs in analysis, number theory, combinatorics,
homotopy theory, and logic, and, on the other hand, to update the
current foundational theories of constructive mathematics with the
powerful computing abstractions that computational side-effects,
partial evaluators and type isomorphisms represent.
In the longer term, we can hope to see the streamlined proof unwinding
methodology becoming an important toolbox across mathematics. We can
also expect to see a synergy of the objectives. For instance, not only
would unwinding efforts across mathematics become possible (Objective
I), but, as the new constructive foundations (Objective II) get
adopted in the community working on proof assistant systems, proof
analysis and development would eventually be carried out even more
efficiently with the help of a proof assistant.
The approach to fulfilling the two objectives would be through carrying
out the five perspectives described in this section. We shall explain each
one of the tasks in the context of its proper state-of-the-art,
objectives, methodology, and feasibility.
\subsection{Perspective 1: Unwinding in Analysis Revisited}
\label{task:mining}
Analysis is essentially the only area of mainstream mathematics to
have benefited from direct application of proof unwinding techniques.
In approximation theory, by using proof theory, Kohlenbach and his
collaborators have managed to obtain explicit moduli of uniqueness,
significantly better than previous ones, for best Chebyshev
approximation, as well as to obtain a first effective rate of strong
unicity in the case of best approximation for the $L_1$-norm
\cite{Kohlenbach1993}. How this works is that first logical
meta-theorems are established~\cite{Kohlenbach2005}, which are on one
hand general enough to be applicable as analytic theorems, and on the
other hand specific enough to enter in a class of statements, such as
the $\Pi^0_2$-class of the arithmetical hierarchy, for which we know
by proof theory that explicit functions or existence witnesses can be
extracted. Moreover, such general logical meta-theorems are not only
good for extracting numerical data from concrete proofs, but also for
analyzing whether a known analytic theorem has optimal form. For
instance, in the fixed point theory for functions of contractive type,
one does not only get effective quantitative forms of theorems, but
one can often also relax the compactness assumption for the metric
space.
Why, then, when Kohlenbach's proof mining approach is already
successful, do we propose a perspective on proof unwinding of analysis? There
are two reasons. The first one is methodological: our form of
unwinding has not been applied outside of logic, and proof mining
provides the perfect test bed to make it grow up in the `real
world'. Second, even if we cannot pretend to analytic skills of the
level of the ones present in mining, we do believe that the general
logical meta-theorems can be unwound in a simpler way; this could lead
to better extracted bounds even if we use the exact same analytic
machinery as in mining.
To explain the difference and simplification mentioned, we briefly
explain how the meta-theorems are established right now. The core is
to show that in classical logic, and in presence of additional axioms
for induction and choice, like the weak Kőnig's lemma, one can turn
the $\forall\exists$ quantifier combination from
$\forall x \exists y A_0(x,y)$, where $A_0$ is a quantifier-free
formula, into an explicit recursive function $f$ such that
$\forall x A_0(x, f(x))$. One can further extend this to formulas
beyond the strict class $\Pi^0_2$ and allow for instance any number of
additional hypotheses of form $\Pi^0_1$. But, to obtain the recursive
functions $f$, which, as explained before in the section
\nameref{methodology:control}, needs an a priori generally recursive
definition schema, one first has to transform by the double-negation
translation all proofs of the original proof system (Peano arithmetic
+ axiom of choice) into proofs of a (semi-)intuitionistic system. This
is a \emph{non-local} transformation of proofs, and in particular the
meaning of formulas can be changed by the transformation (hence the
restriction to the $\Pi^0_2$-class of formulas). Once
a~(semi-)intuitionistic proof is obtained, one can use Gödel's
functional interpretation to obtain a higher-type primitive recursive
function, possibly also needing Spector's generally-recursive schema
of bar recursion. Actually, more redefined versions of the Dialectica
interpretation (monotone and bounded variants) and of bar recursion
are used in practice.
Now, what our approach offers is first to push the technical
complexity of the double negation translation into the realizability
model based on computational side-effects (ex. control
operators). Since these notions have a well-studied operational
semantics, one can perform a more direct reduction of a proof to a
program or a more direct reading off of witnesses (numeric
bounds). With the additional help of a proof assistant like Coq, this
can be further automatized.
In addition, thanks to the reasons already explained in section
\nameref {methodology:control}, our unwinding method makes it likely
that in fact a pure system T witnessing terms can be extracted from
any concrete proof, circumventing bar-recursion-like schemata
altogether.
A third, orthogonal, improvement to the extraction process will be
offered by use of richer data-types for extracted programs and
bounds. Traditionally, one only uses the `negative' function and
product types. Although these can encode `positive' types (for
instance, sum types $\rho+\sigma$ can be encoded by
$(\mathbb{N}\to \rho)\times(\mathbb{N}\to \sigma)$), encoding leads to
an increase of the \emph{degree} of the type. Simpler and more natural
realizers can thus be extracted from disjunction and other inductively
defined positive predicates.
\subsubsection*{Feasibility for Perspective 1} We will need to cope with
semi-intuitionistic principles that we have not treated before,
notably the weak Kőnig's lemma and the independence of premise
schema. For these, we plan to use \nameref{methodology:control}, like
we have done previously for the open induction principle: the fan
theorem, a positive version of the weak Kőnig lemma, is implied by
open induction. At the level of realizers, it will be necessary to use
{\nameref{methodology:typeiso}} to handle extensionality.
The risk for handling the logical part (meta-theorems) is moderate,
hence it is possible to propose this for a subject of a PhD thesis. As
for obtaining better bounds that the ones already obtained in proof
mining, the risk is higher; in fact, it would be a success even if we
manage to obtain the same bounds, since this would mean that our
toolbox is ready to be used in the following, Perspective 2.
\subsection{Perspective 2: Unwinding in Number Theory and Combinatorics}
\label{task:unwinding}
In this task, we should bring in our \nameref{subsec:toolbox} to bear on
highly non-effective proofs from number theory and combinatorics. The
concrete goals will be to unwind landmark proofs in these areas, but
what we see as equally important is to arrive at a situation where a
sufficiently interesting intersection of proof theory and the
application domain area is recognized. This kind of objective is only
possible through a combination of expertise, and for its carrying out,
it would be appropriate to engage two post-doctoral researchers, one in
each application domain.
In number theory, we would intend to unwind Thue-Siegel-Roth's theorem on
Diophantine approximations. Saying that an algebraic irrational number
has only finitely many exceptionally good rational number
approximations, this $\Sigma_2$ statement has first been tackled upon
by Kreisel and Macintyre~\cite{KreiselM1982} using technology for
obtaining Herbrand terms. However the combinatorial explosion arising
from use of Herbrand's theorem apparently did not allow to obtain
useful bounds on the number of rational approximations, and only
Luckhardt~\cite{Luckhardt1989} managed to limit the growth of Herbrand
terms in order to obtain such a bound. This bound is essentially the
same as the one obtained by Bombieri and van der Poorten
\cite{Bombieri}.
In this case, even more advanced existing technology like Gödel's
functional interpretation has not been applied. We suspect this is the
case because, in order to apply it, one would first need to perform a
double-negation translation of an actual proof of Thue-Siegel-Roth
into a semi-intuitionistic theory---something possible to do in
principle, but given the sophistication of the original proof, its
translation would be an order of magnitude more complex. We propose
thus to treat it directly using our approach with computational
side-effects, i.e. without a preliminary double-negation translation
followed by a functional interpretation. Technically, our approach can
be seen as a version of the so called modified realizability technique
but where the language of realizers is enriched to contain delimited
control operators.
In combinatorics, we would intend to unwind Szemerédi's theorem saying that
every subset of the natural numbers with positive upper density
contains arithmetic progressions of arbitrary length. Conjectured by
Erdős and Turán in 1936, this statement was only proved by Szemerédi
in 1975 by an ingenious and complex combinatorial argument. In 1977,
Furstenberg provided a proof using ergodic theory. The interest in
giving a better proof of this theorem is still ongoing, and
applications include for instance the recent work of Green and Tao on
arbitrary long arithmetic progressions in the prime numbers
\cite{green2008primes}.
We first intend to address an important special case of the theorem,
the van der Waerden theorem, saying that if we use a finite number of
colors to color the natural numbers, then there is at least one
color containing arbitrarily long arithmetic progressions. The
current upper bound for van der Waerden's number $W(k,r)$, where $r$
is the number of colors and $k$ is the requested length of an
arithmetic progression, was obtained via Szemerédi's theorem and is
due to Gowers~\cite{Gowers}. What is intriguing in this subject is
that the upper bounds appear to be heavy overestimates: for
instance, the bound for $W(3,3)$ is of the order of $10^{14616}$,
while the exact value is $27$.
Girard has previously analyzed Furstenberg and Weiss's proof of van
der Waerden's theorem using cut elimination~\cite{Girard1987}. But,
the bound that he arrived at was essentially the same upper bound
obtained by Furstenberg and Weiss~\cite{Macintyre2005}. We could
attack the problem by using our modified realizability based on
computational side-effects and attempt to partially evaluate the
latest available proofs for Szemerédi's and van der Waerden's
theorem---that would avoid the need for having a fully formalized proof
on hand.
\subsubsection*{Feasibility for Perspective 2} Although the \emph{statements} of
the mentioned theorems in number theory and combinatorics are
arithmetical, their \emph{proofs} are not arithmetical. The risk for
the objectives of this task is to cope with the highly non-effective
nature of proofs, as well as their considerable complexity (see
Szemerédi's diagram of lemmas from his proof in
\cite{Szemeredi}). After all, proofs of the corresponding theorems
have brought Fields medals to both Roth and Szemerédi. The main proof
theoretic question is which kind of ideal principles are at the core
of arguments and can we provide a direct constructive justification
for them. Sometimes, as in the case of Kruskal's theorem, another
statement of Ramsey theory, the link to the open induction principle
(analyzed previously in joint work with Nakata~\cite{IlikN2014}) turns
out to be direct~\cite{Veldman2001}.
We intend to use proof assistant technology and partial formalization
to cope with the complexity of proofs. Concerning mathematical risk,
given a choice of motivated post-doctoral researchers to work on this
topics, I would say that it is medium. Work on ergodic theory done in
the previous \nameref{task:mining} would serve as preparatory work and
would help to further mitigate the risk. This task demands more
resources than the other ones.
\subsection{Perspective 3: Unwinding Incompleteness Theorems}
\label{task:incompleteness}
A statement is said to be independent from a theory if it can neither
be proved nor disproved from the axioms of the theory. The
\emph{incompleteness} phenomenon is the fact that for \emph{any}
theory, under the assumption that it is consistent, there exist
statements that are independent of the theory. One might wonder what
is the nature of these statements, and whether they are relevant in
practice. Indeed, the first such statement discovered by Gödel in 1931
has an `artificial' flavor since it encodes the Epimenides' liar
paradox. But, later, natural examples from Ramsey theory have been
found, first by Paris and Harrington~\cite{paris1977mathematical}, and
include important results like Kruskal's tree theorem. Finding
concrete mathematical incompleteness statements is nowadays a fruitful
field of research led by Friedman~\cite{Friedman}.
However, what we find especially interesting is the \emph{limit} at
which a statement starts to become independent from a theory. This
phenomenon, called \emph{phase transition} by analogy to
thermodynamics, happens when the \emph{provability} of a theorem,
taking a rational number as parameter, depends on the \emph{value} of
this rational parameter. For instance, a parametrized version of
Kruskal's theorem can be provable in Peano arithmetic (PA) below a
certain value of the parameter, and becomes independent above that
value---this is in fact a real number, often found by use of analytic
combinatorics, and provides a measure of the strength of the axiom
system. Phase transitions are a research program led by Weiermann
\cite{Weiermann2005}.
In this task, we propose to develop a method for unwinding
incompleteness theorems and phase transition phenomena based on
programming language theory. The idea is that an incompleteness
theorem,
$\text{PA}\not\vdash\bot \rightarrow \text{PA}\not\vdash
\text{Con(PA)}$,
saying that no consistent formal system (in this case, Peano
arithmetic (PA)) can prove its own consistency, can be rephrased
positively as
$\text{PA}\vdash \text{Con(PA)} \rightarrow \text{PA}\vdash\bot$.
Translated in programming languages terms,
$\text{PA}\vdash \text{Con(PA)}$ expresses the possibility of writing
an interpreter for Gödel's system T inside system T itself---that is,
a \emph{self-interpreter}. Self-interpreters have not only been
studied in programming languages theory, but they are a standard way
to bootstrap a compiler for a programming language.
Nevertheless, self-interpreters are usually written for a
Turing-complete languages like Scheme and ones without a strong typing
discipline. If one adds a type system on top of Scheme one can
retrieve system T in its $\lambda$-calculus formulation. There are
recent intriguing results on typed self-interpreters. Brown and
Palsberg have recently constructed the first \emph{typed}
self-interpreter~\cite{brown2015self}; their target was Girard's
system U, and this is still `acceptable', since system U is known to
be inconsistent as a logical system. But, their latest result concerns
Girard's system F$_\omega$~\cite{brown2016self}, which is a
higher-order logic and considered to be consistent.
In this task, we would first investigate whether it is possible to
construct a self-interpreter for system T. For the purpose of the
paper~\cite{Ilik2014}, we have already developed a formally verified
interpreter for system T$^+$ inside Martin-Löf type theory. Since this
type theory has a realizability model based on system T, one comes
close to having a self-interpreter. We would also have to study the
recent results of Brown and Palsberg, and attempt to retrieve their
result for system F$_\omega$ in system T.
\subsubsection*{Feasibility for Perspective 3} The proposed methodology involves a
frontal attack on consistency of PA. Although the risk is high, the
fact that the prior works of Brown and Palsberg, and the PI, all
involve formally checked proofs, gives us some confidence. If our
effort succeeds, the gain one may have will be equally high as the
risk. But, even if it turns out to be impossible to write a typed
self-interpreter for T, we can aim to obtain solid results on
interpreting Weiermann's phase transition, and hence characterizing
the strength of a formal system, in terms of notions that are equally
natural from the point of view of computation as analytic
combinatorics are.
\subsection{Perspective 4: Identity of Proofs and Homotopy Type Theory}
\label{task:hott}
Formal proofs are combinatorial objects meant to encode a fully
correct mathematical argument, going down to the smallest details, but
that makes it difficult to spot the most essential parts of an
arguments. Curiously, finding ``criteria of simplicity, or proof of
the greatest simplicity of certain proofs'' was already part of
Hilbert's program, who even planned to include it as the \nth{24} in
his famous list of open problems~\cite{Thiele2003}. In particular,
Hilbert asked for a procedure to decide when two given proof are
essentially the same. This problem known as \emph{identity of proofs}
is still open~\cite{Dosen2003}.
In this task, we would start by tackling the identity of proofs for
constructive logic, before proceeding to a vast generalization of it,
the computational interpretation of Voevodsky's univalence axiom in
homotopy type theory~\cite{Coquand2014} in the case when the
underlying definitional equality has been strengthened to decide
identity of proofs i.e. to convertibility modulo isomorphism for
dependent types.
For intuitionistic \emph{propositional} logic, the difficulty of
deciding identity of proofs comes from the simultaneous presence of
disjunction and implication. Nevertheless, if we follow the analogy
between formulas, types, and exponential polynomials, explained in
section \nameref{methodology:typeiso}, we can re-express the problem
precisely as that of the effective decidability of the
$\beta\eta$-equational theory for the lambda calculus with
coproducts. We have recently proposed a first step in this direction
by showing how to decompose the equational theory for terms, by the
use of the exp-log normal form for types in order to enlarge the
$\beta\eta$-congruence classes of terms~\cite{explog} (in review).
This exp-log normal form of types is extensible to the
\emph{first-order} case, when the quantifiers $\forall$ and $\exists$
are also present. Namely, recent work with Brock-Nannestad \cite{highschool}
shows that it leads to an intuitionistic arithmetical hierarchy, a
classification of formulas that was elusive for intuitionistic logic,
even though it has existed for classical logic since the 1930's where
it is at the basis of results like the completeness theorem.
A further question is whether we can make the technique work for
dependent types, an extension of the first-order case. Martin-Löf Type
Theory has dependent types which allow it to have special treatment of
equality. Basic equality between elements $a,a'$ of a type $A$ is
encoded by the identity type for $A$, $\Id_A(a,a')$. Identity of
proofs in this context means extending the notion of definitional
(computational) equality to cope with $\eta$-equality for coproducts
(and other inductive types).
Pursuing generalization even further, we can talk about \emph{identity
between proofs of identity}, $\Id_{\Id_A}(p,p')$, that, in turn,
endows ever type $A$ with the structure of a groupoid. Iterating this
construction, $\Id_{\Id_{\Id_{\cdots\Id_A}}}$, allows to show that
every type $A$ is in fact endowed with the structure of an
$\infty$-groupoid~\cite{HofmannStreicher}. Using Grothendieck's
correspondence between $\infty$-groupoids and homotopy types has led
Voevodsky to give a homotopy theoretic interpretation of type theory
in his model based on simplicial sets
\cite{kapulkin2012simplicial}. This model satisfies Voevodsky's
\emph{univalence axiom}, generalizing identity of proofs, and
specializing to: equality at the level of propositions, bijection at
the level of sets, categorical equivalence at the level of groupoids,
etc. Adding this axiom on top of Martin-Löf's type theory produces
homotopy type theory, which is a logical system formalizing the
\emph{univalent foundations} of mathematics~\cite{hottbook}.
What we propose to do is to build the convertibility of proof terms
modulo type isomorphism into the definitional equality of Martin-Löf
and homotopy type theory. An identity type then gets to cover equality
between terms of a whole class of isomorphic types instead of only one
type. We hope that in this way it will be possible to strengthen the
notion of \emph{transport of structures} and to show that important
special cases of the univalence axioms satisfy a simple computational
interpretation. The only existing computational interpretation of
homotopy type theory appears in the effective version of the
simplicial set model~\cite{bezem2014model} and works for the standard
(restricted) notion of identity type.
\subsubsection*{Feasibility for Perspective 4} The univalence axiom is known to
imply a form of full functional extensionality in type theory. Given
that extensionality of functions in general is undecidable, the risk
for extending the computational interpretation for the univalence
axiom defined over the notion of identity types strengthened to work
modulo isomorphism is high. Nevertheless, by strengthening the
underlying definitional equality of the type theory, we hope to
diminish the need for resorting to full functional extensionality and
even address important special cases of univalence more simply than
before.
As concerns the identity of proofs for the propositional and
first-order case, based on our preliminary investigations of this
area, we would say that the risk is moderate.
\subsection{Perspective 5: A Next Generation of Constructive
Foundations}
\label{task:metatheory}
This task would serve as an umbrella for more specific but important
problems that need to be tackled in the foundations of constructive
mathematics, as well as an umbrella collecting the foundational
implications of the previous four tasks of this chapter.
For instance, we already know that axioms which are independent of
intuitionistic logic like double negation shift can be safely added to
intuitionistic systems, but we have to establish the outer limits of
the potential given by \nameref{methodology:control}. We need to
develop direct computational interpretations of principles arising
from the work in constructive reverse mathematics, such as the
equivalent forms of the open induction principle~\cite{veldmanarxiv},
the extension of our work~\cite{IlikN2014} to Baire space, and novel
versions of Markov's principle~\cite{fujiwara}.
Another important topic will be to provide a direct constructive proof
of Goodman's theorem. This theorem says that the axiom of choice
presents a conservative extension of higher type Heyting arithmetic
concerning arithmetical formulas; for the meta-theory of constructive
mathematics, it plays the role that Hilbert's $\epsilon$-elimination
theorems play for the proof theory of classical logic. There has
recently been renewed interest about this old result of Goodman by
other researchers as well
\cite{kohlenbach1999note,coquand2013goodman}.
A third important topic will be to find practical decision algorithms
for type isomorphism. As explain in the section
\nameref{methodology:typeiso}, although a decidability result holds
for type isomorphisms~\cite{sumaxioms}, thanks to prior work of
Richardson~\cite{richardson69} and Macintyre~\cite{macintyre81}, it is
not clear at the moment whether a (practical) decision algorithm can
be constructed. Arriving at such an algorithm would not only be useful
for proof theory, but also for symbolic computation.
Finally, we would like to interact with the researchers working on proof
assistant systems like Coq. The logical cores of proof assistants are
lagging behind contemporary proof theory. For instance, program
extraction from proofs in a state-of-the-art proof assistants such as
Coq relies on the simplest possible realizability interpretation, the
so called modified realizability interpretation of
Kreisel. Integrating the techniques from the \nameref{subsec:toolbox}
would be beneficial for users of proof assistants because it would
allow for easier formalization of many apparently ineffective proofs.
\subsubsection{Feasibility for Perspective 5}
The main challenge for this task is that, when we are interpreting
semi-intuitionistic principles, we are working at the limit of
computability: our realizability models for the classical axiom of
choice refute the internal (formal) version of Church's thesis, but
the external weak Church's rule still holds~\cite{Ilik2014} (in
review). It is thus hard to predict upfront how far the outer limits
of constructive foundations can be extended. As concerns Goodman's
theorem, we think the risk involved is not very high, since after all
this result has been establish by non-direct methods. Finally, the
risk on finding a practical algorithm deciding type isomorphism is
hard to estimate; but even if we manage to find ones that only work
for special cases, the benefits could spread also beyond proof theory.
\bibliographystyle{alpha}
|
\section{Introduction}
International migration is a major contributor to population change, but is hard to project, making proper quantification of uncertainty especially important.
Existing global models for migration are well-calibrated marginally,
i.e.~for individual countries \citep{azose2015}, but typically rely on an unrealistic modeling assumption that forecast errors are uncorrelated across countries.
If correlations exist, but are not modeled, the resulting projections may
still be well calibrated for countries individually, but can under- or overestimate variance in projections of migration for regions that span multiple countries.
We present a method for estimating a correlation matrix from a small number
of data points that uses informative priors, shrinking elements of the correlation matrix which we expect \emph{a priori} to be small.
In applying this method to migration, we choose priors based on empirical evidence of non-zero correlations among classes of countries which are ``close'' to one another according to a variety of distance covariates.
Our method improves projections of migration for regional aggregates while mitigating the issue of spurious correlations that arises from trying to estimate a large correlation matrix based on many short time series.
\subsection{Illustrative example}
In this section we focus on six selected countries---Estonia, Latvia, Lithuania, South Africa, Zimbabwe, and Zambia---to highlight the need for regularization of the correlation matrix.
Migration rates in Estonia, Latvia, and Lithuania over the period from 1950 to 2010 look quite similar (top row of Figure \ref{fig:sixCountries}.)
All three countries share a spike in out-migration during the 1990--1995 time period, which appears as a large negative forecast error in a first-order
autoregressive (AR(1)) model.
This sudden jump in out-migration among the Baltic states shares a common cause, namely the fall of the Soviet Union, which both induced westward migration and prompted many ethnic Russians to return to Russia \citep{fassmann1994, okolski1998}
Meanwhile, several countries in Southern Africa also experienced big shifts in migration rates during the 1990--1995 time period (bottom row of Figure \ref{fig:sixCountries}.)
From 1990 to 1995, South Africa received substantially more in-migration than it had in previous decades, while Zimbabwe and Zambia both switched from being net receivers of migrants to net senders.
For these three countries, at least some of the change in migration was due to political shifts related to the end of South Africa's apartheid policy.
For example, the number of legal entrants to South Africa who overstayed their visas grew dramatically during the 1990s, with many such entrants coming from other countries of the Southern African Development Community \citep{crush1999}.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{sixCountries_rates.pdf}
\caption[Net migration rates for six countries]{Net migration rates (net annual migrants per thousand individuals) for six countries.}\label{fig:sixCountries}
\end{figure}
Because all six countries experienced pronounced changes in migration rates during the same time period, the usual Pearson estimates of the correlation in forecast errors are relatively large for these six countries (left panel of Figure \ref{fig:heatmaps}.)
Knowledge of world affairs, however, suggests that some of these correlations may be spurious.
There are plausible explanations for the correlations within the three Baltic nations and within the Southern African nations, but the cross-regional correlations are suspect.
In fact, the cross-regional correlations seem to have arisen largely from a coincidental synchrony in the timing of disparate geopolitical events, and do not represent correlations that we would expect to continue to exist in future migration data.
Our method is designed to shrink these seemingly spurious cross-regional correlations, producing the estimated correlation matrix shown in the right panel of Figure \ref{fig:heatmaps}.
Cross-regional correlations decrease substantially in magnitude, while correlations within regions remain largely unchanged.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{heatmaps}
\caption[Estimated correlations among forecast errors for migration]{Estimated correlations among forecast errors for migration. Left panel shows Pearson correlation estimates. Right panel shows our regularized estimates.}\label{fig:heatmaps}
\end{figure}
\subsection{Background}
Country-specific projections of international migration are an important input in policy-making decisions \citep{bijak2007, brownBean2012}.
Projected migration figures are commonly used in long-term planning of social welfare programs \citep{socialSecurity2013, wright2010}.
However, projection of migration is difficult---\cite{bijak2010} describe migration as ``barely predictable''---and global modeling of migration remains somewhat rudimentary.
The United Nations Population Division produces global projections of fertility, mortality, and migration for all countries \citep{wpp2012}.
For most countries, the 2012 revision of the World Population Prospects (WPP) deterministically projects net migration to persist at current levels until 2050 and decline linearly thereafter.
To produce fully probabilistic population projections, one must incorporate probabilistic projections of fertility and mortality with a global probabilistic model of migration. It follows from the demographic balancing equation that the contribution of migration to population change is given by \emph{net} migration (that is, in-migration minus out-migration.) Probabilistic models exist for both net migration \citep{azose2015,azose2016} and in- and out-migration separately \citep{wisniowski2015}. Both of these models are Bayesian hierarchical autoregressive models which treat forecast errors in migration as independent across countries, conditional on model parameters.
This leads to projections that are well calibrated for individual countries,
but may not be for multi-country aggregates.
Our method aims to relax this independence assumption.
It is worth noting that a strong correlation in migration rates themselves need not translate to a strong correlation in forecast errors.
For example, from 1960 through 2000, Mexico was consistently either the largest or second-largest source of migration flows to the US, with nearly 5 million Mexicans migrating to the US during the 1990's \citep{abel2013}.
While we estimate that net migration rates for the USA and Mexico have a correlation of -0.56 based on quinquennial WPP data from 1950-2010, we estimate a correlation in forecast errors of only -0.07.
That is, most of the relationship between the USA and Mexico is already captured by the autoregressive model parameters, and the ``random'' components of migration rates for the two countries are nearly independent conditional on the AR(1) model.
In this high-dimensional setting with short time series,
the empirical correlation matrix is a poor estimator, in that it can include many spuriously large estimated correlations.
Our goal is to use regularization to improve an empirical correlation matrix for forecast errors in migration.
There is a large body of literature on regularized estimation of covariance matrices, with applications in genomics, image processing, and finance, among other fields \citep{fan2014}.
The novelty of our method is that it allows the incorporation of
available prior information in an easily interpretable way.
Existing covariance estimators based on penalized likelihood maximization are typically maximum {\em a posteriori} (MAP) estimates under some prior belief about covariance, but these formulations are not well suited to specifying beliefs directly about elements of the correlation matrix. Perhaps the most similar method to ours is that of \cite{bien2011}, which allows informative priors on elements of the covariance matrix rather than the correlation matrix. Their method is not directly applicable to our setting, as our goal is to augment existing marginal variances with a suitable correlation structure. Other proposed MAP estimators include the graphical lasso \citep{friedman2008}, which can be used to place an informative prior on the inverse covariance, and the method of \cite{chi2014}, which penalizes covariance estimates that have very large or very small eigenvalues. An extreme example is given by \cite{chaudhuri2007}, who provide a method for covariance estimation in the presence of known zeroes. \cite{zhangZou2014} propose a variant on penalized likelihood maximization that replaces the negative log-likelihood with a simpler loss function.
A related class of covariance estimators relies on shrinkage of an empirical covariance matrix towards a simpler estimator, typically trading some bias for lower mean squared error \citep{ledoit2003,ledoit2004,ledoit2012}.
A strength of these methods is that so long as the empirical covariance matrix is positive semi-definite and the shrinkage target is positive definite, a linear combination of the two will naturally be positive definite. Applying a shrinkage method to the migration setting would be difficult, as the elements we would like to penalize do not define a positive definite shrinkage target.
A form of regularization that is straightforward to implement is applying thresholding directly to elements of a covariance or correlation matrix \citep{bickelLevina2008Thresholding, elKaroui2008}; these authors show that a hard-thresholded covariance matrix is consistent in operator norm. Generalized thresholding \citep{antoniadis2001}, developed in the context of wavelet applications, provides a class of related regularized estimators. A key difficulty with such estimators is that care must be taken to ensure that the resulting estimator is positive definite. In some problems, this can be handled by selecting a thresholding constant from an appropriate range \citep{fanLiaoMincheva2013}. Unfortunately, such an approach is not easily adapted to our problem. The structure of the elements we wish to penalize is such that we can tolerate only a small amount of shrinkage of all penalized elements before our estimated correlation matrix loses positive definiteness.
One fully Bayesian treatment is proposed by \cite{liechty2004}, who include substantive prior information by specifying clusters of correlations which they expect to be similar. This is unfortunately unsuitable to our setting, since geographical and cultural proximity can give rise to either positive or negative correlations. \cite{huang2013} describe a computationally attractive \emph{non-informative} prior on covariances which does not easily extend to the informative priors we would like to include. Other fully Bayesian treatments are given by \cite{barnard2000}, who propose a prior on the correlation matrix which is either marginally or jointly uniform, and \cite{leonard1992} and \cite{deng2013}, who propose Bayesian estimation of the logarithm of the covariance matrix, which is unfortunately hard to interpret.
In scenarios where there is a natural ordering to the variables, it is often reasonable to make the assumption that large values of $|i-j|$ imply near independence or conditional independence. When this is the case, one can regularize by banding or tapering of the covariance or inverse covariance matrix \citep{bickel2008, fan2007, furrer2007, chen2013, levina2008}. These approaches are not suitable to our problem, as there is no natural ordering of countries.
Good overviews of other methods in covariance estimation are given by \cite{fan2015} and \cite{pourahmadi2011}.
\section{Methods}
We start with an established, well-calibrated autoregressive model on net migration rates for all countries \citep{azose2015}.
This model has the form:
\begin{align}
\boldsymbol{g}_t - \boldsymbol{\mu} &= \textrm{diag}(\boldsymbol{\phi}) (\boldsymbol{g}_{t-1} - \boldsymbol{\mu}) + \boldsymbol{\varepsilon}_t,\\
\boldsymbol{\varepsilon}_t &\overset{iid}{\sim} \mathcal{N}_C\left( \boldsymbol{0}, \textrm{diag}(\boldsymbol{\sigma}) \cdot I_C \cdot \textrm{diag}(\boldsymbol{\sigma}) \right) , \\
\phi_c &\overset{\textrm{iid}}{\sim} U(0,1) , \\
\mu_c &\overset{\textrm{iid}}{\sim} N(\lambda, \tau^2) , \\
\sigma^2_c &\overset{\textrm{iid}}{\sim} IG(a,b).
\end{align}
Notationally, $\boldsymbol{g}_t$ is a length-$C$ vector of net migration rates for all countries during the time period from $t$ to $t+1$, where $C$ is the number
of countries analyzed.
The quantities $\boldsymbol{\mu}$, $\boldsymbol{\phi}$, and $\boldsymbol{\sigma}$ are vectors of model parameters, and $\boldsymbol{0}$ is a length-$C$ vector of zeroes.
(We have omitted here the specifics of hyperpriors on $a$, $b$, $\lambda$, and $\tau$, which Azose and Raftery selected to reflect the ranges of plausible values.)
Notably, forecast errors in their model are treated as independent,
conditional on the model's other parameters.
Our method augments this model with an estimated correlation structure.
Although this paper focuses on the migration context, the same technique
could be applied to probabilistic models of other demographic indicators.
From this point forward, we refer to Azose and Raftery's model as the Bayesian Hierarchical Model with Independent Forecast Errors (BHM+IFE).
In principle, the methodology we describe here provides a means of estimating a correlation matrix to be adjoined to any probabilistic model with conditionally independent forecast errors.
The outline of our procedure for estimating a correlation matrix is as follows:
\begin{enumerate}
\item From the BHM+IFE model, draw a posterior sample of $m$ realizations of model parameters, $\boldsymbol{\mu}^{(1)}$, $\boldsymbol{\phi}^{(1)}$, $\boldsymbol{\sigma}^{(1)}$, \ldots, $\boldsymbol{\mu}^{(m)}$, $\boldsymbol{\phi}^{(m)}$, $\boldsymbol{\sigma}^{(m)}$.
\item Convert the estimated forecast errors from the posterior sample of model parameters to a single empirical correlation matrix, $\tilde R$.
\item Combine the empirical correlation matrix with informative priors on correlations to obtain a maximum \emph{a posteriori} (MAP) correlation estimate, $\hat R$.
\end{enumerate}
This procedure can be viewed as performing a single step of the Monte Carlo EM (MCEM) algorithm \citep{wei1990}.
The posterior sampling in stage 1 can be performed using any reasonable sampling procedure. In practice, we performed our posterior sampling with a combination of Gibbs sampling and Metropolis-Hastings steps.
In the following sections, we first discuss the details of obtaining an MAP estimator (Section \ref{sec:MAP}) and then the question of what to use for an empirical correlation matrix (Section \ref{sec:RTilde}).
This is followed by an algorithm for computing the MAP estimator (Section \ref{sec:minimization}), and finally discussion of a criterion for selecting a regularization parameter (Section \ref{sec:lambda}).
\subsection{MAP correlation estimate}\label{sec:MAP}
Our goal is to estimate the correlation structure, $R$, of forecast errors, $\boldsymbol{\varepsilon}_t$. We assume a model of the form
\begin{equation}\label{eqn:normalLikelihood}
\boldsymbol{\varepsilon}_t \overset{iid}{\sim} \mathcal{N}_C\left( \boldsymbol{0}, \Sigma \right),
\end{equation}
where the variance matrix, $\Sigma$, decomposes into standard deviations, $\boldsymbol{\sigma}$, and a correlation matrix, $R$, as $\Sigma = \textrm{diag}(\boldsymbol{\sigma}) \cdot R \cdot \textrm{diag}(\boldsymbol{\sigma})$.
To determine a MAP estimator for $R$, we express the posterior distribution for $R$ as a product of likelihood and prior.
\subsubsection{Data Likelihood}
Equation (\ref{eqn:normalLikelihood}) implies a likelihood function for $R$ of the form
\begin{equation}
p(\boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{T-1} | R, \boldsymbol{\sigma}) \propto_R \det(R)^{-(T-1)/2} \exp \left(
-\frac{1}{2} \sum_{t=1}^{T-1} \boldsymbol{\varepsilon}'_t \textrm{diag}(\boldsymbol{\sigma})^{-1} R^{-1} \textrm{diag}(\boldsymbol{\sigma})^{-1} \boldsymbol{\varepsilon}_t
\right),
\end{equation}
restricted to the space $\Omega$ of valid correlation matrices (i.e.~positive semi-definite matrices with ones on the diagonal.) Matrix trace identities simplify this likelihood to
\begin{equation}
p(\boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{T-1} | R, \boldsymbol{\sigma}) \propto_R \det(R)^{-(T-1)/2} \exp \left(
-\frac{1}{2} \textrm{tr}(R^{-1} \tilde R)
\right),
\end{equation}
where
\begin{equation}\label{eqn:RTilde}
\tilde R := \frac{1}{T-1} \sum_{t=1}^{T-1} \textrm{diag}(\boldsymbol{\sigma})^{-1} \boldsymbol{\varepsilon}_t \boldsymbol{\varepsilon}'_t \textrm{diag}(\boldsymbol{\sigma})^{-1}.
\end{equation}
The evidence from the data is encapsulated in $\tilde R$, which is something akin to an empirical correlation matrix.
Note that $\tilde R$ would be a sufficient statistic for $R$ if the $\boldsymbol{\varepsilon}_t$'s and $\boldsymbol{\sigma}$ were known.
In fact neither the $\boldsymbol{\varepsilon}_t$'s nor $\boldsymbol{\sigma}$ are known, and $\tilde R$ must be replaced with a sensible estimate in order to proceed.
Details of the estimation of $\tilde R$ are given in Section \ref{sec:RTilde}.
\subsubsection{Prior}
Our choice of prior distribution on $R$ is motivated by a desire to incorporate informative prior beliefs about which country pairs are likely to be nearly uncorrelated.
As such, we choose a prior of the form
\begin{equation}
\pi(R) \propto_R \prod_{0 \leq i < j \leq C} \exp(-\lambda P_{ij} |R_{ij}|),
\end{equation}
again restricted to $\Omega$. The matrix $P$ with entries $P_{ij}$ is a penalty matrix that encodes the extent to which we believe that countries $i$ and $j$ may be correlated. In our application to migration, we constrain all the entries
in $P$ to be equal to 0 or 1, although in general $P$ may be allowed to have arbitrary non-negative entries. The parameter $\lambda$ is an overall regularization parameter that encodes how strongly we want to penalize correlations.
The key benefit of this prior is its ease of interpretability. Setting $P_{ij}=1$ expresses a belief that $R_{ij}$ should be close to zero, with the strength of that belief controlled by $\lambda$. Setting $P_{ij}=0$ implies that all values of $R_{ij}$ are equally believable, \emph{a priori}. Other penalized likelihood estimators have been proposed, corresponding to MAP estimators under implied priors on precision \citep{friedman2008}, covariance \citep{bien2011}, or eigenvalues of the covariance matrix \citep{chi2014}.
None of these allow one to specify prior beliefs about correlations directly.
Note that under this specification, the prior distribution of the correlation
$R_{ij}$ is either uniform or truncated Laplace conditional on the rest of the correlation matrix, but marginal distributions will not be uniform or double exponential.
Although it is possible to specify a marginally uniform prior on all elements of the correlation matrix \citep{barnard2000}, we know of no way to specify a distribution that is marginally uniform for some elements and marginally peaked at zero for others.
Because the prior density is a product of Laplace densities on correlations, we will refer to our eventual correlation estimator as the LPoC (Lapalace Prior on Correlations) estimator. Augmenting the BHM+IFE with the LPoC correlation estimate produces the BHM+LPoC model.
\subsubsection{Posterior}
Combining the likelihood and prior, we obtain the log posterior distribution for $R$, equal to
\begin{equation}
\log p(R | \boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{T-1} , \boldsymbol{\sigma}) = - \frac{T-1}{2}\log \det(R) - \frac{T-1}{2} \textrm{tr}(R^{-1} \tilde R) - \frac{\lambda}{2}\|P*R\|_1 + c(\boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{T-1}, \boldsymbol{\sigma}),
\end{equation}
where $*$ denotes elementwise matrix multiplication, and $\| \cdot \|_1$ gives the sum of the absolute value of the elements of a matrix.
Thus, finding the MAP estimator for $R$ is equivalent to solving the minimization problem
\begin{equation}\label{eqn:minProblem}
\textrm{Minimize}_{R \in \Omega} \left\{
\log \det(R) + \textrm{tr}(R^{-1} \tilde R) + \frac{1}{T-1} \lambda \cdot \|P*R\|_1.
\right\}
\end{equation}
Algorithmic details of a numerical solution are given in Section \ref{sec:minimization}.
Note that if the penalty parameter, $\lambda$, is zero, then this minimization problem yields the maximum likelihood estimator (MLE) of R conditional on $\boldsymbol{\sigma}$. So long as $\tilde R$ is itself positive definite, this MLE is just $\tilde R$, the empirical correlation matrix. Similarly, if $\lambda$ is held fixed as $T$ grows, the penalty term in (\ref{eqn:minProblem}) goes to zero and the LPoC estimator converges to $\tilde R$. Since $\tilde R$ is consistent for $R$, the LPoC estimator is also consistent.
\subsection{Estimating $\tilde R$}\label{sec:RTilde}
Since the forecast errors and model parameters of the BHM+IFE model are unknown, we do not have access to the true value of $\tilde R$.
Instead we use an estimate of $\tilde R$.
For practical reasons, we would prefer to have $\tilde R$ itself be a valid correlation matrix so that (\ref{eqn:minProblem}) will have a known analytic solution in the limiting scenarios where $T$ grows or $\lambda$ goes to zero.
Accordingly, we might choose an estimator $\tilde R^{basic}$ with elements defined by
\begin{equation}
\tilde R_{ij}^{basic} := \frac{\sum_{t=1}^{T-1} \hat\varepsilon_{i,t} \hat\varepsilon_{j,t}}
{\sqrt{\sum_{t=1}^{T-1} \hat\varepsilon_{i,t}^2} \sqrt{\sum_{t=1}^{T-1} \hat\varepsilon_{j,t}^2}},
\end{equation}
where $\hat \boldsymbol{\varepsilon}_t$ is the posterior mean of $\boldsymbol{\varepsilon}_t$ from the BHM+IFE model.
This estimate, $\tilde R^{basic}$, is the MLE for estimating the correlation matrix of a multivariate normal random variable with mean known to be zero and unknown marginal variance terms.
By construction, $\tilde R^{basic}$ is guaranteed to be positive semi-definite and to have ones on the diagonal.
However, in our application, $\tilde R^{basic}$ is low rank, since $T$ is small relative to the dimension of the matrix. For computational reasons, we would prefer to have a strictly positive definite matrix, so we estimate $\tilde R$ by
\begin{equation}
\tilde R^{PD} = 0.99 \cdot \tilde R^{basic} + 0.01 \cdot I_C.
\end{equation}
This change can be viewed as augmenting our estimates of $\boldsymbol{\varepsilon}_t$ with a small amount of additional uncorrelated data.
\subsection{Solving the minimization problem}\label{sec:minimization}
We apply a majorize-minimize algorithm similar to that used by \cite{bien2011} to the minimization problem in (\ref{eqn:minProblem}).
The function being minimized over is the sum of a convex and a concave component.
The majorize-minimize algorithm repeatedly iterates through the following steps:
\begin{enumerate}
\item Replace the concave component with its tangent plane to obtain a fully convex function.
\item Find the global minimum of the convex function from Step 1.
\item Update the estimate of the tangent plane.
\end{enumerate}
Notationally, we label our starting point for this algorithm as $R_0$ and subsequent iterations of this majorize-minimize algorithm are denoted with subscripts $R_1, R_2, \ldots$.
In (\ref{eqn:minProblem}), the concave component is $\log \det (R)$, which we replace with the tangent plane $\log \det R_i + \textrm{tr}(R_i^{-1}(R-R_i))$.
After simplifying and removing terms which are constant in $R$, the convex minimization problem in the $i$th iteration of the algorithm is
\begin{equation}\label{eqn:innerMinimization}
\textrm{Minimize}_{R \in \Omega} \left\{ \textrm{tr} (R_i^{-1} R)+ \textrm{tr}(R^{-1} \tilde R) + \lambda \| P * R\|_1 \right\}.
\end{equation}
Now all of the terms the objective function in (\ref{eqn:innerMinimization}) are convex, and all but $\lambda \|P * R\|_1$ are differentiable, so we can apply the generalized gradient descent algorithm \citep{beck2009}.
Each generalized gradient descent step takes the form
\begin{equation}\label{eqn:equivalent}
R_{new}=\textrm{argmin}_{\omega \in \Omega} \{(2t)^{-1} \| \omega - (R_{current}- t(R_i^{-1} -R_{current}^{-1} \tilde R R_{current}^{-1}))\|_F^2 + \frac{\lambda}{T-1} \|P * \omega \|_1 \}.
\end{equation}
If the restriction to $\Omega$ were not present, this problem would have a simple analytic solution, given by
\begin{equation}\label{eqn:proposedStep}
R_{new}=\mathcal{S}\left(R_{current}- t(R_i^{-1} -R_{current}^{-1} \tilde R R_{current}^{-1}), \frac{\lambda}{T-1}tP\right),
\end{equation}
where $\mathcal{S}$ is the element-wise soft-thresholding operator defined by
\begin{equation}
\mathcal{S}(X,\alpha)_{ij}=\textrm{sign}(X_{ij}) \cdot (|X_{ij}|-\alpha_{ij}) \cdot \mathbbm{1}(|X_{ij}|>\alpha_{ij}).
\end{equation}
(This move is actually restricted to the off-diagonal elements only, as the diagonal elements of a correlation matrix are constrained to equal 1.)
Thus, if there were no positive definiteness constraint, each update step would consist of a gradient descent step according to the gradient of the differentiable component followed by soft-thresholding the result.
Although we do have to satisfy a positive definiteness constraint, we can start by trying the update step in (\ref{eqn:proposedStep}).
If this update results in a valid correlation matrix, then that matrix is our solution to (\ref{eqn:equivalent}), and we replace $R_{current}$ with $R_{new}$.
However, sometimes the soft-thresholded gradient step results in a matrix that is not positive definite.
In that case, it is possible to appeal to a slower, iterative solution to (\ref{eqn:equivalent}).
One such solution is given by \cite{cui2016}.
In practice, as long as we are looking for a solution in the interior of $\Omega$, it is good enough to simply reduce step size rather than appealing to the relatively costly iterative algorithm whenever the generalized gradient descent suggestion lies outside of $\Omega$.
Step size selection has a large impact on performance and convergence of this algorithm. Details of step size selection are discussed in Appendix \ref{sec:stepSize}.
\subsection{Selecting the regularization parameter $\lambda$}\label{sec:lambda}
Although the penalty matrix $P$ can be selected on the basis of world knowledge, we are less likely to have genuine prior beliefs about the value of the regularization parameter $\lambda$.
Accordingly, we need some procedure for selecting a value for $\lambda$.
In regularization problems, it is common to select the regularization parameter via cross-validation \citep{bien2011,chi2014,huang2006}.
This approach is too computationally intensive to be feasible for our application.
Among shrinkage estimators, it is common to choose the amount of shrinkage in order to minimize an expected loss function \citep{james1961,ledoit2003}.
However, no suitable analytic result exists that allows us to approximately minimize expected loss in our scenario.
Consequently, we developed a heuristic criterion that selects $\lambda$ in a way that aligns with the goal of our regularization process.
Our method's intent is to shrink the magnitude of penalized elements of the correlation matrix while leaving unpenalized elements more or less unchanged.
In practice, although we succeed at bringing penalized elements towards zero, this shrinkage usually comes at the cost of inflating other elements.
We have observed that this inflation tends to grow more pronounced as $\lambda$ grows.
For very large values of $\lambda$, our estimated correlation matrix may shrink nearly all penalized entries to zero at the expense of inflating a few elements (both penalized and unpenalized) to nearly $\pm 1$. This is not a desirable outcome.
Although it may seem counterintuitive at first,
the observed inflation is not an artifact of a coding error or poor convergence of our algorithm. A simple reproducible example of inflation in a $3 \times 3$ matrix is provided in Appendix \ref{sec:inflation}. In this low-dimensional setting, standard numerical optimization routines agree with the results from our code and both display inflation of unpenalized elements.
Our criterion for selecting $\lambda$ compares the off-diagonal elements of $\tilde R$ and $\hat R(\lambda)$. We choose the value of $\lambda$ which maximizes the difference between average shrinkage and average inflation. Formally, our criterion is defined by
\begin{equation}
k(\tilde R, \lambda) =
\underset{i,j \textrm{ s.t. } |\hat R(\lambda)_{ij}| < |\tilde R_{ij}|}{\textrm{mean}} \left(|\tilde R_{ij}| - |\hat R(\lambda)_{ij}|\right)
-
\underset{i,j \textrm{ s.t. }|\hat R(\lambda)_{ij}| > |\tilde R_{ij}|}{\textrm{mean}} \left(|\hat R(\lambda)_{ij}| - |\tilde R_{ij}|\right).
\end{equation}
Large positive values of $k$ are desirable, as they correspond to values of $\lambda$ for which we induce a lot of shrinkage and not much inflation.
\section{Results}
In this section,
we first report results from applying our method to global migration data in Section \ref{sec:application}.
Section \ref{sec:simStudy} then provides a simulation study which demonstrates that our method outperforms Pearson correlations and the Ledoit-Wolf shrinkage estimator \citep{ledoit2003} in the scenario where the penalty matrix $P$ is appropriate to the true correlation structure.
\subsection{Application to migration}\label{sec:application}
\subsubsection{Data}
We use data on net migration from the 2012 revision of the World Population Prospects (WPP) \cite{wpp2012}.
The WPP contains estimates of net migration for all countries in five-year time periods from 1950 until 2010, a total of 12 time periods. We compute the net migration rate $g_{c,t}$ as the net number of migrants in country $c$ over the five year period starting at time $t$, divided by thousands of individuals in country $c$ at time $t$.
Because we want to express prior beliefs as a function of distance covariates, we restrict the set of modeled countries to the 191-country overlap between the WPP 2012 and the set of countries included in CEPII's GeoDist database, a database of bilateral distance covariates defined on pairs of countries \citep{mayer2011}.
\subsubsection{Selection of $P$}
Our estimation technique requires that we choose a penalty matrix, $P$, that reflects our prior beliefs about which country pairs are likely to be correlated. Although it would be possible to elicit expert opinion about each of the roughly 18,000 country pairs, we instead choose a $P$ that can be characterized in terms of just a few covariates. Our matrix $P$ penalizes a pair of countries if \emph{none} of the following conditions is met:
\begin{enumerate}
\item The two countries are contiguous.
\item The two countries' most important cities are located less than 3000 km apart.
\item The two countries are in the same region according to the United Nations
Population Division's division of the world into 22 regions, based on
both geographical contiguity and cultural affinity \citep{wpp2012}.
\item The two countries are currently in a colonial relationship.
\end{enumerate}
This definition of $P$ is in line with migration theory, which suggests that migrant flows are more likely when monetary and social costs of movement are low \citep{harris&1970, lee1966, sjaastad1962, stark&1985}, as will be the case with countries which are geographically proximate or share administrative ties.
This definition penalizes 85\% of country pairs, leaving 15\% unpenalized. The average country is considered to be ``close'' to 29 other countries, and ``distant'' from the remaining 161.
In selecting these conditions, we examined nine candidate distance covariates.
The first eight such covariates come from CEPII's GeoDist database \citep{mayer2011}, while the ninth is derived from the United Nations division
into 22 regions.
The left column of Table \ref{tab:ksTest} gives the complete list of covariates considered. As an empirical basis for determining which criteria to include in defining our penalty matrix, we examined the elements of the sample correlation matrix for all pairs of countries meeting each criterion. Using a Kolmogorov-Smirnov test, we tested whether the distribution of these sample correlations was different from the distribution of elements of the sample correlation matrix under a null hypothesis of uncorrelated errors. The right column of Table \ref{tab:ksTest} shows the $p$-values from these Kolmogorov-Smirnov tests. Our definition of the penalty matrix $P$ includes all covariates with a $p$-value less than 0.05.
\begin{table}
\caption[Results of Kolmogorov-Smirnov test for correlations]{Results of Kolmogorov-Smirnov test that empirical correlations are significantly different from the distribution of elements of a sample correlation matrix when the true error structure is uncorrelated. $p$-values lower than 0.05 are bolded.}\label{tab:ksTest}
\begin{center}
\begin{tabular}{rc}
\hline
Covariate & $p$-value\\
\hline
Contiguous & \textbf{0.019} \\
Common language (official) & 0.23 \\
Common language (spoken by 9\% of pop.) & 0.58\\
Geodesic distance less than 3000 km & \textbf{0.0003}\\
Colonial relationship after 1945 & 0.57\\
Common colonizer after 1945 & 0.11\\
Current colonial relationship & \textbf{0.035}\\
Ever had a colonial link & 0.36\\
Same UN Region & \textbf{0.036}\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Selection of the regularization parameter, $\lambda$}
We computed values of $\hat R(\lambda)$ for all values of $\lambda$ from 0 to 3 in increments of 0.1.
Figure \ref{fig:shrinkageCriterion} shows the value of $k(\tilde R, \lambda)$ over a range of $\lambda$ values.
We found that $k(\tilde R, \lambda)$ peaked at $\lambda=0.6$, where we find average shrinkage of 0.13 compared with average inflation of 0.07.
Increasing $\lambda$ from 0.6 to 0.7 induces additional shrinkage, but at the cost of greatly inflating some correlations.
Accordingly, we choose $\hat R(0.6)$ as our estimate of $R$.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{shrinkageCriterion}
\caption[Regularization criterion as a function of $\lambda$]{Regularization criterion, $k(\tilde R, \lambda)$ as a function of $\lambda$. The regularization criterion is the difference between the average shrinkage among shrunk elements of $\hat R(\lambda)$ and average inflation among inflated elements.}\label{fig:shrinkageCriterion}
\end{figure}
Figure \ref{fig:splitPlot} shows the impact of regularization on the correlation matrix.
Among penalized elements (top panel), we see significant shrinkage towards zero, although many penalized elements remain large in magnitude, even after regularization.
The bottom panel shows the unpenalized elements of the correlation matrix before regularization (solid curve) and after (dashed curve).
On average we induce some shrinkage in the unpenalized elements, but the distribution is largely unchanged.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{splitPlot}
\caption[Elements of the correlation matrix before and after regularization]{Comparison of elements of the correlation matrix before regularization (solid curves) and after (dashed curves). Top panel shows penalized elements; bottom panel shows unpenalized elements.}\label{fig:splitPlot}
\end{figure}
\subsubsection{Projection and evaluation}
We augment the BHM+IFE model with the LPoC estimate $\hat R(0.6)$ to produce probabilistic projections of migration for any collection of countries.
Figure \ref{fig:continents} contains medians and 80\% prediction intervals of projected migration for all continents.
In Africa, negative correlations narrow our projections.
In Europe, positive correlations cause forecasts to widen.
For the other continents, we see little change in projected migration.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{continents_ggplot}
\caption[Probabilistic projections of net migration for continents]{Medians and 80\% prediction intervals for net migration among continents. Projections from the Bayesian hierarchical model with independent forecast errors (BHM+IFE) are given in red. Projections using our estimated correlation matrix (BHM+LPoC) are in blue. Overlap is in purple.}\label{fig:continents}
\end{figure}
For evaluation, we compare true migration rates for regional aggregates in 1995--2010 with projections of the same regional aggregates based only on migration data from 1950--1995.
This procedure entails re-estimation of the BHM+IFE model using only the 1950--1995 data, followed by construction of an empirical correlation matrix, selection of $\lambda$, and extraction of $\hat R(\lambda)$.
We compare the performance of the BHM+IFE model on regional aggregates to a model using the same sampled values of $\boldsymbol{\mu}$, $\boldsymbol{\phi}$, and $\boldsymbol{\sigma}$, but augmented with $\hat R(\lambda)$.
As an evaluation metric, we use the negatively oriented continuous ranked probability score (CRPS) \citep{hersbach2000,gneiting2007}.
The CRPS compares the cumulative distribution function, $F$, of a probabilistic forecast to an observation, $x$, and is defined by
\begin{equation}
\mathrm{CRPS}(F,x) = \int_{- \infty}^{\infty} (F(y)- \mathbbm{1}\{y \geq x\})^2\ dy.
\end{equation}
In our application the two probabilistic forecasts under consideration have the same mean as one other, by design.
One approximate way of looking at CRPS in this setting is that when $g_{c,t}$ is close to the mean of the forecast, we reward $F$ for having low variance; when $g_{c,t}$ is far from the mean, we reward $F$ for having high variance.
Table \ref{tab:CRPS} gives CRPS for projections of aggregate migration for the six continents. Our model improves the quality of projections in Africa and Europe, while projections for the other four continents are more or less unchanged.
Figure \ref{fig:fourRegions} illustrates the change in projections of net migration in 1995--2010 for four subregions of Africa and Europe.
Projections from the BHM+IFE model are in red; projections from BHM+LPoC are in blue.
Our method narrows prediction intervals in Eastern and Western Africa, bringing the width of the 80\% prediction intervals more into line with the range of observed variability.
In both regions, true migration rates for the projected period stayed within our narrower intervals.
In contrast, our method widens projections in Northern and Western Europe, where the 80\% intervals from the BHM+IFE model either miss or nearly miss capturing some of the observed data points.
\begin{table}[ht]
\begin{center}
\caption[Continuous ranked probability score for continental migration projections]{Continuous ranked probability score for all continents evaluated on projections of 1995-2010, where lower is better.
Left column: Projections based on the Bayesian hierarchical model with independent correlation structure (BHM+IFE). Right column: Projections based on the Bayesian hierarchical model with our regularized correlation estimate (BHM+LPoC). Bolded entry in each row indicates the lower value.}\label{tab:CRPS}
\begin{tabular}{rrrr}
\hline
& IFE & LPoC \\
\hline
Africa & 1.66 & \textbf{1.49} \\
Asia & \textbf{0.73} & 0.74 \\
Europe & 3.92 & \textbf{3.76} \\
Latin America and the Caribbean & \textbf{1.62} & \textbf{1.62} \\
Northern America & 5.02 & \textbf{4.99} \\
Oceania & 8.53 & \textbf{8.49} \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fourRegions_ggplot}
\caption[Probabilistic projections of net migration for regional aggregates]{Medians and 80\% prediction intervals for projections of net migration rates for regional aggregates. Projections from the Bayesian hierarchical model with independent forecast errors (BHM+IFE) are given in red. Projections using our estimated correlation matrix (BHM+LPoC) are in blue. Overlap is in purple.}\label{fig:fourRegions}
\end{figure}
\clearpage
\subsection{Simulation study}\label{sec:simStudy}
In this section we show by simulation that our regularization procedure improves correlation estimates in a low-dimensional setting.
To match the application of interest, we simulate 12 observed time points from an AR(1) process with correlated errors.
For computational tractability, we decrease the number of simulated countries from 191 in the real data to 9 in the simulation. For each of 100 simulations, we perform the following procedure:
\begin{enumerate}
\item Generate a set of simulated migration rates $\boldsymbol{g}_1, \ldots, \boldsymbol{g}_{12}$ from an AR(1) process with errors correlated as described below.
\item Produce point estimates of $\boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{11}$ via MCMC sampling of $\boldsymbol{\mu}$, $\boldsymbol{\phi}$, and $\boldsymbol{\sigma}$.
\item Convert $\boldsymbol{\varepsilon}_t$'s to a matrix $\tilde R$ using the procedure in Section \ref{sec:RTilde}.
\item Solve the minimization problem (\ref{eqn:minProblem}) to obtain a regularized estimate for the correlation matrix.
\end{enumerate}
Since the procedure for selecting $\lambda$ is computationally intensive, we perform this procedure only once and use the same value of $\lambda$ for all subsequent simulations.
\subsubsection{Simulation details}
We simulate a collection of nine countries with true migration rates governed by the AR(1) process
\begin{equation}
\boldsymbol{g}_t - \boldsymbol{\mu} = \textrm{diag}(\boldsymbol{\phi}) (\boldsymbol{g}_{t-1} - \boldsymbol{\mu}) + \boldsymbol{\varepsilon}_t.
\end{equation}
For simplicity we take $\boldsymbol{\mu} = \boldsymbol{0}$, $\boldsymbol{\phi} = \frac{1}{2}\boldsymbol{1}$, and
\begin{equation}
\boldsymbol{\varepsilon}_t \overset{iid}{\sim} \mathcal{N}_9(\boldsymbol{0}, \Sigma).
\end{equation}
We fix $\Sigma$ to be block diagonal. Compound symmetric correlation structure within each $3 \times 3$ block is given by
\begin{equation}
\Sigma_{3 \times 3}=\left(
\begin{array}{ccc}
1 & 0.5 & 0.5 \\
0.5 & 1 & 0.5 \\
0.5 & 0.5 & 1
\end{array}
,
\right)
\end{equation}
and the full covariance matrix by
\begin{equation}
\Sigma = \left(
\begin{array}{ccc}
\Sigma_{3 \times 3} & \boldsymbol{0} & \boldsymbol{0} \\
\boldsymbol{0} & \Sigma_{3 \times 3} & \boldsymbol{0} \\
\boldsymbol{0} & \boldsymbol{0} & \Sigma_{3 \times 3} \\
\end{array}
\right).
\end{equation}
We then simulate observations $\boldsymbol{g}_1, \ldots, \boldsymbol{g}_{12}$ and attempt to make inference on the correlation structure of $\Sigma$.
Because we are basing inference on a small number of time points, Pearson estimates of correlation are highly variable.
Solid curves in Figure \ref{fig:simStudyResults} show the distributions of the off-diagonal elements of the unregularized Pearson correlation matrix in the ideal scenario where the values of $\boldsymbol{\varepsilon}_t$ can be perfectly estimated.
The top panel shows the distribution of the elements for which the true correlation is zero.
The bottom panel shows elements for which the true correlation is 0.5.
In both cases, high variability makes inference difficult.
Our method is designed to decrease variability among estimated correlations for those country pairs where prior knowledge suggests that correlation should be close to zero.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{simulationSplitPlot}
\caption[Simulation study: Elements of the correlation matrix before and after regularization]{Simulation study results: Comparison of elements of the correlation matrix before regularization (solid curves) and after (dashed curves). Top panel shows penalized elements; bottom panel shows unpenalized elements. True correlations are indicated with dashed vertical lines.}\label{fig:simStudyResults}
\end{figure}
To illustrate a best case scenario, we choose a penalty matrix $P$ which is well suited to the true correlation structure. The simplest such $P$ is the one which penalizes the off-diagonal elements of the correlation matrix if and only if the true correlation is zero. That $P$ is given by
\begin{equation}
P=\left(
\begin{array}{ccc}
\boldsymbol{0}_{3 \times 3} & \boldsymbol{1}_{3 \times 3} & \boldsymbol{1}_{3 \times 3}\\
\boldsymbol{1}_{3 \times 3} & \boldsymbol{0}_{3 \times 3} & \boldsymbol{1}_{3 \times 3} \\
\boldsymbol{1}_{3 \times 3} & \boldsymbol{1}_{3 \times 3} & \boldsymbol{0}_{3 \times 3} \\
\end{array}
\right).
\end{equation}
\subsubsection{Initial run to select $\lambda$}
Our procedure to select $\lambda$ is computationally expensive, as it requires us to compute $\hat R(\lambda)$ repeatedly as $\lambda$ varies. We therefore perform this procedure only once and use the same $\lambda$ for estimation of $R$ in all subsequent simulated data sets. Figure \ref{fig:shrinkage_simulated} plots our $\lambda$-selection criterion based on a single simulated data set over the range $\lambda=0, 0.1, 0.2, \ldots, 10$. The exact curve, shown in black, exhibits some jumpiness in this low-dimensional setting, a problem which naturally becomes less severe in the high-dimensional setting of interest. Because of this jumpiness, we base our selection of $\lambda$ on a Lowess-smoothed curve, selecting the maximizing value of $\lambda=6.4$.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{shrinkage_simulated}
\caption[Simulation study: Regularization criterion]{Average shrinkage minus average inflation of elements of $\hat R(\lambda)$ as $\lambda$ varies from 0 to 10. Exact curve in black, Lowess-smoothed curve in red.}\label{fig:shrinkage_simulated}
\end{figure}
\subsubsection{Evaluation of repeated estimation of $R$}
We produced 100 estimates of $\hat R(\lambda=6.4)$ from 100 different sets of simulated migration rates, all using the same block diagonal correlation structure. Dashed lines in Figure \ref{fig:simStudyResults} show the distribution of off-diagonal elements of $\hat R$, split into those elements where the true correlation is 0 and elements where the true correlation is 0.5 (top and bottom panel, respectively).
Our method is successful in shrinking penalized elements towards zero. Among elements where the true correlation is zero, we correctly estimate an exact zero in 62\% of cases in this simulation.
Among unpenalized elements, our method produces estimates with slightly more variability (the standard deviation is 0.256 for Pearson correlations versus 0.272 for our estimates).
Both methods produce estimates for unpenalized elements that are within two standard errors of the true mean value of 0.5. The mean estimated correlation is 0.489 for Pearson correlations (standard error 0.009) versus 0.514 for our estimates (standard error of 0.009).
On the whole, the LPoC estimator greatly improves estimates of penalized elements at the expense of slightly increasing variability in unpenalized elements.
Table \ref{tab:simulationEvaluation} compares mean absolute error and mean squared error from our method with two competing estimators. We compare our results against both Pearson correlation matrices and correlation matrices that have been regularized using the Ledoit-Wolf method, which shrinks Pearson estimates towards a spherical correlation structure \citep{ledoit2003}.
In the top panel, we estimate $\boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{11}$ with a Bayesian hierarchical model, as is done in our real application to migration.
In the bottom panel, we assume instead a scenario where we have direct access to $\boldsymbol{\varepsilon}_1, \ldots, \boldsymbol{\varepsilon}_{11}$, as would be suitable in other applications where the interest is in estimating correlations of directly observed quantities.
In both cases, our method provides an overall reduction in mean squared error by at least two thirds when compared against the Pearson sample correlation matrix.
A large reduction in error from shrinking penalized elements is offset by a mild increase in error among unpenalized elements.
We also outperform the Ledoit-Wolf estimator in terms of overall error.
\begin{table}
\caption[Evaluation of correlation matrix estimates from simulation study]{Evaluation of correlation matrix estimates from simulation study. ``LPoC'' refers to our estimator, which uses Laplace priors on correlations. MAE is mean absolute error. MSE is mean squared error. Averages over ``all elements'' exclude diagonal elements, which are fixed at zero by all methods.
The lowest (best) values are shown in bold.}\label{tab:simulationEvaluation}
\begin{center}
\begin{tabular}{cccc}
\multicolumn{4}{c}{Values of $\boldsymbol{\varepsilon}_t$ estimated with MCMC} \\
& Estimator & MAE & MSE\\
\hline
\hline
\multirow{3}{*}{All elements}
& Pearson & 0.253 & 0.098\\
& Ledoit-Wolf & 0.193 & 0.055\\
& LPoC & \textbf{0.090} & \textbf{0.028}\\
\hline
\multirow{3}{*}{True correlation = 0}
& Pearson & 0.270 & 0.109\\
& Ledoit-Wolf & 0.190 & 0.053\\
& LPoC & \textbf{0.049} & \textbf{0.012}\\
\hline
\multirow{3}{*}{True correlation = 0.5}
& Pearson & 0.201 & 0.066\\
& Ledoit-Wolf & \textbf{0.200} & \textbf{0.060}\\
& LPoC & 0.214 & 0.074\\
\hline \hline
\noalign{\vskip 2mm}
\multicolumn{4}{c}{True values of $\boldsymbol{\varepsilon}_t$ used} \\
& Estimator & MAE & MSE\\
\hline
\hline
\multirow{3}{*}{All elements}
& Pearson & 0.227 & 0.079\\
& Ledoit-Wolf & 0.182 & 0.047\\
& LPoC & \textbf{0.078} & \textbf{0.022}\\
\hline
\multirow{3}{*}{True correlation = 0}
& Pearson & 0.244 & 0.089\\
& Ledoit-Wolf & 0.162 & 0.039 \\
& LPoC & \textbf{0.041} & \textbf{0.010}\\
\hline
\multirow{3}{*}{True correlation = 0.5}
& Pearson & \textbf{0.176} & \textbf{0.051}\\
& Ledoit-Wolf & 0.243 & 0.073\\
& LPoC & 0.190 & 0.058\\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\section{Discussion}
Our method augments probabilistic projections of migration
that are well-calibrated for individual countries,
with a correlation structure that reflects prior knowledge of between-country correlations.
By combining a high-dimensional empirical correlation matrix with an informative prior that shrinks spurious correlations, we produce an estimated correlation matrix that is in line with migration theory and improves projections of regional aggregates.
When compared with a simple model that assumes uncorrelated forecast errors, our method narrows projections of net migration for Africa and widens projections for Europe.
Out-of-sample evaluation confirms that these changes produce better probabilistic forecasts as measured by continuous ranked probability score.
Mechanically, the novelty of our method is our prior on correlations, which benefits from being interpretable and simple in form, and converts MAP estimation to an $\ell_1$-penalized regularization problem which is computationally tractable.
Our analysis focuses on modeling net migration, but an attractive alternative would be to model a full matrix of bilateral migration flows.
Such a model would naturally imply correlations in migration---if out-migrants from country $i$ tend to go to country $j$, then net migration in countries $i$ and $j$ will be negatively correlated.
However, modeling the global bilateral flow matrix is currently not feasible.
Flows are hard to estimate, even in countries with good data \citep{deBeer2010, raymer2011}.
\cite{abel2013} produces global estimates of migration flows based on migrant stock data, but for only a small number of time periods at which migrant stock data exist.
His method involves minimizing the total number of migrants subject to
the available data on migrant stocks.
This induces many structural zeroes in his estimates, making modeling difficult.
Because of the lack of good data on migration flows, we choose instead to work with net migration rates.
Although our method produces a MAP estimator in the presence of informative priors, we are not able to leverage any of the usual Bayesian machinery to produce a sample from the posterior distribution.
While it would in theory be possible to use MCMC methods to produce a posterior sample by updating one element of the correlation matrix at a time, an updating procedure would need to iterate through some 18,000 elements of the correlation matrix, checking for positive definiteness after each proposed step.
Such an algorithm is therefore likely to move around the parameter space too slowly to be of any use.
In some settings a Laplace approximation centered at the posterior mode can provide a good approximation of marginal posterior distributions \citep{tierney1986}.
However, the double-exponential priors in our setting render this procedure impracticable.
Within each orthant of the parameter space, a quadratic approximation to the log likelihood is reasonable, but because of the $\ell_1$ penalty term, a different quadratic approximation is required for each of the roughly $2^{18,000}$ orthants, which is not feasible.
Given our interest in combining data with prior beliefs, an inverse Wishart prior on covariance is tempting because it allows easy sampling from the full posterior.
However, the inverse Wishart distribution is restrictive in form \citep{barnard2000} and does not provide a straightforward way to describe prior beliefs about correlations.
Another tempting alternative is that of \cite{liuWangZhao2014}, who give a simple thresholding method for producing a penalized correlation matrix that is guaranteed to be positive definite. Their estimator solves
\begin{equation}
\underset{\omega \succ \delta \cdot I}{\textrm{argmin}}
\frac{1}{2} \|\tilde R - \omega\|_F^2 + \lambda \|W * \omega\|_{1,\textrm{off}},
\end{equation}
to produce an estimator among the set of valid correlation matrices with minimum eigenvalue no smaller than $\delta$. Although the weight matrix, $W$, is in principle arbitrary, they use $W$ to induce greater shrinkage where empirical correlations are weakest, not as a means of conveying prior information.
We would be hesitant to replace $W$ with our penalty matrix $P$, as that off-license use of their method would not incorporate prior information in a principled way.
Our method can be generalized to shrink estimated correlations towards non-zero values by replacing the penalty term $\lambda\|P*R\|_1$ with $\lambda\|P*(R-S)\|_1$ for some target matrix $S$.
This may be desirable in cases where heavily structured estimates of correlations are available, as is the case for modeling of fertility
\citep{fosdick2014}.
Note that we have used the 2012 revision of the WPP here \citep{wpp2012}.
The more recent 2015 revision \citep{wpp2015} contains one additional data point.
It would be of interest to redo the analysis with the newer data, but we expect the results would be similar.
|
\section{Introduction}
\label{intro-section}
This paper examines polynomials in $q$,
generalizing what are sometimes
called {\it Kreweras}
numbers,
as refinements of the {\it Catalan numbers}.
The $q$-Kreweras numbers arose as
by-products of work of the second author \cite{Sommers2}
on nilpotent orbits in
a simple Lie algebra $\ggg$ over $\CM$ under the action of the associated connected algebraic group $G$.
Their definition, using the Lusztig-Shoji algorithm in Springer theory,
is reviewed in Section~\ref{general-Kreweras-definition-section} below.
More specifically, one has polynomials in $q$ defined for
certain positive integral parameters $m$ (see below) and
for each nilpotent orbit $\0$ in $\ggg$ and
each {\it local system} on $\0$ that arises in the Springer correspondence.
Let $\Phi$ be the root system of $\ggg$ relative to a Cartan subalgebra $\mathfrak h$.
What we call here the
{\it $q$-Kreweras numbers} $\krew$ correspond
to the {\it trivial} local system on $\0$.
In this paper we show three new
results about the polynomials $\krew$, namely
Theorems~\ref{q-Kreweras-formula-theorem},
~\ref{divisibility-and-nonnegativity-theorem_all},
~\ref{CSP-theorem} below, which will be proven in
Sections ~\ref{proof-of-Kreweras-formulas-section},
~\ref{proof-of-divisibility-and-nonnegativity-section},
~\ref{proof-of-CSP-theorem-section}, respectively.
In types $A, B, C$, we also discuss a definition of
$q$-Narayana numbers $\Nar(\Phi,m,k;q)$, which are sums of $q$-Kreweras numbers depending on a statistic on $\0$,
and establish Theorem \ref{q-Narayana-formula-theorem} in
Section \ref{proof-of-Narayana-formulas-section}.
We will take up the $q$-Narayana numbers for other types in a sequel paper.
The further parameter $m$ in the definition
of $\krew$ amd $\Nar(\Phi,m,k;q)$
is a positive integer
that is {\it very good for $\Phi$}:
this amounts to $m$ being relatively prime to the {\it Coxeter number} $h$ in
types $A,E,F,G$, and the (weaker) condition of $m$ being
{\it odd} in the classical types $B,C,D$.
Let $W$ be the Weyl group of $\Phi$.
Since $\ggg$ is simple, $W$ acts irreducibly on $\mathfrak h$, which is called the
reflection representation of $W$ and denoted by $V$. Let $r = \dim V$, the rank of $\ggg$.
Let $d_1 \leq \ldots \leq d_r$ be the degrees of any
set of fundamental invariants for the action of $W$ on
the polynomials $S:=\Sym(V^*)$ on $V$.
The Coxeter number $h$ of $\Phi$ is equal to $d_r$.
Define
\begin{equation}
\label{W-q-Catalan-definition}
\Cat(W,m;q):=\prod_{i=1}^r \frac{[m-1+d_i]_q}{[d_i]_q}
\end{equation}
where $[n]_q := 1+q+q^2+\cdots+q^{n-1}$.
This is known to be a polynomial in $\NN[q]$ for all very good $m$.
Results from \cite[\S 5.3]{Sommers2} imply a summation formula
\begin{equation}
\label{q-Kreweras-sum-to-q-Cat}
\Cat(W,m;q)=\sum_{\0} \krew
\end{equation}
as $\0$ runs through the nilpotent orbits in
$\ggg$, which is a generalization of known results for the specialization at $q=1$, as we now recall.
In type $A_{n-1}$, so that $m$ is
very good if $\gcd(m,n)=1$, this $\Cat(W,m;q)$ is the
{\it rational $q$-Catalan number}
considered, for example,
by Armstrong, Rhoades and Williams \cite{ArmstrongRhoadesWilliams}.
At the specializations $m = h+1 = n+1$ and $q=1$, these become
the {\it Catalan numbers}
\begin{equation}
\label{type-A-Catalan}
C_n :=\frac{1}{n+1}\binom{2n}{n}
\end{equation}
which have
a plethora of combinatorial interpretations
(see Stanley \cite{Stanley-Catalan-book} and
\cite[Exer. 6.19]{Stanley-EC2}),
some restricting to interpretations of the successive refinements
by the {\it Narayana} $N(n,k)$ and {\it Kreweras numbers} $\Krew(\lambda)$:
$$
\begin{array}{rcccccl}
C_n &=&\displaystyle \sum_{k=1}^n N(n,k)
& \text{ where }
& N(n,k)
&:= &\displaystyle \frac{1}{k}\binom{n-1}{k-1}\binom{n}{k-1} \\
N(n,k)&=&\displaystyle \sum_{\substack{\lambda \in \Par(n):\\ \ell(\lambda)=k}} \Krew(\lambda)
& \text{ where }
&\Krew(\lambda)
&:=& \displaystyle \frac{1}{n+1} \binom{n+1}{n-k,\mu_1(\lambda),\mu_2(\lambda), \ldots, \mu_n(\lambda)}
\end{array}
$$
Here $\Par(n)$ is the set of all partitions
$
\lambda = (\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_{\ell} > 0)
$
of the number $n$ and $\mu_j(\lambda)$ denotes the multiplicity of the number
$j$ among the parts $\{ \lambda_i \}$.
The number of parts $\ell=:\ell(\lambda)$ of $\lambda$ is called the {\it length} of
$\lambda$.
Kreweras \cite{Kreweras} originally interpreted
$C_n, N(n,k), \Krew(\lambda)$ in terms of the set $NC(n)$
of {\it noncrossing set partitions}
of $\{1,2,\ldots,n\}$ arranged circularly in the plane; that is, partitions
for which the convex hull of the blocks are pairwise disjoint.
The Catalan number $C_n$ counts the whole set $NC(n)$,
while the Narayana number $N(n,k)$ counts those noncrossing set partitions
with exactly $k$ blocks, and the Kreweras number $\Krew(\lambda)$ counts
those for which $\lambda$ lists their block sizes.
The following table illustrates these interpretations for $n=4$.
\vskip.1in
\begin{center}
\begin{tabular}{|c|c|c|c|c|}\hline
$\substack{\text{noncrossing}\\\text{partitions}}$ &
$\lambda$ & $\Krew(\lambda)$ & $k$& $N(n,k)$
\\ \hline\hline
$\substack{1234}$ &
$(4)$ & $\frac{1}{5}\binom{5}{4,1}=1$ &
$1$ & $\frac{1}{1}\binom{3}{0}\binom{4}{0} =1$ \\
\hline\hline
$ \substack{123-4\\124-3\\134-2\\1-234} $ &
$(3,1)$ & $\frac{1}{5}\binom{5}{3,1,1}=4$ & & \\
& & & $2$ & $\frac{1}{2}\binom{3}{1}\binom{4}{1} =6$ \\
$ \substack{12-34\\14-23} $ &
$(2,2)$ & $\frac{1}{5}\binom{5}{3,2}=2$ & & \\
\hline\hline
$ \substack{12-3-4\\13-2-4\\14-2-3\\1-23-4\\1-24-3\\1-2-34} $ &
$(2,1,1)$ & $\frac{1}{5}\binom{5}{2,2,1}=6$ &
$3$ & $\frac{1}{3}\binom{3}{2}\binom{4}{2} =6$ \\
\hline\hline
$\substack{1-2-3-4}$ &
$(1,1,1,1)$ &
$\frac{1}{5}\binom{5}{1,4}=1$ &
$4$ & $\frac{1}{4}\binom{3}{3}\binom{4}{3} =1$\\
\hline
\end{tabular}
\end{center}
\vskip.1in
\subsection{Generalizations to other $W$ and other $m$}
\subsubsection{Generalizing noncrossing partitions to other $W$ and $m = sh+1$}
The set of noncrossing set partitions $NC(n)$
has a generalization to all Weyl groups and for any parameter $m$ of the form $sh+1$ where $s \in \mathbb N$.
The case of $m =h+1$ was introduced by Bessis \cite{Bessis1}.\footnote{In fact, Bessis's work in \cite{Bessis1} deals not just with Weyl groups, but
all finite real reflection groups, and his later work in
\cite{Bessis2} deals more generally with the
class of {\it well-generated complex reflection groups}. See
work of Gordon and Griffeth \cite{GordonGriffeth} for definitions of
Catalan and $q$-Catalan numbers that apply to {\it all}
complex reflection groups.}
\begin{definition}
Consider
the Cayley graph for $W$ with respect to the
generating set of {\it all} reflections in $W$. Fix a {\it Coxeter element $c$} in $W$.
Then $NC(W)$ is defined to be the set of $w$ in $W$ that lie
along a shortest path between the identity
element and $c$ in this Cayley graph. We regard $NC(W)$ as a partially ordered set in
which $x \leq y$ if there is a shortest path from the identity to $c$ in this
Cayley graph that passes first through $x$ and then through $y$.
\end{definition}
Bessis \cite{Bessis1} showed
that the cardinality of $NC(W)$ equals $\Cat(W,h+1,q=1)$,
generalizing Kreweras's original interpretation of the Catalan numbers $C_n$ counting the number of
noncrossing set partitions $NC(n)$ in type $A_{n-1}$.
Armstrong \cite{Armstrong1} defined a generalization of $NC(W)$ for each positive integer $s$,
inspired by Edelman's {\it $s$-divisible noncrossing partitions} \cite[\S 4]{Edelman}.
\begin{definition}
Let $NC^{(s)}(W)$ denote the set of all
$s$-element multichains $w_1 \leq \cdots \leq w_s$
in $NC(W)$.
\end{definition}
Armstrong showed that cardinality of
$NC^{(s)}(W)$ equals $\Cat(W,sh+1,q=1)$.
Note that when $s=1$ one recovers the set $NC(W)=NC^{(1)}(W)$.
\subsubsection{Generalizing Kreweras numbers to other $W$ and very good $m$}
The Kreweras numbers $\Krew(\lambda)$ in type $A_{n-1}$ have a generalization to any $W$ and any very good $m$.
Let $X \subset V$ be the common fixed points of a set of reflections in $W$. Then the pointwise-stabilizer
subgroup $W_X$ of $X$ in $W$ is a parabolic subgroup of $W$.
The normalizer $N(W_X)$ of $W_X$ within $W$ is then the not-necessary-pointwise-stabilizer of
$X$ within $W$. We are interested in $X$ and $W_X$ up to $W$-conjugacy, so we set $[X]:=W \cdot X$ for the $W$-orbit of $X$.
Associated to the subspace $X$ is a hyperplane arrangement,
obtained by considering the hyperplanes in $X$ of the form $V^{w} \cap X$ where $w$ is a reflection in $W$ and
$V^{w}$ denotes the pointwise-stabilizer of $w$ on $V$.
The characteristic polynomial of a hyperplane arrangement
\cite{OrlikTerao} is an important invariant, denoted by $p_X(t)$ for
the hyperplane arrangement we are considering in $X$.
Using work of Orlik-Solomon \cite{OrlikSolomon1}, we know
from \cite{Sommers1} that
when $m$ is a very good for $W$ that
\begin{equation} \label{ungraded_rep}
\sum_{[X]} \tfrac{p_X(m)}{[N(W_X):W_X]} \cdot 1_{W_X}^W,
\end{equation}
is a representation of $W$ whose character takes the value $m^{\dim V^w}$ at $w \in W$.
Moreover, by Shepard-Todd \cite{ShephardTodd}, proved uniformly by Solomon \cite{Solomon},
we know that the multiplicity of the trivial representation in \eqref{ungraded_rep}
is $\Cat(W, m, q=1)$, which implies that
\begin{equation} \label{ungraded_sum_Cat}
\Cat(W, m, q=1) = \sum_{[X]} \tfrac{p_X(m)}{[N(W_X):W_X]}.
\end{equation}
In fact, $p_X(t)$ takes a simple form. It is a monic polynomial with
positive roots
\begin{equation}
\label{Orlik-Solomon-exponents}
m_1(X),m_2(X),\ldots,m_{\dim(X)}(X)
\end{equation}
called the {\it Orlik-Solomon exponents} for $X$.
The fact that $p_X(t)$ has this form is a consequence of this hyperplane arrangement being
{\it free} (see \cite{Terao}), proved by Orlik-Terao \cite{OrlikTerao} and then uniformly by Broer \cite{Broer} and Douglass \cite{Douglass}.
In type $A_{n-1}$, $W_X$ is, up to conjugacy, just a Young subgroup of $S_n$ of the form $S_{\lambda_1} \times S_{\lambda_2} \times \dots S_{\lambda_k}$, which allows us to associate the partition $\lambda = (\lambda_1, \lambda_2, \dots, \lambda_k)$ of $n$ to $W_X$.
Then $p_X(t) = (t-1)(t-2) \cdots (t-k+1)$ and
$\tfrac{p_X(n+1)}{[N(W_X):W_X]}$ coincides with $\Krew(\lambda)$.
Based on this fact and
\eqref{ungraded_sum_Cat}, Athanasiadis and Reiner \cite{AthanasiadisReiner} considered
$$\Krew(W,[X],m) := \tfrac{p_X(m)}{[N(W_X):W_X]} = \tfrac{1}{[N(W_X):W_X]} \prod_{i=1}^{\dim(X)} \left( m-m_i(X) \right)$$
as the Kreweras numbers for arbitrary $W$ and very good $m$. They then showed that
$\Krew(W, [X], h+1)$ equals the cardinality
of the set of $w \in NC(W)$ with $V^w \in [X]$ \cite[Theorem 6.3]{AthanasiadisReiner}. In classical ($A,B,C,D$) and dihedral ($I$) types,
work of Rhoades \cite{Rhoades} showed more generally that
$\Krew(W, [X], sh+1)$ counts the elements $w_1 \leq \cdots \leq w_s$
in $NC^{(s)}(W)$ with $V^{w_1}$ in $[X]$; this remains open
in the exceptional types $E,F, H$.
\subsection{Cyclic Sieving}
Armstrong defined a natural action of the cyclic group $\ZZ/sh\ZZ$ on $NC^{(s)}(W)$.
In \cite{BessisR} it was shown that this gives an instance
of the {\it cyclic sieving phenomenon} introduced in \cite{RSW} on
$NC^{(s)}(W)$.
To state it, for a positive integer $d$, let
$\omega_d:=e^{\frac{2\pi i}{d}}$, a primitive $d^{th}$ root-of-unity.
\begin{theorem}[\cite{BessisR}] \label{BR_CSP}
For $m = sh+1$,
one has that
$\Cat(W,m;q=\omega_{d})$ counts
those
$$w_1 \leq \cdots \leq w_s \in NC^{(s)}(W)$$
that are fixed under the action of
an element of order $d$ in the $\ZZ/sh\ZZ$-action.
\end{theorem}
In \cite[\S 6]{BessisR}, it was asked how to produce $q$-Kreweras numbers,
$\Krew(\Phi,X,m;q)$, polynomials that would evaluate
to the Kreweras numbers $\Krew(W, [X], m)$ at $q=1$,
but more generally would have the following property:
$\Krew(\Phi, X,m; q=\omega_{d})$ counts
the elements $w_1 \leq \cdots \leq w_s \in NC^{(s)}(W)$
with $V^{w_1} \in [X]$ and which are additionally fixed under the action
an element of order $d$ in the $\ZZ/sh\ZZ$-action.
Such a result would generalize Theorem \ref{BR_CSP}.
\subsection{The $q$-Kreweras numbers}
In work of the second author \cite{Sommers2},
a polynomial in the variable $q$, denoted $f_{e,\phi}(m;q)$,
is introduced for a nilpotent element $e \in \ggg$, an irreducible representation $\phi$ of
the component group of $e$ arising in the Springer correspondence,
and a very good $m$ (see \S \ref{general-Kreweras-definition-section}).
The definition involves a graded version of the representation in \eqref{ungraded_rep} and only depends on the nilpotent orbit $\0$ containing $e$.
Given $X$ as before, the centralizer in $\mathfrak g$ of $X \subset \mathfrak h$ is a Levi subalgebra, denoted $\mathfrak l_X$, which contains $\mathfrak h$ and whose Weyl group identifies with $W_X$.
We write $\0_X$ for the unique nilpotent orbit which contains elements that are principal nilpotent in $\mathfrak l_X$.
The definition of $\0_X$ only depends on $[X]$. We say $\0_X$, and each of the elements it contains, is principal-in-a-Levi.
Now when $e \in \0_X$, then $f_{e, \phi}(m; 1) = \Krew(W,[X],m)$.
Setting $\phi$ to be trivial, we also have
$$\Cat(W,m;q) = \sum f_{e, 1}(m;q),$$
where the sum is over a set of representatives $e$ from each nilpotent orbit.
These two results led us to the following definition of $q$-Kreweras numbers for each nilpotent orbit $\0$
$$\krew := f_{e,1}(m ;q) \text{ where } e \in \0$$
and to conjecture
\begin{conjecture}
\label{CSP-conjecture}
For $m = sh+1$,
and for each $W$-orbit $[X]$,
$\Krew(\Phi,\0_X,m;q=\omega_{d})$ counts
those
$$w_1 \leq \cdots \leq w_s \in NC^{(s)}(W)$$
which are fixed under the action an element of order $d$ in the $\ZZ/sh\ZZ$-action
and have $V^{w_1} \in [X]$.
\end{conjecture}
\subsection{Results}
\subsubsection{Formulas for the $q$-Kreweras numbers}
The formulas for $f_{e, \phi}$ for general $\phi$ in the classical groups are given in Propositions~\ref{BC_most_general} and \ref{D_most_general}. The formulas for $f_{e, \phi}$ in the exceptional groups are tabulated in Section~\ref{Exceptional_calcs}.
Theorem~\ref{q-Kreweras-formula-theorem} below
summarizes the formulas for the $q$-Kreweras numbers (that is, $\phi$ trivial)
in the classical types.
Recall that the nilpotent orbits in the classical Lie algebras
can be parametrized by number partitions $\lambda$
obeying certain restrictions by considering the defining representation of $\ggg$
and taking the Jordan form for an element in the orbit.
For such a partition $\lambda$, we write $\0_\lambda$ for the corresponding orbit in $\ggg$.
\begin{itemize}
\item In type $A_{n-1}$, that is, $\ggg=\mathfrak{sl}_n(\CM)$,
nilpotent orbits are parametrized by
all partitions $\lambda$ of $n$;
as before, denote this set of partitions
$\Par(n)$.
\item In type $B_{n}$, that is, $\ggg=\mathfrak{so}_{2n+1}(\CM)$,
nilpotent orbits are parametrized by
partitions $\lambda$ of $2n+1$
having $\mu_j(\lambda)$ even for $j$ even;
denote this set of partitions
$\Par_B(2n+1)$.
\item In type $C_{n}$, that is, $\ggg=\mathfrak{sp}_{2n}(\CM)$,
nilpotent orbits are parametrized by
all partitions $\lambda$ of $2n$
having $\mu_j(\lambda)$ even for $j$ odd;
denote this set of partitions
$\Par_C(2n)$.
\item In type $D_{n}$, that is, $\ggg=\mathfrak{so}_{2n}(\CM)$,
nilpotent orbits under the orthogonal group $\mbox{O}_{2n}(\CM)$ are
parametrized by partitions $\lambda$ of $2n$ having
$\mu_j(\lambda)$ even for $j$ even;
denote this set of partitions
$\Par_D(2n)$.
We denote by $\0_{\lambda}$ the orbit under $\mbox{O}_{2n}(\CM)$.
Then $\0_{\lambda}$ is a single $\mbox{SO}_{2n}(\CM)$-orbit unless
$\lambda$ has only even parts, in which case
$\0_{\lambda}$ splits into two orbits under $\mbox{SO}_{2n}(\CM)$.
Both of these orbits have the same
$q$-Kreweras numbers, given by $1/2$ times the formula shown in
Theorem~\ref{q-Kreweras-formula-theorem}(Type $D_n$).
\end{itemize}
For the classical groups of types $A,B,C,D$
the polynomial $\krew$ is expressed using {\it $q$-multinomials} which are defined as
follows: for $\nu=(\nu_1,\ldots,\nu_t) \in \NN^{t}$ with
$|\nu|:=\sum_i \nu_i \leq n$, let
$$
\qbin{n}{\nu}{q} :=\qbin{n}{\nu,n-|\nu|}{q} :=
\frac{[n]!_q}{[\nu_1]!_q \cdots [\nu_t]!_q [n-|\nu|]!_q },
$$
where $[n]!_q := [n]_q [n-1]_q \cdots [1]_q$,
and define the left side to be zero whenever $|\nu| > n$.
Letting $\lambda^\prime$ denote the {\it conjugate} or {\it transpose} partition of $\lambda$,
define
$$
c(\lambda) :=\sum_{j} \lambda^\prime_j \lambda^\prime_{j+1}.
$$
\begin{theorem}
\label{q-Kreweras-formula-theorem}
(Type $A_{n-1}$)
For $\lambda \in \Par(n)$ and
for $\gcd(m,n)=1$, one has
$$
\Krew(A_{n-1}, \0_\lambda,m;q)
=q^{m(n-\ell(\lambda))-c(\lambda)}
\frac{1}{[m]_q}\qbin{m}{\mu(\lambda)}{q}.
$$
\end{theorem}
\noindent
In types $B, C, D$, the formulas are similar, replacing various
parameters by roughly half their values. Introduce the notation
$\hat{N}:=\lfloor N/2 \rfloor$ for $N \in \NN$, and
for
$\nu=(\nu_1,\nu_2,\ldots)$, set
$\hat{\nu}:=(\hat{\nu}_1, \hat{\nu}_2, \ldots)$.
Using ``$a \equiv b$'' to abbreviate ``$a = b \bmod{2}$'',
define the following quantities
$$
\begin{aligned}
L(\lambda) &:= \# \{j \in \mathbb N : \mu_j(\lambda) \text{ odd}\}, \\
\BCDexponent&
:=m(n-\hat{\ell}(\lambda)) - \frac{c(\lambda)}{2}-\frac{L(\lambda)}{4},\\
\tau_{\epsilon}(\lambda)
&: =\frac{1}{2} \sum_{\substack{j \equiv \epsilon \\ \mu_j \equiv 0}} \mu_j(\lambda)
\text{ where }
\epsilon:=
\begin{cases}0 & \text{ in type }C,\\
1 & \text{ in types }B\text{ and }D.
\end{cases}
\end{aligned}
$$
\vskip.1in
\noindent
{\bf Theorem \ref{q-Kreweras-formula-theorem}}
{\it
(Type $B_n$)
For $\lambda \in \Par_B(2n+1)$ and
for $m$ odd, one has
$$
\label{type-B-Kreweras-formula}
\Krew(B_n, \0_\lambda,m;q) =
q^{
\BCDexponent
+ \tau_{1}(\lambda)+ \frac{1}{4}}
\prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}
$$
}
For the type $C$ formula, additionally define
$$
\delta(\lambda):=
\begin{cases}
\frac{1}{4}-\frac{\ell(\lambda)}{2} &\text{ for }\ell(\lambda) \text{ odd,}\\
0 &\text{ for }\ell(\lambda) \text{ even.}
\end{cases}
$$
\vskip.1in
\noindent
{\bf Theorem \ref{q-Kreweras-formula-theorem}}
{\it
(Type $C_n$)
For $\lambda \in \Par_C(2n)$ and
for $m$ odd, one has
$$
\label{type-C-Kreweras-formula}
\Krew(C_n, \0_\lambda,m;q) =
q^{\BCDexponent
+ \tau_{0}(\lambda) + \delta(\lambda)}
\prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}.
$$
}
\noindent
\vskip.1in
In type $D_n$, the multiplicity
$\mu_1(\lambda)$ of the part $1$ in $\lambda$ plays a special role,
and we also define
$$\mu_{\geq 2}(\lambda):=(\mu_2(\lambda),\mu_3(\lambda),\ldots).$$
\vskip.1in
\noindent
{\bf Theorem \ref{q-Kreweras-formula-theorem}}
{\it
(Type $D_n$)
For $\lambda \in \Par_D(2n)$
and $m$ odd,
$\Krew(D_n, \0_\lambda,m;q)$ is
$
q^{
\BCDexponent
+ \tau_{1}(\lambda)}
$
times this:
$$
\begin{cases}
\displaystyle
q^{m-\hat{\ell}(\lambda)+1}
\prod_{i=1}^{\hat{L}(\lambda)-1} (q^{m-2i+1}-1) \cdot
\qbin{\hat{m}+1 -\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}
& \text{if }\mu_1(\lambda) \text{ is odd},\\
& \\
\displaystyle
q^{\hat{\ell}(\lambda)-\mu_1(\lambda)}
\prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1) \cdot
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}_{\geq 2}(\lambda)}{q^2}
\qbin{\hat{m}+1 - \hat{L}(\lambda)-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)} {q^2}
&\text{if } \mu_1(\lambda)\text{ even and } \hat{L}(\lambda) \geq 1,\\
& \\
q^{\hat{\ell}(\lambda)-\tau_1(\lambda)}
\qbin{\hat{m}}{\hat{\mu}(\lambda)}{q^2} +
q^{\hat{\ell}(\lambda) -\mu_1(\lambda)}
\qbin{\hat{m}}{\hat{\mu}_{\geq 2}(\lambda)}{q^2}
\qbin{\hat{m}+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)} {q^2}
& \text{if } \hat{L}(\lambda) = 0 . \\
\end{cases}
$$
}
\subsubsection{Divisibility and positivity properties of the $q$-Kreweras numbers}
Using our explicit formulas for the $f_{e,\phi}$, we gather
some of their properties in the following
Theorem~\ref{divisibility-and-nonnegativity-theorem_all}. Its statement
will be slightly less precise for a fairly short list of
ill-behaved nilpotent orbits occurring inside the
exceptional types $F_4, E_6, E_7,$ and $E_8$,
given here by their Bala-Carter notation \cite{Carter}:
\begin{equation}
\label{ill-behaved-orbits}
F_4(a_3), \,
E_6(a_3), \,
E_6(a_3)\!+\!A_1, \,
E_7(a_5), \,
E_7(a_3), \,
E_8(a_7), \,
E_8(a_6), \,
E_8(b_5), \,
E_8(a_4), \,
E_8(a_3).
\end{equation}
Let $R = \rank(Z_G(e))$.
Recall that $e$ is principal-in-a-Levi if $e \in \0_X$ for some $X$.
We will denote by $H^*(\BBB_e)$ the cohomology of the Springer fiber
for $e \in \0$, regarded as a $W$-representation,
which will play a central role in the definition of the $q$-Kreweras
numbers in \S \ref{general-Kreweras-definition-section}.
\begin{theorem}
\label{divisibility-and-nonnegativity-theorem_all}
Let $e$ be a nilpotent element {\bf not} among the
ill-behaved orbits from \eqref{ill-behaved-orbits},
and assume that $f_{e,\phi}$ is not identically zero.
Then there exists $L, c \in \NN$, independent of $\phi$, such that
$$f_{e,\phi}(m ;q) = \prod^{L}_{j=1} (q^{m+1-2j} -1) \cdot q^{cm} \cdot g_{\phi}(m; q),$$
where $g_{\phi}(m;q)$ is the sum of at most two products of the form $q^{-z} \displaystyle \prod^{R}_{i=1} \tfrac{[m-a_i]_q}{[b_i]_q}$ for some $a_i, b_i, z \in \NN$.
Moreover,
\begin{enumerate}
\item[(i)] For each very good $m$, the polynomial $q^{cm} \cdot g_{\phi}(m; q)$
lies in $\NN[q]$.
\item[(ii)] The rank $r$ of $\ggg$ equals $L + c + R$.
\item[(iii)] The multiplicity of $V$ in the $W$-representation $H^*(\BBB_e)$
is $r-c$.
\item[(iv)] If $e$ is principal-in-a-Levi, then $L=0$.
In particular, $f_{e,\phi}(m;q) \in \NN[q]$ for each very good $m$.
\item[(v)] If $e$ is not principal-in-a-Levi, then $L \geq 1$. In the exceptional types it always happens that $L=1$.
\end{enumerate}
Even when $e$ is one of the ill-behaved orbits from
\eqref{ill-behaved-orbits}, if one further specializes
to the case $\phi=1$, then the polynomial $f_{e,1}(m;q)$
is always nonzero,
and still has properties (i),(ii),(iv),(v) listed above\footnote{
Examination of the tables in Section~\ref{Exceptional_calcs}
shows that for $e$ in an ill-behaved nilpotent orbit,
for certain $\phi \neq 1$ one still has a factorization of
$f_{e,\phi}(m;q)$ as in the theorem,
but with $-g_\phi(m;q)$ in $\NN[q]$. Also, for such $e$,
property (iii) fails even if $\phi=1$. Instead, the value $r-c$ is the multiplicity of
$W$ in the $A(e)$-invariants in $H^*(\BBB_e)$.}.
\end{theorem}
\begin{comment}
A slightly modified statement holds for the remaining orbits. For simplicity we state a version of the theorem that holds for any $e$ when $\phi=1$.
\begin{theorem}
\label{divisibility-and-nonnegativity-theorem}
The $f_{e,1}$ are always nonzero and
$$f_{e,1}(m ;q) = \prod^{L}_{j=1} (q^{m+1-2j} -1) \cdot q^{cm} \cdot g(m; q)$$
for some $L, c \in \NN$ and where $g(m;q)$ is the sum of at most two products of the form $q^{-z} \displaystyle \prod^{R}_{i=1} \tfrac{[m-a_i]_q}{[b_i]_q}$ for some $a_i, b_i, z \in \NN$.
Moreover,
\begin{itemize}
\item For each very good $m$, the polynomial $q^{cm} \cdot g(m; q)$ lies in $\NN[q]$.
\item The rank $r$ of $\ggg$ equals $L + c + R$.
\item If $e$ is principal-in-a-Levi, then $L=0$.
In particular, $f_{e,1}(m;q) \in \NN[q]$ for each very good $m$.
\item If $e$ is not principal-in-a-Levi, then $L \geq 1$.
In particular, $f_{e,1}$, as polynomial in $q$, is divisible by $q^{m-1}-1$ for each very good $m$. In the exceptional types it turns out that $L$ is always equal to $1$.
\end{itemize}
\end{theorem}
\end{comment}
It follows that the $q$-Kreweras numbers $f_{e,1}(m;q)$
are never identically zero, that they have nonnegative coefficients
as polynomials in $q$ if $e$ is principal-in-a-Levi, and are divisible by $q^{m-1}-1$ otherwise.
These facts are used in establishing the cyclic sieving property.
The proof of Theorem~\ref{divisibility-and-nonnegativity-theorem_all}
is given in Section \ref{proof-of-divisibility-and-nonnegativity-section}.
\subsubsection{Cyclic sieving}
We also show that the cyclic sieving property holds in the classical types in Section \ref{proof-of-CSP-theorem-section}.
\begin{theorem}
\label{CSP-theorem}
Conjecture~\ref{CSP-conjecture}
holds in all of the classical types $A,B,C,D$.
\end{theorem}
\subsubsection{$q$-Narayana numbers}
We are able to establish a $q$-version of the Narayana numbers in types $A,B,C$. We take up the question of the $q$-Narayana numbers for other types in a sequel paper.
From Theorem \ref{divisibility-and-nonnegativity-theorem_all} we have that the lowest degree $q$-monomial in $\krew$ equals $q^{cm-z}$ for some $c,z \in \NN$.
\begin{definition}
Define $d(\0) = r - c$.
\end{definition}
As long as $e$ does not belong to one of the orbits in \eqref{ill-behaved-orbits},
then by Theorem \ref{divisibility-and-nonnegativity-theorem_all} (iii),
$d(\0)$ equals the multiplicity of $V$ in $H^*(\BBB_e)$.
In particular, this holds whenever $\ggg$ is of classical type or $e$ is principal-in-a-Levi.
On the other hand, when $e$ belongs to one of the orbits in \eqref{ill-behaved-orbits}, then
$d(\0)$ equals the multiplicity of $V$ in $A(e)$-invariants of $H^*(\BBB_e)$.
Using the parameter $d(\0)$, we obtain a good definition for a $q$-version of the Narayana numbers in types $A,B,C$.
\begin{definition} (Types $A, B, C$)
The {\it $q$-Narayana number} for $k=0,1,2,\ldots,r$ and very good $m$ is given by
\begin{equation}
\label{q-Narayana-definition}
\Nar(\Phi,m,k;q):=
\sum_{\substack{\text{nilpotent orbits } \0:\\ d(\0)=k}} \Krew(\Phi, \0 ,m;q).
\end{equation}
\end{definition}
\begin{theorem}
\label{q-Narayana-formula-theorem}
The $q$-Narayana polynomials have the following formulas in types $A,B,C$:
\begin{itemize}
\item
For type $A_{n-1}$, when $\gcd(m,n)=1$ one has for $0 \leq k \leq n-1$ that
$$
\Nar(A_{n-1},m,k;q)=
q^{(n-1-k)(m-1-k)}
\frac{1}{[k+1]_q} \qbin{n-1}{k}{q} \qbin{m-1}{k}{q}.
$$
\item
For either of type $B_n, C_n$, when $m$ is odd one has
for $0 \leq k \leq n$ that
$$
\Nar(B_n,m,k;q)=\Nar(C_n,m,k;q)=
(q^{2})^{(n-k)(\hat{m}-k)} \qbin{n}{k}{q^2} \qbin{\hat{m}}{k}{q^2}.
$$
\end{itemize}
\end{theorem}
\noindent
When $m=h+1=n+1$ in type $A_{n-1}$, the polynomial $\Nar(A_{n-1},h+1,k;q)$
equals a $q$-Narayana number considered by Furlinger and Hofbauer \cite{FurlingerHofbauer} and Br\"anden in \cite{Braenden}.
Theorem~\ref{q-Narayana-formula-theorem}
shows that, in
the one instance where two root systems $\Phi=B_n,C_n$ are associated
with the same Weyl group $W$, it turns out that
$\Nar(\Phi,m,k;q)$ depends only on $W$, and not on $\Phi$,
even though the $\krew$ for the two root systems are not
the same (they are not even indexed by the same set).
It is also interesting to note that, although the polynomials $\krew$
can have negative integral coefficients in types $B$ and $C$, the
formulas above for $\Nar(\Phi,m,k;q)$ exhibit them as polynomials
in $q$ with {\it nonnegative} coefficients, that is, lying in $\NN[q]$.
Theorem \ref{q-Narayana-formula-theorem} is proved in Section~\ref{proof-of-Narayana-formulas-section}.
\section{Reviewing $f_{e,\phi}$ and
the $q$-Kreweras numbers $f_{e,\one}$}
\label{general-Kreweras-definition-section}
In this section we detail some results from \cite{Sommers2}.
Recall that
S = \Sym(V^*)
$ is the graded ring of polynomials on the reflection representation $V$ of $W$
and $r = \dim V$.
When $m$ is very good for $\Phi$, it is known \cite{Gordon, BerestEtingofGinzburg}
that $S$ contains a
{\it homogeneous system of parameters} $\theta^{(m)}=(\theta^{(m)}_1,\ldots,\theta^{(m)}_r)$
whose
span is $W$-stable and carries a representation isomorphic to $V$.
This implies by \cite{Solomon} that the $W$-invariant subspace of the finite-dimensional quotient
ring $S/(\theta^{(m)})$ is a graded vector space whose Hilbert series is
$\Cat(W,m;q)$ and in particular shows that $\Cat(W,m;q)$ is polynomial.
A main result of \cite{Sommers2} is that $S/(\theta^{(m)})$,
as graded $W$-module, is an integral combination of certain
finite-dimensional graded $W$-representations (and their graded shifts),
related to the Green functions that arise in the representation theory of Chevalley groups.
To be more precise,
let $G$ be the connected simple algebraic group of adjoint type over an algebraically closed field
$\mathbf{k}$ of good characteristic $p$ attached to the root system $\Phi$. Let $\ggg$ be its Lie algebra.
For a nilpotent element $e \in \ggg$, let $\BBB_e$ be the variety of Borel subalgebras containing $e$, known as a Springer fiber.
Let $Z_G(e)$ be the centralizer of $e$ in $G$ and let $A(e) := Z_G(e)/ Z^{\circ}_G(e)$ be the component group of $e$.
Then the $l$-adic cohomology $H^*(\BBB_e)$ carries a representation of $W \times A(e)$ \cite{Lus1.5, Hotta}, originally defined by Springer \cite{SpringerRep}.
Denote the irreducible $\overline{{\mathbb Q}}_l$-representations of $A(e)$ by $\ar$.
For $\phi \in \ar$
define the finite-dimensional, graded representations $Q_{e, \phi}$ so that
$$H^*(\BBB_e) = \sum_{\phi \in \ar} Q_{e, \phi} \otimes \phi,$$
as graded representations of $W \times A(e)$.
The cohomology of $\BBB_e$ vanishes in odd degrees and
we grade $Q_{e, \phi}$ by putting $q$ in cohomological degree two.
Then (as Frobenius series)
\begin{equation}
\label{parking-space-as-Kreweras-times-Springer-fibers}
S/(\theta^{(m)}) = \sum_{e,\phi} f_{e,\phi}(m; q) Q_{e, \phi}
\end{equation}
\noindent
where $f_{e,\phi}(m; q) \in \ZZ[q]$ and the sum is over a set of representatives $e$ of the nilpotent orbits in $\ggg$
and those $\phi \in \ar$ that appear in the Springer correspondence.
There is an explicit formula \cite[Equation 18]{Sommers2} for $f_{e,\phi}(m; q)$ that involves (1) the cardinality of the rational nilpotent orbits in $\ggg$ for a finite subfield $\FF_q$ of $\mathbf k$;
and (2) the Frobenius series of the reflection representation $V$ in $Q_{e,\phi}$.
Define $\{(m_1, \pi_1), (m_2, \pi_2), \dots, (m_{\kappa}, \pi_{\kappa}) \},$ where $m_j \in \ZZ_{\geq 0}$ and $\pi_j \in \ar$, by
\begin{equation}
\label{nilpotent-orbit-exponents-definition}
\sum_{ j \geq 0} q^{j} \langle H^{2j}(\BBB_e), V \rangle_{W} = q^{m_1} \pi_1 + q^{m_2} \pi_2 + \cdots + q^{m_{\kappa}} \pi_{\kappa},
\end{equation}
where the pairing is the usual inner product for $W$ and the result is viewed as a Frobenius series for $A(e)$.
It turns out that at most one of the $\pi_j$'s is non-trivial and we set
$\pi_{\kappa}$ to be this non-trivial representation of $A(e)$ when it occurs.
When $e$ belongs to $\0_X$, it is known \cite{Spalt, LeSh} that the $m_i$'s coincide with the Orlik-Solomon exponents from \ref{Orlik-Solomon-exponents}.
Let $G(\FF_q) \subset G$ be the $\FF_q$-points of $G$ with respect to a split Frobenius endomorphism $F$.
Let $c$ denote a conjugacy class in $A(e)$ and $e_c$ a representative from the rational
$G(\FF_q)$-orbit in $\ggg$ over $\FF_q$ corresponding to the pair $(e,c)$.
Set $d_\kappa=\dim(\pi_\kappa)$ and
$M = m_1 + \dots + m_{\kappa-1} + d_\kappa \cdot m_{\kappa}$.
Then
\begin{equation} \label{formula:f}
f_{e, \phi}(m; q) = q^{m(r - \kappa - d_\kappa + 1)+ M} \prod_{j=1}^{\kappa-1} (q^{m-m_j} - 1) \cdot
\left( \sum_{i=0}^{d_\kappa} (-1)^{d_\kappa-i} q^{i(m - m_{\kappa})} \sum_c \frac{ \!\wedge^{d_\kappa-i} \pi_{\kappa}(c) \phi(c)} {|Z_{G^F}(e_c)|} \right)
\end{equation}
When $\pi_{\kappa}$ is trivial, the above expression simplifies to
\begin{equation} \label{formula:f_easy}
f_{e, \phi}(m; q) = q^{m(r - \kappa)+ \sum m_j} \prod_{j=1}^{\kappa} \big (q^{m-m_j} - 1 \big)
\left( \sum_c \frac{ \phi(c)} {|Z_{G^F}(e_c)|} \right)
\end{equation}
In the present paper we are primarily concerned with the Frobenius series of
$S/(\theta^{(m)})$ at the trivial representation of $W$. Since the trivial representation of $W$ only
occurs in $Q_{e,\phi}$ for the trivial local system $\phi$ and then only once in degree zero \cite{Borho-Mac_partial},
we obtain the identity
\begin{equation}
\label{original-form-of-q-Kreweras-sums-to-q-Catalan}
\Cat(W,m;q) = \sum_e f_{e,1}(m; q).
\end{equation}
By \cite[Theorem 15]{Sommers2}
\begin{equation
\label{principal-in-Levi-Kreweras-at-q=1}
f_{e,\phi}(m; q=1) = \Krew(W,m,[X])
\end{equation}
when $e$ is conjugate to a principal nilpotent element in $\mathfrak l_X$.
When $e$ is not of that form, on the other hand,
$f_{e,\phi}(m; q=1)=0.$
In light of these results, it is reasonable to think of $f_{e,1}(m; q)$ as $q$-Kreweras numbers, where now there is one such polynomial for each nilpotent orbit in $\ggg$. Thus,
\begin{definition}
\label{q-Kreweras-definition}
The $q$-Kreweras numbers for $\Phi$ are defined to be
$$\krew = f_{e,1}(m; q)$$
for $e \in \0$.
\end{definition}
We will write down the formulas for $f_{e,\phi}(m; q)$, and hence $\krew$ in Section \ref{A_calculations} for type $A_{n-1}$, Section \ref{BCD_calcs} for the other classical types, and Section \ref{Exceptional_calcs} for the exceptional types.
\section{Computing $f_{e,\phi}$ and the Proof of Theorem~\ref{q-Kreweras-formula-theorem}}
\label{proof-of-Kreweras-formulas-section}
We use the following notation in this section. For a partition $\lambda$, let
$\mu_j := \mu_j(\lambda)$, which recall is the number of parts of $\lambda$ of size $j$.
For a nilpotent element $e \in \ggg$,
let $d = \dim Z_G(e)$ and $d^u$ denote the dimension of a maximal unipotent subgroup of $Z_G(e)$.
\subsection{Type $A$}
\label{A_calculations}
For simplicity we will work with $G = GL_n(\bar{\FF}_q)$ and adjust our results for the case where $G$ is adjoint. Recall from the introduction that the nilpotent $G$-orbits in $\ggg$ are parametrized by $\Par(n)$ of partitions of $n$, with $\lambda \in \Par(n)$
corresponding to the sizes of the Jordan blocks of any element in the nilpotent orbit $\0_{\lambda}$
indexed by $\lambda$.
Let the Frobenius map $F$ consist of raising each matrix element to the $q$-th power, giving the standard
split structure on $G$, so that
$G^F = GL_n(\FF_q)$ and $\ggg^F = {\mathfrak gl}_n(\FF_q)$.
Then it is known, say by using rational canonical form, that nilpotent $G^F$-orbits on
$\ggg^F$ are indexed by $\Par(n)$; in other words, the rational points of $\0_{\lambda}$ remain a single orbit under $G^F$.
We wish to compute $$\Krew(A_{n-1}, \0_\lambda,m;q) := f_{e, 1}(m;q)$$
where $e:=e_{\lambda} \in \0_{\lambda}$ is a rational element.
Since all $A(e)$ are trivial in $GL_n(\bar{\FF}_q)$,
the computation of $f_{e,1}$ in \eqref{formula:f_easy}
reduces to calculating $m_1, \dots, m_{\kappa}$
and the value of $|Z_{G^F}(e)|$.
First, according to \cite{LeSh} we have
\begin{equation}\label{type-A-kappa-formula}
\kappa = \ell(\lambda) - 1 \text{ and } m_j = j,
\end{equation}
Thus \eqref{formula:f_easy}, with $r = n-1$, becomes
\begin{equation} \label{formula:typeA}
f_{e,1} =
q^{m(n- \ell(\lambda))+ \binom{\ell(\lambda)}{2}} \cdot
\frac{q-1}{|Z_{G^F}(e)|} \prod_{j=1}^{\ell(\lambda) -1} \big (q^{m-j} - 1 \big)
\end{equation}
where the extra factor of $q-1$ accounts for the center of $GL_n(\FF_q)$ since the formulas in
\eqref{formula:f_easy} are presented relative to an adjoint group.
Next, we need to compute $|Z_{G^F}(e)|$.
As a variety $Z_G(e)$ is isomorphic to the product of
an affine space (its unipotent radical) and a maximal reductive part of the centralizer of $Z_G(e)$.
The reductive part is isomorphic to
$$Z_{red} := \prod_j GL_{\mu_j} (\bar{\FF}_q).$$
Up to isomorphism, the affine space and each factor in the reductive part carry the standard action of $F$,
so $Z_{G^F}(e)$ is isomorphic over $\FF_q$ to
$$\FF_q^{d_1^u} \times \prod_j GL_{\mu_j} (\FF_q).$$
Since
$$|GL_{r} (\FF_q)| = q^{\binom{r}{2}} (q-1)^r [r]!_q,$$
it follows that
\begin{equation}\label{cent_A}
|Z_{G^F}(e)| = q^{d^u} (q-1)^{\ell(\lambda)} \displaystyle\prod_{j} [\mu_j]!_q
\end{equation}
The Borel subgroup of $Z_{red}$ has dimension $\sum \binom{\mu_j\!+\!1}{2}$,
so $d^u = d - \sum \binom{\mu_j\!+\!1}{2}$.
Now $d = \sum (\lambda'_i)^2$ (see \cite{Carter}) and so a calculation gives
\begin{equation} \label{small_calc}
d^u = \sum (\lambda'_i)^2 - \sum \binom{\mu_j+1}{2} =
\binom{\ell(\lambda)}{2} + \crossterms(\lambda)
\end{equation}
where
$$
\crossterms(\lambda):=
\sum_{j \geq 1} \lambda'_j \lambda'_{j+1}.
$$
Plugging (\ref{cent_A}) and (\ref{small_calc}) into (\ref{formula:typeA})
into the above
yields
$$
\begin{aligned}
f_{e,1}
&=
q^{m(n- \ell(\lambda)) - \crossterms(\lambda)}
\frac{1}{\prod_{j} [\mu_j]!_q} \cdot \frac{[m-1]!_q}{[m-\ell(\lambda)]!_q} \\
&=q^{m(n-\ell(\lambda))-c(\lambda)}
\frac{1}{[m]_q}\qbin{m}{\mu(\lambda)}{q}.
\end{aligned}
$$
as asserted in Theorem~\ref{q-Kreweras-formula-theorem}(Type $A$).
\begin{remark}
The formula for $|Z_{G^F}(e)|$ could also be obtained by looking up the value $|\mathcal{O}^F_e| = \frac{|G^F|}{|Z_{G^F}(e)|}$ in, for example, \cite{Crabb}.
\end{remark}
\subsection{Preparation for types $B, C, D$}
From (\ref{formula:f}), in order to compute $f_{e, \phi}$,
we need to evaluate the sum
$$\sum_c \frac{ \phi(c)} {|Z_{G^F}(e_c)|}$$
as $c$ runs over representatives of the conjugacy classes in $A(e)$.
The component group $A(e)$ is elementary abelian, hence we will have occasion to use the following two lemmas.
First, we introduce some notation for working with elementary abelian groups and their characters.
Let $v, w \in (\FF_2)^{s}$. Write $v = (v_i), w=(w_i)$ relative to the standard basis
and denote the usual dot product $\langle v, w \rangle := \sum_{i=1}^s v_iw_i$.
For each $w \in \FF_2^s$,
define the character $\phi_w \in \widehat{\FF_2^{s}}$ by
$$\phi_w(v) = (-1)^{\langle v, w \rangle} \in \QQ.$$
Every character of $\FF_2^{s}$ is of the form $\phi_w$ for a unique $w$.
Let $x_1, x_2, \dots, x_s$ be a set of $s$ variables.
For $a \in \FF_2$ and a variable $y$, we evaluate $y^a$ to $1$ or $y$ according to
whether $a =0$ or $a=1$, respectively, in $\FF_2$.
\begin{lemma} \label{full_lemma}
Let $\phi$ be a character of $\FF_2^s$. Write $\phi=\phi_w$ for the unique such $w \in \FF_2^s$.
Let $t$ be a nonnegative integer with $0 \leq t \leq s$. Then
\begin{eqnarray*}
\sum_{v \in \FF_2^s} \phi(v) \prod_{i=1}^{t} (x_i + (-1)^{v_i}) =
\begin{cases}
2^{s} \displaystyle \prod_{i=1}^t x_i^{w_i + 1} & \text{if } w_j=0 \text{ for all } j>t \\
0 \text{ otherwise}
\end{cases}
\end{eqnarray*}
\end{lemma}
The identity also hold in the degenerate case where $s=0$.
We omit the proof since it is essentially the same (but simpler) as the proof of the next lemma.
Let $K$ denote the subgroup of $\FF_2^{s}$ consisting of those $v \in \FF_2^{s}$ with $\sum v_i = 0$.
Any character $\phi \in \widehat{K}$ is now equal to the restriction of $\phi_w$
for two values of $w\in \FF_2^{s}$, call them $w', w''$, where
$w' + w'' = (1, 1, \dots, 1)$.
\begin{lemma} \label{even_lemma}
Let $\phi$ be a character of $K$ and $s >0$. Let $w \in \FF_2^{s}$ be the unique choice such that $\phi=\phi_w$ and $w_s=0$.
Let $t$ be a nonnegative integer with $0 \leq t \leq s$. Then
\begin{eqnarray*}
\sum_{v \in K} \phi_w(v) \prod_{i=1}^{t} (x_i + (-1)^{v_i}) =
\begin{cases}
2^{s-1} \left(\prod_{i=1}^s x_i^{w_i} + \prod_{i=1}^s x_i^{w_i + 1} \right) & \text{if } t=s\\
2^{s-1} \prod_{i=1}^t x_i^{w_i + 1} & \text{if }t<s \text{ and }w_j=0 \text{ for all } j>t \\
0 \text{ otherwise}
\end{cases}
\end{eqnarray*}
\end{lemma}
\begin{proof}
Embed $\FF_{\tiny 2}^t$ in $\FF_2^s$ via the first $t$ coordinates.
Then
\begin{eqnarray*}
\prod_{i=1}^{t} (x_i + (-1)^{v_i}) = \sum_{u \in \FF_2^{t}} \prod_{i=1}^{t} (x_i^{u_i+1} ((-1)^{v_i})^{ u_i}) =
\sum_{u \in \FF_2^{t}} (\prod_{i=1}^{t} x_i^{u_i+1} ) (-1)^{\sum v_i u_i} & \\
= \sum_{u \in \FF_2^{t}} \prod_{i=1}^{t} x_i^{u_i+1} \cdot (-1)^{\langle v, u \rangle} =
\sum_{u \in \FF_2^{t}} (\prod_{i=1}^{t} x_i^{u_i+1}) \phi_{u} (v)
\end{eqnarray*}
\noindent Writing $\mathbf{x}^{u+\mathbf{1}}$ for $\prod_{i=1}^{t} x_i^{u_i+1}$ and using the identity above
and switching the order of summation gives
\begin{eqnarray*}
\sum_{v \in K} \phi_w(v) \prod_{i=1}^{t} (x_i + (-1)^{v_i}) &
= & \sum \limits_{u \in \FF_2^{t}} \mathbf{x}^{u+\mathbf{1}} \sum_{v \in K} \phi_w(v) \phi_{u}(v).
\end{eqnarray*}
Character theory for $K$ implies that
the inner sum equals zero unless $\phi_w = \phi_{u}$ on $K$, in which case it equals $|K| = 2^{s-1}$.
Now the equality $\phi_w = \phi_{u}$ holds on all of $K$ if and only
$$w_1+w_2 = u_1+u_2, w_2+w_3 = u_2+u_3, \dots, w_{s-1}+w_s = u_{s-1}+u_s.$$
If $t=s$, this happens if and only if $u = w$ or $u = w + (1,1, \dots, 1)$, giving the first part of the result.
Next consider the case where $t<s$. Since $u_{t+1}=u_{t+2}=...=u_{s}=0$, a necessary condition for $\phi_w = \phi_u$
is that $w_{t+1}+w_{t+2} = 0$, $w_{t+2}+w_{t+3} = 0, \dots, w_{s-1}+w_{s}=0$.
Since $w_s=0$ by hypothesis, this means that $w_{t+1} = \dots = w_s = 0$ for $\phi_w = \phi_u$ to hold, giving
the third part of the result.
Moreover, if $\phi_w = \phi_u$, then also
$u_t =w_t$ since both $w_{t+1}=u_{t+1}=0$. Continuing in this fashion
$u_{t-1} = w_{t-1}, \dots, u_{1}= w_{1}$. So the unique solution for $u$ is $u=w$, which is the second part of the result.
\end{proof}
\subsection{Computing the summation.}\label{compute-summation}
Here again we abbreviate $a \equiv b \bmod{2}$ as "$a \equiv b$".
For $\epsilon \in \{ 0, 1\}$, let
\begin{eqnarray*}
S^+_{\epsilon} &:= & \{ j \in \NN \ | \ j \equiv \epsilon, \countp_j \equiv 0, \countp_j \neq 0 \}
\text{ and } \\
S^-_ {\epsilon} &:= & \{ j \in \NN \ | \ j \equiv \epsilon, \countp_j \equiv 1 \} \text{ and } \\
S_{\epsilon} &:= & S^-_{\epsilon} \cup S^+_{\epsilon} = \{ j \in \NN \ | \ j \equiv \epsilon, \countp_j > 0 \}.
\end{eqnarray*}
For types $B,C,D$, we set $q$ to be a power of an odd prime.
\medskip
{\bf Type} $C$
\medskip
For $\lambda \in \Par_C(2n)$, pick $e \in \0_{\lambda}$.
Then working in $G = \mbox{Sp}_{2n}(\aF_q)$,
a maximal reductive part of the centralizer $Z_G(e)$ is isomorphic to
\begin{equation} \label{red_decomp1}
\prod_{j \equiv 1} \mbox{Sp}_{\countp_j} (\aF_q) \times \prod_{j \in S_0} \mbox{O}_{\countp_j} (\aF_q).
\end{equation}
Let $A$ be the elementary abelian $2$-group with basis
$S_0 = S^-_0 \cup S^+_0$.
For $c \in A$, we write
$$c=(c_j)_{j \in S_0} \text{ with } c_j \in \FF_2.$$
We identify $A$ with the component group $A(e)$, where
$(c_j) \in A$ corresponds to taking an element of determinant $(-1)^{c_j}$ in the orthogonal group in (\ref{red_decomp1}) indexed by $j$,
for each $j \in S_0$
Now we assume that $e \in \ggg^F$ and has split centralizer. For $c \in A(e)$,
we twist $e$ by $c$ to get another
rational element $e_c \in \orbit_e$. In this way we obtain representatives from all
the $G^F$-orbits on $\0^F_e$.
Under our identification of $A$ and $A(e)$,
the group of rational points in a maximal reductive subgroup of $Z_G(e_c)$ is isomorphic to
$$\prod_{j \equiv 1} \mbox{Sp}_{\countp_j} (\FF_q) \times \prod_{j \in S^-_0} \mbox{O}_{\countp_j} (\FF_q) \times
\prod_{j \in S^+_0} \mbox{O}^{c_j}_{\countp_j} (\FF_q),$$
where the groups in the last product are either split or twisted orthogonal groups of type $D$
depending on whether $c_j$ is equal to $0$ or $1$, respectively; see Shoji \cite[\S 1]{Shoji}.
At this stage another $q$-analogue notation is helpful:
for a nonnegative integer $n$, let
$$
\begin{aligned}
\eta(N)&:= (q^2-1)(q^4-1) \cdots (q^{2\lfloor\frac{N}{2}\rfloor}-1) \\
&=\begin{cases}
(q^2-1)(q^4-1) \cdots (q^N-1) &\text{ if }N\text{ is even},\\
(q^2-1)(q^4-1) \cdots (q^{N-1}-1) &\text{ if }N\text{ is odd}.\\
\end{cases}
\end{aligned}
$$
or in other words,
$$
\eta(2m+1) = \eta(2m) = \prod_{i=1}^m (q^{2i} - 1).
$$
The cardinality of $Z_{G^F}(e_c)$ is therefore (see, e.g., Carter \cite[\S 2.9, p. 75]{Carter})
\begin{equation} \label{points_C}
q^{d^u} \cdot |A(e)|
\prod_{j \equiv 1} \eta(\countp_j)
\prod_{j \in S^-_0} \eta(\countp_j)
\prod_{j \in S^+_0} (q^{\frac{\countp_j}{2}} - (-1)^{c_j}) \cdot \eta(\countp_j -2).
\end{equation}
Getting a common denominator over all the conjugacy classes in $A(e)$, we obtain
\begin{eqnarray} \label{c_sum1}
\sum_c \frac{ \phi(c)}{|Z_{G^F}(e_c)|} = \frac{\sum_c \phi(c) \prod_{j \in S^+_0} (q^{ \frac{\countp_j}{2}} + (-1)^{c_j})}
{ q^{d^u} |A(e)| \prod_{j} \eta(\countp_j)}.
\end{eqnarray}
To evaluate this sum we use Lemma \ref{full_lemma} with $x_i = q^{\frac{\countp_i}{2}}$, $s = |S_0|$,
and $t = |S^+_0|$.
Choose $w \in A$ to be the unique element so that $\phi = \phi_w$.
Then by the lemma the expression in \eqref{c_sum1}
equals $0$ if $w_j \neq 0$ for any $j \in S^-_0$,
and
\begin{eqnarray} \label{c_sum3}
\sum_c \frac{ \phi(c)}{|Z_{G^F}(e_c)|} = \frac{q^{ -d^u + \sum_{j \in S^+_0, w_j = 0} \frac{\countp_j}{2}} }
{ \prod_{j} \eta(\countp_j)}
\end{eqnarray}
if $w_j = 0$ for all $j \in S^-_0$.
The value of $d^u$ is computed in Section~\ref{value-of-d-section}.
\medskip
{\bf Type $B$}
\medskip
An important feature is that $S^-_1$ is always non-empty, since $|\lambda|$ is odd.
In type $B_n$ we work with $G = \mbox{SO}_{2n+1}(\aF_q)$ and thus
a maximal reductive subgroup of $Z_G(e)$ is isomorphic to
the determinant one elements in
\begin{equation} \label{red_decompB}
\prod_{j \equiv 0} \mbox{Sp}_{\countp_j} (\aF_q) \times \prod_{j \in S_1} \mbox{O}_{\countp_j} (\aF_q).
\end{equation}
Here, we define $A$ to be the elementary abelian $2$-group with basis
$S_1$, with
elements written as $$c=(c_j)_{j \in S_1},$$ where $c_j \in \FF_2$.
Let $K$ be the subgroup of $A$ consisting of elements $(c_j)$ with $\sum c_j = 0$.
We identify $K$ with the component group $A(e)$, where
$(c_j) \in K$ corresponds to taking an element of determinant $(-1)^{c_j}$ in the orthogonal group in (\ref{red_decompB}) indexed by $j$ for each
$j \in S_1$.
Keeping the same notation as in type $C$,
the group of rational points in a maximal reductive subgroup of $Z_G(e_c)$ is isomorphic to
$$\prod_{j \equiv 0} \mbox{Sp}_{\countp_j} (\FF_q) \times \prod_{j \in S^-_1} \mbox{O}_{\countp_j} (\FF_q) \times
\prod_{j \in S^+_1} \mbox{O}^{c_j}_{\countp_j} (\FF_q).$$
The cardinality of $Z_{G^F}(e_c)$ is therefore
\begin{equation} \label{points_B}
q^{d^u} \cdot |A(e)|
\prod_{j \equiv 0} \eta(\countp_j)
\prod_{j \in S^-_1} \eta(\countp_j)
\prod_{j \in S^+_1} (q^{\frac{\countp_j}{2}} - (-1)^{c_j}) \cdot \eta(\countp_j -2) .
\end{equation}
Getting a common denominator over all the conjugacy classes in $A(e)$, we obtain
\begin{eqnarray} \label{b_sum1}
\sum_c \frac{ \phi(c)}{|Z_{G^F}(e_c)|} = \frac{\sum_c \phi(c) \prod_{j \in S^+_1} (q^{ \frac{\countp_j}{2}} + (-1)^{c_j})}
{q^{d^u} |A(e)| \prod_{j} \eta(\countp_j)}
\end{eqnarray}
To evaluate the sum we use Lemma \ref{even_lemma} with $x_i = q^{\frac{\countp_i}{2}}$, $s = |S_1|$,
and $t = |S^+_1|$. Note that $t<s$ since $S^-_1$ is non-empty, so the first scenario in the lemma never occurs.
Write $\phi$ as $\phi_w$ for $w \in A$ with $w_i=0$ for some $i \in S^-_1$; such a $w$ is unique.
Now if $w_j \neq 0$ for any $j \in S^-_1$, then the third scenario in the lemma applies and the expression in \eqref{b_sum1} equals zero.
On the other hand, if $w_j = 0$ for all $j \in S^-_1$, then since $t<s$, the second scenario of the lemma gives
\begin{eqnarray} \label{b_sum3}
\sum_c \frac{ \phi(c)}{|Z_{G^F}(e_c)|} = \frac{q^{ -d^u + \sum_{j \in S^+_1, w_j = 0} \frac{\countp_j}{2}} }
{ \prod_{j} \eta(\countp_j)}
\end{eqnarray}
The value of $d^u$ is computed in Section~\ref{value-of-d-section}.
\medskip
{\bf Type $D$}
\medskip
We proceed as in type $B$ and write $\phi = \phi_w$ for the unique
$w \in A$ with $w_i = 0$ for some $i \in S^-_1$ when $|S^-_1| > 0$.
Now, $|\lambda|$ is even and so $|S^-_1|$ is even, but it could be zero. Therefore,
in evaluating the sum $\sum_c \frac{ \phi(c)}{|Z_{G^F}(e_c)|}$ all three scenarios in Lemma \ref{even_lemma} can occur. When $s >0$, we
therefore have
\begin{align} \label{d_sum3}
\sum_c \frac{ \phi(c)}{|Z_{G^F}(e_c)|} = \frac{q^{-d^u}}{ \prod_{j} \eta(\countp_j) } \cdot
\begin{cases}
q^{ \sum_{j \in S^+_1, w_j = 1} \frac{\countp_j}{2}} + q^{ \sum_{j \in S^+_1, w_j = 0} \frac{\countp_j}{2} } & \text{if } |S^-_1|=0 \\
q^{ \sum_{j \in S^+_1, w_j = 0} \frac{\countp_j}{2}} & \text{if } |S^-_1| > 0, w_j = 0 \text{ for all } j \in S^-_1\\
0 & \text{otherwise}
\end{cases}
\end{align}
\begin{remark} \label{even_orbs}
The case of $s=0$ is equivalent to the partition $\lambda$ having only even parts. In such cases, there are two $\mbox{SO}_{2n}(\aF_q)$-orbits corresponding to the same
$\lambda$. In each of these cases, $A(e)$ is trivial and the formula for $|S^-_1|=0$ above is correct if we interpret it as the sum over two elements, $e_1$ and $e_2$, one in each of the two $\mbox{SO}_{2n}(\aF_q)$-orbits: $\frac{1}{|Z_{G^F}(e_1)|} + \frac{1}{|Z_{G^F}(e_2)|}$.
In other words, the value, when multiplied by $|G|$, gives the number of points in the $\mbox{O}_{2n}(\aF_q)$-orbit through either element.
\end{remark}
\subsection{Value of $d^u$}
\label{value-of-d-section}
Recall that $d^u$ is the dimension of a maximal unipotent subgroup of $Z_G(e)$.
Let $d^u_1$ be the dimension of the unipotent radical of $Z_G(e)$ and $d^{u}_2$
the dimension of a maximal unipotent subgroup of the reductive part $Z_{red}$ of $Z_G(e)$.
Then $d^u = d_1^u + d_2^u$.
Since $d^u_2$ is the number of positive roots for $Z_{red}$, we can compute its value
from the known type of $Z_{red}$ given previously. The value of $d^{u}_1$ can be found in \cite[pp. 398-9]{Carter}.
Recall that $\dualp$ is the dual partition for $\lambda$
and that $c(\lambda):=\sum_j \lambda'_j \lambda'_{j+1}$.
\medskip
\noindent {\bf Type} $C$.
\medskip
We have
$$d^{u}_1 = \frac{1}{2} \left( \sum (\dualp_j)^2 - \sum \countp_j^2 + \sum_{j \equiv 0} \countp_j \right) \text{ and }$$
\begin{eqnarray*}
d^{u}_2 =
\sum_{j \equiv 1} \frac{\mu^2_j}{4} + \sum_{\stackrel{j \equiv 0}{ \countp_j \equiv 1}} \frac{(\mu_j-1)^2}{4} +
\sum_{\stackrel{j \equiv 0}{\countp_j \equiv 0}} \left(\frac{\mu^2_j}{4} - \frac{\mu_j}{2} \right)
= \frac{1}{4} \sum \mu_j^2 - \frac{1}{2} \sum_{j \equiv 0} \mu_j+ \frac{ L(\lambda) }{4}, \label{pos_roots_C}
\end{eqnarray*}
where $L(\lambda)$ is the number of $\countp_j$ that are odd.
Hence, since $\lambda'_1 = \ell(\lambda)$,
\begin{eqnarray}
\nonumber
d^u = d^{u}_1 + d^{u}_2 & = &
{ \frac{1}{2} \sum (\dualp_j)^2 - \frac{1}{4} \sum \mu_j^2 + \frac{ L(\lambda) }{4}} \\ \nonumber
& = & { \frac{1}{2} \sum (\dualp_j)^2 - \frac{1}{4} \sum (\dualp_j -\dualp_{j+1})^2 + \frac{ L(\lambda) }{4}} \\
& = & \frac{\ell(\lambda)^2}{4} + \frac{c(\lambda)}{2} + \frac{ L(\lambda) }{4}.
\end{eqnarray}
\medskip
\noindent {\bf Type} $B$ and $D$.
\medskip
We have
$$d^{u}_1 =\frac{1}{2} \left( \sum (\dualp_j)^2 - \sum \countp_j^2 - \sum_{j \equiv 0} \countp_j \right) \text{ and }
d^{u}_2 = \frac{1}{4} \sum \countp_j^2 - \frac{1}{2} \sum_{j \equiv 1} \countp_j + \frac{ L(\lambda) }{4}.$$
Hence, since $\sum \countp_j = \ell(\lambda)$,
\begin{eqnarray}
\nonumber
d^u = d^{u}_1 + d^{u}_2 & = &
\frac{1}{2} \sum (\dualp_j)^2 - \frac{1}{4} \sum \mu_j^2 + \frac{ L(\lambda) }{4} - \frac{\ell(\lambda)}{2} \\
& = & \frac{\ell(\lambda)^2}{4} - \frac{\ell(\lambda)}{2} + \frac{c(\lambda)}{2} + \frac{ L(\lambda) }{4}. \label{BD_d_u}
\end{eqnarray}
\subsection{Finishing the $f_{e, \phi}$ calculation in types $B, C, D$}
\label{BCD_calcs}
We first handle types $B_n$ and $C_n$.
By \cite{LeSh} and \cite{Spalt}, we have
\begin{equation}
\label{type-BC-kappa-formula}
\kappa = \left\lfloor \frac{\ell(\lambda)}{2} \right\rfloor \text{ and }
\end{equation}
$$
(m_1, \dots, m_{\kappa})=(1, 3, \dots, 2\kappa-1).
$$
Moreover, $\pi_{\kappa}$ is always trivial (see \cite{Sommers2}), so $f_{e, \phi}$ is computed by \eqref{formula:f_easy}, which becomes
\begin{equation} \label{BC_almost}
f_{e, \phi} =
q^{m(n - \sBC)+ \sBC^2} \prod_{j=1}^{\sBC} \big (q^{m-(2j-1)} - 1 \big)
\big( \sum_c \frac{ \phi(c)} {|Z_{G^F}(e_c)|} \big).
\end{equation}
We introduce the following notation
\begin{equation} \label{beta}
\beta_{\epsilon}(\lambda, w) := \sum_{\stackrel{ j \equiv \epsilon, \countp_j \equiv 0}{w_j = 0}} \frac{\countp_j}{2}
\end{equation}
and recall from the introduction that
$$
\delta(\lambda) := \sBC^2 - \frac{\ell(\lambda)^2}{4} =
\begin{cases}
0 & \text{ for }\ell(\lambda)\text{ even}, \\
\frac{1}{4}-\frac{\ell(\lambda)}{2} & \text{ for }\ell(\lambda)\text{ odd}.
\end{cases}
$$
and $\hat{N}:=\lfloor N/2 \rfloor$, and
for
$\nu=(\nu_1,\nu_2,\ldots)$ that
$\hat{\nu}:=(\hat{\nu}_1, \hat{\nu}_2, \ldots)$.
\begin{proposition} \label{BC_most_general}
Write $\phi=\phi_w$ as in Section~\ref{compute-summation}.
For $\ggg$ of type $B_n$ or $C_n$, then $f_{e,\phi}$ equals zero unless $w_j = 0$ for all $j \in S^{-}_{\epsilon}$, where $\epsilon = 0$ for type $C$ and
$\epsilon = 1$ for type $B$,
in which case $f_{e,\phi}$ equals
$$
q^{m\left( n-\hat{\ell}(\lambda) \right) + \frac{1}{4} - \frac{c(\lambda)}{2} - \frac{ L(\lambda) }{4} + \beta_{1}(\lambda, w)} \cdot \prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} \text{ (Type $B_n$) }
$$
$$
q^{m\left( n-\hat{\ell}(\lambda) \right) + \delta(\lambda) - \frac{c(\lambda)}{2} - \frac{ L(\lambda) }{4} + \beta_{0}(\lambda, w)} \cdot \prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} \mbox{ (Type $C_n$) }
$$
\end{proposition}
\begin{proof}
Using \eqref{c_sum3} and \eqref{b_sum3}
and the fact that
$$\prod_{j=1}^{\hat{\ell}(\lambda) } \big (q^{m-(2j-1)} - 1 \big) = \frac{\eta(m)}
{\eta \left(m- \ell(\lambda) \right)},$$
the expression in \eqref{BC_almost}, when nonzero, equals
\begin{equation}
f_{e, \phi} = q^{m(n - \hat{\ell}(\lambda) )+ \hat{\ell}(\lambda) ^2 - d^u + \beta_{\epsilon}(\lambda, w)} \cdot \frac{\eta(m)}
{\eta \left(m- \ell(\lambda) \right)
\cdot \eta(\mu_1 ) \cdot \eta(\mu_2) \cdots }.
\end{equation}
Next, we have
\begin{equation} \nonumber
\label{type-BC-multinomial}
\frac{\eta(m)}
{\eta \left(m- \ell(\lambda) \right)
\cdot \eta(\mu_1 ) \cdot \eta(\mu_2) \cdots }
= \prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2},
\end{equation}
since $\hat{\ell}(\lambda) = \hat{L}(\lambda) + |\hat{\mu}(\lambda)|$.
The results follow after substituting in the appropriate value of $d^u$ from \S \ref{value-of-d-section}.
\end{proof}
Recall from the introduction that
$$\tau_{\epsilon}(\lambda) =
\sum_{\substack{j \equiv \epsilon, \countp_j \equiv 0}} \frac{\countp_j}{2}
$$
which is the value of is $\beta_{\epsilon}(\lambda, 0)$ when $w=0$, which corresponds to the trivial character of $A(e)$.
Thus the results in the Proposition
simplify to those in Theorem~\ref{q-Kreweras-formula-theorem} (Types $B_n$ and $C_n$).
\begin{comment}
\begin{corollary}
\begin{equation}
\label{type-B-q-Kreweras}
\Krew(B_n,\0_\lambda,m;q) = q^{m\left( n-\hat{\ell}(\lambda) \right) +
\frac{1}{4} - \frac{c(\lambda)}{2} - \frac{ L(\lambda) }{4} + \tau_1(\lambda)} \cdot \prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} ,
\end{equation}
and
\begin{equation}
\label{type-C-q-Kreweras}
\Krew(C_n,\0_\lambda,m;q) =q^{m\left( n-\hat{\ell}(\lambda) \right) + \delta(\lambda) - \frac{c(\lambda)}{2} - \frac{ L(\lambda) }{4} + \tau_0(\lambda)} \cdot \prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}.
\end{equation}
\end{corollary}
\end{comment}
\medskip
\noindent {\bf Type $D_n$}.
\medskip
\bigskip
We now turn to type $D_n$. Here the values of $m_1, m_2, \dots, m_{\kappa}$ depend on both $\ell(\lambda)$ and the parity of $\mu_1$ \cite{Spalt}:
\begin{itemize}
\item
When $\mu_1$ is odd, $\kappa = \frac{\ell(\lambda)}{2} -1$ and
$$(m_1,m_2,\ldots,m_{\kappa})=(1, 3, \dots, 2\kappa -1).$$
\item
When $\mu_1$ is even, $\kappa = \frac{\ell(\lambda)}{2}$ and
$$
(m_1, m_2, \dots, m_{\kappa})=
\left(1, 3, \dots, 2\kappa -3,
\ell(\lambda) - \frac{\mu_1}{2} -1\right).
$$
\end{itemize}
What complicates the type $D_n$ picture is that $\pi_{\kappa}$ may be non-trivial when $\mu_1$ is even.
It is always trivial when $\mu_1$ is odd.
Let us now describe when this happens and what $\pi_{\kappa}$ is.
To that end, we define a subgroup $H \subset A(e)$. Suppose that $\mu_1$ is even and nonzero. Then
$e$ lies in a proper Levi subalgebra $\levi$ of $\ggg$ of type $D$ of semisimple rank $n - \frac{\mu_1}{2}$.
Now if $\lambda$ contains an odd part different from $1$, then $A(e)$ will be nontrivial and moreover
the component group of $e$ relative to $\levi$ defines an index two subgroup of $A(e)$, which we denote $H$.
We can now recall
\begin{proposition}{\cite[Proposition 10]{Sommers2}} \label{me:pij}
If $\mu_1$ is odd or $\mu_1=0$ or $\mu_j = 0$ for all odd $j >1$ , then $\pi_{\kappa}$ is trivial.
Otherwise, $\pi_{\kappa}$ is the
nontrivial representation of $A(e)$ which is trivial on the subgroup $H$.
\end{proposition}
\begin{comment}
Let
$$g_{\lambda, m}:= \frac{ \displaystyle \prod_{j=1}^{ \hat{\ell}(\lambda) - 1} (q^{m-2j+1} - 1) }
{ \displaystyle \prod_{j} \eta(\countp_j) } =
\begin{cases}
\displaystyle \prod_{j=1}^{\hat{L}(\lambda)-1} (q^{m-2j+1}-1) \cdot
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} & \text{ if } \hat{L}(\lambda) \geq 1 \\
\tfrac{1}{[\hat{m} +1]_{q^2} }\cdot
\qbin{\hat{m} +1}{\hat{\mu}(\lambda)}{q^2} & \text{ if } \hat{L}(\lambda) = 0.
\end{cases}
$$
\end{comment}
We can now give the formula for $f_{e, \phi}$.
\begin{proposition}
\label{D_most_general}
Write $\phi = \phi_w$ as in the Type D subsection of Section~\ref{compute-summation}.
Then $f_{e, \phi} = 0$ unless $w_j = 0$ for all $j \in S^-_1$.
Otherwise,
\begin{comment}
\begin{equation*} \nonumber
f_{e, \phi} = q^{
\BCDexponent
+ \beta_1(\lambda, w)} \cdot g_{\lambda, m}
\text{ multiplied by } \\
\end{equation*}
\begin{eqnarray} \label{Factor}
\begin{cases}
q^{
m + 1 - \hat{\ell}(\lambda)
}
& \text{ if } \mu_1 \text{ is odd} \\
q^{
m + 1 - \hat{\ell}(\lambda)
}
-
q^{
\hat{\ell}(\lambda) + \mu_1(w_1-1)
}
& \text{ if } \mu_1 \text{ is even and } \hat{L}(\lambda) \geq 1 \\
\left(
q^{
m + 1 - \hat{\ell}(\lambda)
}
-
q^{
\hat{\ell}(\lambda) + \mu_1(w_1-1)
}
\right)
+ \\
\hspace{.3in}
q^{\tau_1(\lambda) - 2\beta_1(\lambda,w)}
\left(
q^{
m + 1 - \hat{\ell}(\lambda)
}
-
q^{
\hat{\ell}(\lambda) - \mu_1 w_1
}
\right)
& \text{ if } \hat{L}(\lambda) = 0
\end{cases}
\end{eqnarray}
\end{comment}
\begin{equation*} \nonumber
f_{e, \phi} = q^{
\BCDexponent
+ \beta_1(\lambda, w)}
\text{ multiplied by } \\
\end{equation*}
\begin{eqnarray*} \label{Factor}
\begin{cases}
q^{
m + 1 - \hat{\ell}(\lambda)
} \cdot \displaystyle \prod_{j=1}^{\hat{L}(\lambda)-1} (q^{m-2j+1}-1) \cdot
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}
& \text{ if } \mu_1 \text{ is odd} \\
q^{\hat{\ell}(\lambda) -\mu_1}
\displaystyle \prod_{j=1}^{\hat{L}(\lambda)} (q^{m-2j+1}-1) \cdot
\qbin{\hat{m}-\hat{L}(\lambda)}{\hat{\mu}_{\geq 2}(\lambda)}{q^2}
\qbin{\hat{m}+ 1- \hat{L}(\lambda) -|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)} {q^2}
& \text{ if } \mu_1 \text{ is even, } w_1 = 0, \text{ and } \hat{L}(\lambda) \geq 1 \\
q^{\hat{\ell}(\lambda) }
\displaystyle \prod_{j=1}^{\hat{L}(\lambda)} (q^{m-2j+1}-1) \cdot
\qbin{\hat{m} - \hat{L}(\lambda)}{\hat{\mu} (\lambda)}{q^2}
& \text{ if } \mu_1 \text{ is even, } w_1 = 1, \text{ and } \hat{L}(\lambda) \geq 1 \\
q^{\hat{\ell}(\lambda) + \tau_1(\lambda) - 2\beta_1(\lambda,w)}
\qbin{\hat{m}}{\hat{\mu}(\lambda)}{q^2}
+
q^{\hat{\ell}(\lambda) - \mu_1}
\qbin{\hat{m}}{\hat{\mu}_{\geq 2}(\lambda)}{q^2}
\qbin{\hat{m}+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)} {q^2}
& \text{ if } w_1=0 \text{ and } \hat{L}(\lambda) = 0 \\
\nonumber q^{\hat{\ell}(\lambda) }
\qbin{\hat{m} }{\hat{\mu}(\lambda)}{q^2}
+
q^{\hat{\ell}(\lambda) - \mu_1 + \tau_1(\lambda) -2 \beta_1(\lambda,w)}
\qbin{\hat{m} }{\hat{\mu}_{\geq 2} (\lambda)}{q^2}
\qbin{\hat{m}+1 - |\hat{\mu}_{\geq 2}(\lambda)|}{\hat{\mu}_{1}(\lambda)}{q^2}
& \text{ if } w_1=1 \text{ and } \hat{L}(\lambda) = 0 \\
\end{cases}
\end{eqnarray*}
In fact, since in the last case $w$ can always be taken to have $w_1=0$, we only need the first formula.
\end{proposition}
\begin{proof}
When $\mu_1$ is odd, then $\pi_{\kappa}$ is trivial, so formula \eqref{formula:f} becomes
\begin{eqnarray} \label{formula:d, odd}
\nonumber
f_{e, \phi} & = & q^{m(n - \frac{\ell(\lambda)}{2} + 1)+ ( \frac{\ell(\lambda)}{2} - 1)^2} \prod_{j=1}^{\hat{\ell}(\lambda) - 1} (q^{m-(2j-1)} - 1)
\sum_c \frac{ \phi(c)} {|Z_{G^F}(e_c)|} \\
\nonumber & = & q^{m(n - \frac{\ell(\lambda)}{2} + 1)+ ( \frac{\ell(\lambda)}{2} - 1)^2 }
\frac{\eta(m)} {\eta \left(m+1- \ell(\lambda) \right) } \cdot
\frac{ q^{- d^u + \beta_1(\lambda,w)} }{\eta(\mu_1 ) \cdot \eta(\mu_2) \cdots }
\end{eqnarray}
using the second part of \eqref{d_sum3} since $|S^-_1| > 0$.
The result follows from the formula for $d^u$ in \eqref{BD_d_u}
and the identity
\begin{equation} \label{D_version_multinomial}
\frac{\eta(m)}
{\eta \left(m+1- \ell(\lambda) \right)
\cdot \eta(\mu_1 ) \cdot \eta(\mu_2) \cdots }
=
\begin{cases}
\displaystyle \prod_{j=1}^{\hat{L}(\lambda)-1} (q^{m-2j+1}-1) \cdot
\qbin{\hat{m} - ( \hat{L}(\lambda) -1)}{\hat{\mu}(\lambda)}{q^2} & \text{ if } \hat{L}(\lambda) \geq 1 \\
\tfrac{1}{[\hat{m} +1]_{q^2} }\cdot
\qbin{\hat{m} +1}{\hat{\mu}(\lambda)}{q^2} & \text{ if } \hat{L}(\lambda) = 0,
\end{cases}
\end{equation}
which uses the equality $\hat{\ell}(\lambda) = \hat{L}(\lambda) + |\hat{\mu}(\lambda)|$.
Next when $\mu_1$ is even, formula \eqref{formula:f} becomes
\begin{equation} \label{formula:d, even}
f_{e, \phi} =
q^{m(n - \frac{\ell(\lambda)}{2})+ \frac{\ell(\lambda)^2}{4} - \frac{\mu_1}{2}}
\prod_{j=1}^{\hat{\ell}(\lambda) - 1} (q^{m-(2j-1)} - 1)
\cdot \left( q^{m- (\ell(\lambda) - \frac{\mu_1}{2} - 1)} \sum_c \frac{ \phi(c)} {|Z_{G^F}(e_c)|} - \sum_c \frac{ \pi_{\kappa}(c) \phi(c)} {|Z_{G^F}(e_c)|} \right),
\end{equation}
since $\pi_{\kappa}$ is always one-dimensional.
When $\mu_1$ is even and $\hat{L}(\lambda) \geq 1$,
since $|S^-_1| > 0$, the second part of \eqref{d_sum3} is used to evaluate each sum in \eqref{formula:d, even}.
In this case, $\pi_k$ is nontrivial if and only if $\mu_1$ is nonzero.
For $\phi = \phi_w$, the vector corresponding to $\pi_k \phi_w$ is the same as $w$ but with the parity of $w_1$ changed.
Thus as in the previous case, \eqref{formula:d, even} becomes
\begin{eqnarray}
\label{first_step1}
&\\
\nonumber
q^{m(n - \frac{\ell(\lambda)}{2})+ \frac{\ell(\lambda)^2}{4} - \frac{\mu_1}{2}}
& \displaystyle \prod_{j=1}^{\hat{L}(\lambda)-1} (q^{m-2j+1}-1) \cdot
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} q^{-d^u +\beta_1(\lambda,w)}
\begin{cases}
q^{m +1 - \ell(\lambda) + \frac{\mu_1}{2} } - q^{ - \frac{\mu_1}{2} } & \text{ if } w_1 = 0 \\
q^{m +1 - \ell(\lambda) + \frac{\mu_1}{2} } - q^{ \frac{\mu_1}{2} } & \text{ if } w_1 = 1. \\
\end{cases}
\end{eqnarray}
The following two identities are easy to verify, for $A \in \NN$ and $\nu$ a partition with $|\nu| \leq A$:
$$
\qbin{A}{\nu}{q} = \qbin{A}{\nu_{\geq 2}}{q} \qbin{A - |\nu_{\geq 2}|}{\nu_{1}}{q}
\text{ and }
(q^{A+1 - |\nu|}-1) \qbin{A+1}{\nu}{q} = (q^{A+1}-1) \qbin{A}{\nu}{q}.
$$
Using these identities and the identity
$\hat{L}(\lambda) + |\hat{\mu}_{\geq 2}(\lambda)| = \hat{\ell}(\lambda) - \hat{\mu}_1(\lambda)$,
we have
\begin{eqnarray} \label{w1=0}
\nonumber
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}
\left( q^{m +1 - \ell(\lambda) + \frac{\mu_1}{2} } - q^{ - \frac{\mu_1}{2} } \right) =
q^{- \hat{\mu}_1 }
((q^2)^{\hat{m} + 1 - \hat{\ell}(\lambda) + \hat{\mu}_1}-1)
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} = \\
\nonumber
q^{- \hat{\mu}_1 }
((q^2)^{\hat{m} + 1 - \hat{\ell}(\lambda) + \hat{\mu}_1}-1)
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}_{\geq 2} (\lambda)}{q^2}
\qbin{\hat{m}+1 - \hat{L}(\lambda) - |\hat{\mu}_{\geq 2}(\lambda)|}{\hat{\mu}_{1}(\lambda)}{q^2} = \\
\nonumber q^{- \hat{\mu}_1 }
((q^2)^{\hat{m} + 1 - \hat{L}(\lambda)}-1)
\qbin{\hat{m} - \hat{L}(\lambda)}{\hat{\mu}_{\geq 2} (\lambda)}{q^2}
\qbin{\hat{m}+1 - \hat{L}(\lambda) - |\hat{\mu}_{\geq 2}(\lambda)|}{\hat{\mu}_{1}(\lambda)}{q^2}
\end{eqnarray}
and
\begin{eqnarray} \label{w1=1}
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}
\left( q^{m +1 - \ell(\lambda) + \frac{\mu_1}{2} } - q^{ \frac{\mu_1}{2} } \right) =
\nonumber q^{ \hat{\mu}_1 }
((q^2)^{\hat{m} + 1 - \hat{\ell}(\lambda)}-1)
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} = \\
\nonumber q^{ \hat{\mu}_1 }
((q^2)^{\hat{m} + 1 - \hat{L}(\lambda) - |\hat{\mu}(\lambda)|}-1)
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2} =
\nonumber q^{ \hat{\mu}_1 }
((q^2)^{\hat{m} + 1 - \hat{L}(\lambda)}-1)
\qbin{\hat{m} - \hat{L}(\lambda)}{\hat{\mu} (\lambda)}{q^2}
\end{eqnarray}
and the result follows for this case by inserting these values into \eqref{first_step1}.
Finally when $\mu_1$ is even and $\hat{L}(\lambda) = 0$, we have $S^-_1 = \emptyset$. Thus there are two possible choices for $w$: $w$ and $w + (1,1, \dots, 1)$ will both give the same $\phi_w$. The formulas will not depend on the choice.
The two summations in \eqref{formula:d, even} will both make use of the first part of \eqref{d_sum3}.
The calculation is similar to the previous case after noting that
$\tau_1(\lambda) - \beta_1(\lambda,w) = \sum_{j \in S^+_1, w_j = 1} \frac{\countp_j}{2}.\qedhere$
\begin{comment}
When $\mu_1$ is odd, $f_{e, \phi}$ is computed using \eqref{formula:d, odd}. Since $|S^-_1| > 0$ the second part of \eqref{d_sum3} is used and the result follows immediately from the formula for $d^u$ and the definition of $\beta_1(\lambda,w)$ in \eqref{beta}.
When $\mu_1$ is even, $f_{e, \phi}$ is computed using \eqref{formula:d, even}.
When $\hat{L}(\lambda) \geq 1$,
since $|S^-_1| > 0$, the second part of \eqref{d_sum3} is used twice in the calculation.
Also $\pi_k$ is nontrivial if and only if $\mu_1$ is nonzero. For $\phi = \phi_w$, the vector corresponding to $\pi_k \phi_w$ is the same as $w$ but with the parity of $w_1$ changed.
so
\begin{eqnarray}
\nonumber f_{e, \phi} = q^{m(n - \frac{\ell(\lambda)}{2})+ \frac{\ell(\lambda)^2}{4} - \frac{\mu_1}{2} -d^u +\beta_1(\lambda,w)}
\cdot & \\
\displaystyle \prod_{j=1}^{\hat{L}(\lambda)-1} (q^{m-2j+1}-1) \cdot &
\qbin{\hat{m}+1 - \hat{L}(\lambda)}{\hat{\mu}(\lambda)}{q^2}
\cdot
\begin{cases}
q^{m +1 - \ell(\lambda) + \frac{\mu_1}{2} } - q^{ - \frac{\mu_1}{2} } & \text{ if } w_1 = 0. \\
q^{m +1 - \ell(\lambda) + \frac{\mu_1}{2} } - q^{ \frac{\mu_1}{2} } & \text{ if } w_1 = 1. \\
\end{cases}
\end{eqnarray}
Hence the formula holds and remains valid even if $\mu_1 = 0$.
Finally, when $\mu_1$ is even and $\hat{L}(\lambda) = 0$ there are two possible choices for $w$, but the formulas do not depend on the choice.
The two summations in \eqref{formula:d, even} will both make use of the first part of \eqref{d_sum3}. The result follows
since
$\tau_1(\lambda) - \beta_1(\lambda,w) = \sum_{j \in S^+_1, w_j = 1} \frac{\countp_j}{2}.$
\end{comment}
\end{proof}
\noindent
At $w=0$, Proposition~\ref{D_most_general} is
Theorem~\ref{q-Kreweras-formula-theorem} (Type $D_n$),
since $\beta_1(\lambda,w) = \tau_1(\lambda)$ and
$\tau_1(\lambda) - 2\beta_1(\lambda,w) = -\tau_1(\lambda)$.
\begin{remark}
When $\lambda$ has only even parts, then in particular $\mu_1 = 0$ and $\hat{L}(\lambda) =0$ and $\tau_1(\lambda)=0$.
We are in the last case. Then the expression in the Corollary simplifies to
$
2 q^{
\BCDexponent
+ \hat{\ell}(\lambda)} \qbin{\hat{m}}{\hat{\mu}(\lambda)}{q^2}.
$
This value is actually twice the value of $f_{e,1}$ for $e$ in either nilpotent orbit associated to $\lambda$. See Remark \ref{even_orbs}.
\end{remark}
\subsection{The $f_{e, \phi}$ for the exceptional groups}
\label{Exceptional_calcs}
The polynomials $f_{e, \phi}$ are listed in the third column of the following tables. The first column is the Bala-Carter notation
for the nilpotent orbit $\0_e$ together with $\phi$, if non-trivial. All $A(e)$ are symmetric groups and
we denote $\phi$ by the corresponding partition for an irreducible representation of a symmetric group, where $[1^k]$ is the sign representation.
Recall that an orbit is principal in a Levi subalgebra when there are no parentheses in the Bala-Carter notation. The letters in the notation denote the semisimple part of the Levi subalgebra. In the second column are the representation exponents
$m_i$.
Exponents are listed according to the value of $\pi_i$, so that if $V$ occurs in the $\phi$-isotypic component, it is listed in the row of $(e, \phi)$. We abbreviate $[a]_q$ by $[a]$ in the last column of the tables.
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$G_{2}$} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$0$ & $1,5$ & $\frac{[m-1] [m-5]}{ [2] [6] } $ \\\hline
$A_1$ & $1$ & $q^{m - 5}\frac{ [m-1]}{ [2] } $ \\ \hline
$\Tilde{A}_{1}$ & $1$ & $q^{m - 3}\frac{ [m-1] }{ [2] } $ \\ \hline
$G_{2}(a_{1})$ & $1$ & $(q^{m-1}-1) \cdot q^{m - 3}$ \\ \hline
$G_{2}(a_{1}), [2,1]$ & $-$ & $0$ \\ \hline
$G_{2}(a_{1}), [1^3]$ & $-$ & $0$ \\ \hline
$G_{2}$ & $-$ & $q^{2m - 2}$ \\ \hline
\end{tabular}
\vskip.2in
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$F_{4}$} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$0$ & $1,5,7,11$ &$\frac{[m-1][m-5][m-7][m-11]}{ [2][6][8][12]} $ \\\hline
$A_1$ & $1,5,7$ & $q^{m - 11}\frac{ [m-1][m-5][m-7]}{ [2][4][6]} $ \\ \hline
$\Tilde{A}_{1}$ & $1,5,7$ & $q^{m - 5}\frac{ [m-1][m-5][m-7]}{ [2][4][6]} $ \\ \hline
$\Tilde{A}_{1}$, $[1^2]$ & $-$ & $q^{m - 8}\frac{[m-1][m-5][m-7]}{[2][4][6]} $ \\ \hline
$A_{1} + \Tilde{A_{1}}$ & $1,5$ & $q^{2m - 14}\frac{ [m-1][m-5]}{[2][2]} $ \\ \hline
$A_2$ & $1,5$ &$q^{2m -8}\frac{ [m-1][m-5]}{ [2][6]} $ \\ \hline
$A_2$, $[1^2]$ & $-$ &$q^{2m -11}\frac{ [m-1][m-5]}{[2][6]} $ \\ \hline
$\Tilde{A}_{2}$ & $1,5$ & $q^{2m -8 }\frac{ [m-1][m-5]}{[2][6]} $ \\ \hline
$A_{2} + \Tilde{A}_{1}$ & $1$ & $q^{3m - 15} \frac{[m-1]}{[2]} $ \\ \hline
$B_2$ & $1,3$ & $q^{2m -6 }\frac{ [m-1][m-3]}{[2][4]} $ \\ \hline
$B_2$, $[1^2]$ & $-$ & $q^{2m -8 }\frac{ [m-1][m-3]}{[2][4]} $ \\ \hline
$\Tilde{A}_{2} + A_{1}$ & $1$ & $q^{3m - 13} \frac{[m-1]}{[2]} $ \\ \hline
$C_{3}(a_{1})$ & $1,3$ & $(q^{m-1} - 1) \cdot q^{2m - 8}\frac{[m-3]}{[2]} $ \\ \hline
$C_{3}(a_{1})$, $[1^2]$ & $-$ & $0$ \\ \hline
$F_{4}(a_{3})$ & $1$ & $(q^{m-1} - 1) \cdot q^{3m- 11}$ \\ \hline
$F_{4}(a_{3}), [3,1]$ & $-$ & 0 \\ \hline
$F_{4}(a_{3}), [2,2]$ & $3$ & $(q^{m-1} - 1) \cdot(-q^{2m- 8})$ \\ \hline
$F_{4}(a_{3}), [2,1^2]$ & $-$ & 0 \\ \hline
$B_3$ & $1$ & $q^{3m -7} \frac{[m-1]}{[2]}$ \\ \hline
$C_3$ & $1$ & $q^{3m-7} \frac{[m-1]}{[2]}$\\ \hline
$F_{4}(a_{2})$ & $1$ & $(q^{m-1} -1) \cdot q^{3m - 7}$ \\ \hline
$F_{4}(a_{2})$, $[1^2]$ & $-$ & $0$ \\ \hline
$F_{4}(a_{1})$ & $1$ & $(q^{m-1} -1) \cdot q^{3m - 5}$ \\ \hline
$F_{4}(a_{1})$, $[1^2]$ & $-$ & $0$ \\ \hline
$F_{4}$ & $-$ & $q^{4m-4}$ \\ \hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{6}$} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$0$ & $1,4,5,7,8,11$ & $\frac{ [m-1] [m-4][m-5][m-7][m-8][m-11]}
{ [2][5][6][8] [9][12] } $ \\\hline
$A_1$ & $1,4,5,7,8$ & $q^{m-11}\frac{ [m-1] [m-4][m-5][m-7][m-8]}
{ [2][3][4][5] [6]} $ \\ \hline
$2A_{1}$ & $1,4,5,7$ & $q^{2m - 16}\frac{ [m-1] [m-4][m-5][m-7]}
{ [1][2][4][6]} $ \\ \hline
$3A_{1}$ & $1,4,5$ & $q^{3m - 21}\frac{ [m-1] [m-4][m-5]}
{ [2][2][3]} $ \\ \hline
$A_2$ & $1,4,5$ &$q^{2m -9}\frac{ [m-1] [m-4][m-5]([m-7]+q[m-3])}
{ [2][3][4][6]}$ \\ \hline
$A_2$, $[1^2]$ & $5$ &$q^{2m -11}\frac{ [m-1] [m-4][m-5]([m-3]+q^5[m-7])}{ [2][3][4][6] }$ \\ \hline
$A_{2} + A_{1}$ & $1,4,5$& $q^{3m -16}\frac{ [m-1] [m-4][m-5]}{ [1][2][3]} $ \\ \hline
$2A_{2}$& $1,5$ & $q^{4m -16}\frac{ [m-1][m-5]}{ [2][6]} $ \\ \hline
$A_{2} + 2A_{1}$ & $1,4$ & $q^{4m-20}\frac{ [m-1] [m-4]}{[1][2]} $ \\ \hline
$A_{3}$ & $1,3,4$& $q^{3m-11}\frac{ [m-1] [m-3][m-4]}{[1][2][4]} $ \\ \hline
$2A_{2} + A_{1}$ & $1$ & $q^{5m-21}\frac{[m-1]}{[2]} $ \\ \hline
$A_{3}+A_1$ & $1,3$& $q^{4m-15}\frac{ [m-1] [m-3]}{[1][2]} $ \\ \hline
$D_{4}(a_1)$ & $1$ & $(q^{m-1}-1) \cdot q^{3m-11 }\frac{[m-1][m-2]}{[2][3]} $ \\ \hline
$D_{4}(a_1), [2,1]$ & $3$ & $(q^{m-1}-1) \cdot q^{3m-10 }\frac{[m-2][m-4]}{[1][3]} $ \\ \hline
$D_{4}(a_1), [1^3]$ & $-$ & $(q^{m-1}-1) \cdot q^{3m-8}\frac{[m-4][m-5]}{[2][3]} $ \\ \hline
$A_4$ & $1,3$& $q^{4m-11} \frac{[m-1][m-3]}{[1][2]}$\\ \hline
$D_4$ & $1,2$ & $q^{4m-10} \frac{[m-1][m-2]}{[2][3]}$\\ \hline
$A_4+A_1$ & $1$ & $q^{5m-14} \frac{[m-1]}{[1]}$\\ \hline
$A_5$ & $1$ & $q^{5m-11} \frac{[m-1]}{[2]}$\\ \hline
$D_5(a_1)$ & $1,2$ & $(q^{m-1} - 1) \cdot q^{4m-10} \frac{[m-2]}{[1]}$\\ \hline
$E_{6}(a_{3})$ & $1$& $(q^{m-1} -1) \cdot q^{5m - 11}$ \\ \hline
$E_{6}(a_{3})$, $[1^2]$ & $2$ & $(q^{m-1} -1) \cdot (-q^{4m - 9})$ \\ \hline
$D_{5}$ & $1$& $q^{5m - 8} \frac{[m-1]}{[1]}$ \\ \hline
$E_{6}(a_{1})$ & $1$ & $(q^{m-1} -1) \cdot q^{5m - 7}$ \\ \hline
$E_{6}$ & $-$ & $q^{6m-6}$ \\ \hline
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{7}$} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$0$ & $1,5,7,9,11,13,17$ &
$\prod_{i=1}^7 \frac{[m-m_i]}{[m_i+1]}$ \\\hline
$A_1$ & $1,5,7,9,11,13$ & $q^{m-17}\frac{ [m-1] [m-5][m-7][m-9][m-11][m-13]}
{ [2][4][6][6] [8][10]} $ \\ \hline
$2A_{1}$ & $1,5,7,9,11$ & $q^{2m-26 }\frac{ [m-1] [m-5][m-7][m-9][m-11]}
{ [2][2][4][6] [8]} $ \\ \hline
$(3A_{1})''$ & $1,5,7,11$ & $q^{3m-27 }\frac{ [m-1] [m-5][m-7][m-11]}
{ [2][6][8][12]} $ \\ \hline
$(3A_{1})'$ & $1,5,7,9$ & $q^{3m-33 }\frac{ [m-1] [m-5][m-7][m-9]}
{ [2][2][4][6]} $ \\ \hline
$A_2$ & $1,5,7,9$ & $q^{2m - 14}\frac{ [m-1] [m-5][m-7][m-9]([m-3]+q^2[m-13])}{ [2][4][6][6][10] }$ \\ \hline
$A_2$, $[1^2]$ & $8$ &$q^{2m - 17}\frac{ [m-1] [m-5][m-7][m-9] ([m-3]+q^8[m-13]) }{ [2][4][6][6][10] }$ \\ \hline
$4A_{1}$ & $1,5,7$ & $q^{4m- 38}\frac{ [m-1] [m-5][m-7]}{ [2][4][6]} $ \\ \hline
$A_{2} + A_{1}$ & $1,5,7$ &
$q^{3m - 25 }\frac{ [m-1] [m-5][m-7](q^2[m-7]+ [m-9])}{ [2][2][4][6] }$ \\ \hline
$A_{2} + A_{1}$, $[1^2]$ & $8$ &
$q^{3m - 26 }\frac{ [m-1] [m-5][m-7]([m-7]+ q^4[m-9])}{ [2][2][4][6] }$ \\ \hline
$A_{2} + 2A_{1}$ & $1,5,7$ & $q^{4m-32} \frac{ [m-1] [m-5][m-7]}{ [2]^3} $ \\ \hlin
$2A_{2}$ & $1,5,7$ & $q^{4m-26} \frac{ [m-1] [m-5][m-7]}{ [2]^2[6]} $ \\ \hlin
$A_{2} + 3A_{1}$ & $1,5$ & $q^{5m-35}\frac{ [m-1] [m-5]}{ [2][6]} $ \\ \hlin
$A_{3}$ & $1,5,5,7$ & $q^{3m-17}\frac{ [m-1] [m-5][m-5][m-7]}{ [2]^2[4][6]} $ \\ \hlin
$(A_{3}+A_1)''$ & $1,5,7$ & $q^{4m-22}\frac{ [m-1] [m-5][m-7]}{ [2][4][6]} $ \\ \hlin
$2A_{2} + A_{1}$ & $1,5$ & $q^{5m-33}\frac{ [m-1] [m-5]}{ [2]^2} $ \\ \hlin
$(A_{3}+A_1)'$ & $1,5,5$ & $q^{4m-24}\frac{ [m-1] [m-5][m-5]}{ [2]^3} $ \\ \hlin
$D_{4}(a_1)$ & $1,5$ &
$(q^{m-1}-1) \cdot q^{3m-17 }\frac{ [m-1][m-3][m-5]}{[2][4][6]} $ \\ \hline
$D_{4}(a_1)$, $[2,1]$ & $5^{}$ & $(q^{m-1}-1) \cdot q^{3m-15 }\frac{[m-3][m-5][m-7]}{[2]^2[6]} $ \\ \hline
$D_{4}(a_1)$, $[1^3]$ & $-$ & $(q^{m-1}-1) \cdot q^{3m-11 }\frac{[m-5][m-7][m-9]}{[2][4][6]} $ \\ \hline
$A_{3}+2A_1$ & $1,5$ & $q^{5m-29}\frac{ [m-1] [m-5]}{ [2]^2} $ \\ \hline
$D_{4}(a_1)+A_1$ & $1,5$ & $(q^{m-1}-1) \cdot q^{4m-22} \frac{[m-3][m-5]}{[2][4]}$ \\ \hline
$D_{4}(a_1)+A_1$, $[1^2]$ & $5$ & $(q^{m-1}-1) \cdot q^{4m-20} \frac{[m-5][m-7]}{[2][4]}$ \\ \hline
$D_4$ & $1,3,5$ & $q^{4m-16} \frac{[m-1][m-3][m-5]}{[2][4][6]}$\\ \hline
$A_3+A_2$ & $1,5$ & $q^{5m -25} \frac{[m-1][m-5]}{[2]^2}$\\ \hline
$A_3+A_2$, $[1^2]$ & $-$ & $q^{5m -26} \frac{[m-1][m-5]}{[2]^2}$\\ \hline
$A_4$ & $1,5$ & $q^{4m -16} \frac{[m-1][m-5](q^2 [m-3]+[m-5])}{[2]^2 [6]}$ \\ \hline
$A_4$, $[1^2]$ & $4$ & $q^{4m -17} \frac{[m-1][m-5]([m-3]+q^4[m-5])}{[2]^2 [6]}$ \\ \hline
$A_3+A_2+A_1$ & $1$ & $q^{6m-30} \frac{[m-1]}{[2]} $\\ \hline
$A_5''$ & $1,5$ & $q^{5m-17} \frac{[m-1][m-5]}{[2][6]} $\\ \hline
$D_4+A_1$ & $1,3$ & $q^{5m-21} \frac{[m-1][m-3]}{[2][4]} $\\ \hline
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{7}$, part 2} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$A_4+A_1$ & $1$ & $q^{5m-21} \frac{[m-1]}{[2]} \cdot \frac{[m-3] + [m-5]}{[2]}$\\ \hline
$A_4+A_1$, $[1^2]$ & $4$ & $q^{5m-22} \frac{[m-1]}{[2]} \cdot \frac{[m-3]+q^2[m-5]}{[2]}$\\ \hline
$D_5(a_1)$ & $1,3$ & $(q^{m-1}-1) \cdot q^{4m-16} \frac{[m-3]^2}{[2]^2}$\\ \hline
$D_5(a_1)$, $[1^2]$ & $4$ & $(q^{m-1}-1) \cdot q^{4m-15} \frac{[m-3][m-5]}{[2]^2} $\\ \hline
$A_4+A_2$ & $1$ & $q^{6m-24} \frac{[m-1]}{[2]}$\\ \hline
$A_5'$ & $1,3$ & $q^{5m-17} \frac{[m-1][m-3]}{[2]^2}$\\ \hline
$D_5(a_1)+A_1$ & $1,3$ & $(q^{m-1}-1) \cdot q^{5m-19} \frac{[m-3]}{[2]}$\\ \hline
$A_5+A_1$ & $1$ & $q^{6m-22} \frac{[m-1]}{[2]}$\\ \hline
$D_6(a_2)$ & $1,3$ & $(q^{m-1}-1) \cdot q^{5m-17} \frac{[m-3]}{[2]}$\\ \hline
$E_{6}(a_{3})$ & $1,3$ & $(q^{m-1}-1) \cdot q^{5m - 17} \frac{[m-3]}{[2]}$ \\ \hline
$E_{6}(a_{3})$, $[1^2]$ & $3$ & $(q^{m-1}-1) (-q^{4m -14}) \frac{[m-3]}{[2]}$ \\ \hline
$E_{7}(a_{5})$ & $1$ &
$(q^{m-1}-1) \cdot q^{6m-20 }$ \\ \hline
$E_{7}(a_{5})$, $[2,1]$ & $3$ &
$(q^{m-1}-1) \cdot (-q^{5m-17 })$ \\ \hline
$E_{7}(a_{5})$, $[1^3]$ & $-$ &
$(q^{m-1}-1) \cdot q^{4m-14 }$ \\ \hline
$D_{5}$ & $1,3$ & $q^{5m - 13} \frac{[m-1][m-3]}{[2][2]}$ \\ \hline
$A_{6}$ & $1$ & $q^{6m - 16} \frac{[m-1]}{[2]}$ \\ \hline
$D_{5}+A_1$ & $1$ & $q^{6m-16} \frac{[m-1]}{[2]}$ \\ \hline
$D_6(a_1)$ & $1,3$ & $(q^{m-1}-1) \cdot q^{5m -13} \frac{[m-3]}{[2]}$\\ \hline
$E_{7}(a_{4})$ & $1$ & $(q^{m-1} -1) \cdot q^{6m -16}$ \\ \hline
$E_{7}(a_{4})$, $[1^2]$ & $-$ & 0 \\ \hline
$E_{6}(a_{1})$ & $1$ & $(q^{m-1} -1) \cdot q^{5m - 11} \frac{[m-1]}{[2]} $ \\ \hline
$E_{6}(a_{1})$, $[1^2]$ & $2$ & $(q^{m-1} -1) \cdot q^{5m - 10} \frac{[m-3]}{[2]}$ \\ \hline
$D_{6}$ & $1$ & $q^{6m - 12} \frac{[m-1]}{[2]}$\\ \hline
$E_{7}(a_{3})$ & $1$ & $(q^{m-1} -1) \cdot q^{6m - 12}$ \\ \hline
$E_{7}(a_{3})$, $[1^2]$ & $2$ & $(q^{m-1} -1) \cdot (-q^{5m - 10})$ \\ \hline
$E_{6}$ & $1$ & $q^{6m - 10} \frac{[m-1]}{[2]}$\\ \hline
$E_{7}(a_{2})$ & $1$ & $(q^{m-1} -1) \cdot q^{6m - 10}$ \\ \hline
$E_{7}(a_{1})$ & $1$ & $(q^{m-1} -1) \cdot q^{6m - 8}$ \\ \hline
$E_{7}$ & $-$ & $q^{7m-7}$ \\ \hline
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{8}$} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$0$ & $1,7,11,13,17,19, 23,29$ &
$\prod_{i=1}^8 \frac{ [m-m_i]}{[m_i+ 1]}$ \\ \hline
$A_1$ & $1,7,11,13,17, 19,23$ & $q^{m-29}
\frac{ [m-1] [m-7][m-11][m-13][m-17][m-19][m-23]}
{ [2][6][8][10] [12][14][18]}$
\\ \hline
$2A_{1}$ & $1,7,11,13, 17,19$ & $q^{2m-46 }
\frac{ [m-1] [m-7][m-11][m-13][m-17][m-19]}
{ [2][4][6][8] [10][12] } $ \\ \hline
$3A_{1}$ & $1,7,11,13,17$ & $q^{3m-57 }
\frac{ [m-1] [m-7][m-11][m-13][m-17]}
{ [2]^2[6][8] [12]} $ \\ \hline
$A_2$ & $1,7,11,13,17$ & $q^{2m - 24 }\frac{ [m-1][m-7][m-11][m-13][m-17]
([m-5]+q^4 [m-23])}
{ [2][6][8][10][12][18]}$ \\ \hline
$A_2$, $[1^2]$ & $14$ & $q^{2m - 29 }\frac{ [m-1][m-7][m-11][m-13][m-17]
([m-5] + q^{14} [m-23])}
{ [2][6][8][10][12][18]}$ \\ \hline
$4A_{1}$ & $1,7,11,13$ & $q^{4m- 68} \frac{ [m-1] [m-7][m-11][m-13]}
{ [2][4][6] [8]} $ \\ \hline
$A_{2} + A_{1}$ & $1,7,11,13$ &
$q^{3m - 43}
\frac{ [m-1][m-7][m-11][m-13] (q^2 [m-11] +[m-17])}
{ [2][4][6]^2[10]}$ \\ \hline
$A_{2} + A_{1}$, $[1^2]$ & $14$ &
$q^{3m - 46}
\frac{ [m-1][m-7][m-11][m-13]([m-11] + q^8 [m-17])}
{ [2][4][6]^2[10]}$
\\ \hline
$A_{2} + 2A_{1}$ & $1,7,11,13$ & $q^{4m-56} \frac{ [m-1] [m-7][m-11][m-13]}
{ [2]^2[4][6] } $ \\ \hline
$A_{3}$ & $1,7,9,11,13$ & $q^{3m-29}\frac{ [m-1] [m-7][m-9][m-11][m-13]}
{ [2][4][6][8][10]} $ \\ \hline
$A_{2} + 3A_{1}$ & $1,7,11$ & $q^{5m-65}\frac{ [m-1] [m-7][m-11]}{ [2]^2[6]} $ \\ \hline
$2A_{2}$ & $1,7,11$ & $q^{4m-44} \frac{ [m-1] [m-7][m-11]([m-5]+ q^4 [m-17])}
{ [2][4][6][12]} $ \\ \hline
$2A_{2}$, $[1^2]$ & $11$ & $q^{4m- 46} \frac{ [m-1] [m-7][m-11]([m-5]+ q^8 [m-17])}
{ [2][4][6][12]} $ \\ \hline
$2A_{2} + A_{1}$ & $1,7,11$ & $q^{5m-57}\frac{ [m-1] [m-7][m-11]}{ [2]^2[6]} $ \\ \hline
$A_{3}+A_1$ & $1,7,9,11$ & $q^{4m-42}\frac{ [m-1] [m-7][m-9][m-11]}
{ [2]^2[4][6]} $ \\ \hline
$D_{4}(a_1)$ & $1,7,11$ &
$(q^{m-1}-1) \cdot q^{3m-29 }\frac{[m-1][m-5][m-7][m-11]}{[2][6][8][12]} $ \\ \hline
$D_{4}(a_1)$, $[2,1]$ & $9^{}$ & $(q^{m-1}-1) \cdot q^{3m-25 }\frac{[m-5][m-7][m-11][m-13]}{[2][4][6][12]} $ \\ \hline
$D_{4}(a_1)$, $[1^3]$ & $-$ & $(q^{m-1}-1) \cdot q^{3m-17}\frac{[m-7][m-11][m-13][m-17]}{[2][6][8][12]} $ \\ \hline
$D_4$ & $1,5,7,11$ & $q^{4m-28} \frac{[m-1][m-5][m-7][m-11]}{[2][6][8][12]}$\\ \hline
$2A_{2} + 2A_{1}$ & $1,7$ & $q^{6m-66}\frac{ [m-1][m-7]}{[2] [4]} $ \\ \hline
$A_{3}+2A_1$ & $1,7,9$ & $q^{5m-51}\frac{ [m-1][m-7][m-9]}{[2]^2 [4]} $ \\ \hline
$D_{4}(a_1)+A_1$ & $1,7$ & $(q^{m-1}-1) \cdot q^{4m-40} \frac{[m-5][m-7]^2}{[2][4][6]}$ \\ \hline
$D_{4}(a_1)+A_1, [2,1]$ & $9^{}$ & $(q^{m-1}-1) \cdot q^{4m-38} \frac{[m-5][m-7][m-11]}{[2]^2[6]}$\\ \hline
$D_{4}(a_1)+A_1, [1^3]$ & $-$ & $(q^{m-1}-1) \cdot q^{4m-34} \frac{[m-7][m-11][m-13]}{[2][4][6]}$ \\ \hline
$A_3+A_2$ & $1,7,9$ & $q^{5m -45} \frac{[m-1][m-7][m-9]}{[2]^2 [4]}$\\ \hline
$A_3+A_2$, $[1^2]$ & $-$ & $q^{5m -46} \frac{[m-1][m-7][m-9]}{[2]^2 [4]}$ \\ \hline
$A_4$ & $1,7,9$ & $q^{4m -26} \frac{[m-1][m-7][m-9](q^2[m-5] + [m-11])}
{[2][4][6][10]}$ \\ \hline
$A_4$, $[1^2]$ & $8$ & $q^{4m -29} \frac{[m-1][m-7][m-9]([m-5] + q^8[m-11])}
{[2][4][6][10]}$ \\ \hline
$A_3+A_2+A_1$ & $1,7$ & $q^{6m-54} \frac{[m-1][m-7]}{[2]^2} $\\ \hline
$D_4+A_1$ & $1,5,7$ & $q^{5m-39} \frac{[m-1][m-5][m-7]}{[2][4][6]} $ \\ \hline
$D_{4}(a_1)+A_2$ & $1,7$ & $(q^{m-1} -1) \cdot q^{5m- 43} \frac{[m-5][m-7]}{[2][6]} $ \\ \hline
$D_{4}(a_1)+A_2$, $[1^2]$ & $8$ & $(q^{m-1} -1) \cdot q^{5m-40}\frac{[m-7][m-11]}{[2][6]} $ \\ \hline
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{8}$, part 2} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$A_4+A_1$ & $1,7$ & $q^{5m-37} \frac{[m-1][m-7] ([m-5] + q^2 [m-11])}
{[2]^2[6]}$ \\ \hline
$A_4+A_1$, $[1^2]$ & $8$ & $q^{5m-38} \frac{[m-1][m-7]([m-5] + q^4 [m-11])}
{[2]^2 [6]}$ \\ \hline
$2A_3$ & $1,7$ & $q^{6m-46} \frac{[m-1][m-7]}{[2][4]}$\\ \hline
$D_5(a_1)$ & $1,5,7$ & $(q^{m-1} -1) \cdot q^{4m-28} \frac{[m-5]^2[m-7]}{[2][4][6]} $\\ \hline
$D_5(a_1)$, $[1^2]$ & $8$ & $(q^{m-1} -1) \cdot q^{4m-25} \frac{[m-5][m-7][m-11]}{[2][4][6]}$\\ \hline
$A_4+2A_1$ & $1,7$ & $q^{6m-44} \frac{[m-1][m-7]}{[2]^2}$ \\ \hline
$A_4+2A_1$, $[1^2]$ & $-$ & $q^{6m-45} \frac{[m-1][m-7]}{[2]^2}$ \\ \hline
$A_4+A_2$ & $1,7$ & $q^{6m-42} \frac{[m-1][m-7]}{[2]^2}$\\ \hline
$A_5$ & $1,5,7$ & $q^{5m-29} \frac{[m-1][m-5][m-7]}{[2]^2[6]}$\\ \hline
$D_5(a_1)+A_1$ & $1,5,7$ & $(q^{m-1} -1) \cdot q^{5m-35} \frac{[m-5][m-7]}{[2]^2}$ \\ \hline
$A_4 + A_2+A_1$ & $1$ & $q^{7m-49} \frac{[m-1]}{[2]}$\\ \hline
$D_4 + A_2$ & $1,5$ & $q^{6m-36} \frac{[m-1][m-5]}{[2][6]}$ \\ \hline
$D_4 + A_2$, $[1^2]$ & $-$ & $q^{6m-39} \frac{[m-1][m-5]}{[2][6]}$ \\ \hline
$E_{6}(a_{3})$ & $1,5,7$ & $(q^{m-1} -1) \cdot q^{5m - 29} \frac{[m-5][m-7]}{[2][6]}$ \\ \hline
$E_{6}(a_{3})$, $[1^2]$ & $5$ & $(q^{m-1} -1)\cdot (-q^{4m -24}) \frac{[m-5][m-7]}{[2][6]}$ \\ \hline
$D_{5}$ & $1,5,7$ & $q^{5m - 23} \frac{[m-1][m-5][m-7]}{[2][4][6]}$ \\ \hline
$A_4+A_3$ & $1$ & $q^{7m-45} \frac{[m-1]}{[2]}$\\ \hline
$A_5+A_1$ & $1,5$ & $q^{6m-36} \frac{[m-1][m-5]}{[2]^2}$\\ \hline
$D_5(a_1)+A_2$ & $1,5$ & $(q^{m-1}-1) \cdot q^{6m-38} \frac{[m-5]}{[2]}$\\ \hline
$D_6(a_2)$ & $1,5$ & $(q^{m-1}-1) \cdot q^{5m-29} \frac{[m-3][m-5]}{[2][4]}$\\ \hline
$D_6(a_2)$, $[1^2]$ & $5$ & $(q^{m-1}-1) \cdot q^{5m-27} \frac{[m-5][m-7]}{[2][4]}$\\ \hline
$E_{6}(a_{3})+A_1$ & $1,5$ & $(q^{m-1}-1) \cdot q^{6m - 36} \frac{[m-5]}{[2]}$ \\ \hline
$E_{6}(a_{3})+A_1$, $[1^2]$ & $5$ & $(q^{m-1}-1) \cdot (-q^{5m -31}) \frac{[m-5]}{[2]}$ \\ \hline
$E_{7}(a_{5})$ & $1,5$ & $(q^{m-1}-1) \cdot q^{6m-34} \frac{[m-5]}{[2]}$ \\ \hline
$E_{7}(a_{5})$, $[2,1]$ & $5$ &
$(q^{m-1} - 1)\cdot (-q^{5m-29 })\frac{[m-5]}{[2]}$ \\ \hline
$E_{7}(a_{5})$, $[1^3]$ & $-$ &
$(q^{m-1} - 1) \cdot q^{4m-24 }\frac{[m-5]}{[2]}$ \\ \hline
$D_{5}+A_1$ & $1,5$ & $q^{6m-30} \frac{[m-1][m-5]}{[2]^2}$ \\ \hline
$E_{8}(a_{7})$ & $1$ & $(q^{m-1} -1) \cdot q^{7m - 39}$ \\ \hline
$E_{8}(a_{7}), [4,1]$ & $5$ & $(q^{m-1} -1) \cdot (-q^{6m - 34})$ \\ \hline
$E_{8}(a_{7}), [3,2]$ & $-$ & $0$ \\ \hline
$E_{8}(a_{7}), [3,1^2]$ & $-$ & $(q^{m-1} -1) \cdot q^{5m - 29}$ \\ \hline
$E_{8}(a_{7}), [2^2,1]$ & $-$ & $0$ \\ \hline
$E_{8}(a_{7}), [2,1^3]$ & $-$ & $(q^{m-1} -1) \cdot (-q^{4m - 24})$ \\ \hline
$A_{6}$ & $1,5$ & $q^{6m - 28} \frac{[m-1][m-5]}{[2]^2}$ \\ \hline
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{8}$, part 3} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$D_6(a_1)$ & $1,5$ & $(q^{m-1} - 1) \cdot q^{5m -23} \frac{[m-3][m-5]}{[2][4]}$\\ \hline
$D_6(a_1)$, $[1^2]$ & $5$ & $(q^{m-1} - 1) \cdot q^{5m -21} \frac{[m-5][m-7]}{[2][4]}$\\ \hline
$A_{6}+A_1$ & $1$ & $q^{7m - 33} \frac{[m-1]}{[2]}$ \\ \hline
$E_{7}(a_{4})$ & $1,5$ & $(q^{m-1} -1) \cdot q^{6m - 28} \frac{[m-5]}{[2]} $ \\ \hline
$E_{7}(a_{4})$, $[1^2]$ & $-$ & 0 \\ \hline
$E_{6}(a_{1})$ & $1,5$ & $(q^{m-1} -1) \cdot q^{5m - 19} \frac{[m-1][m-5]}{[2][6]} $ \\ \hline
$E_{6}(a_{1})$, $[1^2]$ & $4$ & $(q^{m-1} -1) \cdot q^{5m -16} \frac{[m-5][m-7]}{[2][6]}$ \\ \hline
$D_{5}+A_2$ & $1$ & $q^{7m-31} \frac{[m-1]}{[2]}$ \\ \hline
$D_{5}+A_2$, $[1^2]$ & $-$ & $q^{7m-32} \frac{[m-1]}{[2]}$ \\ \hline
$D_{6}$ & $1,3$ & $q^{6m -22} \frac{[m-1][m-3]}{[2][4]}$\\ \hline
$E_{6}$ & $1,5$ & $q^{6m - 18} \frac{[m-1][m-5]}{[2][6]}$\\ \hline
$D_{7}(a_{2})$ & $1$ & $(q^{m-1} -1) \cdot q^{6m-26} \frac{[m-3]}{[2]}$ \\ \hline
$D_{7}(a_{2})$, $[1^2]$ & $4$ & $(q^{m-1} -1) \cdot q^{6m-25} \frac{[m-5]}{[2]}$ \\ \hline
$A_{7}$ & $1$ & $q^{7m - 27} \frac{[m-1]}{[2]}$\\ \hline
$E_{6}(a_{1})+A_1$ & $1$ & $(q^{m-1} -1) \cdot q^{6m-24} \frac{[m-3]}{[2]} $ \\ \hline
$E_{6}(a_{1})+A_1$, $[1^2]$ & $4$ & $(q^{m-1} -1) \cdot q^{6m-23} \frac{[m-5]}{[2]} $ \\ \hline
$E_{7}(a_{3})$ & $1,3$ & $(q^{m-1} - 1) \cdot q^{6m-22} \frac{[m-3]}{[2]}$\\ \hline
$E_{7}(a_{3})$, $[1^2]$ & $4$ & $(q^{m-1} - 1) \cdot (-q^{5m-18}) \frac{[m-3]}{[2]}$\\ \hline
$E_{8}(b_{6})$ & $1$ & $(q^{m-1} - 1) \cdot q^{7m-27}$\\ \hline
$E_{8}(b_{6})$, $[2,1]$ & $-$ & $0$ \\ \hline
$E_{8}(b_{6})$, $[1^3]$ & $-$ & $0$ \\ \hline
$D_7(a_1)$ & $1,3$ & $(q^{m-1} - 1) \cdot q^{6m - 20} \frac{[m-3]}{[2]}$ \\ \hline
$D_7(a_1)$, $[1^2]$ & $-$ & $(q^{m-1} - 1) \cdot q^{6m - 21} \frac{[m-3]}{[2]}$ \\ \hline
$E_{6}+A_1$ & $1$ & $q^{7m-23} \frac{[m-1]}{[2]}$ \\ \hline
$E_{7}(a_{2})$ & $1,3$ & $(q^{m-1} - 1) \cdot q^{6m - 18} \frac{[m-3]}{[2]}$ \\ \hline
$E_{8}(a_{6})$ & $1$ & $(q^{m-1} - 1) \cdot q^{7m-23}$ \\ \hline
$E_{8}(a_{6})$, $[2,1]$ & $3$ & $(q^{m-1} - 1) \cdot (-q^{6m-20})$ \\ \hline
$E_{8}(a_{6})$, $[1^3]$ & $-$ &$(q^{m-1} - 1) \cdot q^{5m-17}$ \\ \hline
$D_7$ & $1$ & $q^{7m -19 } \frac{[m-1]}{[2]}$ \\ \hline
$E_{8}(b_{5})$ & $1$ & $(q^{m-1} - 1) \cdot q^{7m-21}$ \\ \hline
$E_{8}(b_{5})$, $[2,1]$ & $3$ & $(q^{m-1} - 1) \cdot (-q^{6m-18})$ \\ \hline
$E_{8}(b_{5})$, $[1^3]$ & $-$ & $ (q^{m-1} - 1) \cdot q^{5m-15}$ \\ \hline
$E_{7}(a_{1})$ & $1,3$ & $(q^{m-1} - 1) \cdot q^{6m-14} \frac{[m-3]}{[2]}$ \\ \hline
$E_{8}(a_{5})$ & $1$ & $(q^{m-1} -1) \cdot q^{7m -19}$ \\ \hline
$E_{8}(a_{5})$, $[1^2]$ & $-$ & $0$\\ \hline
$E_{8}(b_{4})$ & $1$ & $(q^{m-1} -1) q^{7m -17}$ \\ \hline
$E_{8}(b_{4})$, $[1^2]$ & $-$ & $0$\\ \hline
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{|c|c|c|} \hline
\multicolumn{3}{|c|}{$E_{8}$, part 4} \\ \hline
\multicolumn{1}{|c|}{$(e, \phi)$}
& \multicolumn{1}{|c|}{$m_i$}
& \multicolumn{1}{|c|}{$f_{e, \phi}$}
\\ \hline \hline
$E_{7}$ & $1$ & $q^{7m -13} \frac{[m-1]}{[2]}$ \\ \hline
$E_{8}(a_{4})$ & $1$ & $(q^{m-1} -1) \cdot q^{7m -15}$ \\ \hline
$E_{8}(a_{4})$, $[1^2]$ & $2$ & $ (q^{m-1} -1) \cdot (-q^{6m -13})$ \\ \hline
$E_{8}(a_{3})$ & $1$ & $(q^{m-1} -1) \cdot q^{7m -13}$ \\ \hline
$E_{8}(a_{3})$, $[1^2]$ & $2$ & $(q^{m-1} -1) \cdot (-q^{6m -11})$ \\ \hline
$E_{8}(a_{2})$ & $1$ & $(q^{m-1} -1) \cdot q^{7m - 11}$ \\ \hline
$E_{8}(a_{1})$ & $1$ & $ (q^{m-1} -1) \cdot q^{7m - 9}$ \\ \hline
$E_{8}$ & $-$ & $q^{8m-8}$ \\ \hline
\hline
\end{tabular}
\end{center}
\section{Proof of Theorem~\ref{divisibility-and-nonnegativity-theorem_all}}
\label{proof-of-divisibility-and-nonnegativity-section}
Before recalling here the statement of the theorem, and giving
its proof, let us review some of the terminology.
We denote $R = \rank(Z_G(e))$, while
$H^*(\BBB_e)$ denotes the cohomology of the Springer fiber
for $e \in \0$, regarded as a $W$-representation.
Lastly, the ill-behaved nilpotent
orbits from \eqref{ill-behaved-orbits} are
$$
F_4(a_3), \,
E_6(a_3), \,
E_6(a_3)+A_1, \,
E_7(a_5), \,
E_7(a_3), \,
E_8(a_7), \,
E_8(a_6), \,
E_8(b_5), \,
E_8(a_4), \,
E_8(a_3).
$$
\vskip.1in
\noindent
{\bf Theorem~\ref{divisibility-and-nonnegativity-theorem_all}.}
{\it
Let $e$ be a nilpotent element {\bf not} among the
ill-behaved orbits from \eqref{ill-behaved-orbits},
and assume that $f_{e,\phi}$ is not identically zero.
Then there exists $L, c \in \NN$, independent of $\phi$, such that
$$f_{e,\phi}(m ;q) = \prod^{L}_{j=1} (q^{m+1-2j} -1) \cdot q^{cm} \cdot g_{\phi}(m; q),$$
where $g_{\phi}(m;q)$ is the sum of at most two products of the form $q^{-z} \displaystyle \prod^{R}_{i=1} \tfrac{[m-a_i]_q}{[b_i]_q}$ for some $a_i, b_i, z \in \NN$.
Moreover,
\begin{enumerate}
\item[(i)] For each very good $m$, the polynomial $q^{cm} \cdot g_{\phi}(m; q)$
lies in $\NN[q]$.
\item[(ii)] The rank $r$ of $\ggg$ equals $L + c + R$.
\item[(iii)] The multiplicity of $V$ in the $W$-representation $H^*(\BBB_e)$
is $r-c$.
\item[(iv)] If $e$ is principal-in-a-Levi, then $L=0$.
In particular, $f_{e,\phi}(m;q) \in \NN[q]$ for each very good $m$.
\item[(v)] If $e$ is not principal-in-a-Levi, then $L \geq 1$. In the exceptional types it always happens that $L=1$.
\end{enumerate}
Even when $e$ is one of the ill-behaved orbits from
\eqref{ill-behaved-orbits}, at least
for the case when $\phi=1$, the polynomials $f_{e,1}(m;q)$
are always nonzero,
and still have properties (i),(ii),(iv),(v) listed above.
}
\vskip.1in
Let us embark on the proof.
The fact that $f_{e,\phi}$ takes the form asserted in the theorem follows
from inspection of the formulas for the $f_{e, \phi}$.
The formula in part (ii) is a consequence of \eqref{formula:f}
and the fact that
$$\sum_x \frac{\phi(x)} {|Z_{G^F}(e_x)|} \neq 0,$$
whenever $\phi=1$ or $e$ does not belong to one of the orbits in \eqref{ill-behaved-orbits}.
This also explains why $f_{e,1}$ is always nonzero.
The formula in part (iii) is a consequence of \eqref{formula:f}
and the fact that
$$\sum_x \frac{\left( \wedge^{d_{\kappa}} \pi_{\kappa} \right) (x) \cdot \phi(x)} {|Z_{G^F}(e_x)|} \neq 0,$$
whenever $e$ does not belong to one of the orbits in \eqref{ill-behaved-orbits}, so that
$c = r - (\kappa -1 + d_\kappa)$.
For (i), (iv), (v), it remains to show that $L = 0$ if and only if $e$ is principal-in-a-Levi and that $g_{\phi}(m;q)$ has the desired positivity property. We do this case-by-case.
\subsection{Type $A$}
Since every nilpotent orbit in type $A_n$ is principal-in-a-Levi,
we need to show that for every $\lambda \in \Par(n)$ that
when $\gcd(m,n)=1$ one has
$$
\frac{1}{[m]_q}\qbin{m}{\mu(\lambda)}{q} \in \NN[q].
$$
According to \cite[Corollary 10.4]{RSW},
this follows if all the $\mu_j(\lambda)$'s together
with $m$ have trivial greatest common divisor.
But this is true since a common divisor of
all the $\mu_j(\lambda)$'s would also be a divisor of
$
n=|\lambda|=\sum_j j \mu_j(\lambda),
$
and we assumed that $\gcd(m,n)=1$.
\subsection{Types $B, C$}
For $\lambda$ in $\Par_B(2n+1)$ or $\Par_C(2n)$,
the orbit $\0_{\lambda}$ is principal-in-a-Levi if and only if $\hat{L}(\lambda) = 0$.
Thus whenever $\0_\lambda$ is not principal-in-a-Levi,
so that $\hat{L}(\lambda) > 0$,
the formula for $f_{e,\phi}$ in Proposition \ref{BC_most_general}
contains as a factor the
product $\prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)$.
On the other hand, if $\hat{L}(\lambda) = 0$, this product
is empty, and $f_{e,\phi} \in \NN[q]$
because it is a power of $q$ times a $q$-multinomial.
For the same reason in all cases, aside from this product, the remaining factor lies in $\NN[q]$.
\subsection{Type $D$}
For $\lambda \in \Par_D(2n)$,
the orbit $\0_\lambda$ is principal-in-a-Levi if and only if
$L(\lambda)=0$ or $L(\lambda) =2$ with $\mu_1$ odd.
Note that $|L(\lambda)|$ is always even since $\lambda$ is partition of $2n$.
We examine separately the three conditions on $\lambda$ in
Proposition~\ref{D_most_general}.
\begin{itemize}
\item If $\mu_1$ is odd, then $L(\lambda) \geq 2$,
or equivalently ${\hat{L}(\lambda}) \geq 1$.
Thus in this case, $\0_\lambda$ is principal-in-a-Levi if and only if $\hat{L}(\lambda) = 1$.
Thus the product $\prod_{i=1}^{\hat{L}(\lambda)-1} (q^{m-2i+1}-1)$ is non-empty
exactly when $\0_\lambda$ is not principal-in-a-Levi.
The remaining factors in $f_{e,\phi}$ all lie in $\NN[q]$.
\item When $L(\lambda) \geq 2$ and $\mu_1$ is even,
then $\0_\lambda$ is never principal-in-a-Levi.
Since $\hat{L}(\lambda) \geq 1$,
the product $\prod_{i=1}^{\hat{L}(\lambda)} (q^{m-2i+1}-1)$
is always non-empty.
The other terms in $f_{e, \phi}$ lie in $\NN[q]$.
\item When $\hat{L}(\lambda)=0$,
$\0_\lambda$ is always principal-in-a-Levi, and
$f_{e,\phi} \in \NN[q]$ because it a sum
of a $q$-multinomial and a product of two $q$-multinomials, shifted by powers of $q$.
\end{itemize}
\begin{remark}
In types $A, B, C,$ as well as in the first case of type $D$,
those $\0_\lambda$ which are principal-in-a-Levi not only have
$\Krew(\Phi,\0_\lambda,m;q)$ in $\NN[q]$, but also have
their coefficient sequence symmetric-- this follows in type $A$
from the same result \cite[Corollary 10.4]{RSW} quoted earlier,
and follows in the other types because $q$-multinomials have this
property.
However, this is {\it not} in general true for the third
case in type $D$, even though they are always
principal-in-a-Levi. For example, when $\lambda=(3,3,1,1)$ in $\Par_D(8)$
one has
$$
\Krew(D_4,\0_{(3,3,1,1)},m;q)
=
q^{14} \left( \qbin{\hat{m}}{1,1}{q^2} +
\qbin{\hat{m}}{1}{q^2}^2 \right)
$$
which equals
$
2 q^{14} + 4q^{16} + 6q^{18} + 7q^{20} + 5q^{22} + 3q^{24} + q^{26}
$
when $m=9$.
\end{remark}
\subsection{Exceptional Types}
In the exceptional types many $f_{e,\phi}(m;q)$ can be related to $\Cat(W',m; q)$
for some Weyl group $W'$, which has the desired positivity property.
Most of the remaining cases
can be handled by writing
\begin{equation} \label{basic_form}
\prod^{R}_{i=1} \frac{[m-a_i]_q}{[b_i]_q}
\end{equation}
as a product of polynomials in $q$ with positive coefficients
as in the paper of Krattenthaler-M\"uller \cite{KrattenthalerMuller2}. This is accomplished by restricting $m$ to a fixed congruence class modulo the least common multiple of the $b_i$'s (with $m$ also relatively prime to $h$).
We wrote a program in Sage \cite{sage}, posted on the second author's webpage, that accomplishes this task, except for a handful of cases, making use of \cite[Corollary 6]{KrattenthalerMuller2}, which states that
\begin{equation} \label{KM}
\frac{[\gamma]_q [ab]_q} { [a]_q [b]_q }
\end{equation}
is a polynomial in $q$ with positive coefficients when $\gcd(a,b)=1$ and $\gamma \geq (a-1)(b-1)$.
\begin{example}
Let $e$ be of type $A_1$ in $E_8$. When $m \equiv 17$ modulo $2520$, we find that
$$
\frac{ [m-1] [m-7][m-11][m-13][m-17][m-19][m-23]}
{ [2][6][8][10] [12][14][18]}
$$
is equal to
$$\left [ \frac{m- 17 }{504} \right ]_ {q^{504}}
\left( \frac{ [ \frac{m- 11}{6}]_ {q^6} [ 84 ]_ {q^6} }{ [ 3 ]_ {q^6} [ 28 ]_{q^6} } \right)
\left( \frac{ [ \frac{m- 19}{2} ]_ {q^2} [ 84 ]_{q^2} }{ [ 7 ]_{q^2} [ 12 ]_{q^2} } \right)
\left( \frac{ [ \frac{m- 13}{4} ]_ {q^4} [ 6 ]_ {q^4} }{ [ 3 ]_{q^4} [ 2 ]_{q^4} } \right)
\left[ \frac{m- 7}{10} \right ]_ {q^{10}}
\left[ \frac{m- 23}{6} \right ]_ {q^6}
\left[ \frac{m- 1}{2} \right ]_ {q^2}.$$
Each term is a polynomial with positive coefficients, using \eqref{KM} for the expressions in parentheses.
\end{example}
The remaining cases are a few of those where $g_{\phi}(m;q)$ is a sum of two terms of the form in \eqref{basic_form}. For some congruence classes, each expression of the form \eqref{basic_form} alone will not even be polynomial, let alone positive. We give an example that illustrates how these cases are handled.
\begin{example}
Let $e$ be of type $A_2$ in $E_8$. The expression for $f_{e,1}(m;q)$, up to a power of $q$, is
$$
\frac{ [m-1] [m-7][m-11][m-13][m-17]\left( [m-5] +q^4[m-23] \right)}
{ [2][6][8][10] [12][18]}.
$$
As long as $\gcd(m, 30)=1$ and $m \not\equiv 29$ modulo $30$, the program returns $f_{e,1}$ as a sum of polynomials
with positive coefficients, possibly after rewriting $[m-5] +q^4[m-23]$ as $[m-19]+q^4[m-9]$.
Otherwise, neither summand as in \eqref{basic_form} is polynomial and this also holds even if we rewrite
$[m-5] +q^4[m-23]$ as $[m-19]+q^4[m-9]$. In such cases we need to deal with the full expression $[m-5] +q^4[m-23]$.
For example, when $m \equiv 29$ modulo $360$, we can write
$ \left[ m-1 \right] \left( \frac{[m-5] +q^4[m-23]}{[6][10]} \right)$ as
$$\left[ \frac{m-1}{2} \right]_{q^2} \left( (q^{10}+q^{24})
\cdot \left[ \frac{m- 29}{30} \right]_ {q^{30}} \cdot
\frac{[ 15 ]_{q^2} }{ [ 3 ]_{q^2} [ 5]_{q^2} }
+ \frac{[12]_{q^2}+q^4 [3]_{q^2} }{ [3]_{q^2} [5]_{q^2} } \right).$$
Then $ \frac{[12]_{q^2}+q^4 [3]_{q^2} }{ [3]_{q^2} [5]_{q^2}} = q^{10} - q^{8} + q^4 - q^2+1$, so the product of this polynomial
with $\left[ \frac{m-1}{2} \right]_{q^2}$ has positive coefficients as in the proof of Corollary 6 in \cite{KrattenthalerMuller2}. The remaining terms in $f_{e,1}$
are
$$
\left[ \frac{m- 7}{2} \right ]_ {q^2}
\left[ \frac{m- 11}{18} \right ]_ {q^{18}}
\left[ \frac{m- 13}{8} \right ]_ {q^{8}}
\left[ \frac{m- 17}{12} \right ]_ {q^{12}},
$$
and so $f_{e,1}(m;q)$ has positive coefficients when $m \equiv 29$ modulo $360$.
\end{example}
All cases in the exceptional groups can be handled by reducing to the Catalan case or the case of one of these two examples. It would be nice to have a uniform proof, or at least one that makes use of the fact that $f_{e,\phi}(m;q)$ is polynomial for all very good $m$.
This completes the proof of
Theorem~\ref{divisibility-and-nonnegativity-theorem_all}.
\section{$q$-Narayana formulas}
\label{proof-of-Narayana-formulas-section}
In this section we prove, in types $A,B,C$, that
the $q$-Kreweras numbers,
when summed over nilpotent orbits $\0$ with a fixed
value of the statistic $d(\0)$ as in \eqref{q-Narayana-definition},
give the $q$-Narayana formulas in Theorem~\ref{q-Narayana-formula-theorem}.
In type $A_{n-1}$ we have $d(\0_\lambda) = \ell(\lambda)-1$ from the formula
for $\Krew(A_{n-1},\0_{\lambda},m;q)$ since $r = n-1$.
Thus we want to show that for $k$ in the range $0 \leq k \leq n-1$ that
\begin{equation}
\label{desired-q-Narayana-summation-in-type-A}
\sum_{\substack{\lambda \in \Par(n):\\ \ell(\lambda)=k+1}} \Krew(A_{n-1},\0_{\lambda},m;q)
=
\frac{q^{(n-1-k)(m-1-k)}}{[k+1]_q} \qbin{n-1}{k}{q} \qbin{m-1}{k}{q}.
\end{equation}
In types $B_n$ and $C_n$,
we have $d(\0) = \hat{\ell}(\lambda)$, so we wish to show
for $k$ in the range $0 \leq k \leq n$ that
\begin{equation}
\label{desired-q-Narayana-summation-in-types-BC}
\sum_{\substack{\lambda \in \Par_B(2n+1):\\ \hat{\ell}(\lambda) = k}} \Krew(B_n, \0_{\lambda},m;q)
=\sum_{\substack{\lambda \in \Par_C(2n):\\ \hat{\ell}(\lambda) = k}} \Krew(C_n, \0_{\lambda},m;q)
=
(q^{2})^{(n-k)(\hat{m}-k)} \qbin{n}{k}{q^2} \qbin{\hat{m}}{k}{q^2}.
\end{equation}
We now give a proof that relies on counting the number of nilpotent elements (over a finite field) of certain prescribed rank.
In a sequel paper we give some alternative proofs.
\subsection{Type A}
The sum on the left in \eqref{desired-q-Narayana-summation-in-type-A}
is over nilpotent orbits $\0_\lambda$ with $\ell(\lambda)=k+1$.
In the formula \eqref{formula:typeA} for $f_{e,1}$, all the terms depend only on $\ell(\lambda)$
except for $|Z_G^F(e)|$, where $e=e_\lambda \in \0_\lambda$. Now $\ell(\lambda)=k+1$ means that
$e_\lambda$ is of rank $n-k-1$ when viewed as an $n \times n$-matrix.
It follows that the number of nilpotent $n \times n$-matrices of rank $n-k-1$ over $\FF_q$ is given by
$$
\sum_{\substack{\lambda \in \Par(n):\\ \ell(\lambda)=k+1}}
\frac{|G^F|}{|Z_{G^F}(e_{\lambda})|},
$$
and thus
by \eqref{formula:typeA}
the sum in \eqref{desired-q-Narayana-summation-in-type-A} becomes
\begin{equation}
\label{type-A-nilpotents-by-rank-equation}
q^{m(n- k-1) + \binom{k+1}{2}}\frac{(q-1)^{k+1} [m-1]!_q}{[m-k-1]!_q}
\cdot \frac{\#\{\text{nilpotent $n \times n$ matrices of rank }n-k-1 \}}{|G^F|}.
\end{equation}
The number of nilpotent matrices of rank $n-k-1$ equals
(see \cite{Crabb, Lus1})
$$q^{\binom{n-k-1}{2}} \frac{(q-1)^{n-k-1}[n]!_q}{[k+1]!_q} \qbin{n-1}{k}{q}$$
and $|G^F| = q^{\binom{n}{2}} (q-1)^n [n]!_q$. Substituting these into
\eqref{type-A-nilpotents-by-rank-equation} gives
\eqref{desired-q-Narayana-summation-in-type-A}.
\subsection{Types B, C}
The sums in \eqref{desired-q-Narayana-summation-in-types-BC}
are over orbits $\0_\lambda$
where $\hat{\ell}(\lambda)$ is fixed.
As in type $A_{n-1}$, the formula for $f_{e,1}$ in \eqref{BC_almost} depends only on
$\hat{\ell}(\lambda)$ except for $|Z_G^F(e)|$, where $e=e_\lambda \in \0_\lambda$.
Now $\ell(\lambda)$ is the dimension of the kernel of $e$ in the standard representation of $\ggg$.
Thus the rank of $e_\lambda$ is $2n+1- \ell(\lambda)$ in type $B_n$ and $2n-\ell(\lambda)$ in type $C_n$.
Therefore the condition that $\hat{\ell}(\lambda) = k$ means
that the rank of $e_\lambda$ is either $2n - 2k$
or $2n-2k-1$. Now in type $B_n$, $\ell(\lambda)$ is always odd and so in particular there are no elements of odd rank.
Using this interpretation and \eqref{BC_almost},
the sums in \eqref{desired-q-Narayana-summation-in-types-BC} become
\begin{equation}
\label{type-BC-nilpotents-by-rank-equations}
\begin{array}{rcl}
\displaystyle\sum_{\substack{\lambda \in \Par_B(2n+1):\\ \hat{\ell}(\lambda) = k}} \Krew(B_n,e_{\lambda},m;q)
&=& q^{m(n-k)+k^2} \displaystyle \prod_{j=1}^k (q^{m-(2j-1)}-1) \cdot
\frac{
\# \left\{ \begin{matrix}
\text{nilpotent elts in $\ggg$ of rank } \\
2(n\!-\!k)
\end{matrix} \right\}
}{|G^F|} \\
\displaystyle\sum_{\substack{\lambda \in \Par_C(2n):\\
\hat{\ell}(\lambda) = k}} \Krew(C_n,e_{\lambda},m;q)
&=& q^{m(n-k)+k^2} \displaystyle\prod_{j=1}^k (q^{m-(2j-1)}-1) \cdot
\frac{
\#\left\{ \begin{matrix}
\text{nilpotent elts in $\ggg$ of rank } \\
2(n\!-\!k) \text{ or } 2(n\!-\!k)\!-\!1
\end{matrix} \right\}}{|G^F|}
\end{array}
\end{equation}
\begin{lemma}
The number of nilpotent elements of type $B$ of rank $2s$ is equal to the
number of nilpotent elements of type $C$ of rank $2s$ or $2s-1$. Both are
equal to
$$q^{s^2-s} \frac{\eta(2n)}{\eta(2n-2s)} \qbin{n}{s}{q^2}.$$
\end{lemma}
\begin{proof}
We use the formulas in \cite[Theorems 3.1, 3.2]{Lus1}.
The stated formula is exactly \cite[Theorems 3.1]{Lus1} for the number of nilpotent elements of rank $2s$ in type $B_n$.
The number of rank $2s$ and rank $2s-1$ elements in type $C_n$ are, respectively,
$$q^{s^2+s} \qbin{n}{s}{q^2} \frac{\eta(2n-2)) (q^{2n-2s} - 1)}{\eta(2n-2s)}
\text{ and } q^{s^2-s} \qbin{n}{s}{q^2} \frac{\eta(2n-2))(q^{2s}-1)}{\eta(2n-2s)}.$$
Adding these together gives
\begin{eqnarray*}
q^{s^2-s} \qbin{n}{s}{q^2} \frac{\eta(2n-2)}{ \eta(2n-2s)}
\left( q^{2s}(q^{2n-2s)}-1) + q^{2s}-1 \right) =
q^{s^2-s} \qbin{n}{s}{q^2} \frac{\eta(2n)}{ \eta(2n-2s))}.
\end{eqnarray*}
\end{proof}
From the lemma, it is immediate that the two expressions
in \eqref{type-BC-nilpotents-by-rank-equations}
are equal. To evaluate them, we write
$$\displaystyle \prod_{j=1}^k (q^{m - (2j-1)}-1) = \frac{\eta(m)}{\eta(m - 2k)}.$$
With $s=n-k$,
the expressions in \eqref{type-BC-nilpotents-by-rank-equations}
evaluate to
$$q^{m(n-k)+k^2} \frac{\eta(m)}{\eta(m - 2k)} \cdot
q^{(n-k)^2-(n-k)}\frac{\eta(2n)}{\eta(2k)} \qbin{n}{k}{q^2} \cdot \frac{1}{|G^F|}.$$
Since $|G^F| = q^{n^2} \eta(2n)$ in both types, this becomes
$$q^{m(n-k)+k^2 + (n-k)^2-(n-k) - n^2} \qbin{\hat{m}}{k}{q^2} \qbin{n}{k}{q^2}$$ and
the exponent of $q$ simplifies to $2(n-k)(\hat{m}-k)$ as desired, where as before $\hat{m} = \frac{m-1}{2}$.
\begin{remark}
From the description of special nilpotent pieces in \cite{Lus2},
special nilpotent pieces in type $B$ and $C$
consist of some nilpotent orbits whose elements are of rank $2s$ or $2s-1$ for some fixed $s$.
It follows that the set of nilpotent elements of rank $2s$ and $2s-1$ is a union of special nilpotent pieces.
Moreover, a special nilpotent piece in type $B$ corresponds to a special nilpotent piece in type $C$ for the same value of $s$. Therefore the equality between the number of elements of rank $2s$ and $2s-1$ in type $B$ and in type $C$ also follows from Lusztig's work that corresponding special pieces in $B$ and $C$ have the same cardinality \cite[\S 6.9]{Lus2}.
\end{remark}
\section{Proof of Theorem~\ref{CSP-theorem}}
\label{proof-of-CSP-theorem-section}
Recall the statement of the theorem:
\vskip.1in
\noindent
{\bf Theorem~\ref{CSP-theorem}.}
{\it
In types $A,B,C,D$, for $m \equiv 1$ mod $h$, say $m = sh+1$,
and for each $W$-orbit $[X]$ of intersection subspaces
of reflecting hyperplanes, $\Krew(\Phi,\0_X,m;q=\omega_{d})$ counts
those $w_1 \leq \cdots \leq w_s$ in $NC^{(s)}(W)$
which are both
\begin{enumerate}
\item[$\bullet$] fixed under the action
an element of order $d$ in the $\ZZ/sh\ZZ$-action, and
\item[$\bullet$]
have the subspace $V^{w_1}$ lying in the $W$-orbit $[X]$.
\end{enumerate}
}
\begin{remark}
We mention here how recent work has generalized the type $A_{n-1}$ special case of
Theorem~\ref{CSP-theorem} from the special case $m=sn+1$
to the case of {\it all} very good $m$ in type $A_{n-1}$, that is, where $\gcd(m,n)=1$.
In \cite{ArmstrongRhoadesWilliams},
Armstrong, Rhoades and Williams introduced for all $m > n$ with
$\gcd(m,n)=1$ the set $NC(n,m)$ of {\it rational}
or {\it $(n,m)$-noncrossing partitions}. These $(n,m)$-noncrossing partitions
are a subset of $NC(m-1)$, specializing to $NC^{(s)}(W)$ when $m=sn+1$.
One might ask whether the subset $NC(n,m) \subset NC(m-1)$
is closed under the natural dihedral symmetry group of
order $2(m-1)$ acting on $NC(m-1)$; this was left open in \cite{ArmstrongRhoadesWilliams},
but later resolved affirmatively in work of
Bodnar and Rhoades \cite{BodnarRhoades}. In fact,
recent work of Bodnar\footnote{B. Rhoades, personal communication, 2016.}
has shown how to define $NC(n,m)$ when $m < n$, again with a dihedral group of
order $2(m-1)$ acting.
In particular,
considering the cyclic $\ZZ/(m-1)\ZZ$-action via rotations,
the Bodnar and Rhoades also proved a cyclic sieving phenomenon
\cite[Thm. 5.1]{BodnarRhoades} whose $m=sn+1$ special case is
equivalent to the type $A_{n-1}$ special
case of Theorem~\ref{CSP-theorem}. These results
involve a $q$-Kreweras number that differs
slightly from the one in Theorem~\ref{q-Kreweras-formula-theorem},
in that it is
missing the factor of $q^{m(n-\ell(\lambda))-c(\lambda)}$.
However, this power of $q$ makes no difference in the proof of the cyclic sieving phenomenon,
as it happens to equal $1$ whenever the $q$-Kreweras number
evaluated at an $(m-1)^{st}$ root-of-unity is nonvanishing-- see
Lemma \ref{c(lambda)-congruences-lemma} below.
The aforementioned work of Bodnar also extends this cyclic sieving phenomenon
to the case $m < n$.
\end{remark}
\vskip.1in
In proving Theorem~\ref{CSP-theorem},
our plan will be to first compute the
evaluations $\Krew(\Phi,\0_X,m;q=\omega_{d})$, and
then compare them with the combinatorial models for
$NC^{(s)}(W)$ in the classical types $A,B,C,D$.
\subsection{Root-of-unity evaluation lemmas}
We collect here a few well-known
observations on evaluating certain
polynomials in $q$ at a primitive $d^{th}$ root of unity $\omega_d$.
The proofs are straightforward and omitted.
\vskip.1in
\noindent
{\bf Warning:}
For the remainder of the paper, we abandon the convention
that ``$a \equiv b$'' means $a \equiv b \bmod{2}$, as we will
now need to often consider equivalence modulo $d$ for
other moduli $d \neq 2$.
\begin{lemma}
\label{root-of-unity-facts}
Let $\omega_d:=e^{{2 \pi i}{d}}$ or any other primitive $d^{th}$ root-of-unity.
\begin{enumerate}
\item[(i)] The polynomial
$$
[m]_q:=\frac{1-q^m}{1-q}
$$
has $\omega_d$ as a root with multiplicity $1$ or $0$,
depending on whether $d$ divides $m$ or not.
\item[(ii)]
For any positive integer $m$, the $q$-factorial
$$
[m]!_q:= [1]_q [2]_q \cdots [m]_q
$$
has $\omega_d$ as a root of multiplicity
$\lfloor \frac{m}{d} \rfloor$.
\item[(iii)]
For $d$ dividing $N$, the product
$$
[N]_q [N-1]_q \cdots [N-k+1]
$$
has $\omega_d$ as a root of multiplicity $\lceil \frac{k}{d} \rceil$.
\item[(iv)] If $a,b$ are positive integers with $a \equiv b \mod d$, then
$$\lim_{q \rightarrow \omega_d} \frac{[a]_q}{[b]_q} =
\begin{cases}
\frac{a}{b} & \text{ if } a \equiv b \equiv 0 \mod d \\
1 & \text{ if } a \equiv b \not\equiv 0 \mod d.
\end{cases}
$$
\item[(v)]
For nonnegative $n,k$ expressed uniquely as
$
n=d \cdot \hat{n} + \hat{\hat{n}}
$
and
$
k=d \cdot \hat{k} + \hat{\hat{k}}
$
with $0 \leq \hat{\hat{k}}, \hat{\hat{n}} < d$, one has
$$
\lim_{q \rightarrow \omega_d} \qbin{n}{k}{q} =
\binom{\hat{n}}{\hat{k}} \cdot
\lim_{q \rightarrow \omega_d} \qbin{\hat{\hat{n}}}{\hat{\hat{k}}}{q}
$$
\end{enumerate}
\end{lemma}
In types $B, C, D$, we will need to evaluate certain
polynomials in $q^2$ at $q=\omega_d$.
As notation, let
$$
d^-:=\frac{d}{\gcd(2,d)} =
\left\{
\begin{aligned}
d & \text{ for }d\text{ odd},\\
\frac{d}{2}& \text{ for }d\text{ even}.
\end{aligned}
\right\}
\qquad \text{ and } \qquad
d^+:=\lcm(2,d) =
\left\{
\begin{aligned}
d & \text{ for }d\text{ even},\\
2d& \text{ for }d\text{ odd}
\end{aligned}
\right\}
=2 d^-.
$$
Some facts that will be used frequently without mention are that, for an even integer $2N$,
$$
d\text{ divides }2N \quad \Leftrightarrow \quad
d^+\text{ divides }2N \quad \Leftrightarrow \quad
d^-\text{ divides }N,
$$
and in this situation,
$$
\frac{2N}{d^+} = \frac{N}{d^-}.
$$
The proof of the following assertions are then straightforward.
\begin{lemma}
\label{BCD-root-of-unity-facts}
Assume throughout that {\bf $d$ divides $2N$}.
\begin{enumerate}
\item[(i)] For a sequence of nonnegative integers
$(\alpha_1,\ldots,\alpha_\ell)$, one has
$$
\lim_{q \rightarrow \omega_d}
\qbin{N}{\alpha_1,\ldots,\alpha_\ell}{q^2}=
\begin{cases}
\binomial{\frac{2N}{d^+}}{\frac{2\alpha_1}{d^+},\ldots,\frac{2\alpha_\ell}{d^+}}
=\binomial{\frac{N}{d^-}}{\frac{\alpha_1}{d^-},\ldots,\frac{\alpha_\ell}{d^-}}
&\text{ if }d\text{ divides }2\alpha_i\text{ for each }i,\\
0&\text{ otherwise.}
\end{cases}
$$
\item[(ii)]
Given a nonnegative integer $k$, for $d=1,2$ one has
$$
\lim_{q \rightarrow \omega_d}
\qbin{N+1}{k}{q^2}=
\binomial{N+1}{k}
$$
but for $d \geq 3$ one has
$$
\lim_{q \rightarrow \omega_d}
\qbin{N+1}{k}{q^2}=
\begin{cases}
\binomial{\frac{2N}{d^+}}{\lfloor\frac{2k}{d^+}\rfloor}
=\binomial{\frac{N}{d^-}}{\lfloor\frac{k}{d^-}\rfloor}
&\text{ if }k \equiv 0,1 \bmod{d^-} \\
0&\text{ otherwise.}
\end{cases}
$$
\end{enumerate}
\end{lemma}
\subsection{The $q$-Kreweras numbers evaluated at roots of unity}
We next use Lemmas~\ref{root-of-unity-facts} and
\ref{BCD-root-of-unity-facts} together with our formulas for
$\Krew(\Phi,\0_\lambda,m,q)$ to evaluate them at $q=\omega_d$
whenever $\0_\lambda$ is principal-in-a-Levi, and $d$ divides
$m-1=sh$ for some $s \geq 1$.
We first compute only
their complex modulus, ignoring
multiplicative factors of powers of $q$ (Proposition~\ref{evaluations-prop}).
Then we check they are correct on the nose,
not just up to modulus (Proposition~\ref{q-powers-become-plus-1}).
\begin{proposition}
\label{evaluations-prop}
Let $s$ be a positive integer, and assume that $d$ divides $m-1=sh$.
\begin{itemize}
\item[(i)]
In type $A_{n-1}$ one has $h=n$, so that $m-1=sn$.
Every $\lambda$ in $\Par(n)$ has $\0_\lambda$ principal-in-a-Levi.
Then
$
\Krew(A_{n-1},\0_\lambda,m;q=\omega_d)
$
is nonvanishing if and only if at most one $\mu_{j_0}(\lambda)$
is not divisible by $d$, and if such a $j_0$ exists,
then $\mu_{j_0}(\lambda) \equiv 1 \bmod{d}$. Furthermore, in this situation
$$
\left\lVert
\Krew(A_{n-1},\0_\lambda,m;q=\omega_d)
\right\rVert
=
\left\lVert
\lim_{q \rightarrow \omega_d}
\frac{1}{[m]_q}\qbin{m}{\mu(\lambda)}{q}
\right\rVert
=
\begin{cases}
\displaystyle \frac{1}{m}\binomial{m}{\mu(\lambda)} & \text{ if }d=1,\\
\binomial{ \frac{sn}{d} }
{ \lfloor \frac{\mu(\lambda)}{d}\rfloor } & \text{ if } d \geq 2.
\end{cases}
$$
\item[(ii)]
In type $B_n, C_n$ one has $h=2n$, so that $m-1=2sn$.
Then $\lambda$ in $\Par_B(2n+1)$ or $\Par_C(2n)$
has $\0_\lambda$ principal-in-a-Levi if and only if
at most one part $j_0$ has $\mu_{j_0}(\lambda)$ odd,
that is $\hat{L}(\lambda)=0$.
Then
$
\Krew(\Phi,\0_\lambda,m;q=\omega_d)
$
is nonvanishing if and only if $d$ divides $2\hat{\mu}_j(\lambda)$ for each $j$,
in which case
$$
\left\lVert
\Krew(\Phi,\0_\lambda,m;q=\omega_d)
\right\rVert
=
\left\lVert
\lim_{q \rightarrow \omega_d}
\qbin{sn}{\hat{\mu}(\lambda)}{q^2}
\right\rVert
=\binomial{\frac{sn}{d^-}}
{ \frac{\hat{\mu}(\lambda)}{d^-} }
$$
\item[(iii)]
In type $D_n$ one has $h=2(n-1)$, so that $m-1=2s(n-1)$.
Then $\lambda$ in $\Par_D(2n)$,
has $\0_\lambda$ principal-in-a-Levi if and only if either
\begin{itemize}
\item[$\bullet$] there are two part sizes, namely $1$ and
some odd $j_0 \geq 3$, with odd multiplicity, so $\hat{L}(\lambda)=1$, or
\item[$\bullet$] there are no parts with odd multiplicity, that is, $\hat{L}(\lambda)=0$.
\end{itemize}
In the former $\hat{L}(\lambda)=1$ case,
$
\Krew(D_n,\0_\lambda,m;q=\omega_d)
$
is nonvanishing if and only if $d$ divides $2\hat{\mu}_j(\lambda)$ for all $j$,
in which case
$$
\left\lVert
\Krew(D_n,\0_\lambda,m;q=\omega_d)
\right\rVert
=
\left\lVert
\lim_{q \rightarrow \omega_d}
\qbin{s(n-1)}{\hat{\mu}(\lambda)}{q^2}
\right\rVert
=
\binomial{\frac{s(n-1)}{d^-}}
{ \frac{\hat{\mu}(\lambda)}{d^-} }.
$$
In the latter $\hat{L}(\lambda)=0$ case,
$
\Krew(D_n,\0_\lambda,m;q=\omega_d)
$
is nonvanishing if and only if
both
\begin{itemize}
\item[(a)] $\mu_j(\lambda) \equiv 0 \bmod{d^+}$ for all $j \geq 2$, and
\item[(b)] $\mu_1(\lambda) \equiv 0 \text{ or } 2 \bmod{d^+}$,
\end{itemize}
in which case
$$
\begin{aligned}
&\left\lVert
\Krew(D_n,\0_\lambda,m;q=\omega_d)
\right\rVert \\
&=\left\lVert
\lim_{q \rightarrow \omega_d}
\left(
\qbin{s(n-1)}{\hat{\mu}_{\geq 2}(\lambda)}{q^2}
\qbin{s(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|)}
{\hat{\mu}_1(\lambda)} {q^2}
+
q^{\mu_1(\lambda)-\tau_1(\lambda)}
\qbin{s(n-1)}{\hat{\mu}(\lambda)}{q^2}
\right)
\right\rVert \\
&=
\begin{cases}
\binomial{s(n-1)}{\hat{\mu}_{\geq 2}(\lambda)}
\binomial{s(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)}
+ \binomial{s(n-1)}{\hat{\mu}(\lambda)}
&\text{ if }d=1, \\
\binomial{s(n-1)}{\hat{\mu}_{\geq 2}(\lambda)}
\binomial{s(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)}
+ (-1)^n \binomial{s(n-1)}{\hat{\mu}(\lambda)}
&\text{ if }d=2, \\
\left( 1+(-1)^{\frac{2n}{d}} \right)
\binomial{\frac{2s(n-1)}{d^+}}
{ \frac{\mu(\lambda)}{d^+} }
& \text{ if }d\geq 3\text{ and }\mu_1(\lambda) \equiv 0 \bmod{d^+},\\
\binomial{\frac{2s(n-1)}{d^+}}
{ \frac{\mu_1(\lambda)-2}{d^+}, \frac{\mu_{\geq 2}(\lambda)}{d^+} }
& \text{ if }d\geq 3\text{ and }\mu_1(\lambda) \equiv 2 \bmod{d^+}. \\
\end{cases}
\end{aligned}
$$
\end{itemize}
\end{proposition}
\begin{proof}
\vskip.1in
\noindent
{\sf Type A.}
The $d=1$ case is clear, since one is setting $q=\omega=1$.
Thus without loss of generality, $d \geq 2$.
We know from Theorem~\ref{divisibility-and-nonnegativity-theorem} that
$$
\frac{1}{[m]_q}\qbin{m}{\mu(\lambda)}{q}
=\frac{[m-1]_q [m-2]_q \cdots [m-\ell(\lambda)+1]_q}
{\prod_j [\mu_j(\lambda)]!_q}
$$
is a polynomial in $q$.
Lemma~\ref{root-of-unity-facts}(ii,iii) tell us that it
has $\omega_d$ as a root of multiplicity
$\lfloor \frac{\ell(\lambda)-1}{d} \rfloor$
in the numerator, and of multiplicity
$\sum_{j \geq 1} \lfloor \frac{\mu_j(\lambda)}{d} \rfloor$
in the denominator. Hence one must always have the
inequality
\begin{equation}
\label{type-A-numerator-denominator-inequality}
\left\lfloor \frac{\ell(\lambda)-1}{d} \right\rfloor
\geq \sum_{j \geq 1} \left\lfloor \frac{\mu_j(\lambda)}{d} \right\rfloor
\end{equation}
and this must be an equality whenevever this polynomial
is nonvanishing at $q=\omega_d$.
Writing $r_j$ for the
remainder of $\mu_j(\lambda)$ on division by $d$
with $0 \leq r_j \leq d-1$, equality
in \eqref{type-A-numerator-denominator-inequality}
would force
$$
\frac{\ell-1}{d}
\leq
\left\lceil \frac{\ell-1}{d} \right\rceil
=
\sum_{j \geq 1} \left\lfloor \frac{\mu_j(\lambda)}{d} \right\rfloor =
\sum_{j \geq 1} \frac{\mu_j-r_j}{d} =
\frac{\ell-\sum_{j \geq 1} r_j}{d}.
$$
Thus $\sum_j r_j \leq 1$, or in other words, at most one
of the $\mu_j(\lambda)$ is not divisible by $d$, and its remainder is $1$.
In this situation, one can use Lemma~\ref{root-of-unity-facts}(iv) to match up
numerator and denominator factors yielding the asserted
evaluation in (i).
\vskip.1in
\noindent
{\sf Types B, C.}
This follows from Lemma~\ref{BCD-root-of-unity-facts}(i) with $N=sn$.
\vskip.1in
\noindent
{\sf Type D.}
The first case, where $\hat{L}(\lambda)=1$,
similarly to types $B/C$, follows from
Lemma~\ref{BCD-root-of-unity-facts}(i) with $N=s(n-1)$.
In the second case, where $\hat{L}(\lambda)=0$, we must set $q=\omega_d$ in
\begin{equation}
\label{type-D-case-4-without-q-power}
\qbin{s(n-1)}{\hat{\mu}_{\geq 2}(\lambda)}{q^2}
\qbin{s(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)} {q^2}
+ q^{\mu_1(\lambda)-\tau_1(\lambda)} \qbin{s(n-1)}{\hat{\mu}(\lambda)}{q^2}.
\end{equation}
Note that Lemma~\ref{BCD-root-of-unity-facts}(i)
with $N=s(n-1)$ shows that,
whenever condition (a) above fails, both summands in
\eqref{type-D-case-4-without-q-power} vanish-- the first summand vanishes
because its first factor vanishes. Similarly,
Lemma~\ref{BCD-root-of-unity-facts}(ii) with
$N=s(n-1)-|\mu_{\geq 2}(\lambda)|$ shows that whenever
condition (b) above fails, both summands
in \eqref{type-D-case-4-without-q-power}
vanish-- the second factor in
the first summand vanishes
unless $\mu_1(\lambda) \equiv 0$ or $2 \bmod{d^+}$.
Hence without loss of generality we may assume
both conditions (a),(b) hold, and we examine what
happens when one evaluates
\eqref{type-D-case-4-without-q-power} at $q=\omega_d$
for various values of $d$.
Note that when $d=1$ so that $q=1$, it gives the asserted
evaluation, so without loss of generality, $d \geq 2$.
\begin{lemma}
A $\lambda$ in $\Par_D(2n)$ with $\mu_j(\lambda) \equiv 0 \bmod{d^+}$ has
$2n \equiv 0 \bmod{d^+}$ and
$
\omega_d^{\mu_1(\lambda)-\tau_1(\lambda)} = (-1)^{\frac{2n}{d}}.
$
\end{lemma}
\begin{proof}
One has
$$
2n=|\lambda|=\sum_j j \mu_j(\lambda) \equiv 0 \bmod{d^+}
$$
and
$$
\mu_1(\lambda)-\tau_1(\lambda)
= \mu_1(\lambda)-\frac{1}{2}\sum_{j \text{ odd }} \mu_j(\lambda)
=\frac{1}{2} \left( \mu_1(\lambda)-
\sum_{\text{ odd }j \geq 3} \mu_j(\lambda)
\right)
=\frac{d}{2} \left(\frac{\mu_1(\lambda)}{d}-
\sum_{\text{ odd }j \geq 3} \frac{\mu_j(\lambda)}{d}
\right)
$$
and therefore
$$
\begin{aligned}
\omega_d^{\mu_1(\lambda)-\tau_1(\lambda)}
&= \left( \omega_{2d}^d
\right)^{
\frac{\mu_1(\lambda)}{d} -
\sum_{\text{ odd }j \geq 3} \frac{\mu_j(\lambda)}{d}}
= (-1)^{ \frac{\mu_1(\lambda)}{d} -
\sum_{\text{ odd }j \geq 3} \frac{\mu_j(\lambda)}{d} }\\
&=(-1)^{
\sum_{\text{ odd }j} \frac{\mu_j(\lambda)}{d}}
=(-1)^{
\sum_{j} \frac{j \cdot \mu_j(\lambda)}{d}}
=(-1)^{\frac{2n}{d}}.
\end{aligned}
$$
\end{proof}
\noindent
Together with Lemma~\ref{BCD-root-of-unity-facts}(ii), this gives the
asserted evaluation for $d=2$, as in that case, $(-1)^{\frac{2n}{d}}=(-1)^n.$
Now assume $d \geq 3$ and both conditions (a) and (b) hold.
If $\mu_1(\lambda)\equiv 0 \bmod{d^+}$, then
setting $q=\omega_d$ in \eqref{type-D-case-4-without-q-power} gives,
after applying Lemma~\ref{BCD-root-of-unity-facts}(i) with
$N=s(n-1)-|\hat{\mu}_{\geq 2}(\lambda)|$,
$$
\binomial{\frac{2s(n-1)}{d^+}}{\frac{\mu_{\geq 2}(\lambda)}{d^+}}
\binomial{
\frac{2s(n-1)-|\mu_{\geq 2}(\lambda)| }{ d^+ }
}
{
\frac{\mu_1(\lambda)}{d^+}
}
+ (-1)^{\frac{2n}{d}}
\binomial{ \frac{2s(n-1)}{d^+} }{ \frac{\mu(\lambda) }{d^+} }
= \left( 1 + (-1)^{\frac{2n}{d}} \right)
\binomial{ \frac{2s(n-1)}{d^+} }{ \frac{\mu(\lambda) }{d^+} },
$$
as desired. On the other hand, if $\mu_1(\lambda)\equiv 2 \bmod{d^+}$
then the first summand in \eqref{type-D-case-4-without-q-power} vanishes
because its second factor is zero, and hence
Lemma~\ref{BCD-root-of-unity-facts}(i) gives (up to sign), the stated
answer.
\end{proof}
\begin{proposition}
\label{q-powers-become-plus-1}
In types $A, B, C, D$ and for a positive integer $s$ and
a divisior $d$ of $m-1=sh$, whenever a principal-in-a-Levi
nilpotent orbit $\0_\lambda$ has
$\Krew(\Phi,\0_\lambda,m;q=\omega_d)$ nonvanishing,
it equals its (nonnegative integer) complex modulus
$\left\lVert \Krew(\Phi,\0_\lambda,m;q=\omega_d) \right\rVert$
given in Proposition~\ref{evaluations-prop}.
\end{proposition}
As preparation for the proof of this proposition, we recall that
the power of $q$ appearing as factor in front of the $q$-Kreweras
formula in Theorem~\ref{q-Kreweras-formula-theorem}(Type A)
involves the quantity
$
c(\lambda) = \sum_j \lambda_j' \lambda_{j+1}',
$
while the corresponding powers of $q$ in
Theorem~\ref{q-Kreweras-formula-theorem}(Types B,C,D)
involve $c(\lambda)/2$. Thus the following lemma on values
modulo $d$, that is, in $\QQ/d\ZZ$, will be helpful.
\begin{lemma}
\label{c(lambda)-congruences-lemma}
Assume $\0_\lambda$ is principal-in-a-Levi
and $\Krew(\Phi,\0_\lambda,m;q=\omega_d) \neq 0$.
\begin{itemize}
\item[$\bullet$] In type $A_{n-1}$,
\begin{enumerate}
\item[(i)]
one either has $\mu_j(\lambda) \equiv 0 \bmod{d}$ for all $j$, in which case,
$$
c(\lambda) \equiv 0 \bmod{d}, \text{ or }
$$
\item[(ii)]
one $j_0$ has
$\mu_j(\lambda) \equiv 0 \bmod{d}$ for all $j \neq j_0$
and $\mu_{j_0}(\lambda) \equiv 1 \bmod{d}$,
in which case
$$
c(\lambda) \equiv j_0-1 \bmod{d}.
$$
\end{enumerate}
\item[$\bullet$] In types $B_n, C_n$,
\begin{enumerate}
\item[(i)]
one either has all $\mu_j(\lambda)$ even and divisible by $d$,
in which case
$$
\frac{c(\lambda)}{2} \equiv 0 \bmod{d}, \text{ or }
$$
\item[(ii)]
one $j_0$ has $\mu_{j_0}(\lambda)$ odd,
$\mu_j(\lambda)$ even for $j \neq j_0$,
and $2\hat{\mu}_j(\lambda) \equiv 0 \bmod{d}$ for all $j$, in which case
$$
\frac{c(\lambda)}{2} \equiv \frac{j_0-1}{2} +\hat{\ell}(\lambda) - \hat{\mu}_{j_0}(\lambda) \bmod{d}.
$$
\end{enumerate}
\item[$\bullet$] In type $D_n$,
\begin{enumerate}
\item[(i)]
one either has all $\mu_j(\lambda)$ are even,
$2\hat{\mu}_j(\lambda) \equiv 0 \bmod{d}$ for $j \geq 2$, and
$2\hat{\mu}_1(\lambda) \equiv 0 \text{ or }2 \bmod{d^+}$,
in which case
$$
\frac{c(\lambda)}{2} \equiv 0 \bmod{d}, \text{ or }
$$
\item[(ii)]
one has $\mu_1(\lambda)$ and $\mu_{j_0}(\lambda)$ odd
for a unique odd part size $j_0 \geq 3$, but $\mu_j(\lambda)$ even
for all $j \neq 1,j_0$,
and $2\hat{\mu}_j(\lambda) \equiv 0 \bmod{d}$ for all $j$,
in which case
$$
\frac{c(\lambda)}{2} \equiv \frac{j_0}{2}
+ \hat{\mu}_1(\lambda) + \hat{\mu}_{j_0}(\lambda) \bmod{d}.
$$
\end{enumerate}
\end{itemize}
\end{lemma}
\begin{proof}[Proof of Proposition~\ref{c(lambda)-congruences-lemma}]
The arguments are all similar, so we show only the last (hardest) case:
the one in type $D$ with a unique odd $j_0 \geq 3$ for which
$\mu_1(\lambda), \mu_{j_0}(\lambda)$ are odd,
$\mu_j(\lambda)$ is even for $j \neq 1,j_0$,
and $2\hat{\mu}_j(\lambda) \equiv 0 \bmod{d}$ for all $j$.
Note that
$$
\mu_i(\lambda)=
\begin{cases}
2 \hat{\mu}_i(\lambda) & \text{ if }i \neq 1,j_0,\\
1+2 \hat{\mu}_i(\lambda) & \text{ if }i = 1,j_0,\\
\end{cases}
\quad \text{ and hence } \quad
\lambda_j'= \sum_{i \geq j} \mu_i(\lambda) =
\begin{cases}
\sum_{i \geq j} 2\hat{\mu}_i(\lambda) & \text{ if }j \geq j_0+1,\\
1+\sum_{i \geq j} 2\hat{\mu}_i(\lambda) & \text{ if }2 \leq j \leq j_0,\\
2+\sum_{i \geq j} 2\hat{\mu}_i(\lambda) & \text{ if }j =1.\\
\end{cases}
$$
Therefore, using ``$\equiv$'' to denote equivalence
modulo $d\ZZ$ in $\QQ/d\ZZ$, one has
$$
\frac{\lambda_j'\lambda_{j+1}'}{2}
=\begin{cases}
\frac{1}{2}
\left( \sum_{i \geq j} 2\hat{\mu}_i(\lambda) \right)
\left( \sum_{i \geq j+1} 2\hat{\mu}_i(\lambda) \right)
\equiv 0 & \text{ if }j \geq j_0+1, \\
& \\
\frac{1}{2}
\left( 1+\sum_{i \geq j_0} 2\hat{\mu}_i(\lambda) \right)
\left( \sum_{i \geq j_0+1} 2\hat{\mu}_i(\lambda) \right)
\equiv \sum_{i \geq j_0+1} \hat{\mu}(\lambda) & \text{ if }j = j_0, \\
& \\
\frac{1}{2}
\left(1+\sum_{i \geq j} 2\hat{\mu}_i(\lambda) \right)
\left(1+ \sum_{i \geq j+1} 2\hat{\mu}_i(\lambda) \right)& \\
\equiv \frac{1}{2}+\sum_{i \geq j+1}\hat{\mu}_i(\lambda)
+\sum_{i \geq j}\hat{\mu}_i(\lambda)& \\
\equiv \frac{1}{2}+\hat{\mu}_j(\lambda) & \text{ if }2 \leq j \leq j_0-1, \\
& \\
\frac{1}{2}
\left(2+ \sum_{i \geq j} 2\hat{\mu}_i(\lambda) \right)
\left(1+ \sum_{i \geq j+1} 2\hat{\mu}_i(\lambda) \right)
\equiv 1+\sum_{i \geq 1}\hat{\mu}_i(\lambda) & \text{ if }j =1.
\end{cases}
$$
Consequently,
$$
\begin{array}{rccccccl}
c(\lambda) &=&\sum_j \frac{\lambda_j'\lambda_{j+1}'}{2} & & & & &\\
&=& \sum_{j \geq j_0+1} \frac{\lambda_j'\lambda_{j+1}'}{2}
&+& \frac{\lambda_{j_0}'\lambda_{j_0+1}'}{2}
&+& \sum_{j=2}^{j_0-1} \frac{\lambda_j'\lambda_{j+1}'}{2}
&+ \frac{\lambda_1'\lambda_{2}'}{2}\\
&\equiv& 0
& +& \sum_{i \geq j_0+1} \hat{\mu}_i(\lambda)
& +& \sum_{j=2}^{j_0-1} \left( \frac{1}{2} + \hat{\mu}_j(\lambda) \right)
& + \left( 1 + \sum_{i \geq 1}\hat{\mu}_i(\lambda) \right)\\
&\equiv&
\hat{\mu}_1(\lambda)
+ \hat{\mu}_{j_0}(\lambda)
+ \frac{j_0}{2}\qedhere.& & & & &
\end{array}
$$
\end{proof}
\begin{proof}[Proof of Proposition~\ref{q-powers-become-plus-1}]
We go through eight cases
from Proposition~\ref{c(lambda)-congruences-lemma}. Here
``$\equiv$'' is equivalence in $\QQ/d\ZZ$.
\vskip.1in
\noindent
{\sf Type $A_{n-1}$.}
Here one needs to show that the exponent
$$
E:=m(n-\ell(\lambda))-c(\lambda)
$$
has $E \equiv 0 $.
In the case $A_{n-1}(i)$, since $d$ divides
$\mu_j(\lambda)$ for all $j$, it also divides
$
n=\sum_j j \mu_j(\lambda)
$
and
$
\ell(\lambda) = \sum_j \mu_j(\lambda).
$
Also $c(\lambda)\equiv 0$
by Proposition~\ref{c(lambda)-congruences-lemma},
so $E \equiv 0 $, as desired.
In the case $A_{n-1}(ii)$, one has
$$
n-\ell(\lambda) = \sum_{j \geq 1}(j-1) \mu_j \equiv j_0-1
$$
and Proposition~\ref{c(lambda)-congruences-lemma}
showed $c(\lambda) \equiv j_0-1 $. But then $m \equiv 1 $ gives
$$
\begin{aligned}
E = m(n-\ell(\lambda))-c(\lambda)
&\equiv (n-\ell(\lambda))-c(\lambda) \\
&\equiv j_0-1 - (j_0-1) \equiv 0.
\end{aligned}
$$
\vskip.1in
\noindent
{\sf Types B, C.}
Here one needs to show that the exponent
$
E=
\BCDexponent+
\sigma(\lambda)
\equiv 0 $,
using the abbreviations
$$
\begin{aligned}
\BCDexponent
&:=m(n-\hat{\ell}(\lambda))-\frac{c(\lambda)}{2}-\frac{L(\lambda)}{4} \\
\sigma(\lambda)
&:=
\begin{cases}
\tau_1(\lambda)+\frac{1}{4} & \text{ in type }B,\\
\tau_0(\lambda) & \text{ in type }C\text{ if }\ell(\lambda)\text{ even},\\
\tau_0(\lambda)+\frac{1}{4}-\frac{\ell(\lambda)}{2} & \text{ in type }C\text{ if }\ell(\lambda)\text{ odd}. \\
\end{cases}
\end{aligned}
$$
Note that since $m \equiv 1$, one has the congruence
$
E\equiv
n-\hat{\ell}(\lambda)-\frac{c(\lambda)}{2}-\frac{L(\lambda)}{4}
+\sigma(\lambda).
$
In case $B_n/C_n (i)$,
Proposition~\ref{c(lambda)-congruences-lemma}
says that $\frac{c(\lambda)}{2} \equiv 0 $.
Furthermore $L(\lambda)=0$, and $\mu_j(\lambda)$ all even implies
both $\ell(\lambda) =\sum_j \mu_j(\lambda)$ and
$|\lambda|=\sum_j j \cdot \mu_j(\lambda)$ are even.
Thus one must be in type $C_n$, with $\sigma(\lambda)=\tau_0(\lambda)$, so
$$
\begin{aligned}
E &\equiv n-\hat{\ell}(\lambda)+\tau_0(\lambda) \\
&\equiv \sum_j j \hat{\mu}_j(\lambda)
- \sum_j \hat{\mu}_j(\lambda)
+ \sum_{j\text{ even }} \hat{\mu}_j(\lambda) \\
&= \sum_{j\text{ odd}} (j-1) \hat{\mu}_j(\lambda)
+ \sum_{j\text{ even}} j \hat{\mu}_j(\lambda).
\end{aligned}
$$
These last two sums are both even because
$2\hat{\mu}_j(\lambda)=\mu_j(\lambda)\equiv 0 $ for all $j$.
In case $B_n/C_n (ii)$,
Proposition~\ref{c(lambda)-congruences-lemma}
says that
$
\frac{c(\lambda)}{2} \equiv
\frac{j_0-1}{2}+\hat{\ell}(\lambda)-\hat{\mu}_{j_0}(\lambda)
.
$
Also $L(\lambda)=1$, so one has
$$
\begin{aligned}
E
&\equiv
n-\hat{\ell}(\lambda)
-\left(
\frac{j_0-1}{2}+\hat{\ell}(\lambda)-\hat{\mu}_{j_0}(\lambda)
\right)
-\frac{1}{4} +\sigma(\lambda)\\
&\equiv
n-\frac{j_0-1}{2}+\mu_{j_0}(\lambda)
-\frac{1}{4} +\sigma(\lambda)
\end{aligned}
$$
since $2 \hat{\ell}(\lambda) \equiv 0 $.
In addition, since $\lambda$ is either in $\Par_B(2n+1)$ or $\Par_C(2n)$,
one can rewrite $n$ as
$$
n = \left\lfloor\frac{|\lambda|}{2} \right\rfloor
= \left\lfloor\frac{j_0}{2} \right\rfloor
+ \sum_j j \hat{\mu}_j(\lambda)
\equiv \left\lfloor\frac{j_0}{2} \right\rfloor + \sum_{j\text{ odd }} \hat{\mu}_j(\lambda).
$$
Therefore
\begin{equation}
\label{processed-BC-expression-for-q-power}
E \equiv
\left\lfloor\frac{j_0}{2} \right\rfloor
-\frac{j_0-1}{2}
+ \sum_{j\text{ odd }} \hat{\mu}_j(\lambda)
+\hat{\mu}_{j_0}(\lambda)
-\frac{1}{4} +\sigma(\lambda).
\end{equation}
If $j_0$ is odd, so we are in type $B_n$, then
$\left\lfloor\frac{j_0}{2} \right\rfloor
=\frac{j_0-1}{2}$, and $\sigma(\lambda)=\tau_1(\lambda)+\frac{1}{4}$, so
\eqref{processed-BC-expression-for-q-power} becomes
$$
E \equiv
\sum_{j\text{ odd }} \hat{\mu}_j(\lambda)
+\hat{\mu}_{j_0}(\lambda)
+\tau_1(\lambda)
= \sum_{j \text{ odd}} 2 \hat{\mu}_j(\lambda)
\equiv 0 .
$$
If $j_0$ is even, so we are in type $C_n$, then
$\left\lfloor\frac{j_0}{2} \right\rfloor-
\frac{j_0-1}{2}=\frac{1}{2}$, and
$\sigma(\lambda)=\tau_0(\lambda)+\frac{1}{4}-\frac{\ell(\lambda)}{2}$,
so \eqref{processed-BC-expression-for-q-power} becomes
$$
E \equiv
\frac{1}{2}
+ \sum_{j\text{ odd }} \hat{\mu}_j(\lambda)
+\hat{\mu}_{j_0}(\lambda)
+\tau_0(\lambda)-\frac{\ell(\lambda)}{2}
= \frac{1-\ell(\lambda)}{2}
+ \sum_j \hat{\mu}_j(\lambda)
= 0.
$$
\vskip.1in
\noindent
{\sf Type D.}
In case $D_n(i)$, one needs to
compare the powers of $q$ in front of
\eqref{type-D-case-4-without-q-power}
and the $q$-Kreweras formula in
Theorem~\ref{q-Kreweras-formula-theorem}(type $D_n$).
Noting that in the case where
$\mu_1(\lambda) \equiv 2 \bmod{d^+}$, the
factor of $1+(-1)^{\frac{2n}{d}}$ vanishes unless $n \equiv 0 \bmod{d}$,
one finds that here one needs to show that this exponent
$$
E:=\BCDexponent + \tau_1(\lambda) + \frac{\ell(\lambda)}{2}
-\left\{
\begin{aligned}
\tau_1(\lambda) &\quad \text{ if }\mu_1(\lambda) \equiv 0 \bmod{d^+}
\text{ and }n \equiv 0 \bmod{d}, \\
\mu_1(\lambda) &\quad \text{ if }\mu_1(\lambda) \equiv 2 \bmod{d^+}
\end{aligned}
\right\}
$$
has $E \equiv 0 \bmod{d}$. Using
$m \equiv 1, L(\lambda)=0, \hat{\ell}(\lambda)=\frac{\ell(\lambda)}{2}$,
and since Proposition~\ref{c(lambda)-congruences-lemma}
gives $\frac{c(\lambda)}{2} \equiv 0$,
one has
$$
E \equiv
\begin{cases}
n &\text{ if }\mu_1(\lambda) \equiv 0 \bmod{d^+}
\text{ and }n \equiv 0 \bmod{d}, \\
n+\tau_1(\lambda) - \mu_1(\lambda) &\text{ if }\mu_1(\lambda) \equiv 2 \bmod{d^+}.
\end{cases}
$$
In the first case, the assumption of case $D_n (i)$ implies $n \equiv 0$.
In the second case, one can compute
$$
\begin{aligned}
E
&\equiv n+\tau_1(\lambda) -\mu_1(\lambda)\\
&= \sum_j j \hat{\mu}_j(\lambda)
+ \sum_{j\text{ odd }} \hat{\mu}_j(\lambda)
- \mu_1(\lambda) \\
&= \sum_{j \text{ even}} j \hat{\mu}_j(\lambda)
+ \sum_{\text{ odd } j\geq 3} (j+1) \hat{\mu}_j(\lambda)
\,\, + \,\, \left( 2\hat{\mu}_1(\lambda) - \mu_1(\lambda)\right) \\
&\equiv 0 + 0 + 0 = 0.
\end{aligned}
$$
In case $D_n(ii)$, one needs to show that the exponent
$$
\begin{aligned}
E&:=\BCDexponent +\tau_1(\lambda) + m-\frac{\ell(\lambda)}{2}+1 \\
&=\left( m(n-\hat{\ell}(\lambda))
-\frac{c(\lambda)}{2}-\frac{L(\lambda)}{4} \right)
+ \tau_1(\lambda)+ m-\frac{\ell(\lambda)}{2}+1
\end{aligned}
$$
has $E \equiv 0 \bmod{d}$.
Using the facts that
$m \equiv 1, L(\lambda)=2, \hat{\ell}(\lambda)=\frac{\ell(\lambda)}{2}$,
and since Proposition~\ref{c(lambda)-congruences-lemma} gives
$
\frac{c(\lambda)}{2} \equiv
\frac{j_0}{2}
+ \hat{\mu}_1(\lambda) + \hat{\mu}_{j_0}(\lambda),
$
one has
\begin{equation}
\label{last-D-case-q-power-expression}
E \equiv n-\ell(\lambda)+\tau_1(\lambda)+\frac{3}{2}
-\left(
\frac{j_0}{2}
+ \hat{\mu}_1(\lambda) + \hat{\mu}_{j_0}(\lambda)
\right).
\end{equation}
Note that since all $\mu_j(\lambda)$ are even except for $j=1,j_0$, one
has
$$
\begin{array}{rccccl}
n &=& \displaystyle \frac{|\lambda|}{2}
&=&\displaystyle \frac{1}{2}+ \frac{j_0}{2}+ \sum_j j \hat{\mu}_j(\lambda)
&\equiv \displaystyle \frac{j_0+1}{2}+ \sum_{j\text{ odd }} \hat{\mu}_j(\lambda) \\
& & & & & \\
\ell(\lambda)
&=& \sum_j \mu_j(\lambda)
&=& 2+\sum_j 2\hat{\mu}_j(\lambda)
&\equiv 2\\
& & & & & \\
\tau_1(\lambda)
&=&\displaystyle \sum_{\substack{j\text{ odd }:\\ \mu_j(\lambda)\text{ even}}} \frac{\mu_j(\lambda)}{2}
&=&\displaystyle\sum_{\substack{j\text{ odd }:\\ j \neq 1,j_0}} \hat{\mu}_j(\lambda)
&
\end{array}
$$
Thus one can rewrite \eqref{last-D-case-q-power-expression} as
$$
E \equiv
\frac{j_0+1}{2}+ \sum_{j\text{ odd }} \hat{\mu}_j(\lambda)
-2
+\sum_{\substack{j\text{ odd }:\\ j \neq 1,j_0}} \hat{\mu}_j(\lambda)
+\frac{3}{2}
-\left(
\frac{j_0}{2}
+ \hat{\mu}_1(\lambda) + \hat{\mu}_{j_0}(\lambda)
\right)
\equiv
\sum_{\substack{j\text{ odd }:\\ j \neq 1,j_0}} 2\hat{\mu}_j(\lambda)
\equiv 0.\qedhere
$$
\end{proof}
\subsection{Combinatorial models for $NC^{(s)}(w)$}
We review here for the classical types $A,B,C,D$
the combinatorial models for the elements of $NC^{(s)}(W)$, that
is, the $s$-element multichains $w_1 \leq \cdots \leq w_s$ in $NC(W)$.
We also review how to read off the $W$-orbit $[X]$ of
the subspace $X=V^{w_1}$, and the $\ZZ/sh\ZZ$-action on $NC^{(s)}(W)$.
\subsubsection{Type $A$}
\label{Type-A-combinatorial-model-section}
One can identify $NC^{(s)}(A_{n-1})$ with the
set of {\it $s$-divisible noncrossing
set partitions} of $\{1,2,\ldots,sn\}$, that is, those whose block sizes
are all divisible by $s$; see \cite[Chapter 3]{Armstrong1}.
Under this identification, if
$w_1 \leq \cdots \leq w_s$ corresponds to an $s$-divisible noncrossing
partition having block sizes $s\lambda=(s\lambda_1,\ldots,s\lambda_\ell)$
for some partition $\lambda$ of $n$, then the fixed space $V^{w_1}$ will lie in
the $W$-orbit $[X]$ where $W_X$ is the parabolic subgroup
$\symm_\lambda = \symm_{\lambda_1} \times \cdots \times \symm_{\lambda_\ell}$
inside $\symm_n$.
Here $h=n$, and Armstrong's $\ZZ/sh\ZZ$-action on $NC^{(s)}(A_{n-1})$ corresponds,
under this identification with $s$-divisible partitions,
to the $\ZZ/sn\ZZ$-action that cycles the label set $\{1,2,\ldots,sn\}$
within the blocks via
$$
\cdots \rightarrow 1 \rightarrow 2 \rightarrow \cdots \rightarrow sn-1
\rightarrow sn \rightarrow 1 \rightarrow \cdots .
$$
\subsubsection{Types $B, C$}
\label{Type-BC-combinatorial-model-section}
One can identify $NC^{(s)}(B_n)=NC^{(s)}(C_n)=$
with the subset of $s$-divisible noncrossing partitions of the label set $\{1,2,\ldots,2sn\}$
that are invariant under the involution
$\iota$ swapping lablels $i \leftrightarrow i+sn \bmod{2sn}$;
see \cite[\S 4.5]{Armstrong1}. For this reason, we relabel
$\{1,2,\ldots,2sn\}$ as
\begin{equation}
\label{type BCD-label-set}
\{+1,+2,\ldots,+sn,-1,-2,\ldots,-sn\}=\{\pm 1,\ldots,\pm n\}
\end{equation}
so that $\iota$ swaps $+i \rightarrow -i$ for each $i$.
Note that, since every block $B$ of the $s$-divisible noncrossing partition
must have $\iota(B)$ another block, the noncrossing condition implies that there
can be at most one block $B_0$ with the property $\iota(B_0)=B_0$; we call such
a block $B_0$, if it exists, a {\it zero block}, and call the pairs $\{B, \iota(B)\}$
with $\iota(B) \neq B$ the {\it nonzero blocks}.
Under the identification, if
$w_1 \leq \cdots \leq w_s$ corresponds to an $s$-divisible noncrossing
partition with nonzero blocks $\{B_1,\iota(B_1)\}, \ldots,\{B_\ell,\iota(B_\ell)\}$
of sizes $s\nu=(s\nu_1,\ldots,s\nu_\ell)$, then the
partition $\nu=(\nu_1,\ldots,\nu_\ell)$ will have $|\nu| \leq n$,
and the fixed space $V^{w_1}$ will lie in
the $W$-orbit $[X]$ where $W_X$ is the parabolic subgroup
$B_{n-|\nu|} \times \symm_{\nu_1} \times \cdots \times \symm_{\nu_\ell}$
inside $B_n$.
Here $h=2n$, and the $\ZZ/sh$-action corresponds to a
$\ZZ/2sn\ZZ$-action cycling the labels $\{\pm 1,\ldots, \pm sn\}$ via
$$
\cdots \rightarrow +1 \rightarrow +2 \rightarrow \cdots \rightarrow +sn
\rightarrow -1 \rightarrow -2 \rightarrow \cdots \rightarrow -sn
\rightarrow +1 \rightarrow \cdots .
$$
\subsubsection{Type $D$}
\label{Type-D-combintorial-model-section}
One can identify $NC^{(s)}(D_n)$
with certain set partitions of the same
label set of size $2sn$ as in
\eqref{type BCD-label-set}, but this time
arranged on the inner and outer
boundary of an {\it annulus}, where the $2s(n-1)$ labels
$$
+1,+2,\ldots,+s(n-1),-1,-2,\ldots,-s(n-1)
$$
appear in this order {\it clockwise} on the outer boundary, and
the remaining $2n$ labels
$$
\begin{aligned}
&+(s(n-1)+1), +s((n-1)+2),\ldots,+(sn-1),+sn,\\
&\quad -(s(n-1)+1), -s((n-1)+2),\ldots,-(sn-1),-sn,
\end{aligned}
$$
appear in this order {\it counterclockwise} on the inner boundary.
Given a block in such any such set partition, say that the block is
{\it entirely inner} (resp. {\it entirely outer}) if it only contains
labels from the inner (resp. outer) boundary; say that the block
is {\it traversing} if it is neither entirely inner nor entirely outer.
Then $NC^{(s)}(D_n)$ is identified with those set partitions that
satisfy these conditions:
\begin{itemize}
\item[NCD1] {\it noncrossing-ness}: the vertices within a block can all be connected by simple closed curves staying within the annulus, in such a way that curves corresponding to distinct blocks do not cross.
\item[NCD2]{\it $\iota$-stability}: if $B$ is a block, then $\iota(B)$ is also a block.
\item[NCD3]{\it zero-block closure}: if there exists a zero-block $B_0=\iota(B_0)$, then it is unique, and contains all the labels on the inner boundary.
\item[NCD4] {\it strong $s$-divisibility}: when reading elements of a block in any clockwise
order coming from the planar embedding, their sequence of absolute values
pass through consecutive residue classes in $\ZZ/s\ZZ$.
\item[NCD5] {\it determinacy}: if there are no traversing blocks, then
the outer blocks completely determine the inner blocks in a certain fashion, whose
details are not important to us here; see \cite[\S 7]{Kim1} or \cite[\S 7]{Rhoades}.
\end{itemize}
The conditions NCD1-4 were given by Krattenthaler-M\"uller \cite[\S 7]{KrattenthalerMuller};
while condition NCD5 was inadvertently omitted, and recorded by
J. S. Kim \cite[\S 7]{Kim1}.
Similarly to types $B_n, C_n$,
under the identification, if
$w_1 \leq \cdots \leq w_s$ corresponds to a partition
with nonzero blocks $\{B_1,\iota(B_1)\}, \ldots,\{B_\ell,\iota(B_\ell)\}$
of sizes $s\nu=(s\nu_1,\ldots,s\nu_\ell)$, then the
partition $\nu=(\nu_1,\ldots,\nu_\ell)$ will either have
$|\nu|=n$ if there is no zero block or
$|\nu| \leq n-2$ if there is a zero block.
Then the fixed space $V^{w_1}$ will lie in
the $W$-orbit $[X]$ where $W_X$ is the parabolic subgroup
$D_{n-|\nu|} \times \symm_{\nu_1} \times \cdots \times \symm_{\nu_\ell}$
inside $D_n$.
Here $h=2(n-1)$, and Rhoades \cite[\S 7]{Rhoades}
observed that the $\ZZ/sh$-action corresponds to
the $\ZZ/2s(n-1)$-action simultaneously
\begin{itemize}
\item cycling the outer labels ({\it clockwise}) via
$$
\cdots \rightarrow -s(n-1)
\rightarrow +1 \rightarrow +2 \rightarrow \cdots \rightarrow +s(n-1)
\rightarrow -1 \rightarrow -2 \rightarrow \cdots \rightarrow -s(n-1)
\rightarrow +1 \rightarrow \cdots
$$
\item cycling the inner labels ({\it counterclockwise}) via
$$
\begin{aligned}
& \cdots \rightarrow -sn
\rightarrow +(s(n-1)+1) \rightarrow +(s(n-1)+2) \rightarrow \cdots \rightarrow +sn \\
&\qquad \rightarrow
-(s(n-1)+1) \rightarrow -(s(n-1)+2) \rightarrow \cdots \rightarrow -sn
\rightarrow +(s(n-1)+1) \rightarrow \cdots .
\end{aligned}
$$
\end{itemize}
\subsection{Putting it together}
We now assemble the proof of Theorem~\ref{CSP-theorem} in each type $A,B,C,D$.
Thus we assume that $m=sh+1$ for a positive integer $s$, and that $d$ is a divisor of $m-1=sh$.
In each case, the strategy will be to first show that if $w_1 \leq \cdots \leq w_s$ is an element in $NC^{(s)}(W)$
having {\it $d$-fold symmetry}, that is, fixed by
an element $c^{\frac{sh}{d}}$ having order $d$ in the cyclic group
$C=\langle c \rangle \cong \ZZ/sh\ZZ$, then the parabolic
subgroup $W_X$ fixing $X=V^{w_1}$ corresponds to a principal-in-a-Levi
nilpotent orbit $\0_\lambda$ that falls into one of the
cases from Proposition~\ref{evaluations-prop}.
where $\Krew(\Phi,\0_\lambda,m,q=\omega_d)$ is nonvanishing.
Then we will count the number of such $d$-fold symmetric elements, generally relying on known formulas or bijections to objects with known formulas, and see that they agree with
the evaluations in Proposition~\ref{evaluations-prop}.
\subsubsection{Type A}
As in Section~\ref{Type-A-combinatorial-model-section},
we have identified the elements of $NC^{(s)}(A_{n-1})$
with the $s$-divisible noncrossing partition of $\{1,2,\ldots,sn\}$,
that is, those noncrossing partitions having block
sizes $s\lambda$ for some $\lambda$ with $|\lambda|=n$.
When $d=1$, one needs to count those which are
fixed by the identity element in $\ZZ/sn\ZZ$,
which are all such elements. In this case, formula
in Proposition~\ref{evaluations-prop}(i) agrees
with Kreweras's original count for such partitions.
If $d \geq 2$ and the $s$-divisible partition
has $d$-fold symmetry, then most of blocks will be in orbits of
size $d$, with at most one block which is itself $d$-fold symmetric--
two such blocks would cross each other.
Thus either $\lambda$ has $\mu_j(\lambda) \equiv 0 \bmod{d}$ for all $j$,
or there exists one $j_0$ having $\mu_j(\lambda) \equiv 1 \bmod{d}$,
matching the description for when
$\Krew(\Phi,\0_\lambda,m,q=\omega_d)$ is nonvanishing from
Proposition~\ref{evaluations-prop}(i).
The proof that in this situation there are exactly
\begin{equation}
\label{Athanasiadis-formula}
\binomial{\frac{sn}{d}}{\lfloor \frac{\mu(\lambda)}{d} \rfloor}
\end{equation}
elements with $d$-fold symmetry,
as in Proposition~\ref{evaluations-prop}(i), was
sketched in \cite[Theorem 6.2]{BessisR}, but we repeat it here
for completeness. Such an element is completely determined by
restricting each of its blocks to its intersection with the subset
$\{1,2,\ldots,\frac{2sn}{d}\}$; relabel these numbers
$
\{ 1,2,\ldots,\frac{sn}{d},-1,-2,\ldots,-\frac{sn}{d} \}.
$
If there is a (unique) $d$-fold symmetric $j_0$-block, then call its
restriction the ``zero block''.
It is easily seen that this gives a bijection to
the type $B_{\frac{sn}{d}}$
noncrossing partitions considered in \cite{Reiner} having
$\hat{\mu}_j(\lambda)$ nonzero blocks of size $sj$ for each $j$.
The formula \eqref{Athanasiadis-formula}
then agrees with the count for such type $B$ noncrossing partitions given
by Athanasiadis \cite[Theorem 2.3]{Athanasiadis-block-sizes}.
\subsubsection{Types B, C}
As in Section~\ref{Type-BC-combinatorial-model-section},
we are identifying the elements of $NC^{(s)}(B_n)$ or
$NC^{(s)}(C_n)$ with the
$s$-divisible $\iota$-stable
noncrossing partition of $\{\pm 1, \pm 2,\ldots,\pm sn\}$.
Such a noncrossing partition
has nonzero blocks $\{B_1,\iota(B_1)\},\ldots,\{B_\ell,\iota(B_\ell)\}$
of sizes $s\nu$ where $\nu=(\nu_1,\ldots,\nu_\ell)$ with $|\nu|\leq n$.
The principal-in-a-Levi nilpotent orbit $\0_\lambda$
corresponding to the same parabolic subgroup $W_X$
has $\lambda=(\nu,\nu,N-2|\nu|)$ where $N=2n+1$ in type $B_n$
and $N=2n$ in type $C_n$.
If the noncrossing partition additionally has $d$-fold symmetry,
then the nonzero blocks will all lie in orbits of
size $d$. That is, only the zero block (if present) can have
fewer than $d$ distinct images under the $d$-fold symmetry, and will in fact,
be itself $d$-fold symmetric.
As there will be $2\mu_j(\nu)=2\hat{\mu}_j(\lambda)$
nonzero blocks of size $sj$
for each $j$, this means that $2\hat{\mu}_j(\lambda) \equiv 0 \bmod{d}$
for each $j$, matching the description for when
$\Krew(\Phi,\0_\lambda,m,q=\omega_d)$ is nonvanishing from
Proposition~\ref{evaluations-prop}(ii).
Similarly to type $A$, there will be exactly
\begin{equation}
\label{Athanasiadis-formula-again}
\binomial{\frac{2sn}{d^+}}{\frac{2\mu(\lambda)}{d^+}}=
\binomial{\frac{sn}{d^-}}{\frac{\mu(\lambda)}{d^-}}
\end{equation}
elements with $d$-fold symmetry,
matching Proposition~\ref{evaluations-prop}(ii):
restricting each block to its intersection with the subset
$\{\pm 1,\pm 2,\ldots,\pm \frac{sn}{d}\}$ gives a bijection to
the type $B_{\frac{sn}{d}}$
noncrossing partitions having $\hat{\mu}_j(\lambda)$
nonzero blocks of size $sj$ for each $j$,
and \eqref{Athanasiadis-formula-again} again agrees with
Athanasiadis' count \cite[Theorem 2.3]{Athanasiadis-block-sizes}
for this.
\begin{comment}
\begin{example}
If $d=4, s=2, n=3$ and $\nu=(1,1)$,
and Figure~\ref{symmetric-ncs-figure} shows one of the
three $2$-divisible partitions of type $B_3$ having $4$-fold
symmetry. Here $d^+=4$ and $2sn=12$, and $\nu$ has only one nonzero
multiplicity, namely $\mu_1(\nu)=2$, so that
Proposition~\ref{evaluations-prop}(ii)
also predicts that there should be
$\binomial{ \frac{12}{4} }{ \frac{2 \cdot 2}{4} }
=\binomial{ 3 }{ 1 }=3$
such partitions.
\begin{figure}
\epsfxsize=50mm
\epsfbox{symmetric-ncs.eps}
\caption{The case $d=4, s=2, n=3$ and $\nu=(1,1)$ of the type $B_n$
noncrossing $s$-divisible partitions with $d$-fold symmetry. There are 3 such
partitions: the one shown, along with its rotations by $30$ and by $60$ degrees.}
\label{symmetric-ncs-figure}
\end{figure}
\end{example}
\end{comment}
\subsubsection{Type D}
As in Section~\ref{Type-D-combintorial-model-section},
we are identifying the elements of $NC^{(s)}(D_n)$
with certain ``annular''
noncrossing partitions.
It is convenient to consider separately the two cases
where a zero block is present or absent; these will correspond to the
two cases in Proposition~\ref{evaluations-prop}(iii)
where $\hat{L}(\lambda)=1$ and $\hat{L}(\lambda)=0$, respectively.
\vskip.1in
\noindent
{\sf The case with a zero block present.}
Due to the {\it zero-block closure condition $NCD3$},
removing the $2s$ elements on the inner boundary
of the annulus from the zero block gives a bijection between
the subset of elements of $NC^{(s)}(D_n)$ having a zero block, and
the whole set $NC^{(s)}(B_{n-1})$. Furthermore, this bijection is
equivariant with respect to the $\ZZ/s(n-1)\ZZ$-action on these sets.
Just as in the type $B/C$ case, let
us assume that the nonzero blocks
$\{B_1,\iota(B_1)\},\ldots,\{B_\ell,\iota(B_\ell)\}$
have sizes $s\nu$ where $\nu=(\nu_1,\ldots,\nu_\ell)$ with $|\nu|\leq n-2$.
Then the principal-in-a-Levi nilpotent orbit $\0_\lambda$
corresponding to the same parabolic subgroup $W_X$
has $\lambda=(\nu,\nu,1,j_0)$ where $j_0=2n-1-2|\nu|$.
Again as in the type $B/C$ case,
$d$-fold symmetry implies that the nonzero blocks all lie in orbits of
size $d$. Hence there will be $2\mu_j(\nu)=2\hat{\mu}_j(\lambda)$
nonzero blocks of size $sj$ for each $j$,
so $2\hat{\mu}_j(\lambda) \equiv 0 \bmod{d}$
for each $j$. Also the number of such $d$-fold symmetric elements
should be the same as the formula
in Proposition~\ref{evaluations-prop}(ii), replacing
$n$ by $n-1$. This exactly matches the conditions
and the formula in the $\hat{L}(\lambda)=1$ case of
Proposition~\ref{evaluations-prop}(iii).
\vskip.1in
\noindent
{\sf The case with no zero block.}
Again assume that the nonzero blocks
$\{B_1,\iota(B_1)\},\ldots,\{B_\ell,\iota(B_\ell)\}$
have sizes $s\nu$ where $\nu=(\nu_1,\ldots,\nu_\ell)$ with $|\nu|= n$.
Then the principal-in-a-Levi nilpotent orbit $\0_\lambda$
corresponding to the same parabolic subgroup $W_X$
has $\lambda=(\nu,\nu)$.
We are claiming that this case will match with the
conditions and formulas appearing in
the $\hat{L}(\lambda)=0$ cases of
Proposition~\ref{evaluations-prop}(iii).
Thus one should expect the analysis to
break into further subcases based on whether $d=1,2$ or at least $3$.
\vskip.1in
\noindent
{\sf The subcase with no zero block and $d=1$.}
Here one wishes to count {\it all} of the elements of $NC^{(s)}(D_n)$ whose
annular noncrossing partition has no zero block and
nonzero block sizes $s\nu$. This is given by a
formula of Krattenthaler and M\"uller \cite[Corollary 16]{KrattenthalerMuller},
\cite[Theorem 1.2]{Kim2}. Using the formulation in
\cite[Theorem 1.2]{Kim2}, and the
notational correspondences $\ell=1$, $b=\mu$, $k=s$, so that $s_1=n-b, s_2=b$,
one obtains a formula equivalent to
the $\hat{L}(\lambda)=0$ case of
Proposition~\ref{evaluations-prop}(iii) with $d=1$.
\vskip.1in
\noindent
{\sf The subcase with no zero block and $d=2$.}
Here one wishes to count the elements of $NC^{(s)}(D_n)$ whose
annular noncrossing partition has no zero block and
nonzero block sizes $s\nu$, and with the additional property
that they are fixed by the
element $c^{s(n-1)}$ of order $2$ inside the cyclic group
$C \cong \ZZ/2s(n-1)\ZZ$. Note that this
element has different effects on the outer and inner boundary
labels: on the outer boundary it rotates $180$ degrees,
acting as the map $\iota: i \leftrightarrow -i$, while on the inner boundary
it iterates the $180$-degree rotation $n-1$ times, acting as $\iota^{n-1}$.
Thus whenever the
annular noncrossing partition has two (nonzero) inner blocks of
size $s$, it is always fixed by $c^{s(n-1)}$: all blocks
will be either entirely inner or outer, and hence
the {\it $\iota$-stability condition $NCD2$} implies the stability
of the blocks under any power of $\iota$.
If the annular noncrossing partition does not have two inner blocks
of size $s$, then all of the labels on its inner boundaries lie
in traversing blocks. We claim that it is then fixed by
$c^{s(n-1)}$ if and only if $n$ is even: on the outer
labels one is acting by $\iota$, and
on the inner labels one is acting by $\iota^{n-1}$,
so one needs $\iota^{n-1}=\iota$ to be fixed, that is,
$n$ even.
Thus for $n$ even, all of these elements are fixed by
$c^{s(n-1)}$, and therefore should have the same formula as in
Proposition~\ref{evaluations-prop}(iii) with $d=1$. Indeed
this agrees with the formula in Proposition~\ref{evaluations-prop}(iii)
when $\hat{L}(\lambda)=0$ and $d=2$ and $n$ even.
Meanwhile for $n$ odd, the elements fixed by
$c^{s(n-1)}$ are those with two (nonzero) inner blocks of size $s$.
Removing these inner blocks gives an easy bijection to the elements
of $NC^{(s)}(B_{n-1})$ with all nonzero blocks, of sizes
$s\lambda-(s,s)$, where $s\lambda-(s,s)$ is obtained from
$s\lambda$ by removing two copies of the part $s$ (corresponding to
these inner blocks). Therefore this should have the same formula
as Proposition~\ref{evaluations-prop}(ii) but replacing $\lambda$ with
$\lambda-(1,1)$, namely
$$
\binomial{s(n-1)}{\hat{\mu}_1(\lambda)-1, \hat{\mu}_{\geq 2}(\lambda)}.
$$
Happily, one has an easily checked identity
\begin{equation}
\label{easily-checked-identity}
\binomial{s(n-1)}{\hat{\mu}_1(\lambda)-1, \hat{\mu}_{\geq 2}(\lambda)}
=
\binomial{s(n-1)}{\hat{\mu}_{\geq 2}(\lambda)}
\binomial{s(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|}{\hat{\mu}_1(\lambda)}
-\binomial{s(n-1)}{\hat{\mu}(\lambda)},
\end{equation}
whose right side agrees with the value given in
Proposition~\ref{evaluations-prop}(iii)
when $\hat{L}(\lambda)=0$ and $d=2$ and $n$ is odd.
\vskip.1in
\noindent
{\sf The subcase with no zero block and $d \geq 3$.}
Note that there is a dichotomy in the conditions and formulas in
Proposition~\ref{evaluations-prop}(iii)
when $\hat{L}(\lambda)=0$ and $d \geq 3$.
This will correspond to the following dichotomy in the
annular noncrossing partitions that model $NC^{(s)}(D_n)$ and have
no zero block.
\begin{proposition}
\label{NCD-traversing-subtlety}
In annular noncrossing partitions modelling
$NC^{(s)}(D_n)$ with no zero block, either
\begin{itemize}
\item[(A)] every label on the inner boundary lies in a traversing block, or
\item[(B)] there are no traversing
blocks, only entirely outer blocks, and two entirely
inner blocks each of size $s$.
\end{itemize}
\end{proposition}
\begin{proof}
Assume case (B) fails, that is, some inner boundary label $j$
lies in a traversing block $B$.
Then $-j$ lies in $\iota(B)$, which will be a {\it different}
traversing block, since we are assuming that there
is no zero block. But then the
the strong $s$-divisibility condition $NCD4$ now
prevents any of the inner boundary labels from lying in an entirely
inner block: $NCD4$ implies that such a traversing block
would contain elements having absolute values that achieve
every residue class in $\ZZ/s\ZZ$, which is impossible
since $j$ and $-j$ are the only inner boundary labels
whose absolute values achieve their residue class.
\end{proof}
An immediate corollary is that if an element of $NC^{(s)}(D_n)$
corresponds to an annular noncrossing partition with no zero blocks and
has $d$-fold symmetry, then either it is in case (A) of
Proposition~\ref{NCD-traversing-subtlety} and its blocks all come in orbits of
size $d$ (as they are all entirely outer or traversing),
or it is in case (B), so that its entirely outer blocks come in orbits of size $d$, leaving only the two entirely inner blocks (each of size $s$).
This means that its block sizes $s\lambda$ will satisfy
$$
\begin{aligned}
\mu_j(\lambda)(=2\hat{\mu}_j(\lambda))
&\equiv 0 \bmod{d^+}\text{ for each }j \geq 2,\\
\mu_1(\lambda)(=2\hat{\mu}_1(\lambda))
&\equiv
\begin{cases}
0 \bmod{d^+}&\text{ in case (A),}\\
2 \bmod{d^+}&\text{ in case (B).}
\end{cases}
\end{aligned}
$$
Note that this matches the dichotomy of conditions in
Proposition~\ref{evaluations-prop}(iii)
when $\hat{L}(\lambda)=0$. It remains only to check
that the formulas there match the number of $d$-fold
symmetric elements for $d \geq 3$ in each case.
In case (B), we claim that the two entirely inner blocks of size $s$
are always stable under the element $c^{\frac{2s(n-1)}{d}}$ of order $d$
inside the cyclic group $C=\langle c \rangle$. This is because
one has
$$
2n= \sum_j j \mu_j(\lambda) \equiv 2 \bmod{d^+}
$$
so that $d^+$ divides $2(n-1)$, and hence $s$ divides $\frac{2s(n-1)}{d}$.
Thus these two inner blocks are always $d$-fold symmetric, and they are completely determined by the entirely outer blocks (according to the {\it determinacy condition} $NCD5$). Therefore
the number of $d$-fold symmetric elements in case (B) is
the same as the number of $d$-fold symmetric elements of type $B_{n-1}$ having
block sizes $s\lambda-(s,s)$, that is, the formula from
Proposition~\ref{evaluations-prop}(ii),
but replacing $\lambda$ with $\lambda-(1,1)$:
$$
\binom{\frac{2s(n-1)}{d^+}}
{\frac{\mu_1(\lambda)-2}{d^+},\frac{\mu_{\geq 2}(\lambda)}{d^+}}.
$$
This matches the desired formula in
Proposition~\ref{evaluations-prop}(iii), with
$\hat{L}(\lambda)=0, d \geq 3$ and
$\hat{\mu}_1(\lambda) \equiv 0 \bmod{d^+}$.
In case (A), we need a further structural observation.
\begin{lemma}
\label{counterclockwise-becomes-clockwise-lemma}
For $d \geq 3$, in case (A),
the annular noncrossing partition
is fixed by $c^{\frac{2s(n-1)}{d}}$
if and only if it has both $d$-fold {\it rotational symmetry} and in
addition, $d$ divides $n$.
\end{lemma}
\begin{proof}
Since in case (A) one has $d$ dividing $2n=\sum_j \mu_j(\lambda)$,
and $d$ also dividing $2s(n-1)$, one concludes that $d$ divides $2s$.
Thus $d$ divides the number of labels on both the inner boundary and the outer boundary.
Given any block $B$ of a $d$-fold symmetric
annular noncrossing partition in case (A),
decompose it uniquely as $B=B_i \sqcup B_o$ with inner boundary labels $B_i$
and outer boundary labels $B_o$.
Then under the action of $g:=c^{\frac{2s(n-1)}{d}}$
one must have $d$ disjoint images
$B_o,g(B_o),\ldots,g^{d-1}(B_o)$ reading {\it clockwise}, and
also $d$ disjoint images
$B_i,g(B_i),\ldots,g^{d-1}(B_i)$ reading {\it counterclockwise}. However,
since each of the sets $g^j(B)=g^j(B_i) \sqcup g^j(B_o)$ is another
block, the {\it noncrossing condition $NCD1$} (and $d \geq 3$)
implies that the counterclockwise ordering
$B_i,g(B_i),\ldots,g^{d-1}(B_i)$ must actually also be {\it clockwise}. In
other words, the partition has $d$-fold rotational symmetry.
Furthermore, since the inner boundary has $2s$ elements
and rotating it $\frac{2s(n-1)}{d}$ steps counterclockwise is the
same as rotating it $\frac{2s}{d}$ steps clockwise, one concludes
that
$$
\frac{-2s(n-1)}{d} \equiv \frac{+2s}{d} \bmod{2s}, \qquad \text{ that is},
\frac{2sn}{d} \equiv 0 \bmod{2s}, \qquad \text{ or }d\text{ divides }n.
$$
Conversely, when $d$ divides $n$ and the partition
has $d$-fold rotational symmetry, one can reverse the above arguments to see that it is fixed by $c^{\frac{2s(n-1)}{d}}$.
\end{proof}
We can now complete the comparison of the number of
$d$-fold symmetric elements in case (A) of
Proposition~\ref{NCD-traversing-subtlety}
with the formula in Proposition~\ref{evaluations-prop}(iii)
at $\hat{L}(\lambda)=0$ and $d \geq 3$
\begin{equation}
\label{last-sub-sub-case-formula}
\left( 1+(-1)^{\frac{2n}{d}} \right)
\binomial{\frac{2s(n-1)}{d^+}}{\frac{\mu(\lambda)}{d^+}}
=\begin{cases}
2\binomial{\frac{2s(n-1)}{d^+}}{\frac{\mu(\lambda)}{d^+}}& \text{ if }d\text{ divides }n,\\
0 & \text{ if }d\text{ does not divide }n.
\end{cases}.
\end{equation}
As noted at the start of the
proof of Lemma~\ref{counterclockwise-becomes-clockwise-lemma},
the assumptions of case (A) imply that
$d$ divides $2n$.
If $d$ does not divide $n$,
Lemma~\ref{counterclockwise-becomes-clockwise-lemma}
implies that there are no $d$-fold symmetric elements,
matching the value of $0$ on the right
in \eqref{last-sub-sub-case-formula}.
So assume that $d$ does divide $n$.
By Lemma~\ref{counterclockwise-becomes-clockwise-lemma} we must count
the annular noncrossing partitions modelling elements in $NC^{(s)}(D_n)$
which lie in case (A) of Proposition~\ref{NCD-traversing-subtlety},
having block sizes $s\lambda$, and which are additionally
$d$-fold rotationally symmetric. The rotational symmetry means that such
a partition is completely determined by restricting its blocks to the
$\frac{2s(n-1)}{d^-}$
outer labels $\{\pm 1,\pm 2,\ldots, \pm \frac{s(n-1)}{d^-} \}$; the
$\frac{2s}{d^-}$ inner labels that accompany these blocks
are determined by the {\it strong $s$-divisibility condition $NCD4$}.
Since $d$ divides $m-1=2s(n-1)$,
it also divides $2s$. Thus, setting $\tilde{s}:=\frac{s}{d^-}$,
the well-defined reduction map $\ZZ/s\ZZ \rightarrow \ZZ/\tilde{s}\ZZ$
shows that these blocks will also satisfy the
{\it strong $\tilde{s}$-divisibility condition $NCD4$}
and determine a unique element of
$NC^{\tilde{s}}(D_n)$. Its blocks still have sizes of the
form $sj$, of course, and the number of blocks of size $sj$ will
be $\frac{\mu_{sj}(s\lambda)}{d^-}=\frac{\mu_{j}(\lambda)}{d^-}$.
This means that this element of $NC^{\tilde{s}}(D_n)$
has block sizes
$\tilde{s} \tilde{\lambda}$ where $\tilde{\lambda}$ only has parts
of the form $d^- \cdot j$. Note that $d \geq 3$ forces $d^- \geq 2$, so that
$\mu_1(\tilde{\lambda})=0$
and $\mu_{\geq 2}(\tilde{\lambda})=\mu(\tilde{\lambda})$. Also,
$
\mu_{d^- \cdot j}(\tilde{\lambda})=\frac{\mu_{sj}(s\lambda)}{d^-}=\frac{\mu_{j}(\lambda)}{d^-}.
$
Hence the number of such elements is counted by
the formula of Krattenthaler and M\"uller already mentioned,
that is, the $\hat{L}(\lambda)=0$ case of
Proposition~\ref{evaluations-prop}(iii) with $d=1$, replacing $s$ with $\tilde{s}$:
$$
\begin{aligned}
&
\binomial{\tilde{s}(n-1)}{\hat{\mu}_{\geq 2}(\lambda)}
\binomial{\tilde{s}(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{\hat{\mu}_1(\lambda)}
+ \binomial{\tilde{s}(n-1)}{\hat{\mu}(\lambda)} \\
&=
\binomial{\tilde{s}(n-1)}{\frac{\hat{\mu}(\lambda)}{d^-}}
\binomial{\tilde{s}(n-1)+1-|\hat{\mu}_{\geq 2}(\lambda)|}
{0}
+ \binomial{\tilde{s}(n-1)}{\frac{\hat{\mu}(\lambda)}{d^-}}
=2 \binomial{\tilde{s}(n-1)}{\frac{\hat{\mu}(\lambda)}{d^-}}
=2 \binomial{\frac{2s(n-1)}{d^+}}{ \frac{\mu(\lambda)}{d^+} },
\end{aligned}
$$
which matches \eqref{last-sub-sub-case-formula} when $d$ divides $n$.
The completes the proof of Theorem~\ref{CSP-theorem}.
\section*{Acknowledgments}
The first author thanks Jang Soo Kim for helpful conversations.
|
\subsection*{Abstract}
Censors of the Internet must continually discover and block
new circumvention proxy servers.
We seek to understand this process; specifically,
the length of the delay between when a proxy first becomes
discoverable and when it is actually blocked.
We measure this delay in the case of obfuscated Tor bridges,
by testing their reachability
before and after their introduction into Tor Browser.
We test from sites in the U.S., China, and Iran,
over a period of five months.
China's national firewall blocked new bridges
after a varying delay of between 2 and 36 days.
Blocking occurred only after end-user software releases,
despite bridges being potentially discoverable earlier through other channels.
While the firewall eventually discovered the bridges of Tor Browser,
those that appeared only in Orbot,
a version of Tor for mobile devices, remained unblocked.
Our findings highlight the fact that censors can behave
in ways that defy intuition,
presenting difficulties for threat modeling but
also opportunities for evasion.
\section{Introduction}
Censors of the Internet must seek out and block
proxy servers that can be used to evade their information controls.
Here we explore the mechanics of this process as it applies
to the blocking of default Tor bridges
after they are published in Tor Browser.
It is known that these default bridges are eventually blocked;
what is not known is exactly how long it takes.
We measure the ``delay'' or ``lag'' of proxy blocking,
by testing the reachability of bridges
before and after their first public disclosure
from sites in the U.S., China, and Iran.
There is prior work on distribution strategies that
prevent the censor from discovering secret proxy addresses in the first place;
examples are
Proximax~\cite{McCoy2011a} and
rBridge~\cite{Wang2013a}.
Our work is different: we study how censors block proxies
that are not secret, but are (in principle) easily discoverable by anyone.
In typical censorship threat models,
such unprotected proxies would be considered to be immediately blocked.
That they are not, in practice,
shows that practical considerations
may prevent censors from exercising all their assumed capabilities.
Zhu et~al.~\cite{Zhu2013a} in 2013 explored a related idea.
They measured how long it took for posts to be censored
on the Chinese microblogging service Sina Weibo.
They found that 30\% of posts are deleted within 30~minutes
and 90\% are deleted within 24~hours, though some posts
survived for weeks or months.
They used their results to hypothesize about the reasons for
and mechanisms of microblog censorship.
In 2014, developers of OONI, the Open Observatory of Network Interference,
a censorship measurement platform, implemented a new test to
check the reachability of Tor bridges
and prepared visualizations~\cite{ooni-bridge-reachability-study-and-hackfest}.
The tests have not run continually since then.
In 2015, a one-off calculation~\cite{tor-dev-censorship-lag}
based on user reports
found loose time bounds for the delay of a single bridge-blocking event in China:
somewhere between 15 and 76 days.
Knowing the proxy blocking delay gives insight
into how censors work:
where they look for new bridges, and
whether their blacklist updates are automatic or manual.
It advances our understanding the operational costs incurred by censors,
and therefore their potential weaknesses.
We think of censors as complex systems,
consisting of interacting human and machine
components, whose goals and
motivations are sometimes in conflict.
\section{Background}
Tor~\cite{tor} is an anonymity network
that is also widely used to circumvent censorship.
In its natural form, Tor is poorly suited to circumvention.
Its nodes' IP addresses are public,
and the protocol itself is fairly distinctive.
But Tor combined with \emph{bridges}
and \emph{pluggable transports} is much harder to block.
Bridges are special, unlisted nodes
whose addresses are not easily discoverable in bulk---users
must acquire a few at a time through an online database called BridgeDB~\cite{bridgedb}.
Pluggable transports
are obfuscation protocols that encapsulate the Tor protocol,
making it difficult to detect.
A censored user uses a pluggable transport
in order to reach a secret bridge,
foiling both deep packet inspection
and IP address blacklisting.
Tor Browser~\cite{torbrowser} is a modified version of Firefox
that features a built-in always-on Tor client.
It is the recommended way of accessing the Tor network for most users.
The browser has a graphical interface
for the configuration of bridges and pluggable transports.
The intended use case has
users acquiring a bridge address through a side channel,
such as email or word of mouth,
and pasting it into the configuration interface.
However, in practice, circumvention is often even easier,
requiring no out-of-band information.
For many users, it suffices to select a pluggable transport name from a menu,
causing Tor to connect to one of
a handful of built-in, \emph{default} bridges
included with the browser.
The concept of a default bridge needs some explanation.
Bridges are supposed to be secret,
so including them with the browser seems to be self-defeating.
The default bridges are in fact open to the world
in the Tor Browser source code---the file housing them is called
bridge\_prefs.js~\cite{torbrowser-bridgeprefs}.
Any censor that can block Tor ought to
be able to block the default bridges as well,
but the strange truth of the matter is that many simply do not.
We can only guess as to why:
it could be that censors are negligent or incompetent;
perhaps Tor traffic is, from their point of view, too inconsequential to bother with;
or maybe they set more stock in deep packet inspection
and dynamic protocol detection than in static IP address blacklists.
Whatever the reason, Tor Browser continues to ship
default bridges for the simple reason that they work for many people.
In fact, the Great Firewall of China is the only state censor
we are aware of that makes some effort to block Tor's default bridges.
This is what we study in this work:
the time delay in blocking a bridge after it is first made public
in a Tor Browser release.
\subsection{Tor Browser releases}
Tor Browser releases follow two tracks: stable and alpha.
The stable track changes slowly,
its minor releases typically including only bugfixes.
The alpha track has new and experimental features.
The alpha track matures until it becomes the basis
of the next major release of the stable track.
(There are also special ``hardened'' releases
that track the alphas---we do not consider them separately.)
In addition to formal major and minor releases,
there are nightly builds that have the latest of everything.
The primary driver of Tor Browser releases is
upstream security fixes in Firefox.
When there is a new version of Firefox
that fixes a security vulnerability,
Tor Browser developers must hurry to
build a new minor release of Tor Browser with the fix.
For this reason, new stable and alpha releases
usually appear at about the same time,
because both are usually
equally affected by Firefox bugs.
Alpha releases are distinguished by a letter ``a''
in the version number;
for example, the stable release 5.5.5
appeared at about the same time as
the alpha release 6.0a5.
During the roughly five-month period of our experiments,
there were
ten stable releases (two major, eight minor), and
seven alpha releases (one major, six minor)~\cite{torbrowser-changelog}.
Each releases was an opportunity to
deploy and measure new bridges.
\subsection{The process of releasing a new bridge}
\label{sec:bridge-process}
There is no single moment when a new bridge becomes public.
The process of adding a bridge involves multiple stages,
each of which potentially reveals it,
depending on how closely the censor pays attention.
\begin{enumerate}
\item \textbf{Ticket filed.}
New default bridges are proposed for inclusion
by the filing of a ticket in Tor's
online bug tracker.
A censor paying attention to the bug tracker
would learn of bridges at this stage.
\item \textbf{Ticket merged.}
When the ticket is merged,
the new bridge is added to Tor Browser's source code.
From there, it will automatically be incorporated
into nightly builds. This is the first time that the code
containing the new bridge is available in executable form.
A censor paying attention to the source code repository,
or following nightly builds,
would learn of bridges at this stage.
\item \textbf{Testing release.}
When it is time for a new release,
Tor Browser developers prepare candidate packages
and send them to a quality assurance mailing list
for testing.
A censor paying attention to the mailing list
would learn of bridges at this stage.
\item \textbf{Public release.}
After testing, new packages are announced on the Tor blog,
and already installed Tor Browsers will automatically update themselves.
This is the stage at which the new bridge
will start to be used by real users.
A censor paying attention to the blog,
or black-box testing an auto-updating installation,
would learn of bridges at this stage.
\end{enumerate}
From ticket filing to public release, the process usually takes a few weeks.
Sometimes the testing release stage is skipped if the new version
only fixes a small error in the previous version, like
a packaging or localization bug.
The releases
of the stable and alpha tracks are separate; they are, however,
usually close in time.
New default bridges are sometimes discussed in private mailing lists
even before a ticket is filed.
A censor could conceivably infiltrate an internal mailing list
and learn about new bridges very early.
We have assumed that this does not happen,
that censors must use the same public channels as everyone else.
\subsection{The special properties of obfs4}
\begin{figure*}
\includegraphics{data/timelines}
\caption{
Timeline of Tor Browser default bridge blocking,
as measured from one probe site in China.
Black vertical lines indicate releases.
All new bridges within a release are blocked
within a few days or weeks.
The ``timed out'' and ``connection refused'' results
for the control bridges on port 1984
were the result of temporary misconfigurations, not blocking.
The vertical stripe on May~20 shows when the probe site
changed its IP address.
}
\label{fig:timelines}
\end{figure*}
\begin{figure*}
\includegraphics{data/comparison}
\caption{
Comparison of reachability of selected bridges
across our three experimental sites.
We detected no blocking from our Iran site.
In China, blocking manifested as timed-out connections.
Blocking in our two China sites was the same,
except for the curious diurnal semi-blocking pattern,
which affected different bridges at each site.
}
\label{fig:comparison}
\end{figure*}
Though Tor Browser supports several pluggable transports,
we concern ourselves only with obfs4~\cite{obfs4},
an advanced transport offering several security features.
We rely critically on these features;
they enable us to limit the means of learning
about new bridges to the ones we control.
The use of obfs4 gives us confidence that the censor
learns of our bridges only in the ways we intend,
namely their inclusion in Tor Browser.
obfs4 resists deep packet inspection by re-encrypting
a Tor stream so that it appears as a stream of random bytes.
More than that, obfs4 resists \emph{active probing} attacks
in which the censor scans suspected proxies
in order to discover what protocols they support.
Every obfs4 server has a per-bridge secret,
which the client must prove knowledge of in its initial message.
The Great Firewall is known to employ active probing against the predecessor protocols
obfs2 and obfs3~\cite{Ensafi2015b},
but the same attack is ineffective against obfs4.
The censor must have the same out-of-band
information as a legitimate client;
merely knowing the IP address of a bridge is insufficient
to confirm that it is, in fact, a bridge.
In addition to its useful properties, obfs4 is relevant to real-world users.
It is marked ``recommended'' in Tor Browser
and has the more users than any other transport,
about 20,000 concurrent on average as of June~2016~\cite{obfs4-users}.
obfs4 has experimental support for obfuscating
packet size and timing,
but none of the default bridges have deployed that feature,
so we did not investigate it.
\section{Method}
We established probe sites in three countries,
using dedicated servers or cloud services:
one in the U.S., two in China, and one in Iran.
The two sites in China were in different autonomous systems.
At each site, we ran a script that tested the TCP reachability
of a variety of destinations every 20 minutes.
For each destination, the script attempted to establish a TCP connection
and then recorded the current time, IP address, port number, status,
and error message if any.
We ran probes for about five months,
from December~12, 2015 to June~4, 2016.
The destinations we tested,
a mixture of fresh obfs4 bridges and other control destinations,
appear in Table~\ref{tab:bridges}.
The probe site in the U.S. was a control
that enabled us to distinguish cases of censorship
from a bridge's temporarily being down.
Owing to the difficulty of accessing network services in China and Iran,
the time periods during which we had access to each site
differ, though they overlap.
In one of the China sites, we lost access to our probing host
partway through the experiment; however, before that happened
we got access to another in the same autonomous system.
The data in Figure~\ref{fig:timelines} and Table~\ref{tab:timeline},
which show the results from this site,
are spliced together from the two series.
\begin{table}
\begin{tabular}{l l}
nickname & ports \\
\hline
\noalign{\smallskip\normalsize\textbf{New Tor Browser default bridges}}
ndnop3 & 24215, 10527 \\
ndnop5 & 13764 \\
riemann & 443 \\
noether & 443 \\
Mosaddegh & 41835, 443, 80 \\
MaBishomarim & 49868, 443, 80 \\
JonbesheSabz & 80 \\
Azadi & 443 \\
GreenBelt & 60873, 80, 443 \\
\noalign{\smallskip\normalsize\textbf{Orbot-only default bridges}}
Mosaddegh & 1984 \\
MaBishomarim & 1984 \\
JonbesheSabz & 1984 \\
Azadi & 1984 \\
\noalign{\smallskip\normalsize\textbf{Already existing Tor Browser default bridges}}
LeifEricson & 41213 \\
\noalign{\smallskip\normalsize\textbf{Never-published bridges}}
ndnop4 & 27668 \\
\noalign{\smallskip\normalsize\textbf{Non-bridge controls}}
github.com & 443, 444 \\
\end{tabular}
\caption{
The destinations, mostly obfs4 bridges, whose reachability we tested.
Bridges are identified by their ``nickname,'' an arbitrary label assigned by the bridge's operator.
Each nickname represents a distinct IP address;
some bridges served obfs4 on multiple ports.
In addition to the destinations listed,
we tested
port 22 (SSH) on the bridges that had it open.
}
\label{tab:bridges}
\end{table}
We were fortunate to run this experiment at a time
when Tor Browser was ramping up its obfs4 capacity,
with new bridges being added in nearly every release.
In some cases these were new ports on existing IP addresses;
in others they were entirely new IP addresses.
We began measurements of each new bridge as soon as we became aware of it.
In some cases we received advance notice of a new bridge
before its ticket was filed,
but in others we started measurements shortly after the ticket.
We coordinated with the Tor Browser developers
to ensure that newly created bridges were not present
in BridgeDB~\cite{bridgedb}, where they might have been
discovered by censors and ordinary users.
We tried, as far as possible, to limit the possible
avenues of discovery to bug tracker tickets and the Tor Browser source code.
We also measured four bridges that appeared only in Orbot~\cite{orbot},
the port of Tor to Android.
Orbot and Tor Browser have most of their default bridges in common,
but a few appear in Orbot only.
They are the ones with port number 1984 in Table~\ref{tab:bridges}.
The Orbot-only bridges remained accessible,
even as the Tor Browser bridges were blocked.
\section{Results}
We recorded over 1.5~million individual probe results
over a period of approximately five months.
The results from our two China sites were the same in most respects,
with blocking occurring at the same time in both.
We found no blocking at all of the default bridges from our site in Iran.
As the two China sites were similar and the Iran site did not show any blocking,
we will mainly present the results from the China site that had more data
(called ``China~1'' in figures).
A graphical summary of the results appears in
Figure~\ref{fig:timelines}
and a textual timeline in Table~\ref{tab:timeline}.
Figure~\ref{fig:comparison} compares
the three sites across a subset of destinations.
\newcommand{\eventheading}[1]{\textbf{#1}\quad}
\newcommand{\eventheading{Ticket filed}}{\eventheading{Ticket filed}}
\newcommand{\eventheading{Ticket merged}}{\eventheading{Ticket merged}}
\newcommand{\eventheading{Testing release}}{\eventheading{Testing release}}
\newcommand{\eventheading{Public release}}{\eventheading{Public release}}
\newcommand{\eventheading{Blocked}}{\eventheading{Blocked}}
\newcommand{\releaseheading}[2]{\noalign{\smallskip\normalsize\textbf{#1:} #2}}
\newcommand{\releaseheadingnull}[1]{\noalign{\smallskip\normalsize\textcolor{gray}{\textbf{#1:} no new bridges}}}
\begin{table}
\footnotesize
\setlength{\tabcolsep}{2pt}
\begin{tabular}{l r >{\raggedright\arraybackslash}p{2.2in}}
\releaseheading{Tor Browser 5.0.5/5.0.6/5.5a5}{1 new bridge}
03 Dec & $-$19 days & \eventheading{Ticket filed} ndnop3:24215 \\
07 Dec & $-$15 days & \eventheading{Ticket merged} ndnop3:24215 \\
12 Dec & $-$10 days & \eventheading{Testing release} 5.0.5 stable \\
14 Dec & $-$8 days & \eventheading{Testing release} 5.5a5 alpha \\
15 Dec & $-$7 days & \eventheading{Public release} 5.0.5 stable \\
17 Dec & $-$5 days & \eventheading{Public release} 5.0.6 stable \\
18 Dec & $-$4 days & \eventheading{Public release} 5.5a5 alpha \\
22 Dec & 0\phantom{ days} & \eventheading{Blocked} ndnop3:24215 \\
\releaseheadingnull{Tor Browser 5.0.7/5.5a6}
\releaseheading{Tor Browser 5.5/6.0a1}{6 new bridges}
16 Jan & $-$13 days & \eventheading{Ticket filed} riemann:443 \\
18 Jan & $-$11 days & \eventheading{Ticket merged} riemann:443 \\
18 Jan & $-$11 days & \eventheading{Ticket filed} ndnop3:10527, ndnop5:13764 \\
18 Jan & $-$11 days & \eventheading{Ticket merged} ndnop3:10527, ndnop5:13764 \\
19 Jan & $-$10 days & \eventheading{Ticket filed} noether:443 \\
19 Jan & $-$10 days & \eventheading{Ticket merged} noether:443 \\
21 Jan & $-$8 days & \eventheading{Ticket filed} Mosaddegh:41835, MaBishomarim:49868 \\
21 Jan & $-$8 days & \eventheading{Ticket merged} Mosaddegh:41835, MaBishomarim:49868 \ \\
22 Jan & $-$7 days & \eventheading{Testing release} 5.5 stable \\
24 Jan & $-$5 days & \eventheading{Testing release} 6.0a1 alpha \\
27 Jan & $-$2 days & \eventheading{Public release} 5.5 stable \\
27 Jan & $-$2 days & \eventheading{Public release} 6.0a1 alpha \\
29 Jan & 0\phantom{ days} & \eventheading{Blocked} ndnop3:10527, riemann:443, ndnop5:13764, noether:443, Mosaddegh:41835, MaBishomarim:49868 \\
\releaseheading{Tor Browser 5.5.1/5.5.2/6.0a2}{2 new bridges}
24 Jan & $-$30 days & \eventheading{Ticket filed} JonbesheSabz:80, Azadi:443 \\
28 Jan & $-$26 days & \eventheading{Ticket merged} JonbesheSabz:80, Azadi:443 \\
03 Feb & $-$20 days & \eventheading{Testing release} 5.5.1 stable \\
05 Feb & $-$18 days & \eventheading{Public release} 5.5.1 stable \\
12 Feb & $-$11 days & \eventheading{Public release} 5.5.2 stable \\
15 Feb & $-$8 days & \eventheading{Public release} 6.0a2 alpha \\
23 Feb & 0\phantom{ days} & \eventheading{Blocked} JonbesheSabz:80, Azadi:443 \\
\releaseheadingnull{Tor Browser 5.5.3/6.0a3}
\releaseheading{Tor Browser 5.5.4/6.0a4}{4 new bridges}
12 Mar & $-$17 days & \eventheading{Ticket filed} Mosaddegh:80, Mosaddegh:443, MaBishomarim:80, MaBishomarim:443 \\
14 Mar & $-$15 days & \eventheading{Ticket merged} Mosaddegh:80, Mosaddegh:443, MaBishomarim:80, MaBishomarim:443 \\
18 Mar & $-$11 days & \eventheading{Public release} 5.5.4 stable \\
18 Mar & $-11$ days & \eventheading{Public release} 6.0a4 alpha \\
29 Mar & 0\phantom{ days} & \eventheading{Blocked} Mosaddegh:80, Mosaddegh:443, MaBishomarim:80, MaBishomarim:443 \\
\releaseheading{Tor Browser 5.5.5/6.0a5/6.0}{3 new bridges}
05 Apr & $-$57 days & \eventheading{Ticket filed} GreenBelt:60873, GreenBelt:80, GreenBelt:443 \\
07 Apr & $-$55 days & \eventheading{Ticket merged} GreenBelt:60873, GreenBelt:80, GreenBelt:443 \\
22 Apr & $-$40 days & \eventheading{Testing release} 5.5.5 stable \\
24 Apr & $-$38 days & \eventheading{Testing release} 6.0a5 alpha \\
26 Apr & $-$36 days & \eventheading{Public release} 5.5.5 stable \\
28 Apr & $-$34 days & \eventheading{Public release} 6.0a5 alpha \\
26 May & $-$6 days & \eventheading{Testing release} 6.0 stable \\
30 May & $-$2 days & \eventheading{Public release} 6.0 stable \\
01 Jun & 0\phantom{ days} & \eventheading{Blocked} GreenBelt:60873, GreenBelt:80, GreenBelt:443 \\
\end{tabular}
\caption{
Timeline of Tor Browser releases
and bridge blocking.
For the meaning of terms, see Section~\ref{sec:bridge-process}.
Time offsets are given from the date of bridge blocking.
}
\label{tab:timeline}
\end{table}
We have batched Tor Browser releases according to what new, unblocked
bridges they contained.
For example, all the releases in the 5.0.5/5.0.6/5.5a5 batch contained
the same new bridge, ndnop3:24215,
while the releases in the 5.5/6.0a1 batch appeared after ndnop3:24215 was blocked
but added six fresh bridges.
There are five of these release batches
(we omitted releases that did not have any new bridges).
Blocking of new bridges was delayed, but abrupt.
When a batch contained more than one bridge,
all were blocked at once (within our probing period of 20 minutes).
Across the five batches, we observed blocking delays of
7, 2, 18, 11, and 36 days after the first public release,
and up to 57 days
after the filing of the first ticket,
when bridges were potentially first discoverable.
This fact suggests, to us, that new default bridges
are loaded into the firewall in groups,
and are not, for example, detected and blocked one at a time.
We found that blocking in China was keyed on both IP address and port,
consistent with an observation of Winter and Lindskog in 2012~\cite{Winter2012a}.
For example, many of the bridges happened to have port 22 (SSH) open,
and it remained accessible even as other ports on the same IP address were blocked.
(See riemann:22 in Figure~\ref{fig:timelines} for an example.)
Per-port blocking is what enabled us to run multiple bridges
on the same IP address.
We never saw a case of a bridge being blocked
before a public release of Tor Browser,
despite their being potentially discoverable
at an earlier stage.
The four bridges that were included only in Orbot,
not in Tor Browser, were never blocked.
From these facts, we infer
that the censors in China probably learn of bridges
not from the bug tracker (which would have revealed Orbot's bridges),
nor from source code inspection (which might have gotten the bridges blocked before release),
but only from public releases.
The 5.5.1/5.5.2/6.0a2 batch is an interesting case because
there was an unusually large gap (about 10 days) between stable and alpha releases:
there were stable releases 18 and 11 days,
and an alpha release 8 days, before
JonbesheSabz:80 and Azadi:443 were blocked.
The bridge ndnop4:27668 did not appear in a release,
but only in BridgeDB, and was not blocked.
\begin{figure}
\includegraphics{data/diurnal}
\caption{
Rates of reachability by time of day
for two bridges from two sites,
between February~1 and March~15, 2016.
There is a diurnal blocking pattern
in both China sites,
though not the same bridges are affected at both sites.
China Standard Time (CST) is UTC+08:00.
}
\label{fig:diurnal}
\end{figure}
There is a conspicuous on--off pattern
in the reachability of certain bridges from China,
for example ndnop3:24215 in China~1
and Mosaddegh:41835 in China~2.
The pattern is roughly periodic with a period of 24~hours.
Figure~\ref{fig:diurnal} averages many 24-hours periods
to show the reachability against time of day of two bridges.
The presence of the diurnal pattern appears to depend on both the bridge and the probing site,
perhaps depending on the network path,
as the same bridges do not show the pattern at both sites.
The pattern can come and go, for example in riemann:443 in China~1.
The China sites also display what are apparently
temporary failures of censorship, stretches of a few hours
during which otherwise blocked bridges were reachable.
Intriguingly, one of these corresponds to a known
failure of the Great Firewall that was documented in the press~\cite{scmp-gfw}.
On March~27, Google services---usually blocked in China---were reachable from
about 15:30 to 17:15 UTC.
This time period is a subset of one in which our bridges
were reachable,
which went from about 10:00 to 18:00 UTC on that day.
\begin{table}
\setlength{\tabcolsep}{3pt}
\begin{tabular}{l l l l l}
\textbf{Batch} & \textbf{Date} & \textbf{Day} & \textbf{Time} \\
\hline
5.0.5/5.0.6/5.5a5 & 22~Dec & Tue & 09:00 {\scriptsize UTC} / 17:00 {\scriptsize CST} \\
5.5/6.0a1 & 29~Jan & Fri & 06:00 {\scriptsize UTC} / 14:00 {\scriptsize CST} \\
5.5.1/5.5.2/6.0a2 & 23~Feb & Tue & 02:40 {\scriptsize UTC} / 10:40 {\scriptsize CST} \\
5.5.4/6.0a4 & 29~Mar & Tue & 06:00 {\scriptsize UTC} / 14:00 {\scriptsize CST} \\
5.5.5/6.0a5/6.0 & 01~Jun & Wed & 02:40 {\scriptsize UTC} / 10:40 {\scriptsize CST} \\
\end{tabular}
\caption{
Day and time of blocking events in China.
Times signify the end of an initial period
of continual or near-continual reachability
of all bridges in a batch of releases.
All bridges in a batch ceased their initial reachability
within 20~minutes after the time shown.
}
\label{tab:blocking}
\end{table}
We wondered whether blocking always occurs on the same
day of the week.
It turns out not to be the case,
as our five blocking events happened on
Tuesday~($\times 3$),
Wednesday~($\times 1$),
and Friday~($\times 1$)---see Table~\ref{tab:blocking}.
However, there may be a pattern in the time of day.
In two cases, the last successful probe happened within the 20~minutes following
02:40~UTC,
and in another two cases,
it was just after 06:00~UTC,
The remaining event happened at 09:00~UTC.
It is hard to make inferences from these limited data,
but they, along with the variable delay in blocking,
suggest a blocking procedure that is part manual and part automatic:
a manual process discovers bridges after an unpredictable delay;
then a periodic, automatic process causes the blocks to take effect.
\section{Further questions}
The data present interesting questions
that call for additional experiments.
Orbot's bridges were blocked only as a side effect
of their being blocked in Tor Browser---bridges
exclusive to Orbot remained reachable.
This suggests a further test
to see whether the censor treats stable and alpha releases differently:
include different sets of bridges in each,
and see whether both or only one of the sets gets blocked.
(This idea has the privacy disadvantage that a network eavesdropper
could infer whether someone is running a stable or an alpha release
by watching the IP addresses they connect to.)
Once in possession of a software release,
how does the censor extract the bridge addresses?
There are several possibilities.
They may have a program automatically parse the file containing bridges;
or a person may have to read the file and enter the bridge addresses manually.
They could simply run the browser in a black-box fashion
and note what addresses it connects to.
In order to distinguish these cases,
one could include commented-out or invalid
bridge addresses in the list.
Table~\ref{tab:blocking} shows that our bridges
got blocked only on weekdays during daylight hours
(China Standard Time).
The five blocking events occurred at three
apparently discrete times of day.
It will be interesting to observe more blocking events
and see whether these patterns continue to hold.
We were surprised not to find any blocking of the default bridges in Iran.
The censorship system in Iran has been documented
to effect blocking through means such as bandwidth throttling~\cite{Anderson2013a,Aryan2013a}
and blocking the IP addresses of the Tor directory authorities~\cite{torbug-12727}.
Throttling would affect obfs4 users but
blocking the directory authorities would not,
as bridges serve Tor directory information in-band.
Iran's censors do not seem to rely on blocking the default bridges,
at least those we tested.
It would be useful to augment our reachability tests with
traceroutes, so we could see where all the routers are between the probe site
and the bridge, and where packets are dropped when bridges are blocked.
Differences in routing might explain why some bridges
were periodically reachable from one China site but not from the other.
The fact that proxy blocking delay is on the order of days
suggests an obvious circumvention strategy:
if the censor blocks new bridges after $n$ days,
introduce new ones every $n-1$ days.
This could be accomplished through more frequent releases,
or obfs4 could be modified to try different destination ports
according to some schedule that depends on the time since release.
The interesting question then becomes the meta-delay:
how long does the censor take to catch on to this new pattern?
\section{Acknowledgments}
We express our thanks to those who helped us by running bridges
or providing us with probe sites, including
Percy Alpha,
Nima Fatemi,
Linus Nordberg,
Henry de~Valence,
and others who remain anonymous.
We thank the developers of Tor Browser and Orbot
for their timeliness in adding new bridges.
We appreciate our discussions with censorship researchers at
the University of California, Berkeley and
the International Computer Science Institute.
\section{Availability}
Our code and data are available from the supporting web page:
\url{https://www.bamsoftware.com/proxy-probe/}.
{\footnotesize \bibliographystyle{acm}
|
\section{Introduction}
For students to be successful quantum mechanics problem solvers, it is insufficient to think about only the features of the physical system. They also need to coordinate different representations by thinking conceptually about the mathematical representations that satisfy the physical system, evaluate the algorithmic steps, and reflect upon their work. Unsurprisingly, students often have trouble unifying these ideas during problem solving.
Researchers in student understanding of quantum mechanics have used ``difficulties" theory to understand student reasoning (e.g. \cite{Singh2015, Emigh2015, Passante2015}), which forms long lists of difficulties that span many topics in quantum mechanics.
However, we posit that these disparate difficulties can be unified through the lens of epistemological framing\cite{Tannen1993}, and errors in transitions between frames\cite{Irving2013framing}. Epistemological frames reveal students'\cite{Redish2004, Tannen1993} ways of thinking and expectations. They govern which ideas students link together and utilize to solve problems. Careful observation of student behaviors, gaze, and discourse can provide clues for determining students' epistemological frames. Productive problem solving requires both an appropriate frame \cite{Scherr2009} and appropriate transitions between frames\cite{Nguyen2016StuFraming}.
Some students, despite having strong numerical tools or skills, still ``get stuck'' in certain problem solving situations\cite{Tuminaro2004a}. This happens particularly in upper-division courses such as quantum mechanics, where mathematics is critical to understanding the subject. Quantum mechanics is a great choice for this study, because students are trying to coordinate difficult, often counter-intuitive concepts and complicated, often novel mathematical formalism.
Bing et al\cite{Bing2012} identified four epistemological frames to aid in understanding the role of math as a reasoning tool as opposed to a numerical tool. They analyzed students' thinking while the students translate physical ideas into informative mathematical forms, or compare a mathematical structure in two similar physics or math scenarios. Though Bing et al identified ``Physical Mapping'', and ``Math Consistency'' frames, their ``Calculation'' frame is biased toward the use of formal math, independent of physical sense making. However, they did not further differentiate between trivial math calculations and conceptual math reasoning.
On the other hand, Kuo et al \cite{Kuo2013Blending} differentiated between the use of equations as an input-output calculator template, instead of attending to the conceptual meaning embedded in the equation to create shortcuts. They referred to ``cognitive elements'' \cite{Sherin2001} to capture students' understanding of equations while they blend their reasoning with symbolic forms and create a shortcut to interpret the situation. They concluded that successful problem solvers are able to make a decision as to which tools they bring into play for an efficient understanding of the problem situation.
Earlier studies identified other possible avenues that students may follow to obtain a correct solution by using conceptual physics and algorithmic math as numerical tools. These problem solving studies often focused on the differences between experts and novices\cite{Larkin1980, Heller1997, Chi1981a}. A novice adopts an inverse strategy by simply attending to the goal of the problem, recalling and manipulating an equation that contains the unknown quantities. In contrast, an expert moves forward based on having a representation for the situation, and then choosing the relevant principles\cite{Larkin1980}. However, this study is limited because experts' expertise far exceeds the difficulty of the end-of-chapter problems, and so such a study can not show the heuristics of expert-like problem solvers.
Heller et al \cite{Heller1997} designed context-rich problems to challenge introductory students beyond end-of-chapter exercises. This method requires students to make sense of the physical system, and justify what strategies to adopt as experts do. Their work assesses how students initially translate the problem statement into a visual representation in order to help them to adopt a proper strategy for determining the implicit unknown physical quantities. The strategy would allow students to translate their physical representation into a mathematical representation in order to do the algorithmic steps, and find the unknown quantity to make sense of their solution. While this problem solving strategy was initially quite prescriptive in the nature and order of the problem solving steps, later research has permitted a less-linear structure to problem solving.
Building on this work, Caballero et al \cite{Caballero2013ACER} worked to explain the common difficulties of upper-division students in problem solving, focusing on four steps in mathematical tool use: activation, construction, execution, and reflection. These steps could be completed in any order, and solutions may vary among students and problems. One student could stay mostly in the execution phase to process the algorithmic steps. Another student might evaluate the solution by staying in the construction phase and skipping the execution elements in favor of conceptual steps, or one could bring into play both components of execution and construction. Students' use of certain steps in this theory does not necessarily imply difficulties with the missing component of their problem solving process. This could become important when the problem statement of the question nudges students toward the use of one of these four components more than the others.
Broadly speaking, these three research traditions -- research into student difficulties in quantum mechanics, research into epistemological framing, and research into student problem solving -- suggest several approaches for understanding how students understand quantum mechanics problems. One approach may consider the initial physical understanding of the problem as more critical, with less emphasis on the mathematical manipulations, whereas other approaches may consider equally a close relationship between math and physics, or focus on the conceptual meaning of math in reasoning. Our present study integrates these three approaches to capture the various facets of students' epistemological framing during problem solving at the upper division. Our theoretical framework\footnote{It is a particular linguistic difficulty of research on students' epistemological frames that the theory used to describe them is a ``framework''. In this paper we follow the convention that a ``framework'' is something that researchers use, while a frame is something that humans (in our case, students) use. A framework is an appropriate technical term for a set of connected theoretical statements (e.g. "Resources Framework") Students -- humans -- frame ideas, have epistemological frames, and participate in framing activities. There are some subtle differences between these three forms of "frame", but all of them are related to the idea "how you know what's going on". }
takes up the idea of epistemic frames to explain student problem solving without prescribed steps. Our model suggests that difficulties are an interaction effect between question asked and students ideas, which implies there may be an underlying structure to identified difficulties in quantum mechanics.
In this paper we develop a theoretical framework which models students' framing in math and physics, expanded through the algorithmic and conceptual space of students' problem solving. We investigate four frames: algorithmic math, conceptual math, algorithmic physics, and conceptual physics, looking for moments where students' problem solving is impeded because they are in an unproductive frame. We applied this theoretical framework to observational data from quantum mechanics classes in which students solve typical quantum problems in pairs and small groups. Our purpose in this paper is to illustrate our theory, not to exhaustively show the prevalence of specific frames or to catalog the methods by which students may transition between them.
\section{Context}
We video recorded the class meetings of one semester of a senior-level quantum mechanics class. The class is taught using Griffith's Introduction to Quantum Mechanics\cite{Griffiths2005QM} using a wavefunctions-first topic order. It meets for four 50-minute sessions each week. During class, lecture is interspersed with small group problem solving. Groups of 2-3 students solve problems collaboratively on shared table-based whiteboards. Most problem-solving sessions last 2-5 minutes, though they can be as long as 15 minutes for more difficult problems. Students are remarkably collaborative, usually working together for the entire duration of each problem-solving session. In our data set, we see about one problem-solving session per class, though this decreases in frequency near the end of the semester.
Generally, these problem-solving sessions begin when the professor halts the lecture to ask the students to attempt to solve a problem related to their current topic, or to introduce a new topic. Occasionally, they also arise when students initiate a class discussion and the professor decides to assign a problem to gauge their understanding.
The groups in this class are somewhat fluid, and students may form different groups on different days. Students occasionally recruit others from nearby groups to help them solve problems. The instructor does not explicitly tell students where to sit or with whom to work (other than ``people near you''). Generally speaking, students work in pairs or threes; occasionally fours.
\section{Methodology for video data}
In learning environments such as group problem solving in upper-division contexts, one way to interpret the high level of interactions within group members is to carefully analyze the discourse and gestures of each member of the group. Ethnography provides an opportunity to understand the detail of students' discourse, behaviors, as well as capture useful information while they are investigating a phenomena\cite{Pirie1997}. One of the methods for data collection in ethnography studies in through video recording of activities. This becomes more important by providing multiple researchers an opportunity to view and analyze the videos\cite{Derry2010}. Previous researchers in education have used ethnography to study the culture of classroom activities \cite{Brown2004} or in more engaging learning environments, such as advanced physics laboratory \cite{Irving2014AdLab}. Our goal was to develop a theoretical lens to enable us to explain problem solving within various topics in quantum mechanics.
We divide class into episodes of problem solving and episodes of lecture, discarding episodes of lecture because they don't help us understand student reasoning. The problem solving episodes have distinct boundaries: they start with the professor explicitly asking students to begin working on their table-based whiteboards and end when the professor either asks for answers or begins explaining the answer.
In our preliminary analysis of the students' group problem-solving activities, we observed that some aspects of the data represent a conceptual approach and other aspects represent an algorithmic approach. We also noticed students' use of conceptual physics and algorithmic math. This distinction is consistent with the ACER\cite{Caballero2013ACER} and framing\cite{Bing2012} literature on problem solving from upper-division physics classes, showing how students' understanding of physical systems maps to algorithmic representations. However, neither theoretical framework adequately captured the richness of our data, prompting us to take further steps to interpret our data set. From the tradition of progressive refinement of hypotheses\cite{Engle2007}, we set out to refine our observations through close interrogation of the video data.
We started with selecting episodes for close analysis based on their duration (longer is better), conceptual richness (more complex is better), and technical quality (more visible and audible are better). We reflected on these episodes, seeking to answer ``what's going on?'' for each of them. Through repeated watching and examining the details of the selected episodes, we sought to capture changes in students' discussion or behavior that might indicate a shift in the students' problem solving processes. We began to focus on instances where students ``got stuck'' in their problem solving processes. This momentary impasse prompted them to try a different kind of reasoning until suddenly they were able to get ``un-stuck''. We examined the interactions immediately preceding and following the unsticking moments to look for regularities in unsticking behavior.
We developed a preliminary theoretical framework\cite{Thompson2016ICLS} which mapped student behavior onto three discrete frames: conceptual physics, conceptual math, and algorithmic math. The two math frames -- concordant with research in mathematics education on concepts and processes\cite{Sfard1991,Dubinsky2001} -- suggested that we expand our ideas to look for the ``missing'' physics frame: algorithmic physics.
Concurrently, we grew troubled with the idea of discrete frames. Sometimes, students seemed exceptionally ``mathy'', operating without regard to any sense of physical meaning. It is possible, however, to blend conceptual ideas from both math and physics domains, or to move fluidly and rapidly between conceptual and algorithmic thinking. We reframed our ideas into two orthogonal axes: conceptual to algorithmic and math to physics, defining a coordinate plane in which students' problem solving roams. In pursuit of evidence to refine this two coordinate-axis framework, we delved again into our observational data, seeking examples of all four quadrants and transitions among them.
After several more iterative cycles of analysis and refinement, we reached a stable point where new episodes did not change the theoretical framework or our application of it. Operating with the newly-stable framework, two independent raters came to consensus on every episode; two additional raters checked a selection of episodes with agreement of $>90\%$. We selected episodes for analysis based on frequency of students' discussion regarding concepts and processes, as well as displays and features of potential frame transitions. We categorized episodes with conceptually rich discussions and frame negotiations as strong examples, and established inter-rater reliability about the content of the episodes and regarding which episodes strongly or weakly evidenced frame transitions.
We also identified very weak examples when it was hard to find evidence of students' framing from the group discussion. This could happen due to noisy or garbled audio, or when students were writing on part of the whiteboard that was not in the view of the camera, or in general the raters did not have enough information to determine students' framing.
Once we identified students' frames, we looked for transitions in those frames to help us to interpret the dynamic of students' problem solving behaviors or identify the impasse students reach when they fail to notice certain factors that could have triggered a transition to a more appropriate state.
We acknowledge the existence of other frames that could describe students' behavior while they are engaged with other kinds of activities in a classroom e.g. turning in home work to the instructor, taking break within solving several parts of a long problem, or discussing upcoming social events\cite{Nguyen2016Dynamics}. While these other frames can be important for problem solving more broadly\cite{Irving2013framing}, in this study our focus is on investigating students' topical discussions during problem solving sessions which last about 2-5 minutes.
Epistemological frames are context dependent\cite{Redish2004}. For example, by walking into a restaurant relevant resources consistent with behaving in the situation are activated to read a menu, order food, pay the tip, etc. However, in the setting of the restaurant we don't access our resources for behaving in a library.
Students' perceptions of the problem context affect their framing of what subset of their knowledge to activate. In an interactive class environments such as group problem solving the instructor's framing can also affect the students framing of the situation\cite{Scherr2009, Irving2013framing, Nguyen2016Dynamics}. Even within group problem solving, other students' framing can affect an individual's framing as well. The instructor can nudge students to frame the problem more conceptually by asking about the physics of the situation, or more algorithmically by asking about formulae\cite{Nguyen2016Dynamics}.
\section{Theoretical framework}
Our theoretical framework consists of two axes: an algorithmic versus conceptual axis, and a math versus physics axis (Figure \ref{fig:framework}).
\begin{figure}
\includegraphics[width=\figwide]{framework.png}
\caption{Math-physics-algorithmic-conceptual theoretical framework. The horizontal axis indicates algorithmic and conceptual directions. The vertical axis represents the math versus physics directions. Each quadrant is labeled.\label{fig:framework}}
\end{figure}
The two axes divide different aspects of students' problem solving into four regions: algorithmic math, conceptual math, algorithmic physics, and conceptual physics. It is important to note that none of these frames are inherently ``good'', ``bad'', or even universally useful. At different times, different frames may be productively used to solve problems in physics, and often more complicated problems require multiple transitions between frames.
\subsection{Algorithmic and conceptual math}
In algorithmic frames, students are focused on following a known series of steps to solve a problem. Since the result of each step is used to process the following step, students are focused on their task to prevent errors. They tend to be in a writing mode, and have less discussion as compared to conceptual frames. In their discussions, they tend to focus on error-checking (``what did you get for part c?'' or minutiae of their steps (``you are missing a minus sign'').
When the problem statement requires explicit algorithmic calculations to find an answer, students enter an algorithmic math frame to spend a considerable amount of time setting up a series of algebra-based steps to evaluate integrals, take derivatives and simplify their solutions by dividing or multiplying a term on both sides of their solution.
Algorithmic math can be a quick and powerful problem-solving mindset as they may take several fast steps over a long period of time. However, without other quadrants it is quite difficult to check whether or not the solution makes sense. This frame leaves students with a narrow\cite{Irving2013framing, Engle2012Expansive} discussion mostly to check the signs, or to alert each other of the missing symbols, while they are focused in their numerical calculations. For these features, we consider the ``just math'' frame\cite{Wolf2014JustMath} is an example of students' prolonged use of the algorithmic math frame. Narrow framing which focuses only on the task at hand\cite{Engle2012Expansive, Irving2013framing} is not exclusive to algorithmic frames, but it is common within them.
In contrast, students in a conceptual math frame use a conceptual approach to understand the mathematics. They reason based on general properties of a class of information in math. This could help them to apply practical ideas about the behavior of the mathematical functions, and determine the result of an operation without actually computing it. For example, by knowing that sine functions are independent of each other, and discussing the orthogonality properties of the sine functions, one can shortcut the integration of product of two sine functions of different periods, preventing the use of many trigonometric identities, and simply ``see'' that the integral equals zero.
Creating a ``shortcut'' solution\cite{Kuo2013Blending} to the problem reduces the procedures and lessens the writing. Concurrently, discussing mathematical problems conceptually gives students more opportunities for sense-making discussions with other members of the group\cite{Scherr2009, Irving2013framing}. This kind of thinking is generally more expansive\cite{Engle2012}, as students connect general cases of mathematics to the specifics in this problem or bring in connections to other problems.
Attention to conceptual mathematics is a large part of numeracy, and as such is an important part of learning mathematics\cite{Dubinsky2001} and physics\cite{Kuo2013Blending}, especially at the upper-division\cite{Sayre2008,Kustusch2013}.
\subsection{Algorithmic and conceptual physics}
Just as we find algorithmic and conceptual frames in math, we find them in physics as well. Students in a conceptual physics frame try to think in terms of the features of the physical system and might coordinate between different representations such as graphical, geometric or gestural to visualize the physical system. They coordinate different physical laws and concepts to explain the situation. We provide examples from the context of Electromagnetic fields course as motivating examples to show the broader phenomena. In the next section, we will provide several quantum mechanics examples from our own data.
For example, students might argue that the total charge on a spherical shell whose surface charge density $\sigma$ is proportional to $\sin(2\theta)$ (where $\theta$ is the azimuthal angle) is equal to zero because the northern hemisphere is positive while the southern is negative, and those two halves must be equal and opposite. In this case, students use conceptual reasoning to map charges to a sphere, employing balancing resources to come to the conclusion that the net charge equals zero. One could move to algorithmic math frame to write the integral of the charge density over the surface area:
\begin{align}
\int{\sigma}{dA}{\propto}{\int{\sin(2\theta)r^2\sin(\theta){d\theta}{d\phi}}}
\end{align}
It's possible, of course, to solve this problem algorithmically (using trigonometric substitutions) or conceptually without reference to physical systems (via the orthogonality of sine functions). In either the conceptual math or conceptual physics frames, discussing the problem plan in the conceptual physics frame can make later algorithmic calculations easier.
Thinking conceptually about the underlying physics of the situation encourages students to create connections to real word situations and other classes of problems as well. For example, to estimate the far distant electric field of a uniformly charged disk, one method is to expand the solution by mostly engaging in algebra to get the answer. Or one can visualize that far from the disk a continuous charge looks similar to a point charge, and by knowing the electric field of a point charge, the leading terms in the solution can be guessed. In each case, students are engaged in an activity to find an answer, but the nature of the activities are different. The latter case is more expansive, as students are open to make connection between the current situation and another class of problem.
In contrast, students in the algorithmic physics frame tend to recall equations, facts, and properties of physical quantities without conceptual justifications. They use math as a tool to adjust equations via a series of algebra-based steps to relate physics quantities to each other, or to check the correctness of the physical quantities in the problem. For example, by doing dimensional analysis students can check the correctness of their answer. Just as with algorithmic math, students tend to frame their work narrowly in algorithmic physics and focus on following procedures to find answers. Someone who applies normalization conditions for wavefunctions by rote, for example, is operating algorithmically.
It's important to note that framing problems algorithmically can be fast. An expert doesn't need to engage in extensive conceptual thinking about the steps of a trivial problem; she can just solve it.
\subsection{Continua vs. categories}
A careful reader might be concerned because we started this section by claiming that there are two axes, implying a continuous distribution of possible framings, yet continued by identifying four frames which appear to be discrete. We chose the axes for theory-driven reasons: it's possible that students' framing exists on a continuum between very mathy and very physicsy, or very algorithmic to very conceptual, and discrete frames cannot capture this sense. We kept it for practical reasons: on occasion, students appear to move fluidly and rapidly among frames, and there's not enough evidence to assign them a single, quasi-stable frame before they move to the next. We're interested in quasi-stable frames because we want to study how students transition between frames, and it is practically very difficult to find transitions between frames without first identifying (quasi-)stable frames. We use the words mathy and physicsy to denote students' framings which are more in the math direction or which are more in the physics direction respectively.
By using axes, we hope to capture a sense of directionality from more mathy to more physicsy and more conceptual to more algorithmic. We do not imply that these axes constitute a formal metric or scale.
While some prior work in student framing of problem solving in physics has used discrete frames (e.g. \cite{Scherr2009,Bing2012}), other work has used continua in the same way (e.g. \cite{Irving2013framing}, building on \citep{Engle2006}).
\begin{figure}
\includegraphics[width=\figwide]{Snapshot2.png}
\caption{Group problem solving in algorithmic math frame\label{fig:snap}}
\end{figure}
\begin{figure}
\includegraphics[width=\figwide]{diagram.png}
\caption{Diagram of students' solution in algorithmic math frame\label{fig:diag}}
\end{figure}
\section{Illustrative episodes}
In this section, we present four brief episodes which illustrate the four quadrants in our framework. Before the examples, we divert into a brief review of the quantum mechanics of free particles and a typographic note on how we present transcript.
\subsection{Physics of free particles}
Most of the examples in this section are chosen from the same physics context of the free particle system, so we review the physics of this system briefly for the reader. In quantum mechanics, the free particle is characterized by a zero potential energy, thus the Hamiltonian has just one term, the kinetic energy. This problem is a good candidate for understanding basic properties of the wave function and the Schr{\"o}dinger equation. Since the Hamiltonian is in form of $p^2/2m$, the eigenfunction solution can be considered as a plane wave, which can be expanded in terms of sinusoidal wave functions. However, for one particle the wave function with a determined momentum is not normalized over all the space. Thus a linear combination of all solutions is considered as a normalizable wave function.
This is physically interpreted as a traveling wave pocket.
Because the Hamiltonian has only the kinetic energy term, the time independent Schr{\"o}dinger (TISE) equation results in a homogeneous second-order differential equation. In order to write the eigenvalue equation in terms of a differential equation students might need to recall some relations from algorithmic physics. Solving and finding the eigenfunctions of the equation leaves room for either algorithmic math calculations, or conceptual math discussions.
\subsection{Typographic note}
Before we present data, here is a brief typographic note. In these interactions, students very frequently speak the names of mathematical symbols. We could have typeset their words as if they were equations or as if they were the names of isolated symbols. Equations are more compact -- importing algebra to Europe caused a scientific revolution -- but they lose some of the nuance of students' speech. Isolated symbol names, on the other hand, tend to be difficult to follow in text in a way that they are not difficult to follow in speech, especially as oftentimes students write as they speak. We have chosen a middle path, seeking to maximize clarity for the reader.
Additionally, we typeset a comma for brief pauses, a period for longer ones, and ellipses (\dots) for the longest ones. Stage directions are denoted by parentheses. Should we omit or alter some students' speech for clarity, changed words are denoted by square brackets and omitted ones by ellipses in square brackets ([\dots]).
\subsection{Episode: Algorithmic math\label{sec:AM}}
In this example a group of three students are solving the Schr{\"o}dinger equation to find the wave function of a free particle. They treat the space part and the time part separately. They start with the TISE for the space part. Guess a solution in the form of $e^{kx}$, and substitute it into the TISE to find the constant $k$.
Adam and Emma work together quickly to solve the problem, while their third groupmate (Eric) stays silent. (All student names are pseudonyms.)
Figure \ref{fig:snap} shows a snapshot of their whiteboard during group problem solving, taken while Emma is pointing to the both sides of their written equation to review the taken algorithmic steps in search for the missing sign. The numbers indicated on the figure show the order of the students' actions and narrations and reference the numbers in the transcript below. Figure \ref{fig:diag} shows a transcript of their writing on the whiteboard.
In this problem, Adam takes the derivative of the time solution of the Schr{\"o}dinger equation, and replaces it into the equation using the whiteboard in front of both students. He continues to replace the factors that the group has manipulated earlier in their solution and setting both sides of the equation equal to each other to verify if their solution satisfies the Schr{\"o}dinger equation.
\begin{description}
\item[Adam] Minus $E$ equal
\item[Emma] Minus $\hbar$ squares over two. (1)
\item[Adam] Minus $2mE$ over $\hbar$ \dots squared \dots Boom \dots Boom \dots Boom (while canceling the same quantities from two sides of the equation) (2)
\item[Emma] Cancel, cancel, cancel, and we are off by a negative (2)
\item[Adam] Yeah [unintelligible] sign
\item[Emma] With a negative up here, because these two are negative.
\item[Adam] Yeah that is true
\item[Emma] So something happened here (pointing to the two sides of their equation) (3)
\item[Adam] Or we lost a sign (3)
\end{description}
Emma and Adam use short sentences and talk quickly to be able to proceed to the next step of their algorithmic evaluation. They speak primarily of mathematical terms and operations, and do not talk explicitly or extensively about the physical quantities these symbols represent. At the end, they come up with an extra negative sign in one side of their solution. After reviewing their solution, Adam removes a negative sign in the earlier line of his solution, which he thinks is extra, but he does not further discuss the reason behind his decision.
At this point of their problem solving session, neither Adam nor Emma try further to make a transition to another frame to resolve thier error. In a low voice, Adam points to the power of the time phase factor exponential and says ``oh wait this [$e^{i\frac{E}{\hbar}t}$] gonna be a negative''. Emma says ``but I think we can start normalizing\dots'', then she starts to normalize their wave function.
In this brief moment of problem solving session the group is in an algorithmic math frame. Students are operating in algorithmic math frame because they are talking about manipulating mathematical symbols rapidly and alegebraically, running through a series of brief steps.
Prior to this episode, Adam makes a transition from algorithmic math to algorithmic physics. In so doing, he is able to resolve the cause of their group error. After, he returns to algorithmic math to continue the solution. (discussed in section \ref{sec:Adam})
The group does not spend further time to find the ``dropped negative" since students are toward the end of their problem solving session and are asked to normalize the wave function, which is discussed in the next episode.
\subsection{Episode: Conceptual math}
In this example, the same group is working on normalization of the free particle wave function by considering the general solution $\Psi$ as the sum of two functions of $Ae^{ikx}$ and $Be^{-ikx}$. They initially set the algorithmic steps, take the modulus square of the wave function ($\mid\Psi^*\Psi\mid$) and insert the limits of the integral.
\begin{widetext}
\begin{eqnarray}
\int_{-\infty}^{\infty} \mid\Psi^*\Psi\mid dx =& \int_{-\infty}^{\infty} \mid A\mid^2 + \mid B \mid^2 + AB^*e^{-2ikx} + A^*Be^{2ikx} dx \label{eqn:1}\\
=& x( \mid A\mid^2 + \mid B \mid^2) \Biggr|_{-\infty}^{\infty}
+ \int_{-\infty}^{\infty} AB^*e^{-2ikx} dx + \int_{-\infty}^{\infty} A^*Be^{2ikx}dx \\
=& AB^*\frac{1}{-2ikx}e^{-2ikx}\Biggr|_{-\infty}^{\infty}
+ A^*B\frac{1}{2ikx}e^{-2ikx}\Biggr|_{-\infty}^{\infty} \label{eqn:2}
\end{eqnarray}
\end{widetext}
There are a few errors in the students' solution which will not affect their conceptual discussion later in the episode. Emma leaves just one differential elements of length $(dx)$ for the whole expression right after the last term in the integrand (Equation \ref{eqn:1}). However, Emma is mindful of her incorrect notation, as she will take the integral of each term of the integrand with respect to $x$ in the following lines. The other error takes place after taking the integral of the $e^{2ikx}$: Emma leaves an extra negative sign in the power of the exponential, and none of the group members notice the extra sign for the second exponential integral (Equation \ref{eqn:2}). Eric notes that the result of the constant integrals are infinite; Emma will not further discuss and skip those terms in the last line of the solution.
They end up with the variable $x$ in both the denominator and the power of the exponential function in the numerator of the fraction (Equation \ref{eqn:2}). We acknowledge that the frame of students is mixed ahead of time while they are setting up their integrals algoritmically. But after this setup they switch to a purely conceptual reasoning and start their discussion again.
Without evaluating the integral numerically, they realize that the answer of the integrals might be infinite. They start in a conceptual conversation in the math context by arguing based on words and properties of the wave functions rather than working out equations to justify their answer as being finite.
\begin{description}
\item[Emma] We don't have to worry, this [the exponential term in the numerator] is gonna blow up faster than this [the denominator], right!
\item[Eric] They both blow up
\item[Emma] Yea, But one blows up faster and that matters
\item[Adam] Definitely the exponential (points to numerator)
\item[Emma] Yeah\dots
\end{description}
Emma has a discussion about which term ``blows up'' faster than the other. Emma gives more evidence of her conceptual understanding of the behavior of the two functions in the second exponential integral, when Eric says ``but both terms blow up''. She then compares the decay rate of the functions as an important factor that ``matters''. Emma conveys her generalized expectation\cite{Sayre2015BESM} of the situation by saying ``We don't have to worry''.
Although Emma only uses the term ``blow up'' briefly, there is a conceptual meaning embedded in this phrase which shows her understanding of the situation by comparing the rate. Adam seems to agree with Emma when pointing to the exponential function as blowing up faster.
\subsection{Episode: Conceptual physics \label{Ex:3}}
The Instructor asks the students to solve the Schr{\"o}dinger equation to find the wave function for the free particle. Robert begins to write down the time independent Schr{\"o}dinger equation ($\hat{H}\psi=E\psi$). At this point Alex states that the equation written by Robert is time independent. On the other hand, Robert seems certain that the wave function is time independent. Robert pauses writing, and both of the students start a conceptual discussion before continuing any algorithmic manipulations. Their arguments are based upon reasoning and discussion rather than working out specific equations:
\begin{description}
\item[Robert] I think\dots.
\item[Alex] That's time independent\dots
\item[Robert] Yeah\dots Why do we need time? hmm?
\item[Alex] Hmm\dots Because the wave function might have it.
\item[Robert] If there is no force then\dots um\dots why would anything about the wave function change over time?
\item[Alex] Because the wave function might depend on time\dots from its initial condition.
\item[Robert] I don't think it did. At least\dots [unintelligible].
\item[Alex] Oh, okay.
\end{description}
Robert refers to the properties of the physical system to reason that the wave function is time independent because there is no force acting on it, whereas Alex has doubts about how the initial condition can affect the evolution of the system over time. However, Alex does not have enough evidence to justify his reasoning.
In this episode, both students are in a conceptual physics frame, justifying their reasoning (albeit briefly) with arguments about physical quantities instead of mathematics or procedures, and bringing in more expansive reasoning. Alex thinks in terms of the feature of the problem by mentioning the initial condition. Thinking about the features of this problem also helps Robert to set the force equal to zero. This helps him to think more deeply about the underlying concepts and justify a zero change of the wave function over time.
\subsection{Episode: Algorithmic physics}
In contrast to the prior example, here Robert and Alex shift into a more algorithmic frame.
Robert continues to rearrange the equation of the Hamiltonian into kinetic and potential energy. He then sets the value of the potential energy equal to zero, and continue to recall the physical equations for the momentum based on the velocity and substitute them into the equations.
\begin{description}
\item[Robert] (writing math as he speaks) Whole definition is\dots $T+U $\dots Zero (crosses out U) is, uh $1/2$\dots $mv$ squared. This is $p$\dots $\frac{1}{2}pv$\dots and $v$ equals $\dot x$
So H is $1/2p\dot x$
\item[Alex] You can have that $H$ equals\dots or $T$ equals $\frac{p^2}{2m}$. It's skipping all of this.
\end{description}
Robert goes through multiple steps of algebra to remember the other physics equations in order to relate the kinetic energy to the momentum. However, Alex directly recalls the equation of the kinetic energy in terms of the momentum. Alex's framing is distinguished from algorithmic mathematics because he's not performing mathematical manipulations, merely recalling general physics formulae. Robert is also in a recall mode, as evidenced by the words which start his observation: ``whole definition is''. Both of them together implicitly agree that the goal of this part of the interaction is to lay out physical laws using mathematical formalism, not to discuss the applicability of those laws or derive them from first principles, as evidenced by Alex's comment that they can ``[skip] all of'' Robert's more elaborate efforts.
\section{Accounting for frame transitions}
The idea of the math-physics-algorithmic-conceptual framework is itself a development in how we model student thinking about math in physics contexts. However, only the briefest of problems (Heller \& Heller's ``exercises''\cite{Heller1992a}) require only one frame to solve them. To better model longer problems in upper-division physics, we must look at how students transition between frames in the course of problem solving.
Frame transitions -- or inability to transition -- in students' problem solving illuminates the connections among ideas and procedures in longer problems.
We identify transitions by first identifying preceding and following frames. The transition, definitionally, occurs between two different frames. Broadly speaking, we notice that the timing of transitions is relatively short (on the order of a few seconds, less than length of a few turns at talk).
\subsection{Example: Conceptual math to algorithmic math\label{sec:Eric}}
In the following example, the group transitions from a conceptual math frame to an algorithmic frame, which is a move from an expansive to a narrow frame.
In the previous class session, students discussed that the probability density of a stationary state is time independent. Immediately prior to this example, the instructor asks the class to work in groups and find if the probability density of a superposition of two stationary states ($\Psi_{1}$ and $\Psi_{2}$) is time dependent or independent.
Eric decides to talk through the solution to the problem with his group. His preference is to start in the conceptual math frame by comparing this problem to previous problems, and his initial conclusion is that the probability density is time independent.
\begin{description}
\item[Eric] I think\dots Cause when you do the, um, absolute value, you have to multiply by the complex conjugate, so I'm pretty sure that e thing [complex exponential part of the wave function] will just go to one, cause you'll replace that with\dots that e to the minus blah blah blah with e to the plus blah blah blah, and then when you multiply the\dots 1 over, you know, $x$ over $x$. That's what I'm thinking.
\end{description}
We believe Eric is in the conceptual math frame because he uses reasoning based on the behavior of the exponential function and complex conjugate to determine the ``form'' of the answer; namely, that the complex conjugate causes complex exponential terms to drop out when multiplied together. He isn't working on an algorithmic solution; he's arguing from the nature of these functions that his solution is reasonable. After he outlines the reasoning behind his conclusion, he begins working this problem out to check his answer.
After about two minutes the instructor mentions that the answer is time dependent, which confuses Eric, who proclaims his violated expectation loudly.
\begin{description}
\item[Eric] It is time dependent? Why? (While the instructor is explaining, he works on his paper) There's cross terms! Stupid\dots (smacks himself on forehead) ugh\dots That's why. Okay. Ugh, so stupid.
\end{description}
(From Eric's tone of voice, we interpret that Eric uses ``stupid'' to mean that his reasoning was thoughtless, not that he is personally stupid.)
The instructor's answer violates Eric's previous conclusion, prompting him to shift to another frame to explain the new answer. He realizes that his conceptual shortcut that exponentials will cancel with each other caused him to make a mistake. By viewing the problem algorithmically, Eric is able to review his work and determine what went wrong. He tries to find an answer for his question by transitioning to the algorithmic math frame and noticing that the ``cross terms'' are non-zero in this case.
\subsection{Example: Algorithmic physics to algorithmic math}
The next example illustrates a transition from algorithmic physics to algorithmic math. Both of these frames are narrow and actions in them are taken rapidly and frequently.
In this session the instructor asks the students to find the wave function of the free particle by solving the Schr{\"o}dinger equation.
Emma begins to write down time dependent Schr{\"o}dinger equation (TDSE). She tries to remember where to put $\hbar$ in the time dependent side of the equation and asks Eric if he remembers. Eric then recalls and writes the TDSE on the whiteboard ($\frac{-{\hbar}^2}{2m}\frac{\partial^2\Psi}{\partial x^2} =i\hbar\frac{\partial\Psi}{\partial t}$). After they both finish writing the equations, Emma compares them.
\begin{description}
\item[Emma] This is what I have, good we agree. (very quickly reviewing the facts) F= 0, V=0, \dots Separable.
\end{description}
Emma is in an algorithmic physics frame, recalling equations and matching them term-by-term in preparation for solving the problem using a known procedure: separation of variables.
However, Eric is in a different frame. He bids to begin their problem solving by thinking about the physical system:
\begin{description}
\item[Eric] So this is like the infinite [square well], except for we don't have boundaries.
\end{description}
Eric's comment compares the current problem to a well-understood system and provides opportunities for further thinking about their current system. Here Eric is using reasoning about a physical system by analogy to a previous problem, which is indicative of conceptual physics thinking. However, Emma does not take up Eric's bid to use the conceptual physics frame.
After Eric's comment, Emma asks Adam to join their group. Adam's involvement transitions the group into the algorithmic math frame, picking up Emma's earlier work on algorithmic physics to state known equations.
Adam begins by checking that the conditions for this problem satisfy the time independent Schr{\"o}dinger equation. He spends a short time in algorithmic math frame to justify that the energy in the spatial part of the Schr{\"o}dinger equation could be anything. This makes him ready to solve the bulk of the problem algorithmically and in a math frame via separation of variables.
Adam leads the group through his solution, which takes about five minutes. Adam has already taken a course on partial differential equations from the math department, and he feels very comfortable with this mathematical procedure. Adam's process begins in algorithmic math with finding the general form for $\Psi$ and the separation constant $K$. Adam uses the letter $\lambda$ as a known constant in the solution and explains it's relation to the separation constant $K$.
\begin{align}
\ \hat{H}\psi=E\psi \label{eqn:4}\
\end{align}
\begin{align}
\frac{-{\hbar}^2}{2m}\frac{\partial^2}{\partial x^2} \Psi = E\psi\label{eqn:5}\
\end{align}
\begin{align}
\ {k^2}=\frac{-2mE}{{\hbar}^2}
\end{align}
\begin{align}
\ \psi''=\frac{-2mE}{{\hbar}^2}\psi
\end{align}
\begin{align}
\ {k}=\frac{i}{\hbar}\sqrt{2mE}
\end{align}
\begin{align}
\ \psi=Ae^{{\lambda}x} + Be^{-{\lambda}x}
\end{align}
\begin{align}
\ \psi=Ae^{\frac{i}{\hbar}\sqrt{2mE}} + Be^{-{\frac{i}{\hbar}\sqrt{2mE}}}
\end{align}
\begin{description}
\item[Adam and Emma] $K$ squared equals minus E\dots
\item[Emma] (interjecting) and then we do the e to the $\lambda$ sign and then we find $\lambda$%
\item [Adam and Emma] K squared equals minus E\dots
\item[Emma] (interjecting) and then we do the e to the sign
and then we nd \dots So now we want to find out what $\lambda$ is
\item [Adam] It is just square root of $K$ (Adam points to equation 8). No wait it's $K$, if we define the square [of] $K$ by that\dots (Adam points to equation 6)
\item [Emma] It is k\dots yea
\item [Adam] So our space part is just $e$ to the $i$, square root of $2mE$ over $\hbar$\dots (Adam writes equation 10)
\item [Emma] Just write $K$\dots Just write $K$\dots
\item [Adam] Plus $B$ to the\dots
\item [Emma] Why would you not just write as $K$.
\item [Eric] Because this is the real name\dots I don't [unintelligible] (Adam finishes writing equation 10)
\end{description}
The students appear to stay in algorithmic math after Adam transitions them there: there is no discussion about how functions behave, what the physical system looks like, the effects on the final solution, or whether this makes sense. The students are purely focused on how to define $K$, the separation constant, and its relationship to $\lambda$, a known constant in the solution. This episode shows that algorithmic math can be a quick and powerful problem-solving mindset, but without other quadrants it is quite difficult to check if solutions makes sense.
\subsection{Example: Algorithmic math to algorithmic physics\label{sec:Adam}}
The group in the previous example continued their calculations to find out the solution of the wave function for the space part and time part. They find that the space part is equal to $Ae^{kx}$ + $Be^{-kx}$.
and the solution of the wave function for the time part as $e^{-i(\mu/\hbar){t}}$. The next step in their algorithmic calculations is to find how $\mu$ is related to $k$. They multiply both functions and substitute the ``whole thing'' into the TDSE. Earlier in their solution, they have found $k^2$ as $\frac{-2mE}{{\hbar}^2}$.
Part of this calculation happens while they are writing on the part of the board that is not visible in the camera.
As they are taking the derivatives with respect to time and space, they forget a sign, which they will not notice it until later in their problem solving session (Section \ref{sec:AM}).
\begin{description}
\item[Adam] So just $\mu$ equals ${\hbar}^2$ over $2m$ times $k^2$ [$\mu=(\frac{{\hbar}^2}{2m}){k^2}$].
\item[Eric] What's $\mu$?
\item[Emma] $\mu$ was our constant from when we were doing this part (pointing to the time derivative part of the TDSE).
\item[Eric] For time?
\item[Emma and Adam] Yea.
\item[Emma] Because we have $e^{-i(\mu/\hbar){t}}$.
\end{description}
While Emma is explaining to Eric where the coefficient $\mu$ comes from, Adam plugs in the value of $k^2$. However, he makes an error in the denominator, and only writes $\hbar$ instead of $\hbar^2$. Emma and Adam continue with algorithmic simplifications. This mistake causes their final $\mu$ to have an extra coefficient of $\hbar$.
\begin{description}
\item[Adam] So, boom\dots boom.
\item[Emma] cancel \dots cancel\dots cancel \dots cancel. We get \dots {$\hbar$}{E}
\item[Adam] Hah \dots (writes ${-E}{\hbar}$ and taps his finger on the board)
\item[Emma] Is that wrong?
\item[Adam] \dots Yeah.
\item[Emma] Because you have the $\hbar$ \dots
\item[Adam] \dots Joules (pointing to the $E$) \dots joules-second (pointing to the $\hbar$). Yeah I don't know if that's right.
\end{description}
Emma and Eric start algorithmic checking on the other side of the board, while Adam is silent after checking units by doing dimensional analysis.
At this point, we can hardly hear the conversation between Emma and Eric, since the instructor has paused the problem solving session and is giving feedback to the class. The group ignores the instructor's explanation and continue to work quietly on their own.
\begin{description}
\item[Emma] Oh wait \dots wait. Isn't that \dots ?
\item[Emma] No $\hbar$ is here (pointing to the $e^{-i(\mu/\hbar){t}}$ to explain something to Eric).
\item[Emma] The $\hbar$ is here (pointing to the $\hbar$ in the time solution), and that will cancel with this one (pointing to the extra $\hbar$ in the final answer)
\item[Emma] And then we will have \dots
\item[Adam] Oh wait a minute \dots hold on \dots hold on \dots
\item[Adam] Wait \dots I know what I did \dots I did not square this (pointing to the $\hbar$ in the denominator).
\end{description}
Adam adds the missing $\hbar$ power to his solution, and cancels it with the remaining $\hbar$ in the final answer.
\begin{description}
\item[Adam] So it's just minus E \dots Yes \dots (while raising his fist in a triumphant gesture).
\end{description}
In the previous episode we showed Adam's participation caused the whole group to shift from algorithmic physics to algorithmic math. In this episode only Adam shifts. He transitions from algorithmic math to algorithmic physics to find the source of his error by thinking in terms of units of the physical quantities. In contrast to Emma and Eric, Adam does not recheck his derivatives. Instead he checks units to make sense of his answer as a physics quantity and not just a symbolic answer. After finding the source of his error he continues the simplifications in the algorithmic math frame to find the final answer for $\mu$. This episode shows that by coordinating multiple frames students can better monitor their calculation process, saving time and/or making sense of their final answers.
\section{Discussion}
In this study we identified the state of students' thinking associated with four discrete frames including algorithmic math, conceptual math, conceptual physics and algorithmic physics. We presented several examples of students' group problems solving switching frames to productively and correctly solve a problem.
While upper-division students are generally facile at problem solving, on occasion they get stuck. By observing students' behaviors we noticed moments that students change the nature of their activities to make a decision that affects the future of their problem solving, to find the source of an error in their solution, or to get ``un-stuck''.
Epistemological framing is a window to individual's implicit state of thinking. This internal state can alter as a result of interaction with external artifacts in the environment such as the instructor's framing\cite{Irving2013framing, Nguyen2016StuFraming}. ``Eric's''(Section \ref{sec:Eric}) shift from conceptual math to algorithmic math is responsive to the instructor's correct answer to the class. In group problem solving, shifts can also be internal to the group: when members of a group disagree, one student might cause the group to shift to another frame\cite{Nguyen2016StuFraming}. Even in individual problem solving, students may shift to another frame in the ordinary course of solving a problem.
Epistemic games have been previously used for studying problem solving behaviors at the introductory level\cite{Tuminaro2004a}. However, at the upper division the strict move structure of these introductory e-games breaks down, and it may be more productive to look at which frames students operate in\cite{Bing2009,Irving2013framing,Nguyen2016StuFraming}.
In a similar contrast,
Sherin\cite{Sherin2001} compares the conceptual schemata associated with symbolic forms with Larkin's\cite{Larkin1980} ``principled-based schemata view''. He explains that the goal of his schemata is conceptual understanding and the goal of the latter schemata is step-based problem solving. However, these two views are again situated at the introductory level where conceptual mathematics is rare. In our framework, both conceptual understanding and algorithmic thinking can be mathematical or physical, allowing for greater freedom in modeling upper-division student thinking. As evidenced by ``Eric'' (Section \ref{sec:Eric}), conceptual thinking in not the only productive aspect of thinking about physics.
``Eric" (Section \ref{sec:Eric}) switched from conceptual math to algorithmic math, we do not mean to imply that algorithmic math is universally more productive than conceptual math. Rather, what counts as productive framing depends strongly on the problem context, and different frames may be productive at different times within a problem. Students' difficulties in quantum mechanics -- such as thinking that the probability density is time independent for a superposition of two stationary states, as this student does -- may simply be the result of unproductive framing and not fundamental inability to solve these problems or conceptual ``difficulties''\cite{Singh2015, Emigh2015, Passante2015}. Modeling students' problem solving as movement in the math-physics-algorithmic-conceptual plane allows for a richer description of students' problem solving behavior than mere difficulty identification, even as difficulty identification may more exactly specify the particular confusion or incorrect reasoning students exhibit.
There is another external factor that is more important in influencing students' framing in a problem solving context, even before being affected by other humans such as groupmates or the instructor. As soon as students read a problem, the problem statement framing interacts with the students' framing. In future work, we will use this theoretical framework to categorize students' framing and then analyze their responses as an artifact of the problem statement and not just due to the final correct answer or correct reasoning path.
\section{Conclusion}
In this study our goal was to examine students' problem solving behaviors in the often-messy setting of the classroom. We're particularly interested in how students solve problems collaboratively in groups, and especially in the ways they connect math and physics reasoning to solve upper-division problems.
We identified four epistemological frames: algorithmic math, conceptual math, conceptual physics and algorithmic physics. We presented several examples of students' group problems solving switching frames to productively and correctly solve a problem. This framework divides possible student errors into three different categories as displacement error, transition error and content error. The displacement error reveals students unproductive frame of the situation. Content error shows what pieces of knowledge have to be activated to understand all the ideas incorporated in the problem frame. The last error is transition, where students have ideas in different mental spaces but do not coordinate them.
This framework developed as a result of analysis of spontaneous and natural moments of in-class activities during one semester. There was no constraint in the problem solving sessions of the class, except, that the time duration of the problem solving session was limited. However, still students had enough time to illustrate natural moments of problem solving, to become creative, to get enough engaged with the problem to ``get stuck'' and then ``un-stuck'', or become so deeply engaged in the group problem solving to ignore the instructor's comment for several minutes after he has already paused the problem solving session. This kind of problem solving is more ecologically valid than the problem solving in individual clinical interviews\cite{Russ2012}, and thus as a field we should attend more carefully to it.
Instructionally, this framework is a useful tool for instructors to assess and facilitate different moments of problem solving sessions in their classroom settings. Some students like ``Eric''(Section \ref{sec:Eric}), are self reflective and get ``un-stuck'' by noticing the missing parts of their solution. However, not all of the students might act as ``Eric'' does. Using this framework may help instructors notice when students are stuck because of unproductive framing, and give them tools to nudge students into a more productive frame. Instructors can tip students into different frames\cite{Irving2013framing,ChariInsFraming} or gently nudge students to use additional resources\cite{Singh2015,Vygotsky1978} to resolve content errors.
\begin{acknowledgments}
The authors gratefully acknowledge the contributions of the KSUPER group who participated in inter-rater reliability testing and codebook development discussions. An earlier version of this theoretical framework appeared in the ICLS proceedings\cite{Thompson2016ICLS}, and we are grateful to those anonymous reviewers for their feedback on the theoretical framework. Portions of this research were funded by NSF DUE-1430967, the KSU Office of Undergraduate Research and Creative Inquiry, and the KSU Physics Department.
\end{acknowledgments}
|
\section{Introduction}
Ising spin chain in a transverse field%
\begin{equation}
H=J\sum_{j=1}^{N}\sigma_{j}^{x}\sigma_{j+1}^{x}-h\sum_{j=1}^{N}\sigma_{j}^{z}
\label{H1}%
\end{equation}
with Pauli matrices $\sigma_{j}^{\alpha}$ ($\alpha=x,z$) is a well-known
prototype for demonstrating quantum phase transition \cite{Sachdev}.
Jordan-Wigner transformation~can be employed to solve it
\cite{J-W,Lieb,Pfeuty}. By neglecting the boundary effect in thermodynamic
limit, Lieb \textit{et al}. defined and solved a "c-cycle" problem. While the
original problem without any approximation is called an "a-cycle" one
\cite{Lieb,Suzuki}. In the c-cycle problem, the thermodynamic limit is
performed at the beginning, which brings the model to a free fermion problem.
While in the a-cycle problem, if we consider a perfect periodic boundary
condition (PBC) for the original spin model, we get a constraint fermion
problem and have to keep an arbitray $N$ during the calculation. We only have
the opportunity to take the limit $N\rightarrow\infty$ at the end of calculation.
The theoretical properties of the model (\ref{H1}) have been related to real
materials since decades ago \cite{Suzuki,Dutta}. Researchers have also been
looking for nowadays state-of-art techniques, such as the ones based on
laser-cooled and trapped atomic ions, to mimic this model \cite{Edwards,Kim}.
But these artificial systems can only produce finite lattices in principle,
through which we hope to see the trend for large enough systems. The system
with perfect PBC can be realized through a ring geometric optical lattices
\cite{Amico}, which demands a thorough comprehension of the a-cycle problem.
To the best of our knowledge, it has not been treated seriously up to now
\cite{Lieb}. In this work, we shall develop a systematic method of band
structure analysis to handle it and produce exact result that can match the
full degrees of freedom of the spin model. The results will also be confirmed
by an alternative method of exact diagonalization on small systems.
\begin{figure}[h]
\begin{center}
\includegraphics[width=3.2in,angle=0]{figure1.eps}
\end{center}
\caption{The main focus of this paper: transverse Ising ring with odd number
of lattice sites: (a) $N=3$, (b) $N=2L+1$ ($L=1,2,3,\cdots$). The Hamiltonian
is shown in (\ref{H1}), which exhibits a ring frustration due to
antiferromagnetic seam ($J>0$).}%
\label{fig1}%
\end{figure}
On the other hand, frustration is an intriguing topic. Very few frustrated
models can be solved exactly \cite{Diep}. To seek for nontrivial phenomenon,
we mainly focus on the antiferromagnetic ($J>0$) system with PBC and large
enough $N\in Odd$, because it suffers a \textit{ring frustration}
\cite{Solomon,Owerre} as a result of antiferromagnetic seam \cite{McCoy Wu}
(please see Fig. \ref{fig1}). Notice the ring frustration here is not a short
range type of the usual case. One may ask whether the odevity of $N$ plays a
meaningful character in the a-cycle problem when $N\rightarrow\infty$. The
answer is affirmative. By the rigorous solution, we demonstrate that the
combination of the a-cycle problem and ring frustration does result in a
dramatical consequence. To understand the fascinating result in a contrastive
manner, we also discuss the system without any frustration, i.e. with $N\in
Even$.
We arrange the contents of this paper as follows. In Section \ref{sec:JW}, we
discuss the details about how the periodic spin problem is turned into a
fermionic a-cycle one. In Section \ref{sec:even}, we dwell on the a-cycle
problem without ring frustration ($N\in Even$). We develop the method of band
structure analysis. In Section \ref{sec:odd}, we work on the a-cycle problem
with ring frustration ($N\in Odd$). We demonstrate that the presence of ring
frustration will induce an interesting gapless spectrum above the ground state
in the strong antiferromagnetic region. We
demonstrate that the ground state exhibits a strong longitudinal spin-spin
correlation and possesses a considerably large entropy of entanglement.
We also give finite-temperature properties of the gapless region,
including the density of states (DOS) and the specific heat. In Section
\ref{sec:exp}, we propose an experimental protocol with special concern of the
realization of ring frustration. At last, we give a discussion in Section
\ref{sec:conclusion}.
\section{Jordan-Wigner fermions and the statement of the a-cycle problem}
\label{sec:JW}
It is convenient to convert the Pauli matrices to the raising and lowering
operators,
\begin{equation}
\sigma_{j}^{x}=\sigma_{j}^{+}+\sigma_{j}^{-},\sigma_{j}^{z}=2\sigma_{j}%
^{+}\sigma_{j}^{-}-1.
\end{equation}
By introducing the Jordan-Wigner fermions that abide by the non-local
relations,%
\begin{equation}
\sigma_{1}^{+}=c_{1}^{\dag},\quad\sigma_{j}^{+}=c_{j}^{\dag}\mathrm{exp}%
(\mathrm{i}\pi\sum_{l<j}c_{l}^{\dag}c_{l}),
\end{equation}
the spin model, (\ref{H1}), can be transformed to
\begin{eqnarray}
H = Nh-2h\sum_{j=1}^{N}c_{j}^{\dag}c_{j}+J\sum_{j=1}^{N-1}(c_{j}^{\dag
}-c_{j})(c_{j+1}^{\dag}+c_{j+1})\nonumber\\
\quad-J\exp(\mathrm{i}\pi M)(c_{N}^{\dag}-c_{N})(c_{1}^{\dag}+c_{1}),
\label{H2}
\end{eqnarray}
where the total number of fermions,
$
M=\sum_{j=1}^{N}c_{j}^{\dag}c_{j},
$
does not conserve. But the parity of the system does, which is defined as
\begin{equation}
P=\exp(\mathrm{i}\pi M)=(-1)^{M}. \label{P}%
\end{equation}
The vacuum state, devoid of any fermions, corresponds to the full polarized
spin state (spin down),%
\begin{equation}
\left\vert 0\right\rangle =\left\vert \downarrow\downarrow\downarrow
\cdots\downarrow\right\rangle . \label{vacuum}%
\end{equation}
(\ref{H2}) defines the full a-cycle problem \cite{Lieb,Dutta}.
As a comparison, the c-cycle problem is defined by neglecting the last term,
$-J[\exp($i$\pi M)+1](c_{N}^{\dag}-c_{N})(c_{1}^{\dag}+c_{1})$, in \cite{Lieb}%
\begin{eqnarray}
H = Nh-2h\sum_{j=1}^{N}c_{j}^{\dag}c_{j}+J\sum_{j=1}^{N}(c_{j}^{\dag}%
-c_{j})(c_{j+1}^{\dag}+c_{j+1})\nonumber\\
\quad-J[\exp(\mathrm{i}\pi M)+1](c_{N}^{\dag}-c_{N})(c_{1}^{\dag}+c_{1}).
\label{H-c-cycle}%
\end{eqnarray}
In doing so, one has accomplished the thermodynamic limit. Thus the c-cycle
problem becomes a free fermion one \cite{Pfeuty}.
While for a system with perfect ring geometry, there holds a precise condition
on the spins, $\sigma_{N+j}^{\alpha}=\sigma_{j}^{\alpha}$. With no ends (or
boundaries) existing, nothing in (\ref{H2}) could be neglected. We have to
keep an arbitrary $N$ in the calculation. We hope to get a result containing
$N$\quad as a variable thus it facilitates us to take the limit,
$N\rightarrow\infty$. Then, we can discern the different consequences of the
limits, $N$($\in Even$)$\rightarrow\infty$ and $N$($\in Odd$)$\rightarrow
\infty$.
One should notice that, although there holds a PBC for the spin operators, a
priori PBC should not be imposed on the fermions since an anti-PBC is also a
reasonable choice. We will demonstrate both of them, $c_{N+j}=c_{j}$ (PBC) and
$c_{N+j}=-c_{j}$ (anti-PBC), are indispensable to restore the full degrees of
freedom of the original spin model, (\ref{H1}), exactly. We will see that the
parity, $P$, will be fixed by the boundary condition of the fermions.
\section{A-cycle problem without ring frustration ($N\in Even$)}
\label{sec:even}
Let us see the case without ring frustration first. To make the Fourier
transformation
\begin{equation}
c_{q}=\frac{1}{\sqrt{N}}\sum_{j=1}^{N}c_{j}\exp(\mathrm{i\thinspace}q\,j)
\end{equation}
available for solving the fermionic problem, we found the boundary condition
must be bound up with the parity. So there are two routes to be followed. When
$M\in odd$, we call it the \emph{odd channel (o)} and when $M\in even$, the
\emph{even channel (e)} respectively. The procedure inevitably becomes a
little tedious. In the following, we delicately use notations to make the
deductions as clear as possible. For example, we use the notations $N\in Even
$ and $M\in even$, although $Even$ and $even$ are the same thing.
In fact, for the case of $N\in Even$, Schultz \textit{et al}. \cite{Schultz}
had discussed the contribution of the two channels in context of the classical
two-dimensional Ising model in the same essence. But their discussion on the
thermal states of the classical two-dimensional Ising model only corresponds
to the ground state property of the quantum transverse Ising model. In this
section, we discuss the quantum model directly and develop a method of band
structure analysis for both ground states and excitations.
\subsection{Diagonalization in the odd channel}
In the odd channel ($M\in odd$), the Jordan-Wigner fermions must obey PBC:
$c_{N+j}=c_{j}$, and the momentum in the first Brillouin zone ($1$st BZ) must
take a value in the set
\begin{equation}
q^{(E,o)}=\{-\frac{N-2}{N}\pi,\ldots,-\frac{2}{N}\pi,0,\frac{2}{N}\pi
,\ldots,\frac{N-2}{N}\pi,\pi\},
\end{equation}
where the superscript $(E,o)$ denotes $N\in Even$ $(E)$ and $M\in odd$ $(o)$.
After the Fourier transformation, the Hamiltonian can be diagonalized by the
Bogoliubov transformation
\begin{equation}
\eta_{q}=u_{q}c_{q}-\mathrm{i}v_{q}c_{-q}^{\dagger}%
\end{equation}
as%
\begin{eqnarray}
H^{(E,o)}=(J-h)\left( 2c_{0}^{\dagger}c_{0}-1\right) -(J+h)\left( 2c_{\pi
}^{\dagger}c_{\pi}-1\right) \nonumber\\
\quad\quad\quad\quad+\sum_{q\in q^{(E,o)},q\neq0,\pi}\omega(q)\left(
2\eta_{q}^{\dagger}\eta_{q}-1\right) , \label{H(E,o)}%
\end{eqnarray}
where
\begin{eqnarray}
u_{q}^{2} & =\frac{1}{2}\left( 1+\frac{\epsilon(q)}{\omega(q)}\right)
,v_{q}^{2}=\frac{1}{2}\left( 1-\frac{\epsilon(q)}{\omega(q)}\right)
,2u_{q}v_{q}=\frac{\Delta(q)}{\omega(q)},\nonumber\\
\omega(q) & =\sqrt{\epsilon(q)^{2}+\Delta(q)^{2}},\epsilon(q)=J\cos
{q}-h,\Delta(q)=J\sin{q.}%
\end{eqnarray}
Notice there is no need of Bogoliubov transformation for $q=0$ and $\pi$.
\subsection{Diagonalization in the even channel}
In the even channel ($M\in even$), the Jordan-Wigner fermions must obey
anti-PBC: $c_{N+j}=-c_{j}$, and the momentum in the $1$st BZ must take a value
in the set
\begin{equation}
q^{(E,e)}=\{-\frac{N-1}{N}\pi,\ldots,-\frac{1}{N}\pi,\frac{1}{N}\pi
,\ldots,\frac{N-1}{N}\pi\}.
\end{equation}
The diagonalized Hamiltonian is%
\begin{equation}
H^{(E,e)}=\sum_{q\in q^{(E,e)}}\omega(q)\left( 2\eta_{q}^{\dagger}\eta
_{q}-1\right) . \label{H(E,e)}%
\end{equation}
\subsection{Band structure of the energy levels}
\subsubsection{The ground state(s) and energy gap}
The lowest energy state in the odd channel is%
\begin{equation}
|E_{0}^{(E,o)}\rangle=c_{\pi}^{\dag}|\phi^{(E,o)}\rangle,
\label{state E0(E,o)}%
\end{equation}
where $|\phi^{(E,o)}\rangle$ is a pure BCS-like function,%
\begin{equation}
|\phi^{(E,o)}\rangle=\prod_{\substack{q\in q^{(E,o)},0<q<\pi}}\left(
u_{q}+\mathrm{i}v_{q}c_{q}^{\dag}c_{-q}^{\dag}\right) |0\rangle, \label{BCS-Eo}%
\end{equation}
where the vacuum state $|0\rangle$ is (\ref{vacuum}). Its energy reads
\begin{equation}
E_{0}^{(E,o)}=\left\vert J-h\right\vert -(J-h)-\sum_{q\in q^{(E,o)}}\omega(q).
\end{equation}
Notice that the fermionic BCS state $|\phi^{(E,o)}\rangle$ itself can not be a
valid state for the original spin model because of the parity constraint.
Likewise, The lowest energy state in the even channel reads%
\begin{equation}
|E_{0}^{(E,e)}\rangle=|\phi^{(E,e)}\rangle, \label{state E0(E,e)}%
\end{equation}
where%
\begin{equation}
|\phi^{(E,e)}\rangle=\prod_{\substack{q\in q^{(E,e)},q>0}}\left(
u_{q}+\mathrm{i}v_{q}c_{q}^{\dag}c_{-q}^{\dag}\right) |0\rangle. \label{BCS-Ee}%
\end{equation}
Its energy reads%
\begin{equation}
E_{0}^{(E,e)}=-\sum_{q\in q^{(E,e)}}\omega(q) \label{E0(E,e)}%
\end{equation}
If $N$ is small, we always have $E_{0}^{(E,e)}<E_{0}^{(E,o)}$, so
$|E_{0}^{(E,e)}\rangle$ is the ground state.
If $N\rightarrow\infty$, $|E_{0}^{(E,e)}\rangle$ is still the ground state for
$J<h$ and there is a gap, $\Delta_{gap}=2(h-J)$, to the first excited state
$|E_{0}^{(E,o)}\rangle$. Above $|E_{0}^{(E,o)}\rangle$, there is a continuum
band of excitations. While for $J>h$, $|E_{0}^{(E,e)}\rangle$ and
$|E_{0}^{(E,o)}\rangle$ become the degenerate ground states and there is a
gap, $\Delta_{gap}=4(J-h)$, above them. Now, the sum in the ground state
energy, (\ref{E0(E,e)}), can be replaced with an integral that can be worked
out, so we get
\begin{equation}
\left. \frac{E_{0}^{(E,e)}}{N}\right\vert _{N\rightarrow\infty}%
\longrightarrow\frac{-2\left\vert J-h\right\vert }{\pi}E\left( \frac
{-4Jh}{(J-h)^{2}}\right) , \label{E0(E,e)/N}%
\end{equation}
where $E(x)$ is the complete elliptic integral of the second kind.
(\ref{E0(E,e)/N}) is non-analytic at $J/h=1$, because its second derivative in
respect of $J/h$ has a logarithmic divergent peak $\sim(1/\pi)\ln|J/h-1|$. So
in fact, we have a critical point at $J=h$. These conclusions are the same as
the ones in previous investigations \cite{SachdevScience}.
We have checked that the two states, (\ref{state E0(E,o)}) and
(\ref{state E0(E,e)}), in the limit $h\rightarrow0$ correspond to two GHZ spin
states in $\sigma^{x}$ representation,%
\begin{eqnarray}
\lim_{h\rightarrow0}|E_{0}^{(E,e)}\rangle=\frac{1}{\sqrt{2}}(|\cdots
\leftarrow_{j-1},\rightarrow_{j},\leftarrow_{j+1},\rightarrow_{j+2}%
,\cdots\rangle \nonumber\\
\quad\quad\quad\quad\quad\quad-|\cdots\rightarrow_{j-1},\leftarrow_{j},\rightarrow
_{j+1},\leftarrow_{j+2},\cdots\rangle),\\
\lim_{h\rightarrow0}|E_{0}^{(E,o)}\rangle=\frac{1}{\sqrt{2}}(|\cdots
\leftarrow_{j-1},\rightarrow_{j},\leftarrow_{j+1},\rightarrow_{j+2}%
,\cdots\rangle \nonumber\\
\quad\quad\quad\quad\quad\quad+|\cdots\rightarrow_{j-1},\leftarrow_{j},\rightarrow
_{j+1},\leftarrow_{j+2},\cdots\rangle),
\end{eqnarray}
respectively.
\subsubsection{Analysis of bands}
\label{332}
The degrees of freedom (DOF) of the fermionic problem are $2^{N}$ for both
channels, so we get $2^{N+1}$ DOF totally, which is redundantly twice of the
DOF of the original spin model. However, the odd channel requires an odd
parity and the even channel an even parity. This parity constraint helps us to
obliterate the redundant DOF in each channel exactly and reconstruct the band
structure of the original spin problem.
We can construct all excited energy levels by the BCS functions (\ref{BCS-Ee})
and (\ref{BCS-Eo}) precisely. All energy levels can be grouped into bands that
are labelled by a set of indexes $(P,Q^{(E)},n_{0},n_{\pi})$, where $P$ is the
parity defined in (\ref{P}), $Q^{(E)}$ is a quasi-particle number defined as
\begin{equation}
Q^{(E)}=\sum_{q\in q^{(E,o)}\cup q^{(E,e)}}n_{q}, \label{Q}%
\end{equation}
$n_{0}=c_{0}^{\dagger}c_{0}$, and $n_{\pi}=c_{\pi}^{\dagger}c_{\pi}$. For
example, the band indexes of the two states discussed above are $(1,0,0,0)$
and $(-1,1,0,1)$ for $|E_{0}^{(E,e)}\rangle$ and $|E_{0}^{(E,o)}\rangle$
respectively. These two bands contains only one level each. From all bands of
the fermionic problem, we can pick out the valid ones for the spin model
according to the parity constraint. Several valid bands of low energy states
are listed in Table \ref{table:E}. The energy value of each state is readily
read out from the diagonalized Hamiltonian, (\ref{H(E,o)}) and (\ref{H(E,e)}).
The band structure is available for arbitrary $N$ ($N\in Even$, $2\leq
N<\infty$). It is noteworthy that the invalid bands, for example such as
$(1,1,0,0)$ and $(-1,2,0,1)$, are prohibited by the parity constraint of each channel.
\begin{table}[ptb]
\begin{center}%
\begin{tabular}
[c]{|c|c|c|}\hline
Valid Bands & Fermionic states & Number\\
$(P,Q^{(E)},\eta_{0},\eta_{\pi})$ & $(q\neq0,\pi)$ & of states\\\hline
\multicolumn{3}{|c|}{odd channel ($M\in odd$)}\\\hline
$(-1,1,0,1)$ & $c_{\pi}^{\dagger}|\phi^{\left( E,o\right) }\rangle
=|E_{0}^{\left( E,o\right) }\rangle$ & $1$\\\hline
$(-1,1,1,0)$ & $c_{0}^{\dagger}|\phi^{\left( E,o\right) }\rangle$ &
$1$\\\hline
$(-1,1,0,0)$ & $\eta_{q}^{\dagger}|\phi^{\left( E,o\right) }\rangle$ &
$\mathrm{C}_{N-2}^{1}$\\\hline
$(-1,3,0,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}\eta_{q_{3}%
}^{\dagger}|\phi^{\left( E,o\right) }\rangle$ & $\mathrm{C}_{N-2}^{3}%
$\\\hline
$(-1,3,0,1)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}c_{\pi}^{\dagger
}|\phi^{\left( E,o\right) }\rangle$ & $\mathrm{C}_{N-2}^{2}$\\\hline
$(-1,3,1,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}c_{0}^{\dagger
}|\phi^{\left( E,o\right) }\rangle$ & $\mathrm{C}_{N-2}^{2}$\\\hline
$(-1,3,1,1)$ & $\eta_{q_{1}}^{\dagger}c_{0}^{\dagger}c_{\pi}^{\dagger}%
|\phi^{\left( E,o\right) }\rangle$ & $\mathrm{C}_{N-2}^{1}$\\\hline
$\vdots$ & $\vdots$ & $\vdots$\\\hline
\multicolumn{3}{|c|}{even channel ($M\in even$)}\\\hline
$(1,0,0,0)$ & $|\phi^{\left( E,e\right) }\rangle=|E_{0}^{\left( E,e\right)
}\rangle$ & $1$\\\hline
$(1,2,0,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}|\phi^{\left(
E,e\right) }\rangle$ & $\mathrm{C}_{N}^{2}$\\\hline
$(1,4,0,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}\eta_{q_{3}%
}^{\dagger}\eta_{q_{4}}^{\dagger}|\phi^{\left( E,e\right) }\rangle$ &
$\mathrm{C}_{N}^{4}$\\\hline
$\vdots$ & $\vdots$ & $\vdots$\\\hline
\end{tabular}
\end{center}
\caption{The valid bands satisfying the odd or even parity constraint for the
case $N\in Even$. Invalid bands are not included. The energy value of each
state is readily read out from the diagonalized Hamiltonian, (\ref{H(E,o)}) or
(\ref{H(E,e)}).}%
\label{table:E}%
\end{table}
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=3.2in,angle=0]{figure2.eps}
\end{center}
\caption{The band structure of low-energy levels for a system of $N=12$. The
purpose of this figure is to check the band structure for $N\in Even$
disclosed in the text. The dingbat diamonds ("$\diamond$") denote the results
from exact diagonalization on the original spin model, which are in perfect
coincidence with the bands of levels. The true ground state,
(\ref{state E0(E,e)}), with band indexes $(1,0,0,0)$ is set as a reference.}%
\label{N=12}%
\end{figure}
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=3.2in,angle=0]{figure3.eps}
\end{center}
\caption{The band structure of low-energy levels for a system of $N=50$. From
this figure, one can figure out the trend for $N\rightarrow\infty$\ (we still
hold $N\in Even$). The dashed black line is the lower bound of excitations
when $N\rightarrow\infty$, whose intersecting point at $J/h=1$ is a critical
point as disclosed by (\ref{E0(E,e)}). The bottom levels of many bands will
touch this critical point, which results in a divergent DOS. At both sides of
the critical point, the system is gapped. Not all bands above the dashed black
line are shown.}%
\label{N=50}%
\end{figure}
To testify the band structure further, we compare it with the result by an
exact diagonalization of the transverse Ising model with a small size, say
$N=12$. The comparison is shown in Fig. 2, where a perfect coincidence can be
clearly seen. So we see our method can restore the full degrees of freedom of
the spin model.
The band structure for a larger system, say $N=50$, is exemplified in Fig. 3,
through which we can see the trend for a large enough system, $N\rightarrow
\infty$. At the critical point, the bottom levels of many bands satisfying
$P=1$ or $P\times n_{\pi}=-1$ will touch the critical point, which will result
in a divergent DOS.
\section{A-cycle problem with ring frustration ($N\in Odd$)}
\label{sec:odd}
Now we turn to the interesting case with frustration (Fig. \ref{fig1}). The
procedure is almost the same. But the story is totally different. In the
strong antiferromagnetic region, we find a gapless spectrum above the ground
state if the system is large enough.
\subsection{Diagonalization in the odd channel}
In the odd channel ($M\in odd$), the Jordan-Wigner fermions must obey PBC:
$c_{N+j}=c_{j}$, and the momentum in the $1$st BZ must take a value in the
set
\begin{equation}
q^{(O,o)}=\{-\frac{N-1}{N}\pi,\ldots,-\frac{2}{N}\pi,0,\frac{2}{N}\pi
,\ldots,\frac{N-1}{N}\pi\}.
\end{equation}
The diagonalized Hamiltonian is%
\begin{equation}
H^{(O,o)}=(J-h)\left( 2c_{0}^{\dagger}c_{0}-1\right) +\sum_{q\in
q^{(O,o)},q\neq0}\omega(q)\left( 2\eta_{q}^{\dagger}\eta_{q}-1\right) .
\label{H(O,o)}%
\end{equation}
\subsection{Diagonalization in the even channel}
In the even channel ($M\in even$), the Jordan-Wigner fermions must obey
anti-PBC: $c_{N+j}=-c_{j}$, and the momentum in the $1$st BZ must take a value
in the set
\begin{equation}
q^{(O,e)}=\{-\frac{N-2}{N}\pi,\ldots,-\frac{1}{N}\pi,\frac{1}{N}\pi
,\ldots,\frac{N-2}{N}\pi,\pi\}.
\end{equation}
The diagonalized Hamiltonian is%
\begin{equation}
H^{(O,e)}=-(J+h)\left( 2c_{\pi}^{\dagger}c_{\pi}-1\right) +\sum_{q\in
q^{(O,e)},q\neq\pi}\omega(q)\left( 2\eta_{q}^{\dagger}\eta_{q}-1\right) .
\label{H(O,e)}%
\end{equation}
\subsection{Band structure of the energy levels}
\subsubsection{The ground state}
The lowest energy state in the odd channel reads%
\begin{equation}
|E_{0}^{(O,o)}\rangle=c_{0}^{\dag}|\phi^{(O,o)}\rangle, \label{state E0(O,o)}%
\end{equation}
where%
\begin{equation}
|\phi^{(O,o)}\rangle=\prod_{\substack{q\in q^{(O,o)},q>0}}\left(
u_{q}+\mathrm{i}v_{q}c_{q}^{\dag}c_{-q}^{\dag}\right) |0\rangle. \label{BCS-Oo}%
\end{equation}
Its energy reads
\begin{equation}
E_{0}^{(O,o)}=\left\vert J-h\right\vert +(J-h)-\sum_{q\in q^{(O,o)}}\omega(q).
\end{equation}
The lowest energy state in the even channel reads%
\begin{equation}
|E_{0}^{(O,e)}\rangle=\eta_{\frac{\pi}{N}}^{\dag}c_{\pi}^{\dag}|\phi
^{(O,e)}\rangle, \label{state E0(O,e)}%
\end{equation}
where%
\begin{equation}
|\phi^{(O,e)}\rangle=\prod_{\substack{q\in q^{(O,e)},0<q<\pi}}\left(
u_{q}+\mathrm{i}v_{q}c_{q}^{\dag}c_{-q}^{\dag}\right) |0\rangle. \label{BCS-Oe}%
\end{equation}
Its energy reads%
\begin{equation}
E_{0}^{(O,e)}=2\omega(\frac{\pi}{N})-\sum_{q\in q^{(O,e)}}\omega(q).
\end{equation}
If $N$ is small, we always have $E_{0}^{(O,o)}<E_{0}^{(O,e)}$, so
$|E_{0}^{(O,o)}\rangle$ is the ground state and $|E_{0}^{(O,e)}\rangle$ is the
first excited state.
If $N\rightarrow\infty$, the state $|E_{0}^{(O,o)}\rangle$
(\ref{state E0(O,o)}) is the ground state. In the region $J<h$, there is a
gap, $\Delta_{gap}=2(h-J)$, to the first excited state $|E_{0}^{(O,e)}\rangle$
(\ref{state E0(O,e)}). In the region $J>h$, the energy gap between
$|E_{0}^{(O,o)}\rangle$ and $|E_{0}^{(O,e)}\rangle$ disappears. But we notice
that there is no energy gap between $|E_{0}^{(O,e)}\rangle$ and the next
excitation, and so forth. In fact there appears a gapless spectrum above the
ground state $|E_{0}^{(O,o)}\rangle$. We will discuss this gapless spectrum
later in detail in Section \ref{gapless spectrum}. While at $J=h$, the ground
state energy,%
\begin{equation}
\left. \frac{E_{0}^{(O,o)}}{N}\right\vert _{N\rightarrow\infty}%
\longrightarrow\frac{-2\left\vert J-h\right\vert }{\pi}E\left( \frac
{-4Jh}{(J-h)^{2}}\right) +\frac{2}{N}(J-h)\theta(J-h), \label{E0(O,o)/N}%
\end{equation}
with a Heaviside step function $\theta(x)$ is still non-analytic. In fact, the
self-duality still holds for the frustrated ring system with odd $N$ and
ensures the ocurring of quantum phase transition at $J=h$. One can see this
clear by defining new Ising-type operators,
\begin{equation}
\tau_{j}^{z}=-\sigma_{j}^{x}\sigma_{j+1}^{x},\tau_{j}^{x}=(-1)^{j}\prod
_{l<j}\sigma_{l}^{z},
\end{equation}
to get a dual form of Hamiltonian \cite{Kogut}%
\begin{equation}
H=-J\sum_{j=1}^{N}\tau_{j}^{z}+h\sum_{j=1}^{N}\tau_{j}^{x}\tau_{j+1}^{x}.
\end{equation}
\subsubsection{Analysis of bands}
By defining the quasi-particle number
\begin{equation}
Q^{(O)}=\sum_{q\in q^{(O,o)}\cup q^{(O,e)}}n_{q},
\end{equation}
we can use the set of indexes $(P,Q^{(O)},n_{0},n_{\pi})$ to label all the
fermionic bands as we have done in Section \ref{332}. From the fermionic bands
in each channel, we can pick out the valid ones for the original spin model
according to the parity constraint. The valid bands of several low energy
levels are listed in Table \ref{table:O}. The energy value of each state is
readily read out from the diagonalized Hamiltonian, (\ref{H(O,o)}) or
(\ref{H(O,e)}). The band structure is available for arbitrary $N$ ($N\in Odd$,
$3\leq N<\infty$).
\begin{table}[ptb]
\begin{center}%
\begin{tabular}
[c]{|c|c|c|}\hline
Valid Bands & Fermionic state & Number\\
$(P,Q^{(O)},\eta_{0},\eta_{\pi})$ & $(q\neq0,\pi)$ & of states\\\hline
\multicolumn{3}{|c|}{odd channel ($M\in odd$)}\\\hline
$(-1,1,0,0)$ & $\eta_{q}^{\dagger}|\phi^{\left( O,o\right) }\rangle$ &
$\mathrm{C}_{N-1}^{1}$\\\hline
$(-1,1,1,0)$ & $c_{0}^{\dagger}|\phi^{\left( O,o\right) }\rangle
=|E_{0}^{\left( O,o\right) }\rangle$ & $1$\\\hline
$(-1,3,0,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}\eta_{q_{3}%
}^{\dagger}|\phi^{\left( O,o\right) }\rangle$ & $\mathrm{C}_{N-1}^{3}%
$\\\hline
$(-1,3,1,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}c_{0}^{\dagger
}|\phi^{\left( O,o\right) }\rangle$ & $\mathrm{C}_{N-1}^{2}$\\\hline
$\vdots$ & $\vdots$ & $\vdots$\\\hline
\multicolumn{3}{|c|}{even channel ($M\in even$)}\\\hline
$(1,0,0,0)$ & $|\phi^{\left( O,e\right) }\rangle$ & $1$\\\hline
$(1,2,0,0)$ & $\eta_{q_{1}}^{\dagger}\eta_{q_{2}}^{\dagger}|\phi^{\left(
O,e\right) }\rangle$ & $\mathrm{C}_{N-1}^{2}$\\\hline
$(1,2,0,1)$ & $\eta_{q}^{\dagger}c_{\pi}^{\dagger}|\phi^{\left( O,e\right)
}\rangle$ & $\mathrm{C}_{N-1}^{1}$\\\hline
$\vdots$ & $\vdots$ & $\vdots$\\\hline
\end{tabular}
\end{center}
\caption{The valid bands satisfying the odd or even parity constraint for the
case $N\in Odd$. The energy value of each state is readily read out from the
diagonalized Hamiltonian, (\ref{H(O,o)}) or (\ref{H(O,e)}).}%
\label{table:O}%
\end{table}
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=0.7\textwidth]{figure4.eps}
\end{center}
\caption{The band structure of low-energy levels for a system of $N=13$. The
purpose of this figure is to check the band structure for $N\in Odd$ disclosed
in the text. The bands of levels are in perfect coincidence with the results
from exact diagonalization on the original spin model (shown as dingbat
diamonds "$\diamond$"). The true ground state, (\ref{state E0(O,o)}), with
band indexes $(-1,1,1,0)$ is set as a reference.}%
\label{N=13}%
\end{figure}
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=0.7\textwidth]{figure5.eps}
\end{center}
\caption{The band structure of low-energy levels for a system of $N=51$. From
this figure, one can figure out the trend for $N\rightarrow\infty$\ ($N\in
Odd$ is still hold). In contrast with the case of $N\in Even$ in Fig.
\ref{N=50}, the low excitations is dramatically changed in the region $J/h>1$
due to the presence of ring frustration. If $N\rightarrow\infty$, we can still
draw a dashed black line, whose intersecting point at $J/h=1$ is a critical
point as disclosed by (\ref{E0(O,o)/N}). And the bottom levels of many bands
satisfying $P=-1$ or $P\times n_{\pi}=1$ will touch this critical point, which
results in a divergent DOS. There is an energy gap for the region $J/h<1$.
However, the excitations for the region $J/h>1$ are gapless, where the energy
interval of width $4h$ is depleted by $N+1$ levels occupied by $2N$ states
involving 4 bands: $(-1,1,1,0)$, $(1,2,0,1)$, $(-1,1,0,0)$, and $(1,0,0,0)$.
Not all bands above the dashed black line are shown.}%
\label{N=51}%
\end{figure}
In Fig. 4, we testify the band structure further by comparing it with the
result by the exact diagonalization on a system of $N=13$. Perfect coincidence
is observed. So we see our method restores the full degrees of freedom of the
spin model.
In Fig. 5, the band structure for a larger system with $N=51$ is exemplified,
through which we can see the trend for a large enough system, $N\rightarrow
\infty$. At the critical point, the bottom levels of many bands satisfying
$P=-1$ or $P\times n_{\pi}=1$ will touch the critical point, which will result
in a divergent DOS.
\subsubsection{ Gapless spectrum in the region $J>h$}
\label{gapless spectrum}
In the strong antiferromagnetic region $J>h$, there forms a gapless spectrum
when $N\rightarrow\infty$. It contains $2N$ states involving $4$ interwoven
bands. They occupy $N+1$ energy levels. The ground state $|E_{0}%
^{(O,o)}\rangle$ with band indexes $(-1,1,1,0)$ lies at the bottom. We relabel
it as
\begin{equation}
|E_{0}\rangle=|E_{0}^{(O,o)}\rangle=c_{0}^{\dag}|\phi^{(O,o)}\rangle.
\label{state E0}%
\end{equation}
The upper-most state is $|\phi^{(O,e)}\rangle$ with indexes $(1,0,0,0)$. We
relabel it as
\begin{equation}
|E_{\pi}\rangle=|\phi^{(O,e)}\rangle. \label{state Epi}%
\end{equation}
The other two bands are:
\begin{equation}
|E_{q}\rangle=\eta_{q}^{\dag}c_{\pi}^{\dag}|\phi^{(O,e)}\rangle,(q\in
q^{(O,e)},q\neq\pi) \label{state EOe}%
\end{equation}
with indexes $(1,2,0,1)$ and
\begin{equation}
|E_{q}\rangle=\eta_{q}^{\dag}|\phi^{(O,o)}\rangle,(q\in q^{(O,o)},q\neq0)
\label{state EOo}%
\end{equation}
with indexes $(-1,1,0,0)$. If $N$ is finite, $|E_{0}\rangle$ and $|E_{\pi
}\rangle$ are nondegenerate, while other $|E_{q}\rangle$'s are doubly
degenerate. When $N\rightarrow\infty$, these $2N$ states deplete the energy
interval of width
\begin{equation}
\Delta_{w}=E_{\pi}-E_{0}\overset{N\rightarrow\infty}{\longrightarrow}4h
\end{equation}
between $|E_{0}\rangle$ and $|E_{\pi}\rangle$. This result is beyond the
familiar schematic picture for quantum phase transition \cite{SachdevScience}.
\paragraph{Perturbative theory}
To understand the formation of gapless spectrum, let us see a perturbative
picture in the strong antiferromagnetic region $J>h$. The first term of
(\ref{H1})%
\begin{equation}
H_{0}=J\sum_{j=1}^{N}\sigma_{j}^{x}\sigma_{j+1}^{x}%
\end{equation}
is a classical Ising model, whose ground states are highly degenerate as an
effect of antiferromagnetic seam \cite{McCoy Wu}. By choosing the
representation of $\sigma^{z}$, i.e. $\sigma_{j}^{z}\left\vert \uparrow
_{j}\right\rangle $=$\left\vert \uparrow_{j}\right\rangle $ and $\sigma
_{j}^{z}\left\vert \downarrow_{j}\right\rangle $=$-\left\vert \downarrow
_{j}\right\rangle $, and denoting the two eigenstates of $\sigma_{j}^{x}$ as
$\left\vert \rightarrow_{j}\right\rangle $=$(\left\vert \uparrow
_{j}\right\rangle $+$\left\vert \downarrow_{j}\right\rangle )/\sqrt{2}$ and
$\left\vert \leftarrow_{j}\right\rangle $=$(\left\vert \uparrow_{j}%
\right\rangle -\left\vert \downarrow_{j}\right\rangle )/\sqrt{2}$
\cite{Sachdev}, we can express its $2N$-fold degenerate ground states as kink
states \cite{Solomon}:
\begin{equation}
\eqalign{|K(j),\rightarrow\rangle=|\cdots,\leftarrow
_{j-1},\rightarrow_{j},\rightarrow_{j+1},\leftarrow_{j+2},\cdots
\rangle \cr
|K(j),\leftarrow\rangle=|\cdots,\rightarrow_{j-1},\leftarrow
_{j},\leftarrow_{j+1},\rightarrow_{j+2},\cdots\rangle.}
\label{kinkstates}
\end{equation}
where kinks occur between sites $j$ and $j$+$1$. The classical Ising system
falls into one of these states by spontaneous symmetry breaking \cite{Diep}.
But they are not eigenstates of the full quantum system. The second term of
(\ref{H1}),%
\begin{equation}
V=-h\sum_{j=1}^{N}\sigma_{j}^{z},
\end{equation}
as a source of quantum fluctuation, plays the role of perturbation when $h/J$
is small. We relabel the kink states as
\begin{equation}%
\eqalign{
|2j-1\rangle=|K(j),\rightarrow\rangle \cr
|2j\rangle=|K(j),\leftarrow\rangle.
}
\end{equation}
Then by the simplest perturbative scheme based on these levels, we can deduce
the matrix form of the full spin Hamiltonian $H=H_{0}+V$ ($2N\times2N$), whose
diagonal elements read
\begin{equation}
H_{2j-1,2j-1}=H_{2j,2j}=J
\end{equation}
and off-diagonal elements read%
\begin{equation}
H_{2j,2j+1}=H_{2j-1,2j+2}=H_{2j+1,2j}=H_{2j+2,2j-1}=-h.\nonumber
\end{equation}
Other elements are zero. We can arrive at an effective Hamiltonian
approximately,%
\begin{eqnarray}
H & \approx H_{\mathrm{eff}}=J\sum_{j=1}^{N}\left( |2j-1\rangle\left\langle
2j-1\right\vert +|2j\rangle\left\langle 2j\right\vert \right) \nonumber\\
& -h\sum_{j=1}^{N}\left( |2j\rangle\left\langle 2j+1\right\vert
+|2j-1\rangle\left\langle 2j+2\right\vert +\mathrm{h.c.}\right) .
\end{eqnarray}
Now by introducing a Fourier transformation%
\begin{equation}
|2j-1\rangle=\frac{1}{\sqrt{N}}\sum_{k}|a_{k}\rangle\mathrm{e}%
^{\mathrm{i}k\,j},|2j\rangle=\frac{1}{\sqrt{N}}\sum_{k}|b_{k}%
\rangle\mathrm{e}^{\mathrm{i}k\,j}%
\end{equation}
with%
\begin{equation}
k=-\frac{N-1}{N}\pi,\cdots,-\frac{2}{N}\pi,0,\frac{2}{N}\pi,\cdots,\frac
{N-1}{N}\pi,
\end{equation}
we get%
\begin{equation}
\fl H_{\mathrm{eff}}=\sum_{k}[J\left( |a_{k}\rangle\left\langle a_{k}\right\vert
+|b_{k}\rangle\left\langle b_{k}\right\vert \right) -2h\left( \cos
k|b_{k}\rangle\left\langle a_{k}\right\vert +\cos k|a_{k}\rangle\left\langle
b_{k}\right\vert \right) ].
\end{equation}
We can diagonalize it as%
\begin{equation}
H_{\mathrm{eff}}=\sum_{k}[\left( J-2h\cos k\right) |A_{k}\rangle\left\langle
A_{k}\right\vert +\left( J+2h\cos k\right) |B_{k}\rangle\left\langle
B_{k}\right\vert ]
\end{equation}
by denoting%
\begin{equation}
|A_{k}\rangle=\frac{1}{\sqrt{2}}\left( |a_{k}\rangle+|b_{k}\rangle\right)
,|B_{k}\rangle=\frac{1}{\sqrt{2}}\left( -|a_{k}\rangle+|b_{k}\rangle\right)
.
\end{equation}
Thus the degenerate ground states of $H_{0}$ is dispersed by $V$ and form a
band of $N+1$ levels, i.e. the degeneracy is partly lifted. It is easy to
check that the states $|A_{k}\rangle$ have odd parity and the states
$|B_{k}\rangle$ have even parity. If $N\rightarrow\infty$, they form a gapless
spectrum of width $4h$. They are good approximations of the lowest $2N$
rigorous energy states. For example, the ground state of the system is%
\begin{equation}
|A_{0}\rangle=\frac{1}{\sqrt{2N}}{\sum\nolimits_{j,\tau}}|K(j),\tau
\rangle\label{sum of kink states}%
\end{equation}
approximately. It is highly entangled. The excited eigenstates are
recombinations of the $2N$ kink states likewise. If $N\rightarrow\infty$, the
low-lying excitations form a gapless spectrum of width $4h$ as the schematic
plot in Fig. \ref{perturbative pic}.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=0.7\textwidth]{figure6.eps}
\end{center}
\caption{Schematic diagram of the formation of gapless specrum above the
ground state from a perturbative point of view.}%
\label{perturbative pic}%
\end{figure}
As a relevant issue, we found that, if one deduces an effective
two-dimensional classical Ising model for the quantum Ising chain by the
first-order Suzuki-Trotter decomposition in the usual way
\cite{Suzuki,MSuzuki}, the model will fail to capture the lifting of
degeneracy of the kink states.
\subsection{Correlation function of the ground state}
Now we concern the longitudinal correlation function of the ground state. We
still follow the strategy: try to work out the correlation function as a
function of $N\in Odd$, then set the limit, $N\rightarrow\infty$, to see if
there is any surprising result. For the gapless region, we find a new type of
Toeplitz determinant that needs to be evaluated rigorously.
The two-point longitudinal spin-spin correlation function of the ground state
is defined as
\begin{equation}
C_{r,N}^{xx}=\langle\phi^{(O,o)}|c_{0}\sigma_{j}^{x}\sigma_{j+r}^{x}%
c_{0}^{\dag}|\phi^{(O,o)}\rangle.
\end{equation}
By introducing the operators, $A_{j}$=$c_{j}^{\dag}$+$c_{j}$ and $B_{j}%
$=$c_{j}^{\dag}-c_{j}$, with the relations, $A_{j}^{2}$=$1$ and $A_{j}B_{j}%
$=$\exp(-$i$\pi c_{j}^{\dag}c_{j})$, we get
\begin{equation}
C_{r,N}^{xx}=\langle\phi^{(O,o)}|c_{0}B_{j}A_{j+1}B_{j+1}\ldots B_{j+r-1}%
A_{j+r}c_{0}^{\dag}|\phi^{(O,o)}\rangle. \label{Cxx}%
\end{equation}
By making use of the Wick's theorem and the contractions in respect of
$|\phi^{(O,o)}\rangle$: $\langle c_{0}c_{0}^{\dag}\rangle$=$1$, $\langle
A_{j}c_{0}^{\dag}\rangle$=$-\langle B_{j}c_{0}^{\dag}\rangle$=$\frac{1}%
{\sqrt{N}}$, $\langle A_{j}A_{j+r}\rangle$=$-\langle B_{j}B_{j+r}\rangle
$=$\delta_{r,0}$, and $\langle B_{j}A_{j+r}\rangle$= $\mathscr{D}_{r+1}$ with
\begin{equation}
\mathscr{D}_{r}=\frac{1}{N}\sum_{_{\substack{q\in q^{\left( O,o\right)
},q\neq0}}}\exp\left( -\mathrm{i}qr\right) D(\mathrm{e}^{\mathrm{i}%
q})-\frac{1}{N} \label{Dr}%
\end{equation}
where
\begin{equation}
D(\mathrm{e}^{\mathrm{i}q})=-\frac{J-h\mathrm{e}^{\mathrm{i}q}}%
{\sqrt{\left( J-h\mathrm{e}^{\mathrm{i}q}\right) \left(
J-h\mathrm{e}^{-\mathrm{i}q}\right) }}.
\end{equation}
We arrive at a Toeplitz determinant%
\begin{equation}
C_{r,N}^{xx}=\left\vert
\begin{array}
[c]{cccc}%
\mathscr{D}_{0}+\frac{2}{N} & \mathscr{D}_{-1}+\frac{2}{N} & \cdots &
\mathscr{D}_{-r+1}+\frac{2}{N}\\
\mathscr{D}_{1}+\frac{2}{N} & \mathscr{D}_{0}+\frac{2}{N} & \cdots &
\mathscr{D}_{-r+2}+\frac{2}{N}\\
\cdots & \cdots & \cdots & \cdots\\
\mathscr{D}_{r-1}+\frac{2}{N} & \mathscr{D}_{r-2}+\frac{2}{N} & \cdots &
\mathscr{D}_{0}+\frac{2}{N}%
\end{array}
\right\vert . \label{Cxxdet}%
\end{equation}
It can be evaluated for arbitrary $r$ and $N$ directly. Notice that
$C_{r,N}^{xx}=C_{N-r,N}^{xx}$ due to the ring geometry. Next, we define%
\begin{equation}
D_{r}=\frac{1}{N}\sum_{_{\substack{q\in q^{\left( O,o\right) }}}}\exp\left(
-\mathrm{i}qr\right) D(\mathrm{e}^{\mathrm{i}q}) \label{Drgapless}%
\end{equation}
with appropriate predefined $D(\mathrm{e}^{\mathrm{i}0})$. Thus we have%
\begin{equation}
\mathscr{D}_{r}=D_{r}-\frac{D(\mathrm{e}^{\mathrm{i}0})}{N}-\frac{1}{N}%
\end{equation}
In the gapped region ($J$ $<h$), we have $D(\mathrm{e}^{\mathrm{i}0})=1$
and $\mathscr{D}_{r}=D_{r}-\frac{2}{N}$. The correlation function is given by%
\begin{equation}
C_{r,N}^{xx}=\left\vert
\begin{array}
[c]{cccc}%
D_{0} & D_{-1} & \cdots & D_{-r+1}\\
D_{1} & D_{0} & \cdots & D_{-r+2}\\
\cdots & \cdots & \cdots & \cdots\\
D_{r-1} & D_{r-2} & \cdots & D_{0}%
\end{array}
\right\vert .
\end{equation}
This is the conventional Toeplitz determinant that has been investigated in
the previous works \cite{McCoy Wu}, the correlation function decays
exponentially with a finite correlation length $\xi=-1/\ln\left( J/h\right)
$.
While in our focused gapless region ($J/h$ $>$ $1$), we have
$D(\mathrm{e}^{\mathrm{i}0})=-1$ and $\mathscr{D}_{r}=D_{r}$. Then the
correlation function is given by%
\begin{equation}
C_{r,N}^{xx}=\Theta(r,N)=\left\vert
\begin{array}
[c]{cccc}%
D_{0}+\frac{2}{N} & D_{-1}+\frac{2}{N} & \cdots & D_{-r+1}+\frac{2}{N}\\
D_{1}+\frac{2}{N} & D_{0}+\frac{2}{N} & \cdots & D_{-r+2}+\frac{2}{N}\\
\cdots & \cdots & \cdots & \cdots\\
D_{r-1}+\frac{2}{N} & D_{r-2}+\frac{2}{N} & \cdots & D_{0}+\frac{2}{N}%
\end{array}
\right\vert \label{Cxxdet1}%
\end{equation}
The extra term $\frac{2}{N}$ in each element makes it a totally new Toepolitz
determinant. If one erases the term $\frac{2}{N}$ when taking the limit
$N\rightarrow\infty$, the conventional Toeplitz determinant is arrived. But we
will show its non-local information is omitted in doing so. To this purpose,
we retain the term $\frac{2}{N}$ and keep $N$ as a variable. Another reason
for retaining the term $\frac{2}{N}$ is the fact that the dimension of the
determinant is $r\times r$, which can lead to a total contribution
proportional to $\frac{r}{N}$.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=0.7\textwidth]{figure7.eps}
\end{center}
\caption{Longitudinal correlation functions for several selected values of
$J/h$ in the gapless region. The dashed black line in (a) is an exact result,
(\ref{Cxx(h=0)}), for $J/h=\infty$ (i.e. $h=0$). In both (a) and (b), the
colored dingbat data are direct rigorous evaluations of (\ref{Cxxdet}), while
the black straight lines show the asymptotic behavior described by
(\ref{Cxx2}). (b) is a zoom-in plot for the case $J/h=1.05$.}%
\label{fig7}%
\end{figure}
First, let us see an exact result in the case $h=0$. By (\ref{Drgapless}), we
have $D_{r}=-\delta_{r,0}$. Then (\ref{Cxxdet1}) is reduced to%
\begin{equation}
C_{r,N}^{xx}=(-1)^{r}(1-2\alpha), \label{Cxx(h=0)}%
\end{equation}
where $\alpha=\frac{r}{N}$. If one takes the limit $N\rightarrow\infty$ first
and gets $C_{r,N}^{xx}\approx(-1)^{r}$, one would think this is a simple
antiferromagnets. But if we take a value of $\alpha\in(0,1/2)$, we see the
exact result, (\ref{Cxx(h=0)}), measures a non-local correlation because
$r=\alpha N\rightarrow\infty$. Please notice that the ground state,
(\ref{sum of kink states}), is an exact superposition of kink states, whose
correlation function is exactly given by (\ref{Cxx(h=0)}). This is purely a
theoretical consequence of the model. The exact result is shown as the dashed
black line in Fig. \ref{fig7}(a).
Second, we work out the asymptotic behaviour for arbitrary $J$ $>h$. For a
large enough system, we can substitute the sum in (\ref{Drgapless}) with an
integral to get%
\begin{equation}
D_{r}\overset{N\rightarrow\infty}{\longrightarrow}\int_{-\pi}^{\pi}\frac
{dq}{2\pi}\exp\left( -\mathrm{i}qr\right) D(\mathrm{e}^{\mathrm{i}q}).
\end{equation}
Now we need to evaluate the new type of Toeplitz determinant in Eq.
(\ref{Cxxdet1}). Following the earlier procedure by McCoy and Wu
\cite{Wu,McCoy,McCoy Wu}, we have proved a theorem for this special case
in Appendix A:
\textit{Theorem: Consider a Toeplitz determinant} $\Theta(r,N)$ \textit{in}
(\ref{Cxxdet1}) \textit{with}
\begin{equation}
D_{n}=\int_{-\pi}^{\pi}\frac{dq}{2\pi}\,D(\mathrm{e}^{\mathrm{i}%
q})\,\mathrm{e}^{-\mathrm{i}qn}.
\end{equation}
\textit{If the generating function }$D(\mathrm{e}^{\mathrm{i}q}%
)$\textit{\ and }$\ln D(\mathrm{e}^{\mathrm{i}q})$\textit{\ are
continuous on the unit circle }$\left\vert \mathrm{e}^{\mathrm{i}%
q}\right\vert =1$\textit{, then the behavior for large }$N$\textit{\ and
}$\mathit{r}$\textit{\ of }$\Theta(r,N)$\textit{\ is given by }%
\begin{equation}
\Theta(r,N)=\Delta_{r}(1+\frac{2\alpha}{D(\mathrm{e}^{\mathrm{i}%
0})}),
\end{equation}
\textit{\ where}%
\begin{eqnarray}
\Delta_{r} & =\mu^{r}\exp(\sum_{n=1}^{\infty}nd_{-n}d_{n}),\\
\mu & =\exp[\int_{-\pi}^{\pi}\frac{dq}{2\pi}\,\ln D(\mathrm{e}%
^{\mathrm{i}q})],\\
d_{n} & =\int_{-\pi}^{\pi}\frac{dq}{2\pi}\,\mathrm{e}%
^{-\mathrm{i}qn}\ln D(\mathrm{e}^{\mathrm{i}q}),
\end{eqnarray}
\textit{if the sum }$\sum_{n=1}^{\infty}nd_{-n}d_{n}$\textit{ is convergent.}
By applying the above theorem to the gapless region ($J$ $>h$), we get an
asymptotic behavior
\begin{equation}
C_{r,N}^{xx}=(-1)^{r}(1-\frac{h^{2}}{J^{2}})^{1/4}(1-2\alpha). \label{Cxx2}%
\end{equation}
It is clear (\ref{Cxx2}) coincides with (\ref{Cxx(h=0)}). This asymptotic
behavior is depicted in Fig. \ref{fig7}, which is perfectly coincident with
the direct evaluations of (\ref{Cxxdet}). This surprising result is totally
different from the conventional findings \cite{Sachdev}.
\subsection{Entanglement entropy of the ground state}
Entanglement entropy is another powerful quantity for exhibiting the entangled
nature of a system. We define the reduced density matrix $\rho_{l}$%
=tr$_{N-l}|E_{0}^{(O,o)}\rangle\langle E_{0}^{(O,o)}|$ and the entanglement
entropy (EE) $S_{l}$=$-$tr$(\rho_{l}\log_{2}\rho_{l})$, where the trace is
performed on the spin states of contiguous sites from $j$=$1$ to $N-l$. We can
evaluate the EE numerically by utilizing the matrix
\cite{Vidal,Latorre,Amico-RMP}
\begin{equation}
\Gamma_{l}=\left\vert
\begin{array}
[c]{cccc}%
\Pi_{0} & \Pi_{1} & \cdots & \Pi_{l-1}\\
\Pi_{-1} & \Pi_{0} & \cdots & \Pi_{l-2}\\
\cdots & \cdots & \cdots & \cdots\\
\Pi_{1-l} & \Pi_{2-l} & \cdots & \Pi_{0}%
\end{array}
\right\vert \mathrm{ with }\Pi_{l}=\left\vert
\begin{array}
[c]{cc}%
0 & -g_{l}\\
g_{-l} & 0
\end{array}
\right\vert ,
\end{equation}
where $g_{l}$=$\mathscr{D}_{l-1}$+$\frac{2}{N}$. Let $V\in SO(2l)$ denote an
orthogonal matrix that brings $\Gamma_{l}$ into a block diagonal form such
that $\Gamma_{l}^{C}$=$V\Gamma_{l}V^{T}$=$\bigoplus_{m=0}^{l-1}($i$v_{m}%
\sigma_{y})$ with $v_{m}\geq0$. Then $S_{l}$ is given by $S_{l}$=$\sum
_{m=0}^{l-1}H_{2}(\frac{1+v_{m}}{2})$ with $H_{2}(x)$=$-x\log_{2}%
x-(1-x)\log_{2}(1-x)$. The numerical results for $l$=$(N-1)/2$ are shown in
Fig. \ref{fig8}. We observe the EE in the gapped region ($J<h$) is small until
near the critical point, where it abruptly tends to become divergent as
predicted by CFT \cite{Holzhey,Korepin,Calabrese,Lin}. While in the gapless
region ($J>h$), we observe $S_{(N-1)/2}$ with $N\rightarrow\infty$ approaches
its minimal value $2$ when $h\rightarrow0$. In fact, one can verify that
$|E_{0}^{(O,o)}\rangle$ evolves with $h\rightarrow0$ adiabatically into a
superposition of all kink states in Eq. (\ref{sum of kink states}), whose EE
is exactly $2$. As a comparison, the EE of the well-known GHZ state is
$\log_{2}2$=$1$.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=0.7\textwidth]{figure8.eps}
\end{center}
\caption{(Color
online) Entanglement entropy $S_{(N-1)/2}$ as a function of $J/h$ for a
sequence of number of lattice sites $N$. The inset shows examples of finite
size scaling for extrapolating to $N\rightarrow\infty$. At the critical point,
the numerical data fit a divergent behavior, $S_{(N-1)/2}\sim\frac{1}{6}%
\log_{2}N$, coincident with the prediction by CFT \cite{Holzhey}.}%
\label{fig8}%
\end{figure}
\subsection{Finite temperature properties in the gapless region}
In the gapless region, the dispersed but neatly aligned lowest $2N$ states,
(\ref{state E0})-(\ref{state EOo}), dominate the system's properties at low
temperatures ($T$ $\ll4h/k_{B}$), where $k_{B}$ is the Boltzmann constant.
This fact facilitates us to work out some quantities at low temperatures based
on the partition function
\begin{equation}
Z=\sum_{q\in q^{(O,o)}\cup q^{(O,e)}}\mathrm{e}^{-\beta E_{q}},
\end{equation}
where $\beta=\frac{1}{k_{B}T}$. The DOS is defined as
\begin{equation}
\rho(E)=\frac{1}{N}\sum_{q\in q^{(O,o)}\cup q^{(O,e)}}\,\delta(E-E_{q}).
\end{equation}
If $N\rightarrow\infty$, the summation in the $1$st BZ can be replaced with
integral, so we get a DOS,%
\begin{equation}
\rho(x)=\frac{4(x+2J-2h)}{\pi\sqrt{x(x-4h)(4h-x-4J)(x+4J)}},
\end{equation}
where $x=E-E_{0}$. It can be expanded as%
\begin{equation}
\rho(x)=ax^{-1/2}+bx^{1/2}+O(x^{3/2})
\end{equation}
with
\begin{equation}
a=\frac{(J-h)^{1/2}}{\pi(Jh)^{1/2}},b=\frac{(h^{2}+Jh+J^{2})}{8\pi
(Jh)^{3/2}(J-h)^{1/2}}.
\end{equation}
So we get the specific heat per site at low temperature,
\begin{equation}
\frac{C_{M}(T)}{N}\approx\frac{k_{B}}{2}\left[ 1+\frac{2bk_{B}T(4a+bk_{B}%
T)}{(2a+bk_{B}T)^{2}}\right] . \label{CM}%
\end{equation}
\section{Experimental proposal}
\label{sec:exp}
We can design a large enough one to see the effect of ring frustration with
nowadays state-of-art techniques based on laser-cooled and trapped atomic
ions. In fact, the case for $N=3$ has been experimentally realized
\cite{Edwards}. To generate a system with larger $N\in Odd$ and ensure that
the frustration comes from the ring geometry not from short-range
interactions, we provide another proposal.
In our proposal as shown in Fig. \ref{fig9}, there are two key points. The
first point is to produce a ring potential with odd number of traps. In
$x$-$y$ plane, we impose $N$ beams of independent standing wave lasers which
are obtained by frequency selection. Then, each standing wave will contributes
an optical potential along $\overrightarrow{k_{i}}$\ direction that can be
expressed as $V_{x\mathrm{-}y}\cos^{2}(\overrightarrow{k_{i}}\cdot
\overrightarrow{r_{i}}-\phi)$\ for the i-th beam, where $\overrightarrow{k_{i}%
}$ is the strength of beams and $\phi$ is the phase shift. The angle between
two neighboring lasers is $2\pi/N$. Thus, by adopting appropriate
$V_{x\mathrm{-}y}$\ and $\phi$, we can obtain a circular lattice potential with
$N$ traps in $x$-$y$ plane (Fig. \ref{fig9}(a) and (b)).
The second point is to realize the antiferromagnetic transverse model
robustly. In $z$ direction we apply two independent standing wave lasers,
$V_{z1}\cos^{2}(k_{z}z)$ and $V_{z2}\cos^{2}(2k_{z}z)$, where the former has
twice wave length of the latter. Eventually, we obtain a periodical two-leg
ladder potential by forming a double-well potential in $z$ direction (Fig.
\ref{fig9}(c) and (d)). In real experiment, there is additional harmonic
trapping potential $V_{trap}\left( x^{2}+y^{2}\right) $. The total potential
can be written as%
\begin{eqnarray}
V\left( x,y,z\right) & =V_{trap}\left( x^{2}+y^{2}\right) +V_{z1}%
\cos^{2}\left( k_{z}z\right) +V_{z2}\cos^{2}\left( 2k_{z}z\right)
\nonumber\\
& +V_{x\mathrm{-}y}\sum_{i=1}^{N}\cos^{2}\left( \overrightarrow{k_{i}}%
\cdot\overrightarrow{r_{i}}-\phi\right) .
\end{eqnarray}
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=0.7\textwidth]{figure9.eps}
\end{center}
\caption{(a) Scheme of the proposed experimental setup in $x$-$y$ plane. Each
arrow depicts a wave vector of a standing-wave laser. The angle between any
two neighboring lasers is $2\pi/N$. (b) The exemplified color map of optical
potential where a ring of $13$ trapping wells is shown by the dark blue
potential wells. (c) The arrangement of lasers in $z$ direction, where two
standing wave lasers form an isolated double wells potential. (d) The total
two-leg ladder potential.}%
\label{fig9}%
\end{figure}
Then, let us consider loading into the ladders with cold atoms
which have two relevant internal states denoted as pseudo-spin states
$\lambda=\uparrow,\downarrow$. The lattice potential experienced by cold atoms
depends on which of those two internal states are located. For sufficiently
deep potential and low temperatures, the system will be described by the
following bosonic or fermionic Hubbard model \cite{Duan},%
\begin{eqnarray}
H_{Hub} & =\sum_{j,\lambda,s}(-t_{\lambda})(a_{j\lambda,s}^{\dag
}a_{(j+1)\lambda,s}+\mathrm{h.c.})+\sum_{j,\lambda}(-t_{\lambda})(a_{j\lambda
,1}^{\dag}a_{j\lambda,2}+\mathrm{h.c.})\nonumber\\
& +\frac{1}{2}\sum_{j,\lambda,s}U_{\lambda}n_{j\lambda,s}(n_{j\lambda
,s}-1)+\sum_{j,s}U_{\uparrow\downarrow}n_{j\uparrow,s}n_{j\downarrow,s},
\label{exphhub}%
\end{eqnarray}
where $s=1,2$ is the leg index. With the conditions of Mott insulator limit
$t_{\lambda}\ll U_{\lambda}$, $U_{\uparrow\downarrow}$ and half filling
$\langle n_{j\uparrow,s}\rangle+\langle n_{j\downarrow,s}\rangle\approx1$, the
low-energy Hamiltonian of (\ref{exphhub}) is mapped to the $\mathrm{XXZ}$ model
by second-order perturbation,%
\begin{eqnarray}
H_{s} & =\sum_{j,s}\pm J_{\perp}(S_{j,s}^{x}S_{j+1,s}^{x}+S_{j,s}%
^{y}S_{j+1,s}^{y})+J_{z}S_{j,s}^{z}S_{j+1,s}^{z}\nonumber\\
& +\sum_{j}\pm K_{\perp}(S_{j,1}^{x}S_{j,2}^{x}+S_{j,1}^{y}S_{j,2}^{y}%
)+K_{z}S_{j,1}^{z}S_{j,2}^{z}, \label{exphspin}%
\end{eqnarray}
where the pseudo-spin operator $\mathbf{S}=a^{\dag}\overrightarrow{\sigma}%
a/2$, $\overrightarrow{\sigma}=(\sigma_{x},\sigma_{y},\sigma_{z})$ are the
Pauli matrices and $a^{\dag}=\left( a_{\uparrow}^{\dag},a_{\downarrow}^{\dag
}\right) $. The positive signs before $J_{\perp},K_{\perp}$ are for fermionic
atoms and negative signs for bosonic one. The interaction coefficients for
bosons are given by,
\begin{eqnarray*}
J_{\perp} & =\frac{4t_{\uparrow}t_{\downarrow}}{U_{\uparrow\downarrow}%
},J_{z}=\frac{2\left( t_{\uparrow}^{2}+t_{\downarrow}^{2}\right)
}{U_{\uparrow\downarrow}}-\frac{t_{\uparrow}^{2}}{U_{\uparrow}}-\frac
{t_{\downarrow}^{2}}{U_{\downarrow}},\\
K_{\perp} & =\frac{4t_{\uparrow}^{\prime}t_{\downarrow}^{\prime}%
}{U_{\uparrow\downarrow}},K_{z}=\frac{2t_{\uparrow}^{\prime2}+t_{\downarrow
}^{\prime2}}{U_{\uparrow\downarrow}}-\frac{t_{\uparrow}^{\prime2}}%
{U_{\uparrow}}-\frac{t_{\downarrow}^{\prime2}}{U_{\downarrow}}.
\end{eqnarray*}
For fermions, we only need to omit the last two terms in $J_{z}$ and $K_{z}$.
By modulating the intensity, the phase shift of the trapping laser beams, and
the $s$ wave scattering length through Feshbach resonance, we can obtain a
desired Hamiltonian from (\ref{exphspin}),%
\begin{equation}
H_{s}=\sum_{j,s}J_{z}S_{j,s}^{z}S_{j+1,s}^{z}+\sum_{j}K_{{}}\vec{S}_{j,1}%
\cdot\vec{S}_{j,2}.
\end{equation}
The properties of this system are dominated by the pseudo-spin singlet
$\left\vert s\right\rangle _{j}=\left( \left\vert \uparrow\downarrow
\right\rangle _{j}-\left\vert \downarrow\uparrow\right\rangle _{j}\right)
/\sqrt{2}$ and triplet $\left\vert t_{0}\right\rangle _{j}=\left( \left\vert
\uparrow\downarrow\right\rangle _{j}+\left\vert \downarrow\uparrow
\right\rangle _{j}\right) /\sqrt{2}$\ on the rung of the ladders in low
energy. At this time, the system can be mapped to the transverse Ising ring,
(\ref{H1}), that we desired \cite{ChenQH}.
\section{Conclusion and discussion}
\label{sec:conclusion}
In this work we have treated a special system, the transverse Ising ring, with
perfect PBC. The main focus is placed on the case with ring frustration due to
antiferromagnetic seam. We have demonstrated how the fermionic a-cycle problem
is applied for solving the transverse Ising ring based on a method of band
structure analysis. We have shown it is crucial to project out the redundant
DOF of the fermions to restore the full DOF of the original spin model. The
odevity of the number of lattice sites triggers or shuts the presence of ring
frustration. The most intriguing result is that the system in strong
antiferromagnetic region develops a gapless spectrum when the ring frustration
is turned on no matter how large the system is. To the best of our knowledge,
this gapless spectrum is totally unaware in previous investigations. The
non-local nature of the longitudinal correlation function of the ground state
is uncovered in detail. To understand all the fascinating properties of the
system, we presented a treatment of perturbative theory for a simple but
reliable cartoon picture of the formation of gapless spectrum. As an
approximation for large enough system and low enough temperature, the DOS and
specific heat are worked out. We also proposed an experimental protocol for
observing the fascinating phenomenon due to the ring frustration.
There are some issues need to be specified. First, the method of band
structure analysis encounters a problem for evaluating thermodynamic
quantities at arbitrary temperature, because the fermionic a-cycle problem is
not a free fermion one, to which the Fermi distribution can not be applied and
the summation on the constraint fermionic states can not be accomplished in a
closed form. But for finite system, it is just a matter of amount of
computation. The finite system can be utilized to produce highly entangled
states. Second, the method for the a-cycle problem can not be applied to the
system with OBC, where the c-cycle problem is applicable.
The odevity-induced phenomenon is reminiscent of the one in the well-known
spin ladders \cite{Dagotto}. But the situation here is more dramatical because
the difference between the consequences of $N\in Even$ and $N\in Odd$ is
robust even when $N\rightarrow\infty$.
We acknowledge useful discussions with Yan He. This work was supported by the
NSFC under Grants no. 11074177, SRF for ROCS SEM (20111139-10-2).
|
\section{Introduction}
\label{section:introduction}
Massive stars undergo vigorous convective shell burning at the end of their lives \cite[e.g.,][]{arnett:09,
takahashi:14, couch:15b, mueller:16b, chatzopoulos:16}. The associated
non-radial dynamics and the deviations from spherical symmetry
can grow further during collapse \citep{lai:00n,takahashi:14}. Recent
works by \citet{couch:13d, couch:15a} and \citet{mueller:15} demonstrate
that such asphericities facilitate supernova explosion. According to
\citet{couch:15a, mueller:15}, this is a result of increased turbulent
activity in the post-shock region driven by the passage of the upstream
fluctuations through the shock. The non-radial dynamics in the
post-shock region is an important factor that aids the expansion of the
supernova shock \citep[e.g.,][]{herant:95, bhf:95, janka:96, blondin:03,
foglizzo:06, foglizzo:07, hanke:12, hanke:13, janka:12b, dolence:13,
murphy:13, burrows:13a, takiwaki:14, ott:13a, abdikamalov:15,
radice:15a, melson:15a, melson:15b, lentz:15, fernandez:15a,
foglizzo:15, cardall:15, radice:16a, bruenn:16, janka:16, roberts:16n}.
In this work, we investigate the physics of the interaction of the upstream turbulence with the supernova shock and its effect on the post-shock flow using a linear perturbation theory commonly known as the linear interaction approximation (LIA) theory. The LIA, which we extend to include the nuclear dissociation at the shock, is a powerful tool originally developed in the 1950s by \cite{ribner:53}, \cite{moore:54}, and \cite{chang:57}, followed by other works \citep[e.g.,][]{ribner:54, chang:57, mckenzie:68, jackson:90, mahesh:96, mahesh:97, duck:97, fabre:01, wouchuk:09, huete:11, huete:12}.
In the LIA, the shock is modeled as a planar discontinuity with
no intrinsic scale and the flow is decomposed into the mean and
fluctuating parts. Both components can be specified arbitrarily in the
upstream flow. Once the upstream field is specified, the downstream
field can be fully determined using the Rankine-Hugoniot jump conditions
at the shock \citep[e.g.,][]{sagaut:08}. The LIA is valid in
the regime of sufficiently small fluctuations such that the mean flow
satisfies the usual jump conditions, while the turbulent fluctuations
satisfy the linearized jump conditions. Numerical simulations by
\cite{lee:93b} suggest that this approximation is valid when
\begin{equation}
\label{eq:lia}
{\cal M}'^2\lesssim 0.1 ({\cal M}^2_1-1),
\end{equation}
where ${\cal M}'$ and ${\cal M}_1$ are the Mach number of upstream turbulence and mean flow, respectively \citep[see also][]{ryu:14}. In massive star shell convection, ${\cal M}' \sim 0.1$ \citep[e.g.,][]{mueller:16b}, which at most can increase by a factor of several during contraction (more precise calculation of this is given below in Section \ref{sec:implications}). Since ${\cal M}_1 \gtrsim 5$ in core-collapse supernovae (CCSNe), condition (\ref{eq:lia}) is well satisfied and we expect the LIA to be an excellent approximation for studying the interaction of CCSN shocks with progenitor asphericities.
\section{The Linear Interaction Approximation}
\label{sec:lia}
The LIA employs the \citet{kovasznay:53} decomposition of the
fluctuating field, according to which any small fluctuations in a
turbulent flow can be decomposed individual Fourier modes that are
characterized by their type, wavenumber, and frequency. There are three
types of modes: vorticity, entropy, and acoustic modes. The vorticity
mode is a solenoidal velocity field that is advected with the mean flow.
It has no pressure or density fluctuations. The entropy mode is also
advected with the flow and it represents density and temperature
fluctuations with no associated pressure or velocity variations. The
acoustic mode represents sound waves that travel relative to the mean
flow. It has isentropic pressure and density fluctuations and
irrotational velocity field. All Kovasznay modes evolve independently in
the limit of weak fluctuations and the interaction of each mode with the
shock wave can be studied independently. Integration over all individual
modes yields the full statistics of the turbulent flow
\citep[e.g.,][]{sagaut:08}.
We assume that the shock wave is a planar discontinuity and we choose our $x$-axis ($y$-axis) to be perpendicular (parallel) to the shock front. The average shock position is assumed to be at $x=0$ and the mean flow is in the positive $x$ direction. The quantities $U$, $\bar{\rho}$, $\bar{p}$, $\bar{T}$, and $\mathcal{M}$ represent the mean velocity, density, pressure, temperature, and Mach number. We choose the values of these parameters to approximate the CCSN shock by requiring vanishing Bernoulli parameter for the upstream flow, as described in Appendix \ref{app:shock}. We employ a gamma-law equation of state with $\gamma=4/3$. The quantities $u'$, $\upsilon'$ $\rho'$, $p'$, $T'$ denote the perturbation in the $x$- and $y$-components of velocity, density, pressure, and temperature, respectively. Hereafter, subscripts 1 and 2 will denote the upstream and downstream states. The upstream vorticity mode is modeled via a planar shear wave with wavenumber $(m\kappa,l\kappa)$ and angular frequency $\kappa mU_1$:
\begin{eqnarray}
\label{eq:u1}
\frac{u_1'}{U_1} &=& lA_\upsilon e^{i\kappa(mx+ly-U_1mt)}, \\
\frac{\upsilon_1'}{U_1} &=& -mA_\upsilon e^{i\kappa(mx+ly-U_1mt)},
\label{eq:v1}
\end{eqnarray}
while the upstream entropy mode is given by another planar sinusoidal wave with the same wavenumber and frequency,
\begin{eqnarray}
\frac{\rho_1'}{\bar{\rho}_1} &=& A_e e^{i\kappa(mx+ly-U_1mt)}, \\
\label{eq:rho1}
\frac{T_1'}{\overline{T}_1} &=& -\frac{\rho'}{\bar{\rho}_1},
\label{eq:T1}
\end{eqnarray}
where $m=\cos\psi_1$ and $l=\sin\psi_1$ and $\psi_1$ is the angle
between the $x$-axis and the direction of propagation of the incident
perturbation. $A_\upsilon$ and $A_e$ are the amplitudes of the incident
vorticity and entropy waves, respectively. In the present work, we
ignore acoustic waves in the pre-shock region, which corresponds to the assumption of zero
pressure fluctuations in the upstream flow ($p_1' = 0$). The effect of the upstream acoustic
component will studied in our future work.
When vorticity and/or entropy waves hit a shock wave, the former responds by changing its position and shape. In the framework of the LIA, for a perturbation of form (\ref{eq:u1})-(\ref{eq:T1}), the shock surface deforms into a shape of a sinusoidal wave propagating in the $y$-direction:
\begin{equation}
\label{eq:shock_lia0}
\xi(y,t) = -\frac{L}{i\kappa m} e^{i\kappa (ly-U_1mt)}, \\
\end{equation}
where $\xi(y,t)$ is the $x$-coordinate of the shock position at time $t$ and ordinate $y$ and $L$ is a quantity that characterizes the amplitude of the shock oscillations (cf. Fig.~\ref{fig:shocklia}).
The interaction of the vorticity and entropy waves with the shock generates a downstream perturbation field consisting of vorticity, entropy, and acoustic waves given by~\citep{mahesh:96,mahesh:97}
\begin{eqnarray}
\label{eq:u2}
&&\mkern-60mu\frac{u_2'}{U_1} = F e^{i\tilde{k}x}e^{i\kappa(ly-U_1mt)} + Ge^{ik(\mathcal{C}mx+ly-U_1 mt)}, \\
\label{eq:v2}
&&\mkern-60mu\frac{\upsilon_2'}{U_1} = H e^{i\tilde{k}x}e^{i\kappa(ly-U_1mt)} + Ie^{i\kappa(\mathcal{C}mx+ly-U_1 mt)}, \\
\label{eq:p2}
&&\mkern-60mu\frac{p_2'}{\bar{p}_2} = K e^{i\tilde{k}x}e^{ik(ly-U_1mt)} \\
&&\mkern-60mu\frac{\rho_2'}{\bar{\rho}_1} = \frac{K}{\gamma} e^{i\tilde{\kappa}x}e^{ik(ly-U_1mt)} + Qe^{i\kappa(\mathcal{C}mx+ly-U_1 mt)}, \\
\label{eq:T2}
&&\mkern-60mu\frac{T_2'}{\overline{T}_1} = \frac{(\gamma-1)K}{\gamma} e^{i\tilde{\kappa}x}e^{i\kappa(ly-U_1mt)} -Qe^{i\kappa(\mathcal{C}mx+ly-U_1mt)}.
\end{eqnarray}
The schematic representation of this process is depicted in
Fig.~\ref{fig:shocklia}. Note that these waves have the same angular
frequencies and $y$-components of wavenumbers as those of the upstream
waves (\ref{eq:u1})-(\ref{eq:T1}). The coefficients $F$, $H$, and $K$ are the amplitudes of the
acoustic component, while coefficients $G$, $I$, and $Q$ are associated
with the entropy and vorticity components. The former two components
have the same wavenumber vector $(m{\cal C}
\kappa,l\kappa)$ and angular frequency $\kappa mU_1$, for which reason
they are often referred to as entropy-vorticity waves. The acoustic
component has the same angular frequency but different wavenumber
$(\tilde{\kappa},l\kappa)$, where $\tilde{\kappa}$ is calculated in
Appendix \ref{app:lia}. The parameter $\mathcal{C}$ is the compression
factor at the shock, $\mathcal{C}=\bar{\rho}_2/\bar{\rho}_1=U_1/U_2$,
which can be obtained from the Rankine-Hugoniot condition as (cf.
Appendix \ref{app:shock}):
\begin{equation}
\label{eq:c}
{\mathcal C} = \frac{\gamma+1}{\gamma+\frac{1}{M_1^2} - \sqrt{\left(1-\frac{1}{M_1^2}\right)^2+(\gamma+1)\frac{(\gamma-1){\cal M}_1+2}{{\cal M}_1^2} \bar \epsilon}}.
\end{equation}
Here, $\bar \epsilon$ is the dimensionless nuclear dissociation parameter, which characterizes nuclear dissociation energy, as explained in Appendix \ref{app:shock}. It typically ranges from $0$, which represents the limit corresponding to zero nuclear dissociation, to $0.4$, which represents strong nuclear dissociation. We adopt $\bar\epsilon=0.2$ and ${\cal M}_1=5$ as our fiducial values.
In order to obtain the coefficients $F$, $G$, $H$, $I$, $K$, $Q$, $L$,
we first expand the Rankine-Hugoniot conditions to the first order in
amplitudes of incoming perturbations \citep{ribner:53, chang:57,
mahesh:97}. The solution of the resulting equations yield the
coefficients $F$, $G$, $H$, $I$, $K$, $Q$, $L$, as we demonstrate in
Appendix~\ref{app:lia}.
The downstream acoustic component depends strongly on the
incidence angle $\psi_1$. If $\psi_1$ is smaller than the critical angle
\begin{equation}
\label{eq:psic}
\psi_c = \cot^{-1} \sqrt{\frac{c_{\mathrm{s},2}^2}{U_1^2} - \frac{U_2^2}{U_1^2}},
\end{equation}
where $c_{\mathrm{s},2}$ is the downstream speed of sound, then $\tilde{\kappa}$ is real and the sound waves represent freely propagating planar sine waves. On the other hand, if $\psi_c<\psi_1$, $\tilde{\kappa}$ is complex and the solution represents an exponentially-damping planar sine wave \citep{mahesh:96,mahesh:97}.
A detailed derivation of the LIA equations, including angle $\psi_c$
(\ref{eq:psic}) and wavenumber $\tilde \kappa$ in
Eqs.~(\ref{eq:u2})-(\ref{eq:T2}), is presented in
Appendix~\ref{app:lia}.
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth,clip=false]{shocklia.pdf}
\caption{Schematic representation of the interaction of an entropy and/or vorticity waves with a shock wave in the context of the LIA formalism. The average position of the shock is aligned with the $y$-axis and the mean flow is in the positive $x$ direction. The upstream mean flow is characterized by velocity $U_1$, density $\bar{\rho}_1$, pressure $\bar p_1$, and temperature $\overline T_1$, while the corresponds downstream quantities are $U_2$, $\bar{\rho}_2$, $\bar p_2$, and $\overline T_2$. When vorticity and/or entropy waves of form (\ref{eq:u1})-(\ref{eq:T1}) hit a shock wave, the latter responds by changing its position and shape. In the framework of LIA, for such perturbations, the shock surface deforms into a sinusoidal planar wave propagating in the $y$-direction described by formula (\ref{eq:shock_lia0}). The downstream perturbation field consists of entropy, vorticity, and acoustic waves given by Eqs.~(\ref{eq:u2})-(\ref{eq:T2}).
\label{fig:shocklia}}
\end{figure}
\section{Results}
\label{sec:results}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth,clip=false]{compression_factor_eps_vs_mach2.pdf}
\caption{The compression factor $\mathcal{C}$ at the shock as a function of nuclear dissociation parameter $\bar\epsilon$ for four values of upstream Mach number ${\cal M}_1$: $2.5$, $5$, $10$, and $100$. In all of these cases, the compression factor $\mathcal{C}$ increases with $\bar\epsilon$, signifying that the nuclear dissociation leads to stronger compression.
\label{fig:c}}
\end{figure}
The key quantity affecting the evolution of the flow through a shock
wave is the compression factor $\mathcal{C}$. Fig.~\ref{fig:c} shows
$\mathcal{C}$ as a function of the nuclear dissociation parameter $\bar
\epsilon$ for four values of upstream Mach number ${\cal M}_1$: $2.5$,
$5$, $10$, and $100$. For all of these values, the compression factor
$\mathcal{C}$ grows with increasing $\bar\epsilon$, meaning that the
nuclear dissociation leads to stronger compression.
Note that the values of the compression factor
$\cal C$ are very close to each other for ${\cal M}_1=5$, $10$, and
$100$. This is a generic property of shock waves, in which the
compression factor depends on ${\cal M}_1$ very weakly when ${\cal
M}_1\gtrsim 5$.
In the following, we present our results in two parts. In the first part (Section~\ref{sec:wave}), we discuss interaction of a shock wave with individual incident waves and explore how it depends on shock and perturbation parameters. In the second part (Section~\ref{sec:turbulence}), we investigate the interaction of a shock wave with incident turbulence fields, which we model as sets of random entropy and vorticity waves.
\subsection{Interaction with a Single Wave}
\label{sec:wave}
Figure~\ref{fig:psic} shows the values of the critical angle $\psi_\mathrm{c}$ as a function of nuclear dissociation parameter $\bar\epsilon$ for four values of upstream Mach number ${\cal M}_1$: $2.5$, $5$, $10$, and $100$. Recall that the critical angle $\psi_c$ separates two regions of the solution: propagative ($\psi_1 < \psi_c$) and non-propagative ($\psi_1 > \psi_c$). The first is characterized by acoustic waves in the post-shock flow, while in the second sound waves do not propagate. In all cases, $\psi_\mathrm{c}$ increases with $\bar\epsilon$. However, this increase is rather modest. For example, for ${\cal M}_1=5$, $\psi_\mathrm{c}$ increases from $67.6^\circ$ to only $71.5^\circ$ as $\bar\epsilon$ increases from $0$ to $0.4$. These values do not change much with ${\cal M}_1$ after $\mathcal{M}_1=5$. This is a simple reflection of the above-mentioned fact that the compression factor does not depend strongly on the upstream Mach number for ${\cal M}_1 \gtrsim 5$.
For an incident vorticity wave of form given by Eqs.~(\ref{eq:u1})-(\ref{eq:v1}), the velocity field is $u_1' \, \propto \, \sin \psi_1$ and $\upsilon_1' \, \propto \, \cos \psi_1$. If the perturbation wavenumber vector $\vec{k}$ is perpendicular to the shock wave ($\psi_1 = 0$), the $x$-component of the fluctuation field is zero. When such a field hits the shock, the solution is trivial: the shock wave is not affected and the the velocity passes through the shock without any modifications, i.e., $\upsilon_2'=\upsilon_1'$ and $u_2'=u_1'=0$. The only property that changes is the $x$-component of the wavenumber: it increases by a factor of $\cal C$. Correspondingly, the wavelength of the wave decrease by the same factor.
The situation is drastically different when $\psi_1 > 0$. The perturbation velocity field now has non-zero $x$-component, which forces the shock surface to oscillate according to Eq.~(\ref{eq:shock_lia0}). Because of this, the downstream field now consists of not only vorticity waves, but also of entropy and acoustic waves, as described by Eqs.~(\ref{eq:u2})-(\ref{eq:T2}). Both entropy and vorticity waves in the post-shock region have the same wavenumber vector $({\cal C} m\kappa, l\kappa)$. Thus, the magnitude of the wavenumber vector increases by factor
\begin{equation}
\label{eq:kappa_ratio}
\frac{\kappa_2}{\kappa_1} = \sqrt{{\cal C}^2m^2+l^2}.
\end{equation}
as the wave crosses the shock. Accordingly, the wavelength of the mode decreases by the same factor across the shock.
Figure~\ref{fig:w2w1_phi} shows the ratio of averaged pre-shock and post-shock vorticities $\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle}$ as a function of angle $\psi_1$ for $\mathcal{M}_1=5$ and $\bar\epsilon=0$\footnote{Note that since the flow is restricted to $x$-$y$ plane, vorticity has only the $z$-component:
\begin{equation}
\omega_1' = \partial_x \upsilon_1' - \partial_y u_1'=-ikA_\upsilon e^{ik(mx+ly-U_1mt)}.
\end{equation}
As we show below, this is not a restriction because a general 3D problem can be expressed in terms of a 2D LIA problem.}. Here, brackets $\langle\rangle$ mean averaging over time $t$ and the $y$-coordinate. The solid black line represents the case of incident vorticity wave. The case of incident entropy and vorticity waves of the same amplitude and phase (i.e., $A_e=A_\upsilon$) is shown with red line, while the same with $180^\circ$ phase difference (i.e., $A_e=A_\upsilon e^{i\pi}$) is represented by the blue line. As we can see, when the vorticity and entropy waves are in phase, we get a significantly stronger amplification. When they are out of phase, we get the weakest amplification. Finally, in the case of pure vorticity incident wave, the amplification is roughly the average of these two regimes. For example, for $\psi_1=60^\circ$, we get $\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle}$ of $6.14$, $2.77$, and $4.46$ in these three cases. The spike in $\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle}$ around $\psi_1\simeq 69^\circ$ corresponds to the critical angle $\psi_1=\psi_\mathrm{c}$.
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{psic_vs_eps_vs_mach2.pdf}
\caption{The critical angle $\psi_\mathrm{c}$ as a function of nuclear dissociation parameter $\bar\epsilon$ for four values of upstream Mach number ${\cal M}_1$: $2.5$, $5$, $10$, and $100$.
\label{fig:psic}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{w2w1_vs_psi_vs_phi.pdf}
\caption{The amplification of vorticity across the shock as a function of angle $\psi$ for incident vorticity wave, in-phase vorticity-entropy wave (i.e., $A_e=A_\upsilon$), and out-of-phase vorticity-entropy wave (i.e., $A_e=A_\upsilon e^{i\pi}$). The upstream Mach number is ${\cal M}_1=5$. When the vorticity and entropy waves are in phase, we get strongest amplification. When they are out of phase, we get the weakest amplification.
\label{fig:w2w1_phi}}
\end{figure}
In order to explain the behavior of the vorticity fluctuations $\omega'$, \citet{mahesh:96,mahesh:97} developed a simple model, which we present here for completeness. Linearizing the Euler equations about the mean flow and neglecting the incident pressure perturbations, we get the following equation for the vorticity fluctuations $\omega'$ \citep{mahesh:96}
\begin{equation}
\label{eq:omega2}
\omega_t'+U\omega_x'=-\omega'U_x - \frac{\rho_y'}{\bar{\rho}^2}\bar{p}_x,
\end{equation}
where subscripts $t$ and $x$ mean partial derivatives with respect to these variables. The first term on the right-hand side of this equation ($-\omega'U_x$) represents the effect of the bulk compression. Since velocity drops across the shock, it amplifies the vorticity. The second term ($-\frac{\rho_y'}{\bar{\rho}^2}\bar{p}_x$) represents the baroclinic processes, which produce vorticity even from pure entropy perturbations. It can either amplify or weaken the effect of bulk compression depending on the relative phase between the vorticity and entropy waves. For incident vorticity and entropy waves of form (\ref{eq:u1})-(\ref{eq:T1}), Eq.~(\ref{eq:omega2}) reduces to
\begin{equation}
\label{eq:omega3}
-\omega'U_x - \frac{\rho_y'}{\bar{\rho}^2}\bar{p}_x \sim A_\upsilon
UU_x - A_e l \frac{p_x'}{\bar{\rho}}
\end{equation}
Since $U_x<0$ and $\bar{p}_x>0$ at the shock, the two sources of vorticity have the same sign if $A_e$ and $A_\upsilon$ have the same sign. In this case, the entropy wave enhances the amplification of vorticity across the shock. If the signs are opposite, the entropy wave weakens the vorticity amplification.
Using Eq.~(\ref{eq:omega3}), we can derive an approximate expression for the value of the downstream vorticity in terms of its upstream value (see Section 3.6 \cite{mahesh:96} for full derivation):
\begin{equation}
\label{eq:omega2_scaling}
\omega_2' \sim {\mathcal C} \omega_1' + \frac{ik\sin \psi_1}{3}A_e U_1
\frac{1-{\mathcal C}^3}{{\cal C}^2}
\end{equation}
where ${\mathcal C}$ is the compression factor (\ref{eq:c}). This suggests that the incident vorticity wave amplifies by a factor of $\mathcal{C}$ due to shock-compression, while the vorticity created by the incident entropy wave is $\propto kA_e\sin \psi_1 (1-{\mathcal C}^3)/{\mathcal C}^2$.
A word of caution is in order here. The effects due to the change of shock position and shape are absent in Eq.~(\ref{eq:omega2_scaling}) and thus it has a limited quantitative accuracy. Nevertheless, as we will see below, it describes well some key qualitative aspects of our results.
In the limit of small incidence angle $\psi_1$, Eq.~(\ref{eq:omega2_scaling}) yields $\omega_2' \sim {\mathcal C} \omega_1'$. For a $\mathcal{M}_1=5$ shock, ${\mathcal C}\simeq5.6$, which is precisely what we observe in Fig.~\ref{fig:w2w1_phi} for $\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle}$ for all the three curves. As predicted by Eq.~(\ref{eq:omega2_scaling}), the three curves gradually diverge with increasing $\psi_1$. Figure~\ref{fig:w2w1} shows the vorticity amplification $\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle}$ for a incident vorticity wave across the shock as a function of angle $\psi_1$ for various values of nuclear dissociation parameter $\bar{\epsilon}$ for $\mathcal{M}_1=5$. Due to larger compression with increasing $\bar{\epsilon}$, the amplification $\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle}$ also grows with $\bar{\epsilon}$, in agreement with the prediction of Eq.~(\ref{eq:omega2_scaling}).
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{w2w1_vs_psi_vs_eps.pdf}
\caption{Amplification of vorticity across the shock as a function of angle $\psi$ for various values of nuclear dissociation parameter $\bar{\epsilon}$ for incident vorticity waves for ${\cal M}_1=5$.
\label{fig:w2w1}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{q2q1_vs_psi_vs_phi.pdf}
\caption{The amplification of turbulent kinetic energy across the shock with $\mathcal{M}_1=5$ as a function of angle $\psi$ for purely vorticity wave, in-phase vorticity-entropy wave (i.e., $A_e=A_\upsilon$), and out-of-phase vorticity-entropy wave (i.e., $A_e=A_\upsilon e^{i\pi}$). When the vorticity and entropy waves are in phase, we get the strongest amplification. When they are out of phase, there is no significant amplification.
\label{fig:q2q1_phi}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{q2q1_vs_psi_vs_eps.pdf}
\caption{Amplification of turbulent kinetic energy across the shock as a function of angle $\psi_1$ for various values of nuclear dissociation parameter $\bar{\epsilon}$ for ${\cal M}_1=5$ for incident vorticity waves.
\label{fig:q2q1}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{q2q1_vs_psi_vs_eps_acoustic.pdf}
\caption{The ratio of kinetic energy association with sound wave to the total kinetic energy of the fluctuating field in the post-shock region as a function of the incidence angle $\psi_1$ for ${\cal M}_1=5$ for incident vorticity waves.
\label{fig:q2q1_ac_ratio}}
\end{figure}
Figure~\ref{fig:q2q1_phi} shows the ratio of turbulent kinetic energy $E'$ across the shock $E'_2/E'_1$ as a function of angle $\psi_1$ for $\mathcal{M}_1=5$ and $\bar\epsilon=0$ for the same three types of incident perturbations. We define $E'$ as
\begin{equation}
E'=\frac{1}{2}\left(\langle u'u'^*\rangle+\langle \upsilon'\upsilon'^* \rangle \right),
\label{eq:eturb}
\end{equation}
where $\langle\rangle$ means averaging over $t$ and $y$, while sign ${}^*$ denotes complex conjugate. Similarly to the case of $\langle\omega'^2\rangle$, we observe the largest (smallest) $E'_2/E'_1$ when entropy and vorticity waves are in phase (out of phase), while for pure vorticity wave, $E'_2/E'_1$ is roughly the average of the two cases. For example, at $\psi_1=60^\circ$, for in-phase waves, we get $E'_2/E'_1=4.51$, while for out-of-phase waves, we get $E'_2/E'_1=0.90$, which means that the total kinetic energy of the perturbations actually decreases across the shock in this case for this value of $\psi_1$. For pure vorticity wave, we get $E'_2/E'_1=2.36$ for the same $\psi_1$.
At $\psi_1\simeq\! 0$, we see no amplification of $E'$, while
the vorticity, as shown above, scales as
$\sqrt{\langle\omega_2'^2\rangle}/\sqrt{\langle\omega_1'^2\rangle} \sim
\mathcal{C}$. This is due to the fact that in this limit, the
$x$-component of the velocity perturbation $u_1'$ is $\simeq 0$, while
$\upsilon_1'\simeq A_\upsilon$, i.e., the velocity perturbation has only
$y$-component, which is tangential to the shock. The tangential
component of the velocity does not changes across the shock
\citep{landau:59}. Hence, there is no amplification of turbulent kinetic
energy in this limit. On the other hand, the vorticity $\omega'$ still
changes because it depends on the wavelength, which decreases by a
factor given by Eq.~(\ref{eq:kappa_ratio}).
The $x$-component of velocity of incident vorticity wave grows with
$\psi_1$ as $u_1' \propto \sin\psi_1$ (cf. Eq.~\ref{eq:u1}). The shock
responds sensitively to $u_1'$ by changing its position and shape
according to Eq.~(\ref{eq:shock_lia0}). Due to the deformation of the
shock, both $x$- and $y$-components of velocity will be perpendicular to
the shock at some $y$ and $t$. In this case, both $u'$ and $\upsilon'$ undergo
significant amplifications across the shock. The amplification factor
gradually grows with $\psi_1$ reaching, e.g., $\simeq\! 1.4$ and
$\simeq\! 2.3$ for $\psi_1=45^\circ$ for purely vorticity and in-phase
entropy-vorticity waves, respectively. The largest amplification is reached at $\psi_1 \simeq \psi_\mathrm{c}$, with amplification factors of $\simeq\! 6.2$ and $\simeq\! 13$ for these
two cases. However, such a large amplification is confined to a
narrow range of values of $\psi_1$ around $\psi_\mathrm{c}$ with width
$\lesssim 5^\circ$.
Figure~\ref{fig:q2q1} shows the amplification of kinetic energy as a function of $\psi_1$ for various values of nuclear dissociation parameter $\bar\epsilon$, ranging from $0$ to $0.4$. As we can see, $E'_2/E'_1$ exhibits only minor change with $\bar\epsilon$. The spike in $E'_2/E'_1$ around $\psi_1\simeq\!69^\circ$, which corresponds to the critical angle $\psi_1=\psi_c$, shifts towards slightly higher $\psi$ with $\bar\epsilon$ due to the fact the $\psi_c$ increases with $\bar\epsilon$ (cf. Fig.~\ref{fig:c}).
Figure~\ref{fig:q2q1_ac_ratio} shows the ratio of the kinetic energy of the acoustic component to the total kinetic energy of the entire fluctuating velocity field as a function of the incidence angle $\psi_1$ for $\bar\epsilon=0,\ 0.2,\ 0.3,$ and $0.4$ for ${\cal M}_1=5$. This ratio can reach up to $0.08$ around $\psi\sim 50^\circ$, which is a non-negligible amount.
\subsection{Interaction with Turbulence}
\label{sec:turbulence}
So far, our analysis has focused on interaction of shocks with
individual fluctuation modes. In the following, we consider interaction
with turbulent fields, which we model as sets of
random three-dimensional vorticity and entropy
waves. In the LIA, each of these waves interact independently
with the shock. The full turbulent statistics behind the shock can be
obtained by integrating over the interactions of each of these waves
with the shock.
In order to achieve this goal, we first need to establish how the 2D LIA presented so far is
related to the general three-dimensional problem. In 3D Cartesian
coordinate system $(x,y,z)$, consider an incident planar wave with
wavenumber $\vec{\kappa}_1$ that makes angle $\psi_1$ with the $x$-axis.
The latter is assumed to be perpendicular to the shock. The dynamics in
the plane spanned by vector $\vec{\kappa}_1$ and the shock normal is
identical to that of the 2D LIA problem. The component of the velocity
field perpendicular to this plane passes unchanged through the shock,
while the components parallel to the plane change according to LIA
\citep{ribner:54,mahesh:96,wouchuk:09}. In the following, we refer to
this plane as the LIA plane.
We consider two types of turbulent fields. The first is an anisotropic turbulence characterized by relation
\begin{equation}
R_{rr}=R_{\theta\theta}+R_{\phi\phi} \quad \mathrm{and} \quad R_{\theta\theta}=R_{\phi\phi},
\label{eq:rrr}
\end{equation}
where $R_{ij}$ is the $ij$-component of the Reynolds stress tensor. This means an equipartition between radial and non-radial components of turbulent kinetic energy and it was observed in buoyancy-driven turbulent convection in stellar interiors~\citep[e.g.,][]{arnett:09}. The second type is a fully isotropic turbulence represented by
\begin{equation}
R_{rr}=R_{\theta\theta}=R_{\phi\phi}.
\label{eq:rrr2}
\end{equation}
Estimate of how well Eq.~(\ref{eq:rrr2}) describes the turbulence in stellar convective shells is beyond the scope of this work. Instead, we use it as an alternative prescription in order to test the sensitivity of our results to the properties of upstream turbulence.
In order to model turbulence characterized by these relations, we randomly sample the velocity field $(v_x, v_y, v_z)$ with a statistics that satisfies these relations. Here, the $x$-component $v_x$ plays a role similar to that of the radial component in CCSNe since both of them are perpendicular to the shocks in their respective contexts. By the same rationale, $v_y$ and $v_z$ play the roles of angular components in CCSNe.
The wavenumber vectors $\vec{\kappa}_1$ of incident waves are sampled
randomly with uniform distribution on a 2D sphere. For each wave, we
decompose the velocity field into three components: the first being
perpendicular to the LIA plane, the second being perpendicular to
$\vec{\kappa}_1$ on the LIA plane, and the third being parallel to
$\vec{\kappa}_1$ on the LIA plane. The first component passes through
the shock unchanged, while the second changes according to LIA. The
third component represents the non-solenoidal part of the
velocity field, which we set to zero when constructing the
vorticity waves.
Using this velocity field, we first investigate how the spectrum of turbulence changes as it crosses the shock. Each turbulent eddy is characterized by its wavenumber $\vec{\kappa}$. According to LIA, when a turbulent eddy with a wavenumber $\vec{\kappa}_1$ passes through the shock, the $x$-component of $\vec{\kappa}_1$ increases from $\kappa_{1,x}$ to ${\cal C} \kappa_{1,x}$. The other two components, $\kappa_{1,y}$ and $\kappa_{1,z}$, do not change. Thus, the wavenumber vector increases from
\begin{equation}
\kappa_1 = \sqrt{\kappa_{1,x}^2+\kappa_{1,y}^2+\kappa_{1,z}^2}
\end{equation}
to
\begin{equation}
\kappa_2 = \sqrt{{\cal C}\kappa_{1,x}^2+\kappa_{1,y}^2+\kappa_{1,z}^2}=\sqrt{({\cal C}-1)\kappa_{1,x}^2+\kappa_1^2}.
\end{equation}
Hence,
\begin{equation}
\label{eq:kratio}
\frac{\kappa_2}{\kappa_1} =\sqrt{({\cal C}-1)\cos^2\psi_1+1},
\end{equation}
where we used the definition $\cos\psi_1=\kappa_{1,x}/\kappa_1$. Since $\kappa = 2\pi/\lambda$, where $\lambda$ is the spatial scale of our eddy, the eddy becomes smaller by factor $\kappa_2/\kappa_1$ as it passes through the shock. At most, $\lambda$ can decrease by a factor of ${\cal C}$, which happens when $\vec{\kappa}_1$ is perpendicular to the shock. When $\vec{\kappa}_1$ is parallel to the shock, there is no change in the size of the eddy.
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{kratio_eps_vs_mach2.pdf}
\caption{The average ratio of downstream and upstream wavenumbers of incident vorticity and/or entropy waves as a function of nuclear dissociation parameter $\bar\epsilon$ for various values of upstream Mach number ${\cal M}_1$. For all the values of $\bar\epsilon$ and ${\cal M}_1$ considered here, the average wavenumber of the upstream turbulent field increases as it crosses the shock, meaning that the spectrum of the turbulent motion shifts towards smaller wavelengths.
\label{fig:kratio}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{eratio_eps_vs_mach_M5_A0.pdf}
\caption{The amplification of turbulent kinetic energy across the shock as a function of nuclear dissociation parameter $\bar\epsilon$ for incident vorticity waves. The black line represents the amplification of the total kinetic energy, while red and blue lines represent the amplifications of angular and radial components of the kinetic energy. The solid lines correspond to anisotropic turbulence represented by relation $R_{rr}=R_{\theta\theta}+R_{\phi\phi}$, while the dashed lines correspond to fully isotropic turbulence represented by relation $R_{rr}=R_{\theta\theta}=R_{\phi\phi}$. \label{fig:eratioA0}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{eratio_eps_vs_mach_M5_A1.pdf}
\caption{The amplification of turbulent kinetic energy across the shock as a function of nuclear dissociation parameter $\bar\epsilon$ for incident entropy vorticity waves of the same phase. The black line represents the amplification of the total kinetic energy, while red and blue lines represent the amplifications of angular and radial components of the kinetic energy. The solid lines correspond to anisotropic turbulence represented by relation $R_{rr}=R_{\theta\theta}+R_{\phi\phi}$, while the dashed lines correspond to fully isotropic turbulence represented by relation $R_{rr}=R_{\theta\theta}=R_{\phi\phi}$. \label{fig:eratioA1}}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{eratio_eps_vs_mach_M5_randomphi.pdf}
\caption{The amplification of turbulent kinetic energy across the shock as a function of nuclear dissociation parameter $\bar\epsilon$ for incident entropy and vorticity waves. The phase difference between the two waves are chosen randomly with a uniform distribution in $(0,2\pi)$. The black line represents the amplification of the total kinetic energy, while red and blue lines represent the amplifications of angular and radial components of the kinetic energy. The solid lines correspond to anisotropic turbulence represented by relation $R_{rr}=R_{\theta\theta}+R_{\phi\phi}$, while the dashed lines correspond to fully isotropic turbulence represented by relation $R_{rr}=R_{\theta\theta}=R_{\phi\phi}$. \label{fig:eratio}}
\end{figure}
In order to obtain an average behavior of $\kappa_2/\kappa_1$, we average it over a random set of vectors $\vec{\kappa}_1$ with uniform distribution on a 2D sphere. This is equivalent to sampling $\cos\psi_1$ uniformly in interval $[0,1]$, which, in turn, is equivalent to solving integral
\begin{equation}
\langle \frac{\kappa_2}{\kappa_1} \rangle = \int_0^1 \sqrt{({\cal C}-1)x+1 } \ dx.
\end{equation}
The latter can be calculated analytically:
\begin{equation}
\langle \frac{\kappa_2}{\kappa_1} \rangle = \frac{2}{3}\frac{1+\sqrt{\cal C}+{\cal C}}{1+\sqrt{\cal C}},
\end{equation}
Figure~\ref{fig:kratio} shows the average ratio $\langle \kappa_2/\kappa_1 \rangle$ as a function of nuclear dissociation parameter $\bar\epsilon$ for four value of upstream Mach number: 2.4, 5, 10, and 100. In all cases, $\langle \kappa_2/\kappa_1 \rangle$ increases mildly with $\bar \epsilon$. This is a simple reflection of the fact that the compression factor increases with $\bar \epsilon$, as we discussed above. For our fiducial values of $\bar\epsilon=0.2$ and ${\cal M}_1=5$, $\langle \kappa_2/\kappa_1 \rangle \simeq 2$. As expected, this result does not change much with further increasing ${\cal M}_1$.
We note that this decrease in the radial extent may be partially compensated and possibly offset by the previous radial stretching experienced by the turbulent fluctuations during the collapse \citep{lai:00n,takahashi:14}.
Now consider the post-shock turbulent kinetic energy. The total specific kinetic energy for a single wave is given by Eq.~(\ref{eq:eturb}), which we rewrite as
\begin{equation}
E' = \frac{1}{2}\left(\langle u'^2 \rangle + \langle \upsilon'^2 \rangle \right),
\end{equation}
where $\langle u'^2 \rangle = \langle u' u'^* \rangle$. For an incident vorticity-entropy wave of form (\ref{eq:T1}), we obtain
\begin{equation}
E'_1 = \frac{1}{2}U_1^2 |A_\upsilon|^2,
\end{equation}
For the downstream vorticity field (\ref{eq:u2})-(\ref{eq:v2}), in the far-field region ($x \gg1/\kappa$), we have\footnote{Note that Eq.~(\ref{eq:eturb2}) does not include the contribution from acoustic waves. In order to include that, have to add $\frac{1}{2}U_1^2 \left(|F|^2+|H|^2 \right)$ to the right-hand side of Eq.~(\ref{eq:eturb2}) in the propagative regime ($\psi_1<\psi_\mathrm{c}$). In the non--propagative regime, there is no contribution from the acoustic component.}
\begin{equation}
\label{eq:eturb2}
E'_2 =\frac{1}{2} U_1^2 \left(|G|^2+|I|^2 \right),
\end{equation}
Thus, the ratio of upstream and downstream turbulent kinetic energies are
\begin{equation}
\frac{E'_2}{E'_1} = |\tilde G|^2+|\tilde I|^2 ,
\label{eq:eturbratio}
\end{equation}
where $\tilde G=G/A_\upsilon$ and $\tilde I=I/A_\upsilon$.
Note that formula (\ref{eq:eturbratio}) depends only on $A_e/A_\upsilon$ and the incidence angle $\psi_1$ of upstream vorticity-entropy waves, but not on their wavenumbers $\kappa_1$. This is an important result because it means that the amplification factor of turbulent kinetic energy across the shock is independent of the spectrum of upstream turbulence.
The black line in Fig.~\ref{fig:eratioA0} shows the amplification of the total kinetic energy across the shock as a function of the nuclear dissociation parameter $\bar \epsilon$ for incident vorticity waves. Here, we use an anisotropic turbulent field represented by relation (\ref{eq:rrr}) and each point on this graph is calculated using a sample of 150,000 random incident waves. The amplification of the total energy $E'_2/E'_1$ does not change much with $\bar\epsilon$, remaining at $\simeq\! 2.14$ as $\bar\epsilon$ grows from $0$ to $0.4$. On the other hand, the amplification of the angular and radial components, defined as $(E'_{y,2}+E'_{z,2})/(E'_{y,1}+E'_{z,1})$ and $E'_{x,2}/E'_{x,1}$, exhibit noticeable dependence on $\bar\epsilon$. The angular component, shown with the red line in Fig.~\ref{fig:eratioA0}, increases from $2.85$ to $3.15$ as $\bar\epsilon$ grows from $0$ to $0.4$. Contrary to this, the amplification of the radial component, shown with the blue line in Fig.~\ref{fig:eratioA0}, decreases from $1.30$ to $0.90$ for the same values of $\bar\epsilon$.
Similar to the behavior of individual waves discussed earlier in Section~\ref{sec:wave}, the change of kinetic energy of the incident vorticity waves across the shock is very sensitive to the presence of incident entropy waves. If we add entropy waves with the same phase and amplitude as the incident vorticity waves (i.e., $A_e=A_\upsilon$), the amplification of total kinetic energy of turbulent field becomes $\simeq\! 3.95$ (cf. the dashed black line in Fig.~\ref{fig:eratioA1}). This is $\simeq\! 1.85$ times larger than what we get in the case of pure vorticity waves shown in Fig.~\ref{fig:eratioA0}. On the other hand, if they are out of phase (i.e., $A_e=A_\upsilon e^{i\pi}$), we find that the total energy does not change much and $E'_2/E'_1 \sim 1$ (not shown here). Such a dependence on entropy waves is a direct manifestation of the simple scaling law (\ref{eq:omega2_scaling}) discussed above.
We also consider the case when the phase difference between the incident entropy and vorticity waves are chosen randomly with uniform distribution between $0$ and $2\pi$. This case is presented in Fig.~\ref{fig:eratio}. We find that in this case, the overall behavior of the turbulent kinetic energy is qualitatively and quantitatively similar to that in the case of incident pure vorticity waves shown in Fig.~\ref{fig:eratioA0}. For fiducial parameters, ${\cal M}_1=5$ and $\bar\epsilon=0.2$, we get $E'_2/E'_1=2.46$, $(E'_{y,2}+E'_{z,2})/(E'_{y,1}+E'_{z,1}) = 3.38$, and $E'_{x,2}/E'_{x,1}=1.32$.
We can summarize these findings as follows. If the phases of incident entropy and vorticity are strongly correlated, then the total kinetic energy of the turbulent field will increase by a factor of $\sim\! 4$. If they are strongly anti-correlated, then there is no amplification. If there is no correlation in the phases, then the amplification is $\sim\! 2$.
In order to test the sensitivity of our results to the particular form of (\ref{eq:rrr}), we repeat this exercise for isotropic turbulence represented by Eq.~(\ref{eq:rrr2}). The dashed black lines in Figs.~\ref{fig:eratioA0} and \ref{fig:eratioA1} show the amplification of the total turbulent kinetic energy of the field of incident vorticity and in-phase entropy-vorticity waves, respectively. In both cases, the amplification is again insensitive to $\bar\epsilon$, remaining at $\simeq\! 1.8$ and $\simeq\! 3.1$ for incident vorticity and in-phase entropy-vorticity waves. These values are $\simeq\! 15\%$ and $\simeq\! 21\%$ smaller than those in the case of anisotropic turbulence. Despite similarity of the of the behavior of the total energy across the shock, we see differences in the behavior of radial and angular components.
For isotropic turbulence, the amplification factors of the angular and radial components are closer to each other than those for the anisotropic turbulence. For example, for $\bar\epsilon=0.2$ and incident vorticity wave, the ratios of the amplifications of angular and radial components are $2.67$ and $2.16$ for anisotropic and isotropic turbulence models. The reason for larger amplification of angular component in anisotropic turbulence is due fact that a smaller fraction of the kinetic energy is contained in a component tangential to the shock, which does not undergo amplification.
Our analysis shows that, downstream of the shock, acoustic waves contribute at most $\sim 2\%$ of the total turbulent kinetic energy, which is a tiny amount. This may seem surprising in the
light of the fact that the ratio of the kinetic energy of sound waves to
the kinetic energy of the total fluctuating velocity field can reach
$\sim 0.08$ for $\psi \sim 50^\circ$, as we saw in
Fig.~\ref{fig:q2q1_ac_ratio}. However, the total kinetic energy of the
fluctuating field in this region is small compared to that at larger
$\psi$. This is easily visible in Fig~\ref{fig:q2q1_ac_phi}, which shows, for incident vorticity waves,
the ratio of upstream and downstream kinetic energies with and without
the contribution of the downstream acoustic field with solid and dashed
lines, respectively. As we can see, if we average over all values of
$\psi_1$, the contribution of the acoustic component should be
negligibly small, in agreement with our findings above.
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{q2q1_vs_psi_vs_eps_accont.pdf}
\caption{The ratios of upstream and downstream kinetic energies of turbulent field as a function of the incidence angle $\psi_1$ for ${\cal M}_1=5$ for incident vorticity waves. The solid lines include the contribution of sound waves to the kinetic energy, while the dashed lines do not.
\label{fig:q2q1_ac_phi}}
\end{figure}
The shock surface responds to upstream velocity perturbations by oscillating according to formula (\ref{eq:shock_lia0}). Fig.~\ref{fig:xit} shows the normalized RMS velocity of the shock $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ as a function of nuclear dissociation parameter $\bar\epsilon$ for ${\cal M}_1=5$. Here, $\langle u_1'^2 \rangle$ is the RMS value of the $x$-component of the perturbation velocity. The black line corresponds to incident vorticity waves, while the red and blue lines correspond to incident entropy and vorticity waves with the same phase (i.e., $A_e=A_\upsilon$) and $180^\circ$ phase difference ($A_e=A_\upsilon e^{i\pi}$), respectively. Finally, the green line represents the case where the entropy and vorticity waves have randomly sampled phase differences from $0$ to $2\pi$ with uniform distribution (i.e., $A_e=A_\upsilon e^{i\pi r}$ where $r$ is a random number with uniform distribution in $[0,2]$). Similar to the amplification of turbulent kinetic energy, the shock velocity does not change much with $\bar\epsilon$, but it is very sensitive to the presence of entropy waves. For our fiducial value $\bar\epsilon=0.2$, we get the largest $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ of $\sim 0.8$ for $A_e=A_\upsilon$, while for case $A_e=A_\upsilon e^{i\pi}$, we get the smallest $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ of $\sim 0.2$. In the case of incident entropy-vorticity waves with randomly distributed phase differences and in the case of incident vorticity waves, we get similar values of $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ $\sim 0.58$ and $\sim 0.47$, respectively. Note that, due to the employed normalization, these values do not depend on weather we use anisotropic (Eq. \ref{eq:rrr}) or isotropic (Eq. \ref{eq:rrr2}) turbulence prescriptions.
\begin{figure}
\includegraphics[angle=0,width=1\columnwidth, clip=false]{xit_vs_eps_vs_phi.pdf}
\caption{The normalized RMS velocity of the shock oscillations $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ as a function of nuclear dissociation parameter $\bar\epsilon$ for ${\cal M}_1=5$. The black line corresponds to incident vorticity waves, while the red and blue lines correspond to incident entropy and vorticity waves with the same phase (i.e., $A_e=A_\upsilon$) and $180^\circ$ phase difference ($A_e=A_\upsilon e^{i\pi}$), respectively. Finally, the green line represents the case where the entropy and vorticity waves have randomly sampled phase differences from $0$ to $2\pi$ with uniform distribution (i.e., $A_e=A_\upsilon e^{i\pi r}$ where $r \in [0,2]$ is random number). We get the strongest velocities $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ of $\sim 0.8$ in the case of in-phase entropy-vorticity waves. When they are out of phase, we get the weakest amplification $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle} \sim 0.2$. In the case of randomly distributed phase differences and in the case of incident vorticity waves, we get similar values of $\sqrt{\langle \xi_t^2 \rangle / \langle u_1'^2 \rangle}$ of $\sim 0.58$ and $\sim 0.47$, respectively.
\label{fig:xit}}
\end{figure}
\section{Implications for the Explosion Condition}
\label{sec:implications}
We now discuss the implications of the above results on the conditions for producing explosion using the concept of critical luminosity \citep{burrows:93}. According to \citet{mueller:15}, in the presence of post-shock turbulence, the critical luminosity for producing explosion is
\begin{equation}
\label{eq:critlum}
L_\nu E_\nu^2 \propto (\dot M M)^{3/5}r_\mathrm{gain}^{-2/5}\left(1+\frac{4\langle {\cal M'}_2^2\rangle}{3}\right)^{-3/5},
\end{equation}
where $\langle {\cal M'}_2^2 \rangle$ is the RMS post-shock turbulent Mach number. Following \citet{mueller:15}, we define it as
\begin{equation}
\langle {\cal M'}^2 \rangle = \frac{\langle \upsilon_\mathrm{a}^2 \rangle}{\langle c_\mathrm{s}^2 \rangle},
\end{equation}
where $\langle \upsilon_\mathrm{a}^2 \rangle$ is RMS angular velocity, which can be expressed in terms of specific kinetic energy of angular turbulent motion as $\langle \upsilon_\mathrm{a}^2 \rangle = 2E'_\mathrm{a}$. Using this, we can write
\begin{equation}
\langle {\cal M'}_2^2 \rangle = \frac{2E'_\mathrm{a,2}}{\langle c_\mathrm{s}^2 \rangle} = 2 \frac{E'_\mathrm{a,2}}{E'_\mathrm{a,1}}\frac{E'_\mathrm{a,1}}{\langle c_\mathrm{s}^2 \rangle} = \frac{E'_\mathrm{a,2}}{E'_\mathrm{a,1}}\frac{\langle \upsilon_\mathrm{a,1}^2 \rangle}{\langle c_\mathrm{s,2}^2 \rangle} = \frac{E'_\mathrm{a,2}}{E'_\mathrm{a,1}} \frac{\langle c_\mathrm{s,1}^2 \rangle}{\langle c_\mathrm{s,2}^2 \rangle} \langle {\cal M'}_1^2 \rangle
\end{equation}
Substituting this into Eq.~(\ref{eq:critlum}), we get
\begin{equation}
L_\nu E_\nu^2 \propto \left(1+\frac{4}{3}
\frac{E'_\mathrm{a,2}}{E'_\mathrm{a,1}} \frac{\langle c_\mathrm{s,1}^2 \rangle}{\langle c_\mathrm{s,2}^2 \rangle} \langle {\cal M'}_1^2 \rangle \right)^{-3/5} \simeq 1-\frac{4}{5}
\frac{E'_\mathrm{a,2}}{E'_\mathrm{a,1}} \frac{\langle c_\mathrm{s,1}^2 \rangle}{\langle c_\mathrm{s,2}^2 \rangle} \langle {\cal M'}_1^2 \rangle,
\end{equation}
Note that we linearized in $\langle {\cal M'}_1^2 \rangle$ in the last step. Subtracting this from the critical luminosity in the absence of post-shock turbulence, we obtain an expression for the relative reduction of the critical luminosity due to upstream turbulence:
\begin{equation}
\label{eq:critlum1}
\delta (L_\nu E_\nu^2) \simeq \frac{4}{5}
\frac{E'_\mathrm{a,2}}{E'_\mathrm{a,1}} \frac{\langle c_\mathrm{s,1}^2 \rangle}{\langle c_\mathrm{s,2}^2 \rangle} \langle {\cal M'}_1^2 \rangle,
\end{equation}
For our fiducial parameters $\bar\epsilon=0.2$ and ${\cal M}_1=5$, $\langle c_\mathrm{s,1}^2 \rangle / \langle c_\mathrm{s,2}^2 \rangle \simeq 0.25$ and $E'_\mathrm{a,2}/E'_\mathrm{a,1} \sim 3$ for anisotropic turbulence represented by relation (\ref{eq:rrr}) for an incident field of vorticity or entropy-vorticity waves with uncorrelated phases. For these values, Eq.~(\ref{eq:critlum1}) reduces to
\begin{equation}
\label{eq:critlum2}
\delta (L_\nu E_\nu^2) \simeq 0.6 \langle {\cal M'}_1^2 \rangle,
\end{equation}
Thus, the critical luminosity decreases by $\simeq\!0.6 \langle {\cal M'}_1^2 \rangle$ compared to the case with no post-shock turbulence.
In convective shells, we expect $\sqrt{\langle {\cal M'}^2 \rangle} \sim\!0.1$ \citep[e.g.,][]{mueller:16b}. During collapse, the Mach number of non-radial fluctuations increases as $\propto r^{(3\gamma-7)/4}$ \citep{lai:00n}. Assuming that convective shells fall from a radius of $\sim\!1500\,\mathrm{km}$ to $\sim\!200\,\mathrm{km}$ before it hits the shock, in the absence of turbulent dissipation, the turbulent Mach number should increase to $\sim 0.45$ before hitting the shock, which yields $\langle {\cal M'}_1^2\rangle\sim\!0.21$. This results in a reduction of the critical luminosity by $\sim\!12\%$ compared to the case with no upstream turbulence.
Note that the estimate (\ref{eq:critlum2}) is of limited accuracy for a number of reasons. First, it neglects turbulent dissipation in the post-shock region. Second, it is based on a comparison to the hypothetical case with no post-shock turbulence. However, by the time a nuclear-burning shell hits the shock, the post-shock region is expected to have a fully-developed neutrino-driven turbulent convection \citep{couch:15b}, which we cannot include in our estimate. Both of these effects overestimate the reduction of critical luminosity. Therefore, the above estimate is expected to yield an approximate upper limit for the reduction of the critical luminosity.
\section{Conclusion}
\label{sec:summary}
In this paper, we studied the interaction of the shock waves in
core-collapse supernovae (CCSNe) with turbulent convection arising from
nuclear shell burning. We used a first-order perturbation theory called
the linear interaction approximation (LIA), which we extended to include
nuclear dissociation at the shock. In the LIA, the shock wave
is modeled as a planar discontinuity with no intrinsic scale. The
upstream flow, which consists of mean and fluctuating part, fully
determines the downstream flow via the Rankine-Hugoniot conditions at the
shock. In the LIA, turbulent field is decomposed into
individual Fourier modes. Each mode interacts independently with the
shock. Integration over all modes yields the full statistics of the
turbulent flow (cf. Section~\ref{sec:lia}).
In order to approximate the situation in CCSNe, we required the mean flow to have the vanishing Bernoulli parameter in the pre-shock region. We considered two types of upstream incident perturbations: the vorticity and entropy waves, both of which are advected with the mean flow. The vorticity mode is a solenoidal velocity field that has no pressure or density fluctuations, while the entropy mode represents density and temperature fluctuations with no associated pressure or velocity variations. Once the incident perturbations hit the shock, the downstream fluctuation field consists of vorticity, entropy, and acoustic waves (cf. Section~\ref{sec:lia}).
The compression factor $\cal C$ at the shock is the key quantity that
affects the flow through the shock. In particular, it determines
by how much the $x$-component of the wavenumber of incident waves increase as they cross the shock. Nuclear
dissociation leads to stronger compression: $\cal C$ increases from
$5.56$ to $10.15$ as nuclear dissociation parameter $\bar\epsilon$
increases from $0$ (inefficient nuclear dissociation) to $0.4$
(efficient nuclear dissociation). The compression factor $\cal C$ grows
fast with the upstream Mach number ${\cal M}_1$ until ${\cal M}_1 \sim\!
5$, after which it does not change much with further increasing of
${\cal M}_1$. We find most of the quantities that characterize the
downstream perturbation field have a similar dependence on ${\cal M}_1$
for ${\cal M}_1 \gtrsim 5$ (cf. Section~\ref{sec:results}).
The critical angle $\psi_c$ separates two regions of the solution: $\psi_1 < \psi_c$ and $\psi_1 > \psi_c$, where $\psi_1$ is the incidence angle. The $\psi_1 < \psi_c$ region is called the propagative regime and it is characterized by acoustic waves in the post-shock flow, while $\psi_1 > \psi_c$ is called the non-propagative region, in which sound waves do not propagate. We investigated how $\psi_c$ depends on the free parameters of the mean flow: the upstream Mach number ${\cal M}_1$ and the efficiency of nuclear dissociation (cf. Section~\ref{sec:lia}). We find that $\psi_c$ depends weakly on both of these parameters. For our fiducial parameters, we get $\psi_c=69.2^\circ$ (cf. Section~\ref{sec:results}).
We explored how individual vorticity and entropy waves affect the shock and the downstream flow (Section~\ref{sec:wave}). In particular, we analyzed the amplification of the kinetic energy of individual incident waves as they cross the shock. The amplification of kinetic energy does not change much with $\bar\epsilon$ and ${\cal M}_1$ for ${\cal M}_1\! \gtrsim\! 5$. On the other hand, it is highly sensitive to the relative phase between the entropy and vorticity waves: when they are in phase, we get the strongest amplification, while when they are out of phase, the kinetic energy does not amplify much. In fact, it may even decrease for some values of $\psi_1$. For example, for our fiducial parameters, ${\cal M}_1=5$ and $\bar\epsilon=5$, we get the amplification factors of $4.51$ and $0.90$ for in-phase and out-of-phase entropy-vorticity waves for $\psi_1=60^\circ$. The amplification for incident vorticity waves is roughly the average of these two regimes. For example, for the same values of $\bar\epsilon$, ${\cal M}_1$, and $\psi_1$, we get an amplification of $2.36$ (cf. Section~\ref{sec:wave}).
For an incident field of turbulent fluctuations, we calculated the amplification of total turbulent kinetic energy. We find that the amplification is not sensitive to the nuclear dissociation parameter and the upstream Mach number beyond ${\cal M}_1 \gtrsim\! 5$. We again observe strong dependence on the phase difference between the incident vorticity and entropy waves. When they are in phase, the total kinetic energy increases by a factor of $\sim\!4$, while when they are out of phase, there is almost no amplification. When the phase is randomly distributed, the amplification is $\sim\!2$. When there is only incident vorticity wave perturbations, the amplification is again $\sim\!2$ (cf. Section~\ref{sec:turbulence}).
When a turbulent eddy crosses the shock, it shrinks in size due to shock
compression. We find that for our fiducial values, the average linear
size of a turbulent eddy shrinks by a factor of $\simeq 2$ (cf.
Section~\ref{sec:turbulence}). This values does not change much with the
upstream Mach number ${\cal M}_1$ and the nuclear dissociation
parameter. This is somewhat disappointing news from the point of
producing explosion because smaller eddies are perhaps less likely to
become buoyant and help explosion \citep[e.g.,][]{couch:13b}.
However, this effect may be partially or completely offset by
the fact that perturbations in the nuclear-burning shells experience
significant radial stretching before reaching the shock
\citep{lai:00n,takahashi:14}.
When a turbulent field crosses the shock, the post-shock turbulence exerts additional pressure behind the shock. This reduces the critical neutrino luminosity necessary to drive the explosion \citep{mueller:15}. We find that, compared to the case with no post-shock turbulence, the critical luminosity decreases by a factor of $\simeq\!0.6 \langle {\cal M}_1'^2\rangle$, where $\langle {\cal M}_1'^2\rangle$ is the RMS turbulent Mach number in the pre-shock region. If the turbulent Mach number in convective shells is $\sim\! 0.1$, it may increase to $\langle {\cal M}_1'^2\rangle \sim\! 0.21$ during collapse prior to hitting the shock. This results in $\lesssim\!12\%$ reduction in the critical luminosity (cf. Section~\ref{sec:implications}).
\section*{Acknowledgements}
We thank Bernhard M\"uller for valuable discussions and comments. This
work is partially supported by ORAU and Social Policy grants at
Nazarbayev University and by the Sherman Fairchild Foundation.
|
\section{Introduction}
The
understanding of the effects and mechanisms by which quarks
interact among themselves
is a necessary step to provide a complete description of
hadron structure and dynamics and the phase diagram
of Strong Interactions.
In low and intermediary energies these interactions can be
parametrized in terms of realistic effective quark interactions
that usually
provide important information to establish
the needed relations
between QCD and
hadron dynamics
\cite{brambilla-etal,greensite}.
The basic and fundamental mechanisms that give rise to
each of the effective interactions and parameters present
in
effective models and theories
should be expected to be well
understood, although
a quite large amount of different quark effective interactions are expected to emerge
due to the intrincated structure of QCD.
The Nambu Jona Lasinio (NJL) model is known to describe
qualitatively well several important
effects in hadron phenomenology \cite{NJL1,NJL} in spite of its known
limitations.
A large variety of possible corrections to the NJL coupling
can be expected to emerge from
QCD, and higher order quark interactions were shown to provide relevant effects
for the ground state
\cite{8th,osipov,andrionov-andrionov},
chiral phase transition
(flavor SU(2) and SU(3)) and higher energies
\cite{phtrans-su3,ho-su2,ho-su2-PL-eNJL,ho-su2-scd,osipov-ph-trans}
and eventually they might contribute to multiquark structures \cite{tetraquarks}.
In FAIR-GSI the high density phase diagram will be tested eventually
providing relevant information also about the role of multiquark interactions
in different regions of the phase diagram.
Few
mechanisms have been shown
to drive quark effective interactions by gluon exchange
\cite{thooft-det,creutz,NJL-derivations,kondo -2010,simonov-plb,qqq-instantons,simonov-prd,wetterich-etal,PRD-2014,vanderbossche,kashiwa-etal}.
Instanton mediation have been shown to provide one
of the most investigated mechanisms for effective quark interactions
for example
by means of the Kobayashi-Maskawa-'t Hooft interaction or instanton gas model.
It depends strongly on flavor and, for flavor SU(2), it yields
a second order quark interaction different from the usual chiral NJL interaction,
producing the axial anomaly and its phenomenological consequences
\cite{thooft-det,KKM,NJL1,NJL}.
Polarization effects were shown to
produce low energy and higher order
effective interactions \cite{PRD-2014}.
In the present work, flavor SU(2) higher order
quark effective interactions
are calculated from polarization effects
by departing from
a dressed one gluon exchange
(i.e. a global color model)
along the lines of Refs. \cite{PRD-2014}.
Simple gluon exchange is a basic mechanism that cannot describe low energy
hadron properties, including dynamical breakdown of
chiral symmetry (D$\chi$SB), although it can be dressed by
gluon interactions producing enough strength for
D$\chi$SB
\cite{cornwall-2011,kondo,SD1}.
This work is organized as follows.
In the next section the method is shortly described
according to which the quark bilinears are separated into two
components,
i.e. $\bar{\psi} \Gamma \psi \to (\bar{\psi} \Gamma \psi)_1 + (\bar{\psi} \Gamma \psi)_2$,
as done in the background field method \cite{background}.
The background field ($\psi_1$) remains as interacting
quarks and the field $\psi_2$ is integrated out.
Instead of introducing auxilary fields (a.f.) for the component that is integrated out,
a weak field approximation is considered such that:
$(\bar{\psi} \psi)_1^2 >> (\bar{\psi} \psi)_2^2$.
Results are the same as by introducing a.f. in the leading order since the a.f.,
for example as shown in Ref. \cite{PRD-2014,PRC,ERV},
play no role in the resulting leading
quark-quark effective interactions.
The quark determinant is expanded in powers
of quark bilinears yielding chiral invariant
and also symmetry breaking terms proportional to the Lagrangian quark mass.
The corresponding effective couplings are resolved.
This expansion is performed up to the eighth order for all the bilinears
and up to the tenth order for the scalar-pseudoscalar ones.
Some ratios between the effective coupling constant
are shown to provide simple numerical values.
Some numerical estimations are also shown.
\section{Diquark interaction and quark field splitting}
The departing point is the following
quark effective interaction:
\begin{eqnarray} \label{Seff}
S_{eff} [\bar{\psi}, \psi] &=&
\int_x \left[
\bar{\psi} \left( i \slashed{\partial}
- m \right) \psi
-
\frac{g^2}{2}\int_y j_{\mu}^b (x)
{\tilde{R}}^{\mu \nu}_{bc} (x-y) j_{\nu}^{c} (y)
\right]
,
\end{eqnarray}
Where $b,c$ stand for color indices,
the color quark current is
$j^{\mu}_b = \bar{\psi} \lambda_b \gamma^{\mu} \psi$,
the sum in color, flavor and Dirac indices are implicit, $\int_x$
stands for
$\int d^4 x$,
the kernel ${\tilde{R}}^{\mu \nu}_{bc}$
can be written in terms of
transversal and longitudinal components
($R_T$ and $R_L$) as:
$\tilde{R}^{\mu\nu}_{ab} \equiv \tilde{R}^{\mu\nu}_{ab} (x-y) = \delta_{ab} \left[
R_T \left( g^{\mu\nu} - \frac{\partial^\mu \partial^\nu}{\partial^2}
\right)
+ R_L \frac{\partial^\mu \partial^\nu}{\partial^2} \right]
$
with implicit Dirac delta functions $\delta(x-y)$.
With a Fierz transformation \cite{NJL1,NJL,PRC,ERV},
by picking up the color singlet sector only,
the above effective quark interaction can be expressed
in terms of
bilocal quark bilinears,
$j_i^q(x,y) = \bar{\psi} (x) \Gamma^q \psi (y)$ where
$q=s,p,v,a$
and $\Gamma_q$ stands for Dirac and flavor SU(2)
operators
$\Gamma_{s} = I$ for the 2x2 flavor and 4x4 identities,
$\Gamma_{p} = \sigma_i i \gamma_5$,
$\Gamma_{v}^\mu = \gamma^\mu \sigma_i $ and
$\Gamma_{a}^\mu = i \gamma_5 \gamma^\mu \sigma_i $,
being $\sigma_i$ are the flavor SU(2) Pauli matrices.
The Fierz transformed interaction is written as:
$\Omega = \alpha \sum_q j_i^q(x,y) R_q (x-y) j_i^q(y,x)$, where
$\alpha=8/9$,
$R_q$ are the kernels in each of the $q$ channel of the interaction.
Next the quark field is separated into two components,
one of them associated with polarization virtual processes
eventually to the formation of quark bound states such as light mesons
and the chiral condensate and the other component remains as (constituent) quark.
This procedure is basically the one loop background field method \cite{background},
and this will be done by rewritting the quark bilinears above as:
\begin{eqnarray} \label{split-Q}
\bar{\psi} \Gamma^q \psi \to (\bar{\psi} \Gamma^q \psi)_2
+ (\bar{\psi} \Gamma^q \psi)_1.
\end{eqnarray}
The Fierz transformed non local interaction above can then be written as:
$\Omega \to \Omega_1 + \Omega_2 + \Omega_{12}$
where $\Omega_{1}$ and $\Omega_2$ stand for the interactions
of each of the quark components,
and
$\Omega_{12}$ for the mixed terms.
The component $\psi_2$ will be integrated out
and the fourth order terms can be eliminated in different approximated ways.
Firstly by simply considering a weak field approximation and therefore
by neglecting $\Omega_2 << \Omega_1$.
This yields the same results
as the leading terms resulting from the
auxiliary field method which eliminates the fourth order interactions
$\Omega_2$,
as discussed in Refs. \cite{PRD-2014,PRC,ERV}.
In this case, bilocal
auxiliary fields ($S,P_i,V_\mu^i, \bar{A}_\mu^i$) are introduced
which couple to the remaining
quark component.
These couplings
encode the non linearities of the
initial model.
However in this work we are interested only in the quark self interactions
and these couplings can be neglected.
Even if one were interested in
the effective interactions induced by these couplings to the auxiliary fields (a.f.),
the resulting
quark-quark effective
interactions induced by the a.f.
would be of higher order and
numerically smaller.
By integrating out the
component $(\psi)_2$,
and by writing the determinant as:
$\det (A) = \exp \left( Tr \ln A \right)$,
the following non linear non local effective action
for quarks $(\psi)_1$ is obtained:
\begin{eqnarray} \label{Seff-q4}
S_{eff} &=& - i \; Tr \ln \left\{ i
({S_0})^{-1} (x-y)
\right.
\nonumber
\\
&+&
\left.
-
i \alpha g^2 \bar{R}^{\mu\nu} (x-y) \gamma_\mu \sigma_i
\left[
(\bar{\psi}_y \gamma_\nu \sigma_i \psi_x)
- i \gamma_5 (\bar{\psi}_y
i \gamma_5 \gamma_\nu \sigma_i \psi_x)
\right]
\right.
\nonumber
\\
&+&
\left.
2 i \alpha g^2 R(x-y)
\left[ (\bar{\psi}_y \psi_x)
+ i \gamma_5 \sigma_i (\bar{\psi}_y i \gamma_5 \sigma_i \psi_x) \right]
\right\}
- I_0
,
\end{eqnarray}
where
$Tr$ stands for traces of discrete internal quantum numbers indices
and integration of spacetime coordinates/momentum
and
\\
$I_0 = \int_x
\left[
\bar{\psi} \left( i \gamma \cdot \partial
- m \right)\psi
- \frac{g^2 }{2} \int_y j_{\mu}^{a}(x) R^{\mu\nu}_{ab} (x-y) j_{\nu}^{b}(y)
\right]$.
In this expression the label $_1$ for the quark field was omitted
because it is the only one remaining from here on.
$({S_0})^{-1} = ({S_0})^{-1}(x-y) \equiv ( i \gamma \cdot { \partial}
- m )$, with an implicit Dirac delta function,
and where instead of $m$ one could introduce an effective mass ($m^*$)
which arise from the coupling to the scalar auxiliary variable $s$
which produces
the dynamical chiral symmetry breaking as discussed at length in Refs.
\cite{NJL1,NJL,PRD-2014,PRC,ERV}.
The following kernels have also been defined from the Fierz transformation:
$ R = R(x-y) =
3 R_T + R_L$ and
$\bar{R}^{\mu\nu} =\bar{R}^{\mu\nu}(x-y)
= g^{\mu\nu} (R_T+R_L) +
2 \frac{\partial^{\mu} \partial^{\nu}}{\partial^2} (R_T - R_L)$
with implicit Dirac delta functions.
By neglecting the derivative couplings,
with a shorthand notation for which the non local character of all the kernels is omitted, i.e
$R=R(x-y)$, $\bar{R}^{\mu\nu}=\bar{R}^{\mu\nu}(x-y)$ and $S_0=S_0(x-y)$,
the quark determinant above can be rewritten \cite{mosel} as:
\begin{eqnarray} \label{Ideterm}
I_d &\equiv &
- \frac{i}{2}
Tr \ln \left[ S^{-1} {S^\dagger}^{-1} \right]
= - \frac{i}{2} Tr \ln [ \tilde{S}_0^{-1} ]
\\
&-& \frac{i}{2} Tr \ln \left[1 + \beta \tilde{S}_0
\left( 2 R \bar{\psi} \psi
- \bar{R}^{\mu\nu} \gamma_\mu \sigma_i \bar{\psi} \gamma_\nu \sigma_i \psi \right)
+ g^4 \sum_{q,q'} \tilde{S}_0 a_{q,q'}
(\Gamma_q \bar{\psi} \Gamma_q \psi )
( \Gamma_{q'}^\dagger
\bar{\psi} \Gamma_{q'} \psi )
\right]
,
\nonumber
\end{eqnarray}
where
$\beta= 2 m g^2 \alpha$ was defined for the quark mass (symmetry breaking term),
$\tilde{S}_0 \equiv \tilde{S}_0 (x-y) = - 1/ (\partial^2 + m^2) \delta (x-y)$
was factorized producing an irrelevant
multiplicative
constant in the generating functional,
$a_{q,q'}$ are coefficients
for each of the flavor channels,
and crossed terms ($q,q'=s,p,v,a$)
with the corresponding operators $\Gamma_q$
and kernels $R_q$.
This expression still has a strong non local character which is not written explicitely.
This determinant
will be expanded for small $\tilde{S}_0$, i.e. large quark (effective) mass
by considering that $m$ may be an effective (constituent) quark mass.
A small coupling $g^2$ or weak quark field $\psi_1$
yields essentially the same results
such that
the final polynomial quark effective interactions are written in terms of effective coupling constants
in the local limit of the resulting couplings.
It can be noticed that all the chiral invariant interactions
only appear from the contributions exclusively of the last term inside of the
determinant.
Therefore chiral invariant terms for this
$SU(2)$ flavor will be $ {\cal O} [(\bar{\psi} \psi)^2 ]^n$.
All the interactions for which the second term contributes
(proportional to
the quark mass) will be not chiral invariant.
One of the
first order terms yields a contribution for the quark effective mass \cite{PRD-2014}
of the form: $\Delta m^* = - i 2 \alpha g^2 m \; Tr \; \tilde{S}_0 R$.
\section{ SU(2) quark effective interactions }
\label{sec:eff-param}
The leading terms,
by resolving the effective coupling constants in the longwavelength limit
and
the zero order derivative expansion, are:
\begin{eqnarray} \label{4quark}
{\cal L}_{4} =
g_4
\left[ ( \bar{\psi} \psi )^2
+ ( \bar{\psi} \sigma_i i \gamma_5 \psi )^2 \right]
-
g_{v4}
\left[ ( \bar{\psi} \sigma_i \gamma_\mu \psi )^2
+ ( \bar{\psi} \sigma_i \gamma_5 \gamma_\mu \psi )^2
\right]
+ {\cal L}_4^{sb}
\end{eqnarray}
where ${\cal L}_4^{sb} = g_{4,sb}
( \bar{\psi} \psi )^2
+
g_{4,v,sb} ( \bar{\psi} \sigma_i \gamma_\mu \psi )^2$
are symmetry breaking terms which emerge from
the second order expansion although they are of the same order of magnitude as
the first one, as it can be noted in the next expressions.
These effective coupling constants
were resolved as:
\begin{eqnarray} \label{g4}
g_{4} \; (1 \; ; \; \delta_{i_j}) &=&
- i 2 (g^2 {\alpha})^2 N_c\; Tr'' \; {\tilde{S}_{0}} R^2 \;
(1 \; ; \; \sigma_i \sigma_j ),
\\ \label{g4sb}
g_{4,sb} &=&
i 4 (g^2 {\alpha})^2 N_c \; Tr'' \; {m}^2 ({\tilde{S}_{0}} R)^2 ,
\\
g_{v4} \; \delta_{ij} g^{\mu\nu}
&=&
- \frac{i}{2} (g^2 \alpha)^2 N_c
\; Tr'' \; {\tilde{S}_{0}} \bar{R}^{\mu\rho}
\bar{R}^{\nu\sigma}
(\sigma_i \sigma_j)
\gamma_\rho \gamma_\sigma
\\ \label{g4vsb}
g_{4,v,sb} \; \delta_{ij} g^{\mu\nu}
&=&
i (g^2 \alpha)^2 N_c
\; Tr'' \; {m}^2 {\tilde{S}_{0}} \bar{R}^{\mu\rho}
{\tilde{S}_{0}} \bar{R}^{\nu\sigma}
(\sigma_i \sigma_j)
\gamma_\rho \gamma_\sigma ,
\end{eqnarray}
where
where $Tr''$ includes all the traces in internal and spacetime indices except
the trace in color indices that has already been done.
The couplings with $g_4$ and $g_{v4}$ are
the usual NJL and vector NJL couplings respectively
with dimension $1/M^2$ for a mass scale $M$.
For the class of diagrams of this one fermion loop level,
by considering that
$g^2 \sim \tilde{g}^2/N_c$,
the resulting $n$-quark coupling constants
are
of the order of $N_c^{1-n}$ in agreement with
\cite{witten}.
The non derivative sixth order terms, after resolving the effective coupling
constants, are all symmetry breaking and they were found to be:
\begin{eqnarray} \label{6th-1}
{\cal L}^{(6)}
&=&
g_{6,sb}^{(1)} (\bar{\psi} \psi )
\left[ (\bar{\psi} \psi )^2
+ (\bar{\psi} \sigma_i i \gamma_5 \psi)^2 \right]
- g_{6,sb,a} \epsilon_{ijk}
(\bar{\psi} \sigma_i i\gamma_5 \gamma_\mu \psi) (\bar{\psi} \sigma_j \gamma^\mu \psi)
( \bar{\psi} \sigma_k i \gamma_5 \psi )
\nonumber
\\
&-&
g_{6,sb,a}
\left[
( \bar{\psi} \sigma_i \gamma_\mu \psi )^2 +
( \bar{\psi} \sigma_i i \gamma_5 \gamma_\mu \psi )^2
\right]
(\bar{\psi} \psi )
+ g_{6,sb}^{(3)} (\bar{\psi} \psi)^3
\end{eqnarray}
where
\begin{eqnarray} \label{g6-all}
g_{6,sb}^{(1)} \; (1 \; ; \; \delta_{ij}) &=&
i 2 (\alpha g^2)^3 N_c
\; Tr'' \; {m} {\tilde{S}_{0}} \; R ({\tilde{S}_{0}} \; R^2) \;
(1 \; ; \; \gamma^2_5 \sigma_i \sigma_j)
,
\\
g_{6,sb}^{(3)} \; (1 \; ; \; \delta_{ij}) &=& -
i \frac{32}{3} (\alpha g^2)^3 N_c
\; Tr'' \; {m}^3 ({\tilde{S}_{0}} \; R)^3
\; (1 \; ; \; \sigma_{i} \sigma_j)
,
\nonumber
\\
g_{6,sb,a} \; g^{\nu\sigma}
( \delta_{ij} \; ; \; i \epsilon_{ijk} )
\;
&=&
i ({\alpha} g^2)^3 N_c \;
Tr'' \; {m} {\tilde{S}_{0}} R {\tilde{S}_{0}} R^{\mu\nu}
R^{\rho \sigma}\;
\gamma_\mu \gamma_\rho \gamma_5^2
\sigma_i \sigma_j \left( 1 \; ; \; \sigma_k \right),
\nonumber
\end{eqnarray}
where for further calculation one defines
$\bar{R}_2^{\nu\sigma} = \bar{R}^{\mu\nu} \bar{R}_{\mu}^{\sigma}
= (R_T + R_L)^2 g^{\nu\sigma} + 8 R_T (R_T - R_L)
\frac{\partial^\nu \partial^\sigma}{\partial^2}$.
There are several chiral invariant and
symmetry breaking
non derivative eighth order interactions.
They were found to be:
\begin{eqnarray} \label{8th-1}
{\cal L}^{(8)}
&=&
g_8
\left[
(\bar{\psi} \psi)^2
+ (\bar{\psi} i \gamma_5 \sigma_i \psi)^2 \right]^2
+
g_{8,sb}^{(2)} ( \bar{\psi} \psi )^2
\left[
( \bar{\psi} \psi )^2
+ ( \bar{\psi} \sigma_i i \gamma_5 \psi )^2
\right]
+ g_{8,sb}^{(4)} ( \bar{\psi} \psi )^4
\nonumber
\\
&+&
g_{8v}
\left[
(\bar{\psi} \gamma_\mu \sigma_i \psi)^2
+ (\bar{\psi} \gamma_5 \gamma_\mu \sigma_i \psi)^2
\right]^2
- g_{8v,sb}
(\bar{\psi} \psi)^2
\left[
(\bar{\psi} \gamma_\mu \sigma_i \psi)^2
+ (\bar{\psi} \gamma_5 \gamma_\mu \sigma_i \psi)^2
\right]
\nonumber
\\
&-& g_{8vs} \left[
(\bar{\psi} \gamma_\mu \sigma_i \psi)^2
+ (\bar{\psi} \gamma_5 \gamma_\mu \sigma_i \psi)^2
\right]
\left[ (\bar{\psi} \psi)^2
+ (\bar{\psi} i \gamma_5 \sigma_i \psi)^2
\right]
\nonumber
\\
&+&
g_{8}^{s}
(\bar{\psi} \gamma_\nu \sigma_j \psi)^2 \left[
(\bar{\psi} \gamma_\mu \sigma_i \psi)^2
+ (\bar{\psi} \gamma_5 \gamma_\mu \sigma_i \psi)^2
\right]
,
\end{eqnarray}
where the chiral invariant terms are of second order of the expansion,
and the symmetry breaking are
of third and fourth orders in the expansion
of $I_d$.
Up to this order of the expansion,
terms in odd powers of the pseudoscalar and axial bilinears
naturally disappear due to the traces such as
$tr (\gamma_5) = 0$.
The effective coupling constants are the following:
\begin{eqnarray} \label{g8}
g_{8}\; (1 \; ; \delta_{ij} ) &=&
4 i (\alpha g^2)^4 N_c
\; Tr'' \; ({\tilde{S}_{0}} \; R^2)^2
(1 \; ; \; \gamma^2_5 \sigma_i \sigma_j),
\nonumber
\\
g_{8,sb}^{(2)} (1\; ; \;\delta_{ij} ) &=&
128 i (\alpha g^2)^4 N_c
\; Tr'' \; m^2 (\tilde{S}_0 R)^2 ({\tilde{S}_{0}} \; R^2)
(1 \; ; \; \gamma^2_5 \sigma_i \sigma_j) ,
\nonumber
\\
g_{8,sb}^{(4)} &=&
64 i (\alpha g^2)^4 N_c
\; Tr'' \; m^4 ({\tilde{S}_{0}} \; R)^4 ,
\nonumber
\\
g_{8v} \Gamma^{\mu_1\nu_1\rho_1\sigma_1}
\Gamma_{ijkl}
&=& \frac{i}{2}
(\alpha g^2)^4 N_c \; Tr'' \; \tilde{S}_0 \bar{R}^{\mu_1\mu_2} \bar{R}^{\nu_1\nu_2}
\tilde{S}_0
\bar{R}^{\rho_1\rho_2} \bar{R}^{\sigma_1\sigma_2}
(\gamma_{\mu_2}\gamma_{\nu_2} \gamma_{\rho_2} \gamma_{\sigma_2})
( \sigma_i \sigma_j \sigma_k \sigma_l),
\nonumber
\\
g_{8vs} g^{\mu\rho} \delta_{ij}
&=& -
i 4 (\alpha g^2)^4 N_c \; Tr'' \; (\tilde{S}_0 R^2) (\tilde{S}_0 R^{\mu\nu}
R^{\rho\sigma} )
\sigma_i \sigma_j \gamma_\nu \gamma_\sigma,
\nonumber
\\
g_{8v,sb} g^{\mu\rho} \delta_{ij}
&=&
- i 8 (\alpha g^2)^4 N_c \; Tr'' \; m^2 (\tilde{S}_0 R)^2 (\tilde{S}_0 R^{\mu\nu}
R^{\rho\sigma} )
\sigma_i \sigma_j \gamma_\nu \gamma_\sigma,
\\
g_{8}^{(s)}
\Gamma^{\mu_1\rho_1}_{\mu\rho}
\delta_{ijkl}
&=&
- \frac{i}{2}
(\alpha g^2)^4 N_c \; Tr'' \; m^2 (\tilde{S}_0 R^{\mu_1\nu_1})
(\tilde{S}_0 R^{\rho_1\sigma_1})
(\tilde{S}_0 R_{\mu}^{\nu}
R_{\rho}^{\sigma} )
\gamma_{\nu_1} \gamma_{\sigma_1}
\gamma_\nu \gamma_\sigma
(\sigma_i \sigma_j \sigma_k \sigma_l),
\nonumber
\end{eqnarray}
where $\Gamma_{ijkl}= \delta_{ij}\delta_{kl} + \delta_{il}\delta_{jk}
- \delta_{ik}\delta_{jl}$
and
$\Gamma_{\mu\nu\rho\sigma}= g_{\mu\nu}g_{\rho\sigma} + g_{\mu\sigma}g_{\nu\rho}
+ g_{\mu\rho}g_{\nu\sigma}$.
Some of these terms were considered
in Ref. \cite{ho-su2}.
The tenth order interaction terms
(leading terms from expansion up to the fifth order)
are all symmetry breaking and
the scalar-pseudoscalar terms can be written as:
\begin{eqnarray} \label{10th-1}
{\cal L}^{(10)}
=
g_{10}^{(1)}
( \bar{\psi} \psi ) \left[
( \bar{\psi} \psi )^2
+ ( \bar{\psi} \sigma_i i \gamma_5 \psi )^2 \right]^2
+
g_{10}^{(3)} ( \bar{\psi} \psi )^3 \left[ ( \bar{\psi} \psi )^2
+ ( \bar{\psi} \sigma_i i \gamma_5 \psi )^2 \right]
+ g_{10}^{(5)} ( \bar{\psi} \psi )^5 ,
\end{eqnarray}
where:
\begin{eqnarray} \label{g10}
g_{10}^{(1)} &=&
- \frac{i}{2} (4\alpha g^2)^5 N_c
\; Tr'' \; m ({\tilde{S}_{0}} \; R) ( \tilde{S}_0 R^2 )^2 ,
\nonumber
\\
g_{10}^{(3)} &=&
\frac{i 3}{4} (4 \alpha g^2)^5 N_c
\; Tr'' \; {m}^3 ({\tilde{S}_{0}} \; R)^3
({\tilde{S}_{0}} \; R^2) ,
\nonumber
\\
g_{10}^{(5)} &=&
- \frac{i}{10} (4\alpha g^2)^5 N_c
\; Tr'' \; {m}^5 ({\tilde{S}_{0}} \; R)^5 ,
\end{eqnarray}
The symmetry breaking terms of the scalar-pseudoscalar channel
can be written
in a general form for the $n-$term of the expansion
in terms of a number (combinatorial) $a_m$,:
\begin{eqnarray}
g_{2n,sb}^{(m)} \; = \;
\frac{i}{n} a_m (2\alpha g^2)^n
\; Tr \; {m}^m ({\tilde{S}_{0}}^{\frac{(n+m)}{2}} \; R^{n})
.
\end{eqnarray}
One can consider two particular limits for calculating
ratios of the quark effective coupling constants
depending on the gluon propagator components.
These ratios are obtained by assuming a large quark mass
and by choosing one of the two following
limits:
(I) $R_L=0$ ($^T$),
or (II) $R_T=0$ ($^L$).
With the expressions shown above which turns out to
depend on the vector or axial bilinears,
the moduli of some ratios yield:
\begin{eqnarray} \label{ratio}
&&
\left| \frac{g_{4}}{g_{v4}} \right|^T
\sim \left| 4 \frac{g_{4}}{ g_{4,v,sb}} \right|^T
\sim 3,
\hspace{.8cm}
\left|
\frac{g_{6,sb}^{(1)}}{g_{6,sb,a}}
\right|^T
\sim 6 ,
\hspace{.8cm}
\left|
\frac{g_{8}}{g_{8vs}}
\right|^T
\sim \frac{3}{4} ,
\\
&&
\left|\frac{g_{4}}{g_{v4}}\right|^L
\sim \left|4 \frac{g_{4}}{g_{4,v,sb}} \right|^L
\sim 1 ,
\hspace{.8cm}
\left|
\frac{g_{6,sb}^{(1)}}{g_{6,sb,a}}\right|^L
\sim 2 ,
\hspace{.8cm}
\left|\frac{g_{8}}{g_{8vs}}
\right|^L
\sim \frac{1}{4} .
\end{eqnarray}
The ratios between the chiral invariant fourth order coupling constants
$(\frac{g_{4}}{g_{v4}})$ are in good agreement with phenomenology
\cite{vector-NJL,sugano,deborah}.
These ratios might therefore present
quite strong gauge dependence and this issue will not be discussed
in the present work.
Some ratios are independent of the gluon kernel component
and their moduli are given by:
\begin{eqnarray} \label{ratio-2}
\left| \frac{g_4}{g_{4,sb}} \right| \sim \frac{1}{2},
\hspace{0.3cm}
\left|\frac{g_{6,sb}^{(1)}}{g_{6,sb}^{(3)}}\right|
\sim \frac{3}{4} ,
\hspace{.3cm}
\left|\frac{g_{8}}{g_{8sb}^{(2)}}\right|
\sim
\left|\frac{g_{8}}{2 g_{8sb}^{(4)}}\right|
\sim \frac{1}{32},
\hspace{0.3cm}
\left|\frac{g_{10}^{(1)}}{g_{10}^{(3)}}\right|
\sim 1,
\hspace{0.3cm}
\left|\frac{g_{10}^{(1)}}{g_{10}^{(5)}}\right|
\sim 5,
\end{eqnarray}
the first of this ratios shows that the exclusive contribution of the
explicit chiral symmetry breaking via the Lagrangian quark mass
for the coupling $(\bar{\psi} \psi)^2$ is of the same order of magnitude
as the NJL coupling.
Next, some numerical values are shown by replacing the
traces in spacetime coordinates by momentum integration rotated to
Euclidean space in the limit of zero momentum exchange.
A simplified confining gluon propagator from Ref. \cite{cornwall-2011}
is considered
with the same values for the prescription given by expression (10) of
Ref. \cite{cornwall-2011}.
The only ultraviolet divergent effective parameter presented above
is the one
for the effective mass correction before Section (\ref{sec:eff-param}).
It can be directly renormalized
with the Lagrangian mass counterterms and it will not be estimated here.
The mass for the quark kernel $\tilde{S}_0$
was considered to be an effective mass from D$\chi$SB $m = 0.33$ GeV and
the
coupling constant $g^2$ as the zero momentum limit of the
QCD lattice calculations divided by 1000,
i.e. $g^2 = 17.8 \pi/(10^3 N_c)$ from Ref. \cite{lattice}.
It is reasonable to consider a reduced value because a full
running coupling constant
would reduce the contribution of the higher energy modes.
The resulting values
were found to be $g_4 \simeq 1.2 $ GeV$^{-2}$, $g_6 \simeq - 28.2$ GeV$^{-5}$,
$g_8 \simeq 4.1 \cdot 10^{4}$ GeV$^{-8}$
and $g_{10}^{(1)} \simeq 2.2 \cdot 10^{8}$ GeV$^{-11}$.
These values are comparable to values
obtained in the literature by phenomenological fitting
except the higher order ones.
From Ref.
\cite{ho-su2} some SU(2) flavor coupling constants were considered as:
$g_4 \sim 10 $ GeV$^{-2}$ and $g_8 \sim 100 - 450$ GeV$^{-8}$,
and for the sake of comparison for SU(3)
Refs. \cite{osipov}
$g_4 \sim 10$ GeV$^{-2}$,
$g_6 \sim - 1100$ GeV$^{-5}$,
$g_8 \sim 6000$ GeV$^{-8}$.
The values for the higher order couplings are somewhat larger than
the values obtained from phenomenology and this might be related to the
truncated momentum dependence considered
and to the values of the parameters $m, g^2$ considered above.
The emerging quark-quark potential
is therefore composed by several types of chiral invariant and symmetry breaking
terms
and this intrincated structure is expected from a confining theory
\cite{greensite}.
Obvious corrections to the effective interactions found above
are due to the derivative interactions that were not calculated
and which may be expected to be relevant for a complete effective theory
for quark dynamics.
It is interesting to emphasize two points:
firstly it can be seen in expressions (\ref{g4sb},\ref{g4vsb},\ref{g6-all}) and
the symmetry breaking couplings of expressions (\ref{g8}) and (\ref{g10}),
that all the symmetry breaking effective interactions
have the effective couplings proportional to the Lagrangian quark mass, that is the explicit symmetry breaking
term. If the quark mass were corrected by the quark condensate to an effective quark mass
the same conclusion holds.
Secondly, the strength of the resulting symmetry breaking effective couplings are of the order of
the chiral invariant terms.
The expressions for these effective quark
interactions were obtained
without an explicit form of the gluon propagator
which plays a fundamental role in the resulting relative strength of the
resulting effective
coupling constants.
Furthermore all the expressions for the effective coupling constants were written
in a way to make possible to compute the corresponding form factors.
It is also interesting to emphasize that results of this work allows for systematic computation of
effective coupling constants without performing extensive phenomenological fits
with hadron masses and/or couplings.
Although the gluon propagator and higher order gluon
interactions in the departing quark effective action
might induce different quark-quark effective interactions
they should not be expected to change the shape of the effective interactions
found in the present work.
\subsection*{ Acknowledgement}
F.L.B. thanks short conversation with G.I. Krein.
This work was
partially supported by CNPq-Brazil
(482080/2013-2).
|
\section{Introduction}
\noindent Phase-change materials (PCMs) constitute a class of semiconductors characterized by two allotrope phases, a crystalline and an amorphous one, having distinct physical properties \cite{Lencer11, deringer15}.
Such materials, discovered in the late 60s \cite{Ovshinsky68}, are already used in rewritable optical discs, such as DVDs and Blu-ray discs.
They are also very promising to realize fast, non-volatile electronic memories (PCRAM).
Because of the variety of PCMs a rigorous chemical definition is absent, although design schemes for some ternary compounds are well established \cite{mottdavis71, lencer08, Lencer11}.
A practical requirement for a PCM, as underlined in Ref.~\cite{lencer08}, is that the switching between the two phases must be reversible, efficient and repeatable, due to obvious technological reasons.
Benchmark PCMs are alloys composed of Ge, Sb and Te -- along the so called GeTe-Sb$_2$Te$_3$ pseudobinary line -- thus denoted as “GST” \cite{chen86,yamada91,wuttig07}.
Although the potential for applications of PCMs spans from dynamic memories \cite{wuttig05} to display fabrication \cite{Hosseini14}, at present the limits for the use of GSTs are the speed and the power required to switch between the two phases, being respectively of approximately 100~ns and 100~$\mu$W \cite{lee-burger09}.
Hence a huge effort is devoted to find new solutions, and nanoscale structures seem to offer significant advantages. \cite{Lankhorst05}.
Following this trend, phase-change heterostructures had been produced, showing to function with improved performances \cite{simpson11}.
Here we focus on chalcogenide superlattice (CSL) films made by high temperature deposition of alternating nm-size layers of GeTe and Sb$_2$Te$_3$ \cite{kooi15}. \\%Despite the interest on CSL \cite{Tominaga14, Bang14}, the understanding of the physical properties of these materials is still incomplete.
Prior to investigate the properties of CSLs, we review the charge carrier transport \cite{Siegrist11,Zhang12} and the structural \cite{ovshinsky69, Kolobov04,akola07,Kolobov11,jostwuttig15} properties of GST compounds.
The crystal phase of GSTs is characterized by a strong dependence of the resistivity upon annealing treatment \cite{Siegrist11}.
In fact, by increasing the temperature, the system undergoes an insulator-metal transition (IMT) at 275$^\circ$C, due to disorder reduction. \cite{Siegrist11}
Density functional theory calculations showed that the insulating state is caused by the localization of charge carriers in vacancy-rich areas, and the transition to the metallic state happens when vacancies reconfigure into ordered layers \cite{Zhang12}.
The presence of vacancy layers in the high temperature annealed phase of GSTs was proven by high-resolution TEM and electron diffraction \cite{kooi02}. \\
The lattice structure of the crystalline Ge-Sb-Te based CSL is similar, by growth design, to the high temperature GST phase.
As mentioned before CSL are formed by high temperature deposition -- although lower than the IMT -- of GeTe and Sb$_2$Te$_3$ layers.
One might consequently expect that the as grown structure is characterized by well separated layers of pure GeTe and Sb$_2$Te$_3$.
Recent TEM and EXAFS experiments \cite{kooi15,casarin16} have shown that the structure is indeed layered, but GeTe is not actually pure since it is intercalated into a Sb$_2$Te$_3$ quintuple layer.
Please note that the vacancy layers, when two-dimensionally extended, are labelled van der Waals gaps. \cite{oeckler12}
The similarity between the high-temperature-annealed GSTs \cite{Siegrist11} and the CSLs, besides the structure, is corroborated by the measured resistivity of the latter that decreases with increasing temperature, suggesting a metallic type conductivity \cite{unpublished1}.
Nonetheless, it is not obvious whether also the dielectric properties of CSLs are analogous to those of GST compounds \cite{lucovsky73,littlewood79_diel, yokota89, huang10}.
For instance one of the peculiar characteristics of crystalline GST is the high value of the static dielectric constant, which is evidence of resonance bonding. \cite{shportko08}
Measuring the dielectric function of CLS will provide useful information either about the structure, or about the bonding character of the compounds.
The dielectric function of CSLs can be estimated by transmittance experiments, probing simultaneously the character of the interband transition and the density of free electrons. \\
\noindent In this article we compare the optical transmission behaviour of films grown by DC magnetron sputtering (MS), with those grown by molecular beam epitaxy (MBE) used as reference samples.
Transmission experiments are performed on films with thickness below 100~nm, to guarantee high film quality, and at normal incidence, to simplify the interpretation of the results.
Our analysis supports recent results showing that the GeTe layers are not isolated but intercalated within a Sb$_2$Te$_3$ layer \cite{kooi15, casarin16}.
We explain how intermixing modifies the conductivity of the film.
In addition we show that it is possible to control the electronic transport properties of the CSL not only by an annealing treatment, as in Ref. \cite{Siegrist11}, but also by means of a specific stack design.
In fact, as we will see, the CSL conductivity depends on the layer sequence, giving a new design pathway for the control of the material properties. \\
\section{Experimental}
CSL samples are grown by MS and MBE on Sb-passivated Si(111) surfaces of intrinsic 500~$\mu$m-thick wafers above 210 $^\circ$C.
MS samples are films of 15 repetitions of [GeTe($N$~nm)/Sb$_2$Te$_3$(3~nm)], where $N=1,2,3$.
For the MBE samples we use the intermediate N=2 case.
MS annealed (ANN) samples -- at 250~$^\circ$C for 30 minutes -- are distinguished from the MS as deposited (ASD) ones.
Transmission experiments in the energy range 0.06--1.23~eV were collected at SISSI beamline on the Elettra storage ring \cite{lupi07} by a $Bruker~VERTEX~70v$ spectrometer.
The fitting function is determined for multilayer thin films optics \cite{python_tmm}.
\section{Fitting model and results}
Figure \ref{tall} shows the absorbance $-$ln$(T_{CSL}/T_{Si})$, with $T_{CSL}$ and $T_{Si}$ respectively the sample and the substrate transmission, as a function of the incoming photon energy for all the samples.
The data present few physically relevant features, common to all samples and regardless of growth technique and layer stack.
In particular, we observe
i) the onset of an absorption edge below $\sim$0.15~eV, increasing by decreasing the photon energy;
ii) a minimum between 0.15 and 0.3~eV, and
iii) a broad feature rising above $\sim$0.4~eV, increasing towards higher energies.
In analogy to Ref.~\cite{shportko08} we associate the low frequency absorption component to free Drude electrons, and the higher energy broad band to the onset of the valence-to-conduction absorption.
The absorption modulation of the data, particularly visible on the MS 33 annealed sample, is an artefact due to multilayer interference effects.
We account for interference effects by calculating the transmission coefficient $T_{Data} = T_{CSL}/T_{Si}$ including reflection losses.\\
\begin{figure
\includegraphics[width=\columnwidth]{Fig1.pdf}
\caption{(Color online) Absorbance spectra of MBE 23, MS 13-23-33 as deposited (ASD) and MS 13-23-33 annealed (ANN) samples.} \label{tall}
\end{figure}
The film dielectric function (DF) is modelled with a sum of a Drude and a single Tauc-Lorentz (TL) oscillator term \cite{jellison96,shportko08}:
\begin{equation} \label{eps}
\epsilon(x) = \epsilon^{Drude}(x) + \epsilon^{TL}(x),
\end{equation}
where
$$\epsilon^{Drude}(x) = -\frac{\omega_p^2}{x^2+1/\tau^2} + i \frac{\omega_p^2}{x \tau (x^2+1/\tau^2)},$$
with $x$ the photon energy, $\omega_p$ the plasma frequency and $\tau$ the scattering time.
The real part of $\epsilon^{Drude}$ is a negative Lorentzian function centred at zero while the imaginary part is positive and diverges at zero.
The Tauc-Lorentz dielectric function is obtained by Kramers-Kronig integration ($\epsilon_1(x) = 1 + \frac{1}{\pi}\int_{-\infty}^{\infty} \frac{\epsilon_2(x')}{x'-x} dx'$) from the imaginary part:
\begin{equation} \label{epstl}
\left \{
\begin{array}{rcl}
\epsilon^{TL}_2 (x \leq E_g) &=& 0\\
\epsilon^{TL}_2 (x > E_g) & = & \frac{A~C~E_0 ~(x-E_g)^2 }{ x~[(x^2-E_0^2)^2+C^2 x^2]}\\
\end{array}
\right. ,
\end{equation}
where $A$ is the TL amplitude, $E_g$ is the Tauc --or optical-- gap,
$C$ is the band-width and $E_0$ is the central frequency,
representing the photon energy where the transition probability is at the maximum.
\begin{figure
\includegraphics[width=\columnwidth]{Fig2.pdf}
\caption{(Color online) a) Transmission data MBE~23 and fitting curve $T_{Fit}$.
b) Dielectric function resulting from the fit.
Both the real (red) and the imaginary (blue) components of the DF are shown also subtracting the Drude contribution (dashed).
}\label{fitres}
\end{figure}
For the (intrinsic) silicon substrate we use $\epsilon(x) \sim 3.45$ \footnote{http://refractiveindex.info}.
We verified the assumptions made for the dielectric properties of the substrate by transmittance measurements on virgin wafers.
Regarding the calculation of $T_{Fit}(x)$ we include multiple coherent propagation only for the thin CSL film.
Also the CSL film thickness is a free parameter of the fitting procedure.
A typical fit result is shown in Figure \ref{fitres}a.
The related DF is plotted in Figure \ref{fitres}b.
The misfit between the experimental and calculated data is due to Si phonons (the peak at $\sim$0.3~eV) and probably to the small contribution of interference effect from the substrate that have not been accounted for.
In all cases the fit function $T_{Fit}(x)$ describes all the main features of the transmission data, thus the deduced parameters are meaningful.\\
The fit parameters for the TL oscillator and the Drude term, for each investigated CSL, are given in Table \ref{tab_par}.\\
\begin{figure
\includegraphics[width=\columnwidth]{Fig3.pdf}
\caption{(Color online) Free electron density $n=\frac{m}{e^2} \omega_p^2$ obtained from the Drude term of the fit.
A decrease of $n$ is observed with increasing GeTe content.
}\label{fin_par}
\end{figure}
\begin{table}\centerin
\begin{ruledtabular}
\begin{tabular}{r c c c c c c}
& \multicolumn{3}{c}{TL} & & \multicolumn{2}{c}{Drude}\\
\hline
& $E_g$(eV) & $E_0$(eV) & $\epsilon^{TL}_1(0)$ && $\omega_p$(eV) & $\tau$(fs) \\
\hline
MBE 23 & 0.3(1) & 1.7(2) & 23 && 0.7(1) & 8(2) \\
\hline
13 & 0.1(2) & 1.5(2) & 35 && 1.0(1) & 9(2) \\
ASD 23 & 0.0(1) & 1.5(1) & 29 && 0.9(1) & 6(2) \\
33 & 0.0(1) & 1.5(1) & 24 && 0.7(1) & 7(2) \\
\hline
13 & 0.1(2) & 1.3(1) & 30 && 1.1(1) & 8(2) \\
ANN 23 & 0.0(1) & 1.3(1) & 24 && 0.8(1) & 13(2) \\
33 & 0.5(1) & 1.5(1) & 23 && 0.7(1) & 6(2) \\
\end{tabular}
\end{ruledtabular}
\caption{Paremeters resulting from the fit of Eq.~(\ref{eps}) for all the measured samples.
The error on the fit procedure is shown in parenthesis, representing the deviation of the last digit.
The low-energy limit of the TL DF, $\epsilon^{TL}_1(0)$, is also computed. }\label{tab_par}
\end{table}
\section{DISCUSSION}
We first consider the TL component of the DF, whose $E_g$(eV), $E_0$(eV), $\epsilon^{TL}_1(0)$ parameters are given in Table \ref{tab_par}.
The Tauc gap lays below 0.5~eV.
$E_0$, representing the maximum of the semiconductor optical absorption, has an average value of approximately 1.5~eV.
This value matches quite well with the experimentally obtained value of crystalline GeTe \cite{velea15} and GST \cite{shportko08}.
Moreover, the computed
low frequency value of the real part of the TL DF, $\epsilon^{TL}_1(0)$, has an average value above 20.
This $\epsilon^{TL}_1(0)$ high value is typical of the crystal phase of PCMs \cite{shportko08}, suggesting the presence of resonance bondings.
In summary the TL component of our CSL is approximately independent on the annealing treatment.
The fact that annealing does not affect the as-deposited (ASD) structures might suggest that the films undergo an immediate reorganization during growth, which is more effective than annealing itself. \cite{calarco16}
The analysis of the free electrons in our CSL supports this hypothesis.\\
\begin{figure
\includegraphics[width=0.7\columnwidth]{Fig4.pdf}
\caption{(Color online) a) Sketch of the band diagram of GeTe, Sb$_2$Te$_3$ and GST materials.
The Tauc gap $E_g$ in both GeTe and Sb$_2$Te$_3$ is lower with respect to GST compounds.
Thermal promotion of valence band (VB) electrons to the conduction band (CB) is accordingly reduced in GSTs, hence the free electron density and the conductivity are lower.
b) Formation of GST layers at the interface between GeTe and Sb$_2$Te$_3$. } \label{scheme}
\end{figure}
The plasma frequency and the scattering time, resulting from the fitting procedure, are presented in Table \ref{tab_par}.
While the scattering time ranges between 5-15 fs -- just half with respect to Copper at 300 K \cite{tanner} -- the plasma frequency consistently decreases with the increase of the stacking sequence 13-23-33,
both for the ASD and ANN samples.
The CSL electron density $n=\frac{m}{e^2} \omega_p^2$, derived from $\omega_p$, is shown in Figure \ref{fin_par}.
A decrease of $n$ is observed with the increase of nominal GeTe content in the film.
Also the conductivity $\sigma = \frac{e^2}{m} n \tau$ shows the same trend.
This behaviour can be explained either by an increased defectivity, by a lower conductivity of GeTe with respect to Sb$_2$Te$_3$, or by an increase of the optical gap.
Yet, since the comparison acts within the MS grown samples, we expect the same growth quality and -- consequently -- the same concentration of defects.
In addition, pure GeTe has a \emph{higher} free carrier concentration with respect to Sb$_2$Te$_3$. \cite{GST95,tong15}
Thus, the expected trend with stack sequence 13-23-33 should be opposite than what observed.
It follows that variations of $n$ must result from changes of the electronic band structure. \\
The intermixing of GeTe and Sb$_2$Te$_3$ layers rises the material optical gap, causing a reduction of the free carrier concentration.
Figure \ref{scheme}a sketches the band diagrams for GeTe, Sb$_2$Te$_3$ and GSTs, in analogy to Refs \cite{ovshinsky69, Siegrist11}.
The diagrams emphasize the differences between the optical gaps and the presence of degenerate electrons in the conduction band.
Both GeTe and Sb$_2$Te$_3$ have a low Tauc gap $E_g \sim$0.1~eV \cite{esaki68, chopra69, ibram91}, while for GST $E_g^{GST} \sim$ 0.4~eV \cite{shportko08, kato05, park08}.
In first approximation the Tauc gap of the CSL depends on the constituents volume fraction and on their respective gaps.
If GeTe and Sb$_2$Te$_3$ are ideally separated inside the CSL the film is expected to have small gap and high carrier concentration, independent of the relative content of the two elements.
If instead GST is formed by the intermixing of the two, the material Tauc gap must rise.
The more GST is formed, the higher the Tauc gap (approaching eventually the value of GSTs) and the lower the number of degenerate electrons.
This is consistent with the carrier density of the sequence 13-23-33, shown in Figure \ref{fin_par}.
\section{Electric transport control}
\begin{figure
\includegraphics[width=0.8\columnwidth]{Fig5_b.pdf}
\caption{(Color online) a) Conductivity of the film calculated from Eq.~(\ref{sigma}) for increasing number of GeTe units, and compared with the measured conductivity of the CSLs.
b) Conductivity of a film with stack sequence (GeTe)$_N$/(Sb$_2$Te$_3$)$_M$.
The red dashed diagonal indicates the lowest conductivity region.
The black dashed horizontal lines represent the cut used to generate the graph a).
}\label{NM}
\end{figure}
We extrapolate the trend observed in the present CSL to a stack of general thickness, having (GeTe)$_N$/(Sb$_2$Te$_3$)$_M$ as repeat unit, where $N,M \geq 1$.
This formula is applicable to all CSL structures developed so far in literature:
i) superlattice-like (SSL) PCMs, with $N,~M>2$ \cite{chong06};
ii) interfacial phase change materials (IPCM), with $N \leq 4,~M \leq 2$ \cite{simpson11};
iii) the case $N=2M$, studied in Ref.~\cite{tong15}
and iv) our CSLs.
We assume our stack sequence corresponds to $N=2, 4, 6$ and $M=3$, since 1~nm of GeTe along the (111) direction is composed by at least two bilayers \cite{Tominaga14,kooi15}.
The conductivity of a CSL in the general case is calculated assuming that GeTe and Sb$_2$Te$_3$ layers intermix at the interface, forming GST as shown in Figure \ref{scheme}b.
In particular, we consider that the interface layer consists of a maximum of 2 Sb$_2$Te$_3$ and 4 GeTe layers, resulting in 1 thick layer of GST.
Then, given the bulk conductivities $\sigma_{GeTe}$, $\sigma_{Sb_2Te_3}$, $\sigma_{GST}$ \cite{tong15}, the film conductivity is calculated as:
\begin{equation}\label{sigma}
\sigma^{(N,M)} = \sum_{i} v_i \sigma_i
\end{equation}
where $v_i$ is the volume fraction of the $i$-th component (GeTe, Sb$_2$Te$_3$ and GST). \\
In Figure \ref{NM}a, the conductivity measured for all our samples is compared with the conductivity $\sigma^{(N,M=3)}$, as a function of GeTe layer units $N$.
In agreement with experimental observations the conductivity decreases from $N=1$ down to $N \sim 8$.
The strikingly different behaviour of the ``non-intermixing'' case, calculated simply via Eq.~(\ref{sigma}) with $v_{GST}=0$, demonstrates once more that intermixing effects must be accounted for when working with superlattice structures. \\
For completeness the dependence of $\sigma^{(N,M)}$ is shown in Figure \ref{NM}b.
Interesting to note the conductivity has a minimum along the diagonal $M\sim N/2$ up to $N=8$
that corresponds to the maximum possible volumetric formation of GST, hence forming the material with the lowest conductivity.
For higher values of $N$ and $M$ the conductivity approaches the ``non-intermixing'' case, as also visible for high values of GeTe units in Figure \ref{NM}a.\\
\section{Conclusion}
In this work we deduce important structural details of crystalline GeTe/Sb$_2$Te$_3$ superlattice deposited at high temperature ($>$210~$^\circ$C) by studying the film interband transmission.
We observe that, at the interface between the GeTe and Sb$_2$Te$_3$ deposition layers, a crystalline GST compound is formed already during growth.
GST, having higher band gap with respect to its constituents, lowers the number of degenerate conduction electrons.
By varying the respective number of building block layers it is possible to control the percent of GST in the film, and consequently the film conductivity, with no need of annealing treatment.
In the future it would be interesting to investigate the dynamics deriving from the intermixing and ion diffusion by depositing two single thick layers of GeTe and Sb$_2$Te$_3$ and studying the interface effects at different deposition and annealing conditions.
\section{Acknowledgements}
This work was supported by EU within FP7 project PASTRY [GA 317764].
AC would like to acknowledge Roberta Ciprian for useful comments.
|
\section{Introduction}
\noindent One major success of Plasmonics relies on the extraordinary possibility of electromagnetic field localization at the nanoscale, mediated by collective electron excitations at suitable metal-dielectric interfaces\cite{MaierBook:2007}.
A homodimer of nanoparticles (NPs) with dimensions $ \ll \lambda_{\rm o}$ (free space wavelength) provides the basic element of most plasmonic architectures. The local photonic density of states can be amplified of several orders of magnitude into volumes comparable with molecular length scales into so-called hot-spots\cite{Halas:2011}.
Near- and far-field responses supported by the dielectric geometry of the system are determined by the hybridized localized surface plasmon-polariton (LSP) modes\cite{Nordlander:2004}. Engineering of such optical responses has provided a formidable variety of nanophotonic tools that include nanoantenna devices\cite{Giannini:2011,Brown:2010}, surface-enhanced Raman scattering (SERS) substrates\cite{Zito:2015,Cecchini:2012,DeRosa:2015,Isticato:2013,Matteini:2015}, integrated optical devices\cite{Fang:2015}, metasurfaces\cite{Lin:2013}, plasmon-induced loss devices\cite{Ndukaife:2015}, etc.
In general, when the symmetry of a plasmonic configuration of nanoparticles is broken, a larger variety of coupling mechanisms between NP plasmons are possible\cite{Brown:2010}, which exponentially broadens the capability of light interaction and manipulation. \\
\indent Recently, Chen \textit{et at.} reported on the possibility to light specific hot-spots into NP aggregates on a metal film \cite{Chen:2015}. In fact, the coupling between NPs and film support charge density modes that show special charge accumulations over the metal-dielectric boundaries and determine specific near-field distributions. The possibility to excite special hot-spots therefore arises from the symmetry breaking introduced by the metal film.\\
\indent In this study, we investigate the optical response of a linear chain of six mismatched silver nanoparticles (linear hexamer) under plane wave illumination. The chain modes \cite{Tserkezis:2014} of our linear hexamer provide a rich variety of hot-spot patterns in the gaps between the various NPs, deterministically controlled by the excitation energy. The novelty is that for special antibonding modes, strongly coupled collinear dark plasmons with internal, antiparallel dipole moments can totally quench specific hot-spots, damping the scattered near-field down to the level of the background field. In contrast to cold spots of isolated nanoparticles\cite{Haggui:2012}, we term such an anomalous gap as \textit{dark} spot. Remarkably, we find an exceptional asymmetric near-field distribution around the NPs nearest to the dark gap. The amplitude-gain contrast is up to $\sim$ 228. To the best of our knowledge, this possibility has never been predicted before. \\
\section{Slowly Scaling Linear Chain of Nanoparticles}
Chains of mismatched NPs have been used to increase near-field enhancement in plasmonic applications such as SERS spectroscopy\cite{Ding:2010,Tserkezis:2014} and near-field chemical scanning\cite{Hoppener:2012,Rusciano:2014}. Indeed, chains of multiscale NPs may enable huge enhancement of the local electric field\cite{Li:2003}. In this work, we study a (relative) simpler system with slow scale variation. We calculate its optical response as a function of the overall length scale introducing an isotropic scale factor $f$ (global parameter). The starting geometry consists of six nanoparticles with radii, from the biggest to the smallest, $R_{\{1,2,3,4,5,6\}} = \{13, 10, 7, 5, 4, 3\}$ nm, in this order, and with gaps measured between closer points $g_{\{12,23,34,45,56\}} =\{ 1.5, 1, 1, 1, 1\}$ nm. For this particular geometry, where only $g_{12} = 1.5$ nm, we empirically found an interesting case of LSP modal interference. In Fig. 1(a), the geometry is depicted for $f = 1.0$. \\
\indent We carried out 3$D$ numerical simulations solving the full electrodynamic problem with a finite element method-based commercial software (Comsol 4.3b). We considered spherical nanoparticles of silver with dielectric function from Johnson and Christy\cite{Johnson:1972} embedded in a surrounding vacuum medium. The spherical simulation region was embedded into a perfectly matched layer with outer scattering boundary conditions\cite{Zito:2015}. Minimum mesh element was 0.07 nm. The simulations were carried out with wavelength step of 2.5 nm (0.2 nm for finer inspections). Relative error tolerance was set to 1$\times$10$^{-7}$. With a quad-core processor Intel i5-3570K @3.4GHz and 32 GB of RAM, simulation time was ca. two minutes per wavelength. The dipole moments of the NPs associated to the various chain modes were calculated integrating the polarization density vector over the NP volumes and, for control, by the relation $\vec{p}(\vec{r}) =\int_{\partial V} \sigma(\vec{r_{0}})(\vec{r_{0}}-\vec{r}) {\rm d^2}\vec{r_{0}}$, where $\sigma$ is the surface charge density given by the relation $\sigma (\vec{r_{0}}) = \epsilon_{0}n^2 \hat{u} \cdot \left[\vec{E_{s}} (\vec{r_{0}} )+ \vec{E_{0}} (\vec{r_{0}})\right] $, whereas $\partial V$ is the surface of the nanoparticles with $\hat{u}$ normal versor to the boundaries and $n$ the refractive index. The scattering far field response was calculated by using an auxiliary surface enclosing radiating NPs.\\
\indent The scattering cross section (SCS) spectrum of the linear hexamer excited by a plane wave with polarization parallel to the chain axis is shown in Fig. 1(b). The family of curves is obtained by varying $f$. The SCS presents multiple peaks across the UV-visible range related to high-order LSP resonances. We will refer to these modes indicating the lower energy radiant mode of interest as RM$_{0}$ and progressively higher order radiant modes as RM$_{1,2,3}$. Of course, the spectral positions of these modes evolve with the system size as clearly visible in Fig. 1(b). An increase of the surrounding $n$ produces a redshift of the peaks without affecting their overall structure (not shown). At RM$_{0}$ the near-field is maximized. It is worth noticing that this mode, spectrally located just on the left of the SCS peak, is not a dipole mode. The dipole mode is located on the right of the SCS maximum and shows inferior near-field enhancement. The spectral dip indicated as subradiant mode (SM) in the SCS is of particular interest to our discussion as will be clear later. \\
\begin{figure*}[t]
\centering
\setlength{\unitlength}{\textwidth}
\includegraphics[width = 13cm]{Figure_1_resub.png}
\caption{(a) Schematic layout of the hexamer of silver spheres ($f = 1.0$). (b) Family of SCS spectra parametrized by the isotropic scale factor $f$ for plane wave linearly polarized along the axis chain. (c) Surface charge density modes (wireframed on NP surfaces) and overlaid $G$-map for different LSP modes at $f = 1.0$, calculated in the $xy$ midplane of the chain for polarization parallel to the chain axis. (d) Field enhancement cross-section along the chain axis for the same LSP modes depicted in (c) (curves are vertically shifted for clarity).}
\label{fig:Figure_1}
\end{figure*}
\indent We will focus first on the case $f = 1.0$. The characteristic surface-charge density modes $\sigma(\lambda_{\rm o})$ on each nanoparticle are plotted in Fig. 1(c) for RM$_{i}$ $(i = 1, 2, 3)$ and SM. In the heterochain, translational symmetry is broken. As a consequence, the mode profiles $\sigma (\lambda_{\rm o})$ become characteristic of each NP since determined by the local topology of the system and the interaction with the surrounding NPs. Overlaid in panels of Fig. 1(c), we also report the corresponding spatial distributions of the fourth power of scattered field amplitude $E_{\rm s}$, normalized to the incident plane-wave amplitude $E_{\rm o}$, i.e. the approximated SERS enhancement factor $G(x, y, z) = |E_{\rm s}/E_{\rm o}|^4$ (indicated as $G$-map). The five main states of interest extracted in Fig. 1(c) correspond to the spectral positions indicated in the SCS of Fig. 1(b) for $f = 1.0$. The chain modes of the linear hexamer provide a rich variety of hot-spot patterns in the NP gaps. Such patterns consist of sequences of relative maxima of the local field that can be deterministically controlled by varying the wavelength of the excitation radiation. In fact, the field amplification does not follow a cascade progression towards the smallest NP, which requires a rapidly scaling chain\cite{Li:2003}. We find that local field amplification $|E_{\rm s} / E_{\rm o}|$ in the various gaps of the chain is highly dependent on the particular LSP. In Fig. 1(d), we plot it along the chain axis for the resonant modes illustrated in Fig. 1(c). The left edge of NP$_2$ [Fig 1(a)] is the origin of the coordinate system. As can be seen, the position of maximum amplification moves from one gap to another following a scheme that can be associated to each LSP mode. Indeed, at a fixed wavelength, the local surface charge associated to the mode can accumulate on specific regions of the NPs, in equilibrium with the whole system (of course, the overall net charge must be null), and give rise to gap hot-spots of different strength. \\
\begin{figure*}[t]
\centering
\setlength{\unitlength}{\textwidth}
\includegraphics[width = 13cm]{Figure_2_resub.png}
\caption{(a) Surface charge density modes and overlaid $G$-map at SM for $f=1.5$, $\lambda_{\rm o} =$ 390 nm in Fig. 1(b), for polarization parallel to the chain axis. (b) Close-up around $g_{34}$. (c) Electric field gain along the chain axis: in $g_{34}$, the gain is damped of nearly 2 orders of magnitude. (d) SCS spectrum (blue) and dipole moment ratio between small trimer ($p_{2}$) and large trimer ($p_{1}$) for polarization along the axis: the dipoles become antiparallel in correspondence of SM. (e) Optimum found at $g_{34} = 0.8$ nm ($f = 1.56$): exactly $E_{{\rm s}} = E_{{\rm 0}}$ in the dark gap (colormap inverted for clarity). (f) Scattering and absorption cross section (ACS) in case of polarization orthogonal to the chain axis, for geometric paramters as in (e). (g) Surface charge density and $G$-map at 365 nm (maximum near-field) for orthogonal polarization: gap hot-spots form also in this case.}
\label{fig:Figure_2}
\end{figure*}
\indent Now, let us consider the behavior of the subradiant SM dip that occurs at 382 nm for $f = 1.0$ [Fig. 1(b)]. Inspecting the corresponding mode profile in Fig. 1(c) (SM), we observe a sort of interruption at $g_{\rm 34}$ in the sequence of gap hot spots. In the adjacent gaps, the field is again highly amplified. The local field is reduced by a factor 20 in $g_{\rm 34}$ with respect to the maximum field gain observed along the chain axis, see SM in Fig. 2(c)-(d). Looking into this phenomenon, we observe a further decrease of the field amplification for $f = 1.5$. The corresponding mode and field maps are shown in Fig. 2(a)-(b). Indeed, the SM mode further damps the local gain in $g_{34}$ down to $\sim$ 3, while in the adjacent gaps $g_{23}$, $g_{45}$ and $g_{56}$ the local field is enhanced up to 232-, 110- and 116-fold $E_{\rm o}$, respectively [Fig. 2(c)]. We define the gain contrast as the ratio $E_{\rm s}(g_{23}) / E_{\rm s}(g_{34})$ between the field enhancements observed in the adjacent gaps around NP$_{3}$. This ratio evaluates the asymmetry of the amplification on that NP. The maximum contrast of the field gain is therefore equal to a factor $232/3 \simeq 77$, which would correspond to an approximated SERS amplification quenching from $2.9\times 10^9$ to 81, i.e. a decrease of more than seven orders of magnitude ($3.6\times10^7$), followed by a rapid spatial increase to $1.5\times10^8$ in the next gap of the chain [Fig. 2(b)]. A further fine tuning of $f$ allows us increasing the contrast to $229/2 = 114.5$ with $f = 1.56$, always at SM ($\lambda_{\rm o} = 385$ nm). Therefore, when spectrally approaching the SM-dip, the local field gain is nearly turned off in $g_{34}$ and the actual damping depends on the tuning of the modes triggered by the coupling between the NPs. The otherwise bright hot-spot is highly damped. In Fig. 2(c), such a hot-spot has been indicated as dark hot-spot for the reason clarified in the following. \\
\indent To shed light on this phenomenon we calculated the dipole moments of the system as a function of $\lambda_{\rm o}$. After inspecting the surface-charge density $\sigma$ on the NPs in Fig. 2(a), we focused on the partial plasmon modes formed by the first and last three NPs - say large- and small-trimer, respectively. While the first has a weakly radiative symmetry, the second is a bright mode but with smaller dipole moment because of the smaller NP size. Both dipole moments of these trimers, respectively $\vec{p}_{1,2}$ [Fig. 2(a)], are weak at SM. In fact, $p_{2} \simeq -0.06 p_{1}$ and their sum is $\sim 1/10$ of the hexamer dipole moment induced at the RM$_{0}$, mainly contributed from the weakly-radiative large trimer. We can say that the overall hexamer LSP behaves as a \textit{binary} dark plasmon mode since formed by a dark (large trimer) and a bright mode (small trimer) that have \textit{collinear} antiparallel dipole moments at SM, which decreases the overall dipole moment of the system. Thus the hexamer mode is subradiant. In Fig. 2(d), we can observe the reversal of the polarity of the trimer plasmons in coincidence with the spectral dip SM (blue line). The spectral dip reflects the minimum of total dipole moment of the hexamer. Therefore, the opposite dipole moments, induced on average on both trimer plasmons, produce a minimal charge density on the NP surfaces forming the gap $g_{34}$ (nearly neutral). This produces an apparent decoupling between the trimers despite their close proximity ($g_{34} \sim 1$ nm) responsible for the damping of the field in $g_{34}$. This mechanism is therefore at the origin of what we term dark hot-spot in the interjunction between the trimers [Fig. 2(b)-(c)]. Although being a relative maximum of near-field, it is surprisingly weak if compared to the bright hot-spots that the very same NPs on the side of $g_{34}$ create with the next NPs of the chain. \\
\indent At this stage, we explored the influence of the coupling also varying the size of the gap. Fixed $f = 1.56$, we explored the range $g_{34} =$ 0.6$ - $2.0 nm leaving unchanged the other geometric parameters of the chain. Please note that since $f = 1.56$ all parameters rescale, so that $g_{34} = $ 0.936$-$3.12 nm in the rescaled structure. We find that the contrast reaches a maximum for $g_{34} = 0.8$ nm (i.e. 1.248 nm after rescaling). In fact, as shown in Fig. 2(e), it is possible to have a scattered field exactly equal to the background field, $|E_{\rm s} / E_{\rm o}| = 1$. In this case, the maximum contrast (near field enhancement $g_{23} : g_{34}$) becomes equal to $228/1$ (i.e., $2.7 \times 10^9$ in the SERS enhancement factor). Since the field enhancement is now totally quenched, the dark hot-spot more properly may be indicated as dark spot [Fig. 2(e)]. Changing the value of $g_{12}$ from 1.5 nm reduces the optimal quenching (we explored the range 1.2 - 1.7 nm). \\
\indent In case of polarization perpendicular to the chain axis, shown in Fig. 2(f)-(g), the coupling between adjacent NPs gives rise to multipolar splitting of the charge density close to the gaps. As a consequence, also in this case, gap-hot spots can form. Therefore, it is not possible to quench the near-field in the gaps simply rotating the incident polarization. These hot-spots are characterized by two distinct, symmetric maxima in each gap. A $G_{{\rm max}} = 4.4 \times 10^6$ is found at 365 nm (wavelength of maximum near-field) [Fig. 2(g)].
\section*{Conclusion}
We have theoretically investigated the optical response of mismatched silver nanoparticles forming a slowly scaling linear chain. Our study sheds light on novel phenomena that may arise into plasmonic systems. Patterns of local field gain can be deterministically excited in the gaps of the chain. The local field gain can be totally quenched in association to a dark chain plasmon with internal antiparallel dipole components along the chain, despite the close proximity ($\sim$ 1 nm) between NPs. In principle, this effect could be used to manipulate the emission properties of quantum emitters in close proximity to the bright and dark gaps. Addressable spatial control of field localization might be employed for novel nanophotonic devices. These results have potential impact for plasmon-enhanced spectroscopies, nanosensing and plasmon-induced loss or plasmonic-force devices. \\
\indent We acknowledge financial support from Italian Ministry of Education, University and Research, Grant No. FIRB 2012-RBFR12WAPY; and in part from University of Naples Federico II, Compagnia di San Paolo e Istituto Banco di Napoli - Fondazione (LARA).
\end{document}
|
\section{Introduction}
\label{sect:intro}
Visual tracking plays an important role in computer vision and has many applications such as video surveillance, robotics, motion analysis and human computer interaction.
Even though various algorithms have come out, it is still a challenge problem due to complex object motion, heavy occlusion, illumination change and background clutter.
Visual tracking algorithms can be roughly categorized into two major categories:
discriminative methods and generative methods.
Discriminative methods (\emph{e.g.},~\cite{co/training/iccv/LiuCL09,MIL/cvpr/BabenkoYB09,Struck/iccv/HareST11}) view object tracking as a binary classification problem in which the goal is to separate the target object from the background.
Generative methods (\emph{e.g.},~\cite{Appearance/Models/pami/JepsonFE03,IVT/ijcv/RossLLY08,L1PCA/Liu14,LRF/Zhang14,LSP/Liu14})
employ a generative appearance model to represent the target's appearance.
We focus on the generative one and will briefly review the relevant work below.
Recently, sparse representation has been successfully applied to visual tracking (\emph{e.g.},~\cite{l1/tracker/mei,Liu/sparse/represnetation/tracking/application,MTT/ijcv/ZhangGLA13,conf/icmcs/Jinwei}).
The trackers based on sparse representation are under the assumption that the appearance of a tracked object can be sparsely represented by a over-complete dictionary which can be dynamically
updated to maintain holistic appearance information. Traditionally, the over-complete dictionary is a series of redundant object templates, however, a set of basis vectors from target subspace as dictionary is also used because an orthogonal dictionary performs as efficient
as the redundant one. In visual tracking, we will call the $L_1$ regularized object representation "sparse coding" (\emph{e.g.},~\cite{l1/tracker/mei}), and the $L_0$ regularized object representation "sparse counting" (\emph{e.g.},~\cite{panl0}).
\cite{l1/tracker/mei} has been shown to be robust against partial occlusions,
which improves the tracking performance. However, because of using redundant dictionary, heavy computational overhead in $L_1$ minimization hampers the tracking speed.
Very recent efforts have been made to improve this method in
terms of both speed and accuracy by using accelerated proximal gradient (APG) algorithm~\cite{APGL1/bao}
or modeling the similarity between different candidates~\cite{MTT/ijcv/ZhangGLA13}.
Different from~\cite{l1/tracker/mei}, IVT~\cite{IVT/ijcv/RossLLY08} incrementally learns a low-dimensional PCA subspace representation, which adapts online to the appearance changes of the target.
To get rid of image noise, Lu \emph{et al.}~\cite{Lu/tip13/Wang13} introduce $L_1$ noise regularization into the PCA reconstruction, which is able to handle partial occlusion and other challenging factors.
Pan \emph{et al.}~\cite{panl0} employs $L_0$ norm to regularize the linear coefficients of incrementally updated linear basis (sparse counting) to remove the redundant features of the basis vectors.
However, sparse counting will cause unstable solutions because of its nonconvexity and discontinuity. Although the sparse coding has good performance, it may cause biased estimation since it penalizes true large coefficients more, and produce over-penalization. Consequently, it is necessary to find a
way to overcome the disadvantages of spare coding and sparse counting.
From the viewpoint of statistics, sparse representation are similar to variable selection when the dictionary is fixed. Besides, it is a bonus that Bayesian framework has been successfully applied to select variables by enforcing appropriate priors. Laplace priors were used to avoid overfitting and enforce sparsity in sparse linear model, which derives sparse coding problem. To further enforce sparsity and reduce over-penalization of sparse coding, each coefficient is assigned with a Bernoulli variable. Therefore, a novel model interpreted from a Bayesian perspective by carrying maximum a posteriori (MAP) is proposed, which turns out to be a combination of sparse coding and counting model. In paper \cite{Lu2013Sparse}, Lu \emph{et al.} ~also consider $L_0$ and $L_1$ norm under a Bayesian perspective. However, considering that there will be occlusion, illumination change and background clutter in tracking, we restraint the noise with $L_1$ norm. Besides, We use an orthogonal dictionary to replace the redundant object templates as similar atoms of redundant templates may cause mistake of coefficients and huge computational complexity. Lastly, We propose closed solution of regularization which is the combination of the $L_0$ norm and $L_1$ norm. However Lu \emph{et al.} ~obtain the approximate solution by using he Greedy Coordinate Descent.
\begin{figure*}[htb!]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.98\textwidth,keepaspectratio]{Fig1.jpg}
\end{tabular}
\end{center}
\caption
{\label{fig:coefficient}
{\bf The comparison of coefficients, optimal candidates and reconstruction.} The top is the coefficients of our method versus unconstrained,
spars coding and sparse counting regularization, respectively. The bottom is the optimal candidates and reconstruction results by using unconstrained, sparse coding, sparse counting and our method under same dictionary, respectively. }
\end{figure*}
Tracking results by using unconstrained regularization, sparse counting, sparse coding and our model under the same dictionary $D$ are shown in Fig.~\ref{fig:coefficient}, respectively. As shown in Fig.~\ref{fig:coefficient}, one can see that the coefficients of unconstrained regularization and sparse coding are actually not sparse and the target object is not tracked well.
Similarly, sparse counting with sparsity coefficients sometimes cannot obtain appropriate linear combination of the orthogonal basis vectors, which will interfere with the tracking accuracy. However, we note that our method is able to reconstruct the object well and find the good candidate, then facilitating the tracking results. We also compare our model with unconstrained regularization, sparse counting, sparse coding over all 50 sequences in benchmark, the precision and success plots are shown in Fig.~\ref{fig:precision_compare}. One can see the parameter setting in the section Experimental Results.
{\flushleft \bf Contributions}: The contributions of this work are threefold.
(1) We propose a sparse coding and counting model from a novel Bayesian perspective for visual tracking. Compared to the state-of-the-art algorithms, the proposed method achieves more reliable tracking results
(2) We propose closed solution of combining the $L_0$ norm and $L_1$ norm based regularization in a unique one.
(3) Although the sparse coding and counting related minimization is an NP-hard problem,we show that the proposed model can be efficiently estimated by the proposed APG method.
This makes our tracking method computationally attractive in general and
comparable in speed with SP method~\cite{Lu/tip13/Wang13} and the accelerated $L_1$ tracker~\cite{APGL1/bao}.
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.48\textwidth,keepaspectratio]{Fig2.jpg}
\end{tabular}
\end{center}
\caption
{ \label{fig:precision_compare}
{\bf Precision and success plots of overall performance comparison among unconstrained regularization, sparse counting, sparse coding and ours for the 50 videos in the benchmark.} The mean precision scores are reported in the legends. }
\end{figure}
\section*{Visual Tracking based on the Particle Filter}
\label{sec: Visual Tracking Algorithm based on the Particle Filter}
In this paper, we employ a particle filter to track the target object. The particle filter provides an estimate of posterior distribution
of random variables related to Markov chain. Given a set of observed image vectors $\mathbf{Y}_t = \{\mathbf{y}_1, \mathbf{y}_2, ..., \mathbf{y}_t\}$ up to the $t$-th frame and target state
variable $\mathbf{x}_t$ that describes the six affine motion parameters, the posterior distribution $p(\mathbf{x}_{t}|\mathbf{Y}_{t})$ based on the Bayesian theorem is estimated by:
\begin{equation}
p(\mathbf{x}_{t}|\mathbf{Y}_{t})\propto p(\mathbf{y}_{t}|\mathbf{x}_{t})\int p(\mathbf{x}_{t}|\mathbf{x}_{t-1})p(\mathbf{x}_{t-1}|\mathbf{Y}_{t-1})d\mathbf{x}_{t-1},
\end{equation}
where $p(\mathbf{y}_t | \mathbf{x}_{t})$ is the observation model that estimates the likelihood of an observed image patch $\mathbf{y}_t$ belonging to the object class, and $p(\mathbf{x}_t | \mathbf{x}_{t-1})$ is the motion model that describes the state transition between consecutive frames.
\textbf{The Motion Model: }The motion model $p(\mathbf{x}_t | \mathbf{x}_{t-1}) = N (\mathbf{x}_t ; \mathbf{x}_{t-1}, \mathbf{\Sigma})$ models the parameters by independent Gaussian distribution around the counterpart in $\mathbf{x}_{t-1}$, where $\mathbf{\Sigma}$ is a diagonal covariance matrix whose elements are the variances of the affine parameters.
In the tracking framework, the optimal target state $\hat{\mathbf{x}}_t$ is obtained by the maximal approximate posterior (MAP) probability: $\hat{\mathbf{x}}_t = \arg\max_{\mathbf{x}_t^i}p(\mathbf{x}_t^i | \mathbf{Y}_t )$,
where $\mathbf{x}_t^i$ indicates the $i$-th sample of the state $\mathbf{x}_t$.
\textbf{The observation model: }In this paper, we assume that the
tracked target object is generated by a subspace
(spanned by $\mathbf{D}$ and centered at ${\bm{\mu}}$) with corruption (i.i.d Gaussian Laplacian noise),
$$\mathbf{y} = \mathbf{D}{\bm{\alpha}}+\mathbf{\epsilon}+\mathbf{e},$$
where $\mathbf{y}\in \mathbb{R}^{N}$ denotes an observation vector centered at ${\bm{\mu}}$, the columns of $\mathbf{D}=\{\d_1, \d_2,\ldots, \d_K\} \in \mathbb{R}^{N \times K}$ are orthogonal basis vectors of the subspace, ${\bm{\alpha}}$ indicates the coefficients of basis vectors, $\mathbf{\epsilon}$ and $\mathbf{e}$ stand for the Gaussian noise and Laplacian noise vector respectively. the Gaussian component models small dense noise and the Laplacian one aims to handle outliers. As proposed by~\cite{Lu/CVPR13/Wang13}, under the i.i.d
Gaussian-Laplacian noise assumption, the distance between
the vector $\mathbf{y}$ and the subspace $(\mathbf{D}, {\bm{\mu}})$ is the least soft threshold squares distance:
$$d({\bm{\alpha}}, \mathbf{e}) = \min_{{\bm{\alpha}}, \mathbf{e}}\frac{1}{2}\| \mathbf{y}-\mathbf{D}{\bm{\alpha}} - \mathbf{e}\|_2^2+ \lambda\|\mathbf{e}\|_1.$$
Thus, for each observation $\mathbf{y}_t$ corresponding to a predicted state $\mathbf{x}_t$, the observation model $p(\mathbf{y}_t | \mathbf{x}_{t})$ that is set to be
\begin{equation}\label{eq:observation-likelihood}
p(\mathbf{y}_t | \mathbf{x}_{t}) = \exp(-\tau d({\bm{\alpha}}^*, \mathbf{e}^*)),
\vspace{-0.66ex}
\end{equation}
where ${\bm{\alpha}}^*$ and $\mathbf{e}^*$ are the optimal solution of Eq.~\eqref{eq:tracking modal} which will be introduced in detail
in next section, and $\tau$ is a constant controlling the shape of the Gaussian kernel.
\textbf{Model Update: }It is essential to update the observation model for handling appearance change of the target in visual tracking.
Since the error term $\mathbf{e}$ can be used to identify some outliers (\emph{e.g.}, Laplacian noise, illumination),
we adopt the strategy proposed by~\cite{Lu/CVPR13/Wang13} to update the appearance model
using the incremental PCA with mean update~\cite{IVT/ijcv/RossLLY08} as follows,
\begin{equation}
\label{eq: model-update}
y_i = \left\{\begin{array}{ll}y_i, & \mbox{}\ e_i = 0,\\ \mu_i, & \mbox{}\
$otherwise$, \end{array}\right.
\end{equation}
where $y_i$, $e_i$, and $\mu_i$ are the i-th elements of $\mathbf{y}$, $\mathbf{e}$, and ${\bm{\mu}}$, respectively,
${\bm{\mu}}$ is the mean vector computed the same as~\cite{IVT/ijcv/RossLLY08}.
\section*{Object Representation under Bayesian Framework}
Based on the discussion in aforementioned Section, If $\mathbf{y}$ is viewed as the vectorized target region, it can be represented by an image subspace with corruption,
\begin{equation}\label{eq:1}
\mathbf{y} = \mathbf{D}{\bm{\alpha}}+\mathbf{\epsilon}+\mathbf{e}.
\end{equation}
~\cite{panl0} shows that sparse counting can remove redundant features (\emph{e.g.}, background portions) while selecting useful parts in the subspace. However, sparse counting will cause unstable solutions because of its nonconvexity and discontinuity. Sparse coding may produce over-penalization, despite its good stability.
Considering that Bayesian framework has the capacity to encode prior knowledge and to make valid estimation of uncertainty, a novel model combining sparse coding and sparse counting is proposed for visual tracking. The model is
\begin{equation}\label{eq:tracking modal}
\min_{{\bm{\alpha}}\,\mathbf{e}}\frac{1}{2}||\mathbf{y}-\mathbf{D}{\bm{\alpha}}-\mathbf{e}||_2^2+\beta||\mathbf{e}||_1+\lambda\gamma||{\bm{\alpha}}||_1+\lambda(1-\gamma)||{\bm{\alpha}}||_0,
\end{equation}
where $\mathbf{D}^\top \mathbf{D} = \mathbf{I}$, $\|\cdot\|_0$ denotes the $L_0$ norm which counts the number of non-zero elements,
$\|\cdot\|_2$ and $\|\cdot\|_1$ denote $L_2$ and $L_1$ norms, respectively,
$\gamma$, $\lambda$ and $\beta$ are regularization parameters, and $\mathbf{I}$ is an identity matrix.
The term $\|\mathbf{e}\|_1$ is used to reject outliers (\emph{e.g.}, occlusions), while $\|{\bm{\alpha}}\|_0$ and $\|{\bm{\alpha}}\|_1$ are used to select the useful subspace features.
Next we will introduce the aforementioned model under Bayesian framework in detail.
The joint posterior distribution of ${\bm{\alpha}}, \r,\mathbf{e}$ and $\sigma^2$ based on the Bayesian theorem can be written as
\begin{align}\label{eq:PDF}
\begin{split}
&p({\bm{\alpha}},\r,\mathbf{e},\sigma^2|\mathbf{D},\mathbf{y},\tilde{\mu},\tau_1,\tau_2,\kappa,\hat{\sigma})\propto \\ &p(\mathbf{y}|,\mathbf{D},{\bm{\alpha}},\r,\mathbf{e},\sigma^2)p({\bm{\alpha}}|\sigma^2,\tilde{\mu})p(\r|\kappa)p(\mathbf{e}|\hat{\sigma})p(\sigma^2|\tau_1,\tau_2),\\
\end{split}
\end{align}
where $p(\mathbf{y}|\mathbf{D},{\bm{\alpha}},\r,\mathbf{e},\sigma^2)$, $p({\bm{\alpha}}|\sigma^2,\tilde{\mu})$, $p(\r|\kappa)$, $p(\mathbf{e}|\hat{\sigma})$, $p(\sigma^2|\tau_1,\tau_2)$, denote the priors on the noisy vectorized target region, the coefficient vector ${\bm{\alpha}}=[\alpha_{1},\alpha_{2},\ldots,\alpha_{K}]$, the index vector $\r = [r_{1}, r_{2},\ldots, r_{K}]$ ($r_{l} = \mathbb{I}(\alpha_{l} \neq 0),l = 1, 2,\ldots, K$), the Laplacian noise, and the noise level, respectively. In Eq.~\eqref{eq:PDF}, the parameters
$\tilde{\mu}$, $\tau_1$, $\tau_2$, $\hat{\sigma}$, and $\kappa$ are the relevant constant parameters of the priors.
With the definition of the index variable $r_{l}$ , Eq.~\eqref{eq:1} can be rewritten as
\begin{equation}
y_{j}=\sum_{l=1}^K d_{jl}r_{l}\alpha_{l}+\epsilon_{j}+e_{j},~j=1,2,\ldots,N.
\end{equation}
We generally assume that the noise $\epsilon_{j}$ follows the Gaussian distribution, $\emph{i.e.},~p(\epsilon_{j})= N(0,\sigma^{2})$. We treat the Laplacian noise term $e_{j}$ as missing values with the same Laplacian prior. Therefore, the Prior $p(\mathbf{y}|,\mathbf{D},{\bm{\alpha}},\r,\mathbf{e},\sigma^2)$ has the follow distribution:
\begin{align}
\begin{split}\label{eq:y}
&p(\mathbf{y}|,\mathbf{D},{\bm{\alpha}},\r,\mathbf{e},\sigma^2)=\\
&\prod_{j=1}^NP(y_{j}|,\d_j,{\bm{\alpha}},\r,e_{j},\sigma^{2})\\
&=\prod_{j=1}^NN(\sum_{l=1}^K d_{jl}r_{l}\alpha_{l}+e_{j},\sigma^{2})\\.
\end{split}
\end{align}
To enforce sparsity, the coefficients ${\bm{\alpha}}$ are assumed to follow Laplace distribution.
\begin{equation}\label{eq:balpha}
p({\bm{\alpha}}|\sigma^2,\tilde{\mu})=\prod_{l=1}^Kp(\alpha_{l}|\sigma^2,\tilde{\mu})= \prod_{l=1}^K\frac{1}{2\sigma^2\tilde{\mu}^{-1}}\exp(-\frac{|\alpha_{l}|}{\sigma^2\tilde{\mu}^{-1}}).
\end{equation}
Our goal is to remove redundant features while preserving the useful parts in the dictionary.
As Laplace priors resulting sparse coding may lead to over penalization on the large coefficients, we assume the index variable
$r_{l}$ of each coefficient $\alpha_{l}$ to be a Bernoulli variable to enforce sparsity and reduce over penalization.
\begin{equation}\label{eq:r}
p(\r|\kappa)=\prod_{l=1}^K\kappa^{r_{l}}(1-\kappa)^{1-r_{l}},
\end{equation}
where $\kappa \leq 1/2$. Here, the Bernoulli prior on $r_{l}$ means that $r_{l}$ will have probability $\kappa$ to be 1 and $1-\kappa$ to be 0, if the prior information is known.
The noise $e_{j}$ is aims at handling outliers, so it follows Laplace distribution:
\begin{equation}\label{eq:e}
p(\mathbf{e}|\hat{\sigma})=\prod_{j=1}^{N}p(e_{j}|\hat{\sigma})= \prod_{j=1}^{N}\frac{1}{2\hat{\sigma}}\exp(-\frac{|e_{j}|}{\hat{\sigma}}).
\end{equation}
The variances of noises are assigned with Inverse Gamma prior as follow:
\begin{equation}\label{eq:sigma}
p(\sigma^2|\tau_1,\tau_2)=\frac{\tau_2^{\tau_1}}{\Gamma(\tau_1)}\sigma^{-2(\tau_1+1)}\exp(-\frac{\tau_2}{\sigma^2}),
\end{equation}
where $\Gamma(\cdot)$ denotes the gamma function.
Then, the optimal ${\bm{\alpha}}, \r,\mathbf{e},\sigma^2$ are obtained by the MAP probability. After taking the negative logarithm, the formula is
\begin{equation}
\begin{aligned}
&({\bm{\alpha}}^*,\r^*,\mathbf{e}^*,\sigma^{*2})=\arg\min\limits_{{\bm{\alpha}},\r,\mathbf{e},\sigma^2} \\
& \{-2logp({\bm{\alpha}},\r,\mathbf{e},\sigma^2|\mathbf{D},\mathbf{y},\tilde{\mu},\tau_1,\tau_2,\kappa,\hat{\sigma})\}.
\end{aligned}
\end{equation}
Combining the aforementioned Eq.~\eqref{eq:PDF}, Eq.~\eqref{eq:y}, Eq.~\eqref{eq:balpha}, Eq.~\eqref{eq:r}, Eq.~\eqref{eq:e}, Eq.~\eqref{eq:sigma}, we have
\begin{align}\label{eq:2}
\begin{split}
&-2logp({\bm{\alpha}},\r,\mathbf{e},\sigma^2|\mathbf{D},\mathbf{y},\tilde{\mu},\tau_1,\tau_2,\kappa,\hat{\sigma})=\\
&\frac{1}{\sigma^2}\sum_{j=1}^N(y_{j}-\sum_{l=1}^{K}d_{jl}r_{l}\alpha_{l}-e_{j})^2+\frac{1}{\sigma^2}\frac{2\sigma^2}{\hat{\sigma}}\sum_{j=1}^{N}|e_{j}|\\
&+\frac{2\tilde{\mu}}{\sigma^2}\sum_{l=1}^K|\alpha_{l}|+(2N+2K+2\tau_1+2)log\sigma^2+\frac{2\tau_2}{\sigma^2}\\
&+\sum_{l=1}^{K}r_{l}log\frac{(1-\kappa)^2}{\kappa^2}+const.\\
\end{split}
\end{align}
With fixing $\sigma^2 = 1$, Eq.~\eqref{eq:2} can be rewritten as
\begin{equation}\label{eq:3}
||\mathbf{y}-\mathbf{D} {\bm{\alpha}}-\mathbf{e}||_2^2+2\beta||\mathbf{e}||_1+2\tilde{\mu}||{\bm{\alpha}}||_1+\rho_{\kappa}||{\bm{\alpha}}||_0+const,
\end{equation}
where $\rho_{\kappa}= log (1 - \kappa)^2/\kappa^2, \beta=\sigma^2/\hat{\sigma}$. With $\gamma \in [0, 1], \lambda = \tilde{\mu} + 1/2\rho_{\kappa}$ and $\gamma = 4\tilde{\mu}/(2\tilde{\mu} + \rho_{\kappa})$, Eq. \eqref{eq:3} can be rewritten as
\begin{equation}\label{eq:4}
\frac{1}{2}||\mathbf{y}-\mathbf{D}{\bm{\alpha}}-\mathbf{e}||_2^2+\beta||\mathbf{e}||_1+\lambda\gamma||{\bm{\alpha}}||_1+\lambda(1-\gamma)||{\bm{\alpha}}||_0+const,
\end{equation}
By observing the objective function in Eq.~\eqref{eq:4}, it can be found that the essential regularization in Eq.~\eqref{eq:4} is a combination of the sparse coding and the sparse counting.
With a fixed appropriate orthogonal dictionary D, Eq.~\eqref{eq:4} can be written as an optimization problems Eq.~\eqref{eq:tracking modal}.
\section*{Theory of Fast Numerical Algorithm}
As we know, APG is an excellent algorithm for convex programming~\cite{APG/lin,APG/Tseng} and has been used in visual tracking. In this section, we propose a fast numerical algorithm for solving the proposed nonconvex and nonsmooth model by using APG approach. The experimental results show that it can converge to a solution quickly and achieve attractive performance.
Besides, the closed solution of the combining $L_0$ and $L_1$ based regularization is provided.
\subsection*{APG Algorithm for Solving Eq.~\eqref{eq:tracking sub-modal}}
Eq.~\eqref{eq:tracking modal} contains two subproblem: one is solving ${\bm{\alpha}}$ given fixed $\mathbf{e}$, the other one is solving $\mathbf{e}$ given fixed ${\bm{\alpha}}$, the formula is shown as follow
\begin{equation}\label{eq:tracking sub-modal}
\left\{\begin{array}{l}
{\bm{\alpha}}=\arg\min\limits_{{\bm{\alpha}}}\frac{1}{2}||\mathbf{y}-\mathbf{D}{\bm{\alpha}}-\mathbf{e}||_2^2+\lambda\gamma||{\bm{\alpha}}||_1+\lambda(1-\gamma)||{\bm{\alpha}}||_0,\\
\mathbf{e}=\arg\min\limits_{\mathbf{e}}\frac{1}{2}||\mathbf{y}-\mathbf{D}{\bm{\alpha}}-\mathbf{e}||_2^2+\beta||\mathbf{e}||_1.
\end{array}\right.
\end{equation}
Solving Eq.~\eqref{eq:tracking sub-modal} is an NP-hard problem because it involves a discrete counting metric.
We adopt a special optimization strategy based on the APG approach~\cite{APG/lin}, which ensures each step be solved easily. In APG Algorithm, we need to solve
\begin{equation}\label{eq:apg-model-minimization-1}
\left\{\begin{array}{l}
{\bm{\alpha}}_{k+1}^*=\arg\min\limits_{{\bm{\alpha}}} \lambda\gamma\|{\bm{\alpha}}\|_1+\lambda(1-\gamma) \|{\bm{\alpha}}\|_0+\frac{L}{2} \| {\bm{\alpha}} - \mathbf{z}_{k+1}^{{\bm{\alpha}}} + \frac{\nabla_{\bm{\alpha}} F(\mathbf{z}_{k+1})}{L}\|_2^2,\\
\mathbf{e}_{k+1}^*= \arg\min\limits_{\mathbf{e}}\beta\| \mathbf{e} \|_1 + \frac{L}{2}\| \mathbf{e} - \mathbf{z}_{k+1}^{\mathbf{e}} + \frac{\nabla_\mathbf{e} F(\mathbf{z}_{k+1})}{L}\|_2^2,
\end{array}\right.
\end{equation}
where $\mathbf{z}_{k+1}=(\mathbf{z}_{k+1}^{{\bm{\alpha}}},\mathbf{z}_{k+1}^{\mathbf{e}})$, $\nabla_{{\bm{\alpha}}} F({\bm{\alpha}}, \mathbf{e}) = \mathbf{D}^\top(\mathbf{D}{\bm{\alpha}} + \mathbf{e} - \mathbf{y})$,
$\nabla_\mathbf{e} F({\bm{\alpha}}, \mathbf{e}) = \mathbf{e} - (\mathbf{y} - \mathbf{D}{\bm{\alpha}})$, and $L$ is a Lipschitz constant.
The solutions of Eq.~\eqref{eq:apg-model-minimization-1} can be obtained by
\begin{equation}\label{eq:apg-model-minimization-1-sol}
\left\{\begin{array}{l}
{\bm{\alpha}}_{k+1}^*= \mathcal{E}_{(\lambda\gamma/L,\lambda(1-\gamma)/L)}\left( \mathbf{z}_{k+1}^{{\bm{\alpha}}} - \frac{\nabla_{\bm{\alpha}} F(\mathbf{z}_{k+1})}{L}\right),\\
\mathbf{e}_{k+1}^{*}= \mathcal{S}_{\beta/L}\left( \mathbf{z}_{k+1}^{\mathbf{e}} - \frac{\nabla_\mathbf{e} F(\mathbf{z}_{k+1})}{L}\right),
\end{array}\right.
\end{equation}
where $\mathcal{S}_{\theta}(y) = \text{sign}(y)\max(|y| - \theta,0)$,
and $\mathcal{E}_{(\delta,\eta)}(y)$ is defined as
\begin{equation}
\mathcal{E}_{(\delta,\eta)}(y)=\left\{
\begin{array}{ll}
y-\delta, & \mbox{}\ y > \delta+\sqrt{2\eta},\\
y+\delta, & \mbox{}\ y < -\delta-\sqrt{2\eta},\\
0,& \mbox otherwise.
\end{array}\right.
\end{equation}
The numerical algorithm for solving Eq.~\eqref{eq:tracking sub-modal} is summarized in Algorithm~\ref{alg:our-apg-alogrithm}. Due to the orthogonality of $\mathbf{D}$, Algorithm~\ref{alg:our-apg-alogrithm} converges fast, and its computation cost does not increase compared to the solver of $L_1$ regularized model.
\begin{algorithm}
\caption{Fast numerical algorithm for solving Eq.~\eqref{eq:tracking sub-modal}}
\label{alg:our-apg-alogrithm}
\begin{algorithmic}
\STATE \textbf{Initialize:} Set initial guesses ${\bm{\alpha}}_0 = {\bm{\alpha}}_{-1} = \textbf{0}$, $\mathbf{e}_0 = \mathbf{e}_{-1} = \textbf{0}$, and $t_0 = t_{-1} = 1$.
\STATE \textbf{while} not convergence or termination \textbf{do}
\STATE \textbf{Step 1:} $\mathbf{z}_{k+1}^{{\bm{\alpha}}}:= {\bm{\alpha}}_k + \frac{t_{k-1} -1 }{t_{k}}({\bm{\alpha}}_k - {\bm{\alpha}}_{k-1})$;
\STATE \textbf{Step 2:} $\mathbf{z}_{k+1}^{\mathbf{e}}:= \mathbf{e}_k + \frac{t_{k-1} -1 }{t_{k}}(\mathbf{e}_k - \mathbf{e}_{k-1})$;
\STATE \textbf{Step 3:} ${\bm{\alpha}}_{k+1}= \mathcal{E}_{(\lambda\gamma/L,\lambda(1-\gamma)/L)}\left( \mathbf{z}_{k+1}^{{\bm{\alpha}}} - \frac{\nabla_{\bm{\alpha}} F(\mathbf{z}_{k+1})}{L}\right)$;
\STATE \textbf{Step 4:} $\mathbf{e}_{k+1}= \mathcal{S}_{\beta/L}\left( \mathbf{z}_{k+1}^{\mathbf{e}} - \frac{\nabla_\mathbf{e} F(\mathbf{z}_{k+1})}{L}\right)$;
\STATE \textbf{Step 5:} $t_{k+1}:= \frac{1+\sqrt{1+4t_k^2}}{2}$, $k\leftarrow k+1$.
\STATE \textbf{end while}\\
\end{algorithmic}
\end{algorithm}
\subsection*{Closed Solution of combining $L_1$ and $L_0$ regularization}
This subsection mainly focus on a sparse combinatory model which combines $L_0$ and $L_1$ norm together as the regularizer term
\begin{equation}\label{eq:problem}
\min_x \frac{1}{2}(x-y)^2+\delta|x|+\eta|x|_0,
\end{equation}
where $x,y\in\mathbb{R}^1$, and $|x|$ denotes $L_0$ norm: if $x = 0$, then$|x|_0=0$, and $|x|_0=1$, otherwise.
\vspace{2ex}\noindent{\footnotesize\textbf{lemma.}
\label{lemma:1}
The optimal solution $x^*$ of the Eq.~\eqref{eq:problem} is defined as
\begin{equation}
\label{eq: proof}
x^* = \left\{
\begin{array}{ll}
y-\delta, & \mbox{}\ y > \delta+\sqrt{2\eta},\\
y+\delta, & \mbox{}\ y < -\delta-\sqrt{2\eta},\\
0,& \mbox otherwise.
\end{array}\right.
\end{equation}}
The proof can be found in Supporting Information. If $x\in\mathbb{R}^N$, the Eq.~\eqref{eq:problem} changes into
\begin{equation}\label{eq:regression}
\min_\mathbf{x} \frac{1}{2}||\mathbf{x}-\mathbf{y}||_2^2+\delta||\mathbf{x}||_1+\eta||\mathbf{x}||_0,
\end{equation}
where $||\mathbf{x}||_1=\sum_{i=1}^N|x_i|$ and $||\mathbf{x}||_0=\sum_{i=1}^N|x_i|_0$. It is obvious that Eq.~\eqref{eq:problem} can be turned into
\begin{equation}
\min_{x_i} \sum_{i=1}^N \frac{1}{2}(x_i-y_i)^2+\delta|x_i|+\eta|x_i|_0.
\end{equation}
So it can be seen as a sequence of optimization of $x_i,i=1,\ldots,n$, and each can be solved by Lemma.
More analysis about combination of $L_1$ and $L_0$ regularization can be found in Supporting Information.
\subsection{Analysis of the combinatory model Eq.~\eqref{eq:regression}}
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.49\textwidth,keepaspectratio]{Fig9.jpg}\label{fig:Regularized Regression}
\end{tabular}
\end{center}
\caption{{\bf Analysis about combination of $L_1$ and $L_0$ regularization.} (a) shows the closed solutions of linear regression, $L_0$, $L_1$, $L_0+L_1$ regularized regression, respectively. (b) shows the sparsity threshold changes of $L_0$, $L_1$ and $L_0+L_1$ regularized regression, respectively.}\label{fig: Regression}
\end{figure}
In Eq.~\eqref{eq:regression}, if we set $\delta=0$ and $\eta=0$, the model degenerates to the linear regression. If we set $\delta=0$, Eq.~\eqref{eq:regression} reduces to $L_0$ regularized regression, while becoming $L_1$ regularized regression when $\eta=0$. S2 Fig.~\ref{fig: Regression} (a) shows the closed solutions of these four cases. We set $\delta=\eta=0.5$ in Eq.~\eqref{eq:regression} ($L_0+L_1$ regularized regression), $\eta=1$ in $L_0$ regularized regression, and $\delta=1$ in $L_1$ regularized regression. We note that $L_0+L_1$ regularized regression has the same sparsity as $L_0$ regularized regression, while causing little over penalization than $L_1$ regularized regression. In S2 Fig.~\ref{fig: Regression} (b), sparsity threshold changes of $L_0$, $L_1$ and $L_0+L_1$ regularized regression are shown, respectively. When $\delta = 1-\eta$ changes from 0 to 1, the sparsity threshold of $L_0+L_1$ varies from that of $L_0$ to the threshold of $L_1$. Besides, it is obvious that the threshold of $L_0+L_1$ is larger than those of $L_0$ and $L_1$ in interval $(0, 0.8]$.
\section*{Orthogonal Dictionary learning for Visual Tracking}
In this section, we demonstrate dictionary learning in detail through three parts: dictioanry initialization, orthogonal dictionary update and dictionary reinitialization.
\textbf{Dictioanry Initialization:}\label{Dictioanry Initialization}
There are two schemes to initialize the orthogonal dictionary, one is doing PCA for the set of initial first $k$ frames $\mathbf{Y}_k$, the other is doing RPCA for $\mathbf{Y}_k$. When initial frames do not undergo corruption (\emph{e.g.},~occlusion or illumination), we do PCA for $\mathbf{Y}_k$ instead of RPCA. The whole process of PCA is doing skinny SVD for $\mathbf{Y}_k$ and get the basis vectors of column space as the initial dictionary. However, when initial frames have large sparse noise, RPCA is selected to get the intrinsic low-rank features $\mathbf{Z}_k$, which can be obtained by solving~\cite{LRF/Zhang14}:
\begin{equation}
\min\limits_{\mathbf{Z}_k,\mathbf{E}_k} \|\mathbf{Z}_k\|_* + \lambda\|\mathbf{E}_k\|_1, \ s.t. \ \mathbf{Y}_k=\mathbf{Z}_k + \mathbf{E}_k.\label{eq:rpca_ini}
\end{equation}
When solving Eq.~\eqref{eq:rpca_ini}, the skinny SVD of $\mathbf{Z}_k$ is readily available: $\mathbf{Z}_k=\mathbf{U}_k\Sigma_k\mathbf{V}_k^T$, and $\mathbf{D}=\mathbf{U}_k$ is the initial orthogonal dictionary.
Fig. \ref{fig:process/PCA_VS_RPCA} (a) shows that PCA initialization and RPCA initialization both perform well when the initial first $k$ frames have little noise. The initial frames is generally clean, therefore, we choose PCA initialization as the default.
\begin{figure}[htbp]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.49\textwidth]{Fig3.jpg}
\end{tabular}
\end{center}
\caption{{\bf Comparison of PCA process to RPCA process.} The upper portion of the image is the tracking frame. the middle of the image consists of three sub-pictures, the left is the mean image, the middle is the reconstruction result, and the right is the Lapalace noise. the bottom of the image is the top ten basis vectors of dictionary. (a) shows the tracking results of PCA and RPCA dictionary initialization. The tracking performance with and without RPCA reinitialization is shown in (b). }
\label{fig:process/PCA_VS_RPCA}
\end{figure}
\textbf{Orthogonal Dictionary Update:}\label{dictionary update}
As the appearance of a target may change drastically, it is necessary to update the orthogonal dictionary $\mathbf{D}$. Here we adopt an incremental PCA algorithm \cite{Levey2000SKL} to update the dictionary.
\textbf{ Dictionary reinitialization:}\label{Reinitializing Dictionary}
When the tracker is prone to drift, dynamically reinitializing dictionary to obtain the intrinsic subspace features is needed. We adopt the strategy proposed by~\cite{LRF/Zhang14}. The reinitialization is performed at $t$-th frame if $\sigma=\|\mathbf{e}_t\|_0/len(\mathbf{e}_t) > thr$, where $\mathbf{e}_t$ is the noise item at $t$-th frame, $len(.)$ is the length of vector, and $thr>0$ is a threshold parameter (generally 0.5). If $\sigma > thr $, we reinitialize the dictionary in the same way as initialization of dictionary by doing RPCA, but $\mathbf{Y}_t$ in Eq.~\eqref{eq:rpca_ini} is different. Here, $\mathbf{Y}_t$ consists of optimal candidate observations respectively
from the initial $n$ (generally 10) frames and the latest $t-n$ frames (we set $t=30$).
Fig.~\ref{fig:process/PCA_VS_RPCA} (b) compares the tracking performance within and without RPCA reinitialization when the object undergoes variable illumination. After reinitializing dictionary, our tracker retracks the object, so reinitializing dictionary is efficient to improve the reconstruction ability. In Algorithm~\ref{alg:tracker}, we summarize the overall tracking process for frame $t$.
\section*{Experimental Results}
In this section, we compare the performance of our proposed tracker with several state-of-the-art tracking algorithms, such as TLD~\cite{TLD/pami/KalalMM12}, IVT~\cite{IVT/ijcv/RossLLY08}, ASLA~\cite{ALSA/cvpr/Lu/JiaLY12}, $L_1$APG~\cite{APGL1/bao}, MTT~\cite{MTT/ijcv/ZhangGLA13}, SP~\cite{Lu/tip13/Wang13}, SPOT~\cite{zhang2013structure}, FOT~\cite{vojivr2014enhanced}, SST~\cite{zhang2015structural}, SCM~\cite{SCM/cvpr/Lu/ZhongLY12}, MIL~\cite{MIL/cvpr/BabenkoYB09}, and Struck~\cite{Struck/iccv/HareST11}, on a benchmark~\cite{survey2/cvpr/Yang} with 50 challenge video sequences. Our tracker is implemented in MATLAB and runs at 4.2 fps on an Intel 2.53 GHz Dual-Core CPU with 8GB memory, running Windows 7 and Matlab (R2013b). We empirically set $\eta = 0.1$, $\lambda = 0.5$, $\gamma = 0.1$, $\tau = 0.05$ and the Lipschitz constant L = 2. Before solving Eq.~\eqref{eq:tracking modal}, all the candidates $\mathbf{y}$ are centralized.
Considering the efficiency, the updated orthogonal dictionary $\mathbf{D}$ is taken columns corresponding to the $16$ largest eigenvalues of PCA or RPCA, 600 particles are adopted, and the model is incrementally updated every $5$ frames.
In the following, we present both qualitative and quantitative comparisons of above mentioned methods.
\begin{algorithm}[hb]
\caption{Robust Visual Tracking Using Our tracker}
\label{alg:tracker}
\begin{algorithmic}
\STATE {\bfseries Initialization:} Initialize orthogonal dictionary $\mathbf{D}$ by performing PCA on ${{\bf{Y}}_{k}}$.
\STATE {\bfseries Input:} State $\mathbf{x}_{t-1}$ ($t>k$) and orthogonal dictionary $\mathbf{D}$.
\STATE {\bfseries Step 1:} Draw new samples $\mathbf{x}_t^i$ from $\mathbf{x}_{t-1}$ and
obtain corresponding candidates $\mathbf{y}_t^i$.
\STATE {\bfseries Step 2:} Obtain $\mathbf{\alpha}_{t}^i$ and $\mathbf{e}_{t}^{i}$ using (\ref{eq:tracking sub-modal}).
\STATE {\bfseries Step 3:} For each candidate, calculate the observation probability $p(\mathbf{y}_t^i|\mathbf{x}_{t}^i)$ using (\ref{eq:observation-likelihood}).
\STATE {\bfseries Step 4:} Find the tracking result patch $\mathbf{y}_t^*$ with the maximal observation likelihood and its corresponding noise $\mathbf{e}_{t}^*$.
\STATE {\bfseries Step 5:} perform an
incremental PCA algorithm to update the orthogonal dictionary $\mathbf{D}$ every five frames. If $\sigma>thr$, reinitializing Dictionary at $t$-th frame using \eqref{eq:rpca_ini}.
\STATE {\bfseries Output:} State $\mathbf{x}_t^*$ and corresponding image patch; orthogonal dictionary $\mathbf{D}$.
\end{algorithmic}
\end{algorithm}
\subsection*{Qualitative Evaluation}
Fig.~\ref{fig:all} were taken the frames of the 50 videos to show the Qualitative results for our method, compared with the top-performing SP and SST. We choose some examples from part of 50 sequences to illustrate the effectiveness of our method. Fig.~\ref{fig:trackingResult} shows the visualization results.
\begin{figure*}[htb]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.99\textwidth,keepaspectratio]{Fig4.jpg}
\end{tabular}
\end{center}
\caption{\label{fig:all}{\bf Qualitative results for our method, compared with SP and SST.} Reprinted from~\cite{survey2/cvpr/Yang} under a CC BY license, with permission from Yi Wu, original copyright 2013.}
\end{figure*}
\textbf{Heavy Occlusion: }Fig.~\ref{fig:trackingResult} (a) and (b) show four challenging sequences with heavy occlusion.
In \emph{Faceocc1} and \emph{Faceocc2}, the targets undergo
with heavy occlusion and in-plane rotation, it can be seen that our method outperforms the other tracking algorithms.
\emph{Freeman4} and \emph{David3} demonstrate that the proposed
method can capture the accurate location of objects in terms of position, and scale when the target undergoes severe occlusion (\emph{e.g.}, \emph{Freeman4} \#0144 and \emph{David3} \#0085). However, IVT, $L_1$APG, MIL, SP, SCM, ASLA,
TLD, SPOT, FOT, SST, MTT, and Struck methods drift away from the target object when occlusion occurs. For these four sequences, the IVT method performs poorly since conventional PCA is not robust to occlusions. Although $L_1$APG and SP utilize sparsity to model outliers, it is observed that their occlusion detection are not stable when drastic change of appearance happens. In contrast, our method is robust to heavy occlusion. This is because our combination of $L_0$ and $L_1$ regularized appearance model can exactly reconstruct the object.
\begin{figure*}[htb!]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.99\textwidth,keepaspectratio]{Fig5.jpg}
\end{tabular}
\end{center}
\caption{\label{fig:trackingResult}{\bf Sampled tracking results of evaluated algorithms on fourteen challenging image sequences.} Reprinted from~\cite{survey2/cvpr/Yang} under a CC BY license, with permission from Yi Wu, original copyright 2013.}
\end{figure*}
\textbf{Fast Motion: }Fig.~\ref{fig:trackingResult} (c) show
the sequences \emph{Boy} and \emph{Jumping} with fast motion. It is difficult to
predict the locations of the tracked objects when they undergo
abrupt motion.
In \emph{Boy}, the captured images are blurred seriously, but Struck and our method track the target faithfully throughout the images. IVT, MTT, ALSA, SCM and SST methods drift away seriously.
We note that most of the other trackers have drift problem due to the abrupt motion in sequence~\emph{Jumping}. In contrast, the SST and our method successfully track the target for whole video.
\textbf{Drastic Pose, Scale and Illumination Changes: }In Fig.~\ref{fig:trackingResult} (d) and (e), we test five challenging
sequences with drastic pose, scale and illumination change.
\emph{Fish} and \emph{Tiger1} chips contain significant illumination variation. We can see that the $L_1$APG, MTT, and MIL methods are less effective in these cases (\emph{e.g.}, \emph{Fish} \#0305 and \emph{Tiger1} \#0240).
In \emph{Singer2} and \emph{Jogging-2}, other trackers drift away when objects under variable illumination, and pose variation (\emph{e.g.}, \emph{Singer2} \#0110 and \emph{Jogging-2} \#0100 ), however, our method still performs well. Our method also achieves good performance in \emph{CarScale} with scale variation (\emph{e.g.}, \emph{CarScale} \#0204).
For subspace-based approaches, they may fail to update the appearance model as the calculation of coefficients in their models may have redundant background features. Our method can successfully adapt to variable drastic changes since the combination of sparse coding and sparse counting is not merely stable but also applicable to obtain the intrinsic features of the subspace.
\textbf{Background Clutters: }{Fig.~\ref{fig:trackingResult} (f)} demonstrates the tracking
results in \emph{Deer}, \emph{Baskerball}, and \emph{Football} with background clutter.
\emph{Baskerball} is a difficult sequence because it contains cluttered background, illumination change, heavy occlusion and non-rigid pose variation. Unless our tracker, none of the compared algorithms can work well on it(\emph{e.g.}, \emph{Baskerball} \#0486 and \#0614).
As shown in \emph{Deer} and \emph{Football}, our tracker performs relatively well (\emph{e.g.}, \emph{Deer} \#0031 and \emph{Football} \#304) as it has excluded background clutters in the sparse errors, but TLD, FOT, and MIL fail.
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.49\textwidth,keepaspectratio]{Fig6.jpg}
\end{tabular}
\end{center}
\caption
{ \label{fig: precision}
{\bf Precision and success plots over all the 50 sequences.} The mean precision scores are reported in the legends. }
\end{figure}
\subsection*{Quantitative Evaluation}
\begin{table*}[htbp]
\centering
\footnotesize
\caption{Average Overlap Rate (in pixels) and average frame per second (FPS). The best and the second results are shown in \textcolor[rgb]{1.00,0.00,0.00}{\textbf{BOLD}} fonts and \textcolor[rgb]{0.00,0.00,1.00}{\textbf{BOLD}} fonts, respectively.} \label{tab:ORE}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&{TLD} &{IVT} &{ASLA} &{$L_1$APG} &{MTT} &{SP} &{SPOT}&{FOT}&{SST}&{SCM} &{MIL} &{Struck} &{Ours} \\ \hline
Faceocc1 &0.58&0.73&0.32&0.76&0.70&0.79&0.74&0.60&0.79&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.79}}&0.60&0.73&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.80}}\\ \hline
Faceocc2 &0.62&0.73&0.65&0.69&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.75}} &0.59&0.69&0.64&0.63&0.73&0.67&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.79}}&0.69\\ \hline
Freeman4 &0.22&0.15&0.13&0.34&0.22&0.17&0.01&0.11&0.18&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.26}}&0.05&0.17&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.41}}\\ \hline
David3 &0.10&0.48&0.43&0.38&0.10&0.46&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.77}}&0.41&0.30&0.41&0.54&0.29&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.73}}\\ \hline
Boy &0.66&0.26&0.37&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.73}}&0.50&0.36&0.57&0.64&0.36&0.38&0.49&0.76&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.81}}\\ \hline
Jumping &0.66&0.12&0.23&0.15&0.10&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.70}}&0.01&0.20&0.16&0.62&0.12&0.52&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.71}}\\ \hline
Fish &0.81&0.77&0.85&0.34&0.16&0.83&0.83&0.78&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.86}}&0.75&0.45&0.85&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.87}}\\ \hline
Tiger1 &0.38&0.10&0.29&0.31&0.26&0.10&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.70}}&0.19&0.16&0.16&0.12&0.15&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.61}}\\ \hline
Singer2 &0.22&0.04&0.04&0.04&0.04&0.04&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.75}}&0.21&0.04&0.17&0.51&0.04&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.62}}\\ \hline
Jogging-2 &0.66&0.14&0.14&0.15&0.13&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.73}}&0.20&0.12&0.12&0.73&0.14&0.20&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.74}}\\ \hline
CarScale &0.45&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.63}}&0.61&0.50&0.49&0.60&0.01&0.35&0.55&0.59&0.41&0.41&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.81}}\\ \hline
Deer &0.60&0.03&0.03&0.60&0.61&0.72&0.72&016&0.62&0.07&0.12&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.74}}&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.82}}\\ \hline
Basketball&0.02&0.11&0.39&0.23&0.19&0.23&0.01&0.17&0.20&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.46}}&0.22&0.20&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.63}}\\ \hline
Football &0.49&0.56&0.53&0.55&0.58&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.69}}&0.01&0.55&0.40&0.49&0.59&0.53&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.59}}\\ \hline
Average &0.46&0.34&0.36&0.41&0.34&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{0.50}}&0.43&0.37&0.39& 0.44&0.39&0.41&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{0.70}}\\ \hline \hline
FPS &21.74&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{27.83}}&7.48&2.47&0.99&2.35&--&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{376.48}}&2.12&
0.37&28.06&10.01&4.27\\ \hline
\end{tabular}
\end{table*}
We use two metrics to evaluate the proposed algorithm with other state-of-the-art methods.
The first metric is the center location error measured with manually labeled ground truth data.
The second one is the overlap rate, i.e., $score = \frac{area(R_T\bigcap R_G)}{area(R_T\bigcup R_G)}$, where $R_T$ is the tracking bounding box and $R_G$ is the ground truth bounding box. The larger average scores mean more accurate results.
Table \ref{tab:ORE} shows the average overlap rates. Table \ref{tab:CEE} reports the average center location errors (in pixels) where a smaller average error means a more accurate result. As can be seen from the table, the most sequences generated by our method have lower average error and higher overlap rate values. We provide the precision and success plots in Fig.~\ref{fig: precision} to evaluate our performance over all the 50 sequences. The evaluation parameters are set as default in~\cite{survey2/cvpr/Yang}.
We note that the our algorithm performs well for the videos with occlusion, deformation,
in plane rotation, and out of plane rotation based on the precision
metric and the success rate metric as shown in Fig.~\ref{fig: center error}
and Fig.~\ref{fig: overlap rate} respectively. Both table and figures show that our method achieves favorable performance against other state-of-the-art methods.
To further compare the running time of four subspace-based tracking algorithms (i.e. IVT, $L_1$APG, SP and our method), we calculated the average Frames Per Second (FPS) for $32\times32$ image patch (see the last row of Table~\ref{tab:ORE}). For $L_1$APG, we reported FPS for its APG acceleration.
It can be seen that IVT is quite faster than other trackers as its computation only involves matrix-vector multiplication. Both SP and our method are faster than $L_1$APG. It is also observed that our method is much faster than SP. This is due to the different choices of the optimization scheme. SP adopts a naive altering minimization strategy, in contrast, our method is efficiently solved by APG.
\begin{table*}[htbp]
\centering
\footnotesize
\caption{ Average Center Location Error(in pixels) and average frame per second (FPS). The best and the second results are shown in \textcolor[rgb]{1.00,0.00,0.00}{\textbf{BOLD}} fonts and \textcolor[rgb]{0.00,0.00,1.00}{\textbf{BOLD}} fonts, respectively.} \label{tab:CEE}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&{TLD} &{IVT} &{ASLA} &{$L_1$APG} &{MTT} &{SP} &{SPOT}&{FOT}&{SST}&{SCM} &{MIL} &{Struck} &{Ours} \\ \hline
Faceocc1 &27.37&18.42&78.06&17.33&21.00&14.14&17.17&29.00&13.00&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{13.04}}&29.86&18.78&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{12.88}}\\ \hline
Faceocc2 &12.28 &7.42 &19.35 &12.76 &9.836&10.43&11.78&11.94&12.82&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{5.96}} &9.02 &13.60&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{5.50}}\\ \hline
Freeman4 &39.18&43.04&70.24&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{22.12}}&23.55&79.66&108.70&54.66&56.20&56.20&62.07&48.70&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{10.39}}\\ \hline
David3 &208.00&51.95&87.76&90.00&341.33&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{8.74}}&6.27&33.40&104.50&73.09&29.68&106.50&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{5.79}}\\ \hline
Boy &4.49&91.25&106.07&7.03&12.77&58.09&8.93&5.79&66.97&51.02&12.83&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{3.84}}&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{2.57}}\\ \hline
Jumping &5.94&61.56&46.08&83.75&84.57&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{4.72}}&120.37&19.83&45.70&6.54 &65.89&9.99&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{4.99}}\\ \hline
Fish &6.54&5.67&3.85&29.43&45.50&3.99&4.52&6.50&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{3.14}}&8.54&24.14&3.40&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{3.08}}\\ \hline
Tiger1 &49.45&106.61&55.87&58.45&64.39&124.36&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{15.93}}&73.49&93.49&93.49&108.93&128.70&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{18.64}}\\ \hline
Singer2 &58.32&175.46&175.28&180.87&209.69&178.39&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{13.73}}&57.62&175.28&113.63&22.53&174.32&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{14.45}}\\ \hline
Jogging-2 &13.56&138.22&169.87&145.85&157.12&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{3.61}}&72.23&169.16&442.77&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{4.15}}&132.99&107.687&5.88\\ \hline
CarScale &22.60&11.90&24.64&79.78&87.61&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{13.36}}&207.01&106.20&87.05&33.38&33.47&36.43&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{7.66}}\\ \hline
Deer &30.93&182.69 &160.06&24.19&18.91&6.84&13.95&80.30&13.81&103.54&100.73&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{5.27}}&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{4.59}}\\ \hline
Basketball &213.86&107.11&82.64&137.53&106.80&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{39.79}}&169.86&118.02&105.93&52.90&91.92&118.6&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{7.92}}\\ \hline
Football &14.26&14.34&15.00&15.11&13.67&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{5.22}}&202.03&13.36&17.21&16.30&12.09&17.31&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{7.28}}\\ \hline
Average &50.48&72.54&78.20&64.58&85.48&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{39.38}}&69.46&55.66&88.42
&48.26&48.92&49.17&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{7.97}}\\ \hline \hline
FPS &21.74&\textcolor[rgb]{0.00,0.00,1.00}{\textbf{27.83}}&7.48&2.47&0.99&2.35&--&\textcolor[rgb]{1.00,0.00,0.00}{\textbf{376.48}}&2.12&
0.37&28.06&10.01&4.27\\ \hline
\end{tabular}
\end{table*}
\begin{figure*}[htb!]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.98\textwidth,keepaspectratio]{Fig7.jpg}
\end{tabular}
\end{center}
\caption
{ \label{fig: center error}
{\bf The plots of OPE with attributes based on the precision metric. }}
\end{figure*}
\begin{figure*}[htb!]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.98\textwidth,keepaspectratio]{Fig8.jpg}
\end{tabular}
\end{center}
\caption
{ \label{fig: overlap rate}
{\bf The plots of OPE with attributes using the success rate metric. }}
\end{figure*}
\section*{Conclusion}
In this paper, we propose sparse coding and counting method under Bayesian framwork for robust visual tracking. The proposed method combines $L_0$ regularization and $L_1$ regularized sparse representation in a unique one, therefore, it has better ability to sparsely represent an object and the reconstruction result are also better.
Besides, to solve the proposed model, we develop a fast and efficient APG algorithm.
Moreover, the closed solution of the combination of $L_0$ norm and $L_1$ norm regularization is provided.
Extensive experiments testify to the superiority of our method over state-of-the-art methods, both qualitatively and quantitatively.
\section*{Acknowledgment}
This work is partially supported by the National Natural Science Foundation of China (Nos. 61300086, 61432003, 61301270, 61173103, 91230103), the Fundamental Research Funds for the Central Universities (DUT15QY15), the Open Project Program of the State Key Laboratory of CAD\&CG, Zhejiang University, Zhejiang, China (No. A1404), and National Science and Technology Major Project~(Nos. 2013ZX04005-021, 2014ZX001011).
\section*{Appendix: Proof of Lemma~\ref{lemma:1}}
\label{appendix-1}
\begin{proof}
First, we denote $E(x)=\frac{1}{2}(x-y)^2+\delta|x|+\eta|x|_0$. It is obvious that if $x=0$, then $E(0)=\frac{1}{2}y^2$. Then we need to discuss the case that $x\neq0$:
\begin{enumerate}
\item if $x>0$, then $E(x)=\frac{1}{2}(x-y)^2+\delta x+\eta$. Writing its K.K.T condition, we get $x=y-\delta$, and the objective value is $E(y-\delta)=-\frac{1}{2}\delta^2+\delta y+\eta$.
\item if $x<0$, then $E(x)=\frac{1}{2}(x-y)^2-\delta x+\eta$. It is easy to get $x=y+\delta$, and the objective value is $E(y+\delta)=-\frac{1}{2}\delta^2-\delta y+\eta$.
\end{enumerate}
Then, we need to compare these three cases, if $E(0)>E(x-\delta)$, we have $(\delta-y)^2>2\eta$. Combining with $x=y-\delta>0$, we have $y > \delta+\sqrt{2\eta}$. Similarly, if $E(0)>E(x+\delta)$, then we have $y < -\delta-\sqrt{2\eta}$. And $x =0$, otherwise.
\end{proof}
\bibliographystyle{model1b-num-names}
|
\section{Introduction}
We propose a new methodology, forest floor, to visualize regression and classification problems through feature contributions of decision tree ensembles such as random forest (RF). Hereby, it is possible to visualize an underlying system of interest even when the system is of higher dimensions, non-linear, and noisy. 2D or 3D visualizations of a higher-dimensional structure may lead to details, especially interactions, not being identifiable. Interactions in the model structure mean that the model predictions in part rely on the interplay on two or more features. Thus, the interaction parts of a model structure cannot be reduced to additive scoring rules, one for each feature. Likewise, to plot single feature-to-prediction relationships is not a sufficient context for visualizing any interactions. Often a series of complimentary visualizations are needed to produce an adequate representation. It can be quite time consuming to look through any possible low dimensional projection of the model structure to check for interactions. Forest floor guides the user in order to locate prominent interactions in the RF model structure and to estimate how influential these are.
For RF modeling, hyper parameter tuning is not critical and default parameters will yield acceptable model fits and visualizations in most situations \cite{Liaw2002,Svetnik2003}. Therefore, it is relatively effortless to train a RF model. In general, for any system where a model has a superior prediction performance, it should be of great interest to learn its model structure. Even within statistical fields, where decision tree ensembles are far from standard practice, such insight from a data driven analysis can inspire how to improve goodness-of-fit of a given model driven analysis.
Although the RF algorithm by Breimann \cite{Breimann2001} has achieved the most journal citations, other later decision tree ensemble models/algorithms such as ExtraTrees \cite{Maree2005}, conditional inference forest \cite{Hothorn2006}, Aborist \cite{Seligman2015}, Ranger \cite{Ranger2015} and sklearn.random.forest \cite{sklearn} will often outperform the original RF on either prediction performance and/or speed. These models/algorithms differ only in their software implementation, split criterion, agreggation or in how deep the trees are grown. Therefore all variations are compatible with the forest floor methodology. Another interesting variant, rotation forest \cite{Rodriguez2006}, does not make univariate splits and is therefore unfortunately not directly compatible with forest floor visualizations. To expand the use of feature contributions and forest floor, we also experimented with computing feature contributions for gradient boosted trees \cite{friedman2001}. This is possible, as splits still are univariate and trees contribute additively to the ensemble prediction. A proof-of-concept of computing feature contributions on gradient boosted regression trees and visualizations are provided in supplementary materials.
Decision trees, as well as other machine learning algorithms, such as support vector machines and artificial neural networks can fit regression and classification problems of complex and noisy data, often with a high prediction performance evaluated by prediction of test sets, n-fold cross validation, or out-of-bag (OOB) cross validation. The algorithms yield data driven models, where only little prior belief and understanding is required. Instead, a high number of observation are needed to calibrate the adaptive models. The models themselves are complex black-boxes and can be difficult to interpret. If a data driven model can reflect the system with an impressive prediction performance, the visualization of the model may deduce knowledge on how to interpret the system of interest. In particular, a good trade-off between generalization power and low bias is of great help, as this trade-off in essence sets the boundary for what is signal and what is noise. The found signal is the model fit, which can be represented as the mapping from feature space to prediction space (output, target, response variable, dependent variable, y). The noise is the residual variance of the model. The estimated noise component will both be due to random/external effects but also lack of fit.\par
\subsection{Overview of the article}
In this article we introduce the forest floor methodology. The central part is to define a new mapping space visualization, forest floor. Forest floor rely on the feature contributions method \cite{Kuz'min2011}\cite{Palczewska2014}, rather than averaging many projections (partial dependence) \cite{friedman2001} or projecting the average (sensitivity analysis) \cite{Cortez2013}. In Section \ref{intrRepRF} these previous mapping space visualizations are introduced and the challenges to overcome are discussed.
In the theory section, \ref{TheoDefMap}, we discuss the feature space, prediction space and the joined mapping space for any regression or classification model and define local increments as vectors in the prediction space. Properties of the RF algorithm by Breimann \cite{Breimann2001} and the feature contributions method by Kuz'min \textit{et al} \cite{Kuz'min2011} and Palczewska \textit{et al} \cite{Palczewska2014} are highlighted and illustrated in section \ref{TheoRF}. In section \ref{TheoLI} we argue that the prediction of any node in any tree is a point in the prediction space and the local increments are the vectors that connect the nodes of the trees. Any prediction for any observation is basically a summed sequence of local increments plus the grand mean or base rate. Since local increments are vectors and not a tree graph, the sum of vectors is not dependent on the order of the sequence. In Section \ref{TheoDecompose} we show how that feature contributions, a particular reordering of local increments by splitting feature, can be used to decompose the model structure \ref{TheoDecompose}.
We also introduce a new cross-validated variant of feature contributions and provide an elaborated definition of feature contribution to also account exactly for the bootstrapping process and/or stratification.
The materials and methods sections, \ref{materialDataSoft} and \ref{materialToyData}, provide instructions on how to reproduce all visualization in this paper. The result section \ref{resStart} is dedicated to three practical examples of visualizing models with forest floor. The three examples are a simulated toy data set, a regression problem (white whine quality) and a classification problem (contraception method choice). A low-dimensional visualization is not likely to convey all aspects of a given RF mapping surface. For all practical examples, we describe how to find an adequate series of visualizations that do.
\subsection{Representations of random forest models}
\label{intrRepRF}
A RF model fit, like other decision tree based models, can be represented by the graphs of the multiple trees. Few small tree graphs can be visualized and comprehended. However, multiple fully grown trees are typically needed to obtain an optimal prediction performance. Such a representation cannot easily be comprehended and is thus inappropriate for interpretation of model fits. A random forest fit can be seen as a large set of split rules which can be reduced to a smaller set of simpler rules, when accepting a given increase in bias. This approach has been used to reduce the model complexity \cite{Liu2014}. But if the minimal set of rules still contains a large number, e.g. hundreds or thousands, then this simplified model fit is still incomprehensible. It is neither certain which rules have influence on predictions nor which rules tend to cancel each other out. We believe that the rule-set or tree-structure representations are mainly appropriate to understand how a RF algorithm possibly can model data. On the other hand, these representations are indeed inappropriate for interpreting RF model fits and conveying the overall model structure. For that purpose, a mapping space visualization is superior in terms of visualization and communication.
\begin{figure*}[htp]
\centering
\includegraphics{figures/partialExplained.png}
\caption{Illustration of sensitivity analysis and partial dependence plots. The grey response surface depicts a given learned model structure of two input features ($X_1$ and $X_2$) and one prediction axis ($\hat{y}$). 11 data points vs. predictions are depicted as blue dots. 1D-sensitivity analysis (fat red lines): one partial function slice intersects the centroid where $X_2 = \overline{X}_2$ an is projected to the $X_1$-$y$ plane. ICE plot: Multiple function slices (black lines) all parallel to $X-1$ intersects each one data point and all slices are projected to the $X_1$-$y$ plane. Partial dependence plots: Each data point intersected by one black line is projected to any black lines (green points). The green point outline a grid. All green and blue points are projected into the $X_1$-$y$ plane, and the fat green line connects the average prediction values as a function of $X-1$. This illustration can be generalized to any dimensional reduction.
}
\label{partialFigure}
\end{figure*}
If we join the feature space and prediction space, this function will be represented as a geometrical shape of points. Each point represents one prediction for a given feature combination. This geometrical shape is the model structure and is an exact representation of the model itself. Nevertheless, for a given \emph{d}-dimensional problem where $d>3$, this is still difficult to visualize or even comprehend. Instead, one may project/slice or decompose the high-dimensional mapping into a number of marginal visualization where small subsets of features can be investigated in turns. This allows us to comprehend the isolated interplay of one or a few features in the model structure.
Following, we will introduce previous examples of mapping space visualizations to specify what forest floor aims to improve. Different types of sensitivity analysis (SA) were used by Cortez and Embrechts to make such investigations \cite{Cortez2013}, we will here discuss sensitivity analysis and data based sensitivity analysis. First a supervised machine learning model is trained. Next the model is probed. That means to input a set of simulated feature observations (points in feature space) into the model fit and record the output (target predictions). Instead of probing the entire high-dimensional mapping space, only one confined slice of fewer dimensions is probed in order to make feasible visualizations.
The simplest visualization in SA is one dimensional (1D-SA), where a single feature is varied in a range of combinations, and this range will span the X-axis of the visualization. When two features are varied (2D-SA), the resulting grid of combinations will span the XY-plane. All other features must be fixed at e.g. the mean value, the feature centroid of the training set. The model fit is probed with these observations and the resulting predictions will be plotted by the Z-axis. The obtained line/surface will now visualize one particular 2D or 3D slice of the full mapping structure.
In figure \ref{partialFigure}, a non-linear regression model structure ($y = sin(X_1)^8 sin(X_2)^8 + \epsilon$) is represented by the grey transparent surface. The model has two feature axes in the horizontal XY-plan and the prediction axis by the vertical Z-axis. Thus, the mapping space has 3 dimensions and the model structure is some curved 2D-surface which connect any given feature combination with one prediction. The red line/slice in the model structure is the example of an 1D-SA visualization. This single slice is projected into the $X_1$-$Z$ plane. This 1D-SA projection portrays the partial effect of feature $X_1$ in the special case, where other features are set to mean observed value. Notice that the red line almost completely misses the local hill in the model structure. A single low dimensional slice of the mapping structure can easily miss prominent local interactions, when number of model dimensions is high.
A 2D-SA slice can explain a main effect and/or the possible interaction within two selected features. Figure \ref{partialFigure} only illustrates a 1D-SA slice projection, but represents the idea of any projection. The depicted model structure itself could infact be a 2D-SA projection of a higher dimensional model structure. Whether a given slice is a good generalization of the full mapping structure is unknown. A good generalization means that any parallel slices, where the fixed features are set to another combination, yield the same XYZ-visualization, with only perhaps a fixed offset in the prediction axis (Z) \cite{icebox}. We will for now term that such visualization has a high goodness-of-visualization. In section \ref{TheoDecompose} we will propose a metric for goodness-of-visualization. For a data structure with only additive effects and no interactions, the obtained model mapping structure is likely to have no interactions as well as any slice will be identical to its many parallel counterparts. In Figure \ref{partialFigure}, all the black parallel slices to the red slices give different projection lines in the mirror plane which could not be corrected by a simple offset. Therefore the model structure must have an interaction which cannot be seen in this projection alone. The iceBOX package displays multiple projection lines to search for masked interactions and is a good alternative to the forest floor approach \cite{icebox}.
A second concern is whether a given slice or slices extrapolate the training data. For a RF model with a satisfactory cross validated prediction performance, the mapping structure will represent the underlying data structure, but only within the proximity of the training data. Extrapolated areas of the mapping structure are far from guaranteed to represent an underlying data structure. Several different non-linear learners (RF, SVM, ANN, etc.) may easily have comparable model structures in the proximity to training data points, whereas far from the training set the models will heavily disagree. For RF models containing dominant interaction effects, the mapping structure on the borders of the training data becomes noise sensitive, as decision trees only can extrapolate parallel to feature axes, as the splits only are univariate. RF models only containing additive main effects have stable and smooth mapping structure at the borders of the training data. Model extrapolation of random forests with dominant interaction effects have been illustrated in supplementary materials.
SA plots remain a useful tool. When forest floor yield plots of similar structure, these plots generally represents the model mapping well. Visualization of multiple parallel projections, the so called ICE plots (individual conditional expectation) with the ICEbox package, can also reveal interactions. However multiple projection lines cannot directly filter out main effects by other features. These will tend to offset the projection lines on the prediction axis. Centered ICE (c-ICE) visualizations do adjust this offset by centering the prediction axis for all projections in one specific location \cite{icebox}.
A frequently used visualization method proposed by Friedman is the partial dependence plot (PD) which is the same as what Cortez and Embrechts later have termed data-based sensitivity analysis (DSA)\cite{Cortez2013,friedman2001}. In Figure \ref{partialFigure}, the green fat line in the mirror plane represents a partial dependence projection. Whereas 1D-SA and 2D-SA only project the slice intersecting e.g. the training data centroid, the partial dependence plot projects multiple slices. Each projected slice intersects one data point. The partial dependence line is the average prediction values of all slices. Thus, the obtained PD visualization summarizes all parallel slices of the mapping structure by averaging. To summerize, SA averages and then projects, whereas PD projects and then averages. ICE-plot projects many slices and do not aggregate. The PD approach may improve generalization across slices as it up-weighs the parts of mapping structure, that are well represented by data points. Still, interactions between varying and fixed features will be lost by averaging. Furthermore, the PD projections form a regular data grid spanned by the data observations. See the grid of black and green lines on the model structure surface in Figure \ref{partialFigure}. However, for data sets with high feature collinearity, data points will mainly be positioned in one diagonal of the grid, whereas the remaining part of the grid will span extrapolated parts of the model structure. This extrapolation occur for both SA, PD and ICE-plots.
Feature contributions was introduced by Kuz'min \cite{Kuz'min2011} for RF regression and elaborated by Palczewska \emph{et al} \cite{Palczewska2014} to also cover RF multi-classification. Feature contributions are RF predictions split into components by each feature. Feature contributions are essentially computed utilizing information from the tree networks of a RF model. Feature contributions have not before been used or understood in conjunction with the idea of function mapping structures. The contribution of this paper, is to show that feature contributions can be understood as a different way of slicing the mapping structure. From this insight the methodology, forest floor, was developed.
We have developed a number of tools to increase the usefulness of the forest floor methodology. These are: Out-of-bag cross validated feature contributions to increase robustness without increasing computation time, goodness-of-visualization tests to evaluate how well slices generalize the mapping structures and color gradients traversing mapping space to visually identify latent sources of interactions. Furthermore, the methods have been implemented as a freely available R-package, from which all mapping visualizations of this paper originate. The R-package forestFloor \cite{forestFloor2015} aims to assist the user visualizing a given RF model fit through a serious of appropriately chosen visualizations.
\section{Theory and calculation}
Here is provided a new notation for RF regression and classification to combine a mapping space representation with the feature contributions method developed by Kuz'min \cite{Kuz'min2011} and Palczewska \emph{et al.} \cite{Palczewska2014}. Moreover to obtain an exact decomposition of the model structure, we expand the previous notion of feature contribution to also cover the initial bootstrap and/or stratification step for each decision tree. For RF multi-classification we describe a probabilistic (K-1)-simplex prediction space, to improve the interpretation of feature contributions. Lastly we introduce how to calculate out-of-bag cross-validated feature contributions.
\subsection{Defining regression and classification mappings}
\label{TheoDefMap}
Any regression model $f_{r}$ can be seen as a mapping between a $d$-dimensional feature space $X \in \mathbb{R}^d$ and and a prediction scale $\hat{y} \in \mathbb{R}^1$
\begin{equation}
\label{regMap}
\hat{y} = f_{r}(X) \quad ,
\end{equation}
where $X$ represents the infinite set of points in the feature space. A subset of points in $X$ can be notated as e.g. $X_t$ where $t$ is a defined set. Single value entries of a countable subset of $X$ is notated as $x_{ij}$ where $i \in \{1,...,N\}$ ($N$ points) and $j \in \{1,...,d\}$ ($d$ features). $\hat{y}$ represents the entire prediction scale, where $\hat{y}_t$ could be a subset, if countable with point entries $\hat{y}_i$.
The entire mapping can be represented as a $d$-dimensional (hyper)surface $S$ in a $d+1$-dimensional mapping space $V$. $S$ can be understood as a learned model structure trained on a set of training observations, $t$. Obviously, if $d\in \{1,2\}$, then $S$ can conveniently be plotted by Cartesian axes as a 2D function plot or a 3D response surface (prediction as function of two features). Each label of a categorical feature can be assigned an integer value from 1 to \emph{K'} categories and thus also be plotted.
A classification model can be seen as a mapping from $X \in \mathbb{R}^d$ to $\hat{y} \in \{1,2,...,K\}$. Some models, as RF, provides a probabilistic prediction (pluralistic voting) of class membership $\hat{p}_{k}$ for any class $k \in \{1,2,...,K\}$ and assign the class membership hereafter. Thus, the probabilistic classification model $f_{c}$ is a mapping from $X$ to the probability space $P$,
\begin{equation}
\label{probMap}
f_{c}(X) = P \quad .
\end{equation}
Any point in $P$ is a possible prediction $\hat{p}$ with a unique probability distribution over $K$ mutually exclusive classes, such that $\hat{p} = \{ \hat{p}_{1},\hat{p}_{2},...,\hat{p}_{K} \}$. As class memberships are mutually exclusive, the sum of the class probabilities is always one, $|\hat{p}|^1=1$. Therefore the probability space is a \emph{K-1} dimensional simplex \cite{OBrien2008}, which contains any possible combination of assigned probabilities to $K$ mutually exclusive classes, see Figure \ref{mapTypes} . The $K$ axes, which assign probability of 0 to 1, are not orthogonal, meaning it is not possible to modify the assigned probability of one class without affecting at least one other.
The classification mapping can be represented by simply joining the simplex-space with the feature space, but this would only allow a 2D or 3D visualization when $(d+K-1) \in \{2,3\}$, thus either maximally a 2 feature problem for 2 classes, or a 1 feature separation for 3 classes. Instead, this mapping can also be represented as $K$ separate $d$-dimensional surfaces $S_{k}$ in a $d+1$-dimensional space $V$ with $d$ axes representing features and one axis $\hat{p}_k$ representing the probability of either of the $K$ classes. Thus, we align the directions of all $K$ probability axes to reduce the dimensionality of the mapping space with $K-2$ dimensions. Then, any line parallel to the probability axis $\hat{p}_k$, will intersect every $S_{k}$ surface, describing the predicted probability of the $k^{th}$ class at this point of input features. The sum of predicted probabilities of all intersections for any such line will be equal to one. To summarize, multi classification model structures are more difficult to visualize, as each class adds another dimension to the mapping space. It is possible to plot the individual predicted probability of each class and overlay these plots. Figure \ref{mapTypes} summarizes the mapping topology for regression, for binary classification, and for multi classification.
\begin{figure*}[htp]
\centering
\includegraphics{figures/fig01.mapping_types.140mm.pdf}
\caption{Topologies of random forest model represented as a function mapping from $d$-dimensional feature space to one of the following prediction spaces: (a) regression, 1-dimensional scale; (b) binary classification, $K=2-1$ probability simplex reducible to a 1-dimensional probability scale; (c) multi-classification, probabilistic ($K-1$)-simplex. The mapping can be represented as a high-dimensional surface $S$, in a joined feature and prediction space linking any combination of features to a given prediction. For multi-classification, $S$ can be split into multiple $S_k$ surfaces describing predicted probability for each of $K$ individual classes.}
\label{mapTypes}
\end{figure*}
RF mapping for both regression and classification can jointly be defined as
\begin{equation}
\label{RFmap}
\hat{y} = f(X) \quad .
\end{equation}
Here $\hat{y}$ is the \emph{c}-dimensional prediction space. For regression, $c=1$, $f$ maps to a 1-dimensional prediction scale. For classification, $c=K$ classes, and $f$ maps to a prediction vector space, where the $k^{th}$ dimension predicts the probability of class $k$. For classification the predictions $\hat{y}$ can be any point within the ($K-1$)-simplex. On the other hand, the training examples $y$ can only be of one class each, which are the $K$ vertices (corners) of the ($K-1$)-simplex.
We define a local increment vector, $L$, pointing from $\hat{y}_i$ to $\hat{y}_j$ in a prediction space of $c$ dimensions, such that
\begin{equation}
\label{LIdef}
L_{ij}=\hat{y}_j - \hat{y}_i =
\{\hat{y}_{j1}-\hat{y}_{i1},...,\hat{y}_{jc}-\hat{y}_{ic} \} \quad .
\end{equation}
For regression, where $(c=1)$, the local increment is a scalar with either a positive or negative direction. For classification, $(c>1)$, the local increment is a vector with $c$ elements, one for each class. Each node of a RF model fit is a prediction, which is a specific point in the prediction space. Local increments are the connections between nodes, describing the change of prediction. Computing the thousands or millions of local increments for trees and nodes, and sum these individually for each observation and feature is essentially the feature contributions method.
\subsection{Properties of random forest related to feature contributions}
\label{TheoRF}
RF is an ensemble of bootstrapped decision trees for either regression or classification. Figure \ref{forestFigure} illustrates how the RF algorithm operates for regression. For each of the trees (1 to $n_{tree}$) the training set is bootstrapped (random sampling with replacement). In average $ (\frac {N-1} N)^N \approx 0.37$ of $N$ observations will not be included in each bootstrap. These observations are called out-of-bag (OOB). Thus for any tree, a selection of observations will be 'in-bag' and used to train/grow the tree starting from the root node. Any node will have a node prediction which is defined by in-bag observations in that node.
\begin{equation}
\label{nodePredReg}
\hat{y}''_{j} = \frac{1}{n_j} \sum_{i=1}^{n_{j}} y_{ij}
\end{equation}
For a regression tree, the node prediction of the $j^{th}$ node $\hat{y}''_j$ is equal to the mean of in-bag target values in the $j^{th}$ node. Where $y_{ji}$ is the target value of the $i^{th}$ observation in the $j^{th}$ node. $n_j$ is the number of observations in the $j^{th}$ node. Thus we are only computing a node prediction from in-bag elements.\par
For classification, the probabilistic node prediction $p_{jk}$ of the class $k$ of the node $j$ is equal to the number of in-bag observations of class $k$ divided with total number of in-bag observations in the node.
\begin{equation}
\label{nodePredClass}
\hat{p}_{jk} =\frac{n_{jk}}{n_j} \quad .
\end{equation}
A node prediction $\hat{y}''_{j}$ can also describe all class probabilities at once as a vector corresponding to a point in the $(K-1)$-simplex space.
\begin{equation}
\label{nodePredClass2}
\hat{y}''_{j} = \{ \hat{p}_{(j,1)},..., \hat{p}_{(j,K)} \}
\end{equation}
For classification $c>1$, the class probabilities of any node will always sum to 1 for any node:
\begin{equation}
\label{nodePredClassSumToOne}
|\hat{y}''_j|^1 = \sum_{k=1}^{K}{p_{jk}} = 1 \quad .
\end{equation}
Therefore, the elements of any local increment vector for classification, see Equation \ref{LIdef} will always sum to zero. This is not true for the local increment scalars of regression, $c=1$.
For an original RF implementation \cite{Liaw2002}, predictions of terminal nodes of classification trees are reduced to a single majority vote. Other implementations such as sklearn.randomForestClassifier \cite{sklearn} would rather pass on the probabilistic vote from terminals nodes and only on the ensemble level perform reduction by majority vote or just keep the full probabilistic average. In practice, implementations of feature contributions usually have to re-estimate node predictions. A feature contributions implementation such as forest floor should match the specific rule of terminal node predictions of the specific model algorithm.
A node is by default terminal if there are 5 or less in-bag observations left for regression or a single in-bag observation for classification. Any non-terminal node will be split into two daughter nodes to satisfy a loss-function. For regression the loss function is typically the sum of squared residuals.
For classification, a Gini criterion is used as the loss function. That is to select the split yielding the lowest node size weighted Gini impurity. Gini impurity ($g$) is 1 minus the sum of squared class prevalence ratios in nodes, $g = 1- \sum_{k=1}^K \hat{p}_{jk}^2$. Gini impurity is in fact the equation of a K-dimensional hypersphere, where $\sqrt{1-g}$ is the radius and all $\hat{p}_{jk}$ are the coordinates. The $(K-1)$-simplex space intersects this hypersphere where all prevalences sum to one, $1 = \sum_{k=1}^K \hat{p}_{jk}$. Therefore for a $K=3$ classification, a Gini loss function isobar appear as a 2D-circle, when visualized in the ($K-1$)-simplex space. One circular isobar is drawn in Figure \ref{LIclass}. The Gini loss function chooses the split placing two daughter nodes the furthest from the center of the ($K-1$)-simplex.
Splitting numerical features of ratio-, ordinal- or integer-scale is all the same for RF. A break point will direct observations lower or equal to the left node. Splitting by categorical features is to find the best binomial combination of categories designated for either daughter node. A feature with 8 categories will have $2^{8-1}-1 = 63$ possible binary splits. Any available break point are evaluated by the loss-function, but the RF algorithm is constrained to only access a random selection of the features in each node. The amount of features available, \emph{mtry}, can e.g.\ be a third of the total amount of features. This \emph{random variables subspace} and bootstrapping will ensure decorrelation of trees and feature regularization without overly increasing the bias of each fit. Each fully grown tree is most likely highly overfitted, as the individual predictions of each terminal node are dictated by 5 or less observations. Combining the votes of many overfitted but decorrelated trees form an ensemble with lowered variance and without increased bias. Out-of-bag(OOB) predictions are calculated for each terminal nodes. As OOB observations are not used actively in growing the trees of the forest, they can serve as an internal cross validation which yields similar results as a 5 fold cross validation \cite{Svetnik2003}. The prediction of individual trees are written as $\hat{y}'_{ij}$ for $i \in \{1,...,N\}$ observations predicted by $j \in \{1,...,n_{tree}\}$. The ensemble predictions are computed as
\begin{equation}
\label{pred}
\hat{y}_{i} = \frac{1}{n_{tree}} \sum_{j=1}^{n_{tree}} \hat{y}'_{ij} \quad ,
\end{equation}
and the OOB cross validated ensemble predictions $\tilde{y}_i$ are computed as
\begin{equation}
\label{OOBpred}
\tilde{y}_{i} = \frac{1}{|\tilde{J}_i|} \sum_{j \subseteq \tilde{J}_i}\hat{y}'_{ij} \quad ,
\end{equation}
where $\tilde{J}_i$ is the subset of $\{1,...,n_{tree}\}$ trees, where $i^{th}$ observation is OOB. $|\tilde{J}_i|$ is the size of the subset $\tilde{J}_i$. Thus let any training observation $i$ iterate through the $\tilde{J}_i$ subset of trees, defined as those trees where $i$ was not in-bag, and find the mean of terminal node predictions.
To obtain value/class predictions of new observations, the observations will be forwarded through all trees according to the established split rules. A tree prediction is dictated by the terminal node a given observation ends up in. The ensemble prediction of a RF model fit will by default be the average for regression and the majority vote for classification. Figure \ref{forestFigure} explains graphically the structure of a single regression tree by feature $x_1$ and $x_2$. First all bootstrapped observations exist within the node n1. The mean prediction value of n1 is in this example 0.14 a slight offset compared to the training set prediction mean of 0. The first split is over a break point in $x_2$, dividing n1 into n2 with low prediction value and n3 with a high prediction value. Both n2 and n3 are further split by $x_1$. Interestingly, n2 and n3 have almost opposite splits by $x_1$. In n2, high $x_1$ leads to a lower prediction, while reversely in n3. This illustrated tree have only grown 7 nodes. Nonetheless, the tree contains an interaction term, where high $x1$ only contribute positively to the prediction $\hat{y}$ when conditioned by high $x_2$.
\begin{figure*}[htp]
\centering
\includegraphics[width=1.00
\textwidth]{figures/fig02b.RF_visulization.190mm.pdf}
\caption{Random forest and local increments explained. Left, an 3D illustration of a small regression tree of 7 nodes. Right, the same tree described by node means($\overline{u}$), node size($n$) and local increments $L_{ijk}$. $L$ is subsetted by observation, tree, node and feature. A observation falling in e.g.\ node 4, will have a prediction as the sum of the local increments in its path plus the grand mean of the training set.
}
\label{forestFigure}
\end{figure*}
\subsection{Local increments and feature contributions}
\label{TheoLI}
This section explains how feature contributions are computed. This paper expands the feature contributions defined by Palczewska \textit{et al} \cite{Palczewska2014} to also account for bootstrapping and/or stratification and to allow OOB cross validation. Feature contributions summarize the pathways any observation (a given combination of input features) will take through the many decision trees in a RF model. Each sub node of the trees holds a prediction, which is average observed target of observations populating it, see Equations \ref{nodePredReg} \& \ref{nodePredClass}. The sum of the many steps from node to node (local increments) is for regression exactly the resulting large step from the grand mean of the training set to the given numeric target prediction. Likewise for classification, the large step is from base rate to a probabilistic target prediction. A proof hereof is provided in supplementary materials. As these many small steps towards the final prediction is an additive process, it is possible to reorder the sequence of steps and end up by the same prediction. The important implication hereof is that the RF model structure can be decomposed into additive sub models, each with the same dimensionality. As each sub model structure is the sum local increments of decision splits by one specific feature, each sub model structure tend to only describe the main effect of this one specific feature plus perhaps interactions with other features.
In order to efficiently describe how variations of feature contributions are computed, a notation of how to access any local increment in a given RF model fit is formulated. We define $L$ as a list of lists of lists containing all local increments. $L$ is defined in the following three levels (observations, trees, increments):
\begin{enumerate}
\item $L_i$ is a list with $i \in \{1,...,N\}$, and $N$ is the number of observations predicted by the forest. $i$ is the $i^{th}$ observation.
\item Each element of $L_i$, called $L_j$ is a list with $j \in \{1,...,n_{tree} \}$, and $n_{tree}$ is the number of trees in the ensemble.
\item Each element of $L_j$, called $L_k$ is a list with $k \in \{1,...,n_{increment,i,j} \}$, and $n_{increment,i,j}$ is the number of increments encountered by the $i_{th}$ observation in the $j^{th}$ tree.
\end{enumerate}
Note that $L$ can be ordered as a 2-dimensional array ($i$ observation, $j$ tree) where each element is a sequence of local increments specific for the $i^{th}$ observation in the $j^{th}$ tree. Overall, we can access any local increment in $L$ with $L_{ijk}$. Depending on the model type, $L$ will contain local increments as scalars for regression or as vectors for classification. The first local increment $k=1$ for any tree and observation in $L_{ijk}$ is the step from node 0 (training set) to node 1 (root node of tree). Thus the $k^{th}$ local increment steps from the parent node $k-1$ to a daughter node $k$. The local increment $L_{ijk}$ is the change of node prediction $\hat{y}''_{ijk}-\hat{y}''_{ij(k-1)}$
Equation \ref{LIeq2} describes how any prediction can be computed from $L_{ijk}$ as the sum of all local increments plus grand mean or base rate. A proof hereof can be found in the supplementary materials.
The target prediction $\hat{y}_i$ is computed as
\begin{equation}
\label{LIeq2}
\hat{y}_{i} = \frac{
\sum_{j=1}^{n_{tree}} \sum_{k=1}^{n_{increment,i,j}} L_{ijk}}
{ n_{trees}}
+ \overline{y} \quad ,
\end{equation}
where $L_{ijk}$ is a local increment and where $\overline{y}$ is the grand mean or base-rate. The numerator is a scalar for regression and a vector for classification. The denominator, $n_{tree}$, is always a scalar.
So far the prediction of the $i^{th}$ observation is the grand mean (regression) or the base-rate (classification) plus the sum of all local increments $L_{ijk}$ encountered by this $i^{th}$ observation divided by $n_{trees}$.
Figure \ref{LIclass} is a new geometrical representation of local increments for a 3-class classification. Figure \ref{LIclass} is not intended as a model structure visualization, but rather as a representation of how decision trees branch out in the prediction space. Each node in the classification tree can be seen as a probabilistic prediction defining a point in a probabilistic ($K-1$)-simplex. Figure \ref{LIclass} depicts node predictions and local increments for a small tree with four terminal nodes. To this tree graph is appended a node (T) for training set to the root node of the tree. This train node represents the class distribution of the training set. The bootstrap increment leads to the root node. This step is often small and a result of random uniform sampling w/o replacement. If applying class stratification, the length and direction of this step can be controlled. Stratification corresponds to defining a prior expected class distribution, which will be the position of the root nodes in the prediction space. From here all trees will branch out from this point. The following local increments and nodes comprise the entire tree. Any split produces two nodes and two local increments of opposite direction. If not of equal node size, there will be one shorter local increment defined of many in-bag observations and one longer local increment defined of fewer in-bag observations. This is a consequence of that class distributions of daughter nodes multiplied by the node sizes and added together is exactly equal to class distribution of parent node multiplied by its node size. This symmetry effect can be found in Figure \ref{cmc.mainSimplex} in section \ref{resCMC}. For the unbalanced binary features \emph{wives' religion}, \emph{wives working} and \emph{media exposure} the prediction is offset a lot for a few observations, while the prediction of remaining many observations will only change a little in the exact opposite direction. For regression and binary classification such a direction is essentially one-dimensional and can be positive or negative. For multi classification the direction is a vector of K elements with the restriction that the sum of elements is zero. In Figure \ref{LIclass}, the circle represents a Gini loss function isobar. The further away (euclidean distance) nodes are placed from uniform class distribution the better a split according to RF Gini loss function. The best kind of split is one placing both daughter nodes onto two of the K vertices of the (K-1)-simplex.
\begin{figure*}[htp]
\centering
\includegraphics{figures/fig03.LIclass.90mm.pdf}
\caption{A representation of how node predictions and local increments for a small classification tree with four terminal nodes. The first node in center represents the class distribution of a balanced training set (T). The bootstrap increment leads to the root node of the tree (R). The following local increments and nodes comprises the entire tree. Any split produces two local increments of opposite direction. The circle represents Gini loss function isobar. The further the two nodes (weighted by size) are from uniform class distribution the better a split according to the Gini loss function.}
\label{LIclass}
\end{figure*}
For the training set, a cross validated OOB-prediction $\tilde{y}$ can be formulated as
\begin{equation}
\label{LIeq3}
\tilde{y}_{i} = \frac{
\sum_{j \subseteq \tilde{J}_i} \sum_{k=1}^{n_{increments,i,j}} L_{ijk}}
{ |\tilde{J}_i|}
+ \overline{y} \quad ,
\end{equation}
where $\tilde{J}_i$ is the subset of trees where $i^{th}$ sample is OOB. One can reason, that if Equation \ref{LIeq2} is true for any set of trees, then Equation \ref{LIeq3} must also be true for a given subset of any trees, such as the OOB subset $\tilde{J}_i$, see supplementary materials.
When predicting the training set with an RF model, any training observation $i \in \{1,...,N\}$ will have a high proximity to itself, that is, it will in any in-bag tree both define the in-bag node predictions of the terminal node and be predicted by the very same terminal node. For data sets with a high noise level this becomes a problem and the points $S_i$ of model structure $S$ will overfit the sampled training set observations $T_i$, and visualizations hereof will look more noisy. If the RF training parameter minimum terminal node size is increased and/or bootstrap sample size is lowered then training observation $i$ will have a lower influence on its own prediction and visualizations will not look noisy. \par
To compute feature contributions, the summed local increments over each observation and feature, it is necessary to keep a record of splitting features in each parent node. In Equation \ref{LIeq2}, the $i^{th}$ observation in the $j^{th}$ tree encountered the local increments for $k \in \{1,...,n_{increments,i,j}\}$. For this $i^{th}$ observation in $j^{th}$ tree, let $H_{ijl}$ be the subset of local increments where the parent node was split by the $l^{th}$ feature. The local increments of bootstrapping are assigned to \textit{feature 0}. The letter $H$ is used, as $K$ already is used to describe number of classes.
This distinction between OOB-predictions $\tilde{y}$ and regular test predictions $\hat{y}$ of training set now becomes important as how to feature contributions are defined. Previously \cite{Palczewska2014,Kuz'min2011} feature contributions have been defined for regression and classification analogous to this:
\begin{equation}
\label{FC}
F_{il} = \frac {
\sum_{j=1}^{n_{tree}}
\sum_{k \subseteq H_{ijl}}
L_{ijk}
}{n_{tree}} \quad ,
\end{equation}
Here $F_{il}$, the feature contribution of the $i^{th}$ observation for the $l^{th}$ feature, is a subtotal of local increments $L_{ijk}$, where $k$ only iterates over $H_{ijl}$, which is those times the parent nodes were split by feature $l$.
This definition of feature contributions is fine if: (a) the noise level is low or (b) if feature contributions $F$ only is computed for some test set different from training set or (c) if the user is confident, that the model structure is not over fitted. It would be possible to cross validate by segregating the data set in a training set and test set to avoid over fitted visualizations. To discard data points is not desirable for a data set with limited observations. It would be possible to perform an n-fold cross validation, but n-fold random forests would be necessary to train.
We propose to compute feature contributions for the OOB cross validated predictions. OOB cross validated predictions are only the sum of local increments over trees where $i^{th}$ observation was OOB, see Equation \ref{LIeq3}. Analogously, we OOB feature contributions $\tilde{F}_{il}$ as
\begin{equation}
\label{FCoob}
\tilde{F}_{il} = \frac{
\sum_{j \subseteq \tilde{J}_i}
\sum_{k \subseteq H_{ijl}}
L_{ijk}
} {|\tilde{J}_i|} \quad ,
\end{equation}
where $j$ only iterates the subset of trees $\tilde{J}_i$, and where $i^{th}$ observation was OOB. $|\tilde{J}_i|$ is the total number of times the $i^{th}$ observation was OOB and the size of the subset $\tilde{J}_i$. Equation \ref{FCoob} is used in forest floor visualizations to compute cross validated feature contributions of the training set predictions.
\subsection{Decomposing the mapping surface with feature contributions}
\label{TheoDecompose}
We can compute the OOB cross validated set of points $\tilde{S}_i = \{X_i,\tilde{y}_i\}$ for $i \in T$ the training set. That is the combination by training features $X_i$ and the cross validated predictions $\tilde{y}_i$, where $c=1$ for regression and $c>1$ for classification. To decompose $\tilde{S}_i$, then $\tilde{y_ i}\}$ is expanded with $\tilde{F}_{il}$, such that:
\begin{equation}
\label{FCsums2y}
\tilde{y_i} = \sum^{d}_{l=0} \tilde{F}_{il} + \overline{y} \quad .
\end{equation}
Likewise non cross-validated $\hat{y_i}$ is a sum of non cross-validated $F$,
\begin{equation}
\label{FCsums2yhat}
\hat{y}_i = \sum^{d}_{l=0} F_{il} + \overline{y} \quad .
\end{equation}
The ensemble prediction $\hat{y}$ or $\tilde{y}$ is equal to sum of local increments + grand mean / base rate, see Equation \ref{LIeq2},\ref{LIeq3}. As sequences of additive vectors can be rearranged, it is possible to compute sub totals of local increments of the full prediction. Feature contributions is just the subtotal of encountered local increments for the for the $i^{th}$ observation where the parent node was split by the $l^{th}$ feature.
Notice feature 0 ($l=0$) is included to accurately account for the normally small and negligible feature contribution of random bootstrapping. For an increasing number of trees, this bootstrapping feature contribution will approach zero. However, if the bootstrapping is stratified $F_{i0}$ and $\tilde{F}_{i0}$ is equal to local increment from training set base rate $\overline{y}$ to the chosen stratification rate in every root node.
Figure \ref{ffExplained} illustrates OOB cross validated feature contributions and regular feature contributions. A so called “one-way feature contribution plot” is a single feature contribution column plotted against the values of the corresponding feature. In Figure \ref{ffExplained} the "one-way feature contribution plot" can be seen as projections of $\tilde{F}$. Conveniently, the main effects of either feature $x_1$ and $x_2$ have been separated with feature contributions before the projection into the 2D plane. In Figure \ref{ffExplained}, the goodness-of-visualization fit to the projected feature contributions can be seen for both $\tilde{F}_{i1}$ and $\tilde{F}_{i2}$. If it is possible to re-estimate the set feature contributions e.g. $\tilde{F}_{i1}$ with some estimator $f$ only by the feature context of the visualization, it is certain, that no interactions have been missed. Thus the model structure do not contain any interaction effect with feature $x_1$. To quantify this we use a leave-one-out cross validation,
\begin{equation}
\label{GOV}
GOV(\hat{f}_\lambda) = cor(\hat{g_{.l}},\tilde{F}_{.l})^2 \quad ,
\end{equation}
here the goodness-of-visualization ($GOV$), is the pearson correlation between LOO predicted feature contributions. Where $\hat{g}_{il} = \hat{f}_{i}^{-i}(X_{i\lambda})$ is the leave-one-out prediction of the $\tilde{F}_{il}$ feature contribution of the $i^{th}$ observation for the $l^{th}$ feature. $\lambda$ is the features which are used to fit the estimator. When $\lambda = l$, $GOV$ quantifies how well feature contribution of the $l^{th}$ feature $\tilde{F}_{.l}$ is explained as a main effect. In Figure \ref{ffExplained} $\tilde{F}_{.1}$ is predicted by $X_{.1}$ and $\tilde{F}_{.2}$ is predicted by $X_{.2}$. GOV can also quantify other visualization contexts than main effect plots. E.g. in Figure \ref{regplot.3d} of result section the goodness of a visualization context of two features $x_3$ and $x_4$ is quantified, where $\lambda = \{3,4\}$.
\begin{figure*}[htp]
\centering
\includegraphics[width=1\textwidth]{figures/fig00f.forsestFloor_explained.140mm.pdf}
\caption{(1) Simulated data set of 5000 observations,
$y_i = f(X_i) = -(X_{i1})^2 - cos(X_{i2})) + \epsilon_i$ where $X_{i1}$ and $X_{i2}$ are drawn from a uniform distribution such that $X_1 \in [-\frac \pi 2 ;\frac \pi 2], X_2 \in [0;8 \pi]$. For all plotted points, a colour gradient (hue color wheel) is used to mark different combinations of $X_1$ and $X_2$. (2) Out-Of-Bag cross-validated predictions $\tilde{y}$ are plotted. (2a/2b) $\tilde{y}$ is decomposed into feature contributions $\tilde{F}_1$ and $\tilde{F}_2$ and projected into a 2D plane, see Equation \ref{FCoob} and \ref{FCsums2y}. Either contain almost only variance from the two main effects $-(X_1)^2$ or $cos(X_2)$. (3) Blue surface depict the full model structure, $\hat{y} = f(X)$. To either side (3a/3b) $\hat y$ is decomposed into $F_1$ and $F_2$, see Equation \ref{FC}. The sum of cross-validated feature contributions by each observation plus the grand mean $\overline{y}$ is equal to the cross-validated predictions, and vice versa for non-cross validated. $F_0$ is the corrections for random or stratified bootstrapping. If no stratification, $F_0$ will be negligibly small. This illustration also generalizes more input features/dimensions and probabilistic classification.
}
\label{ffExplained}
\end{figure*}
\section{Materials and methods}
\subsection{Data and software}
\label{materialDataSoft}
The real datasets \emph{contraceptive method choice} (cmc) and \emph{white wine quality} (wwq) were acquired from the UCI machine learning repository \cite{Cortez2009,Lim1987}. All algorithms were implemented in R (3.2.4) \cite{R2015} and developed in Rstudio (0.99.892) \cite{Rstudio2015}. The main functionality is available as the R-package, \emph{forestFloor} (1.9.5) \cite{forestFloor2015}, published on the repository CRAN. If not stated otherwise all RF models was trained with the CRAN package \emph{randomForest} \cite{Liaw2002} by default parameters except keep.inbag=TRUE in order to reconstruct the individual pathways of observations through the trees. To reproduce result section, R scripts for each data example have been included in the package.
\subsection{Simulating toy data}
\label{materialToyData}
To demonstrate that the visualizations in the result section \ref{resStart} provide correct representations of the data structure, it is beneficial to use simulated (toy) data from a given hidden function. Such functions as Friedman\#1 and 'Mexican hat' are known examples \cite{Alhamdoosh2014}. To illustrate the principal functionality of forestFloor a new hidden function, $G$ is defined. $G$ is the ideal hidden structure, which cannot be observed directly. The toy function was defined as $G(X) + \epsilon = G^*(X) = y = x_1^2 + \frac 1 2 sin(2 \pi x_2) + x_3 x_4 + \epsilon k$ and was sampled 5000 times. $x_i$ were sampled from a uniform distribution $U(-1,1)$. The noise variable $\epsilon$ was sampled from a normal distribution $N(0,1)$ and $k$ was set such that the Pearson correlation $cor(G(X),G^*(X)) = 0.75$. Thus the true unexplainable variances component is roundly 25\% of the total variance. The level of detail, RF can capture from hidden structure G, declines as the noise increases.
\section{Results}
\label{resStart}
Three data sets were modeled with RF regression or RF classification and subsequently explored with forest floor. The examples demonstrate how feature contributions can be used to visualize the data structure and how to identify unaccounted interactions in a visualization.
\subsection{Random forest regression of \emph{toy data}}
\label{resToy}
A default RF regression model was trained on the toy data set with a hidden structure, $y = x_1^2 + \frac 1 2 sin(2 \pi x_2) + x_3 x_4$. Figure \ref{regplot.main} plots feature contribution of all six features against the training set feature values of the toy data. This type of plotting illustrates the main-effects, as feature contributions by each feature were plotted against their respective feature values. Hereby, the mapping surface $S$ was visualized as the sum of $d$ partial functions(black-lines), one for each feature. As the feature contributions retained any variance (main effects + interactions) associated with the node splits by each feature, it was possible to visually verify and test the goodness-of-visualization. Notice that main effect plots of $x_1$ and $x_2$ form nonlinear patterns representing the underlying additive $x^2_1$ and $\frac 1 2 sin(2 \pi x_2)$ contributions to the target y. Therefore, the leave-one-out $R^2$ goodness-of-visualization was $>0.95$ for both these plots. As the explained variance of feature contributions of $x_1$ and $x_2$ was more than 95\% when fitted as main effects, there was no considerable unaccounted interactions. On the other hand, feature contributions of $x_3$ and $x_4$ were poorly explained in main the effect plots. The GOV was poor, less than $R^2<0.1$. It was hence concluded that plotting the one-way feature contributions of $x_3$ and $x_4$ did not assist to explain the structure of $S$. Feature contributions of $x_5$ and $x_6$ were also poorly explained but contained no large variance and were therefore not interesting to explore further. The features $x_5$ and $x_6$ could also be identified as unrelated to the target $y$ for having a very low variable importance (not shown). To include such uncorrelated/unrelated features illustrated the base line of random fluctuations in the mapping structure. This helped to assess whether a given local structure only was a random ripple.
As the feature contributions of $x_3$ and $x_4$ were inadequately accounted for, a broader context was needed to understand the hidden structure. To identify interactions relevant for the feature contribution of $x_3$ a color gradient (red-green-blue) was applied in mapping space $V$ along the $x_3$ axis. The color of any other observation in any other plot was decided by its projected position on the $x_3$ axis. Low values were assigned red and high values blue. Figure \ref{regplot.main} depicts the main effects feature contribution plot of $x_1$,...,$x_6$ with the applied color gradient to $x_3$. Any main effect feature contribution plot of features who neither correlate and neither interact with $x_3$ will show a random color pattern. Such features were $x1$, $x2$, $x5$ and $x6$, which neither correlated nor interacted with $x_3$. Plots of only correlated features would reproduce the same horizontal color pattern. In the extreme case, a feature identical to $x3$ would reproduce the exact same horizontal color pattern. Plots of only interacting features would reproduce the color gradient vertically along the feature contribution axis. A combination of correlation and interaction would make the color gradient reappear diagonally. In Figure \ref{regplot.main} the color gradient suggests, that $x_3$ interacted with $x_4$ due to the vertical color gradient in the plot of $x_4$. In Figure \ref{regplot.3d} their combined feature contributions were plotted in the context of both feature $x_3$ and $x_4$. In this 3D plot it was observed, that the 2D rule of color gradients of interacting features was a basic consequence of perspective. Both color patterns of $x_3$ and $x_4$ could be reproduced by rotating the 3D plot. In this 3D plot, there was no large deviation of feature contributions from the fitted grey. Thus, it was evident that any structure of $S$ related to $x_3$ and $x_4$ were well explained in the joined context of both features $x_3$ and $x_4$.
The GOV of this fit was $R^2>.9$. Therefore, this second order effect plot was an appropriate representation of how $x_3$ and $x_4$ contribute to the target $y$. The depicted saddle-point structure of Figure \ref{regplot.3d} was expected, as the product of $x_3$ and $x_4$ contributed additively to the target $y$. Overall, the model surface $S$, could be represented by two one-way plot of $x_1$ and $x_2$ and one two-way plot of $x_3$ and $x_4$. Hereby the hidden structure of the toy data was fully recovered.
\begin{figure*}[htp]
\includegraphics[width=1
\textwidth]{figures/fig04.simplot_main.190mm.pdf}
\caption{Forest floor main effect plot of a RF mapping structure trained on hidden function $y={x_1}^2+\frac{1}{2}sin(\pi x_2) + x_3 x_4 + k \epsilon$. $x_5$ and $x_6$ have no relation to $y$ and were included only to illustrate a base line signal. A color gradient parallel to $x3$ is applied to identify latent interaction with $x4$. Leave-one-out k-nearest neighbor gaussian kernel estimation provides goodness-of-visualization(black line \& $R^2$ correlation) to evaluate how well each feature contribution can be explained as a main effect.}
\label{regplot.main}
\end{figure*}
\begin{figure*}[htp]
\centering
\includegraphics{figures/fig05.simplot_3d.90mm.pdf}
\caption{One forest floor interaction plot. XY-plan represent feature values $x_3$ and $x_4$ and Z-axis is the summed feature contributions of $\tilde{F}_{i3} + \tilde{F}_{i4}$. goodness-of-visualization is evaluated with leave-one-out k-nearest neighbor gaussian kernel estimation (grey surface, $R^2= .90$). This indicates no remaining latent interactions related to features $x_3$ and $x_4$.}
\label{regplot.3d}
\end{figure*}
\subsection{Random forest regression of \emph{white wine quality} (wwq)}
The previous example of forest floor visualization was an idealized example with uncorrelated features and either representing clear main effect or clear interaction effects. The white wine quality data set (wwq) is an example of mixed main effects and interactions by most features. The target, consumer panel ratings(1-10) of wines, was predicted on basis of 11 chemical features. A default RF model was trained and explained 56\% of variance and the mean absolute error was 0.42 rating levels matching the previous best found model performance \cite{Cortez2013}. To explore the model structure of $S$, first all main effect plots were inspected. Figure \ref{wwq.1.main} depicts all plots by all 11 features. Features were sorted in reading direction by variable importance to present most influential feature first. A color gradient along the most influential feature, \emph{alcohol}, was applied to search for interactions. Hereby it was observed that \emph{density} was negatively correlated with \emph{alcohol}, that \emph{volatile acidity} interacted with \emph{alcohol} and that \emph{residual sugar} both correlated and interacted with \emph{alcohol}. The observed correlation between \emph{residual sugar}, \emph{density} and \emph{alcohol} is trivial, where low-density \emph{alcohol} linearly lowers \emph{density} while high-density \emph{residual sugar} increases \emph{density}. Close to 98\% of the scaled variance of these three features can be described by two principal components. This information redundancy was expected to affect variable importance of the three implicated features and to lower the general variance of the respective feature contributions. Although the overall structure suggested that alcohol content in general was associated with higher preference scores, there was a local cluster identified as low \emph{alcohol}, high \emph{residual sugar} and low \emph{pH} which was associated with high preference scores also. Figure \ref{wwq.1.main} suggested that wines could achieve a high preference score when \emph{residual sugar}$\approx$17, \emph{pH}$\approx$2.9, \emph{citric acid}$\approx$.35 and \emph{fixed acidity}$<$7 despite a low alcohol content. Such white wines was perhaps by the consumer panel attributed fruity and fresh. Any found interaction could be investigated with several color gradients and two-way forest floor plots. It was chosen to investigate the interactions of \emph{volatile acidity}, as this feature was the third most important feature, whereas the goodness-of-visualization of the one-way forest floor plot was only $R^2=0.69$. Two-way forest floor plot was therefore a more suitable representations of this effect. The color gradient along alcohol content already suggested a notable interaction between \emph{volatile acidity} and \emph{alcohol}. Figure \ref{wwq.volatile2way} depicts the two-way forest floor plot of feature contributions of \emph{volatile acidity} in the context of itself and the feature \emph{alcohol}. The goodness-of-visualization was then $R^2=0.94$. Therefore, the residual variance of feature contributions not explained by this plot was low.
For wines with alcohol content more than 10\% (blue area) \emph{volatile acidity} appeared slightly positively to preference score. For wines with lower than 10\% \emph{alcohol} (red area) \emph{volatile acidity} appeared to contribute negatively to preference score.
\begin{figure*}[htp]
\centering
\includegraphics[width=1
\textwidth]{figures/fig06.wine_main.190mm.pdf}
\caption{Forest floor main effect plots of random forest mapping structure of model predicting panel ratings of 4900 white wines on basis of chemical properties. The plots are arranged according to variable importance. X-axis are variable values and Y-axis the corresponding cross validated feature contributions. Color gradient in all plots are parallel to the feature \emph{alcohol} (content w/w). goodness-of-visualization is evaluated with leave-one-out k-nearest neighbor estimation (black line , $R^2 values$)}
\label{wwq.1.main}
\end{figure*}
\begin{figure*}[htp]
\centering
\includegraphics{figures/fig07.wine_volatile_3d.90mm.pdf}
\caption{Forest floor interaction plot: Feature contribution of \emph{volatile acidity} versus feature values of \emph{volatile acidity} and \emph{alcohol}. Color gradient is parallel to \emph{alcohol} axis. goodness-of-visualization is evaluated with leave-one-out k-nearest neighbor estimation (grey surface and $R^2 = 0.93$)}
\label{wwq.volatile2way}
\end{figure*}
\subsection{Random forest multi-classification: \emph{Contraceptive method choice} (cmc)}
\label{resCMC}
To illustrate the capabilities of forest floor for multi-classification the data set cmc was chosen. The data set originates from a survey of 1473 non-pregnant wives in Indonesia in 1987 comparing current choice of contraception with socioeconomic features. These features were, \emph{wives' age} (16-49), \emph{wives' education level} (1-4), \emph{husbands' education} (1-4) , \emph{n\textunderscore children} (0-16), \emph{wives' religion} (0 (not islam), 1 (islam) ), \emph{wives working} (0 (yes), 1(no)), \emph{husbands' occupation} (I,II,III,VI), \emph{standard-of-living index} (1-4), \emph{media exposure} (0=Good, 1=not good) and the target \emph{contraceptive method choice} (1=no-use (629), 2=long term(333), 3=short term (511)).
In the \emph{cmc} data set the choice of contraception was far from fully described by the available features \cite{Lim2000}. The OOB cross validated RF model error-rate was $.44$. Assuming wives did not use contraception (the most prevalent case) yielded a $\frac{629}{1473} =.57$ error rate. Anyhow, if the RF model performance would be regarded as good by domain specialists, the model structure could possibly provide insights to the socioeconomic mechanisms in play. Hyper parameters \emph{Sample size} and \emph{mtry} were tuned to yield the best OOB cross validated performance. Optimal parameters was found to be bootstrap \emph{sample size}= 100 and \emph{mtry} = 2. A lower \emph{sample size} can increase robustness by tree decorrelation but also introduce more bias. To lower \emph{sample size} of trees can be advantageous, when explained variance component is less than 50\%. Thus a RF model different from default settings, was chosen to slightly improve predictions and to simplify/smooth the mapping structure to explore. Hereby the mapping structure may better represent the underlying social/economic mechanisms, that the specific data structure of survey reflects.
Three types of plots were constructed to investigate the mapping structure. As the number of features was $d=9$ and number of classes was $c=3$, a full dimensional mapping space visualization would require 12 dimensions. As shown in Figure \ref{mapTypes}, probability axes can be aligned along the y-axis, to reduce the number of dimensions to represent prediction space to only one. Also, when the cross validated predictions were decomposed into cross validated feature contributions, only 2 dimensions were needed to plot any main-effect. These plots resembled one-way forest floor regression plots although coloring was reserved to identify class of predicted probability. Otherwise each class by each feature would need to be plotted separately. Black assigns no usage. Red assigns long-term usage and green assigns short-term usage. Figure \ref{cmc.mainAlign} illustrated the main effects of each feature of a RF-fit, the y-axis describes the additive change of predicted probability for each observation for each each class. The actual feature value for each observation was depicted by the x-axis. Thus any observation were placed three times in each plot by the same feature value in three colors once for each three classes. The sum of changed probability over classes for any observation must be zero, see Equation \ref{nodePredClassSumToOne}. Overall, Figure \ref{cmc.mainAlign} showed that main effects were dominant, as most variance was explained by the respective features. \emph{n\textunderscore children} was the most important feature strongly predicting (probability change up to +/- .30) that wives with 0 or 1 child tended not to use contraception. On the other hand, more than 4 children predicted a slight increase in either type of contraception. Except for a preference separation for long-term contraception over short-term for wives with more 7 children, the \emph{n\textunderscore children} feature was not found useful to predict the choosing betwen the two types of contraception. \emph{Wives's education} especially separated between no-use of contraception and long-term use, where lowest level predicted up to +/-10\% probability change. With more education the wives tended to use long-term contraception over no usage. The use of short-term contraception was comparably unchanged as a function of \emph{wives' education}. \emph{Wives' age}, the third most important feature, favored short-term contraception for wives younger than 30, while long-term and no contraception for wives elder than 30. After 40 years, either use of contraception declined. \emph{Husbands' education} elicited same pattern as \emph{wives' education} though size of effect was half. A small subgroup of 7\% was reported to have a not good \emph{media exposure} and this predicted a probability increase in no contraception of 8\%. Type of \emph{Husband' occupation} favored for category \emph{I} long-term by 5\% over short-term, whereas category \emph{III} predicted an opposite 3\% effect. Standard of living predicted a pattern much similar to \emph{husband's eduction}. A small subgroup (15\%) of wives were not muslim, and this predicted a 5\% increase in short-term contraception over long-term usage and no usage. Lastly for a subgroup of 25\% working wives was predicted a very slight increase (2\%) of no-usage over short-term.
\begin{figure*}[htp]
\centering
\includegraphics[width=1
\textwidth]{figures/fig08.cmc_main_align.190mm.pdf}
\caption{Cross validated feature contributions for each feature for each class(black, red, green) and for all training observations plotted against the corresponding feature values. Categorical features are coded with integers. Feature contributions can be understood as change of predicted class probability attributed to a given feature.}
\label{cmc.mainAlign}
\end{figure*}
The main effects for this 3-class problem could also be depicted as a series of $(3-1)$-dimensional simplexes, where the position in the triangle depicts the predicted probability distribution for any observation. Colors can either depict true class (black: no-usage, red: long-term and green: short term) or colors can depict a feature (low value (red), middle (green), high(blue)). Figure \ref{cmc.mainSimplex} depicts all main effects in bi-simplex plots, with left simplex colored by cross-validated true class separation, and right simplex colored by feature value distribution across the simplex space. Figure \ref{cmc.mainSimplex} depicts 10 pairs of simplexes. Lines were added to the simplexes to illustrate majority vote. Only 17\% of wives were predicted to use long-term contraception even though 22\% of the sample population did so. Because RF models effectively used the sampled base rate as prior (marked as a blue cross) and the effective separation was weak, predictions tended to be skewed towards largest class away from smallest class. A different prior than the sampled base rate could be set by stratified bootstrapping of each tree in a random forest model. E.g.\ to stratify sampling by target class would move the blue cross to the middle of the simplex, and roughly a third of predictions would fall into either class. Stratified bootstrapping would e.g. be reasonable if the preferred contraception is expected to be different in the full population than in the training population.
In the second total separation simplex, to present an overview of any differences in socioeconomic status, principal component analysis was used to reduce the full feature space to two principal color components. Here a purple cluster indicated no-usage, a green cluster was shifted towards long-term usage, light blue cluster predicted short-term usage, and a dark-blue cluster predicted short-term or no usage. The color separation was not perfect, partly because the separation problem was difficult and partly because PCA cannot fully characterizes a potential nonlinear mapping surface of randomforest. To colour be several features at the same time, seemed to be most useful for data sets with high linear feature collinearity.
The left of following bi-plots of simplexes depicted the effective separation of true class separation by any feature contribution. The right simplex depicted the separation as a function of the corresponding feature (by color). This second simplex could be used both to illustrate the main effect of each feature and to assess whether higher order effects were present. For features with small set of levels such as womans education, a separation in four clusters (red(1), brown(2), pale blue(3), deep blue(4)) could be seen. Education level 1 and 2 were partly joined. The local centroids of these cluster was interpreted as the main effect, and the deviation from the centroids as higher order effects + unfiltered noise. For all simplexes the global centroid and prior is the (blue cross).
The series of bi-plot simplexes of Figure \ref{cmc.mainSimplex} could illustrate with finer detail the predicted probability distribution for any observation, whereas the precise feature value was depicted with less fidelity than in Figure \ref{cmc.mainAlign}.
The three features \emph{media exposure}, \emph{wives' religion} and \emph{wives working} were binary and showed the largest change of predicted probability in the smallest subgroups. This observation was regarded trivial, as the group size weighted probability change across a binary feature split must have equal size. Thus few observations can change prediction a lot, if many observations only change prediction a little in a opposite direction. This was regarded a property for all binary decision tree models and Figure \ref{LIclass} in Section \ref{TheoLI} depicted a similar pattern of how local increments would propagate in a probability simplex.
\begin{figure*}[htp]
\centering
\includegraphics[width=1
\textwidth]{figures/fig09.cmc_main_simplex.190mm.pdf}
\caption{From top left: Cross validated predicted class probability colored by true class and a PCA color gradient describing observation diversity. Following pairs of plots, were the predicted probability decomposed into feature contributions. Left colored by true class, right colored by corresponding feature value. Red is minimal value, blue is maximal value. Blue cross is class base rate of training set. Dashed lines are drawn manually to assist interpretation of main effects.}
\label{cmc.mainSimplex}
\end{figure*}
To search for higher order effects, similar to forest floor regression, simplex plots can in turn be colored by other features. In Figure \ref{cmc.2ndSimplex} the simplex plots of \emph{wives' age} and \emph{wives' education} was printed 3 times each. From left to right, color gradients illustrated respectively \emph{wives' age}, \emph{wives' education}, and lastly \emph{n\textunderscore children}. The simplexes in the diagonal reproduced the main effect coloring from Figure \ref{cmc.mainSimplex}, whereas other depicted simplexes possibly would detail 2\textsuperscript{nd} order interactions. E.g. \emph{wives' education} of Figure \ref{cmc.2ndSimplex} showed the four clusters, one for each education level. The distance from any point to its local cluster as a mix of higher order effects and a small noise component. It was found that wives with highest education aged 20 were predicted more likely to use contraception than when aged 25. Wives' with highest education and few children (red) preferred short term contraception over long term. As the features \emph{n\textunderscore children} and \emph{wives' age} are correlated, these will both interact with \emph{wives' education}, not only one.
\begin{figure*}[htp]
\centering
\includegraphics{figures/fig10.cmc_2nd_simplex.140mm.pdf}
\caption{Feature contributions for the three most important features plotted row-wise. Each plot is colored column-wise by corresponding feature values. Dash lines are drawn manually to assist interpretation of interactions.}
\label{cmc.2ndSimplex}
\end{figure*}
\section{Discussion}
Forest floor is a methodology to visualize the mapping structure of a RF model using feature contributions. RF can be termed a predictive algorithmic model, designed to have a high predictive accuracy on the expense of model transparency \cite{Shmueli2010, Breimann2001}. RF could also be termed as data driven, as the model can adapt itself to the data with little guidance. The opposite is a theory driven model where the user manually choose an explicitly and clearly stated model to capture the data structure. A practical advantage of using RF, is when the user have little prior knowledge or theory on the subject. The majority of nonlinear machine learning algorithms models have in common, that the resulting model stated as an equation is fairly complex in the eyes of a human user. The complexity may be difficult to avoid if the model should be able to capture an unknown structure. But exactly when little prior theory is given, that is when the model should inspire the interpretation of the data structure. A dualistic approach is to choose both a perhaps linear explanatory model to interpret the system and a machine learning algorithm to get the most accurate predictions \cite{Shmueli2010}. Such an approach may leave a gap between users comprehension and the actual structure of the nonlinear model. If the user is far from understanding a certain data-structure any optimization cannot hardly evolve from brute trial-and-error searches such as grid search or ant-colony-optimization methods.
For nonlinear high-dimensional multivariate models, it is not straight forward to visualize the trained mapping function. The provided visualizations can be understood as slices or projections of the mapping structure. It appears that a given series of 2D and/or 3D projections can jointly explain the structure of a RF mapping surfacesa. The quantifiable goodness-of-visualization measure describes how well the variance of the full structure can be explained in the context of the provided feature axis. If a large component of feature contribution variance remains unexplained, there is likely an unaccounted interaction pattern associated with this feature. Thus an advantage of forest floor is, that it aids the user to learn what local interaction effects are not yet visualized. With feature contributions it is possible to make an interpretation of what variance is attributed main effects, second order effects or higher order effects. Feature contributions can be computed from the training set itself and thus do not extrapolate the training set. The training set is used to set boundaries for model structure, such that extrapolated and unrelated model structures are not visualized. Feature contributions can be combined with the out-of-bag concept allowing cross validation to avoid presenting an overfitted mapping structure. Visualizations of cross validated feature contributions appear less noisy.
Color gradients allowed to include one or two extra dimensions in an illustration thus otherwise limited of three dimension. Color gradients traversing entire mapping space was used to highlight selected latent dimensions in a series of main effect plots to pinpoint missing interactions. We perceive colors as a combination of three channels red, green and blue. Thus, it may seem possible to visualize three additional dimensions in colors. Nonetheless, the ranges of color saturation and brightness should be constrained to avoid indistinguishable grey color tones and to ensure a minimal contrast to the background. Such considerations, limited color gradients to provide only two additional dimensions at maximum. It was possible to summarize a high-dimensional structure with e.g. principal component analysis and apply color gradients along the first 2 loading vectors, such as in Figure \ref{cmc.mainSimplex}. In practice, we found a sequence of 1-dimensional color gradients best suited to uncover latent interaction structures in a RF model fit.
Feature contributions were first described in the context of RF regression, where a given feature can contribute either positively or negatively to a given prediction \cite{Kuz'min2011}. Next, the concept of feature contributions has previously been extended to classification, where the categorical majority vote labeling were replaced with numeric probability predictions \cite{Palczewska2014}. We have argued that these probabilistic predictions are confined in a prediction space defined the ($K-1$)-simplex, for model with $K$ classes. Any node in any tree will itself be a prediction and have a position in this space. We argue local increments are in fact vectors connecting nodes in the ($K-1$)-simplex space. The first local increment (the bootstrap increment) of any tree will be the vector connecting the class distribution of the training set to the class distribution of the root node. As the bootstrap increments will point randomly in any direction, the sum of a large number of such will approach the zero vector if no stratification is chosen. For stratification by true class, the bootstrap increments will connect the training set class distribution point in the ($K-1$)-simplex to the point in the ($K-1$)-simplex chosen by stratification.
The Gini loss function can be understood as maximizing the squared distance of node positions to the center of ($K-1$)-simplex (equal class probability). Therefore any split by Gini will place the daughter nodes the furthest from the center, weighted by node size. As the classification trees are fully grown, the terminal nodes of one pure class can only be positioned on the vertices of the simplex. In Figure \ref{cmc.mainSimplex} was shown that the distribution of classes in the training set will function effectively as the prior of the RF model. If the user do not expect to find the same class distribution in future predictions as in training set, this prior can be moved in the simplex by stratification during the bootstrap process. In Figure \ref{cmc.mainSimplex} the center blue cross marked that the average root node center was skewed towards class 1 (no contraception) as 42\% of the wives did not use any contraception. As class separation by the RF model was not strong the majority of predictions fall close to this prior base rate. In supplementary materials a RF model was trained with bootstrap stratification by true class such that the average root node is positioned in the center of the ($K-1$)-simplex and following predicted class probabilities were also centred around this point. Figure \ref{LIclass} depicted how any node-split will produce two new nodes with local increments in perfectly opposite direction. Thus, training set predictions will always be centred around this point.
Direct plotting of $K$ class probabilities requires $K-1$ dimensions. This is possible for 3 or 4 classes with 2D plot or 3D plot respectively. The context of feature values can only be included as one extra axis or as color gradients. We have shown that the axis of the ($K-1$)-simplex can be aligned such that only one axis is needed to visualize the feature contributions as seen in Figure \ref{cmc.mainAlign}. This frees 1 or 2 axis to provide an adequate feature value context. In such visualization each observation will be plotted one time for each predicted class probability. Colors can be used to distinguish the classes.
In a previous article we trained a molecular descriptor model with RF to predict protein permeation enhancement in an epithelial cell model (Caco-2) \cite{Welling2015}. A diagnostic tool was missed to address why such a model would be credible and to communicate intuitively the found pattern to fellow chemists/biologist with little knowledge of machine learning. We first stumbled upon feature contributions in the two articles \cite{Palczewska2014,Kuz'min2011} and experimented to plot these feature contributions against the feature values. The R package rfFC \cite{rfFC2013} provided the first computations of feature contributions and was an inspiration to the design of the forestFloor package \cite{forestFloor2015}. Hereafter we discovered partial dependence plots and sensitivity analysis \cite{Cortez2013,friedman2001}. Now in hindsight we can report the set of advantages to forest floor, especially the tracking of unaccounted interactions such that no strong interaction will be overlooked when visualizing the mapping structure.
The following citation by Friedman \cite{friedman2001} originates from an article from 2001 discussing the usefulness of partial dependence plots on nonlinear functions: \textit{"Given the general complexity of these generated targets as a function of their arguments, it is unlikely that one would ever be able to uncover their complete detailed functional form through a series of such partial dependence plots. The goal is to obtain an understandable description of some of the important aspects of the functional relationship."} \cite{friedman2001} \par
Indeed the structure of RF models can be highly complex and visualizations are unlikely to present every detail at once. Therefore a visualization tool-set should assist the user to navigate the mapping structure. This has been done by isolating the part of the model structure related to the data structure, by evaluating the goodness-of-visualization of a given plot, and by pointing to where locally in the model structure a sizable latent interaction is not yet visualized. Our goal is to present complex models as adequately detailed visualizations. In a RF model there will likely always be a baseline of random ripples in the mapping structure, that we do not expect to be able to reproduce. These ripples are partly filtered of by using the out-of-bag cross validated feature contributions. Other ripples occur due to biases of the RF algorithm. Especially does the RF model structure surface contain wave like curvature parallel to the feature axes due to the univariate step functions of RF, see RF surfaces in Supplementary Materials.
We predict that 4D projections of a third order interaction rarely would be needed for the RF algorithm. In supplementary materials we have provided a simulation suggesting that RF only poorly can fit interactions higher than second order even when trained on 10.000 observations without any noise. This can be explained as the RF algorithm is limited in its potential complexity as the algorithm only can perform univariate splits decided by an immediate loss function. Another algorithm such as rotation forest \cite{Rodriguez2006} is not limited to perform univariate splits and therefore better on such simulated tasks with higher order interactions. What initially was an interaction effect can be rearranged into a main effect by new combined features. Multivariate split methods are not compatible with forest floor, but they are compatible with the generic methods partial dependence plots and sensitivity analysis \cite{friedman2001,Cortez2013}.
\section{Conclusion}
Forest floor has extended the tool-box to visualize the mapping structure of RF models. The geometrical relationship between random forest models and feature contributions has been described. For RF multi-classification it was useful to understand the prediction space as a ($K-1$)-simplex probability space. Hereby the feature contributions can be interpreted as changes of predicted probability due to a given feature. A ($K-1$)-simplex prediction space can also visualize how the training set stratification affect RF predictions. Target class stratification is effective to modify the prior for the RF model.
We have emphasized that parts of a mapping structure which extrapolates the training set are irrelevant. To extract only the relevant mapping structure, feature contributions are computed only from the training set itself. Two new variants of feature contributions have been introduced to avoid inherent overfitting when using training set predictions. These variants of feature contributions are out-of-bag cross validated feature contributions, and n-fold cross validated feature contributions.
Feature contributions from a single feature can contain variance from main effects and/or interaction effects. A measure of goodness-of-visualization has been introduced to evaluate if the feature contributions of a given feature alone can be explained in the context of itself. If not, color gradients traversing the mapping space can be used to pin-point overlooked interactions within feature contributions and features. Sizable interactions can be visualized in two-way interaction plots in the context of two features and perhaps even a third feature as color gradient. Again a goodness-of-visualization can be computed and evaluated for such a visualization.
Ultimately, it is difficult to communicate a context of more than 2 or 3 dimensions + target dimension(s). Thus fourth order interactions would be difficult to visualize and communicate. Anyhow, such visualizations are likely not missed, as the random forest algorithm could not fit fourth order interactions well and had a poor efficiency already with third order interactions.
As forest floor can break down a RF model fit into effects attributed to each feature and assist to find adequate context to understand these effects. It is intended that RF no longer should be seen as a non interpretable model. Learned associations between features and targets should inspire new ideas of the underlying possible causality structure.
\clearpage
\end{multicols}
\clearpage
|
\section*{Significance statement}
Variability and covariability in neural activity are some of the most puzzling aspects of the otherwise impressively effective neural computations. For efficient decoding and prediction, models of information encoding in neural populations hinge upon an appropriate model of neural variability. Our work shows that stimulus-dependent changes in the pairwise statistics but not in single-cell statistics can differentiate between two widely used models of neuronal variability. Importantly, contrasting model predictions with neuronal data provides hints on the sources of noise in neural spiking. These findings provide essential constraints on statistical models of population activity.
\section{Introduction}
Variability in the nervous system is ubiquitous and affects perception, decision making, and motor control \citep{Faisal:2008cp}. Characterizing the properties \citep{Ecker:2014cl, Goris:2014jg, Kohn:2005um, Lin:2015dw} and identifying the sources \citep{Renart:2014dr} of this variability is critical for understanding the computations taking place in the nervous system. In particular, in the visual system, understanding how cellular, network, or bottom-up processes contribute to variability helps to assess the interaction between the stimulus, internal states of the brain and possible noise. In order to tackle variability, the goal is to be able to consistently predict response statistics of neurons beyond the mean response under different conditions.
The overwhelming majority of data recorded with a focus on the assessment of response variability comes from recordings of spiking activity of neurons \citep{Churchland:2010he}. Spike count measurements are characterized by a distinctive pattern that relates variability to firing rates: there seems to be a tendency that spike count variability grows linearly with mean firing rate \citep{Tolhurst:1981ju}. This form of single-cell statistics is central to many theories of coding \citep{Churchland:2011hd, Ecker:2016hc, Jazayeri:2006fk, Ma:2014in, Simoncelli:2004ue, Pillow:2007wh, Froudarakis:2014fs}: the resemblance of this statistics to the characteristics of the Poisson process led to a simple statistical model of variability, which assumes that variability arises as a consequence of a renewal process with independent spikes sampled in finite time windows with a given expected value but independent across time windows.
While Poisson spiking statistics seems to be a sound model of variability, it suggests a very specific assumption on the source of noise in neural responses: stochasticity is assumed to arise at the level of spike generation. Variability can arise at different levels of cellular and network dynamics: at the level of membrane potential (for instance because of channel noise), at the level synaptic transmission, at the level of uncontrolled variables either in the stimulus, in the movements of the animal, or internal to the processing (e.g. attention, or top-down processing) \citep{Ruff:2014fa}. Recent studies have focused on identifying different sources of response variability by distinguishing the contributions of Poisson-like stochasticity and additional fluctuations that further increase variability \citep{Goris:2014jg, Churchland:2011hd}.
The assumption that it is a stochastic process in spike generation that underlies spiking variability has been challenged by the highly reliable spike generation mechanisms found in vitro \citep{Mainen:1995uz}. An alternative approach, the Rectified Gaussian model (RG), which is compatible with this result but can also account for the characteristic mean-variance relationship of spike count statistics, has been proposed by \citep{Carandini:2004ee}. According to this approach, firing rate nonlinearity, the mapping between membrane potential to firing rate, by itself can achieve the scaling of spike count variance with spike count mean. As opposed to a Poisson-like account, spiking variability in the RG model originates from variability present at the level of membrane potentials. A linear relationship between spike count mean and variance is achieved by increased stretching of the normally distributed membrane potential by the convex firing rate nonlinearity when the mean membrane potential is higher. This model has also proven successful in accounting for patterns in mean spiking responses of V1 neurons \citep{Finn:2007hc}.
Stimulus-dependence of response mean and variance, often characterized by the Fano factor, provides important insights into the forms of stochasticity responsible for response variability. Going beyond single-cell response statistics by analyzing joint statistics of the responses of multiple neurons can provide the constraints necessary to dissect possible mechanisms responsible for spiking variability. Collective changes, or correlations in their simplest form, have recently gained intense attention \citep{Ecker:2010dn, Cohen:2011eh}. Studies have demonstrated that bottom-up \citep{Kohn:2005um}, top-down \citep{Ruff:2014fa, Haefner:2016vr} and even cellular \citep{deLaRocha:2007go} factors can affect spike count correlations and others have revealed that both multiplicative and additive forms of correlations contribute to these changes \citep{Lin:2015dw}. It is unclear, however, how well alternative models of spike count variability can predict the patterns in response correlations as well as the patterns in mean responses and response variance.
In this study we set out to contrast competing approaches proposed for describing spike count variability and use their predictions on response statistics. We define the Doubly stochastic Poisson (DSP) model and the Rectified Gaussian model to analyze the relationship between membrane potential statistics and spike count statistics. In particular, we focus on the joint spiking statistics of a pair of neurons to demonstrate a dissociation between the two models based on changes in spike count correlation resulting from changes in membrane potential statistics. Using these analyses as a starting point, we simulate stimulus change-related modulation of spike count statistics for both of the models in a population of model neurons and compare the predictions to orientation-dependent and contrast-dependent changes in the activity statistics of extracellularly recorded V1 neurons in awake monkeys. We argue that changes in stimulus attributes give rise to distinctive patterns in response correlations that are compatible with the Rectified Gaussian model but contradict the Doubly Stochastic Poisson model.
\section{Materials and Methods}
\subsection{Model of membrane potential responses}
We model the responses of simple cells of the primary visual cortex at the level of the membrane potentials. The membrane potential of a simple cells shows systematic variations with changes in the stimulus. The most widely studied of these changes, the orientation dependence was considered as changes in the trial-by-trial mean of the membrane potential response. Trial-averaged responses provide an incomplete description of the membrane potential statistics since there are both within-trial fluctuation and across trial fluctuations \citep{Tomko:1974ul}. The source of these fluctuations are not considered. A neuron is characterized by the probability distribution of membrane potential values at each time point within a trial, which random samples were drawn from. Importantly, besides stimulus-driven covariations in membrane potential responses of multiple neurons, stimulus-independent, so called noise covariances are also characteristic features of cortical neurons \citep{Yu:2010p314}. As a consequence, a population of neurons in the visual cortex is also characterized by the joint probability distribution of the membrane potentials. Similar to systematic variations in trial-averaged mean membrane potential responses, membrane potential variances are characterized by systematic changes \citep{Finn:2007hc}, which we also take into account in the model.
Instead of aiming for a complete description of such probability distributions, we focus on the second-order statistics of neural responses. This is motivated by two considerations: (i) experimental designs are typically limited in terms of the number of trials, rendering higher-order joint statistics of neurons hard to estimate (ii) pairwise statistics are the simplest measure of population activity going beyond individual cell response properties, and are already able to capture definitive signatures of population-level cortical computation \citep{Karklin:2009hl, Haefner:2013gh}.
The second-order statistics of a population of $N$ units with temporal dynamics is completely characterized by $N$ autocorrelation functions and $N(N-1) / 2$ cross-correlation functions, specifying the linear dependence between each pair at every timescale \citep{MorenoBote:2008gg}. Based on the typical support of the autocorrelation function of membrane potentials, we chose to examine interactions on the scale of 20 ms \citep{Azouz:1999wj}. Using this, we can simplify the above picture greatly by specifying a single membrane potential value in each 20 ms bin, assuming that there is no temporal dependence on longer timescales, and omitting variability on shorter ones. Thus, we can consider membrane potentials to be sampled from a Gaussian distribution at each time bin $b$ independently:
\begin{equation}\label{eq:mp_gauss}
\mathbf{u_b} \sim \mathcal{N}(\mathbf{u_b}; \mathbf{\mu},\mathbf{C_{mp}}).
\end{equation}
The mean of the distribution, $\mu \in \mathbb{R}^N$, determines the trial-averaged membrane potential level of each neuron, and the sequences of individual samples give rise to within-trial variability.
Since the focus of the study is to understand the implications of membrane potential fluctuations on spiking statistics, we do not seek to find a match between membrane potential recordings and the model. Rather, membrane potential statistics is assessed in terms of its consequences on spiking statistics.
\subsection{Doubly Stochastic Poisson spike generation model}
Intensity of spike responses of neurons is determined by the instantaneous firing rate function. This mapping from membrane potential to firing rate is achieved by the firing rate nonlinearity, for which we take a parametric form from the literature \citep{Carandini:2004ee}, parametrized identically for each neuron:
\begin{equation}\label{eq:firing_rate}
\mathbf{r}_b = k \left[ \mathbf{u}_b - V_{th} \right]_{+}^\beta.
\end{equation}
$V_{th}$ denotes the membrane potential threshold under which the firing rate is zero. Above the threshold the firing rate is a power law function with exponent $\beta$, where $\beta=1$ corresponds to a linear mapping. The rate is mapped to the Hertz scale using a gain parameter, $k$, which indicates how many spikes are to be expected within a single time bin (which is always 20 ms in our study). Using this mapping, we obtain an instantaneous rate in every time bin, which serves as an intermediate quantity between membrane potentials and spikes. The parameters in the rate model are chosen to be typical to the primary visual cortex based on Carandini, 2004. The values used throughout this paper are $V_{th} = 0$ (the threshold only contributes to the rate through the difference with the membrane potential, thus we need to choose $\mathbf{\mu}$ and $V_{th}$ together to produce realistic membrane potential dynamics), $\beta = 1.4$, and $k = 0.4$, which together with the 20 ms time bin corresponds to a rate of 20 Hz, typically observed in the primary visual cortex in response to a high contrast stimulus of parameters preferred by the tuning curve of the cell \citep{Finn:2007hc}.
The firing rate nonlinearity establishes the link between average membrane potential and average firing rate. A widely used approach \citep{Gur:1997ur} assumes that the firing rate determines the probability with which a spike is generated in any particular time window. When these probabilities are independent across time we formally obtain the Poisson process. In this model, spike counts are sampled from a Poisson distribution, parametrized by the instantaneous rate, independently for each neuron $n$.
\begin{equation}\label{eq:poisson_sampling}
s_b^n \sim \mathrm{Poisson}(s_b^n; r_b^n)
\end{equation}
In summary, the Poisson model of spiking, the Doubly Stochastic Poisson model (DSP), relies on two sources of variability: 1, membrane potentials are stochastically generated and these samples are correlated to represent covariability of neuronal responses; 2, a second source of stochasticity comes from the generation of spikes which introduces noise that is independent across neurons.
\subsection{Rectified Gaussian model}
The DSP model assumes that the firing rate defines the probability with which spikes are generated in a given time window. Inspired by the relatively little stochasticity found in sensory neurons in vitro \citep{Mainen:1995uz}, an alternative model can be formulated, which assumes a deterministic process for spike generation \citep{Carandini:2004ee}. Self-consistency requires that the number of spikes with constant stimulus on average is proportional to the firing rate. A model that assumes no further source of variability but fulfils the self consistency criterion can be formulated by integrating the firing rate over time and generating spikes whenever the integral crosses integer values. If we consider the rate to be constant within a time bin $b$, scaled appropriately with the base rate parameter $k$, the integral becomes a finite sum, for neuron $n$:
\begin{equation}\label{eq:sc_integration}
s_t^n = \lfloor \sum_{b=1}^{\tau} r_b^n \rfloor
\end{equation}
The spiking model defined this way, the Rectified Gaussian model (RG), is formally equivalent to an integrate-and-fire neuron model without refractory period with the addition of the firing rate nonlinearity. The principal source of variability in the RG model is coming from the variability present in the membrane potentials. An additional, though minor, source of variability for the timing of spikes originates from an uncertainty of the state of the integrator at the beginning of a trial.
A cartoon depicting the procedure of generating firing rates from Gaussian membrane potentials, and then spike counts using the two models, is summarized in Fig.~\ref{fig:cartoon}.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig1_dsp_rg_cartoon.pdf}
\end{center}
\caption{\textbf{Schematic illustration of the RG and DSP models.} \textbf{\textit{A}}, Membrane potential responses are characterized by a mean activation and a stochastic component which varies over time. The mean is assumed to be stimulus-dependent, and is determined by the orientation tuning curve. Importantly, the stochastic component is not independent but is correlated among neurons. \textbf{\textit{B}}, Firing rates are calculated from membrane potentials by transforming them using the firing rate nonlinearity. \textbf{\textit{C}}, The way spike counts are obtained from the firing rate is determined by the spiking model. \textbf{\textit{D}}, According to the Rectified Gaussian (RG) spiking model, spikes are generated by integrating the firing rate, and deterministically registering a spike every time the integrated rate crosses an integer value. \textbf{\textit{E}}, In the Doubly Stochastic (DSP) spiking model, spike counts are stochastically generated: the time varying firing rate (normalized by the time window being considered, diamonds, left panel) determines the mean of the Poisson distribution (diamonds, right panel), which assigns probabilities to the number of spikes to be generated in the time window (right panel).}
\label{fig:cartoon}
\end{figure}
\subsection{Simulation of population measurements}
In order to compare the predictions of the RG and DSP models with experimental recordings of activity of a population of neurons, population models were constructed and spike response statistics of the two models were contrasted with experimental data. The primary goal of these comparisons was to test predictions of the two models on the population level, where summary statistics are less susceptible to sampling noise, and their measurements do not require the precise control of receptive field contents. We simulated full-field gratings as stimuli in the simulated experimental paradigm.
Determining the membrane potential statistics for the population of $N$ neurons requires the specification of their means, variances and correlations, as specified in Eq.~\ref{eq:mp_gauss}. The mean membrane potential response $\mu^n$ for cell $n$ is determined by stimulus orientation through the tuning curve that is characteristic to the particular neuron. The stimulus orientation at which the peak of the tuning curve is located corresponds to the preferred orientation of the neuron and preferred orientations are sampled independently for each neuron from a uniform distribution. The height of the tuning curve is varying from neuron to neuron and is sampled from a Gaussian distribution with a standard deviation of 0.1. Widths of the tuning curve is fixed at $SD_{TC} = 0.2\pi$. The time varying stochastic component of the membrane potential response is coming from a Gaussian distribution which is characterized by its variance. The level of variance was inhomogeneous across the population and was set randomly by sampling an inverse gamma distribution. Parameters of the inverse gamma distribution are the shape and scale parameters, with values three and four, respectively, chosen to reproduce the scale of spike count Fano factors observed in experiments. Membrane potential correlation matrices are generated algorithmically by specifying the width of the distribution with a scalar parameter \citep{Lewandowski:2009gd}. Distribution of membrane potential correlations were tuned such that the distributions of spike count correlations were matched for the RG and DSP models.
Simulation of changes on stimulus orientation is straightforward since tuning curves define the changes in membrane potential mean, while other aspects of the statistics are assumed to be unchanged. This choice is motivated by studies on membrane potential variance, which demonstrated that variance is relatively intact by changes in stimulus orientation \citep{Finn:2007hc}. The question of the orientation dependence of membrane potential correlations is still open but we took a conservative approach by assuming orientation independence at the level of membrane potential correlations. Modulation of stimulus contrast was simulated by adjusting both the gain of the tuning curve and the level of variance. Lowered contrast caused the mean membrane potential response to scale down to a level half of that at high contrast, while the variance increased to 1.4 times the original level.
When the membrane potential statistics of the cell populations are fully defined, we simulate membrane potential responses by taking membrane potential samples from the population. A single trial was 500 ms long, the number of trials was 1000. The samples are then transformed to instantaneous firing rates by the rectifier nonlinearity (Eq.~\ref{eq:firing_rate}), and then spike counts are obtained by using the DSP and RG spiking models.
\subsection{Electrophysiological data}
To test the predictions of the models, we used publicly available data recorded in the labs of Matthias Bethge and Andreas Tolias \citep{Ecker:2010dn}. Detailed description of the recording settings are available at the original publication. Briefly, unit recordings were obtained by extracellular electrode arrays from the primary visual cortex of awake monkeys. Stimuli consisted of static and moving full-field gratings. We constrained our analysis to static gratings because the static grating data set featured multiple contrast levels besides eight grating orientations. We used 400 ms segments extracted from the evoked activity period of the trials in which the spike counts were calculated.
In order to reliably estimate pairwise correlations, we needed to exclude some of the recordings. We only considered pairs in which both units had an average firing rate over 0.1 Hz to avoid biased correlation estimates due to the insufficient number of spiking events. Pilot analyses (data not shown) have demonstrated that low number of stimulus repetitions can lead to highly inconsistent estimates of the spike count correlations, therefore we only included recording sessions that consisted of at least 39 repetitions.Thus, we included five sessions in the analysis, with repetition numbers (39, 40, 85, 72, 39). The filtering criteria for firing rates and trial numbers allowed us to use 41 units from the recordings.
When comparing spike count correlations between subpopulations observing a stimulus with a preferred or a non-preferred orientation, we defined orientations as preferred when the firing rate of a unit averaged over the trials using the orientation was higher than the average firing rate over all trials. This binary classification scheme helped us to avoid errors in orientation preference estimation, and in the same time lead to more reliably estimated of correlations in the non-preferred condition, as including only an orientation perpendicular to the most preferred one would have produced very low firing rates. Preferred-orientation correlations were calculated between pairs both observing a preferred stimulus.
\subsection{Analysis of neural responses}
We characterize the distributions of spiking responses of neurons up to second-order statistics, similarly to the descriptions of membrane potentials. However, due to specific properties of spike trains, the applied measures are slightly different. Spike count responses are characterized by variances that grow linearly with spike count means. Therefore, we are interested in changes in the variance that are independent of changes in the mean. By using Fano factor,we can control for this effect and can obtain a trial-by-trial measure of response variability for neuron $n$:
\begin{equation}\label{eq:ff}
\mathrm{FF} \left[ s^n_t \right]_t = \frac{\mathrm{Var} \left[ s^n_t \right]_t}{\mathrm{E} \left[ s^n_t \right]_t}
\end{equation}
The choice is also supported by the fact that systematic changes in the membrane potential variance are similarly observed in the spike count Fano factor, as described by \citep{Churchland:2010he}.
The pairwise co-activation of a pair of neurons given their spike trains is fully characterised by the cross-correlogram of their spike counts. While correlations may occur at different time scales, it is a typical choice in experiments to use the correlation of spike counts over entire trials. Doing so has the advantage of taking interactions with different delays into account similarly, by sacrificing the finer temporal structure of co-activations \citep{Smith:2008gv}. In order to account for irregularities in the firing rates and individual variances of experimentally recorded spike trains, correlations are calculated between z-scored spike counts, defined as follows:
\begin{equation}\label{eq:sc_corr}
\varrho^{ij} = \mathrm{Corr} \left[\frac{s^i_t - \mathrm{E} \left[s^i_t \right]_t }{\mathrm{SD} \left[s^i_t \right]_t}, \frac{s^j_t - \mathrm{E} \left[s^j_t \right]_t }{\mathrm{SD} \left[s^j_t \right]_t} \right]_t
\end{equation}
Aiming for population-level characterisation of responses puts constraints on what kind of measurements are applicable for comparison. As anesthesia is known to introduce significant biases in neuronal response correlations \citep{Ecker:2014cl}, we sought to test model predictions against data recorded from awake animals.
\section{Results}
Extensive data on the linear relationship between spike count mean and spike count variance \citep{Softky:1993uj, Britten:1993wv, Tolhurst:1983wa} motivated a model of spiking activity that assumes a Poisson process at spike generation. In this model, the Poisson-like single-cell spiking statistics can be attributed to private noise, a form of noise that is not shared by neurons. As a consequence, it does not necessarily contain any other forms of variability. Such a setting constrains the model to a Fano factor that is equal to one (when signal variance related to changing stimuli is ignored) and no variability that is correlated across neurons. V1 simple cells, however, can express correlated activity \citep{Ecker:2010dn} and are characterized by Fano factors that both deviate from one and can change with changing stimulus attributes. In order to accommodate these effects, we extended the simple Poisson model: the Doubly Stochastic Poisson (DSP) model assumes that the membrane potential has a stochastic component (Fig.~\ref{fig:cartoon}E). The membrane potential was sampled from a multivariate normal distribution, and consecutive samples were assumed to be independent across 20 ms time bins. Independence of samples is a simplifying assumption that is motivated by the fast-decaying autocorrelation function of V1 neurons \citep{Azouz:1999wj}. The membrane potential of a model simple cell was transformed by a nonlinearity to obtain a firing rate \citep{Carandini:2000ud} and spiking activity was obtained by the Poisson process, which ensured an expected value for the number of spikes proportional to the rate. The threshold-power-law firing rate nonlinearity ensured the correct mapping between mean membrane potential and firing rate \citep{Carandini:2004ee}.
In the alternative Rectified Gaussian (RG) model, there was a single source of stochasticity which was a stochastic process at the level of membrane potential. Again, the distribution at any given time bin was a multivariate normal distribution and time bins were independent (Fig.~\ref{fig:cartoon}A). Membrane potentials were mapped through the same nonlinearity as in the DSP model which ensured a similar evolution of firing rate with increased mean membrane potential (Fig.~\ref{fig:cartoon}B). Spikes were obtained by a process that bears as little stochasticity as possible: firing rate was integrated over time and spikes were generated when the integral crossed integer values (Fig.~\ref{fig:cartoon}D). While this model does not rely on a Poisson process to ensure the scaling of spike count variance with the mean, it has been demonstrated to account for the linear relationship between spike count mean and spike count variance solely as a result of the interaction of the subthreshold variability and the firing rate nonlinearity \citep{Carandini:2004ee}.
The expressive power of the two models is similar, which is also confirmed by the equal number of parameters characterizing the two models. These parameters were established based on previously published data (see Materials and Methods) \citep{Carandini:2004ee}. There are important differences, however, in the statistics of spiking activities the two models predict (Fig.~\ref{fig:mpmatch}A,B). Assuming identical distributions for membrane potentials for a pair of model neurons, the firing rate nonlinearity and consistent spike generation processes ensure similar mean spike counts (Fig.~\ref{fig:mpmatch}C). The variances, however, differ for the two models: while the variance of responses of RG model neurons is dominantly determined by the appropriately scaled variance of the membrane potentials, the variance of the DSP model is the sum of the membrane potential variance and a term coming from the Poisson stochasticity. As a result, the spike count variance, as well as the Fano factor, of the DSP model exceeds that of the RG model (Fig.~\ref{fig:mpmatch}C). The correlation measured from the spike count distribution can differ from the membrane potential correlations as a result of multiple factors \citep{Cohen:2011eh, Ecker:2010dn}. Most importantly, the firing rate nonlinearity can truncate the subthreshold part of the membrane potential distribution causing a decrease in spike count correlations relative to the membrane potential correlations, which is evident for the RG model (Fig.~\ref{fig:mpmatch}B). Another important consequence of the excess private variability introduced by the Poisson spike generation process is a further drop in the spike count correlation (Fig.~\ref{fig:mpmatch}C). A simple intuition for this effect can be obtained by considering that the covariance matrix of the spike count correlation is the sum of the membrane potential covariance and the covariance of the spike generation. The latter, however, is a diagonal matrix, since this source of noise is independent among neurons, therefore it only increases the diagonal elements of the resulting covariance matrix, and since correlation is the ratio of the covariance and the geometric mean of the variances, there is an overall decrease in correlations.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig2_mpmatch_stats.pdf}
\end{center}
\caption{\textbf{Spike response statistics for the Doubly-stochastic Poisson (DSP) and Rectified Gaussian (RG) models.} \textbf{\textit{A-B}}, Spike count distribution for a pair of model neurons with the same membrane potential statistics (mean, variance, correlation) for DSP (\textbf{\textit{A}}) and RG models (\textbf{\textit{B}}) in a simulated 100-ms time window. Membrane potential means and variances were identical for the two neurons and membrane potential correlation was set to 0.25. Circles indicate spike counts for individual trials. Cross shows the across-trial mean while ellipse represents the across-trial covariance ellipse of the joint spike count distribution. Small jitter was added to spike counts for illustration purposes. \textbf{\textit{C}}, While the spiking models are consistent in predicting equal mean spike count responses for both DSP (dark bars) and RG models (light bars) at matching membrane potential distributions, Fano factors and spike count correlations show characteristic differences. Note that spike count correlations are systematically lower than membrane potential correlations at both models. }
\label{fig:mpmatch}
\end{figure}
In the coming sections we analyze the consequences of these differences on simplified model of a pair of neurons before moving to the analysis of the responses statistics of populations of neurons. When using the simplified model we do not simulate the tuning curve-mediated changes in membrane potentials, rather we directly investigate the effects of changes in membrane potential statistics. These analyses provide predictions on changes in spike count statistics expected in response to changes in stimulus orientation and contrast.
\subsection{Matching spike response statistics}
In order to be able to contrast the effects of stimulus change on response statistics of the competing models, we first establish a method for matching the spiking statistics of the DSP and RG models in a pair of neurons. Since equal membrane potential statistics lead to different spiking statistics, it is clear that either membrane potential statistics or parameters of the firing rate nonlinearity need to be adjusted to have matching firing rates, Fano factors and spike count correlations. In order to keep our arguments simple, we keep the firing rate nonlinearity unchanged. In fact, the scale of membrane potential (which is determined together by the distance of the membrane potential from the firing rate threshold and the variance of the membrane potential) and the scale of firing rate nonlinearity (parameter $k$) can be altered largely interchangeably (see Fig.~\ref{fig:corrdep}), therefore the argument can be translated into changes in the scale parameter of the firing rate.
By exploring membrane potential parameters for the two models, we can obtain a parameter setting where firing rates (Fig.~\ref{fig:statmatch}A) and Fano factors (Fig.~\ref{fig:statmatch}B) are matched. Both of these criteria constrain the parameter sets up to a linear combination of the tested parameters, therefore the intersection of the lines allows matched firing rate and Fano factor. Difference between the spike count correlations along the explored parameter range is relatively untouched (Fig.~\ref{fig:statmatch}C), and can be adjusted by tuning membrane potential correlations. The resulting statistics-matched models (Fig.~\ref{fig:statmatch}D-F) have markedly different membrane potential variances (0.2 $\mathrm{mV^2}$ and 2.75 $\mathrm{mV^2}$ for the DSP and RG models, respectively) and different membrane potential correlations (0.95 and 0.13 for the DSP and RG models, respectively). Because of the extra variance and the decorrelation effect of the DSP model, these differences are expected and show that the Poisson process introduces a private variability which can easily wash out membrane potential correlations. While the membrane potential correlation level required in the DSP model seems to be extreme, it only serves the purpose of matching the spiking statistics. At lower firing rates the excess variance added to the membrane potential covariance would be lower and therefore statistics matching would require considerably lower membrane potential correlations.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig3_statmatch.pdf}
\end{center}
\caption{\textbf{Matching the output statistics of DSP and RG models.} \textbf{\textit{A-C}}, Spike count (\textbf{\textit{A}}), Fano factor (\textbf{\textit{B}}) and spike count correlation (\textbf{\textit{C}}) of the neurons as a function of membrane potential parameters for the DSP (dark plane) and RG model (light plane). Simulations show responses in a 100 ms time window. Intersections of the plane indicate equal statistical measure for the two models which then help to determine the spike count statistics. Parameters explored for matching spiking statistics were the mean membrane potential of the DSP model and membrane potential variance of the RG model. \textbf{\textit{D-F}}, By tuning membrane potential statistics for the DSP (\textbf{\textit{D}}) and RG (\textbf{\textit{E}}) models individually, the spike count statistics of the models can be matched up to the mean, variability (as measured by the Fano factor) and correlation. Spike count distribution representation is the same as that on Fig.~\ref{fig:mpmatch}. Matched means. \textbf{\textit{F}}, spiking statistics for the matched models. Colours are identical to panels \textbf{\textit{A-C}}.}
\label{fig:statmatch}
\end{figure}
\subsection{Dependence of spiking statistics on the membrane potential mean}
The membrane potential mean in V1 simple cells is sensitive to stimulus orientation. In order to understand the effects incurred by orientation change on spiking statistics, we need to separate the effects on different aspects of the response statistics. Changes in membrane potential mean have obvious effects on the firing rate. Effects of mean membrane potential on other characteristics of the response statistics are less straightforward. In order to be able to see how the membrane potential mean changes affect Fano factors and spike count correlations in DSP and RG models, we aim to see changes in these measures independent of changes in firing rates.
Membrane potential statistics are different in the statistics-matched DSP and RG models. Therefore, we first establish rates of change for the membrane potentials in the two models, which guarantee that even upon deviating from the statistics-matched levels of membrane potentials, the firing rates change at a similar rate (Fig.~\ref{fig:orientchange}A,D). Using this firing rate-matched scenario, we can directly contrast mean-related changes in Fano factors (Fig.~\ref{fig:orientchange}B,E) and spike count correlations (Fig.~\ref{fig:orientchange}C,F). While Fano factors are approximately equal (Fig.~\ref{fig:orientchange}B) across the range of membrane potentials for the DSP and RG models, spike count correlations deviate for the two models (Fig.~\ref{fig:orientchange}C). Independence of Fano factors from membrane potentials (Fig.~\ref{fig:orientchange}E), and from firing rates as well, confirms the original results of the RG model \citep{Carandini:2004ee} and is expected for the DSP model. Membrane potential dependence of spike count correlations (Fig.~\ref{fig:orientchange}F) reveals a decreasing tendency for the DSP and an increasing tendency of the RG with increasing membrane potential levels. Growing firing rate resulting from increased mean membrane potential has a differential effect in the two models. In the DSP, increased firing rate incurs increased private variability which suppresses the contribution of the membrane potential covariance to the total covariance and therefore spike count correlation diminishes. In the RG model, however, a different mechanism dominates: increased mean activation results in a higher proportion of the membrane potential covariance to be above firing threshold, and consequently, smaller truncation of this distribution boosts the magnitude of the measured correlation \citep{deLaRocha:2007go}. Taken together, under controlled change in mean activity, spiking variability is relatively insensitive to the choice of DSP or RG models. Analysis of spike count correlations, however, provide opposing predictions in the case of DSP and RG models for manipulations that affect mean responses.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig4_matched_orientchange.pdf}
\end{center}
\caption{\textbf{Dependence of spike count statistics on the membrane potential mean.} \textbf{\textit{A-C}}, Predictions of the two spiking models directly contrasted against each other for firing rate (\textbf{\textit{A}}), Fano factor (\textbf{\textit{B}}), and spike count correlation (\textbf{\textit{C}}) upon changes in membrane potential mean. Matched-statistics DSP and RG are tested with changing membrane potential means but constant membrane potential variances. \textbf{\textit{A}}, With appropriate linear scaling of the membrane potentials, approximately equal firing rates can be achieved across a range of membrane potentials resulting in a wide range of firing rates (0 to 60 Hz). \textbf{\textit{B}}, Fano factors are close to equal and invariant across the whole range of membrane potentials. \textbf{\textit{C}}, While firing rates and Fano factors cannot explicitly differentiate between the DSP and RG models, the correlation shows systematic differences. \textbf{\textit{D-F}}, Same as \textbf{\textit{A-C}} but respective statistics for the DSP (dark lines) and RG (light lines) models as a function of the membrane potentials. Note that in order to match the spiking statistics, the membrane potential statistics (including the mean, variance, and correlation of neurons) differ in the two models (see different horizontal axes at the bottom and the top of panels). \textbf{\textit{D}}, Firing rates from the two different model neurons show characteristic nonlinearity of firing rates in the two models. \textbf{\textit{E}}, Fano factors in both models show only minimal dependence on membrane potential (and thus on firing rate). \textbf{\textit{F}}, Spike count correlations show sensitivity to changes in the membrane potential: while pairwise correlation in the DSP model decreases with increasing membrane potential levels, it increases in the RG model. Except for the changing mean membrane potential, parameters of the models are the same as those used on Fig.~\ref{fig:statmatch} (matched mean is denoted by black circle).}
\label{fig:orientchange}
\end{figure}
Differential effects of the membrane potential mean on spike count correlations in the two models highlight an opportunity to distinguish between the models upon changes in stimulus orientation. A detailed exploration of the evolution of spike count correlation with changing membrane potential mean can reveal the generality of the effect seen on Fig.~\ref{fig:orientchange}F. We tested this question by assessing spike count correlations at different levels of membrane potential correlations (Fig.~\ref{fig:corrdep}). Analysis of the RG model reveals a monotonic rise of the magnitude of spike count correlations from zero towards the level of the membrane potential correlation as membrane potential mean increases (Fig.~\ref{fig:corrdep}B). In the DSP model, the effect of a less truncated membrane potential joint distribution, when a larger proportion of the distribution gets above threshold, is shown at a low membrane potential regime (Fig.~\ref{fig:corrdep}A): at low levels of activations, a rise similar to the RG model can be observed. This range, however, is severely limited (Fig.~\ref{fig:corrdep}C) and can be observed at moderate firing rates. Beyond that point, a steady decline of spike count correlation takes place which converges to zero (Fig.~\ref{fig:corrdep}A,C). The biphasic profile of membrane potential dependence of spike count correlation raises the possibility that an increased gain in the firing rate nonlinearity can simply scale the firing rate profile. Thus, the regime where the spike count correlation is positively correlated with membrane potential mean could possibly overcome the limited range shown on Fig.~\ref{fig:corrdep}A and could reach higher firing rates. We tested this question by scaling the firing rate gain together with inverse scaling of the membrane potential (Fig.~\ref{fig:corrdep}D). In order to keep not only the firing rate but also the Fano factor constant, the variance of the membrane potential was also scaled together with the mean and rate gain parameters (Fig.~\ref{fig:corrdep}E). In the resulting setting we could test a wide range of the gain parameter $k$, while keeping the mean, the Fano factor, and the correlation of spiking responses constant (Fig. F-H). The firing rate profile was identical for the different parameter settings (Fig.~\ref{fig:corrdep}I). Importantly, the evolution of correlations was very close at different settings of the gain parameter $k$, with no visible shift in the membrane potential (or alternatively firing rate) value maximizing the correlation (Fig.~\ref{fig:corrdep}J). This analysis demonstrates that the regime where increasing correlations are present with increasing firing rates are constrained to low firing rate levels and high Fano factors.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig5_corrdep.pdf}
\end{center}
\caption{\textbf{Relationship between spike count correlations and membrane potential mean in the two spiking models.} Evolution of spike count correlations with changing membrane potentials at different membrane correlation levels in the matched statistics DSP (\textbf{\textit{A}}) and RG (\textbf{\textit{B}}) models. \textbf{\textit{A}}, In the DSP model, after a brief increase of the correlation magnitude towards the level of membrane potential correlation, the tendency reverses and a decline towards zero spike count correlation takes place. Range of membrane potential correlations tested spans from -0.95 to 0.95. \textbf{\textit{B}}, The RG model is characterized by a steadily increasing magnitude of spike count correlation, saturating at the correlation of the membrane potentials. Range of membrane potential correlations tested spans from -0.4 to 0.4. \textbf{\textit{C}}, Exploration of spike count statistics in the DSP model to assess the regimes in which correlation magnitude increases with the firing rate. Positive slope regimes (red hues) are constrained to low firing rates and high Fano factors, with the white separating blue and red regimes denoting the peak of the curve in \textbf{\textit{A}}. \textbf{\textit{D-J}}, Analysis of the dependence of spike count statistics on the base rate. Membrane potential statistics (\textbf{\textit{D-E}}) are tuned for a range of values of the base rate parameter $k$ in order to produce similar spiking statistics (\textbf{\textit{F-H}}). \textbf{\textit{I}}, Membrane potential dependence of firing rates is identical for the different gain, $k$ (lines are overlapping). Membrane potentials are normalized to the values established on \textbf{\textit{D}} (black circle), diamond denotes the maximum on \textbf{\textit{J}}. \textbf{\textit{J}}, Membrane potential dependence of correlations is similar across different levels of the gain, $k$ (grey lines). }
\label{fig:corrdep}
\end{figure}
\subsection{Dependence of spiking statistics on membrane potential variance}
Stimulus contrast was demonstrated to have a combined effect on membrane potential mean and variance \citep{Finn:2007hc}: while the mean of the membrane potential response shrinks as contrast goes to zero, membrane potential variance grows. Thus, changes in membrane potential response variance have relevant consequences on the spiking statistics. In order to get insights into the effects of joint changes in membrane potential mean and membrane potential variance, we first explored these characteristics separately and then turned to the combined effects of parallel changes.
After the mean-dependent changes discussed in the previous section, we set out to analyze the membrane potential variance-dependence of spiking statistics (Fig.~\ref{fig:variancechange}). Similarly to the protocol followed at testing the effects of membrane potential mean, we started from the matched-statistics DSP and RG models, and set the range of variance scaling such that the resulting firing rate changes in the two models are approximately equal (Fig.~\ref{fig:variancechange}A,D). Again, this mean firing rate-matched approach ensures that changes seen in the Fano factors and spike count correlations are not related to differences in firing rates. Contrasting the Fano factors at firing rate-matched settings of the DSP and RG models revealed similar tendencies but slightly differing values for the Fano factors (Fig.~\ref{fig:variancechange}B). Increased membrane potential variance translated into increased Fano factors in both of the spiking models, but the DSP model was characterized by systematically larger Fano factors at higher membrane potential variances (Fig.~\ref{fig:variancechange}E). This difference is due to the excess variance of the DSP model coming from the increased spike count variance of the Poisson stochasticity at higher mean spike counts. Changing membrane potential variance showed conflicting effects in the two models on spike count correlations (Fig.~\ref{fig:variancechange}C). While the spike count correlation in the RG model was relatively insensitive to changes in membrane potential variance and thus firing rate, the DSP model was shown to exhibit increased spike count correlation with increased membrane potential variance (Fig.~\ref{fig:variancechange}F). This increase can be easily understood by recognizing that a scaled membrane potential covariance results in a larger relative contribution of the membrane potential covariance to the total covariance, thus the spike count correlation will be more dependent on the correlated membrane potential stochasticity than uncorrelated spiking stochasticity.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig6_matched_variancechange.pdf}
\end{center}
\caption{\textbf{Dependence of spike count statistics on the membrane potential variance.} Spiking statistics of the two spiking models directly contrasted against each other for the two models upon changes in membrane potential variance. Matched-statistics DSP and RG are tested with changing membrane potential variances but constant membrane potential means. Firing rates (\textbf{\textit{A}}) and Fano factors (\textbf{\textit{B}}) are approximately equal in both models, but spike count correlations (\textbf{\textit{C}}) show systematic differences. \textbf{\textit{D-F}}, Plotting the statistical measures against membrane potential variance reveals that both firing rates (\textbf{\textit{D}}) and Fano factors (\textbf{\textit{E}}) show a tendency to increase with increased membrane potential variance. Dark lines: DSP model; light lines: GR model. Note that in order to match the spiking statistics, the membrane potential statistics (including the mean, variance, and correlation of neurons) differ in the two models (see different horizontal axes at the bottom and the top of panels). While spike count correlations are invariant across the range of variances for the RG model, the poisson model shows a steady increase with increasing membrane potential variance (\textbf{\textit{F}}). Except for the changing mean membrane potential, parameters of the models are the same as those used on Fig.~\ref{fig:statmatch} (matched variance is denoted by black circle).}
\label{fig:variancechange}
\end{figure}
Contrast change incurs concomitant changes in membrane potential mean and variance. Based on the previous analyses we can conclude how these parallel changes interact when the spiking statistics are considered in response to a stimulus at lower contrast. In terms of firing rate, decreased mean membrane potential and increased variance due to reduced contrast have opposing effects, but the effect of decreased membrane potential dominates patterns in firing rate (Fig.~\ref{fig:lchc}A). In terms of Fano factor, it is the change in variance that causes increased Fano factors in both models (Fig.~\ref{fig:lchc}B). In terms of spike count correlations, the two models have distinct predictions (Fig.~\ref{fig:lchc}C). It is only the change in the mean membrane potential that contributes to a shrinking magnitude of correlations in the RG model (Fig.~\ref{fig:orientchange}F, 5B). In the DSP model, one component contributing to contrast-related changes is the increased variance, which causes larger spike count correlations (Fig.~\ref{fig:variancechange}F). The effect of decreased mean seems to be more complex: it results in increased correlations in a wide range of parameters and decreased correlations in a specific subspace of the parameters (Fig.~\ref{fig:corrdep}C). Remarkably, this subspace is characterized by low firing rates and relatively high Fano factors. In summary, contrast modulation related changes in membrane potential variance introduce changes in Fano factors and spike count correlations in both models. The magnitude of the variance change determines the difference in Fano factors between high contrast and low contrast conditions but the direction of deviation is the same for both models. Contrast has opposing effects on spike count correlations in the two models. In the DSP model both increased membrane potential variance and decreased membrane potential mean incur higher spike count correlations at lower contrast levels. In the RG model, however, excess variance does not affect spike count correlations and therefore changes in spike count correlations are solely determined by changes in mean membrane potential which ultimately results in decreasing correlations with decreasing contrast.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig7_demo_lchc.pdf}
\end{center}
\caption{\textbf{Contrast dependence of spiking statistics.} Spike counts, Fano factors and spike count correlations in the two models (dark and light lines for the DSP and RG models, respectively) with concomitant changes in membrane potential mean and variance stimulating a switch from high contrast (HC) to low contrast (LC): decreased mean was followed by increased variance. Parameters for high and low contrast levels were set such that firing rate changes (\textbf{\textit{A}}) and Fano factor changes (\textbf{\textit{B}}) approximated the changes expected in V1 recordings. Under such conditions the spike count correlation is expected to decrease for the DSP model but increase for the RG model when contrast is increased (\textbf{\textit{C}}). }
\label{fig:lchc}
\end{figure}
\subsection{Contrast- and orientation-dependent modulation of spiking correlations in a population of simple cells}
In order to test the predictions of the two models against experimental data, we simulated the activity of a population of V1 simple cells in response to changes in stimulus. A critical motivation for using population-level analyses instead of the analysis of pairs of neurons is that pairwise analysis of response statistics requires much more precise control of the recording conditions to curb confounding factors.
As demonstrated by our analysis of the two-neuron model, stimulus change-related changes in private variability do not distinguish between the DSP and RG models. Therefore it is expected that population distributions of Fano factor are indistinguishable too. Spike count Fano factors show little dependence on stimulus orientation and are increased when stimulus contrast is lowered (data not shown), as predicted by both spiking models. As a consequence, we focus on the analysis of stimulus-dependence of correlated variability in spiking responses of a population of V1 neurons.
First, we tested how decreasing stimulus contrast affects the population distribution of spike count correlations in the two models, and compared the results to experimental data. We set up two populations of 100 model neurons each, implementing different spiking profiles corresponding to the DSP and the RG models. We simulated membrane potential using tuning curve responses to a full-field grating stimulus (see Materials and Methods). Single-cell spiking statistics in the model populations were matched (mean firing rates were 6.5 Hz and 6.3 Hz, mean Fano factors were and 1.3 and 1.2 in the high contrast condition in the DSP and RG populations, respectively). The membrane potential correlations in the two populations were chosen such that mean and width of the two spike count correlation distributions are matched at high contrast stimulus presentation (Fig 8A-B, insets). The distribution of spike count correlation was tuned to have width and mean comparable to physiological values observed in the data recorded (-0.02 and 0.16 for the mean and standard deviation in high contrast conditions, respectively) \citep{Ecker:2010dn}. Width of the distribution was wider for membrane potentials than spike counts both in the case of the DSP and RG models and, as expected from previous analyses, the width of the membrane potential correlation was wider for the DSP model (0.56 and 0.23 for the DSP and RG models, respectively). Lowering the stimulus contrast made the spike count correlation distribution wider in the DSP population (Fig.~\ref{fig:contrast}A). This is caused by the same phenomenon as the increase in pairwise correlation from higher to lower contrast (Fig.~\ref{fig:lchc}C), namely that lower firing rates and increased membrane potential variance tilts the balance between correlated membrane potential and independent spiking variability towards the former. Conversely, in the RG population, the correlation distribution gets narrower with decreasing stimulus contrast (Fig.~\ref{fig:contrast}B). This effect is again in agreement with the results on pairwise correlation (Fig.~\ref{fig:lchc}C).
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig8_pop_contrast_data.pdf}
\end{center}
\caption{\textbf{Comparison of model predictions to experimental data - effect of stimulus contrast.} \textbf{\textit{A}}, Distributions of pairwise spike count correlations in a simulated population of 100 DSP neurons in response to a low contrast (LC) and high contrast (HC) stimulus, using membrane potentials (MP) from tuning curves proportional to the contrast. \textbf{\textit{B}}, Distributions of pairwise spike count correlations in a simulated population of 100 RG neurons in response to the same stimuli. \textbf{\textit{C}}, Distributions of measured spike count correlations in response to grating stimuli of low and high contrasts, from \citep{Ecker:2010dn}. \textbf{\textit{D}}, Differences in the means of the two correlation distributions in the two simulated and experimentally recorded populations. Stars represent significant difference of the mean from zero at the p=0.05 level. \textbf{\textit{E}}, Differences in the standard deviations (SD) of the two correlation distributions obtained under the high and low contrast condition. Mean of the differences are shown for the two models across different populations and for experimental recordings across different sets of left-out neurons. The SD mean is significantly different from zero in all populations, with the RG agreeing with data regarding the change of direction and the DSP not. \textbf{\textit{F}}, Dependence of the SD difference in the DSP model on the width of MP correlation distribution. \textbf{\textit{G}}, SD difference in the DSP model at different values of MP mean and variance, shown as a function of the spike count Fano factor, testing whether low mean and high variance in MP can reverse contrast effects in spike count correlations.}
\label{fig:contrast}
\end{figure}
Analyzing population responses from V1 reveals contrast-dependent changes in the distribution of spike count correlations. The distribution of spike count correlations is narrower in response to low contrast stimuli than in response to high contrast (standard deviations of 0.16 for high and 0.12 for low contrast, respectively, Fig.~\ref{fig:contrast}C). The uncertainty of these estimates was assessed by bootstrapping (discarding 20\% of the pairs 5 times), providing an estimate for the standard error of mean for the change in the standard deviation of the spike count correlation distribution. The same analysis was performed for both of the models to obtain a comparable estimate of uncertainty using 5 simulations using 41 randomly selected units, similarly to the number of units available from the experiment. Shrinking of the width of the distribution was significant (Fig.~\ref{fig:contrast}E, one-sample t-test t(4) = 7.06, p = 0.002), similar to the shrinking of the width of the distribution of correlations in the RG model (standard deviation of 0.11 and 0.08 in the high and low contrast conditions respectively, one-sample t-test t(4) = 5.51, p = 0.005, Fig 8E). This result is in contrast with the increased width of spike count correlation distribution at the DSP model (standard deviations of 0.106 and 0.111 in the high and low contrast conditions respectively, one-sample t-test t(4) = -4.06, p = 0.015, Fig.~\ref{fig:contrast}E). Changes in the mean of correlations are relatively small and are consistent across both models and experimental data (in the experimental population, -0.015 and 0.001 in the high and low contrast condition respectively, one-sample t-test t(4) = -0.92, p = 0.41; in the DSP population, 0.009 and 0.015, one-sample t-test t(4) = -7.7, p = 0.002; in the RG population, 0.013 and 0.017, one-sample t-test t(4) = -4.5, p = 0.01; Fig.~\ref{fig:contrast}D).
In order to test the robustness of the dissociation of the two models by contrast related changes in spike count correlations, we explored how much the changes in correlation distributions depend on the specific settings we used in the model. In the case of the RG model, intuitions obtained from the two-neuron analysis confirm that the increased width is robust against changes in parameters. In the case of the DSP model, however, the specific settings might affect the direction of the change (see Fig.~\ref{fig:corrdep}B,C). We varied the width of the membrane potential correlation distribution for the DSP model to assess whether the contrast dependency of spike count correlations could be reversed at specific settings. Simulations confirmed that this manipulation is not capable of reproducing the experimentally observed pattern (the changes being significantly or non-significantly negative at all tested values, Fig.~\ref{fig:contrast}F). The analysis described in Fig.~\ref{fig:corrdep} motivated us to investigate whether the difference between spike count correlation widths in the high and low contrast conditions could also be positive in the DSP model population when intensively decreasing means and increasing variances concomitantly in the membrane potential. Such changes did not produce a narrowing spike count distribution in response to lower contrast stimuli within (and neither way above) the Fano factor range plausibly observed in measurements (the changes being significantly or non-significantly negative at all tested values, Fig.~\ref{fig:contrast}G). These analyses highlight that the population response statistics to contrast-varied stimuli reliably discriminate between the DSP and RG models of neural spiking and it is the RG model that provides predictions compatible with contrast change related data.
Next, we analyzed population responses to changes in stimulus orientation. Population models are constructed similarly to the analysis of contrast dependence. Preferred orientations are uniformly distributed, therefore a simple orientation change is not expected to cause changes in population response statistics. In order to assess orientation related changes in population responses, we classified any particular orientation as preferred or non-preferred orientation based on whether the response of the neuron was above or below its average response intensity. In order to construct distribution of correlations for preferred and non-preferred directions, pairs of cells were selected based on whether both of the cells had the actual stimulus among their preferred orientation or both had it among their non-preferred orientations. Simulation results were contrasted with recordings of populations of V1 units \citep{Ecker:2010dn} in which we also separated pairs of units observing preferred and non-preferred stimuli (see Materials and Methods). In the experimental dataset, we observed a narrower distribution of spike count correlations in response to non-preferred stimuli compared to the response to preferred ones (standard deviations of 0.15 for preferred and 0.10, for non-preferred stimuli, one-paired t-test t(4)=4.83, p=0.009, Fig.~\ref{fig:orient}C,E). In the DSP population, the width of the spike count correlation distribution did not change significantly between the two conditions (standard deviations of 0.11 and 0.1, one-sample t-test t(4)=1.65, p=0.17, Fig.~\ref{fig:orient}A,E). In the RG population, the non-preferred stimulus elicited a narrower distribution of correlations than the preferred one (standard deviations of 0.12 and 0.09, one-sample t-test t(4)=4.77, p=0.009, Fig.~\ref{fig:orient}B,E). Similar to the simulations and experimental data of contrast dependence of responses, changes in the means of correlation distributions were small (in the experimental population, -0.001 and 0.001 in the preferred and non-preferred condition respectively, one-sample t-test t(4) = -0.16, p = 0.88; in the DSP population, 0.006 and 0.011, one-sample t-test t(4) = -2, p = 0.12; in the RG population, 0.016 and 0.01, one-sample t-test t(4) = -1.7, p = 0.17; Fig.~\ref{fig:orient}D). Analysis of the membrane potential parameter space revealed that most correlation, mean and variance values yield non-significant changes between the spike count correlation widths in the preferred and non-preferred conditions, with some small, significantly positive values. (Fig.~\ref{fig:orient}F-G). While experimental results do not provide a direct dissociation between the two models in terms of predictions related to the dependence of correlations on orientation preference, they are reproduced without any particular tuning of model parameters in the case of the RG model, but the DSP model can only account for the patterns in experimental data with specific parameter tuning.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig9_pop_orient_data.pdf}
\end{center}
\caption{\textbf{Comparison of model predictions to experimental data - effect of stimulus orientation.} \textbf{\textit{A}}, Distributions of pairwise spike count correlations in a simulated population of 100 DS neurons from pairs observing stimuli of preferred orientation (PO) and non-preferred orientation (NO), using membrane potentials (MP) from tuning curves distributed uniformly over orientations. \textbf{\textit{B}}, Distributions of pairwise spike count correlations in a simulated population of 100 RG neurons in response to the same stimuli. \textbf{\textit{C}}, Distributions of measured spike count correlations in response to grating stimuli of different orientations, from \citep{Ecker:2010dn}. \textbf{\textit{D}}, Differences in the means of the two correlation distributions in the two simulated and the measured populations. Stars represent significant difference of the mean from zero at the 0.05 level. \textbf{\textit{E}}, Differences in the standard deviations (SD) of the two correlation distributions obtained from pairs of neurons observing stimuli with preferred or non-preferred orientations. Mean of the differences are shown for the two models across different populations and for experimental recordings across different sets of left-out neurons. The SD mean is significantly different from zero in the experimental and the RG populations, but not in the DSP population. \textbf{\textit{F}}, Dependence of the SD difference in the DSP model on the width of MP correlation distribution. \textbf{\textit{G}}, SD difference in the DSP model at different values of MP mean and variance, shown as a function of the spike count Fano factor, testing whether low mean and high variance in MP can reproduce experimentally measured orientation effects in spike count correlations.}
\label{fig:orient}
\end{figure}
\section{Discussion}
We analyzed widely used models of spiking to investigate their power to predict stimulus-dependent changes in not only in single-cell statistics but in joint statistics of activity too. The Doubly Stochastic Poisson model \citep{Gur:1997ur} assumes stochastic spiking to account for linear scaling of spike count variance with spike count mean and relies on a separate stochastic process to model the covariance structure of spiking responses. The Rectified Gaussian model \citep{Carandini:2004ee} assumes a single source of stochasticity at the level of membrane potentials but relies on a quasi-deterministic process of spike generation. We demonstrated that while in terms of single-cell statistical measures the models make similar predictions for stimulus change-related modulations in response statistics, predictions on pairwise correlations are distinct. Using a model of a pair of neurons, simulated changes in stimulus orientation and stimulus contrast revealed opposing changes in spike count correlations. The key intuition behind this finding is that the level of spike count correlation with a given level of membrane potential correlation is determined by two phenomena: 1, higher mean membrane potential implies higher magnitude spike count correlations \citep{deLaRocha:2007go} and 2, higher firing rate implies lower level of spike count correlations when Poisson spiking stochasticity is present. The interaction of these two processes result in opposing changes in the correlation structure of the population. In order to assess which of the two models is compatible with neural recordings, response statistics of a population of neurons was simulated in both models and contrasted with response statistics of neurons recorded in V1 of awake monkeys. We have shown that the predictions of the Rectified Gaussian model are in line with electrophysiological data while the Doubly Stochastic Poisson model is not capable of reproducing the patterns of noise correlations in V1 neurons. Our analyses highlight the necessity to assess joint responses of neurons when constructing models of population activity and also highlight that assessment of the width of the distribution of the correlations beyond the mean of correlations can be an important factor. These analyses provide important constraints on the models that can be used effectively to characterize computations in neural populations.
In our analysis we adopted an approach where we focused on matching the spike count statistics of the models without particular emphasis on the interpretation of the actual levels of membrane potentials. This approach is motivated by the aim to directly contrast changes in spiking statistics between the two models. It is obvious that separate tuning of the membrane potential parameters could result in membrane potential values that are easier to interpret physiologically. Nonetheless, the assumptions of the DSP model necessitate choices that are hard to reconcile with neuronal data. For instance, in the analysis of the pair of neurons we used a membrane potential correlation of 0.95 in the DSP model in order to obtain a spike count correlation level matching that of the RG model (approximately 0.1) when firing rate was about 35 Hz. In the population model of spiking statistics less extreme values were used, still the tendency of the DSP model to wash out correlations necessitated a wider correlation distribution for the DSP than for the RG model.
Recent studies using approaches close to our DSP model \citep{Ecker:2016hc, Rabinowitz:2015fa} discuss top-down modulation of response correlations in contrast with the bottom-up modulations discussed here. The important contribution of these models is that different forms of correlations in spiking responses are explained based on simple computational principles. The differences between those approaches and ours are important and can be instructive for future work. First, correlations are introduced at the level of firing rates instead of membrane potentials. Since the variability is introduced prior to spike generation, the conclusions of those studies can be easily translated to our approach solely by assessing the effect of the firing rate nonlinearity. Second, these studies used a single scalar (or a low dimensional) stochastic variable to model gain modulation of visual cortical neurons. The collective gain modulation implies a firing rate correlation of one between those neurons that share the modulatory signal. In our case correlated activity was introduced by a multivariate stochastic process which implements a softer coupling between neurons. Nevertheless, the analysis of the DSP model reveals that realistic firing rates, Fano factors, and spike count correlations require a surprisingly high level of membrane potential correlation. Third, the use of Poisson stochasticity in these models implies that only variability beyond the Poisson variability are meant to be accounted for. Our analysis shows that the RG model can be a more effective model for variability than the DSP. Since the RG model only assumes a single source of variability at the level of membrane potentials, it provides an opportunity to account for a larger portion of variability and therefore provides an opportunity for models with better predictive power.
We chose to use independent temporal evolution for membrane potentials, which meant independent membrane potential samples in fixed, short time intervals. This was motivated by the relatively fast decay of membrane potential correlations \citep{Azouz:1999wj}. This treatment explicitly incorporates within-trial variability to model the shared variability of neurons. This is in contrast with other models of response variability \citep{Ecker:2016hc}, which consider processes that produce trial-to-trial variability since changes in attentional modulation occur on a slower time scale. We argued in the paper that spiking statistics provide constraints on the spiking models, therefore we believe that patterns in the auto-, and cross-correlation functions \citep{Smith:2008gv} provide further constraints on neural models of spiking.
Poisson-like firing has been used extensively in the literature \citep{Goris:2015hp, Ma:2006bh, Jazayeri:2006fk, Ma:2014in, Simoncelli:2004ue, Pillow:2007wh, Froudarakis:2014fs}. Besides its capability to provide a parsimonious explanation of the relationship between spiking intensity and spiking variability, Poisson neurons have much theoretical appeal too. First, Poisson-like spiking distribution ensures that fitting network parameters is a convex optimization problem, or in other words, there is a single (global) maximum in optimization \citep{Paninski:2004to}. Second, in theories of encoding information via populations of neurons, a Poisson-like likelihood function provides a representation in which the log likelihood contains linear terms which enables simple, neurally plausible computations \citep{Ma:2006bh, Jazayeri:2006fk}. In contrast with models based on the Poisson assumption, alternative approaches have applied models that have a closer relationship to the Rectified Gaussian model \citep{Paninski:2004jc, Brette:2005ke, Lin:2015dw} and argued for the capability of such models to predict both for single-cell response properties \citep{Brette:2005ke} and population statistics \citep{Lin:2015dw}. Response variability in these two model classes is approached in two markedly different ways: in one, stochasticity is part of the spike generation process, in the other it originates prior to spike generation and can be related to membrane potential-level processes. In this context, our study contributes to the field by a direct and controlled comparison of the predictions of these approaches on spiking statistics.
Recent advances in modelling data recorded from a large number of neurons have helped to assess neural responses in a trial-by-trial manner and to relate them to variances in behavior \citep{Churchland:2010he, Yu:2009ex}, disentangle mixed sensitivities \citep{Kobak:2016eo}, eliminate noise by tracing neural variability back to changes in latent factors \citep{Machens:2010in} and to implement closed-loop brain-computer interfacing \citep{Sadtler:2015kx}. Modelling and predicting correlations is a critical factor in population-level analyses of neuronal data \citep{Cunningham:2014ev}. Those are precisely the correlated changes in neuronal activity that help to eliminate the effect of uncontrolled variables, to reduce apparent noise in the measurements, and to predict the activity of missing neurons. Therefore, it is crucial to have an adequate model of response variability that can predict the effects of changes in stimulus on the correlation structure. The stochasticity assumed to underlie observed spiking variability can take on the form of Poisson variability \citep{Archer:2014wn, Macke:2011ut} or Gaussian noise \citep{Yu:2009ex, Sadtler:2015kx}. Our study aims to provide constraints on the forms of stochasticity in these models by emphasising that bottom-up driven changes in the response statistics can differentiate between alternative models. As demonstrated in the paper, even when single-cell response statistics have limited power to distinguish between alternative models, joint statistics can reveal properties that are incompatible with the predictions of one or the other. We expect that proper understanding and characterization of stochasticity in the nervous system helps to better interpret joint statistics and especially correlations present in the activity of neural populations.
\section*{Acknowledgements}
We thank M. Lengyel for useful discussions, D. G. Nagy for comments on the manuscript and especially A. Ecker, P. Berens, M. Bethge and A. Tolias for making their data publicly available. This work was supported by a Lend\"ulet Award of the Hungarian Academy of Sciences (G.O., M.B.) and an award from the National Brain Research Program of Hungary (NAP-B). The authors declare no competing financial interests.
\bibliographystyle{apalike}
|
\section{Introduction}
\label{introduction}
Let $E\subset \mathbb{C}$ be a discrete subset.
In \cite{fujino2}, to study the Teichm\"uller space of
the punctured plane $\mathbb{C}\setminus \mathbb{Z}$,
the author gave some criteria for
$\mathbb{C}\setminus E$ to be quasiconformally equivalent to
$\mathbb{C}\setminus \mathbb{Z}$ (that is, there exists a quasiconformal
mapping $F:\mathbb{C}\rightarrow \mathbb{C}$ such that $F(E)=\mathbb{Z}$).
In this paper, furthermore, we investigate the correspondences between
$E$ and $\mathbb{Z}$ which are the restrictions of global quasiconformal mappings
$F:\mathbb{C}\rightarrow \mathbb{C}$ such that $F(E)=\mathbb{Z}$.
A motivation of this attempt is to
study the Teichm\"uller modular group of $\mathbb{C}\setminus \mathbb{Z}$
and its action.
Let $\eta:[0,+\infty )\rightarrow [0,+\infty )$ be a homeomorphism and
$f:X\rightarrow \mathbb{R}^n$ be an $\eta-$quasisymmetric embedding from
a subset $X\subset \mathbb{R}^n$ into $\mathbb{R}^n$.
The theory of the quasisymmetry and its quasiconformal extension originated
from the well known study for $X=\mathbb{R}$ and $n=1$
by Beurling--Ahlfors \cite{beurling1}.
They proved that a homeomorphism $f:\mathbb{R}\rightarrow
\mathbb{R}$ admits a quasiconformal extension $F:\mathbb{C}\rightarrow
\mathbb{C}$ if and only if $f$ is quasisymmetric.
This result enables us to treat the universal Teichm\"uller space, the Teichm\"uller
space of the unit disk, as the space of all orientation preserving
quasisymmetric homeomorphisms of the unit circle which fix given three points.
Later, V\"ais\"al\"a posed the following question in \cite[Question 8]{vaisala4}
which is still open;
can $f$ be extended to a $K-$quasiconformal mapping $F:\mathbb{R}^{2n}
\rightarrow \mathbb{R}^{2n}$ with a constant $K=K(n,\eta)\geq 1$
which depends only on $n$ and $\eta$?
For example, Alestalo--V\"ais\"al\"a showed
that if $f:X\rightarrow \mathbb{R}^n$ is $M-$ biLipschitz, then there always
exists a $\sqrt{7}M^2-$ biLipschitz extension $F:\mathbb{R}^{2n}\rightarrow
\mathbb{R}^{2n}$ of $f$ (see \cite[Theorem 5.5]{alestalo1}).
On the other hand, for quasisymmetric embeddings, there is an obstacle;
Trotsenko--V\"ais\"al\"a proved in \cite[Theorem 6.6]{trotsenko1} that
if $X\subset \mathbb{R}^n$ is not relatively connected, then
there exists a quasisymmetric embedding $f:X\rightarrow \mathbb{R}^n$
which cannot be extended to a quasisymmetric embedding
$F:\mathbb{R}^n\rightarrow\mathbb{R}^N$ for any $N\geq n$.
Since global quasiconformal mappings $F:\mathbb{R}^{2n}\rightarrow
\mathbb{R}^{2n}$ are also quasisymmetric (see \cite[Theorem 11.14]{heinonen1}),
this fact implies that the V\"ais\"al\"a problem cannot be solved
affirmatively for general subsets $X$ even if $n=1$.
According to the recent study by Vellis \cite{vellis1}, he showed that
if $X\subset \mathbb{R}$ is $M-$relatively connected, then
every $\eta-$quasisymmetric embedding $f:X\rightarrow \mathbb{R}^n$
can be extended to an $\eta'-$quasisymmetric embedding $F:\mathbb{R}
\rightarrow \mathbb{R}^N$, where $\eta'$ depends only on $\eta$
and $M$, and $N(\geq n)$ depends only on $n$, $\eta$, and $M$.
Considering the one dimensional case of the V\"ais\"al\"a problem,
it is interesting to find out whether we can choose $N=2$ uniformly
when $n=1$ in the Vellis's result.\\
Let us consider the case of $X=\mathbb{Z}$ and $n=1$.
In this paper, we would like to give
detailed observations on quasisymmetric embeddings $f:\mathbb{Z}\rightarrow
\mathbb{R}$, as an example of a relatively connected set
for which the V\"ais\"al\"a problem can be solved affirmatively;
\setcounter{alpthm}{0}
\begin{alpthm}{\rm (Extensibility of quasisymmetric embeddings of
$\mathbb{Z}$)} \label{T5.1}\\
Every $\eta-$quasisymmetric embedding $f:\mathbb{Z}\rightarrow
\mathbb{R}$ admits a $K=K(\eta)$ $-$quasiconformal extension
$F:\mathbb{C}\rightarrow \mathbb{C}$ where
$K=K(\eta)$ is a constant depending only on $\eta$.
\end{alpthm}
Compared to the Beurling--Ahlfors extension theorem, the difficulty
in our case is that $f$ can change the magnitude relation.
To prove Theorem \ref{T5.1}, first, we will observe the extensibility
of quasisymmetric automorphisms $f:\mathbb{Z}\rightarrow \mathbb{Z}$
in Section \ref{automorphism} and \ref{proofT1}.
\setcounter{alpthm}{1}
\begin{alpthm}{\rm (Extensibility of quasisymmetric automorphisms of $\mathbb{Z}$)}
\label{T3.2}
For a bijection $f:\mathbb{Z}\rightarrow \mathbb{Z}$, the following
conditions are quantitatively equivalent;
\begin{enumerate}
\item $f$ is $\eta-$quasisymmetric.
\item $\{a_n:=f(n)\}_{n\in\mathbb{Z}}$ satisfies the $\lambda-$three point condition.
\item $f$ admits a $K-$quasiconformal extension $F:
\mathbb{C}\rightarrow \mathbb{C}$.
\end{enumerate}
\end{alpthm}
\noindent
We say that a sequence $\{a_n\}_{n\in \mathbb{Z}}$ satisfies the $\lambda-$three point
condition for $\lambda \geq 1$ if $|a_n-a_m|/|a_n-a_k|\leq \lambda$ holds
for any integers $n<m<k$. Thus Theorem \ref{T3.2} does not only state
every quasisymmetric automorphism of $\mathbb{Z}$ is
quasiconformally extensible, but also characterizes the quasisymmetry by a
simple geometric condition. Further, an analogous theorem holds for
quasisymmetric automorphisms of $E=\{e^n\}_{n\in \mathbb{Z}}$
(see Theorem \ref{T3.3}).
Next, we will observe a subset of $\mathbb{R}$ which is an image of a
quasisymmetric embedding $f:\mathbb{Z}\rightarrow \mathbb{R}$
in Section \ref{images}, to complete
the proof of Theorem \ref{T5.1}. In this case, such subsets can also be
characterized by a simple geometric condition as follows;
\setcounter{alpthm}{2}
\begin{alpthm}{\rm (Characterizetion of quasisymmetric images)} \label{T4.1}
For a subset $E\subset \mathbb{R}$, the following conditions
are quantitatively equivalent;
\begin{enumerate}
\item There exists an $\eta-$quasisymmetric bijection
$f:\mathbb{Z}\rightarrow E$.
\item $E$ can be written as a monotone increasing sequence $E=\{a_n\}_{n\in \mathbb{Z}}$
with $a_n\rightarrow\pm \infty\ (n\rightarrow\pm \infty)$, and
there exists a constant $M\geq 1$ such that the following inequality holds
for all $n\in \mathbb{Z}$ and $k\in \mathbb{N}$;
\[
\frac{1}{M} \leq \frac{a_{n+k}-a_n}{a_n-a_{n-k}} \leq M.
\]
\item There exists a $K-$quasiconformal mapping $F:\mathbb{C}\rightarrow\mathbb{C}$,
such that $F(\mathbb{Z})=E$.
\end{enumerate}
\end{alpthm}
\medskip
\section{Definitions and Basic properties}
First, let $\eta:[0,+\infty )\rightarrow [0,+\infty )$ be a homeomorphism
and $X\subset \mathbb{C}$ be a subset.
An injection $f:X\rightarrow \mathbb{C}$ is said to be
$\eta-$quasisymmetric if the following inequality holds for
any three points $x,y,z\in X\ (x\neq z)$;
\begin{align}
\left|\frac{f(x)-f(y)}{f(x)-f(z)}\right|\leq
\eta\left(\left|\frac{x-y}{x-z}\right|\right). \tag{QS} \label{qs}
\end{align}
If $x\neq y$, replacing $y$ and $z$, the following lower estimate holds;
\[
\left|\frac{f(x)-f(y)}{f(x)-f(z)}\right|\geq
\eta\left(\left|\frac{x-y}{x-z}\right|^{-1}\right)^{-1}.
\]
Notice that if there exists at least one $\eta-$quasisymmetric mapping
(and $X$ contains at least two elements),
applying (\ref{qs}) to $y=z$,
it turns out that $\eta$ must satisfy $\eta(1)\geq 1$.
Next, let $K\geq 1$ and $\Omega \subset \mathbb{C}$ be a domain.
An orientation preserving homeomorphism $f:\Omega \rightarrow \mathbb{C}$
into $\mathbb{C}$ is said to be $K-$quasiconformal if its distributional
derivatives $f_z$ and $f_{\bar{z}}$ are in the locally integrable class, and
\[
K(f):=\underset{z\in\Omega}{{\rm ess.sup}}
\frac{|f_z(z)|+|f_{\bar{z}}(z)|}{|f_z(z)|-|f_{\bar{z}}(z)|}\leq K.
\]
$K(f)$ is called the maximal dilatation of $f$.
These two concepts are closely related by the so-called egg-yolk principle
(see \cite[Theorem 11.14]{heinonen1}). In particular, for orientation
preserving homeomorphisms from $\mathbb{C}$ onto itself, the quasiconformality
and the quasisymmetry are quantitatively equivalent.
\medskip
\section{Key Observation}
\label{observation}
We would like to start from a simple observation which is trivial for ones
who are familiar with quasiconformal mappings. However, this observation
will play a central role in the construction of quasiconformal
extensions in later sections. \\
Let us consider a rectangle $R_{a,b}=\left\{ z \in \mathbb{C};\
|{\rm Re} z|<a,\ |{\rm Im} z|<b\right\}$ for $a,b>0$.
For a real number $c\in (0,a)$, we set
\[
f(z):=\left\{
\begin{array}{ll}
z & (z\not \in R_{a,b})\\
\displaystyle \left(a+c-\frac{2c}{b}|y|\right)\frac{x+a}{a-c}-a+iy &
(z=x+iy\in R_{a,b},\ -a\leq x\leq -c)\\
\displaystyle \left(a-c+\frac{2c}{b}|y|\right)\frac{x-a}{a+c}+a+iy &
(z=x+iy\in R_{a,b},\ -c< x\leq a).
\end{array}
\right.
\]
\begin{figure}[h]
\begin{center}
\includegraphics[width=12cm]{obs1.eps}
\end{center}
\caption{} \label{obs1}
\end{figure}
Then $f$ defines a quasiconformal homeomorphism of $\mathbb{C}$
(see Figure \ref{obs1}).
In particuler $f=id$ on $\mathbb{C}\setminus R_{a,b}$,
$f(-c)=c$, and its maximal dilatation depends only on $a,b$ and $c$.
By using this flexible deformation, we have the following preliminary lemma.
\begin{deflem}\label{L1}
Let $n\in \mathbb{N}$ and $\delta \in (0,+\infty]$, and
let
\[
R(n,\delta)=\left\{z\in \mathbb{C};\
\frac{1}{2}<{\rm Re} z< n+\frac{1}{2},\ |{\rm Im} z|< \delta\right\}.
\]
Then, for any bijection
$f:\{1,2,\cdots,n\}\rightarrow \{1,2,\cdots,n\}$,
there exists a $K=K(n,\delta)-$quasiconformal extension
$\widetilde{f}:\mathbb{C}\rightarrow \mathbb{C}$ of f,
such that $\widetilde{f}=id$ on $\mathbb{C}\setminus R(n,\delta)$,
where $K=K(n,\delta)$ is a constant depending only on $n$ and $\delta$.
\end{deflem}
\begin{proof}
We prove the claim by induction with respect to $n\in \mathbb{N}$.
Clearly, the claim holds for $n=1$. We assume the claim holds
for $n-1\ (n\geq 2)$.
Let $f:\{1,2,\cdots,n\}\rightarrow\{1,2,\cdots,n\}$
be a bijection, and let $m:=f(n)$. By the preceding observation,
we can easily construct a quasiconformal mapping $g_1:\mathbb{C}
\rightarrow \mathbb{C}$ which maps $m\mapsto m-i\delta/2,\ n\mapsto n+i\delta/2$,
fixes the other integers, and is identity on $\mathbb{C}\setminus
R(n,\delta)$ (see Figure \ref{obs2}).
Similarly we construct global quasiconformal mappings
$g_2$ which maps $m-i\delta/2\mapsto n-i\delta/2,\ n+i\delta/2\mapsto m+i\delta/2$,
and $g_3$ which maps $n-i\delta/2 \mapsto n,\ m+i\delta/2 \mapsto m$.
Then $g:=g_3 \circ g_2 \circ g_1$ is a quasiconformal mapping which permutes
$n$ and $m$, fixes the other integers, and is identity on $\mathbb{C}
\setminus R(n,\delta)$.
Since the possible values of $m$ are only $n-$kinds, the maximal dilatation of $g$
is bounded by a constant depending only on $n$ and $\delta$.
\begin{figure}[h]
\begin{center}
\includegraphics[width=13cm]{obs3.eps}
\end{center}
\caption{: construction of $g_1$.} \label{obs2}
\end{figure}
By the construction, $f_1:=g^{-1}\circ f$ fixes $n$.
Thus $f_1$ defines a permutation of the set
$\{1,2,\cdots,n-1\}$. From the assumption, $f_1$ extends to a
global quasiconformal mapping which satisfies the conditions of the claim.
Thus, we have a desired extension $\widetilde{f}=g\circ f_1$.
\end{proof}
Each bijection $f:\mathbb{Z}\rightarrow \mathbb{Z}$ can be regarded as
a permutation of countably many elements. On the other hand, Lemma \ref{L1}
guarantees that any permutation of $\{1,2,\cdots,n\}\subset \mathbb{C}$, a finite set,
can be represented by a quasiconformal homeomorphism of $\mathbb{C}$ which deforms only
a small neighborhood $R(n,\delta)$ of $\{1,2,\cdots,n\}$.
\medskip
\section{Extensibility of quasisymmetric automorphisms of $\mathbb{Z}$}
\label{automorphism}
The aim of this section is to prove Theorem \ref{T3.2}.
For this purpose, it is useful to consider bijections
$f:\mathbb{Z}\rightarrow \mathbb{Z}$ as sequences.
\subsection{Splittable bijective sequence}\
We say that a sequence $A=\{a_n\}_{n\in\mathbb{Z}}= \mathbb{Z}$ is \emph{bijective}
if the correspondence $n\mapsto a_n\ (\mathbb{Z}\rightarrow \mathbb{Z})$
is bijective. For two integers $k,\ell \in\mathbb{Z}\ (k\leq \ell)$, we use
the following notations;
\begin{eqnarray*}
[k,\ell]_{\mathbb{Z}}&:=&[k,\ell]\cap \mathbb{Z}=\{k,k+1,\cdots ,\ell-1,\ell \},\\
\bigl| [k,\ell]_{\mathbb{Z}}\bigr|&:=&\# [k,\ell]_{\mathbb{Z}}\ \ =\ \ell-k+1.
\end{eqnarray*}
Remark that we allow the case of $k=\ell$ in the above notations, and in this case,
$[k,k]_{\mathbb{Z}}=\{k\}$ and $\bigl| [k,k]_{\mathbb{Z}}\bigr|=1$.
\begin{defdef}\label{D3.1}
Let $A=\{a_n\}_{n\in\mathbb{Z}}=\mathbb{Z}$ be bijective.
We say that an interval $I=[k,\ell]_{\mathbb{Z}}$ splits $A$ if
$\displaystyle a_n>\max_{j\in I} a_j$ holds for all $n>\ell$,
and $\displaystyle a_n<\min_{j\in I} a_j$ holds for all $n<k$.
Further, we say that $A$ is $C-$splittable for a constant $C\geq 1$ if
there exists a strictly monotone increasing sequence $\{k_n\}_{n\in\mathbb{Z}}
\subset \mathbb{Z}$ which satisfies the following conditions
for all $n\in \mathbb{Z}$;
\begin{itemize} \setlength{\itemsep}{1ex}
\item the interval $I_n:=[k_n +1,k_{n+1}]_{\mathbb{Z}}$ splits $A$,
\item $|I_n|=k_{n+1}-k_n\leq C$.
\end{itemize}
\end{defdef}
By the definition, the following is immediately confirmed.
\begin{deflem} \label{L2}
If an interval $I=[k,\ell]_{\mathbb{Z}}$ splits a
bijective sequence $A=\{a_n\}_{n\in \mathbb{Z}}= \mathbb{Z}$,
then $\{a_n\}_{n\in I}$ is an interval $I'=[k',\ell ']_{\mathbb{Z}}$ with
$|I|=|I'|$. Further $\{a_n\}_{\mathbb{Z}_{>\ell}}=[\ell '+1,+\infty)_{\mathbb{Z}}
=\{\ell'+1,\ell'+2,\cdots\}$,
and $\{a_n\}_{n\in\mathbb{Z}_{<k}}=(-\infty,k'-1]_{\mathbb{Z}}
=\{k'-1,k'-2,\cdots\}$.
\end{deflem}
\begin{proof}
Let $k'= \min_{j\in I} a_j$ and $\ell'=\max_{j\in I} a_j$. By the definition,
we have $\{a_n\}_{n\in I}\subset [k',\ell']_{\mathbb{Z}}$,\ \
$\{a_n\}_{\mathbb{Z}_{>\ell}}\subset [\ell '+1,+\infty)_{\mathbb{Z}}$,
and $\{a_n\}_{n\in\mathbb{Z}_{<k}}\subset (-\infty,k'-1]_{\mathbb{Z}}$.
Since $A=\{a_n\}_{n\in \mathbb{Z}}$ is bijective, the above implications
must be equalities. Further $|I'|=\# \{a_n\}_{n\in I}=|I|$.
\end{proof}
Note that if an interval $I$ splits a bijective sequence
$A=\{a_n\}_{n\in \mathbb{Z}}=\mathbb{Z}$,
then $a_n\rightarrow \pm \infty$ as
$n\rightarrow \pm \infty$.
\begin{defex}
Define a bijective sequence $A=\{a_n\}_{n\in\mathbb{Z}}$ by
\begin{eqnarray*}
a_{6n}=6n,&\ a_{6n+1}=6n+2,&\ a_{6n+2}=6n-2,\\
a_{6n+3}=6n+3,&\ a_{6n+4}=6n+5,&\ a_{6n+5}=6n+1.
\end{eqnarray*}
Then $a_n\rightarrow \pm \infty$ as $n\rightarrow \pm \infty$, but there is no
interval which splits $A$.
\end{defex}
\begin{figure}[h]
\begin{center}
\includegraphics[width=10.5cm]{sbs1.eps}
\end{center}
\caption{\empty} \label{sbs1}
\end{figure}
In Figure \ref{sbs1}, each arrow represents the orbit of the sequence
$A=\{a_n\}_{n\in \mathbb{Z}}$, that is, each arrow starts from $a_n$ and
gets to $a_{n+1}$ for some $n\in\mathbb{Z}$. Such diagrams as Figure \ref{sbs1}
are useful for our argument, and will be used frequently in later sections.
\begin{deflem} \label{L3}
Let $A=\{a_n\}_{n\in\mathbb{Z}}= \mathbb{Z}$ be a bijective sequence,
$C\geq 1$, and $\delta \in (0,+\infty]$.
If $A$ is $C-$splittable, then there exists a $K=K(C,\delta)-$quasiconformal mapping
$\tilde{f}:\mathbb{C}\rightarrow \mathbb{C}$ which satisfies the following conditions;
\begin{itemize}
\item $\widetilde{f}(n)=a_n$ for all $n\in \mathbb{Z}$,
\item $\widetilde{f}=id$ on $\mathbb{C} \setminus \{z\in\mathbb{C};\ |{\rm Im} z|< \delta\}$,
\end{itemize}
where $K=K(C,\delta)$ is a constant depending only on $C$ and $\delta$.
\end{deflem}
\begin{proof}
Since $A$ is $C-$splittable, there is a strictly monotone increasing sequence
$\{k_n\}_{n\in\mathbb{Z}}\subset\mathbb{Z}$ such that each interval
$I_n:=[k_n +1,k_{n+1}]_{\mathbb{Z}}$ splits $A$, and satisfies $|I_n|\leq C$.
Since translation $z\mapsto z+\alpha\ (\mathbb{C}\rightarrow\mathbb{C})$
is conformal for any $\alpha\in \mathbb{C}$,
we may assume $\{a_j\}_{j\in I_n}=[k_n +1,k_{n+1}]_{\mathbb{Z}}=I_n$
for all $n\in\mathbb{Z}$. Namely, the correspondence $j\mapsto a_j
\ (\mathbb{Z}\rightarrow\mathbb{Z})$ splits
into permutations $j\mapsto a_j\ (I_n\rightarrow I_n)\ (n\in \mathbb{Z})$.
By Lemma \ref{L1}, for each $n\in \mathbb{Z}$, there
exists a $K=K(C,\delta)-$quasiconformal mapping $\widetilde{f}_n:\mathbb{C}\rightarrow
\mathbb{C}$ such that
\begin{itemize}
\item $\widetilde{f}_n(j)=a_j$ for all $j\in I_n$,
\item $\widetilde{f}_n=id$ on $\mathbb{C}\setminus \displaystyle
\left\{z;\ k_n +\frac{1}{2} < {\rm Re} z < k_{n+1} +\frac{1}{2},\
|{\rm Im} z|< \delta \right\}$.
\end{itemize}
Let $\widetilde{f}(z):=\widetilde{f}_n(z)$ if
$z\in \{z;\ k_n+1/2< {\rm Re} z \leq k_{n+1}+1/2\}.$
Then, clearly $\widetilde{f}$ defines a homeomorphism of $\mathbb{C}$.
Since domains $\{z;\ k_n+1/2< {\rm Re} z < k_{n+1}+1/2\}
\ (n\in \mathbb{Z})$ are disjoint, the maximal
dilatation of $\widetilde{f}$ is also $K$. Thus $\widetilde{f}$ is a desired mapping.
\end{proof}
\subsection{Three point condition for bijective sequences}\
\begin{defdef}
Let $A=\{a_n\}_{n\in\mathbb{Z}}=\mathbb{Z}$ be a
bijective sequence, and let $\lambda \geq 1$.
We say that $A$ satisfies the $\lambda-$three point condition if
for any integers $n<m<k$, it holds that
\begin{equation}
\left|\frac{a_n-a_m}{a_n-a_k}\right|\leq \lambda. \tag{3PC} \label{3PC}
\end{equation}
\end{defdef}
Suppose $A$ satisfies the $\lambda-$three point condition. Then for
any integers $n<m<k$, it holds from the triangle inequality that
\[
\lambda \geq \left|\frac{a_n-a_m}{a_n-a_k}\right| \geq \left| 1-
\left|\frac{a_k-a_m}{a_n-a_k}\right|\right|.
\]
Thus, we have a symmetric condition;
\begin{equation}
\left|\frac{a_k-a_m}{a_k-a_n}\right| \leq \lambda +1. \tag{3PC'} \label{3PC'}
\end{equation}
\begin{defremark} \label{remark1}
We would like to emphasise that the condition {\rm (\ref{3PC})} holds
even if $n=m$, and the condition {\rm (\ref{3PC'})} holds even if $m=k$.
This makes our arguments concise.
\end{defremark}
The conditions (\ref{3PC}) and (\ref{3PC'}) have a simple geometrical meaning;
suppose the orbit starts from a certain point $a_n$ and
goes far away, say $a_m\ (m>n)$, then the orbit
$\{a_j\}_{j>m}$ cannot return to a point
near to $a_n$ above a certain rate (see Figure \ref{3pc1}).
\begin{figure}[h]
\begin{center}
\includegraphics[width=10cm]{3pc1.eps}
\end{center}
\caption{: $\lambda-$three point condition} \label{3pc1}
\end{figure}
\begin{defprop}\label{P1}
If a bijective sequence $A=\{a_n\}_{n\in\mathbb{Z}}=\mathbb{Z}$ satisfies
the $\lambda-$three point condition for some $\lambda \geq 1$, then
$a_n\rightarrow \pm\infty\ (n\rightarrow \pm \infty)$ or
$a_n\rightarrow \mp\infty\ (n\rightarrow \pm \infty)$ holds.
\end{defprop}
\begin{proof}
First, we prove $a_n\rightarrow +\infty$ or $a_n\rightarrow -\infty$ as
$n\rightarrow +\infty$. To obtain a contradiction, we assume
$a_n \not \rightarrow \pm \infty$ as $n\rightarrow +\infty$. Further,
we suppose $a_0=0$ for simplicity.
By the assumption, $\{a_n\}_{n=0}^{+\infty}$ is unbounded from above
and below. Thus there is an integer $n_1>0$ such that $|a_{n_1}|>
\lambda+1$ holds, and $a_{n_1}+1 \in \{a_n\}_{n=-\infty}^{-1}$ or
$a_{n_1}-1\in \{a_n\}_{n=-\infty}^{-1}$ holds. Indeed if such an integer
does not exist, $\{a_n\}_{n=0}^{+\infty}$ must contain $\{m\in \mathbb{Z};
\ |m|>\lambda\}$. This cannot occur since $\{a_n\}_{n=-\infty}^{-1}
= \mathbb{Z} \setminus \{a_n\}_{n=0}^{+\infty}$ is an infinite subset.
Let $n_{-1}<0$ be an integer which satisfies $a_{n_{-1}}=a_{n_1}+1$ or
$a_{n_{-1}}=a_{n_1}-1$. Then, by the three point condition,
We have a contradiction;
\[
\lambda \geq \left| \frac{a_{n_{-1}}-a_0}{a_{n_{-1}}-a_{n_1}}\right|
=|a_{n_{-1}}|>\lambda.
\]
Similarly, we can show $a_n\rightarrow +\infty$ or $a_n\rightarrow -\infty$
as $n\rightarrow -\infty$. Obviously, if $a_n\rightarrow +\infty\ (n
\rightarrow +\infty)$, then $a_n \not \rightarrow +\infty\ (n\rightarrow -\infty)$.
Thus we have the claim.
\end{proof}
\subsection{Key Theorem and Extensibility of Quasisymmetric automorphisms}\
The following theorem will be proved in Section \ref{proofT1}.
\begin{defthm}\label{T1}
Let $A=\{a_n\}_{n\in\mathbb{Z}}=\mathbb{Z}$ be bijective,
$\lambda \geq 1$, and $\delta\in (0,+\infty]$. If $A$ satisfies the $\lambda-$three
point condition and $a_n\rightarrow \pm\infty\ (n\rightarrow \pm\infty)$,
then there exists a $K=K(\lambda,\delta)-$quasiconformal mapping
$\widetilde{f}:\mathbb{C}\rightarrow \mathbb{C}$ such that
\begin{itemize}
\item $\widetilde{f}=id$ on $\mathbb{C}\setminus \{z;\ |{\rm Im} z|< \delta\}$,
\item $B=\{b_n:=\widetilde{f}(a_n)\}_{n\in\mathbb{Z}}$ is
$(2\lambda +3)-$splittable,
\end{itemize}
where $K=K(\lambda,\delta)$ is a constant depending only on $\lambda$ and $\delta$.
\end{defthm}
As a direct corollary of Theorem \ref{T1}, we have the following;
\setcounter{alpthm}{1}
\begin{alpthm}{\rm (Extensibility of quasisymmetric automorphisms of $\mathbb{Z}$)}
\label{T3.2}
For a bijection $f:\mathbb{Z}\rightarrow \mathbb{Z}$, the following
conditions are quantitatively equivalent;
\begin{enumerate}
\item $f$ is $\eta-$quasisymmetric.
\item $\{a_n:=f(n)\}_{n\in\mathbb{Z}}$ satisfies the $\lambda-$three point condition.
\item $f$ admits a $K-$quasiconformal extension $\widetilde{f}:
\mathbb{C}\rightarrow \mathbb{C}$.
\end{enumerate}
\end{alpthm}
\begin{proof}
First, $(1)\Rightarrow (2)$ is clear. Indeed, for any integers $n<m<k$, we have
\[
\left| \frac{f(n)-f(m)}{f(n)-f(k)} \right|
\leq \eta \left( \left|\frac{n-m}{n-k}\right|\right) \leq \eta(1).
\]\vspace{0.5ex}
Next, $(3)\Rightarrow (1)$ is also clear, since $K-$quasiconformal
self-homeomorphisms of $\mathbb{C}$
are $\eta-$quasisymmetric with an $\eta$ depending only on $K$
(thus the restrictions to $\mathbb{Z}$
are also $\eta-$quasisymmetric with the same $\eta$).\vspace{1ex}
Last, we prove that $(2)\Rightarrow (3)$. By Proposition \ref{P1}, $a_n
\rightarrow \pm \infty\ (n\rightarrow\pm \infty)$ or $a_n\rightarrow \mp
\infty\ (n\rightarrow \pm \infty)$ holds. Since $z\mapsto -z\ (\mathbb{C}
\rightarrow \mathbb{C})$ is conformal, we may assume the former case holds.
Then we can apply Theorem \ref{T1}, that is, there exists a
$K_1=K_1(\lambda)-$quasiconformal mapping $\widetilde{f}_1:\mathbb{C}
\rightarrow \mathbb{C}$ such that
$B=\{b_n:=\widetilde{f}_1(a_n)\}_{n\in\mathbb{Z}}$ is $(2\lambda +3)-$splittable.
Further, applying Lemma \ref{L3}, we have $K_2=K_2(\lambda)-$quasiconformal
mapping $\widetilde{f}_2:\mathbb{C}\rightarrow\mathbb{C}$ such that
$\widetilde{f}_2(n)=b_n=\widetilde{f}_1(a_n)=\widetilde{f}_1 \circ f(n)$.
Therefore we obtain a desired extension $\widetilde{f}=\widetilde{f}_1^{-1}
\circ \widetilde{f}_2$.
\end{proof}
Next, we consider quasisymmetric automorphisms of $E:=\{e^n\}_{n\in\mathbb{Z}}$.
In this case, we analogously obtain the following theorem;
\begin{defthm} \label{T3.3}
For a bijection $f:E=\{e^n\}_{n\in\mathbb{Z}}\rightarrow E$, the
following conditions are quantitatively equivalent;
\begin{enumerate}
\item $f$ is $\eta-$quasisymmetric.
\item $b_n:=f(e^n) \rightarrow 0$ as $n\rightarrow -\infty$, and
$\{b_n\}_{n\in\mathbb{Z}}$ satisfies the $\lambda-$three point
condition, that is, for any integers $n<m<k$ it holds that
\[
\left| \frac{b_n-b_m}{b_n-b_k} \right|
=\left| \frac{f(e^n)-f(e^m)}{f(e^n)-f(e^k)}\right|
\leq \lambda.
\]
\item $f$ admits a $K-$quasiconformal extension $\widetilde{f}:
\mathbb{C}\rightarrow\mathbb{C}$.
\end{enumerate}
\end{defthm}
Assume $f:E\rightarrow E$ is $\eta-$quasisymmetric. Since quasisymmetric
mappings map Cauchy sequences to Cauchy sequences, we have
$f(e^n)\rightarrow 0\ (n\rightarrow -\infty)$. Further, for any integers $n<m<k$
\[
\left| \frac{f(e^n)-f(e^m)}{f(e^n)-f(e^k)}\right|
\leq \eta\left( \left| \frac{e^n-e^m}{e^n-e^k}\right|\right)<\eta(1).
\]
Thus $(1)\Rightarrow (2)$ is valid, and $(3)\Rightarrow(1)$ is clear
for the same reason as the preceding proof.
\begin{defremark}
In the condition $(2)$, $f(e^n)\rightarrow 0\ (n\rightarrow -\infty)$ is
necessary. More precisely, the $\lambda-$three point condition does not imply this
property. In fact, for $f:e^n\mapsto e^{-n}\ (E\rightarrow E)$, the sequence
$\{ b_n:=f(e^n)\}_{n\in \mathbb{Z}}$ satisfies the $1-$three point condition,
but $b_n=f(e^n)\rightarrow +\infty\ (n\rightarrow -\infty)$.
\end{defremark}
Thus we only need to show $(2)\Rightarrow (3)$. To prove this,
we prepare some lemmas. Let us assume $f:E\rightarrow E$ satisfies the condition $(2)$.\\
Let $a_n:=\log \circ f \circ \exp (n)$ (then $b_n=e^{a_n}$ holds).
Since $b_n \rightarrow 0\ (n\rightarrow -\infty)$, we have $a_n\rightarrow -\infty\
(n\rightarrow -\infty)$. Note that $A:=\{a_n\}_{n\in \mathbb{Z}}=\mathbb{Z}$
is a bijective sequence.
\begin{deflem} \label{L3.4}
There exists a constant $C_{\lambda}\geq 0$ depending only on $\lambda$,
such that $a_k-a_{\ell} \leq C_{\lambda}$ holds if $k<\ell$ and $a_{\ell}<a_k$.
\end{deflem}
\begin{proof}
Assume that integers $k,\ell$ satisfy
$k<\ell$ and $a_{\ell}<a_k$. Since $a_n\rightarrow -\infty\
(n\rightarrow -\infty)$, there exists an integer $j<k$ such that
$a_j<a_{\ell}$. Thus, by the three point condition,
\begin{eqnarray*}
\lambda &\geq & \left|\frac{b_j-b_k}{b_j-b_{\ell}}\right| \\
&=& \left| \frac{e^{a_j}-e^{a_k}}{e^{a_j}-e^{a_{\ell}}}\right|
=\frac{e^{a_k-a_{\ell}}-e^{-(a_{\ell}-a_j)}}{1-e^{-(a_{\ell}-a_j)}}.
\end{eqnarray*}
Since $0<e^{-(a_{\ell}-a_j)}<1$, we have $\lambda \geq e^{a_k-a_{\ell}} -1$.
Therefore $a_k-a_{\ell}\leq \log (\lambda +1)=:C_{\lambda}$.
\end{proof}
\begin{deflem} \label{L3.5}
$\{a_n\}_{n\in \mathbb{Z}}$ satisfies the $(C_{\lambda}+1)-$three point condition,
where $C_{\lambda}$ is a constant in Lemma \ref{L3.4}.
\end{deflem}
\begin{proof}
Let $n<m<k$. If $|a_n-a_k|\geq |a_n-a_m|$, then $|a_n-a_m|/|a_n-a_k|
\leq 1 \leq C_{\lambda}+1$. Thus we consider the case of $|a_n-a_k|< |a_n-a_m|$.
First, if $a_n>a_m$, then we have
\[
\left| \frac{a_n-a_m}{a_n-a_k} \right| \leq a_n-a_m \leq C_{\lambda}
<1+C_{\lambda}.
\]
In this estimation, remark that $a_n$ and $a_k$ are distinct integers,
that is, $|a_n-a_k|\geq 1$ holds.
Next, if $a_n <a_m$, by $|a_n-a_k|<|a_n-a_m|$ it holds $a_k<a_m$.
Thus, by Lemma \ref{L3.4} we have $0<a_m-a_k\leq C_{\lambda}$ and
\[
\left|\frac{a_n-a_m}{a_n-a_k}\right| \leq
\frac{|a_n-a_k|+|a_k-a_m|}{|a_n-a_k|} \leq 1+ C_{\lambda}.
\]
\end{proof}
\begin{defprop} \label{P3.2}
$(2) \Rightarrow (3)$ in Theorem \ref{T3.3} holds.
\end{defprop}
\begin{proof}
By Lemma \ref{L3.5} and Proposition \ref{P1}, the sequence $A=
\{a_n:=\log b_n\}_{n\in\mathbb{Z}}=\mathbb{Z}$ satisfies the
$(C_{\lambda}+1)-$three point condition and $a_n\rightarrow \pm \infty\
(n\rightarrow \pm \infty)$. Thus by Theorem \ref{T1} and Lemma \ref{L3}, there exists
a $K=K(\lambda)-$quasiconformal mapping $\widetilde{G}:\mathbb{C}\rightarrow
\mathbb{C}$ such that
\begin{itemize}
\item $\widetilde{G}(n)=a_n$,
\item $\widetilde{G}=id$ on $\mathbb{C}\setminus \{z;\ |{\rm Im} z|< \pi\}$.
\end{itemize}
Define a homeomorphism $\widetilde{F}:\mathbb{C}\rightarrow \mathbb{C}$ by
\[
\widetilde{F}(z+2n\pi i):=\widetilde{G}(z)+2n\pi i,
\]
for $n\in\mathbb{Z}$, and $z\in \{z;\ -\pi<{\rm Im} z\leq \pi\}$.
Clearly, $\widetilde{F}$ is $K=K(\lambda)-$quasiconformal
and $\widetilde{F}(n)=a_n\ (n\in \mathbb{Z})$.
Thus, the projection of $\widetilde{F}$ with respect to the
universal covering $\pi:z\mapsto e^z\ (\mathbb{C}\rightarrow \mathbb{C}^{\ast})$,
that is, the mapping $\widetilde{f}:\mathbb{C}^{\ast}\rightarrow
\mathbb{C}^{\ast}$ defined by $\widetilde{f}\circ \pi =\pi \circ \widetilde{F}$
gives a $K-$quasiconformal extension
of $f$. Since $b_n=\widetilde{f}(e^n) \rightarrow 0\
(n\rightarrow -\infty)$, we
obtain a desired extension by the removable singularity theorem for
quasiconformal mappings.
\end{proof}
\begin{defremark}
By the construction of $\tilde{f}$, it turns out that
we can choose the quasiconformal extension in Theorem \ref{T3.3}
so that it is identity on the negative real axis.
\end{defremark}
\medskip
\section{Proof of Theorem \ref{T1}}
\label{proofT1}
We devote this section to the proof of Theorem \ref{T1}.
The statement of Theorem \ref{T1} is the following;
\begin{empthm}
Let $A=\{a_n\}_{n\in\mathbb{Z}}=\mathbb{Z}$ be bijective,
$\lambda \geq 1$, and $\delta\in (0,+\infty]$. If $A$ satisfies the $\lambda-$three
point condition and $a_n\rightarrow \pm\infty\ (n\rightarrow \pm\infty)$,
then there exists a $K=K(\lambda,\delta)-$quasiconformal mapping
$\widetilde{f}:\mathbb{C}\rightarrow \mathbb{C}$ such that
\begin{itemize}
\item $\widetilde{f}=id$ on $\mathbb{C}\setminus \{z;\ |{\rm Im} z|< \delta\}$,
\item $B=\{b_n:=\widetilde{f}(a_n)\}_{n\in\mathbb{Z}}$ is
$(2\lambda +3)-$splittable,
\end{itemize}
where $K=K(\lambda,\delta)$ is a constant depending only on $\lambda$ and $\delta$.
\end{empthm}
We would like to start proving this claim. Throughout this section,
we assume $A=\{a_n\}_{n\in \mathbb{Z}}=\mathbb{Z}$ satisfies the $\lambda-$three
point condition $(\lambda \geq 1)$ and
$a_n\rightarrow \pm \infty\ (n\rightarrow \pm \infty)$.
Further, let $\delta\in (0,+\infty]$.
\subsection*{Step1}
By translation, we may assume $a_0=0$.
Since $a_n\rightarrow \pm \infty\ (n\rightarrow \pm \infty)$,
there uniquely exist integers $k_0$ and $k_1$ such that
\begin{align}
a_{k_0+1}\geq 0\ \ \ &{\rm and}\ \ \ n<k_0+1\Rightarrow a_n<0,
\tag{C0} \label{A}\\
a_{k_1}\leq 0\ \ \ &{\rm and}\ \ \ n>k_1\Rightarrow a_n>0.
\tag{C1} \label{B}
\end{align}
Remark that since $a_0=0$, it holds that $k_0<0\leq k_1$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=12cm]{pT11.eps}
\end{center}
\caption{: Orbit of $A=\{a_n\}_{n\in\mathbb{Z}}$} \label{pT11}
\end{figure}
Let $I_0:=[k_0+1,k_1]_{\mathbb{Z}}$, and let $\ell_0, r_0\in I_0$
satisfy $a_{\ell_0}=\displaystyle \max_{j\in I_0} a_j$, and
$a_{r_0}=\displaystyle \min_{j\in I_0} a_j$ (see Figure \ref{pT11}).
Further, let $\lambda':=\lambda +1$.
\begin{defclaim} \label{claim1}
$|a_{r_0}|,\ |a_{\ell _0}|\leq \lambda'$ and
$|I_0|\leq 2\lambda'+1$ hold.
\end{defclaim}
\begin{proof}
By the three point condition, the following hold for any
integers $n,m,k\in\mathbb{Z}$
(see Remark \ref{remark1});
\begin{eqnarray*} \displaystyle
n\leq m<k\ \ &\Longrightarrow & \ \ \left|\frac{a_n-a_m}{a_n-a_k}\right|
\leq \lambda <\lambda' ,\\
n< m\leq k\ \ &\Longrightarrow & \ \ \left|\frac{a_k-a_m}{a_k-a_n}\right|
\leq \lambda'.
\end{eqnarray*}
Let $j$ be an integer such that $a_j=a_{\ell _0}+1$. Then by the condition
(\ref{A}), we have $j>k_1(\geq \ell _0)$. Thus,
\[
\lambda' \geq \left| \frac{a_{\ell_0}-a_{k_1}}{a_{\ell_0}-a_j}\right|
=|a_{\ell_0}-a_{k_1}|.
\]
Since $a_{\ell_0}\geq 0$ and $a_{k_1}\leq 0$, we have $|a_{\ell_0}|\leq \lambda'$.
Similarly, we can show $|a_{r_0}|\leq \lambda'$.
Further, we have $|I_0|\leq a_{\ell_0}-a_{r_0}+1\leq 2\lambda'+1$.
\end{proof}
Next, we sort the interval $[a_{r_0},a_{\ell_0}]_{\mathbb{Z}}$ appropriately
by a global quasiconformal mapping. Let $\{c_n\}_{n=m_0}^{m_1}$ be the
unique sequence such that
\[
\left\{
\begin{array}{l}
\{c_n\}_{n=m_0}^{m_1}=\{a_n\}_{n=k_0+1}^{k_1},\vspace{1ex}\\
a_{r_0}=c_{m_0}<c_{m_0+1}<\cdots
<c_0=a_0=0<\cdots<c_{m_1}=a_{\ell_0}.
\end{array}
\right.
\]
That is, $\{c_n\}_{n=m_0}^{m_1}$ is the ascending sort of $\{a_n\}_{n\in I_0}$
normalized by $c_0=a_0(=0)$. Remark that $m_0$ and $m_1$ are also
uniquely determined by the above conditions.
Similarly, let $\{d_n\}_{n=p_0}^{-1}$ and $\{d_n\}_{n=1}^{p_1}$ be the
unique sequences such that
\[
\left\{
\begin{array}{l}
\{d_n\}_{n=p_0}^{-1}\cup \{d_n\}_{n=1}^{p_1}=
[a_{r_0},a_{\ell_0}]_{\mathbb{Z}}\setminus \{a_n\}_{n\in I_0},\vspace{1ex}\\
d_{p_0}<d_{p_0+1}<\cdots<d_{-1}<0<d_{1}<\cdots<d_{p_1}.
\end{array}
\right.
\]
$p_0$ and $p_1$ are also uniquely determined by the conditions $m_0+p_0=a_{r_0}$ and
$m_1+p_1=a_{\ell_0}$. (If $p_0=0$ or $p_1=0$, then we assume the corresponding
sequences are empty.)
By Lemma \ref{L1} and Claim \ref{claim1}, there exists a
$K=K(\lambda,\delta)-$quasiconformal mapping $\widetilde{f}_0:
\mathbb{C}\rightarrow\mathbb{C}$ such that
\begin{enumerate}
\item $\widetilde{f}_0=id$ on $\mathbb{C}\setminus
\left\{ z\in\mathbb{C};\ \displaystyle
a_{r_0}-\frac{1}{2}<{\rm Re} z< a_{\ell_0}+\frac{1}{2},\ |{\rm Im} z|< \delta
\right\}$,
\item $\widetilde{f}_0(d_j)=m_0+j\ \ \ (j=-1,-2,\cdots,p_0)$,\vspace{1ex}\\
$\widetilde{f}_0(c_j)=j\hspace{8ex} (j=m_0,m_0+1,\cdots,m_1)$,\vspace{1ex}\\
$\widetilde{f}_0(d_j)=m_1+j\ \ \ (j=1,2,\cdots,p_1)$.
\end{enumerate}
Using this mapping, we set $A_0:=\{a_n^0:=\widetilde{f}_0(a_n)\}_{n\in \mathbb{Z}}$
(see Figure \ref{pT12}).
Then $I_0=[k_0+1,k_1]_{\mathbb{Z}}$ splits $A_0$ and $|I_0|\leq 2\lambda'+1=
2\lambda+3$.\vspace{3ex}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=13cm]{pT12.eps}
\end{center}
\caption{: permutation by $\widetilde{f}_0$} \label{pT12}
\end{figure
\subsection*{Step2}
By Lemma \ref{L2}, $\{a_n^0\}_{n\in I_0}=[m_0,m_1]_{\mathbb{Z}}$ and
$\{a_n^0\}_{n\in \mathbb{Z}_{>k_1}}=[m_1+1,+\infty)_{\mathbb{Z}}
=\{m_1+1,m_1+2,\cdots\}$.
Let $k_2$ be the maximum integer which satisfies
$a_{k_2}^0\in [m_1+1,a_{\ell_0}+1]_{\mathbb{Z}}$,
and let $I_1:=[k_1+1,k_2]_{\mathbb{Z}}$. Further
we let $\ell_1 \in \mathbb{Z}$
be the integer such that $a_{\ell_1}^0=\displaystyle\max_{j\in I_1}a_j^0$.
Remark that since $\widetilde{f}_0=id$ on $\{z;\ {\rm Re} z>a_{\ell_0}+1/2\}$,
it holds $a_n^0=a_n$ if $a_n^0\geq a_{\ell_0}+1$.
In particular, $a_{\ell_1}^0=a_{\ell_1}$.
\begin{defclaim} \label{claim2}
$|a_{\ell_1}-(a_{\ell_0}+1)|\leq \lambda'$, and $|I_1|\leq 2\lambda'+1$ hold.
\end{defclaim}
\begin{proof}
Let $j\in \mathbb{Z}$ satisfy $a_j^0=a_{\ell_1}^0+1=a_{\ell_1}+1$.
Then $j>k_2(\geq \ell_1)$. Further by the preceding remark, it holds $a_j^0=a_j$.
By the definition of $\widetilde{f}_0$, we have $a_{\ell_0}+1\geq a_{k_2}^0
=\widetilde{f}_0(a_{k_2})\geq a_{k_2}$. Therefore,
\[
\lambda'\geq \left|\frac{a_{\ell_1}-a_{k_2}}{a_{\ell_1}-a_j}\right|
\geq|a_{\ell_1}-(a_{\ell_0}+1)|.
\]
Moreover by Claim \ref{claim1}, we have
\begin{eqnarray*}
|I_1|\leq |[m_1+1,a_{\ell_1}]_{\mathbb{Z}}|&=& a_{\ell_1}-m_1\\
&\leq &a_{\ell_1}-(a_{\ell_0}+1)+a_{\ell_0}+1 \leq 2\lambda' +1.
\end{eqnarray*}
\end{proof}
Similarly to Step1, we sort the interval $[a_{\ell_0}+1,a_{\ell_1}]_{\mathbb{Z}}$
appropriately by a global quasiconformal mapping. Let
$\{c_n^1\}_{n=a_{\ell_0}+1}^{m_2}$ be the unique sequence such that
\[
\left\{
\begin{array}{l}
\{c_n^1\}_{n=a_{\ell_0}+1}^{m_2}=[a_{\ell_0}+1,a_{\ell_1}]_{\mathbb{Z}}
\cap \{a_n^0\}_{n\in I_1},\vspace{1ex}\\
a_{\ell_0}+1=c_{a_{\ell_0}+1}^1<c_{a_{\ell_0}+2}^1<\cdots<c_{m_2-1}^1
<c_{m_2}^1=a_{\ell_1},
\end{array}
\right.
\]
and let $\{d_n^1\}_{n=1}^{p_2}$ be the unique sequence such that
\[
\left\{
\begin{array}{l}
\{d_n^1\}_{n=1}^{p_2}=[a_{\ell_0}+1,a_{\ell_1}]_{\mathbb{Z}}
\setminus \{a_n^0\}_{n\in I_1},\vspace{1ex}\\
d_1^1<d_2^1<\cdots<d_{p_2}^1.
\end{array}
\right.
\]
Again, we remark that $p_2$ is automatically determined by the
equation $m_2+p_2=a_{\ell_1}$, and if $p_2=0$, we assume
$\{d_n^1\}_{n=1}^{p_2}$ is empty.
By Lemma \ref{L1} and Claim \ref{claim2}, there exists a $K-$quasiconformal
mapping $\widetilde{f}_1:\mathbb{C}\rightarrow\mathbb{C}$ which satisfies
\begin{enumerate}
\item $\widetilde{f}_1=id$ on $\mathbb{C}\setminus
\left\{ z\in\mathbb{C};\ \displaystyle
a_{\ell_0}+\frac{1}{2}<{\rm Re} z< a_{\ell_1}+\frac{1}{2},\ |{\rm Im} z|< \delta
\right\}$,
\item $\widetilde{f}_1(c_j)=j\hspace{8.5ex} (j=a_{\ell_0}+1,
a_{\ell_0}+2,\cdots,m_2)$,\vspace{1ex}\\
$\widetilde{f}_1(d_j)=m_2+j\ \ \ (j=1,2,\cdots,p_2)$,
\end{enumerate}
where $K$ is the same constant appeared in the construction of $\widetilde{f}_0$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=11cm]{pT13.eps}
\end{center}
\caption{: permutation by $\widetilde{f}_1$} \label{pT13}
\end{figure}
Let $A_1:=\{a_n^1:=\widetilde{f}_1(a_n^0)\}_{n\in\mathbb{Z}}$. Then intervals
$I_0=[k_0+1,k_1]_{\mathbb{Z}}$ and $I_1=[k_1+1,k_2]_{\mathbb{Z}}$ split $A_1$,
and $|I_0|,\ |I_1|\leq 2\lambda'+1=2\lambda +3$ hold.
Furthermore, since $\widetilde{f}_0$ and $\widetilde{f}_1$ deform disjoint domains,
the maximal dilatation of $\widetilde{f}_1 \circ \widetilde{f}_0$
does not increase. Namely,
$\widetilde{f}_1 \circ \widetilde{f}_0$ is also $K-$quasiconformal (see Figure
\ref{pT13}).
\subsection*{Step3}
Applying Step2 repeatedly, for each $m\geq 2$, we can construct
an interval $I_m=[k_m+1,k_{m+1}]_{\mathbb{Z}}$,
a $K-$quasiconformal mapping $\widetilde{f}_m:\mathbb{C}
\rightarrow \mathbb{C}$, and a sequence
$A_m=\{a_n^m:=\widetilde{f}_m(a_n^{m-1})\}_{n\in\mathbb{Z}}$ such that
\begin{itemize}
\item intervals $I_0,\ I_1,\cdots,\ I_m$ split $A_m$,\vspace{1ex}
\item $|I_0|,\ |I_1|,\cdots,\ |I_m|\leq 2\lambda+3$,\vspace{1ex}
\item $\widetilde{f}_j$ are identity on $\mathbb{C}\setminus
\{z;\ |{\rm Im} z|< \delta\}$.
\end{itemize}
Furthermore, by the construction of $\widetilde{f}_j$, the mappings
$\widetilde{f}_j\ (j=0,1,\cdots,m)$ deform disjoint domains. Thus
$\widetilde{f}_m\circ \cdots \circ \widetilde{f}_0$ converges to
a $K-$quasiconformal mapping uniformly on each compact subset
of $\mathbb{C}$ as $m\rightarrow +\infty$.
Further, we can apply the same argument to the negative direction
of $\{a_n\}_{n\in\mathbb{Z}}$. Consequently we have a desired
$K-$quasiconformal mapping $\widetilde{f}:\mathbb{C}\rightarrow
\mathbb{C}$.
\hfill $\Box$
\medskip
\section{Characterization of quasisymmetric images}
\label{images}
In this section, we characterize subsets $E\subset \mathbb{R}$
which are images of some quasisymmetric embeddings
$f:\mathbb{Z}\rightarrow \mathbb{R}$. On the other hand,
the author have characterized images of quasiconformal
mappings as follows;
\begin{defthm} \label{oldthm}
{\rm (F. 2015 \cite[Theorem A]{fujino2})}
For a subset $E\subset \mathbb{R}$, the following conditions are
quantitatively equivalent.
\begin{enumerate}
\item There exists a $K-$quasiconformal mapping $F:\mathbb{C}\rightarrow\mathbb{C}$,
such that $F(\mathbb{Z})=E$.
\item $E$ can be written as a monotone increasing sequence $E=\{a_n\}_{n\in \mathbb{Z}}$
with $a_n\rightarrow\pm \infty\ (n\rightarrow\pm \infty)$, and
there exists a constant $M\geq 1$ such that the following inequality holds
for all $n\in \mathbb{Z}$ and $k\in \mathbb{N}$;
\[
\frac{1}{M} \leq \frac{a_{n+k}-a_n}{a_n-a_{n-k}} \leq M.
\]
\end{enumerate}
Further, if $E$ satisfies the second condition,
there exists a quasiconformal mapping $F:\mathbb{C}\rightarrow\mathbb{C}$
such that $F(n)=a_n$ for all $n\in \mathbb{Z}$.
\end{defthm}
We will see that the above conditions are desired characterizations.
To see this, we can use almost the same proof as \cite[Theorem A]{fujino2}.
However, we would like to give proofs here for completeness and convenience.
First, we prepare some preliminary lemmas.
\begin{defremark}
If $E\subset \mathbb{R}$ is an image of a quasisymmetric mapping $f:\mathbb{Z}
\rightarrow \mathbb{R}$, since quasisymmetric mappings take Cauchy sequences
to Cauchy sequences, $E$ must be closed and discrete in $\mathbb{R}$.
\end{defremark}
\begin{deflem} \label{L4.1}
Let $f:\mathbb{Z}\rightarrow \mathbb{R}$ be an $\eta-$quasisymmetric mapping,
and let $E:=f(\mathbb{Z})$. Then $\sup E=+\infty$ and $\inf E=-\infty$.
\end{deflem}
\begin{proof}
To obtain a contradiction, we assume $\inf E>-\infty$. Since $E$ is closed
and discrete, we have $\sup E=+\infty$.
Thus $E$ can be written as a monotone increasing sequence $E=\{a_n\}_{n\in
\mathbb{N}}$ with $a_n\rightarrow +\infty$ as $n\rightarrow +\infty$.
Let $g:=f^{-1}:E\rightarrow \mathbb{Z}$. By translation, we may assume
$g(a_1)=0$. Further, note that $g$ is $\eta'-$quasisymmetric where
$\eta'(t)=1/\eta^{-1}(1/t)$. Let $\mu :=\eta'(1)$ and consider the set
\[
S:=\left\{k\in\mathbb{N};\ g(a_k)=\max_{j=1,2,\cdots,k} g(a_j)\geq \mu \right\}.
\]
Since $g:E\rightarrow \mathbb{Z}$ is bijective,
$S$ consists of infinitely many elements.
We number $S=\{k_j\}_{j\in \mathbb{N}}$ in ascending order.
Then the sequence $\{g(a_{k_j})\}_{j\in\mathbb{N}}\subset \mathbb{Z}$
is monotone increasing.
On the other hand, there exist infinitely many $n\in \mathbb{N}$ with $g(a_n)<0$.
Thus we can find $j, \ell \in \mathbb{N}$ such that $k_j < \ell < k_{j+1}$ and
$g(a_{\ell})<0$. Moreover since $g(a_n)\leq g(a_{k_j})$ for all
$n=1,2,\ldots, k_{j+1}-1$, if $g(a_m)=g(a_{k_j})+1$ then $m\geq k_{j+1}$.
Consequently we confirmed that there exists $k\in S$ and exist
$\ell,m \in \mathbb{N}$ such that
\begin{itemize}
\item $k<\ell<m$,
\item $g(a_{\ell})<0$ and $g(a_m)=g(a_k)+1$ (see Figure \ref{img1}).
\end{itemize}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=12cm]{img1.eps}
\end{center}
\caption{} \label{img1}
\end{figure}
Therefore, we have a contradiction;
\begin{eqnarray*}
\mu > \eta'\left(\left|\frac{a_k-a_{\ell}}{a_k-a_m}\right|\right)
&\geq &\left| \frac{g(a_k)-g(a_{\ell})}{g(a_k)-g(a_m)}\right|\\
&=&g(a_k)-g(a_{\ell})>g(a_k)\geq \mu.
\end{eqnarray*}
\end{proof}
\begin{deflem} \label{L4.2}
Let $E=\{a_n\}_{n\in\mathbb{Z}}\subset \mathbb{R}$ be a monotone
increasing sequence with $a_n\rightarrow \pm \infty$ as
$n\rightarrow \pm \infty$. If $g:E\rightarrow \mathbb{Z}$
is an $\eta'-$quasisymmetric bijection, then there exists
a constant $L\geq 1$ depending only on $\mu:=\eta'(1)$
which satisfies the following inequality for all
$n\in\mathbb{Z}$ and $k\in \mathbb{N}$;
\[
\frac{1}{L}< \left|\frac{g(a_{n+k})-g(a_n)}{g(a_n)-g(a_{n-k})}
\right| < L.
\]
\end{deflem}
To prove Lemma \ref{L4.2}, first, we prove the following estimation;
\setcounter{defclaim}{0}
\begin{defclaim} \label{subclaim1}
For any $n\in\mathbb{Z}$, it holds $|g(a_n)-g(a_{n+1})|< 2\mu$.
\end{defclaim}
\begin{proof}
Since $\mu=\eta'(1)\geq 1$, it suffices to consider the case that
$|g(a_n)-g(a_{n+1})|\geq 2$. Then we may assume $g(a_{n+1})>g(a_n)$
since the same argument mentioned below can be applied to the case
$g(a_n)>g(a_{n+1})$.
Letting $m\in \mathbb{Z}_{\leq n}$ satisfy
\[
g(a_m)=\max \left\{ g(a_j);\ j\in \mathbb{Z}_{\leq n}
\ \text{such that}\ g(a_n)\leq g(a_j)< g(a_{n+1})\right\}
\]
and $\ell \in \mathbb{Z}$ satisfy $g(a_{\ell})=g(a_m)+1$ (
then $\ell \geq n+1$ by the construction),
we can construct $m, \ell \in \mathbb{Z}$ which satisfy the following conditions
(see Figure \ref{img2});
\begin{enumerate}
\item $m\leq n$ and $n+1 \leq \ell$,
\item $g(a_n)\leq g(a_m) < g(a_{\ell})=g(a_m)+1 \leq g(a_{n+1})$.
\end{enumerate}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=12cm]{img2.eps}
\end{center}\ \vspace{3ex}
\caption{} \label{img2}
\end{figure}
First, suppose $g(a_m)-g(a_n)\geq \left( g(a_{n+1})-g(a_n)\right)/2(\geq 1)$.
Then $g(a_m),g(a_n),g(a_{\ell})$ are distinct and
\begin{eqnarray*}
\mu &> & \eta'\left(\left|\frac{a_m-a_n}{a_m-a_{\ell}}\right|\right)\\
&\geq & \left|\frac{g(a_m)-g(a_n)}{g(a_m)-g(a_{\ell})}\right|\\
&=& g(a_m)-g(a_n) \geq \frac{g(a_{n+1})-g(a_n)}{2}.
\end{eqnarray*}
Thus we have $g(a_{n+1})-g(a_n)< 2\mu$.
Next, suppose $g(a_m)-g(a_n)< \left( g(a_{n+1})-g(a_n)\right)/2$.
Then $g(a_{n+1})-g(a_m)>(g(a_{n+1})-g(a_n))/2$ holds.
Since $g(a_{n+1})-g(a_n)\geq 2$,
\[
g(a_{n+1})-g(a_{\ell})> \frac{g(a_{n+1})-g(a_n)}{2}-1 \geq 0,
\]
that is, $\ell \neq n+1$. Therefore $g(a_m),g(a_{n+1}),g(a_{\ell})$ are distinct.
Similarly we have $g(a_{n+1})-g(a_n)< 2\mu$.
\end{proof}
\begin{defclaim} \label{subclaim2}
For any $n\in \mathbb{Z}$ and $k\in\mathbb{N}\ (k\neq 1)$, the following
inequality holds;
\[
\frac{k-1}{2\mu}< |g(a_n)-g(a_{n+k})|< 2\mu k.
\]
\end{defclaim}
\begin{proof}
(\textit{Upper bound})
By the triangle inequality, it immediately follows from Claim \ref{subclaim1} that
$|g(a_n)-g(a_{n+k})|< 2\mu k$.\\
(\textit{Lower bound}) Suppose $k\neq 1$.
Since the open interval
\[
\left(g(a_n)-\frac{k-1}{2},\ g(a_n)+\frac{k-1}{2}\right)
\]
contains at most $(k-1)$ integer points,
there exists an integer $m\in \mathbb{Z}\ (n<m<n+k)$ such that
\[
|g(a_n)-g(a_m)|\geq \frac{k-1}{2}.
\]
By the quasisymmetry, we obtain
\begin{eqnarray*}
\mu &>& \eta'\left(\left|\frac{a_n-a_m}{a_n-a_{n+k}}\right|\right)\\
&\geq& \left| \frac{g(a_n)-g(a_m)}{g(a_n)-g(a_{n+k})}\right|
\geq \frac{k-1}{2|g(a_n)-g(a_{n+k})|},
\end{eqnarray*}
that is, $|g(a_n)-g(a_{n+k})|> (k-1)/2\mu$.
\end{proof}
\begin{defclaim} \label{subclaim3}
Lemma \ref{L4.2} holds.
\end{defclaim}
\begin{proof}
If $k\neq 1$, it immediately follows from Claim \ref{subclaim2} that
\[
\frac{1}{L} < \left| \frac{g(a_{n+k})-g(a_n)|}{g(a_n)-g(a_{n-k})}
\right| < L
\]
for $L=8\mu^2$. Moreover, even if $k=1$, it follows from Claim \ref{subclaim1}
\[
8\mu^2>2\mu > \left| \frac{g(a_{n+1})-g(a_n)}{g(a_n)-g(a_{n-1})}
\right| > \frac{1}{2\mu} >\frac{1}{8\mu^2}.
\]
\end{proof}
By the above lemmas, we obtain the following;
\setcounter{alpthm}{2}
\begin{alpthm} \label{T4.1}
For a subset $E\subset \mathbb{R}$, the following conditions
are quantitatively equivalent;
\begin{enumerate}
\item There exists an $\eta-$quasisymmetric bijection
$f:\mathbb{Z}\rightarrow E$.
\item $E$ can be written as a monotone increasing sequence $E=\{a_n\}_{n\in \mathbb{Z}}$
with $a_n\rightarrow\pm \infty\ (n\rightarrow\pm \infty)$, and
there exists a constant $M\geq 1$ such that the following inequality holds
for all $n\in \mathbb{Z}$ and $k\in \mathbb{N}$;
\[
\frac{1}{M} \leq \frac{a_{n+k}-a_n}{a_n-a_{n-k}} \leq M.
\]
\item There exists a $K-$quasiconformal mapping $F:\mathbb{C}\rightarrow\mathbb{C}$,
such that $F(\mathbb{Z})=E$.
\end{enumerate}
\end{alpthm}
\begin{proof}
The equivalence $(2)\Leftrightarrow (3)$ is already confirmed by Theorem \ref{oldthm}
(see \cite[Theorem A]{fujino2}). Further, for the same reason as the proof of
Theorem \ref{T1}, $(3)\Rightarrow (1)$ follows. Thus it suffices to show
$(1)\Rightarrow (2)$.
Let us assume that there exists an $\eta-$quasisymmetric bijection
$f:\mathbb{Z}\rightarrow E$. By Lemma \ref{L4.1}, $E$ can be
written as a monotone increasing sequence
$E=\{a_n\}_{n\in\mathbb{Z}}$ with $a_n\rightarrow \pm \infty$ as $n\rightarrow
\pm \infty$ (recall $E$ must be closed and discrete in $\mathbb{R}$).
Let $g:=f^{-1}$. Then $g$ is $\eta'-$quasisymmetric where
$\eta'(t)=1/\eta^{-1}(1/t)$. By Lemma \ref{L4.2}, there exists a constant
$L\geq 1$ depending only on $\eta'(1)=1/\eta^{-1}(1)$
which satisfies the following inequality
for any $n\in\mathbb{Z}$ and $k\in\mathbb{N}$;
\[
\frac{1}{L}< \left|\frac{g(a_{n+k})-g(a_n)}{g(a_n)-g(a_{n-k})}\right|
< L.
\]
Therefore we obtain
\[
\left|\frac{a_{n+k}-a_n}{a_n-a_{n-k}}\right|
\leq \eta\left( \left|\frac{g(a_{n+k})-g(a_n)}{g(a_n)-g(a_{n-k})}
\right|\right) < \eta(L).
\]
and
\[
\left|\frac{a_{n+k}-a_n}{a_n-a_{n-k}}\right|
\geq \eta\left( \left|\frac{g(a_{n+k})-g(a_n)}{g(a_n)-g(a_{n-k})}
\right|^{-1}\right)^{-1} > \frac{1}{\eta(L)}.
\]
\end{proof}
\medskip
\section{Extensibility of quasisymmetric embeddings}
\label{extension}
We would like to complete this paper, proving the following theorem;
\setcounter{alpthm}{0}
\begin{alpthm} \label{T5.1}
Every $\eta-$quasisymmetric embedding $f:\mathbb{Z}\rightarrow
\mathbb{R}$ admits a $K=K(\eta)-$quasiconformal extension
$\widetilde{f}:\mathbb{C}\rightarrow \mathbb{C}$ where
$K=K(\eta)$ is a constant depending only on $\eta$.
\end{alpthm}
\begin{proof}
Let $f:\mathbb{Z}\rightarrow \mathbb{R}$ be an $\eta-$quasisymmetric
embedding, and let $E:=f(\mathbb{Z})$. Then, by Theorem \ref{T4.1},
there exists a $K'-$quasiconformal mapping $F:\mathbb{C}\rightarrow
\mathbb{C}$ such that $F(\mathbb{Z})=E$, where $K'$ depends only
on $\eta$. Since compositions of quasisymmetric mappings are
also quasisymmetric, $F^{-1}\circ f:\mathbb{Z}\rightarrow\mathbb{Z}$
becomes an $\eta'-$quasisymmetric automorphism where $\eta'$ depends
only on $\eta$. By Theorem \ref{T3.2}, $F^{-1}\circ f$ admits a
$K''-$quasiconformal extension $G:\mathbb{C}\rightarrow \mathbb{C}$,
where $K''$ depends only on $\eta$.
Therefore, we obtain a $K=K'K''-$quasiconformal extension
$\widetilde{f}=F\circ G:\mathbb{C}\rightarrow \mathbb{C}$ of $f$.
The proof is completed.
\end{proof}
\subsection*{Acknowledgements}
I am deeply grateful to Professor Takeo Ohsawa for his guidance and helpful advices.
This research is partially supported by Grant-in-Aid for JSPS Fellow 16J02185.
\bibliographystyle{amsxport}
\begin{bibdiv}
\begin{biblist}
\bib{alestalo1}{article}{
author={Alestalo, P.},
author={V{\"a}is{\"a}l{\"a}, J.},
title={{Uniform domains of higher order. {III}}},
date={1997},
ISSN={0066-1953},
journal={Ann. Acad. Sci. Fenn. Math.},
volume={22},
number={2},
pages={445\ndash 464},
review={\MR{1469802}},
}
\bib{beurling1}{article}{
author={Beurling, A.},
author={Ahlfors, L.~V.},
title={{The boundary correspondence under quasiconformal mappings}},
date={1956},
ISSN={0001-5962},
journal={Acta Math.},
volume={96},
pages={125\ndash 142},
review={\MR{0086869 (19,258c)}},
}
\bib{fujino2}{article}{
author={Fujino, H.},
title={{The existence of quasiconformal homeomorphism between planes
with countable marked points}},
date={2015},
ISSN={0386-5991},
journal={Kodai Math. J.},
volume={38},
number={3},
pages={732\ndash 746},
url={http://dx.doi.org/10.2996/kmj/1446210604},
review={\MR{3417531}},
}
\bib{heinonen1}{book}{
author={Heinonen, J.},
title={{Lectures on analysis on metric spaces}},
series={{Universitext}},
publisher={Springer-Verlag, New York},
date={2001},
ISBN={0-387-95104-0},
url={http://dx.doi.org.ejgw.nul.nagoya-u.ac.jp/10.1007/978-1-4613-0131-8},
review={\MR{1800917}},
}
\bib{trotsenko1}{article}{
author={Trotsenko, D.~A.},
author={V{\"a}is{\"a}l{\"a}, J.},
title={{Upper sets and quasisymmetric maps}},
date={1999},
ISSN={1239-629X},
journal={Ann. Acad. Sci. Fenn. Math.},
volume={24},
number={2},
pages={465\ndash 488},
review={\MR{1724387}},
}
\bib{vaisala4}{incollection}{
author={V{\"a}is{\"a}l{\"a}, J.},
title={{Questions on quasiconformal maps in space}},
booktitle={{Quasiconformal mappings and analysis ({A}nn {A}rbor, {MI},
1995)}},
review={\MR{1488460}},
}
\bib{vellis1}{unpublished}{
author={Vellis, V.},
title={{Quasisymmetric extension on the real line}},
date={2015},
note={arXiv:1509.06638 [math.MG]},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
We consider the coexistence and implications
between periodic objects of maps on the cylinder
$\Omega = \SI\times \I,$ of the form:
\[
\map{F}{\begin{pmatrix} \theta \\ x\end{pmatrix}}[
{\begin{pmatrix} R_\omega(\theta)\\ \zeta(\theta,x)\end{pmatrix}}
],
\]
where $\SI = \R / \Z$, $\I$ is an interval of the real line,
$R_\omega(\theta) = \theta + \omega \pmod{1}$ with
$\omega \in \R \setminus \Q$
and $\zeta(\theta,x) = \zeta_{\theta}(x)$ is continuous on both variables.
The class of all maps of the above type will be denoted by \cSO.
In this setting a very basic and natural question is the following:
\emph{is it true that any map in the class $\cSO$ has an invariant
curve?}
In \cite{FJJK}, the authors created an appropriate topological
framework that allowed them to obtain the following extension of
the Sharkovski\u{\i} Theorem to the class {\cSO}\footnote
As already remarked in \cite{FJJK}, instead of $\SI$ we could take any
compact metric space~$\Theta$ that admits a minimal homeomorphism
$\map{R}{\Theta}$ such that $R^{\ell}$ is minimal
for every $\ell > 1.$
However, for simplicity and clarity we will remain in the class $\cSO.$}.
Let $X$ be a compact metric space.
We recall that a subset $G \subset X$ is \emph{residual} if it
contains the intersection of a countable family of open dense subsets
in $X.$
In what follows, $\map{\pi}{\Omega}[\SI]$ will denote the
standard projection from $\Omega$ to the circle.
Given a set $B \subset \SI,$ for convenience we will use the following
notation:
\[
\setsilift{B} := \pi^{-1}(B) = B \times \I \subset \Omega
\]
In the particular case when $B = \{\theta\},$ instead of
$\setsilift{\{\theta\}}$ we will simply write $\setsilift{\theta}.$
Also, given $A \subset \Omega,$ we will denote by $\setfib{A}{B}$ the set
\[
A \cap \setsilift{B} = \set{(\theta,x) \in \Omega}{\theta \in B \text{ and } (\theta,x) \in A}.
\]
In the particular case when $B = \{\theta\},$ instead of
$\setfib{A}{\theta}$ we will simply write $\setfibth{A}.$
Instead of periodic points we use objects that project over the
whole~$\SI,$ called \emph{strips} in \cite[Definition~3.9]{FJJK}.
A set $B \subset \Omega$ such that $\pi(B) = \SI$
(i.e., $B$ projects on the whole $\SI$) will be called a \emph{circular set}.
\begin{definition}
A \emph{strip in $\Omega$} is a compact circular set $B \subset \Omega$
such that $\setfibth{B}$ is a closed interval (perhaps degenerate to
a point) for every $\theta$ in a residual set of $\SI.$
\end{definition}
Given two strips $A$ and $B,$ we will write $A < B$ and $A \le B$
(\cite[Definition~3.13]{FJJK}) if there exists a residual set
$G \subset \SI,$ such that
for every $(\theta,x) \in \setfib{A}{G}$ and
$(\theta,y) \in \setfib{B}{G}$
it follows that $x < y$ and,
respectively, $x \le y$.
We say that the strips $A$ and $B$ are \emph{ordered}
(respectively \emph{weakly ordered})
if either $A < B$ or $A > B$
(respectively $A \le B$ or $A \ge B$).
\begin{definition}[\protect{\cite[Definition~3.15]{FJJK}}]
A strip $B \subset \Omega$ is called \emph{$n$-periodic} for $F \in \cSO$
if $F^{n}(B) = B$ and the image sets
$B,\ F(B),\ F^{2}(B),\dots, F^{n-1}(B)$
are pairwise disjoint and pairwise ordered
(see Figure~\ref{fig-example-periodicorbits} for examples).
\end{definition}
\begin{figure}[ht]
\begin{center}
\hfill \subfloat[$3.28 x(1-x) + \tfrac{4}{100}\cos(2\pi\theta)$]{\includegraphics[width=0.46\textwidth]{a328-004}}
\hfill \subfloat[$3.85 x(1-x)(1 + \frac{111}{10^5}\cos(2\pi\theta))$]{\includegraphics[width=0.46\textwidth]{3strips}}
\hfill \strut
\end{center}
\caption{In the left picture we show an example two periodic orbit of curves,
and in the second we show a possible example of a three periodic orbit solid strips.}\label{fig-example-periodicorbits}
\end{figure}
To state the main theorem of \cite{FJJK} we need to recall the
\emph{Sharkovski\u{\i} Ordering} (\cite{Shar, Shartrans}).
The \emph{Sharkovski\u{\i} Ordering} is a linear ordering of $\N$
defined as follows:
\begin{align*}
& 3 \gtso{\Sho} 5 \gtso{\Sho} 7 \gtso{\Sho} 9 \gtso{\Sho} \dots \gtso{\Sho} \\
& 2 \cdot 3 \gtso{\Sho} 2 \cdot 5 \gtso{\Sho} 2 \cdot 7 \gtso{\Sho} 2 \cdot 9 \gtso{\Sho} \dots \gtso{\Sho} \\
& 4 \cdot 3 \gtso{\Sho} 4 \cdot 5 \gtso{\Sho} 4 \cdot 7 \gtso{\Sho} 4 \cdot 9 \gtso{\Sho} \dots \gtso{\Sho}\\
& \hspace*{7em} \vdots \\
& 2^n \cdot 3 \gtso{\Sho} 2^n \cdot 5 \gtso{\Sho} 2^n \cdot 7 \gtso{\Sho} 2^n \cdot 9 \gtso{\Sho} \dots \gtso{\Sho} \\
& \hspace*{7em} \vdots \\
& \cdots \gtso{\Sho} 2^n \gtso{\Sho} \dots \gtso{\Sho}
16 \gtso{\Sho} 8 \gtso{\Sho} 4 \gtso{\Sho} 2 \gtso{\Sho} 1.
\end{align*}
In the ordering $\geso{\Sho}$ the least element is 1 and the largest
one is 3. The supremum of the set $\{1,2,4,\dots,2^n,\dots\}$ does not
exist.
\begin{ST}[\cite{FJJK}]\label{SharTheo}
Assume that the map $F \in \cSO$ has a $p$-periodic strip.
Then $F$ has a $q$-periodic strip for every $q \ltso{\Sho} p.$
\end{ST}
In view of this result, the new following natural question
(that is stronger that the previous one) arises:
\emph{Does Theorem~\ref{SharTheo} holds when restricted to curves?}
where a curve is defined as the graph of a continuous map from $\SI$ to $\I$.
More precisely, \emph{is it true that if $F$ has a $q$-periodic
curve and $p\leso{\Sho} q$ then does there exists a $p$-periodic curve of $F$?}
The aim of this paper is to answer both of the above questions in the negative
by constructing a counterexample.
This is done by the following result which is the main result of the paper.
\begin{MainTheorem}\label{MainTh}
There exists a map $T \in \cSO$ with
$f(\theta,\cdot)$ non-increasing for every $\theta \in \SI,$
such that $T$ permutes the upper and lower circles of $\Omega$
(thus having a periodic orbit of period two of curves),
and $T$ does not have any invariant curve.
\end{MainTheorem}
The construction will be done in two steps.
First, in Section~\ref{pseudo-curve}, we construct a strip $A$
which is a pseudo-curve which is not a curve.
This strip is obtained as a \emph{limit} of sets defined inductively
by using of a collection of \emph{winged boxes} $\wbasicbox{i} \subset \Omega.$
Second, we construct a Cauchy sequence $\{T_m\}_{m=0}^{\infty}$ that gives
as a limit the function $T$ from Theorem~\ref{MainTh} having $A$ as invariant set.
To this end, in Section~\ref{FunctionsGi} we define a collection of auxiliary
functions $G_i$ defined on the winged boxes $\wbasicbox{i}$.
Next, in Section~\ref{stratification} we introduce a notion of \emph{depth}
in the set of winged boxes $\wbasicbox{i}$ which defines a convenient
stratification in the set of winged boxes $\wbasicbox{i}.$
In Section~\ref{BoxesintheWings} we study the wings of box
and its interaction with boxes of higher depth.
In Section~\ref{skew-product},
by using the auxiliary functions from Section~\ref{FunctionsGi},
the stratification from Section~\ref{stratification} and the technical results
from Section~\ref{BoxesintheWings}
we construct the Cauchy sequence $\{T_m\}_{m=0}^{\infty} \subset \cSO,$
we define the map $T = \lim_{m\to\infty} T_m,$
and we prove Theorem~\ref{MainTh}.
For clarity, we omit the proofs of all results from Section~\ref{skew-product}.
These proofs will be provided in Sections~\ref{proofofT0mapProperties},
\ref{proofofseqTmProperties} and \ref{proofofdistTmTm-1}.
Section~\ref{definitions} is devoted to introduce the necessary
definitions and, in particular, to introduce the notion of
pseudo-curve and some necessary results on the space of pseudo-curves.
\section{Definitions and preliminary results}\label{definitions}
The main aim of this section is to introduce the definition and
basic results about pseudo-curves.
Given $G \subset \SI$ and a map $\map{\varphi}{G}[\I]$,
$\Graph(\varphi)$ denotes the \emph{graph of $\varphi$}.
Also, given a set $A$ we will denote the closure of $A$
by $\overline{A}$.
\begin{definition}[Pseudo-curve]\label{PCDefinition}
Let $G$ be a residual set of $\SI$ and let $\map{\varphi}{G}[\I]$
be a continuous map from $G$ to $\I.$
The set $\overline{\Graph(\varphi)},$ denoted by $\pc,$
will be called a \emph{pseudo-curve}.
Notice that every pseudo-curve is a compact circular set.
Also, $\pcs$ will denote the class of all pseudo-curves.
\end{definition}
A set $A \subset \Omega$ is \emph{$F$-invariant}
(respectively \emph{strongly $F$-invariant})
if $F(A) \subset A$ (respectively $F(A) = A$).
Observe that if $F \in \cSO,$ every compact
$F$-invariant set is circular.
A closed invariant set is called \emph{minimal}
if it does not contain any proper closed invariant set.
An \emph{arc of a curve} is the graph of a continuous
function from an arc of $\SI$ to $\I$.
The pseudo-curves have the following properties which are easy to prove:
\begin{lemma}\label{PC-properties}
Given a pseudo-curve $\pc \in \pcs$ the following statements hold.
\begin{enumerate}[(a)]
\item $\setfibth{\pc}$ consists of a single point for every $\theta \in G.$
Consequently, \[ \setfib{\pc}{G} = \Graph(\varphi). \]
\item Every circular compact set contained in a pseudo-curve
coincides with the pseudo-curve.
\item $\pc = \overline{\Graph(\varphi\evalat{\widetilde{G}})}$
for every $\widetilde{G} \subset G$ dense in $\SI.$
\item If $\pc$ contains a curve then it is a curve.
\end{enumerate}
\end{lemma}
\begin{proof}
We start by proving (a).
By the definition of a pseudo-curve we have
$\Graph(\varphi) \subset \setfib{\pc}{G}.$
To prove the other inclusion fix $\theta \in G$ and $x \in \I$
such that $(\theta,x) \in \pc.$
Then, there exists a sequence
$
\{(\theta_n,\varphi(\theta_n))\}_{n=1}^{\infty} \subset \Graph(\varphi)
$
such that
$\lim_{n\to\infty}(\theta_n,\varphi(\theta_n)) = (\theta,x).$
The continuity of $\varphi$ in $G$ (and hence in $\theta$)
implies $x = \varphi(\theta)$ and, therefore,
$(\theta,x)\in \Graph(\varphi).$
Now we prove (b).
Assume that $B \subset \pc$ is a circular compact set.
From the assumptions and statement (a) we get
$\setfib{\pc}{G} = \setfib{B}{G}.$
Hence,
\[
\pc = \overline{\Graph(\varphi)} = \overline{\setfib{\pc}{G}} =
\overline{\setfib{B}{G}} \subset B.
\]
Now (d) follows directly from (b) and the fact that a curve
is compact since it is the graph of a continuous function.
Statement (c) also follows from (b) because
$
\overline{\Graph\left(\varphi\evalat{\widetilde{G}}\right)} \subset \pc
$
and
$
\overline{\Graph\left(\varphi\evalat{\widetilde{G}}\right)}
$
is a circular set (since $\widetilde{G}$ is dense in $\SI$).
\end{proof}
We also will be interested in the pseudo-curves as a possible invariant
objects of maps from $\cSO.$ The next lemma studies their properties
in this case.
\begin{lemma}\label{PC-properties-invariant}
Let $F \in \cSO$ and assume that $\pc \in \pcs$
is an $F$-invariant pseudo-curve. Then,
\begin{enumerate}[(a)]
\item $\pc$ is strongly $F$-invariant and minimal.
\item If $\pc$ contains an arc of a curve then it is a curve.
\end{enumerate}
\end{lemma}
\begin{proof}
We start by proving (a).
Let $B \subset \pc$ be a closed invariant set.
We have that $B$ is circular and, by Lemma~\ref{PC-properties}(b),
$B = \pc.$ Hence, $\pc$ is minimal.
On the other hand, $F(\pc) \subset \pc$ implies
$F^{2}(\pc) \subset F(\pc)$ and, hence,
$F(\pc)$ is a compact $F$-invariant set.
Therefore, by the part already proven,
$F(\pc) = \pc.$
Now we prove (b).
Let $S$ be an (open) arc of $\SI$ and let
{\map{\xi}{S}[\I]} be a continuous map such that
$\Graph(\xi) \subset \pc.$
Clearly, there exists $m\in \N$ such that
$\bigcup_{i=0}^m R^i_\omega(S) = \SI.$
Now we set $\xi_0 := \xi$ and, for $i=1,2,\dots,m,$
we define {\map{\xi_i}{R^i_\omega(S)}[\I]} by
\[
\xi_i(\theta) :=
f\left(R^{-1}_\omega(\theta),\xi_{i-1}\left(R^{-1}_\omega(\theta)\right)\right).
\]
The continuity of $f$ implies that every $\xi_i$
is an arc of a curve and
$\Graph(\xi_i) = F(\Graph(\xi_{i-1})).$
Hence,
\[
\bigcup_{i=0}^m \Graph(\xi_i) = \bigcup_{i=0}^m F^i(\Graph(\xi)) \subset \pc
\]
because $\pc$ is $F$-invariant.
In view of Lemma~\ref{PC-properties}(d) we only have to show that
$\bigcup_{i=0}^m \Graph(\xi_i)$
is a curve. We will prove prove this by induction.
Assume that $\emptyset \ne M \varsubsetneq \{0,1,2,\dots,m\}$
verifies that
$S_M:= \bigcup_{i \in M} R^i_\omega(S)$
is an (open) arc of $\SI$ and
$\bigcup_{i \in M} \Graph(\xi_i)$ is an arc of a curve
(initially we can take $M$ to be any unitary subset of $\{0,1,2,\dots,m\}$).
Then, there exists a continuous map {\map{\xi_{_M}}{S_M}[\I]}
such that $\Graph(\xi_{_M}) = \bigcup_{i \in M} \Graph(\xi_i).$
Clearly, there exists $j \in \{0,1,2,\dots,m\}\setminus M$
such that
$
S_{M,j} := S_M \cap R^j_\omega(S) \ne \emptyset.
$
The set $S_{M,j}$ is an open arc of $\SI$ and, by Lemma~\ref{PC-properties}(a),
$\xi_{_M}\evalat{S_{M,j}\cap G} = \xi_j\evalat{S_{M,j}\cap G}$
because
$\Graph(\xi_{_M}), \Graph(\xi_j) \subset \pc.$
Since $S_{M,j}\cap G$ is dense in $S_{M,j},$
given $\theta \in S_{M,j}\setminus G,$
there exists a sequence
$\{\theta_n\}_{n=0}^\infty \subset S_{M,j}\cap G$
converging to $\theta.$
The continuity of $\xi_{_M}$ and $\xi_j$ on $S_{M,j}$ implies that
$
\xi_{_M}(\theta) = \lim_{n\to\infty} \xi_{_M}(\theta_n) =
\lim_{n\to\infty} \xi_j(\theta_n) = \xi_j(\theta).
$
Consequently, $\xi_{_M}\evalat{S_{M,j}} = \xi_j\evalat{S_{M,j}}$
and $\Graph(\xi_{_M}) \cup \Graph(\xi_j)$ is an arc of a curve
(defined on the open arc $S_M \cup R^j_\omega(S)$).
By redefining $M$ as $M \cup \{j\}$ and iterating this procedure until
$M \cup \{j\} = \{0,1,2,\dots,m\}$ we see that the whole
$\bigcup_{i=0}^m \Graph(\xi_i)$ is a curve.
\end{proof}
Next we will introduce and study the space of pseudo-curves.
\begin{definition}\label{PCG-dinfinito}
We define the \emph{space of pseudo-curve generators} as
\[
\C := \set{(\varphi,G)}{\text{$G$ is a residual set in $\SI$ and {\map{\varphi}{G}[\I]} is a continuous map}}.
\]
On $\C$ we also define the \emph{supremum pseudo-metric}
$\map{\dinf}{\C\times\C}[\R^+]$
by:
\[
\dinf\bigl((\varphi,G),(\varphi',G')\bigr) :=
\sup_{\theta\in G\cap G'} \abs{\varphi(\theta)-\varphi'(\theta)}.
\]
Clearly,
$\dinf((\varphi,G),(\varphi',G')) = 0$
if and only if $\varphi\evalat{G\cap G'} = \varphi'\evalat{G\cap G'}$
and, hence, $\dinf$ is a pseudo-metric.
\end{definition}
The next lemma will be useful in using the metric $\dinf.$
\begin{lemma}\label{dinfinito-equiv}
Let $(\varphi,G),(\varphi',G') \in \C.$
Then,
\[
\dinf\bigl((\varphi,G),(\varphi',G')\bigr) =
\sup_{\theta\in \widetilde{G}} \abs{\varphi(\theta)-\varphi'(\theta)}
\]
for every $\widetilde{G} \subset G \cap G'$ dense in $\SI.$
\end{lemma}
\begin{proof}
Set
$
\dist{\infty, \widetilde{G}}\bigl((\varphi,G),(\varphi',G')\bigr) :=
\sup_{\theta\in \widetilde{G}} \abs{\varphi(\theta)-\varphi'(\theta)}.
$
With this notation, we clearly have
$
\dist{\infty, \widetilde{G}}\bigl((\varphi,G),(\varphi',G')\bigr) \le
\dinf\bigl((\varphi,G),(\varphi',G')\bigr).
$
To prove the reverse inequality take
$\theta \in (G \cap G')\setminus \widetilde{G}.$
Since $\widetilde{G}$ is dense in $\SI,$
there exists a sequence
$\{\theta_n\}_{n=0}^\infty \subset \widetilde{G}$
converging to $\theta.$
On the other hand, by definition, the maps $\varphi$ and $\varphi',$
are continuous in $G \cap G'$ (and, hence, in $\theta$).
Consequently,
$
\abs{\varphi(\theta),\varphi'(\theta)} =
\lim_{n\to\infty} \abs{\varphi(\theta_n)-\varphi'(\theta_n)} \le
\dist{\infty, \widetilde{G}}\bigl((\varphi,G),(\varphi',G')\bigr).
$
This ends the proof of the lemma.
\end{proof}
As it is customary we will introduce an equivalent relation
in the space of pseudo-curve generators so that the quotient space
will be a metric space.
\begin{definition}\label{relation}
Two pseudo-curve generators
$(\varphi,G),(\varphi',G')\in \C$ are said to be equivalent,
denoted by $(\varphi,G) \sim (\varphi',G')$
if and only if $\pc = \pc[\varphi',G'].$
Clearly $\sim$ is an equivalence relation in $\C$.
The $\sim$-equivalence class of $(\varphi,G)\in \C$
will be denoted by $[\varphi,G].$
\end{definition}
\begin{remark}\label{equiv-relation}
From Lemma~\ref{PC-properties}(a,c) it follows that
$(\varphi,G) \sim (\varphi',G')$ if and only if
$\varphi\evalat{\widetilde{G}} = \varphi'\evalat{\widetilde{G}}$
for every $\widetilde{G} \subset G \cap G'$ dense in $\SI.$
In particular, by taking $\widetilde{G} = G \cap G',$
we get that $\dinf((\varphi,G),(\varphi',G')) = 0$
if and only if $(\varphi,G) \sim (\varphi',G').$
\end{remark}
\begin{definition}
The space $\C/\kern-3pt\sim$ will be called the
\emph{space of pseudo-curves generator classes}
and denoted by $\SC.$
Also, on $\SC$ we define the \emph{supremum metric}, also denoted
$\map{\dinf}{\SC\times\SC}[\R^+]$
by abuse of notation,
in the following way.
Given $A = [\varphi_A,G_A], B=[\varphi_B,G_B] \in \SC$ we set
\[
\dinf(A,B) := \dinf\bigl((\varphi_A,G_A),(\varphi_B,G_B)\bigr).
\]
Note that $\dinf$ is well defined. To see this take
$[\varphi_A,G_A]=[\varphi_{'A},G_{A'}], [\varphi_B,G_B] \in \C.$
Then, by Lemma~\ref{dinfinito-equiv} and Remark~\ref{equiv-relation}
applied to $\widetilde{G} = G_A \cap G_{A'} \cap G_B$
we get
$
\dinf\bigl((\varphi_A,G_A),(\varphi_B,G_B)\bigr) =
\dinf\bigl((\varphi_{A'},G_{A'}),(\varphi_B,G_B)\bigr).
$
\end{definition}
The next result establishes the basic properties of the
space of pseudo-curves generator classes $(\SC, \dinf).$
\begin{proposition}\label{Ccompleto}
The space of pseudo-curves generator classes $\SC$ is a complete metric space.
\end{proposition}
\begin{proof}
The fact that $\dinf$ is a metric in $\SC$
follows from Remark~\ref{equiv-relation}.
Now we prove that $\SC$ is complete.
Assume that $\{[\varphi_n,G_n]\}_{n=1}^{\infty}$
is a Cauchy sequence in $\SC.$
We have to see that $\lim_{n\to\infty} [\varphi_n,G_n] \in \SC.$
Set, $G := \cap_{i=1}^{\infty} G_n.$
Since this intersection is countable, $G$ is still a residual set.
The definition of $\dinf$ implies that the sequence
$\{\varphi_n(\theta)\}_{n=1}^{\infty} \subset \I$
is a Cauchy sequence in $\I$ for every $\theta\in G.$
So, it is convergent and we can define a map $\map{\varphi}{G}[\I]$
by $\varphi(\theta) := \lim_{n\to\infty} \varphi_n(\theta).$
If $(\varphi,G)\in \C$ we have $[\varphi,G] \in \SC$ and,
from the definition of $\varphi$ it follows that
\[
\lim_{n\to\infty} \dinf([\varphi,G], [\varphi_n,G_n]) =
\leftlimits{\sup}{\theta\in G\cap G_n} \lim_{n\to\infty} \abs{\varphi(\theta)-\varphi_n(\theta)} = 0.
\]
Consequently, $[\varphi,G] = \lim_{n\to\infty} [\varphi_n,G_n].$
Since $\varphi$ is the uniform limit of a sequence of continuous functions on $G,$
it is continuous on $G.$ That is, $(\varphi,G)\in \C.$
\end{proof}
In what follows we want to look at the space $\pcs$ as a metric space and
relate this metric space with $(\SC, \dinf).$
Let $\rho$ denote the euclidean metric in $\Omega.$
Then, the space $(\Omega, \rho)$ is a compact metric space.
We recall that the \emph{Hausdorff metric} is defined
in the space of compact subsets of $(\Omega, \rho),$ by
\[
H_{\rho}(\A,\mathsf{B}) = \max\left\
\leftlimits{\max}{(\theta,x) \in \A} \rho((\theta,x),\mathsf{B}),
\leftlimits{\max}{(\theta,x) \in \mathsf{B}} \rho((\theta,x),\A)
\right\}.
\]
Then, $(\pcs, H_{\rho})$ is a metric space. To study the relation between
$(\SC, \dinf)$ and $(\pcs, H_{\rho})$ we need a couple of simple technical
results.
\begin{lemma}\label{HdimCCS}
Let $\A,\mathsf{B} \subset \Omega$ be compact circular sets.
Then,
\[
H_{\rho}(\A,\mathsf{B}) \le
\max_{\theta \in \SI} H_{\rho}\bigr(\setfibth{\A},\setfibth{\mathsf{B}}\bigl).
\]
\end{lemma}
\begin{proof}
It follows directly from the definitions:
\begin{align*}
H_{\rho}\left(\A, \mathsf{B}\right)
&\le \max\left\{
\leftlimits{\sup}{(\theta,x) \in \A} \rho\Bigl((\theta,x), \setfibth{\mathsf{B}}\Bigr),
\leftlimits{\sup}{(\theta,x) \in \mathsf{B}} \rho\Bigl((\theta,x), \setfibth{\A}\Bigr)
\right\}\\
&= \max\left\{
\sup_{\theta \in \SI}\leftlimits{\max}{\set{x\in \I}{(\theta,x) \in \A}} \rho\Bigl((\theta,x), \setfibth{\mathsf{B}}\Bigr),\right.\\
&\hspace*{15.5em} \left.\sup_{\theta \in \SI}\leftlimits{\max}{\set{x\in \I}{(\theta,x) \in \mathsf{B}}} \rho\Bigl((\theta,x), \setfibth{\A}\Bigr)
\right\}\\
&= \sup_{\theta \in \SI} \max\left\{
\leftlimits{\max}{\set{x\in \I}{(\theta,x) \in \A}} \rho\Bigl((\theta,x), \setfibth{\mathsf{B}}\Bigr),
\leftlimits{\max}{\set{x\in \I}{(\theta,x) \in \mathsf{B}}} \rho\Bigl((\theta,x), \setfibth{\A}\Bigr)
\right\}\\
&= \sup_{\theta \in \SI} H_{\rho}\left(\setfibth{\A}, \setfibth{\mathsf{B}}\right).
\end{align*}
\end{proof}
\begin{proposition}\label{Hdimdinfty}
Let $(\varphi,G), (\widetilde{\varphi},\widetilde{G}) \in \C.$
Then,
\[
H_{\rho}\left(\pc, \pc[\widetilde{\varphi},\widetilde{G}]\right) \le
\sup_{\theta \in \SI} H_{\rho}\left(
\setfibth{\pc},
\setfibth{\pc[\widetilde{\varphi},\widetilde{G}]}
\right) =
\dinf\bigl((\varphi,G),(\widetilde{\varphi},\widetilde{G})\bigr).
\]
\end{proposition}
\begin{proof}
The first inequality follows from Lemma~\ref{HdimCCS}.
Now we prove the second equality.
By Lemma~\ref{PC-properties}(a),
\[
\dinf\bigl((\varphi,G),(\widetilde{\varphi},\widetilde{G})\bigr)
= \leftlimits{\sup}{\theta \in G \cap \widetilde{G}} \abs{\varphi(\theta) - \widetilde{\varphi}(\theta)}
= \leftlimits{\sup}{\theta \in G \cap \widetilde{G}} H_{\rho}\left(
\setfibth{\pc},
\setfibth{\pc[\widetilde{\varphi},\widetilde{G}]}
\right).
\]
So, to end the proof of the lemma, we have to see that
\[
H_{\rho}\left(
\setfibth{\pc},
\setfibth{\pc[\widetilde{\varphi},\widetilde{G}]}
\right) \le
\dinf\bigl((\varphi,G),(\widetilde{\varphi},\widetilde{G})\bigr)
\andq[for every]
\theta \in \SI \setminus (G \cap \widetilde{G}).
\]
Fix $\theta \in \SI \setminus (G \cap \widetilde{G}).$
From the definition of the Hausdorff metric it follows that there exist
$x,y \in \I$ such that
$
H_{\rho}\left(
\setfibth{\pc},
\setfibth{\pc[\widetilde{\varphi},\widetilde{G}]}
\right) = \abs{x-y},
$
$(\theta,x) \in \setfibth{\pc},$ and
$(\theta,y) \in \setfibth{\pc[\widetilde{\varphi},\widetilde{G}]}.$
Since $G \cap \widetilde{G}$ is residual (and thus dense) in $\SI,$
from Lemma~\ref{PC-properties}(a,c) it follows that there exists sequences
$
\{(\theta_n,\varphi(\theta_n))\}_{n=0}^\infty,\
\{(\theta_n,\widetilde{\varphi}(\theta_n))\}_{n=0}^\infty \subset \setsilift{(G \cap \widetilde{G})}
$
such that
$\lim_{n\to\infty} (\theta_n,\varphi(\theta_n)) = (\theta,x)$ and
$\lim_{n\to\infty} (\theta_n,\widetilde{\varphi}(\theta_n)) = (\theta,y).$
Hence,
\[
H_{\rho}\left(
\setfibth{\pc},
\setfibth{\pc[\widetilde{\varphi},\widetilde{G}]}
\right)
= \abs{x-y}
= \lim_{n\to\infty} \abs{\varphi(\theta_n) - \widetilde{\varphi}(\theta_n)}
\le \dinf\bigl((\varphi,G),(\widetilde{\varphi},\widetilde{G})\bigr).
\]
\end{proof}
Proposition~\ref{Hdimdinfty} tells us that that if
$\{[\varphi_n,G_n]\}_{n=1}^{\infty}$ is a Cauchy sequence in $\SC$
then $\pc[\varphi_n,G_n]$ is a Cauchy sequence in $(\pcs, H_{\rho}),$
and if $[\varphi,G] = \lim_{n\to\infty} [\varphi_n,G_n]$ then
$\pc = \lim_{n\to\infty} \pc[\varphi_n,G_n].$
Unfortunately the space $(\pcs, H_{\rho})$ is not complete as the
following simple example shows.
\begin{example}[The space $(\pcs, H_{\rho})$ is not complete]
Consider continuous maps {\map{\xi_n}{\SI}[\I]} with $n\in \N,\ n \ge 2,$
defined by
\[
\xi_n(\theta) = \begin{cases}
2n\theta & \text{if $\theta \in [0,\tfrac{1}{2n}]$,}\\
2(1-n\theta) & \text{if $\theta \in [\tfrac{1}{2n}, \tfrac{1}{n}]$,}\\
0 & \text{if $\theta \ge \tfrac{1}{n}$.}
\end{cases}
\]
Clearly, $(\xi_n,\SI) \in \C$ and
$
H_{\rho}(\pc[\xi_n,\SI], \pc[\xi_m,\SI]) \le \tfrac{1}{\min\{n,m\}}.
$
Hence, $\{\pc[\xi_n,\SI]\}$ is a Cauchy sequence in $\pcs.$
However, the sequence $\{\pc[\xi_n,\SI]\}$ has no limit in $\pcs.$
Indeed,
$
\lim_{n\to\infty} \pc[\xi_n,\SI] = \mathsf{L} =
(\SI \times \{0\}) \cup (\{0\} \times [0,1]),
$
which is not the closure of the graph of a continuous map on a
residual set of $\SI$ (in other words, $\mathsf{L} \notin \pcs$).
This is consistent with the fact that, clearly, $\{[\xi_n,\SI]\}$
is not a Cauchy sequence in $(\SC, \dinf).$
\end{example}
\section{Construction of a connected pseudo-curve}\label{pseudo-curve}
The aim of this subsection is to construct a strip $\A = \pc[\gamma,G]$
as a connected pseudo-curve with certain topological
properties that will allow us to define the map $T \in \cSO$ having this
pseudo-curve as the only proper invariant object.
The pseudo-curve $\pc[\gamma,G]$ will be obtained as a limit in $\SC$
of a sequence of pseudo-curves that will be constructed recursively.
We will start by introducing the necessary notation.
In what follows, for simplicity, we will take the interval $\I$ as the
interval $[-2,2].$ Also, fix $\omega \in [0,1]\setminus\Q.$
For any $\ell \in \Z$ set
$\lstar = \ell\omega \pmod{1}$ and
$\Orbom = \set{\lstar}{\ell\in\Z}.$
That is, $\Orbom$ is the orbit of $0$ by the rotation of angle $\omega.$
We will denote by $\dSI$ the arc distance on $\SI = \R/\Z$.
That is, for $\theta_1, \theta_2 \in \SI,$ we set
\[
\dSI(\theta_1, \theta_2) := \begin{cases}
\theta_2 - \theta_1 & \text{when $\theta_1 \le \theta_2$, and}\\
(\theta_2 + 1) - \theta_1 & \text{when $\theta_1 > \theta_2$.}
\end{cases}
\]
The closed arc of $\SI$ joining $\theta_1$ and $\theta_2$ in the natural direction
will be denoted by $[\theta_1, \theta_2].$
That is,
\[
[\theta_1, \theta_2] = \begin{cases}
\set{t \pmod{1}}{\theta_1 \le t \le \theta_2}
& \text{when $\theta_1 \le \theta_2$, and}\\
\set{t \pmod{1}}{\theta_1 \le t \le \theta_2+1}
& \text{when $\theta_1 > \theta_2$.}
\end{cases}
\]
The open arc of $\SI$ joining $\theta_1$ and $\theta_2$
will be denoted by
$(\theta_1, \theta_2) = [\theta_1, \theta_2] \setminus \{\theta_1, \theta_2\},$
and is defined analogously with strict inequalities
Given an arc $B \subset \SI$, $\Bd(B)$ will denote the set of endpoints of $B.$
We will denote the open (respectively closed) ball (in $\SI$)
of radius $\delta$ centred at $\theta \in \SI$ by
$\ball{\theta}{\delta}$ (respectively $\cball{\theta}{\delta}$):
\begin{align*}
\ball{\theta}{\delta} &=
\set{\widetilde{\theta} \in \SI}{\dSI(\theta, \widetilde{\theta}) < \delta}
= (\theta - \delta \pmod{1}, \theta + \delta \pmod{1}),\text{ and}\\
\cball{\theta}{\delta} &=
\overline{\ball{\theta}{\delta}} =
\set{\widetilde{\theta} \in \SI}{\dSI(\theta, \widetilde{\theta}) \le \delta}
= [\theta - \delta \pmod{1}, \theta + \delta \pmod{1}].
\end{align*}
We consider the space $\Omega$ endowed the metric induced by the maximum of $\dSI$
and the absolute value on $\I$. That is, given $(\theta,x), (\nu, y) \in \Omega$
we set
\[
\dom((\theta,x), (\nu, y)) := \max\left\{\dSI(\theta,\nu), \abs{x-y}\right\}.
\]
Then, given $A \subset \Omega$ we will denote the
\emph{interior of $A$} by $\Int(A)$ and
$\diam(A)$ will denote the \emph{diameter of $A$} whenever $A$ is compact.
To define the sequence of pseudo-curves that will converge to
$\pc[\gamma,G]$ we first need to construct an auxiliary family
$\{\basicbox{\ell}\}_{\ell\in\Z}$ of compact regions in $\Omega$
and a family of compact sets
$\{\Gamma\varphi_{_{\lstar}}\}_{\ell\in\Z}$
such that, for every $\ell \in \Z,$
$
\Gamma\varphi_{_{\lstar}} \subset \basicbox{\ell}
$
and it is the restriction of a pseudo-curve generator to $\pi(\basicbox{\ell}).$
To do this we define the auxiliary functions
$\map{\beta}{[-1,1]}$
and $\map{\phi}{[-1,1]\setminus \{0\}}[{[-1,1]}]$
by (see Figure~\ref{fig1}):
\begin{figure}[tb]
\begin{tikzpicture}[scale=2.5, domain=-1.2:1.2]
\draw[thick] (-1,0) -- (0,1) -- (1,0) -- (0,-1) -- (-1,0);
\draw (-1.2,0) -- (1.2,0);
\draw (0,-1.2) -- (0,1.2);
\node at (0,1) [right] {$1$}; \node at (0,-1) [right] {$-1$};
\node at (0.95,0) [anchor=north west] {$1$}; \node at (-0.95,0) [anchor=north east] {$-1$};
\node (phi) at (1,-0.6) {$\phi(\theta)$}; \draw[->, >=stealth, dashed] (phi) -- (0.33,-0.1);
\draw[color=red, dashed, domain=-1:1] plot (\x,{(1-abs(\x))^2});
\node (beta) at (-1,0.7) {$\beta(\theta)$};
\draw[->, >=stealth, dashed] (beta) -- (-0.26,0.75);
\draw[->, >=stealth, dashed] (beta) .. controls +(1,1) and (0.66,1.45) .. (0.26,0.75);
\draw[color=red, dashed, domain=-1:1] plot (\x,{-(1-abs(\x))^2});
\node (mbeta) at (-1,-0.7) {$-\beta(\theta)$};
\draw[->, >=stealth, dashed] (mbeta) -- (-0.26,-0.75);
\draw[->, >=stealth, dashed] (mbeta) .. controls +(1,-1) and (0.66,-1.45) .. (0.26,-0.75);
\color{blue}
\draw plot file {PseudoCurve_AlsManMor-base.table};
\draw[xscale=-1, yscale=-1] plot file {PseudoCurve_AlsManMor-base.table};
\end{tikzpicture}
\caption{The graphs of the functions $\phi$ (in \textcolor{blue}{blue}) and $\pm\beta$ in thick black.
The \textcolor{red}{red} dashed curve is $(1-\abs{x})^2.$}\label{fig1}
\end{figure}
\[
\beta(x) := 1 - \abs{x}
\qquad\text{and}\qquad
\phi(x) := (1-\abs{x})^2 \sin\left(\frac{\pi}{x}\right).
\]
Note that $-\beta(x) < \phi(x) < \beta(x),$
for all $x\in [-1,1]\setminus\{0\}$ and the graphs of $-\beta$ and $\beta$
intersect the closure of the graph of $\phi$ only at the points
$(0,-1),\ (0,1), (-1,0)$ and $(1,0).$
To define the families $\{\basicbox{\ell}\}_{\ell\in\Z}$
and $\{\Gamma\varphi_{_{\lstar}}\}_{\ell\in\Z}$
we use the following \emph{generic boxes}.
For every $\theta\in\SI$ and $\delta < \tfrac{1}{2},$
$\map{\vartheta_{_\theta}}{[-\delta,\delta]}[\SI]$
denotes the map defined by $\vartheta_{_\theta}(x) = x + \theta \pmod{1}.$
Clearly $\vartheta_{\theta}$ is a homeomorphism between
$[-\delta,\delta]$ and $\cball{\theta}{\delta}.$
Finally
$
\map{\vartheta^{-1}_{\theta}}{\cball{\theta}{\delta}}[{[-\delta,\delta]}]
$
denotes the inverse homeomorphism of $\vartheta_{\theta}.$
\begin{definition}[Generic boxes]\label{GenericBoxes}
Fix $\ell,n\in\Z,\ n \ge \all,\ \alpha \in (0,2^{-n}),\ \delta \in (0, \alpha),$
$a\in[-1,1]$ and $a^+,a^- \in \ball{2^{-n}\beta(\delta)}{a}$
(see Figure~\ref{fig-boxes}).
Now we consider the Jordan closed curve in $\Omega,$
formed by the graphs of the functions
\[
a + 2^{-n}(\beta\circ\vartheta^{-1}_{_{\lstar}}) \evalat{\cball{\lstar}{\delta}}
\andq
a- 2^{-n}(\beta\circ\vartheta^{-1}_{_{\lstar}}) \evalat{\cball{\lstar}{\delta}},
\]
together with the four segments that join the points:
\begin{align*}
& \text{$(\lstar -\alpha,a^-)$ with $\left(\lstar-\delta,a-2^{-n}\beta(-\delta)\right)$},\\
& \text{$(\lstar -\alpha,a^-)$ with $\left(\lstar-\delta,a + 2^{-n}\beta(-\delta)\right)$},\\
& \text{$(\lstar + \alpha,a^+)$ with $\left(\lstar + \delta,a-2^{-n}\beta(\delta)\right)$, and}\\
& \text{$(\lstar + \alpha,a^+)$ with $\left(\lstar + \delta,a + 2^{-n}\beta(\delta)\right)$}.
\end{align*}
We denote the closure of the connected component of the complement of the above
Jordan curve in $\Omega$ that contains the point $(\lstar,a)$
by $\win{\lstar,n,\alpha,\delta,a,a^+,a^-}$
(the coloured region in Figure~\ref{fig-boxes}).
Observe that
$\pi\left(\win{\lstar,n,\alpha,\delta,a,a^+,a^-}\right),$
the projection of $\win{\lstar,n,\alpha,\delta,a,a^+,a^-}$ to $\SI,$
is $\cball{\lstar}{\alpha} = [\lstar-\alpha, \lstar+\alpha].$
\begin{figure}
\begin{tikzpicture}[scale=2.5]
\def\CentRegBound{0.4}
\filldraw[draw=Bittersweet, fill=Bittersweet!50] (-1,-0.25) -- (-\CentRegBound, 1-\CentRegBound) -- (0,1) --
(\CentRegBound, 1-\CentRegBound) -- (1, 0.15) -- (\CentRegBound, \CentRegBound-1) -- (0,-1) --
(-\CentRegBound, \CentRegBound-1) -- (-1,-0.25);
\draw (-1,-1) -- (-1,1) -- (1,1) -- (1,-1) -- (-1,-1);
\foreach \x in {-\CentRegBound,0,\CentRegBound} { \draw[dashed, thin] (\x,-1) -- (\x,1); }
\node at (-1,-1) [anchor=north] {$\lstar-\alpha$};
\node at (-\CentRegBound,-1) [anchor=north] {$\lstar-\delta$};
\node at (0,-1) [anchor=north] {$\lstar$};
\node at (\CentRegBound,-1) [anchor=north] {$\lstar+\delta$};
\node at (1,-1) [anchor=north] {$\lstar+\alpha$};
\draw[dashed, thin] (-1,0) -- (1,0); \node at (-1,0) [left] {$a$};
\node (beta) at (-1,1.2) {\footnotesize $a+\tfrac{1}{2^n}(\beta\circ\vartheta^{-1}_{_{\lstar}})(\theta)$} ;
\draw[->, dashed, >=stealth] (-1,1.15) -- (-0.26,0.75);
\draw[->, dashed, >=stealth] (beta) .. controls +(0.7,0.5) and (0.56,1.25) .. (0.26,0.75);
\node (mbeta) at (-0.87,-1.4) {\footnotesize $a-\tfrac{1}{2^n}(\beta\circ \vartheta^{-1}_{_{\lstar}})(\theta)$};
\draw[->, dashed, >=stealth] (-0.92,-1.32) -- (-0.26,-0.75);
\draw[->, dashed, >=stealth] (mbeta) .. controls +(1.6,-0.7) and (0.9,-0.8) .. (0.35,-0.65);
\filldraw[blue] (-1,-0.25) circle(0.6pt) (1,0.15) circle(0.6pt);
\node at (-1,-0.25) [left] {$a^-$};
\node at (1,0.15) [right] {$a^+$};
\node (graph) at (0.6,-0.65) [anchor=west] {$\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}$};
\draw[->, >=stealth, dashed] (graph.north west)+(0.03,-0.03) -- (0.33,-0.17383);
\color{blue}
\begin{scope} \clip (-\CentRegBound,-1) rectangle (\CentRegBound,1);
\draw plot file {PseudoCurve_AlsManMor-base.table};
\draw[xscale=-1, yscale=-1] plot file {PseudoCurve_AlsManMor-base.table};
\end{scope}
\draw (-1,-0.25) -- (-\CentRegBound, -0.35006);
\draw (1,0.15) -- (\CentRegBound, 0.35006);
\end{tikzpicture}\vspace*{-8ex}
\caption{The region $\win{\lstar,n,\alpha,\delta,a,a^+,a^-}$
is the \textcolor{Bittersweet}{colored filled area}, delimited in the rectangle
$\setsilift{\cball{\lstar}{\delta}}$
by the graphs of the functions $a\pm\tfrac{1}{2^n}(\beta\circ\vartheta^{-1}_{_{\lstar}})(\theta).$
In \textcolor{blue}{blue} the set $\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}$
inductively defining the pseudo-curve.}\label{fig-boxes}
\end{figure}
We denote by
\[
\map{ \varphi_{_{\lstar}} = \varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}}{
\cball{\lstar}{\alpha}\setminus\lstarset}[\I]
\]
the continuous map defined as follows:
\begin{enumerate}[(i)]
\item
$
\varphi_{_{\lstar}}\evalat{\cball{\lstar}{\delta}\setminus\lstarset} =
a + (-1)^{\ell} 2^{-n}(\phi \circ \vartheta^{-1}_{_{\lstar}}).
$
\item
$
\varphi_{_{\lstar}}(\lstar-\alpha) = a^-
$
and
$
\varphi_{_{\lstar}}(\lstar + \alpha) = a^+.
$
\item
$
\varphi_{_{\lstar}}\evalat{[\lstar-\alpha,\lstar-\delta]}
$
and
$
\varphi_{_{\lstar}}\evalat{[\lstar+\delta, \lstar+\alpha]}
$
are affine.
\end{enumerate}
We also denote by
$
\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}\subset
\win{\lstar,n,\alpha,\delta,a,a^+,a^-}
$
the closure in $\Omega$ of the graph of
$\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}.$
\end{definition}
\begin{remark}\label{propiedadesR}
The region
$
\win{\lstar,n,\alpha,\delta,a,a^+,a^-}
$
and the set
$
\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}
$
satisfy the following properties:
\begin{enumerate}[(1)]
\item
$
\win{\lstar,n,\alpha,\delta,a,a^+,a^-} \subset
\cball{\lstar}{\alpha}\times[a-2^{-n}, a+2^{-n}].
$
\item
$
\diam(\win{\lstar,n,\alpha,\delta,a,a^+,a^-}) =
\diam(\setfibpt{\win{\lstar,n,\alpha,\delta,a,a^+,a^-}}{\lstar}) =
2 \cdot 2^{-n}.
$
\item
The sets $\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}$ and
$\partial\win{\lstar,n,\alpha,\delta,a,a^+,a^-}$
only intersect at the points $(\lstar, a-2^{-n}),(\lstar, a+2^{-n}),$
$(\lstar-\alpha,a^-)$ and $(\lstar + \alpha, a^+).$
\item $
\setfibpt{\left(\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}}\right)}{\lstar}
= \win{\lstar,n,\alpha,\delta,a,a^+,a^-}^{\lstar}
$
is an interval.
\item Let $\win{\lstar,n,\alpha,\delta,a,a^+,a^-}$ and
$\win{\kstar,\widetilde{n},\widetilde{\alpha},\widetilde{\delta},\widetilde{a},\widetilde{a}^+,\widetilde{a}^-}$
be two regions, then
$\cball{\lstar}{\alpha} \cap \cball{\kstar}{\widetilde{\alpha}} = \emptyset$
implies
\[\win{\lstar,n,\alpha,\delta,a,a^+,a^-} \cap \win{\kstar,\widetilde{n},\widetilde{\alpha},\widetilde{\delta},\widetilde{a},\widetilde{a}^+,\widetilde{a}^-} = \emptyset.\]
\end{enumerate}
\end{remark}
For every $j\in \Z^+,$ we set
\begin{align*}
Z_{j} &:= \set{i\in\Z}{\ai \le j\} = \{-j,-j + 1,\ldots,-1,0,1,\ldots,j-1,j}\text{ and}\\
\Zstar_{j} & := \set{\istar}{i\in Z_{j}}.
\end{align*}
With the help of the sets
$\win{\lstar,n,\alpha,\delta,a,a^+,a^-}$ and
$\Gamma\varphi_{_{(\lstar,n,\alpha,\delta,a,a^+,a^-)}},$
which are the ``bricks'' of our construction we are ready to define
the sequence of pseudo-curve generators
$\{(\gams{j},\SI\setminus \Zstar_{j})\}_{j=0}^{\infty}$
that we are looking for.
To do this, for every $j \ge 0$ we define
\begin{itemize}
\item a strictly increasing sequence $\{n_j\}_{j=0}^\infty \subset \N,$
\item a strictly decreasing sequence $\{\alpha_j\}_{j=0}^\infty$ such that $2^{-n_{j+1}} < \alpha_j < 2^{-n_{j}}$
\item and a sequence $\{\delta_j\}_{j=0}^\infty$ with $2^{-n_{j+1}} < \delta_{j} < \alpha_j$
\end{itemize}
verifying some technical properties that we will make explicit below,
and we define a sequence of boxes
$\basicbox{j} := \win{\jstar,n_{j},\alpha_{j},\delta_{j},a_j,a_j^+,a_j^-}$ and
$\basicbox{-j} := \win{\sstar{-j},n_{j},\alpha_{j},\delta_{j},a_{-j},a_{-j}^+,a_{-j}^-}$
(for $j=0$ both sets coincide) with projections
\[
\pi\left(\basicbox{j}\right) = \basint{j}
\andq
\pi\left(\basicbox{-j}\right) = \basintneg{j}.
\]
Finally, with the use of all these sequences and objects we can define
our functions $\gams{j}\evalat{\SI\setminus \Zstar_{j}}.$
Observe that we are using the intervals of the form
$\basintabs{\ell},$ $\basintabs[\delta]{\ell}$ and also $\BSG{\ell}{\all-1}$
when $\ell$ is negative.
To ease the use of these intervals we introduce the following notation:
\[
\wbasint{\ell} := \begin{cases}
\basint{\ell} & \text{if $\ell \ge 0$, or}\\
\BSG{\ell}{\abs{\ell+1}} & \text{if $\ell < 0$,}
\end{cases}
\andq
\wobasint{\ell} := \begin{cases}
\obasint{\ell} & \text{if $\ell \ge 0$, or}\\
\OBG{\ell}{\abs{\ell+1}} & \text{if $\ell < 0$.}
\end{cases}
\]
Notice that the ball $\wbasint{\ell}$ has diameter $\alpha_j$ for
$\ell \in \{j, -(j+1)\}.$
\begin{remark}\label{Rotation-intervals-formulas}
With the above notation $\basintabs{\ell} \varsubsetneq \wobasint{\ell}$ for every $\ell < 0.$
Moreover, for $\ell \in \Z$ and $j \in \Z^+,$
\begin{align*}
R_\omega\left(\BSG{\ell}{j}\right) &= \BSG{\ell+1}{j},\text{ and}\\
R_\omega\left(\wbasint{\ell}\right) &= \begin{cases}
\BSG{\ell+1}{\ell} & \text{if $\ell \ge 0$, or}\\
\basintabs{\ell+1} & \text{if $\ell < 0$.}
\end{cases}
\end{align*}
Also, the same formulae holds with $\alpha$ replaced by $\delta$
and for open balls.
\end{remark}
The next crucial definition fixes in detail all quantities and objects mentioned above.
\begin{definition}\label{PCgenerators}
We start by defining
$\basicbox{0} := \win{\sstar{0},n_{0},\alpha_{0},\delta_{0},0,0,0}$ and
$\varphi_{_{\sstar{0}}} := \varphi_{_{(\sstar{0},n_0,\alpha_0,\delta_0,0,0,0)}}$
by choosing (Definition~\ref{GenericBoxes})
$n_{0}=1,$ $\alpha_{0} < \frac{1}{2} = 2 ^{-n_0}$ and $\delta_0 < \alpha_0$
small enough so that the intervals
$\wbasint{0} = \basint{0},$ $\BSG{1}{0}$ and $\wbasint{-1} = \BSG{-1}{0}$ are pairwise disjoint;
and $\sstar{-2}, \sstar{2} \notin \wbasint{-1}$
and, additionally, $\Bd\left(\basint{0}\right) \cap \Orbom = \emptyset.$
We also set $a_{0}^+ = a_{0}^- = a_{0} = 0,$
and we define the map $\map{\gams{0}}{\SI\setminus \{0\}}[\I]$ by
\[
\gams{0}(\theta) =
\begin{cases}
\varphi_{_{\sstar{0}}}(\theta) & \text{if $\theta \in \basint{0}\setminus\{0\}$,}\\
0 & \text{if $\theta \notin \basint{0}$.}
\end{cases}
\]
For consistency with the definition of $\gams{j}$ in the case $j \ge 1,$ we define
the map $\map{\gams{-1}}{\SI\setminus \{0\}}[\I]$ by
$\gams{-1}(\theta) = 0$ for every $\theta \in \SI.$
Then, notice that, $a_{0} = \gams{-1}(\sstar{0}),$
$
a^{\pm}_{0} = \varphi_{_{\sstar{0}}}(\sstar{0}\pm\alpha_{0}) = \gams{-1}(\sstar{0}\pm\alpha_{0}),
$
and $\gams{0}(\theta) = \gams{-1}(\theta)$ for every $\theta \notin \basint{0}.$
Next, for every $j\in \N$ we define $\basicbox{j},$ $\basicbox{-j}$
and $(\gams{j},\SI\setminus \Zstar_{j})$
from the corresponding boxes $\basicbox{i}$ and $\basintabs{i} \subset \wbasint{i}$
for $i \in Z_{j-1},$ and $(\gams{j-1},\SI\setminus \Zstar_{j-1})$ as follows.
We take $n_j,$ $\delta_{j}$ and $\alpha_{j}$ such that
(see Figure~\ref{IterativeBoxes} to fix ideas):
\begin{enumerate}[({\tsfR.}1)]
\item $n_j > n_{j-1}$, $\delta_{j} < \alpha_j < 2^{-n_j} < \delta_{j-1} < \alpha_{j-1}$ and
\[
\Bigl( \Bd\left(\basintneg{j}\right) \cup \Bd\left(\basint{j}\right) \Bigr)
\cap \Orbom = \emptyset.
\]
\item The intervals
\begin{align*}
&\wbasint{j} = \basint{j},\\
&R_\omega\left(\basint{j}\right) = \BSG{j+1}{j},\\
&\wbasint{-j} = \BSG{-j}{j-1}\text{ and}\\
&\wbasint{-(j+1)} = \BSG{-(j+1)}{j}
\end{align*}
are pairwise disjoint,
\[
\gams{j-1} \left(\BSG{\ell}{j}\right) \subset \left[
\gams{j-1}(\lstar) - 2^{-n_{j}},
\gams{j-1}(\lstar) + 2^{-n_{j}}
\right]
\]
for every $\ell \in \{j+1, -(j+1)\},$
\begin{align*}
& \wbasint{\ell} \cap \Zstar_{j+1} = \lstarset
\text{ for }
\ell \in \{j, -(j+1)\} \text{ and}\\
& \BSG{j+1}{j} \cap \Zstar_{j+1} = \sstarset{j+1},
\end{align*}
and $\sstar{-(j + 2)}, \sstar{j+2} \notin \wbasint{-(j+1)} = \BSG{-(j+1)}{j}.$
\item $\Bd\left(\BSG{k+1}{\ak}\right) \cap \left(\basint{j}\cup \basintneg{j}\right) = \emptyset$
for every $k\in Z_{j-1}$.
\item Assume that there exists $k\in Z_{j-1}$ such that
$\BSG{j+1}{j} \cap \wbasint{k} \ne \emptyset$ and
$\ak$ is maximal verifying these conditions. Then,
$\BSG{j+1}{j}$ is contained in one of the two connected components
of $\obasintabs{k} \setminus \kstarset$ when
$\BSG{j+1}{j} \cap \basintabs{k} \ne \emptyset$, and
$\BSG{j+1}{j}$ is contained in one of the two connected components
of $\wobasint{k}\setminus \basintabs{k}$ if
$\BSG{j+1}{j} \cap \basintabs{k} = \emptyset$
(note that, in this case, $k$ must be negative).
\item Let $\ell \in \{j, -(j+1)\}$
(recall that the ball $\wbasint{\ell}$ has diameter $\alpha_j$
for these two values of $\ell$ and only for them).
\begin{enumerate}[{(\tsfR.5.}i)]
\item If $\lstar \notin \bigcup_{i\in Z_{j-1}} \wbasint{i}$ then,
$\wbasint{\ell} \cap \wbasint{i} = \emptyset$
for every $i \in Z_{j-1}.$
\item If $\lstar \in \wbasint{m}$ for some $m \in Z_{j-1}$
such that $\am$ is maximal with these properties, then
\begin{enumerate}[({\tsfR.5.ii.}1)]
\item $\wbasint{\ell} \cap \wbasint{i} = \emptyset$ for every
$i \in Z_{j-1}$ such that $\ai \ge \am,\ i \ne m,$ and
\item $\wbasint{\ell}$ is contained in (a connected component of)\\[-1ex]
\begin{minipage}{25em
\begin{multline*}
\wobasint{m} \setminus \left(\Bd\left(\basintabs{m}\right) \cup \mstarset\right) =\\
\left(\mstar - \alpha_{_{\am - 1}}, \mstar - \alpha_{_{\am}}\right) \cup
\left(\mstar - \alpha_{_{\am}}, \mstar\right) \cup\\
\left(\mstar, \mstar + \alpha_{_{\am}}\right) \cup
\left(\mstar + \alpha_{_{\am}}, \mstar + \alpha_{_{\am - 1}}\right)
\end{multline*}
\end{minipage}\\[1ex]
(observe that $\wbasint{\ell} \subset \wobasint{m} \setminus \basintabs{m}$
can only happen when $m < 0$ since
$\wbasint{m} = \basintabs{m}$ for $m \ge 0$).
\end{enumerate}
\end{enumerate}
\item Let $\ell \in \{j, -j\}.$
If $\wbasint{\ell} \cap \wbasint{m} = \emptyset$
for every $m\in Z_j,$ $m \ne \ell$ then,
to define $\basicbox{\ell}$ and the map $\varphi_{_{\lstar}},$ we set
\[ a_{\ell} = \gams{j-1}(\lstar) = a_\ell^{\pm} = \gams{j-1}(\lstar \pm \alpha_{j}) = 0.\]
Otherwise, there exists $m\in Z_{j-1}$ such that
$\wbasint{\ell}$ is contained in a connected component of
$\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right)$
and $\am$ is maximal with these properties. Then,
to define $\basicbox{\ell}$ and the map $\varphi_{_{\lstar}},$ we set
\begin{enumerate}[{(\tsfR.6.}i)]
\item $a_{\ell} := \gams{\am}(\lstar),$
$a_\ell^{\pm} := \gams{\am}(\lstar \pm \alpha_{j})$ and
$
\Graph\Bigl(\gams{\am}\evalat{\BSG{\ell}{j}}\Bigr) \subset \basicbox{\ell}.
$
\item Assume that there exists $k \in Z_{\am} \subset Z_{j-1}$ such that
$\wbasint{\ell} \subset \obasintabs{k} \setminus \kstarset.$
Then, $\basicbox{\ell}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{k} \setminus \setsilift{\sstar{k}}\right).$
\end{enumerate}
\end{enumerate}
Finally we define $\map{\gams{j}}{\SI\setminus \Zstar_{j}}[\I]$ by
\[
\gams{j}(\theta) =
\begin{cases}
\varphi_{_{\jstar}}(\theta) & \text{if $\theta \in \basint{j} \setminus \sstarset{j},$} \\
\varphi_{_{\sstar{-j}}}(\theta) & \text{if $\theta \in \basintneg{j} \setminus \sstarset{-j},$} \\
\gams{j-1}(\theta) & \text{if $\theta \notin \left( \basint{j} \cup \basintneg{j} \cup \Zstar_{j-1}\right).$}
\end{cases}
\]
(notice that $\Zstar_{j} = \Zstar_{j-1} \cup \{\jstar,\sstar{-j}\}$).
\end{definition}
\begin{figure}[hbt]
\begin{tikzpicture}[domain=-0.5:0.5, x=\textwidth, y=0.3\textwidth]
\def\alphazero{0.12}\def\CentRegBound{0.07}
\def\BasicBox#1{\pgfmathsetmacro{\vertor}{(-1)^#1
\filldraw[draw=Bittersweet, fill=Bittersweet!50]
(-\alphazero,0) -- (-\CentRegBound, 1-\CentRegBound) --
(0,1) -- (\CentRegBound, 1-\CentRegBound) --
(\alphazero, 0) -- (\CentRegBound, \CentRegBound-1) --
(0,-1) -- (-\CentRegBound, \CentRegBound-1) -- cycle;
\foreach \side in { 1, -1 }{
\draw[Bittersweet] (\side*\CentRegBound, 1-\CentRegBound) -- (\side*\CentRegBound, \CentRegBound-1);
\begin{scope} \clip (-\CentRegBound-0.0005,-1) rectangle (\CentRegBound+0.0005,1);
\draw[blue, xscale=\side*\alphazero, yscale=\side*\vertor] plot file {PseudoCurve_AlsManMor-base.table};
\end{scope}
\draw[blue,xscale=\side] (\CentRegBound, -\vertor*\side*0.12791) -- (\alphazero,0) -- (2*\alphazero,0);
};}
\draw[blue] (-0.5,0) -- (0.5,0);
\foreach \j/\xcent/\xs in { 0/0/1, 1/-0.365/0.5, 2/0.21/0.125, 4/0.15/0.083 }{ \pgfmathsetmacro{\ys}{1.0/2^\j}
\begin{scope}[shift={(\xcent,0)}, xscale=\xs, yscale=\ys]
\BasicBox{\j}
\node[below] at (0, -1) {\scriptsize$\win{\sstar{\j}}$};
\end{scope}
\ifnum\j>0
\begin{scope}[shift={(-\xcent,0)}, xscale=\xs, yscale=\ys]
\draw[garnet, line width=3pt] (-2*\alphazero,0) -- (2*\alphazero,0);\BasicBox{\j}
\ifnum\j< 4 \node[below] at (0, -1) {\scriptsize$\win{\sstar{-\j}}$};
\else \node[above] at (0, 1) {\rotatebox{90}{\scriptsize$\win{\sstar{-\j}}$}};
\fi
\end{scope}
\fi
};
\draw[garnet, line width=3pt, join=round] (-0.063-\alphazero/10,0.115 ) -- (-\CentRegBound, 0.12791) --
plot[domain=-\CentRegBound:-\CentRegBound+0.004] (\x,{-((1 + 25*\x/3)^2 - 0.01)*sin(deg(-0.3769911185/\x))});
\draw[garnet, line width=3pt] plot[domain=-0.05685:-0.063+\alphazero/10] (\x,{-((1 + 25*\x/3)^2 - 0.01)*sin(deg(-0.3769911185/\x))});
\foreach \signel in { 1 , -1 }{
\begin{scope}[shift={(\signel*0.063,-\signel*0.02)}, xscale=\signel*0.05, yscale=\signel*0.125]
\filldraw[draw=Bittersweet, fill=Bittersweet!50]
(-\alphazero,0.89) -- (-\CentRegBound, 1-\CentRegBound) --
(0,1) -- (\CentRegBound, 1-\CentRegBound) --
(\alphazero, -0.85) -- (\CentRegBound, \CentRegBound-1) --
(0,-1) -- (-\CentRegBound, \CentRegBound-1) -- cycle;
\foreach \side in { 1, -1 }{
\draw[Bittersweet] (\side*\CentRegBound, 1-\CentRegBound) -- (\side*\CentRegBound, \CentRegBound-1);
\begin{scope} \clip (-\CentRegBound-0.0005,-1) rectangle (\CentRegBound+0.0005,1);
\draw[blue, xscale=\side*\alphazero, yscale=-\side] plot file {PseudoCurve_AlsManMor-base.table};
\end{scope}
};
\draw[blue, join=round] (\CentRegBound, 0.12791) -- (\alphazero,-0.85) -- (0.14,-0.86) -- (0.3,-0.7);
\draw[blue, join=round] (-\CentRegBound, -0.12791) -- (-\alphazero-0.003,0.89) -- (-\alphazero-0.015,1.05);
\end{scope}
};
\draw[blue] plot[domain=-0.05685:-0.063+\alphazero/10] (\x,{-((1 + 25*\x/3)^2 - 0.01)*sin(deg(-0.3769911185/\x))});
\node[above right] at (0.05, 0.07) {\rotatebox{90}{\tiny$\win{\sstar{3}}$}};
\node[above right] at (-0.075, 0.11) {\rotatebox{90}{\tiny$\win{\sstar{-3}}$}};
\end{tikzpicture}
\caption{The boxes $\win{\lstar}$ for $\ell \in \{-4, -3, -2, -1, 0, 1, 2, 3, 4\}$ and the graph of $\gams{4}.$
The wings are represented as a thick \textcolor{garnet}{garnet} curve surrounding the graph of $\gams{4}.$
For clarity the scale and separation between boxes is not preserved.
The circle $\SI$ is parametrized as $[-\tfrac{1}{2}, \tfrac{1}{2}).$
}\label{IterativeBoxes}
\end{figure}
For every $\ell \in \Z$ we define the \emph{winged region associated to $\ell$} as
\[
\wbasicbox{\ell} := \begin{cases}
\basicbox{\ell} & \text{if $\ell \ge 0$, or}\\
\basicbox{\ell} \cup \Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr) & \text{if $\ell < 0$.}
\end{cases}
\]
The next technical lemma shows that the objects from Definition~\ref{PCgenerators}
exist (that is, they are well defined), and studies some of the
basic properties of the family of pseudo-curve generators
$\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty}.$
\begin{remark}[Explicit consequences of Definition~\ref{PCgenerators}]\label{PCgeneratorsExplicitConsequences}
The following statements are easy consequences of Definition~\ref{PCgenerators}.
They are stated explicitly for easiness of usage.
\begin{enumerate}[({\tsfR.}1)]
\item $n_j > j.$
This follows from Definition~\ref{PCgenerators}(\tsfR.1)
and the fact that we have set $n_0 = 1$ and $n_j > n_{j-1}$ for $j\in\N.$
\item For every $j \in \N,$
\[
\wbasint{-j} \cap \Zstar_{j+1} = \sstarset{-j}.
\]
This follows from Definition~\ref{PCgenerators}(\tsfR.2)
for $j-1.$ We get\\[-1ex]
\hspace*{1em}\begin{minipage}{31.5em
\[
\wbasint{-j} \cap \Zstar_{j} = \sstarset{-j}
\andq
\sstar{-(j + 1)}, \sstar{j+1} \notin \wbasint{-j}.
\]
\end{minipage}\\[1ex]
which shows the statement.
\setcounter{enumi}{5}
\item Let $j \in \N$ and $\ell \in \{j, -j\},$
and assume that $\wbasint{\ell} \cap \wbasint{m} = \emptyset$
for every $m\in Z_j,$ $m \ne \ell.$
Then, $\gams{r}\evalat{\wbasint{\ell}} = \gams{0}\evalat{\wbasint{\ell}} \equiv 0$
for $r=1,2, \dots, j-1.$
\begin{enumerate}[{(\tsfR.6.}i)]
\item Assume that here exists $m\in Z_{j-1}$ such that
$\wbasint{\ell}$ is contained in a connected component of
$\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right)$
and $\am$ is maximal with these properties. Then,
$\gams{r}\evalat{\wbasint{\ell}} = \gams{\am}\evalat{\wbasint{\ell}}$
for $r=\am + 1, \am + 2, \dots, j-1.$
\item Assume that there exists $k \in Z_{\am} \subset Z_{j-1}$ such that
$\wbasint{\ell} \subset \obasintabs{k} \setminus \kstarset$ and
$\ak$ is maximal with these properties. Then,
$\gams{r}\evalat{\wbasint{\ell}} = \gams{\ak}\evalat{\wbasint{\ell}}$
for $r=\ak + 1, \ak + 2, \dots, \am.$
\end{enumerate}
\end{enumerate}
To prove (\tsfR.6) notice that when
$
\wbasint{\ell} \cap \basintabs{m}
\subset \wbasint{\ell} \cap \wbasint{m}
= \emptyset
$
for every $m\in Z_j,$ $m \ne \ell,$ from the definition of
$\gams{r}$ for $0 \le r < j$ we get that
$\gams{r}\evalat{\wbasint{\ell}} = \gams{0}\evalat{\wbasint{\ell}} \equiv 0$
for $r=1,2, \dots, j-1.$
\inidemopart{\tsfR.6.i}
The maximality of $\am,$ together with Definition~\ref{PCgenerators}(\tsfR.2),
imply that
$
\wbasint{\ell} \cap \basintabs{i}
\subset \wbasint{\ell} \cap \wbasint{i}
= \emptyset
$ for every $i \in Z_{j-1},$ $\ai \ge \am,$ $i \ne m.$
So, by the definition of the functions $\gams{r}$,
\[
\gams{r}\evalat{\BSG{\ell}{j}} = \gams{\am}\evalat{\BSG{\ell}{j}}
\andq[for]
r=\am + 1, \am + 2, \dots, j-1.
\]
\inidemopart{\tsfR.6.ii}
When $\ak = \am$ (\tsfR.6.ii) holds trivially.
So, assume that $\ak < \am.$
As in the case (\tsfR.6.i),
the maximality of $\ak$ and Definition~\ref{PCgenerators}(\tsfR.2)
imply that
$
\wbasint{\ell} \cap \basintabs{r} = \emptyset
$
for every $r \in Z_{j-1},$ $\abs{r} \ge \ak,$ $r \ne k.$
So, (\tsfR.6.ii) follows from the definition of
the functions $\gams{r}$.
\end{remark}
\begin{lemma}\label{Propertiesvarphi}
For every $j\in \Z^+$ the regions
$\basicbox{j}$ and $\basicbox{-j}$ (and hence $\wbasicbox{j}$ and $\wbasicbox{-j}$),
and the maps $(\gams{j},\SI\setminus \Zstar_{j})$ are well defined.
Moreover, the following statements hold:
\begin{enumerate}[(a)]
\item $(\gams{j}, \SI\setminus\Zstar_{j}) \in \C.$
Furthermore, for every $\ell \in \{j+1, -(j+1)\},$
\[
\gams{j} \left(\BSG{\ell}{j}\right) \subset \left[
\gams{j}(\lstar) - 2^{-n_{j}},
\gams{j}(\lstar) + 2^{-n_{j}}
\right].
\]
\item $\bigcup_{\ell \in \Z} \wbasicbox{\ell} \subset \SI \times [-1,1]$ and
$\Graph\left(\gams{j}\evalat{\SI\setminus\Zstar_{j}}\right) \subset \SI \times [-1,1].$
\item For $\ell \in \{j, -j\}$ we have
$\Graph\Bigl(\gams{j-1}\evalat{\BSG{\ell}{j}}\Bigr) \subset \basicbox{\ell},$
$a_{\ell} = \gams{j-1}(\lstar),$ and
$
a^{\pm}_{\ell} =
\varphi_{_{\lstar}}(\lstar\pm\alpha_{j}) =
\gams{j-1}(\lstar\pm\alpha_{j}).
$
\item $\Graph\Bigl(\gams{n}\evalat{\BSG{\ell}{j} \setminus \Zstar_n}\Bigr) \subset \basicbox{\ell}$
for every $n \ge j$ and $\ell \in \{j, -j\}.$
\item For every $\ell \in \{j, -j\},$
\[\hspace*{1.5em}
\gams{j}\evalat{\left(\wbasint{\ell} \setminus \OBG{\ell}{j}\right) \cup R_\omega\left(\wbasint{\ell} \setminus \OBG{\ell}{j}\right)} =
\gams{j-1}\evalat{\left(\wbasint{\ell} \setminus \OBG{\ell}{j}\right) \cup R_\omega\left(\wbasint{\ell} \setminus \OBG{\ell}{j}\right)}.
\]
Moreover, for every
$\theta \in \Bd(\wbasint{\ell} \setminus \OBG{\ell}{j}) = \Bd(\BSG{\ell}{j}) \cup \Bd(\wbasint{\ell}),$
we have
$\theta \notin \wbasint{n} \cup \wbasint{-n}$ and
$\gams{n}(\theta) = \gams{j}(\theta) = \gams{j-1}(\theta)$
for every $n > j,$ and
$R_\omega(\theta) \notin \basint{n} \cup \basintneg{n}$ and
$\gams{n}\bigl(R_\omega(\theta)\bigr) = \gams{j-1}\bigl(R_\omega(\theta)\bigr)$
for every $n \ge j.$
\item For every $\ell \in \Z,$ $\wbasicbox{\ell}$ is a compact connected set
such that $\pi\left(\wbasicbox{\ell}\right) = \wbasint{\ell},$
$\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}$ is continuous
and
\[
\diam\left(\wbasicbox{\ell}\right) = \begin{cases}
\diam\left(\basicbox{\ell}\right) = \diam\left(\basicbox{-\ell}\right) =
2 \cdot 2^{-n_{\ell}} \le 2^{-\ell} & \text{if $\ell \ge 0$,}\\
2 \cdot 2^{-n_{\abs{\ell+1}}} \le 2\cdot 2^{-{\all}} & \text{if $\ell < 0$.}
\end{cases}
\]
\item Given $\ell,m\in \Z$ such that
$\all \ge \am,$ $\ell \ne m$ and $\wbasint{\ell} \cap \wbasint{m} \ne \emptyset,$
it follows that $\all > \am,$ and either
$\wbasint{\ell} \subset \obasintabs{m} \setminus \mstarset$
and the region $\wbasicbox{\ell}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{m} \setminus \setsilift{\mstar}\right),$
or $m < 0$ and $\wbasint{\ell}$ is contained in one of
the two connected components of $\wobasint{m} \setminus \basintabs{m}$.
\end{enumerate}
\end{lemma}
\begin{proof}
We start by proving the first statement of the lemma and (a) by induction.
Observe that $n_0 = 1$, $\alpha_0,$ $\delta_0$ and $\gams{0}$ are defined so that
Definition~\ref{PCgenerators}(\tsfR.1--2) for $j=0$
and $(\gams{0}, \SI\setminus\Zstar_{0}) \in \C$
are verified except for the obvious fact that $\wbasint{-j} = \wbasint{j}.$
On the other hand, by construction, $\basint{0}$ is disjoint from
$\BSG{1}{0}$ and $\BSG{-1}{0}.$
Then, by the definition of $\gams{0},$
\[
\gams{0} \left(\BSG{\ell}{0}\right) = \{0\} \subset [-\tfrac{1}{2}, \tfrac{1}{2}] =
\left[
\gams{0}(\lstar) - 2^{-n_{0}},
\gams{0}(\lstar) + 2^{-n_{0}}
\right]
\]
for $\ell \in \{1, -1\}.$
Hence, (a) holds.
Fix $j > 0$ and assume that we have defined
$n_{\ell}$, $\alpha_{\ell},$ $\delta_{\ell}$ and $\gams{\ell}$
such that all Definition~\ref{PCgenerators}(\tsfR.1--6) above and (a)
hold for $\ell = 0, 1, \dots, j-1.$
Since the elements of $\Zstar_{j + 2}$ are pairwise different,
we can choose an integer $n_{j} > n_{j-1}$ and $\delta_{j}$ and $\alpha_j$
small enough so that
\begin{itemize}
\item $0 < \delta_{j} < \alpha_j < 2^{-n_j} < \delta_{j-1},$
\item $\sstar{-(j + 2)}, \sstar{j+2} \notin \wbasint{-(j+1)} = \BSG{-(j+1)}{j},$
\item the three intervals
$\wbasint{j} = \basint{j},$
$R_\omega\left(\basint{j}\right) = \BSG{j+1}{j}$ and
$\wbasint{-(j+1)}$
are pairwise disjoint,
\item $\wbasint{\ell} \cap \Zstar_{j+1} = \lstarset$
for $\ell \in \{j, -(j+1)\}$,\newline
$\BSG{j+1}{j} \cap \Zstar_{j+1} = \sstarset{j+1}$
and, additionally,
\item $
\Bigl( \Bd\left(\basintneg{j}\right) \cup
\Bd\left(\basint{j}\right) \Bigr) \cap \Orbom = \emptyset.
$
\end{itemize}
Then, Definition~\ref{PCgenerators}(\tsfR.1) is verified.
Moreover, from the above conditions it follows that
$\BSG{\ell}{j} \cap \Zstar_{j+1} = \lstarset$
for every $\ell \in \{j+1, -(j+1)\}.$
Thus, by statement~(a) for $j-1$,
$\gams{j-1}$ is defined and continuous on $\lstar \in \BSG{\ell}{j}$
because this interval is disjoint from $\Zstar_{j-1}.$
Hence, we can decrease the value of $\alpha_j$
(and, accordingly, the value of $0 < \delta_{j} < \alpha_j$),
if necessary, to get\smallskip
\begin{itemize}
\item \hfill$
\gams{j-1} \left(\BSG{\ell}{j}\right) \subset \left[
\gams{j-1}(\lstar) - 2^{-n_{j}},
\gams{j-1}(\lstar) + 2^{-n_{j}}
\right]
$\hfill\strut\par\par\medskip\par\par\noindent
for every $\ell \in \{j+1, -(j+1)\}.$
\end{itemize}
To see that Definition~\ref{PCgenerators}(\tsfR.2) is verified it remains
to show that the intervals $\wbasint{j},$ $\BSG{j+1}{j}$ and
$\wbasint{-(j+1)}$ are disjoint from $\wbasint{-j}.$
By induction, Definition~\ref{PCgenerators}(\tsfR.2) holds for $j-1$. Thus we see, that
$\sstar{-(j + 1)}, \sstar{j+1} \notin \wbasint{-j},$
and $R_\omega\left(\basint{j-1}\right) = \BSG{j}{j-1}$
is disjoint from $\wbasint{-j}.$
Hence, we can decrease the value of $\alpha_j$
(and, accordingly, the value of $0 < \delta_{j} < \alpha_j$),
if necessary, until $\BSG{j+1}{j}$ and $\wbasint{-(j+1)} = \BSG{-(j+1)}{j}$
are disjoint from $\wbasint{-j}.$
On the other hand we have that
$\alpha_j < 2^{-n_j} < \delta_{j-1} < \alpha_{j-1}.$ So,
$\wbasint{j} = \basint{j} \subset \BSG{j}{j-1}$ is disjoint from $\wbasint{-j}.$
Up to now we have seen that we can choose
$n_{j},$ $\delta_{j}$ and $\alpha_j$ so that Definition~\ref{PCgenerators}(\tsfR.1--2)
hold for $j$.
Let us see that we can choose $\alpha_j$ such that Definition~\ref{PCgenerators}(\tsfR.3) also holds.
Observe that for every $\ell, i \in \Z$ and every $m\ge 0$ it follows that
$\Bd\left(\BSG{\ell}{m}\right) \cap \Orbom \ne \emptyset$ if and only if
$
\Bd\left(R^i_\omega\left(\BSG{\ell}{m}\right)\right) \cap \Orbom =
\Bd\left(\BSG{\ell + i}{m}\right) \cap \Orbom \ne \emptyset.
$
Therefore, by using Definition~\ref{PCgenerators}(\tsfR.1) inductively, we obtain
\[
\bigcup_{k\in Z_{j-1}} \Bd\left(\BSG{k+1}{\ak}\right) \cap \{\sstar{-j}, \jstar\} \subset
\bigcup_{k\in Z_{j-1}} \Bd\left(\BSG{k+1}{\ak}\right) \cap \Orbom = \emptyset.
\]
Consequently, since $\bigcup_{k\in Z_{j-1}} \Bd\left(\BSG{k+1}{\ak}\right)$ is a finite set,
by decreasing again the value of $\alpha_j$, if necessary,
we can achieve that Definition~\ref{PCgenerators}(\tsfR.3) holds for $j$
and Definition~\ref{PCgenerators}(\tsfR.1--2) are still verified.
Next we will take care of Definition~\ref{PCgenerators}(\tsfR.4).
If $\sstar{j+1} \notin \bigcup_{i\in Z_{j-1}} \wbasint{i},$
by decreasing again the value of $\alpha_j$ (and $\delta_j$), if necessary,
we can achieve that
$\BSG{j+1}{j} \cap \left( \bigcup_{i\in Z_{j-1}} \wbasint{i} \right) = \emptyset$
while preserving that Definition~\ref{PCgenerators}(\tsfR.1--3) are verified for $j$.
In this case Definition~\ref{PCgenerators}(\tsfR.4) holds trivially.
Conversely, assume that there exists $k\in Z_{j-1}$ such that
$\sstar{j+1} \in \wbasint{k}$ and $\ak$ is maximal verifying these conditions.
By Definition~\ref{PCgenerators}(\tsfR.2), $k$ is unique (that is, the condition cannot be verified
by $k$ and $-k$ simultaneously).
On the other hand, by the Definition~\ref{PCgenerators}(\tsfR.1) for $\ak$ and $\ak - 1$
and the comment above,
$\sstar{j+1} \notin \Bd\left(\wbasint{k}\right) \cup \Bd\left(\basintabs{k}\right).$
Since $k \in Z_{j-1},$ $\ak \le j-1$ and, hence,
$\sstar{j+1} \notin \Zstar_{\ak}$ (in particular $\jstar \ne \kstar$).
Consequently, $\sstar{j+1}$ is contained in one of the connected components of
$\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \Zstar_{\ak} \right).$
Then, by decreasing again the value of $\alpha_j$, if necessary,
we can get that $\BSG{j+1}{j}$ is contained in the connected component of
$\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \Zstar_{\ak} \right)$
where $\sstar{j+1}$ lies, while preserving that
Definition~\ref{PCgenerators}(\tsfR.1--3) are verified for $j$.
Consequently, Definition~\ref{PCgenerators}(\tsfR.1--4) hold for $j$.
Now we will deal with Definition~\ref{PCgenerators}(\tsfR.5).
If $\lstar \notin \bigcup_{i \in Z_{j-1}} \wbasint{i},$
by decreasing again the value of $\alpha_j$, if necessary,
we can get Definition~\ref{PCgenerators}(\tsfR.5.i) while preserving that
Definition~\ref{PCgenerators}(\tsfR.1--4) are verified for $j$.
Assume that there exists $m \in Z_{j-1}$ such that
$\lstar \in \wbasint{m}$ and $\am$ is maximal with these properties.
As in the above construction, by Definition~\ref{PCgenerators}(\tsfR.1--2),
$\lstar \in \wobasint{m} \setminus \left(\Bd\left(\basintabs{m}\right) \cup \mstarset \right)$
and $m$ is unique (that is, the condition cannot be verified
simultaneously by $m$ and $-m$). Consequently,
$\lstar \notin \wbasint{i}$ for every
$i \in Z_{j-1}$ such that $\ai \ge \am,\ i \ne m.$
Thus, by decreasing again the value of $\alpha_j$, if necessary,
we can get that Definition~\ref{PCgenerators}(\tsfR.1--4) still hold,
Definition~\ref{PCgenerators}(\tsfR.5.ii.1) is verified and the interval
$\wbasint{\ell}$ is contained in the connected component of
$\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right)$
where $\lstar$ lies.
So, Definition~\ref{PCgenerators}(\tsfR.5.ii.2) also holds.
We claim that\\\noindent{\itshape
for every $\ell,m\in \Z$ such that
$\am \le \all \le j,$ $\ell \ne m,$ either
$\wbasint{\ell} \cap \wbasint{m} = \emptyset$
or $\am < \all$ and
$\wbasint{\ell}$ is contained in a connected component of
\[
\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right).
\]}\\[-1ex]
We prove the claim by induction.
Observe that the claim holds trivially for $\am \le \all \le 1$ because
$\wbasint{0},$ $\wbasint{1} = \basint{1} \subset \BSG{1}{0}$ and
$\wbasint{-1}$ are pairwise disjoint by construction.
Assume that the claim holds for every $\am \le \all < j.$
So, to prove the claim, we may assume that
$\ell \in \{j, -j\},$ $m \in Z_{j-1} \cup \{-\ell\}$
and $\wbasint{\ell} \cap \wbasint{m} \ne \emptyset$.
By Definition~\ref{PCgenerators}(\tsfR.2),
$\wbasint{j} \cap \wbasint{-j} = \emptyset.$
Consequently, $m \ne -\ell$
(that is, $m \in Z_{j-1}$ and $\all = j > \am$).
On the other hand, if $\ell = -j,$ Definition~\ref{PCgenerators}(\tsfR.2) for $j-1$ shows that
$\wbasint{j-1},$ $\wbasint{-(j-1)}$ and $\wbasint{-j}$
are pairwise disjoint. Thus, $m \in Z_{j-2}$ in this case.
Hence, by the Definition~\ref{PCgenerators}(\tsfR.5) for $j$ when $\ell = j$ and
for $j-1$ when $\ell = -j,$ there exists $k \in Z_{j-1}$
(in fact when $\ell = -j,$ $k \in Z_{j-2}$) such that $\wbasint{\ell}$
is contained in a connected component of
$\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \kstarset \right)$
and $\all = j > \ak \ge \am.$
If $m = k$ then the claim holds.
Otherwise, $m \ne k$ and since $j=\all > \ak \ge \am$,
by the induction hypotheses, $\ak > \am,$ and $\wbasint{k}$
is contained in a connected component of
$\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right).$
So, the claim holds also in this case. This ends the proof of the claim.
Finally, we consider Definition~\ref{PCgenerators}(\tsfR.6).
The fact that either $\wbasint{\ell} \cap \wbasint{m} = \emptyset$
for every $m\in Z_j,$ $m \ne \ell$ or
there exists $m\in Z_{j-1}$ such that
$\wbasint{\ell}$ is contained in a connected component of
$\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right)$
follows from the claim.
To show that Definition~\ref{PCgenerators}(\tsfR.6.i) can be guaranteed,
it is enough to decrease again the value of $\alpha_j$, if necessary,
until $\BSG{\ell}{j}$ is disjoint from $\Zstar_{\am}$ and
Definition~\ref{PCgenerators}(\tsfR.1--5) are still verified.
Thus by (a) for $\am,$ $\gams{\am}$ is well defined and continuous on
$\BSG{\ell}{j}.$
So, we can set $a_{\ell} := \gams{\am}(\lstar)$ and,
by decreasing again $\alpha_j$ (if necessary), we get
$
\Graph\Bigl(\gams{\am}\evalat{\BSG{\ell}{j}}\Bigr) \subset \basicbox{j}.
$
To show that Definition~\ref{PCgenerators}(\tsfR.6.ii) can be guaranteed we first assume that $k = m.$
As before, if necessary, we can increase the value of $n_{j}$ and, accordingly,
decrease the values of $\alpha_j < 2^{-n_{j}}$ and $0 < \delta_{j} < \alpha_j$
so that Definition~\ref{PCgenerators}(\tsfR.1--5) and (\tsfR.6.i)
are still verified for $j$ and in addition,
\[
(\lstar,a_{\ell}+2^{-n_{j}}), (\lstar,a_{\ell}-2^{-n_{j}}) \in \Int(\basicbox{k})
\]
and the region $\basicbox{\ell}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{k} \setminus \setsilift{\kstar}\right).$
Assume now that $k \ne m$ (recall that $\ak \le \am < j$).
In this case we have $\wbasint{\ell} \subset \wobasint{m} \cap \obasintabs{k}$.
In particular, $\wobasint{m} \cap \obasintabs{k} \ne \emptyset$ and,
by the above claim, $\ak < \am$ and
$\wbasint{\ell} \subset \wbasint{m}$ is contained in a connected component of
$\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \kstarset \right).$
The fact that $\wbasint{\ell} \subset \obasintabs{k} \setminus \kstarset$
implies that
$\wbasint{\ell} \subset \wbasint{m} \subset \obasintabs{k} \setminus \kstarset.$
Then, as above we can increase the value of $n_{j}$ and, accordingly,
decrease the values of $\alpha_j < 2^{-n_{j}}$ and $0 < \delta_{j} < \alpha_j$
so that Definition~\ref{PCgenerators}(\tsfR.1--5) and (\tsfR.6.i) are still verified,
\[
(\lstar,a_{\ell}+2^{-n_{j}}), (\lstar,a_{\ell}-2^{-n_{j}}) \in \Int(\basicbox{k})
\]
and the region $\basicbox{\ell}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{k} \setminus \setsilift{\kstar}\right).$
Now assume that $\ak$ is not maximal verifying the assumptions.
Then, there exists $r \in Z_{\am} \subset Z_{j-1}$ such that
$\wbasint{\ell} \subset \obasintabs{r} \setminus \sstarset{r}$
and $\abs{r}$ is maximal with these properties.
We have $\ak \le \abs{r} \le \am < j$ and
\[
\wbasint{r} \cap \wbasint{k} \supset \obasintabs{r} \cap \obasintabs{k} \ne \emptyset
\]
because
$\wbasint{\ell} \subset \obasintabs{r} \cap \obasintabs{k}.$
Then, by the claim, $\ak < \abs{r}$ and $\wbasint{r}$
is contained in a connected component of
$\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \kstarset \right).$
The fact that $\wbasint{\ell} \subset \obasintabs{k} \setminus \kstarset$
implies that
$\wbasint{r} \subset \obasintabs{k} \setminus \kstarset.$
By the part already proven and Definition~\ref{PCgenerators}(\tsfR.6.ii) for $\abs{r} < j$ we get
that $\basicbox{\ell}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{r} \setminus \setsilift{\sstar{r}}\right)$
and $\basicbox{r}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{k} \setminus \setsilift{\sstar{k}}\right).$
This shows that Definition~\ref{PCgenerators}(\tsfR.6.ii) can be guaranteed.
Let us prove that (a) holds for $j$.
Since the set $\SI\setminus\Zstar_{j}$ is residual,
to prove that $(\gams{j}, \SI\setminus\Zstar_{j}) \in \C$
we have to show that
$\gams{j}\evalat{\SI\setminus\Zstar_{j}}$
is continuous.
Note that, from Definition~\ref{PCgenerators}(\tsfR.6.ii),
$
a^{\pm}_{\ell} = \varphi_{_{\lstar}}(\lstar\pm\alpha_{j}) =
\gams{j-1}(\lstar\pm\alpha_{j}).
$
Hence, the continuity of $\gams{j}\evalat{\SI\setminus\Zstar_{j}}$
follows from the fact that $\gams{j-1}$ is continuous on
$\SI\setminus\Zstar_{j-1} \supset \SI\setminus\Zstar_{j}$ and
the continuity of $\varphi_{_{\jstar}}$ and $\varphi_{_{\sstar{-j}}}$
(Definition~\ref{GenericBoxes}).
This ends the proof of the first statement of the lemma and
the first statement of (a).
For every $\ell \in \{j+1, -(j+1)\},$
from By Definition~\ref{PCgenerators}(\tsfR.1,2) we get:
\begin{align*}
\gams{j-1} \left(\BSG{\ell}{j}\right)
&\subset \left[
\gams{j-1}(\lstar) - 2^{-n_{j}},
\gams{j-1}(\lstar) + 2^{-n_{j}}
\right]\\
\BSG{\ell}{j} &\text{ is disjoint from
$\basint{j}$ and $\BSG{-j}{j-1} \supset \BSG{-j)}{j}$, and}\\
\lstarset & \notin \BSG{\ell}{j} \cap \Zstar_{j-1} \subset
\BSG{\ell}{j} \cap \Zstar_{j+1} = \lstarset.
\end{align*}
So, from the definition of $\gams{j}$
it follows that
\[
\gams{j}\evalat{\BSG{\ell}{j}} = \gams{j-1}\evalat{\BSG{\ell}{j}}
\]
and, thus, (a) holds.
Statement~(c) follows immediately
from Definition~\ref{PCgenerators}(\tsfR.6)
and Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.6).
Next we prove (b,d,e,f,g).
\inidemopart{d} When $n = j,$ we get
$
\BSG{\ell}{j} \setminus \Zstar_j = \BSG{\ell}{j} \setminus \lstarset
$
from Definition~\ref{PCgenerators}(\tsfR.2). Hence,
$
\Graph\Bigl(\gams{j}\evalat{\BSG{\ell}{j} \setminus \Zstar_j}\Bigr) \subset \basicbox{\ell}
$
by the definition of $\gams{j}$ (Definition~\ref{PCgenerators}) and the definition of
$\varphi_{_{\lstar}}$ (Definition~\ref{GenericBoxes}).
Now assume that $n > j$ and fix $\theta \in \BSG{\ell}{j} \setminus \Zstar_n.$
We have to show that the point $(\theta,\gams{n}(\theta)) \in \basicbox{\ell}.$
If $\theta \notin \basintabs{m} $ for every $m$ such that $j < \am \le n$
then, by the iterative use of the definition of
$\gams{i}$ for $i=j+1,j+2,\dots, n$ (Definition~\ref{PCgenerators})
and Definition~\ref{GenericBoxes},
\[
(\theta,\gams{n}(\theta)) = (\theta,\gams{n-1}(\theta)) = \dots =
(\theta,\gams{j+1}(\theta)) = (\theta,\gams{j}(\theta)) =
(\theta,\varphi_{_{\lstar}}(\theta)) \in \basicbox{\ell}.
\]
Otherwise, by Definition~\ref{PCgenerators}(\tsfR.2),
there exists $m \in \Z$ such that $\all < \am \le n,$
$
\theta \in \basintabs{m} \setminus \Zstar_n,
$
and
$\theta \notin \basintabs{s}$ for every $s$ such that $\am < \abs{s} \le n.$
This implies that
$\wbasint{\ell} \cap \wbasint{m} \supset
\BSG{\ell}{j} \cap \basintabs{m} \ne
\emptyset
$ and $\am$ is maximal with these properties.
So, by the claim for $j = \am,$
$\wbasint{m}$ is contained in a connected component of
$\wobasint{\ell} \setminus \left( \Bd\left(\basintabs{\ell}\right) \cup \lstarset \right)$.
Moreover, since $\theta \in \wobasint{m} \cap \BSG{\ell}{j} \ne \emptyset,$
$\wbasint{m} \subset \basintabs{\ell} \setminus \lstarset$.
Thus, by Definition~\ref{PCgenerators}(\tsfR.6.ii) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.6.ii) for $j = \am,$
$\ell$ replaced by $m$ and $k$ replaced by $\ell$,
$\basicbox{m} \subset \basicbox{\ell}$
and (d) follows from the part already proven
by replacing $\ell$ by $m$ and $j$ by $\am.$
\inidemopart{g} By the claim we have that for every $\ell,m\in \Z$ such that
$\all \ge \am,$ $\ell \ne m$ and $\wbasint{\ell} \cap \wbasint{m} \ne \emptyset,$
it follows that $\all > \am,$ and
$\wbasint{\ell}$ is contained in a connected component of
$\wobasint{m} \setminus \left( \Bd\left(\basintabs{m}\right) \cup \mstarset \right)$.
Only it remains to show that if
$\wbasint{\ell} \subset \obasintabs{m} \setminus \mstarset,$
then the region $\wbasicbox{\ell}$ is contained in one of the
two connected components of
$\Int\left(\basicbox{m} \setminus \setsilift{\mstar}\right).$
By Definition~\ref{PCgenerators}(\tsfR.6.ii) we know that this holds for
$\basicbox{\ell}$ instead of $\wbasicbox{\ell}.$
Hence, if $\ell \ge 0,$ (g) holds because $\wbasicbox{\ell} = \basicbox{\ell}.$
Assume now that $\ell < 0.$ Since
$\wbasicbox{\ell} = \basicbox{\ell} \cup \Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)$
is connected,
$\basicbox{\ell} \subset \basicbox{m},$ and
$\Int\left(\basicbox{m} \setminus \setsilift{\mstar}\right)$
has two connected components, it is enough to show that
\[
\Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)
\subset \basicbox{m}.
\]
Since
$
\wbasint{\ell} \setminus \obasintabs{\ell} \subset
\wbasint{\ell} \subset
\obasintabs{m} \setminus \mstarset,
$
statement (g) follows from (d) with $\ell$ replaced by $m$, $j$ by $\am$ and
$n$ replaced by $\all.$
\inidemopart{b}
With (g) in mind we set
\[
\mathsf{D} := \set{\ell \in \Z}{\wbasicbox{\ell} \not\subset \basicbox{i} \text{ for every }i \in \Z\setminus\{\ell\}}.
\]
Clearly,
\begin{align*}
\bigcup_{\ell \in \Z} \wbasicbox{\ell} &=
\left(\bigcup_{i \in \Z\setminus \mathsf{D}} \wbasicbox{i}\right) \cup
\left(\bigcup_{\ell \in \mathsf{D}} \wbasicbox{\ell}\right) \\
&\subset
\left(\bigcup_{i \in \mathsf{D}} \basicbox{i}\right) \cup
\left(\bigcup_{\ell \in \mathsf{D}} \wbasicbox{\ell}\right)
= \bigcup_{\ell \in \mathsf{D}} \wbasicbox{\ell}
\end{align*}
\begin{case}{Claim:}
For every $\ell \in \mathsf{D},$
$\gams{\all-1}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}} \equiv 0.$
\end{case}
First we prove statement (b) from the above claim and then we will prove the claim.
To this end we start by pointing out few elementary facts.
From the definition of $\wbasicbox{\ell}$ we see that
$\wbasicbox{\ell} \setminus \basicbox{\ell} = \emptyset$ for every $\ell \ge 0$
and
$
\wbasicbox{\ell} \setminus \basicbox{\ell} \subset
\Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)
$
for every $\ell < 0.$
So, in any case,
\[
\wbasicbox{\ell} \setminus \basicbox{\ell} \subset
\Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)
\andq[for every]
\ell \in \Z.
\]
On the other hand, the arc
$\wbasint{\ell} \supset \wbasint{\ell} \setminus \obasintabs{\ell}$
is disjoint from the arc
$\wbasint{-\ell} \supset \BSG{-\ell}{\all}$
by Definition~\ref{PCgenerators}(\tsfR.2). Thus,
by Definition~\ref{PCgenerators} and (a),
\[
\gams{\all-1}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}} =
\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}.
\]
Furthermore, by the Claim and Definition~\ref{PCgenerators}(\tsfR.6),
$a_{\ell}^+ = a_{\ell}^- = a_{\ell} = 0$ for every $\ell \in D.$
So, by Remark~\ref{propiedadesR}(1),
\[
\basicbox{\ell} \subset
\basintabs{\ell} \times [-2^{-n_{\all}}, 2^{-n_{\all}}] \subset
\basintabs{\ell} \times [-2^{-\all}, 2^{-\all}] \subset
\basintabs{\ell} \times [-1,1].
\]
Therefore, summarizing and using again by the Claim,
\begin{align*}
\bigcup_{\ell \in \Z} \wbasicbox{\ell} &\subset
\bigcup_{\ell \in \mathsf{D}} \wbasicbox{\ell} \subset
\bigcup_{\ell \in \mathsf{D}} \left( \basicbox{\ell} \cup
\Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)
\right)\\
&=
\left(\bigcup_{\ell \in \mathsf{D}} \basicbox{\ell} \right) \cup \left(\bigcup_{\ell \in \mathsf{D}}
\Graph\Bigl(\gams{\all-1}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)\right)\\
\subset \left(\bigcup_{\ell \in \mathsf{D}} \basintabs{\ell} \right) \times [-1,1] \cup \SI \times \{0\}
\subset \SI \times [-1,1].
\end{align*}
So, the first part of (b) is proved, provided that the claim holds.
Let us prove the second statement of (b).
Observe that, since
\[
\left(\bigcup_{\ell \in \Z} \basicbox{\ell} \right) \cup \SI \times \{0\} \subset
\left(\bigcup_{\ell \in \Z} \wbasicbox{\ell} \right) \cup \SI \times \{0\}
\subset \SI \times [-1,1],
\]
it is enough to show that
\[
\Graph\left(\gams{j}\evalat{\SI\setminus\Zstar_{j}}\right) \subset
\left(\bigcup_{\ell \in \Z} \basicbox{\ell} \right) \cup \SI \times \{0\}
\]
for every $j \in \Z^+.$
We will prove this statement by induction on $j.$
By construction we have
\[
\Graph\left(\gams{0}\evalat{\SI\setminus\{\sstar{0}\}}\right) \subset
\basicbox{0} \cup \SI \times \{0\} \subset
\left(\bigcup_{\ell \in \Z} \basicbox{\ell} \right) \cup \SI \times \{0\}.
\]
So, the statement holds for $j=0$. Now assume that it holds for some $j \ge 0,$
and prove it for $j+1.$
By Definition~\ref{PCgenerators} and (d),
\begin{align*}
\Graph\left(\gams{j+1}\evalat{\SI\setminus\Zstar_{j+1}}\right) &\subset
\basicbox{j} \cup \basicbox{-j} \cup
\Graph\left(\gams{j}\evalat{\SI\setminus\Zstar_{j}}\right) \\
&\subset \basicbox{j} \cup \basicbox{-j} \cup
\left(\bigcup_{\ell \in \Z} \basicbox{\ell} \right) \cup \SI \times \{0\}\\
&\subset \left(\bigcup_{\ell \in \Z} \basicbox{\ell} \right) \cup \SI \times \{0\}.
\end{align*}
To end the proof of (b) it remains to show the Claim.
Let $\ell \in \mathsf{D}$ and $m \in Z_{\all},$ $m \ne \ell.$
Then, either
\begin{equation}\label{DichotomyforD}
\begin{cases}
\wbasint{\ell} \cap \wbasint{m} = \emptyset\text{ or}\\
\all > \am,\ m < 0 \text{ and } \wbasint{\ell} \subset \wobasint{m} \setminus \basintabs{m}.
\end{cases}
\end{equation}
To see this, observe that if $\wbasint{\ell} \cap \wbasint{m} \ne \emptyset$
then, by (g), $\all > \am$ and either $\wbasicbox{\ell} \subset \basicbox{m}$ or
$m < 0$ and $\wbasint{\ell} \subset \wobasint{m} \setminus \basintabs{m},$
and the first possibility is ruled out because $\ell \in \mathsf{D}.$
By using iteratively the dichotomy \eqref{DichotomyforD} we get that, for every $\ell \in \mathsf{D},$
there exists a sequence $m_0,m_1,\dots,m_k = \ell \in \Z$ with $k \ge 0$ such that
$\wbasint{m_0} \cap \wbasint{q} = \emptyset$ for every $q \in Z_{\abs{m_0}},\ q \ne m_0$
and, in the case $k > 0,$
$
\abs{m_0} < \abs{m_1} < \dots < \abs{m_k} = \all
$
and, for every $p=0,1,\dots,k-1,$
\begin{itemize}
\item $m_{p} < 0,$
\item $\wbasint{m_{p+1}} \subset \wobasint{m_p} \setminus \basintabs{m_p}$ and
\item $\wbasint{m_{p+1}} \cap \wbasint{q} = \emptyset$
for every $q\in Z_{\abs{m_{p+1}}},$ $q \ne m_p, m_{p+1}$ and $\abs{m_p} \le \aq.$
\end{itemize}
The condition
$\wbasint{m_0} \cap \wbasint{q} = \emptyset$ for every $q \in Z_{\abs{m_0}},\ q \ne m_0$
implies
\[
\gams{\abs{m_0}-1}\evalat{\wbasint{m_0}} = \gams{\abs{m_0}-2}\evalat{\wbasint{m_0}} =
\dots = \gams{0}\evalat{\wbasint{m_0}} \equiv 0
\]
by Definition~\ref{PCgenerators}(\tsfR.6) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.6)
(with $\ell = m_0$). This ends the proof of the Claim when $k = 0.$
Assume now that $k > 0.$
As before we have
\[
\gams{\abs{m_0}-1}\evalat{\wbasint{m_0} \setminus \obasintabs{m_0}} =
\gams{\abs{m_0}}\evalat{\wbasint{m_0} \setminus \obasintabs{m_0}}.
\]
This, together with the inclusion,
\[
\wbasint{m_{1}} \subset \wobasint{m_0} \setminus \basintabs{m_0}
\]
implies that
\[
\gams{\abs{m_0}}\evalat{\wbasint{m_1}} \equiv 0.
\]
Then, by Definition~\ref{PCgenerators}(\tsfR.6.i) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.6.i) with $\ell = m_1$,
\[
0 \equiv \gams{\abs{m_0}}\evalat{\wbasint{m_1}} = \gams{\abs{m_0}+1}\evalat{\wbasint{m_1}} =
\dots = \gams{\abs{m_1}-1}\evalat{\wbasint{m_1}}.
\]
If $k = 1$ we are done. Otherwise, $k \ge 2$ and, as above,
\[
\gams{\abs{m_1}}\evalat{\wbasint{m_2}} \equiv 0.
\]
By iterating the above arguments at most $k$ times the Claim holds.
This ends the proof of (b).
\inidemopart{e}
By Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2)
it follows that
\[
\theta \notin \Zstar_{j+1} \cup \OBG{\ell}{j} \cup \wobasint{-\ell}
\andq[for every]
\theta \in \wbasint{\ell} \setminus \OBG{\ell}{j}.
\]
So, by (a), $\gams{j-1}(\theta)$ is well defined and
$\gams{j-1}$ is continuous at $\theta.$
Thus, by the definition of $\gams{j}$ (Definition~\ref{PCgenerators})
and the continuity of $\gams{j-1}$ at $\theta,$
$\gams{j}(\theta) = \gams{j-1}(\theta).$
Now assume that
$\theta \in \Bd(\wbasint{\ell} \setminus \OBG{\ell}{j}) = \Bd(\BSG{\ell}{j}) \cup \Bd(\wbasint{\ell}).$
By (g), $\theta \notin \wbasint{n} \cup \wbasint{-n}$
for every $n > j.$
So, by the iterative use of the definition of
$\gams{i}$ for $i=j+1,j+2,\dots, n$ (Definition~\ref{PCgenerators})
we get
\[
\gams{j}(\theta) = \gams{j+1}(\theta) = \dots =
\gams{n-1}(\theta) = \gams{n}(\theta).
\]
Now we prove the part of (e) concerning $R_\omega(\wbasint{\ell} \setminus \OBG{\ell}{j}).$
We first assume that $\ell = j \ge 0.$ Then,
\[
\wbasint{j} = \basint{j},\
\theta \in \Bd(\basint{j})
\andq
R_\omega(\theta) \in \Bd(\BSG{j + 1}{j}).
\]
Again by Definition~\ref{PCgenerators}(\tsfR.2),
$R_\omega(\theta) \notin \Zstar_{j+1} \cup \basint{j} \cup \wbasint{-j}.$
So, by (a) and the definition of $\gams{j}$ (Definition~\ref{PCgenerators}),
$\gams{j-1}\bigl(R_\omega(\theta)\bigr)$ is well defined and
$\gams{j}\bigl(R_\omega(\theta)\bigr) = \gams{j-1}\bigl(R_\omega(\theta)\bigr).$
By Definition~\ref{PCgenerators}(\tsfR.3)
(with $j = n$ and $k = \ell = j$),
$R_\omega(\theta) \notin \basint{n} \cup \basintneg{n}$
for every $n > j.$
So,
$\gams{n}\bigl(R_\omega(\theta)\bigr) = \gams{j}\bigl(R_\omega(\theta)\bigr)$
as above.
Assume now that $\ell = -j < 0.$
In this case we have $\wbasint{\ell} = \BSG{\ell}{\abs{\ell+1}}$
and, hence,
$
R_\omega(\theta) \in \basintabs{\ell + 1} \setminus \OBG{\ell + 1}{j}.
$
By Definition~\ref{PCgenerators}(\tsfR.1) we have
\[
\BSG{\ell + 1}{j} \subset
\BSG{\ell+1}{\abs{\ell+1}} \subset
\wbasint{\ell+1}.
\]
Thus,
$R_\omega(\theta) \in \wbasint{\ell+1} \setminus \sstarset{\ell+1}.$
Again by Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2)
(with $j$ replaced by $-(\ell + 1)$),
\[
R_\omega(\theta) \notin
\Zstar_{\ell} \cup \BSG{-\ell}{-(\ell+1)} \cup \wbasint{\ell} \supset
\Zstar_{j} \cup \basint{j} \cup \wbasint{-j}.
\]
So, by (a) and the definition of $\gams{j}$ (Definition~\ref{PCgenerators}),
$\gams{j-1}\bigl(R_\omega(\theta)\bigr)$ is well defined and
$\gams{j}\bigl(R_\omega(\theta)\bigr) = \gams{j-1}\bigl(R_\omega(\theta)\bigr).$
To end the proof of (e), assume as above that
$\theta \in \Bd(\BSG{\ell}{j}) \cup \Bd(\wbasint{\ell})$
and, hence,
$R_\omega(\theta) \in \Bd(\BSG{\ell + 1}{j}) \cup \Bd\left(\basintabs{\ell + 1}\right).$
We have to show that, in this case,
$R_\omega(\theta) \notin \basint{n} \cup \basintneg{n}$
for every $n > j$
(the fact that $\gams{n}\bigl(R_\omega(\theta)\bigr) = \gams{j}\bigl(R_\omega(\theta)\bigr)$
follows as above).
When $R_\omega(\theta) \in \Bd(\BSG{\ell + 1}{j})$ this follows from
Definition~\ref{PCgenerators}(\tsfR.3) as before.
Assume now that
$R_\omega(\theta) \in \Bd\left(\basintabs{\ell + 1}\right).$
Then, by (g),
$R_\omega(\theta) \notin \wbasint{n} \cup \wbasint{-n}$ for every $n > j.$
\inidemopart{f} If $\ell \ge 0$ then the first two statements of (f)
follow directly from the definitions.
Moreover, by Remarks~\ref{propiedadesR}(2) and
\ref{PCgeneratorsExplicitConsequences}(\tsfR.1),
\[
\diam\left(\wbasicbox{\ell}\right) =
\diam\left(\basicbox{\ell}\right) =
\diam\left(\basicbox{-\ell}\right) =
2 \cdot 2^{-n_{\ell}} \le 2 \cdot 2^{-(\ell+1)} = 2^{-\ell}.
\]
Assume that $\ell < 0.$
From Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2)
we get
$\left(\wbasint{\ell} \setminus \obasintabs{\ell}\right) \cap \Zstar_{\all} = \emptyset$
and, hence, $\gams{\all}$
is continuous in an open neighbourhood of
$\wbasint{\ell} \setminus \obasintabs{\ell}$
by (a). On the other hand, by (d),
$\left(\theta,\gams{\all}(\theta)\right) \in \basicbox{\ell}$
for every
$
\theta \in \Bd\left(\basintabs{\ell}\right)\subset
\wbasint{\ell} \setminus \obasintabs{\ell}.
$
Thus,
\[
\wbasicbox{\ell} = \basicbox{\ell} \cup \Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)
\]
is closed, connected and projects onto the whole $\wbasint{\ell}.$
On the other hand, by (e) and (a)
(since $\ell < 0$, $\abs{\ell+1} = \all -1$),
\begin{align*}
\gams{\all}\left(\wbasint{\ell} \setminus \obasintabs{\ell}\right)
& = \gams{\all-1}\left(\BSG{\ell}{\abs{\ell+1}} \setminus \obasintabs{\ell}\right)\\
& \subset \left[
\gams{\all - 1}(\lstar) - 2^{-n_{\all - 1}},
\gams{\all - 1}(\lstar) + 2^{-n_{\all - 1}}
\right].
\end{align*}
Thus, by Remark~\ref{propiedadesR}(1), (c) and Definition~\ref{PCgenerators}(\tsfR.1),
\begin{align*}
\wbasicbox{\ell}
& = \basicbox{\ell} \cup \Graph\Bigl(\gams{\all}\evalat{\wbasint{\ell} \setminus \obasintabs{\ell}}\Bigr)\\
&\subset \basintabs{\ell} \times \left[\gams{\all - 1}(\lstar) - 2^{-n_{\all}}, \gams{\all - 1}(\lstar) + 2^{-n_{\all}}\right] \cup\\
& \hspace*{2em}\left(\BSG{\ell}{\abs{\ell+1}} \setminus \obasintabs{\ell}\right) \times \left[
\gams{\all - 1}(\lstar) - 2^{-n_{\all - 1}},
\gams{\all - 1}(\lstar) + 2^{-n_{\all - 1}}
\right] \\
&\subset \BSG{\ell}{\abs{\ell+1}} \times \left[
\gams{\all - 1}(\lstar) - 2^{-n_{\all - 1}},
\gams{\all - 1}(\lstar) + 2^{-n_{\all - 1}}
\right].
\end{align*}
Hence, by Definition~\ref{PCgenerators}(\tsfR.1) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.1),
\[
\diam\left(\wbasicbox{\ell}\right) \le
2\cdot \max\{\alpha_{\abs{\ell+1}}, 2^{-n_{\all - 1}}\}
= 2\cdot 2^{-n_{\all - 1}} \le 2 \cdot 2^{-{\all}}.
\]
\end{proof}
The next results allow us to define the limit pseudo-curve
generated by the sequence $\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty}.$
\begin{lemma}\label{convergencia}
The sequence
$\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty} \subset \C$
is convergent in $\C.$
\end{lemma}
\begin{proof}
By Proposition~\ref{Ccompleto} it suffices to show that
$\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty}$
is a Cauchy sequence in $\C.$
By the definition of $\gams{i}$ (Definition~\ref{PCgenerators}) we have
\begin{align*}
\dinf\left(\gams{i-1},\gams{i}\right)
& = \leftlimits{\sup}{\theta\in \SI\setminus \Zstar_{i}} \abs{\gams{i-1}(\theta)-\gams{i}(\theta)} \\
& = \leftlimits{\sup}{\theta \in \left(\basint{i} \setminus \istarset\right) \cup
\left(\basintneg{i}\setminus \sstarset{-i}\right)
} \abs{\gams{i-1}(\theta)-\gams{i}(\theta)}.
\end{align*}
By Lemmas~\ref{Propertiesvarphi}(c,d),
and Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
\[
(\theta,\gams{i-1}(\theta)), (\theta, \gams{i}(\theta)) \in \basicbox{\ell}
\andq[for]
\text{$\theta \in \BSG{\ell}{i} \setminus \lstarset$ and $\ell \in \{i,-i\}.$}
\]
Hence, by Lemma~\ref{Propertiesvarphi}(f),
\[
\dinf(\gams{i-1},\gams{i}) \le \diam(\basicbox{i}) =
\diam(\basicbox{-i}) \le 2^{-i}.
\]
Since $n_{i}$ is a strictly increasing sequence, for every $m \ge 0,$
\[
\dinf(\gams{i+m},\gams{i}) \le
\sum^{i+m}_{k=i+1} 2^{-k} <
2^{-(i+1)} \sum_{k=0}^{\infty} \tfrac{1}{2^k} = 2\cdot 2^{-(i+1)},
\]
and consequently
$
\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty}
$
is a Cauchy sequence in $\C.$
\end{proof}
Lemma~\ref{convergencia} allows us to define the following limit pseudo-curve generator
of the sequence $\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty}.$
\begin{definition}\label{gammalimit}
There exists $(\gamma,\SI\setminus \Orbom) \in \C$ such that
\[
(\gamma,\SI\setminus \Orbom) = \lim_{i\to\infty} (\gams{i},\SI\setminus \Zstar_{i})
\]
(that is,
$\gamma(\theta) = \lim_{i\to\infty} \gams{i}(\theta)$ for every $\theta \in \SI\setminus \Orbom$).
Observe that
\[
\SI\setminus \Orbom = \bigcap_{i=1}^{\infty} \left(\SI \setminus \Zstar_{i} \right)
\]
is a residual set in $\SI.$
\end{definition}
Now, we are ready to define the sequence of pseudo-curves associated to
the sequence
$\{(\gams{i},\SI\setminus \Zstar_{i})\}_{i=0}^{\infty},$
and to the limit pseudo-curve generator $(\gamma,\SI\setminus \Orbom).$
This will finally define the pseudo-curve $\A$ that we want to construct.
\begin{definition}\label{PCAtLast}
We denote by
\[
\A_{j} := \pc[\gams{j},\SI\setminus\Zstar_{j}] =
\overline{\Graph(\gams{j},\SI\setminus \Zstar_{j})}
\]
the pseudo-curve defined by
$(\gams{j},\SI\setminus \Zstar_{j}) \in \C$, and
\[
\A = \pc[\gamma,\SI\setminus \Orbom] := \overline{\Graph(\gamma,\SI\setminus \Orbom)}.
\]
By Definition~\ref{gammalimit} and Proposition~\ref{Hdimdinfty},
$
\A = \lim_{j\to\infty} \pc[\gams{j},\SI\setminus\Zstar_{j}].
$
\end{definition}
The next lemmas study the properties the pseudo-curves $\A_{j}$ and $\A.$
\begin{lemma}\label{propiedadesA}
The following statements hold for every $\ell\in \Z$:
\begin{enumerate}[(a)]
\item $\setfibth{\A_n} \subset \setfibth{\basicbox{\ell}}$
for every $n \ge \all-1$ and $\theta \in \basintabs{\ell}.$
\item $\setfibpt{\A_n}{\lstar} = \setfibpt{\A_{\all}}{\lstar} \subset \setfibbb{\ell}$
for every $n \ge \all.$
Moreover, $\setfibpt{\A_{\all}}{\lstar} = \setfibbb{\ell}$
is a non-degenerate interval.
\item $\setfibth{\A_{\ell}} = \{(\theta, \gams{\ell}(\theta)\}$
for every $\theta \in \SI\setminus\Zstar_{\ell}.$
\item $\A_{\all} \subset \SI \times [-1,1].$
\end{enumerate}
\end{lemma}
\begin{proof}
\inidemopart{a}
By Lemma~\ref{Propertiesvarphi}(c,d),
$
\Graph\Bigl(\gams{n}\evalat{\basintabs{\ell} \setminus \Zstar_n}\Bigr) \subset \basicbox{\ell}.
$
Then, the statement follows from the compacity of $\basicbox{\ell}$.
\inidemopart{b}
From the definition of $\gams{i}$ (Definition~\ref{PCgenerators}) and
Definition~\ref{PCgenerators}(\tsfR.2),
for every $n > \all$ there exists an $\varepsilon(n) > 0$ such that
$\gams{n}(\theta) = \gams{\all}(\theta)$ for every
$\theta \in \ball{\lstar}{\varepsilon(n)} \setminus \lstarset.$
Hence $\setfibpt{\A_n}{\lstar} = \setfibpt{\A_{\all}}{\lstar}.$
Moreover,
$\gams{\all}$ coincides with $\varphi_{_{\lstar}}$ in a neighbourhood of $\lstar.$
Thus, $\setfibpt{\A_{\all}}{\lstar} = \setfibbb{\ell}$ and it is an interval
by Definition~\ref{GenericBoxes} and Remark~\ref{propiedadesR}(4).
Finally statement (c) follows from
Lemma~\ref{PC-properties}(a) and Definition~\ref{PCAtLast},
and (d) from Lemma~\ref{Propertiesvarphi}(b).
\end{proof}
\begin{lemma}\label{propiedadesPCA}
The following statements hold.
\begin{enumerate}[(a)]
\item $\setfibth{\A} \subset \setfibth{\basicbox{\ell}}$
for every $\ell\in\Z$ and $\theta \in \basintabs{\ell}.$
\item $\setfibpt{\A}{\lstar} = \setfibpt{\A_{\all}}{\lstar}$
for every $\ell\in\Z.$
In particular $\setfibpt{\A}{\lstar}$ is a non-degenerate interval.
\item If $\theta\notin \Orbom,$ then
$\setfibth{\A} = \{(\theta,\gamma(\theta))\}.$
\item $\A \subset \SI \times [-1,1].$
\end{enumerate}
\end{lemma}
\begin{proof}
Statement (c) follows directly from Lemma~\ref{PC-properties}(a).
Now we prove (a).
From Lemma~\ref{propiedadesA}(a),
$\setfibth{\A_n} \subset \basicbox{\ell}$
for every $\ell \in \Z$ and $n \ge \all.$
On the other hand,
by Definition~\ref{gammalimit} and Proposition~\ref{Hdimdinfty},
$\setfibth{\A} = \lim_{n\to\infty} \setfibth{\A_n}.$
Hence the result follows from the compacity of $\basicbox{\ell}$.
By Lemma~\ref{propiedadesA}(b)
and the part of the lemma already proved we have
\[
\setfibpt{\A}{\lstar} = \lim_{n\to\infty} \setfibpt{\A_n}{\lstar} = \setfibpt{\A_{\all}}{\lstar}.
\]
Statement (d) follows from Lemma~\ref{propiedadesA}(d),
the compacity of $\SI \times [-1,1]$ and the fact that
$\A = \lim_{j\to\infty} \A_j.$
\end{proof}
The next proposition, summarizes the main properties of the set $\A.$
\begin{proposition}\label{teoremacentral}
The set $\A$ is a connected, does not contain any arc of curve and
$\Omega\setminus \A$ has two connected components.
\end{proposition}
\begin{proof}
From statements (b) and (c) of the previous lemma, we know that
$\setfibth{\A}$ is connected for every $\theta \in\SI.$
If $\A$ is not connected there exist closed (in $\A$) sets $U$ and $V$
such that $U \cap V = \emptyset$ and $U \cup V = \A$.
Observe that $\pi(U) \cup \pi(V) = \pi(\A) = \SI$
because every pseudo-curve is a circular set.
Moreover, since $\A$ is compact,
$U$ and $V$ are also compact sets of $\Omega$.
Hence, $\pi(U)$ are $\pi(V)$ compact in $\SI.$
Since {\SI} is connected, $\pi(U)\cap\pi(V) \ne \emptyset.$
For every $\theta \in \pi(U) \cap \pi(V)$ we have,
\[
\setfibth{\A} = \setfibth{(U \cup V)} = \setfibth{U} \cup \setfibth{V}.
\]
The sets $\setfibth{U}$ and $\setfibth{V}$ are closed, non-empty and disjoint.
Consequently, $\setfibth{\A}$ is not connected; a contradiction.
This proves that $\A$ is connected.
By Lemma~\ref{propiedadesPCA}(b),
$\setfibpt{\A}{\lstar}$ is a non-degenerate interval for every $\ell\in \Orbom.$
Then, since $\Orbom$ is dense in $\SI,$
$\A$ does not contain any arc of curve by
Lemma~\ref{PC-properties-invariant}(b).
To prove that $\Omega\setminus \A$ has two connected components we define
\begin{align*}
\Omega_{-} &:= \set{(\theta,y) \in\Omega}{y < \min \set{x \in \I}{(\theta,x) \in \A}}\text{, and}\\
\Omega_{+} &:= \set{(\theta,y) \in\Omega}{y > \max \set{x \in \I}{(\theta,x) \in \A}}.
\end{align*}
By Lemma~\ref{propiedadesPCA}(d) we know that
\[
-1 \le \min \set{x \in \I}{(\theta,x) \in \A} \le \max \set{x \in \I}{(\theta,x) \in \A} \le 1.
\]
Hence,
$\Omega \setminus \A = \Omega_{-} \cup \Omega_{+},$
$\Omega_{+}$ and $\Omega_{-}$ are disjoint open circular subsets of $\Omega$
and $\Omega_{-} \supset \SI \times [-2, -1]$ and $\Omega_{+} \supset \SI \times [1, 2]$
(in particular, for every $\theta\in \SI,$ $\setfibth{\Omega_{+}}$ and $\setfibth{\Omega_{-}}$ are
non-degenerate intervals).
Thus, $\Omega_{+}$ and $\Omega_{-}$ are arc-wise connected and, hence, connected.
\end{proof}
\section{A collection of auxiliary functions $G_i$ defined on the boxes $\protect\wbasicbox{i}$}\label{FunctionsGi}
In this section we define a family of auxiliary functions
{\map{G_i}{\basicbox{i}}[\Omega]} with $i \in \Z$ and study their properties.
In what follows we consider the supremum metric {\dinf} on the class of all functions
{\map{F}{A}[\Omega]} with $A \subset \Omega$. That is, given {\map{F,G}{A}[\Omega]}
we set
\[
\dinf(F,G) := \leftlimits{\sup}{(\theta,x) \in A} \dom(F(\theta,x), G(\theta,x)).
\]
In the special case when $F$ and $G$ are skew products with the same base,
that is when $F(\theta,x) = (R(\theta),f(\theta,x))$ and
$G(\theta,x) = (R(\theta),g(\theta,x)),$ then
\[
\dinf(F,G) :=
\leftlimits{\sup}{(\theta,x) \in A} \abs{f(\theta,x) - g(\theta,x)}.
\]
Observe that $(\cSO, \dinf)$ is a complete metric space.
Before defining the maps $G_i$ we need to introduce the necessary notation, and
recall and collect some basic facts that we will use in this definition
and to study their properties.
For every $i \in \Z,$ we define
\begin{align*}
& {\map{M_i}{\wbasint{i}}[\I]}
& \text{by}& \qquad
M_i(\theta) := \max \set{x \in \I}{(\theta,x) \in \wbasicbox{i}}, \text{ and}\\
& {\map{m_i}{\wbasint{i}}[\I]}
& \text{by} & \qquad
m_i(\theta) := \min \set{x \in \I}{(\theta,x) \in \wbasicbox{i}}.
\end{align*}
The next simple lemma states the basic properties of the maps $m_i$ and $M_i.$
\begin{lemma}\label{voresdelescaixesalesvores}
The following statements hold for every $i \in \Z$
\begin{enumerate}[(a)]
\item $-1 \le m_i(\theta) \le M_i(\theta) \le 1$ for every $\theta \in \wbasint{i}$.
\item $m_i $ and $M_i$ are continuous.
\item $m_i\evalat{\basintabs{i}}$ and $M_i\evalat{\basintabs{i}}$ are piecewise linear.
\item $m_i(\theta) = M_i(\theta) = \gams{\ai}(\theta)$ if and only if $\theta \in \wbasint{i} \setminus \obasintabs{i}.$
\end{enumerate}
\end{lemma}
\begin{proof}
It follows easily from Definition~\ref{GenericBoxes},
the definition of a winged region and
Lemma~\ref{Propertiesvarphi}(b,f).
\end{proof}
Notice that, for every $i \in \Z,$
\[
\wbasicbox{i} = \LSleftlimits{\bigcup}{\theta \in \wbasint{i}} \setfibth{\wbasicbox{i}}
= \LSleftlimits{\bigcup}{\theta \in \wbasint{i}} \{\theta\} \times [m_i(\theta), M_i(\theta)].
\]
In what follows the interval $[m_i(\theta), M_i(\theta)] \subset \I,$
defined for every $\theta \in \wbasint{i},$
will be denoted by $\I_{i,\theta}.$
Clearly, for every $\theta \in \wbasint{i},$
$\setfibth{\wbasicbox{i}} = \{\theta\} \times \I_{i,\theta}.$
By Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
\[
\wbasint{i} \setminus \istarset
\andq[is disjoint from]
\Zstar_{\ai}.
\]
Hence, Lemmas~\ref{Propertiesvarphi}(a,d) and \ref{propiedadesA}(c)
can be summarized as:
\begin{equation}\label{gammathetaproperties}
\begin{cases}
\gams{\all}\evalat{\wbasint{\ell} \setminus \lstarset}
\quad\text{is continuous,}\\
\gams{\all}(\theta) \in \I_{\ell,\theta}
\andq[for every]
\theta \in \wbasint{\ell} \setminus \lstarset,\text{ and}\\
\setfibth{\A_{\all}} = \{(\theta, \gams{\all}(\theta)\}
\andq[for every]
\theta \in \wbasint{\ell} \setminus \lstarset
\end{cases}
\end{equation}
for $\ell \in \{i, i+1\}.$
Now we define a family of continuous maps
{\map{G_i}{\wbasicbox{i}}[\Omega]} with $i\in \Z,$ by
\[
G_i(\theta,x) = \bigl(R_\omega(\theta), g_i(\theta,x)\bigr)
\]
Also, for every $\theta\in \wbasint{i},$
we will denote the map
{\map{g_i(\theta, \cdot)}{\I_{i,\theta}}[\I]}
by $g_{_{i,\theta}}.$
To define the functions $g_{_{i,\theta}},$ for clarity, we will consider separately two different situations:
\begin{itemize}
\item $i \ge 0,$ when $\wbasicbox{i} = \basicbox{i}$, $\wbasint{i} = \basintabs{i}$ and
$G_i(\basicbox{i})$ strictly contains the smaller box $\basicbox{i+1},$ and
\item $i \le -1,$ when $G_i(\wbasicbox{i})$ is strictly contained in the bigger box $\basicbox{i+1}.$
\end{itemize}
We start by defining $g_{_{i,\theta}}$ for $i \ge 0$
in three different ways, depending on the base point $\theta \in \basint{i}$.
In this definition, for simplicity we will use
$\basicbox{i}$ instead of $\wbasicbox{i}$ and
$\basintabs{i}$ instead of $\wbasint{i}.$
Notice that, by Definition~\ref{PCgenerators}(\tsfR.1) and Lemma~\ref{Propertiesvarphi}(c),
\begin{equation}\label{alphadeltaintervals}
\begin{split}
& \text{for every $i \ge 0$}\\
& \BSG[\delta]{i}{i+1} \subset \OBG{i}{i+1}
\andq
\BSG{i}{i+1} \subset \obasint[\delta]{i} \subset \obasint{i},\text{ and}\\
& \gams{i-1}(\istar) = a_i
\andq
\gams{i}(\sstar{i+1}) = a_{i+1}.
\end{split}
\end{equation}
\begin{definition}[\bfseries Definition of $\boldsymbol{g_{i}}$ for $\boldsymbol{i\ge 0}$]\label{defi-gi-positiva}\strut
\begin{labeledlist}{$\boldsymbol{\theta \in \cball{\istar}{\alpha_{i+1}} \setminus \ball{\istar}{\delta_{i+1}}}$}
\item [$\boldsymbol{\theta \in \cball{\istar}{\delta_{i+1}}}$]
$
g_{_{i,\theta}}(x) := \gams{i}(\sstar{i+1}) + \frac{2^{n_{i}}}{2^{n_{i+1}}}
\left(\gams{i-1}(\istar) - x\right).
$
\item [$\boldsymbol{\theta \in \cball{\istar}{\alpha_{i+1}} \setminus \ball{\istar}{\delta_{i+1}}}$]
we define $g_{_{i,\theta}}$ to be the unique piecewise affine map
with two affine pieces, defined on $\I_{i,\theta},$
whose graph joins
$\left(m_{i}(\theta), M_{i+1}\bigl(R_\omega(\theta)\bigr)\right)$
with
$\left(\gams{i}(\theta),\gams{i+1}\bigl(R_\omega(\theta)\bigr)\right),$
and this with the point
$\left(M_{i}(\theta), m_{i+1}\bigl(R_\omega(\theta)\bigr)\right)$
(in particular,
$g_{_{i,\theta}}\bigl(\gams{i}(\theta)\bigr)= \gams{i+1}\bigl(R_\omega(\theta)\bigr)$),
\item [$\boldsymbol{\theta \in \cball{\istar}{\alpha_{i}} \setminus \ball{\istar}{\alpha_{i+1}}}$]
$g_{_{i,\theta}}(x) := \gams{i+1}\bigl(R_\omega(\theta)\bigr)$
(that is, $g_{_{i,\theta}}$ is constant).
\end{labeledlist}
\end{definition}
The next lemma states the basic properties of the functions $G_i$ for $i \ge 0.$
\begin{lemma}\label{gpositiva}
The following statements hold for every $i \ge 0:$
\begin{enumerate}[(a)]
\item The map $g_{_{i,\theta}}$ is well defined and non-increasing for every $\theta \in \basint{i}.$
Moreover, $-1 \le g_{_{i,\theta}}(x) \le 1$ for every $\theta \in \basint{i}$ and $x \in \I_{i,\theta}.$
Furthermore, the function $G_i$ is continuous.
\item $G_i\evalat{\setfibth{\basicbox{i}}}$ is affine and
$G_i\bigl(\setfibth{\basicbox{i}}\bigr) = \setfibpt{\basicbox{i+1}}{R_\omega(\theta)}$
for every $\theta \in \BSG[\delta]{i}{i+1};$
$G_i\evalat{\setfibth{\basicbox{i}}}$ is piecewise affine with two pieces and
$G_i\bigl(\setfibth{\basicbox{i}}\bigr) = \setfibpt{\basicbox{i+1}}{R_\omega(\theta)}$
for every $\theta \in \BSG{i}{i+1} \setminus \OBG[\delta]{i}{i+1};$ and\\
$G_i\bigl(\setfibth{\basicbox{i}}\bigr) = \setfibpt{\A_{i+1}}{R_\omega(\theta)}$
for every
$\theta \in \basint{i} \setminus \OBG{i}{i+1}.$
\item $G_i(\setfibth{\A_{i}}) = \setfibpt{\A_{i+1}}{R_\omega(\theta)}$ for every $\theta \in \basint{i}.$
\end{enumerate}
\end{lemma}
\begin{proof}
We will prove all statements of the lemma simultaneously and according to the regions
in the definition of the map $g_i.$
\inidemopartfree{\textbullet} We start with the region $\setfibball[\delta]{i}{i+1}.$\\[\medskipamount]
Let $z \in [-\delta_{i}, \delta_{i}] \subset \R$ and let
$\theta = \istar + z \in \basint[\delta]{i}.$
From Definition~\ref{GenericBoxes} and \eqref{alphadeltaintervals} we get
\begin{equation}\label{mMformula}
\begin{split}
m_i(\theta) &= a_i - 2^{-{n_i}}(1-z) =
\gams{i-1}(\istar) - 2^{-{n_i}}(1-z), \text{ and}\\
M_i(\theta) &= a_i + 2^{-{n_i}}(1-z) =
\gams{i-1}(\istar) + 2^{-{n_i}}(1-z).
\end{split}
\end{equation}
In a similar way, for every $\theta \in \BSG[\delta]{i}{i+1}$
(that is, $z \in [-\delta_{i+1}, \delta_{i+1}]$), we have
$R_\omega(\theta) = \sstar{i+1} + z \in \basint[\delta]{i+1},$
and
\begin{equation}\label{mMRotformula}
\begin{split}
m_{i+1}(R_\omega(\theta)) &= a_{i+1} - 2^{-{n_{i+1}}}(1-z) =
\gams{i}(\sstar{i+1}) - 2^{-{n_{i+1}}}(1-z), \text{ and}\\
M_{i+1}(R_\omega(\theta)) &= a_{i+1} + 2^{-{n_{i+1}}}(1-z) =
\gams{i}(\sstar{i+1}) + 2^{-{n_{i+1}}}(1-z).
\end{split}
\end{equation}
Hence, for every $\theta \in \BSG[\delta]{i}{i+1},$
\begin{equation}\label{mappingendpoints}
\begin{split}
g_{_{i,\theta}}(m_i(\theta))
&= \gams{i}(\sstar{i+1}) +
\tfrac{2^{n_{i}}}{2^{n_{i+1}}} 2^{-{n_i}} (1-z)
= \gams{i}(\sstar{i+1}) + 2^{-{n_{i+1}}} (1-z)\\
&= M_{i+1}(R_\omega(\theta)),\\
g_{_{i,\theta}}(M_i(\theta))
&= \gams{i}(\sstar{i+1}) -
\tfrac{2^{n_{i}}}{2^{n_{i+1}}} 2^{-{n_i}} (1-z)
= \gams{i}(\sstar{i+1}) - 2^{-{n_{i+1}}} (1-z)\\
&= m_{i+1}(R_\omega(\theta)).
\end{split}
\end{equation}
So,
$g_{_{i,\theta}}\evalat{\I_{i, \theta}}$
is the affine map whose graph joins the point
$\bigl(m_{i}(\theta), M_{i+1}(R_\omega(\theta))\bigr)$ with
$\bigl(M_{i}(\theta), m_{i+1}(R_\omega(\theta))\bigr).$
In particular, $g_{_{i,\theta}}$ sends the interval
$\I_{i, \theta}$ affinely onto
$\I_{i+1, R_\omega(\theta)}$ or, equivalently,
$G_i$ sends the interval $\setfibth{\basicbox{i}}$ affinely onto
$\setfibpt{\basicbox{i+1}}{R_\omega(\theta)}.$
Then, by Lemma~\ref{Propertiesvarphi}(b), this implies that
$-1 \le g_{_{i,\theta}}(x) \le 1$ for every $x \in \I_{i,\theta}.$
Moreover, the continuity of the maps $m_i,\ M_i,\ m_{i+1} \circ R_\omega$
and $M_{i+1} \circ R_\omega$ imply that $g_{i}$ is well defined
and continuous on
$\setfibball[\delta]{i}{i+1}$
Next we will prove that
$
G_i\bigl(\setfibth{\A_{i}}\bigr) = \setfibpt{\A_{i+1}}{R_\omega(\theta)}
$
for every $\theta \in \BSG[\delta]{i}{i+1}.$
We take
$\theta = \istar + z \in \BSG[\delta]{i}{i+1}\setminus\istarset.$
Then, clearly, $z \in [-\delta_{i+1}, \delta_{i+1}]\setminus\{0\} \subset \R.$
By Definitions~\ref{PCgenerators} and \ref{GenericBoxes} and statement \eqref{alphadeltaintervals},
\begin{align*}
\gams{i}(\theta)
&= \varphi_{_{\istar}}(\theta) =
a_i + 2^{-n_i} d =
\gams{i-1}(\istar) + 2^{-n_i} d \in \I_{i, \theta}, \text{ and}\\
\gams{i+1}(R_\omega(\theta))
&= \varphi_{_{\sstar{i+1}}}(\theta) =
a_{i+1} - 2^{-n_{i+1}} d =
\gams{i-1}(\istar) - 2^{-n_{i+1}} d \in \I_{i+1, R_\omega(\theta)},
\end{align*}
where $d = (-1)^{i} \phi(z).$
So, for every
$\theta \in \BSG[\delta]{i}{i+1}\setminus\istarset,$
\begin{equation}\label{Formula1}
g_{_{i,\theta}}(\gams{i}(\theta))
= \gams{i}(\sstar{i+1}) - \frac{2^{n_{i}}}{2^{n_{i+1}}} 2^{-n_i} d
= \gams{i+1}(R_\omega(\theta)).
\end{equation}
Thus, from \eqref{alphadeltaintervals} and \eqref{gammathetaproperties} we get
\begin{align*}
G_i\bigl(\setfibth{\A_{i}}\bigr)
&= G_i\bigl(\{(\theta, \gams{i}(\theta))\}\bigr)
=\{(R_\omega(\theta), g_{_{i,\theta}}(\gams{i}(\theta)))\}\\
&= \{(R_\omega(\theta), \gams{i+1}(R_\omega(\theta)))\}
= \setfibpt{\A_{i+1}}{R_\omega(\theta)}
\end{align*}
for every
$\theta \in \BSG[\delta]{i}{i+1}\setminus\istarset.$
On the other hand, by the part already proven,
$g_{_{i,\istar}}$ sends the interval
$\I_{i,\istar}$ affinely to
$\I_{i+1,\sstar{i+1}}$
or, equivalently, $G_i$ sends the interval
$
\setfibbb{i} = \istarset \times \I_{i,\istar}
$
affinely onto
$
\setfibbb{i+1} = \iistarset \times \I_{i,\sstar{i+1}}.
$
This implies that
$
G_i\bigl(\setfibpt{\A_{i}}{\istar}\bigr) = \setfibpt{\A_{i+1}}{\sstar{i+1}}
$
by Lemma~\ref{propiedadesA}(b).
Hence,
$
G_i\bigl(\setfibth{\A_{i}}\bigr) = \setfibpt{\A_{i+1}}{R_\omega(\theta)}
$
for every $\theta \in \BSG[\delta]{i}{i+1}.$
\inidemopartfree{\textbullet} Now we study
$
\setfib{\basicbox{i}}{\left(\BSG{i}{i+1} \setminus
\OBG[\delta]{i}{i+1}\right)}.
$\\[\medskipamount]
Observe that
$
R_\omega(\cball{\istar}{\alpha} \setminus \istarset) =
\cball{\sstar{i+1}}{\alpha} \setminus \iistarset
$
for $\alpha \in \{\alpha_i, \alpha_{i+1}\}.$
Then, by \eqref{gammathetaproperties}
\begin{equation}\label{gammaRotthetaontoca}
\begin{split}
& \gams{i+1} \circ R_\omega\evalat{\basint{i} \setminus \istarset}
\quad\text{is continuous, and}\\
& \gams{i+1}(R_\omega(\theta)) \in \I_{i+1, R_\omega(\theta)}
\andq[for every]
\theta \in \BSG{i}{i+1} \setminus \istarset.
\end{split}
\end{equation}
So, the continuity of the maps
$m_i,\ M_i,\ m_{i+1} \circ R_\omega$ and $M_{i+1} \circ R_\omega$
imply that $g_{i}$ is well defined and continuous on
$
\setfib{\basicbox{i}}{\left(\BSG{i}{i+1} \setminus
\OBG[\delta]{i}{i+1}\right)},
$
and
\[
\bigl(\gams{i}(\theta),\gams{i+1}(R_\omega(\theta))\bigr) \in
\I_{i, \theta}
\times
\I_{i+1, R_\omega(\theta)}
\]
for every
$\theta \in \BSG{i}{i+1} \setminus \OBG[\delta]{i}{i+1}.$
Consequently, $g_{_{i,\theta}}$ maps
$\I_{i, \theta}$ piecewise affinely with two pieces onto
$\I_{i+1, R_\omega(\theta)}$ or, equivalently,
$G_i$ sends the interval $\setfibth{\basicbox{i}}$
piecewise affinely with two pieces onto
$\setfibpt{\basicbox{i+1}}{R_\omega(\theta)}.$
Again, by Lemma~\ref{Propertiesvarphi}(b), this implies that
$-1 \le g_{_{i,\theta}}(x) \le 1$ for every $x \in \I_{i,\theta}.$
On the other hand,
from \eqref{alphadeltaintervals} and \eqref{gammathetaproperties} we have
\begin{align*}
G_i\bigl(\setfibth{\A_{i}}\bigr)
&= G_i\bigl(\{(\theta, \gams{i}(\theta))\}\bigr)
=\{(R_\omega(\theta), g_{_{i,\theta}}(\gams{i}(\theta)))\}\\
&= \{(R_\omega(\theta), \gams{i+1}(R_\omega(\theta)))\}
= \setfibpt{\A_{i+1}}{R_\omega(\theta)}
\end{align*}
for every $\theta \in \BSG{i}{i+1} \setminus \OBG[\delta]{i}{i+1}.$
\inidemopartfree{\textbullet} Finally, we study the region
$
\setfib{\basicbox{i}}{\left(\basint{i} \setminus
\OBG{i}{i+1}\right)}.
$\\[\medskipamount]
In this case, by definition and Lemma~\ref{Propertiesvarphi}(b) we have
$-1 \le g_{_{i,\theta}}(x) \le 1$ for every $x \in \I_{i,\theta}.$
By \eqref{gammaRotthetaontoca},
$g_{i}(\cdot, x) = \gams{i+1} \circ R_\omega$ is well defined and continuous in
both variables on
$
\setfib{\basicbox{i}}{\left(\basint{i} \setminus
\OBG{i}{i+1}\right)}
$
because $m_i$ and $M_i$ are continuous.
Moreover, for every
$\theta \in \basint{i} \setminus \OBG{i}{i+1}$
and $x$ such that $(\theta,x) \in \setfibth{\basicbox{i}},$
we have
\[
\{G_i(\theta, x)\}
= \{(R_\omega(\theta), g_{i}(\theta, x))\} =
\{(R_\omega(\theta), \gams{i+1}(R_\omega(\theta))\} =
\setfibpt{\A_{i+1}}{R_\omega(\theta)}
\]
by Definition~\ref{PCAtLast} and Lemma~\ref{PC-properties}(a).
Thus, by Lemma~\ref{propiedadesA}(a),
\[
G_i\bigl(\setfibth{\A_{i}}\bigr) =
G_i\bigl(\setfibth{\basicbox{i}}\bigr) =
\setfibpt{\A_{i+1}}{R_\omega(\theta)}.
\]
From all the previous arguments (b) and (c) follow.
To end the proof of (a) we have to see that $G_i$ is well defined and globally continuous.
This amounts to show that it is well defined on the fibres
\begin{align*}
\setfibpt{\basicbox{i}}{(\istar \pm \delta_{i+1})}
&= \{\istar \pm \delta_{i+1}\} \times \I_{i, \istar \pm \delta_{i+1}} \text{ and}\\
\setfibpt{\basicbox{i}}{(\istar \pm \alpha_{i+1})}
&= \{\istar \pm \alpha_{i+1}\} \times \I_{i, \istar \pm \alpha_{i+1}}.
\end{align*}
We will only show that the two definitions of $g_{i}$ coincide on
$\{\theta\} \times \I_{i, \theta}$
with $\theta \in \{\istar + \delta_{i+1}, \istar + \alpha_{i+1}\}.$
The case $\theta \in \{\istar - \delta_{i+1}, \istar - \alpha_{i+1}\}$
follows analogously.
We start with $\theta = \istar + \alpha_{i+1} \in \obasint[\delta]{i}.$
In this case,
$R_\omega(\theta) = \sstar{i+1} + \alpha_{i+1} \in \Bd(\basint{i+1})$ and,
by Definition~\ref{GenericBoxes} and Lemma~\ref{Propertiesvarphi}(c),
\[
M_{i+1}(R_\omega(\theta)) =
m_{i+1}(R_\omega(\theta)) = a^+_{i+1} =
\gams{i+1}(R_\omega(\theta)).
\]
Thus, the piecewise affine map whose graph joins the points
$\bigl(m_{i}(\theta), M_{i+1}(R_\omega(\theta))\bigr),$
$\bigl(\gams{i}(\theta),\gams{i+1}(R_\omega(\theta))\bigr),$ and
$\bigl(M_{i}(\theta), m_{i+1}(R_\omega(\theta))\bigr)$
is the constant map $\gams{i+1}(R_\omega(\theta)).$
Hence, $g_{_{i,\theta}}$ is well defined
for $\theta = \istar + \alpha_{i+1}.$
Now we deal with the case
$\theta = \istar + \delta_{i+1} \in \basint[\delta]{i}.$
By \eqref{mappingendpoints} and \eqref{Formula1} we know that the points
$\bigl(m_{i}(\theta), M_{i+1}(R_\omega(\theta))\bigr),$
$\bigl(\gams{i}(\theta),\gams{i+1}(R_\omega(\theta))\bigr)$ and
$\bigl(M_{i}(\theta), m_{i+1}(R_\omega(\theta))\bigr)$
belong to
$
\Graph\left(x \mapsto \gams{i}(\sstar{i+1}) +
\frac{2^{n_{i}}}{2^{n_{i+1}}} \left(\gams{i-1}(\istar) - x\right)
\right).
$
Consequently, the map
$
\gams{i}(\sstar{i+1}) +
\frac{2^{n_{i}}}{2^{n_{i+1}}} \left(\gams{i-1}(\istar) - x\right)
$
coincides with the piecewise affine map whose graph joins
$\bigl(m_{i}(\theta), M_{i+1}(R_\omega(\theta))\bigr),$
$\bigl(\gams{i}(\theta),\gams{i+1}(R_\omega(\theta))\bigr)$ and
$\bigl(M_{i}(\theta), m_{i+1}(R_\omega(\theta))\bigr).$
This ends the proof of (a).
\end{proof}
Now we define $g_{_{i,\theta}}$ for $i < 0.$
In this case, since we are going from a smaller box $\wbasicbox{i}$
to a bigger one, we only need to define $g_{_{i,\theta}}$ in two
different ways, depending on the base point $\theta \in \wbasint{i}$.
As in the previous case we need to fix some facts about the elements
that we will use in the definition.
By Definition~\ref{PCgenerators}(\tsfR.1) and Lemma~\ref{Propertiesvarphi}(c),
\begin{equation}\label{alphadeltaintervalsineg}
\begin{split}
& \text{for every $i < 0$\hspace*{1.5em}}\\
& \BSG[\delta]{i+1}{\ai} \subset\
\BSG{i+1}{\ai} \subset
\OBG[\delta]{i+1}{\aii} \subset
\obasintabs{i+1},\\
& R_\omega\left(\wbasint{i}\right) = \basintabs{i+1},\quad
\basintabs[\delta]{i} \subset \obasintabs{i}, \text{ and}\\
& \gams{\aii}(\istar) = a_i \quad
\text{and}\quad
\gams{\abs{i+2}}(\sstar{i+1}) = a_{i+1}.
\end{split}
\end{equation}
Consequently, from \eqref{gammathetaproperties}
and Definitions~\ref{GenericBoxes} and \ref{PCgenerators} we get
\begin{align*}
& m_{i}(\theta) < \gams{\ai}(\theta) < M_{i}(\theta) \text{ and}\\
& m_{i+1}(R_\omega(\theta)) < \gams{\aii}(R_\omega(\theta)) < M_{i+1}(R_\omega(\theta))
\end{align*}
for every $\theta \in \obasintabs{i} \setminus \istarset$
(and $R_\omega(\theta) \in \OBG{i+1}{\ai} \setminus \iistarset$).
Then,
\[
\widetilde{\kappa}_i(\theta) = \min \left\{1,
\frac{m_{i+1}(R_\omega(\theta)) - \gams{\aii}(R_\omega(\theta))
}{\tfrac{2^{n_{\ai}}}{2^{n_{_{\aii}}}} (\gams{\ai} (\theta) - M_{i}(\theta))},
\frac{ M_{i+1}(R_\omega(\theta)) - \gams{\aii}(R_\omega(\theta))
}{\tfrac{2^{n_{\ai}}}{2^{n_{_{\aii}}}} (\gams{\ai}(\theta) - m_{i}(\theta))}
\right\} > 0
\]
defines a continuous function
$
\map{\widetilde{\kappa}_i}{\obasintabs{i} \setminus \obasintabs[\delta]{i}}[{(0,1]}].
$
To define the map $g_{i}$ we need an auxiliary function
\[
\map{\kappa_{i}}{\basintabs{i} \setminus \obasintabs[\delta]{i}}[{[0,1]}]
\]
such that
$\kappa_{i}$ is non-decreasing and continuous,
$\kappa_{i}(\istar\pm \delta_{\ai}) = \widetilde{\kappa}_{i}(\istar\pm \delta_{\ai}),$ and
$\kappa_{i}(\theta) \le \widetilde{\kappa}_{i}(\theta)$ for every
$\theta \in \obasintabs{i} \setminus \obasintabs[\delta]{i}.$
In principle any such function would do, but for definiteness, and to show that such function exists,
we note that we can take, for instance,
\[
\kappa_{i}(\theta) = \begin{cases}
\inf_{t \in [\theta, \istar - \delta_{\ai}] \cap \obasintabs{i}} \widetilde{\kappa}_i(t)
& \text{if $\theta \le \istar - \delta_{\ai}$},\\
\inf_{t \in [\istar + \delta_{\ai}, \theta] \cap \obasintabs{i}} \widetilde{\kappa}_i(t)
& \text{if $\theta \ge \istar + \delta_{\ai}$}.
\end{cases}
\]
It is easy to check that this map verifies the desired properties.
\begin{definition}[\bfseries Definition of $\boldsymbol{g_{i}}$ for $\boldsymbol{i< 0}$]\label{defi-gi-negativa}
For every $(\theta,x) \in \wbasicbox{i}$ we set
\[
g_{_{i,\theta}}(x) := \begin{cases}
\frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}}
\left(\gams{\aii}(\istar) - x\right)
+ \gams{\abs{i+2}}(\sstar{i+1})
& \text{if $\theta \in \basintabs[\delta]{i},$}\\
\frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}} \kappa_{i}(\theta)
\left(\gams{\ai}(\theta) - x\right)
+ \gams{\aii}(R_\omega(\theta))
&\text{if $\theta \in \basintabs{i} \setminus \obasintabs[\delta]{i}$}\\
\gams{\aii}(R_\omega(\theta))
&\text{if $\theta \in \wbasint{i} \setminus \obasintabs{i}.$}
\end{cases}
\]
\end{definition}
The next lemma states the basic properties of the functions $G_i$ for $i < 0.$
\begin{lemma}\label{gnegativa}
The following statements hold for every $i < 0:$
\begin{enumerate}[(a)]
\item The map $g_{_{i,\theta}}$ is well defined and non-increasing for every $\theta \in \basint{i}.$
Moreover, $-1 \le g_{_{i,\theta}}(x) \le 1$ for every $\theta \in \basint{i}$ and $x \in \I_{i,\theta}.$
Furthermore, the function $G_i$ is continuous.
\item $G_i\evalat{\setfibth{\wbasicbox{i}}}$ is affine,
$G_i\bigl(\setfibth{\wbasicbox{i}}\bigr) \subset \setfibpt{\basicbox{i+1}}{R_\omega(\theta)}$
for every $\theta \in \wbasint{i}$ and
$G_i\bigl(\setfibth{\wbasicbox{i}}\bigr) = \setfibpt{\basicbox{i+1}}{R_\omega(\theta)}$
for every $\theta \in \basintabs[\delta]{i}.$
\item $G_i(\setfibth{\A_{\ai}}) = \setfibpt{\A_{\aii}}{R_\omega(\theta)}$ for every $\theta \in \wbasint{i}.$
\end{enumerate}
\end{lemma}
\begin{proof}
First we will prove that the map $G_i$ is continuous and that
$G_i\evalat{\setfibth{\wbasicbox{i}}}$ is affine, according to the
three regions in the definition.
\inidemopartfree{\textbullet}
As in the previous lemma we start with
$\setfib{\wbasicbox{i}}{\basintabs[\delta]{i}} =
\setfibball[\delta]{i}{\ai}.$\\[\medskipamount]
As in the same case of Lemma~\ref{gpositiva}, by using
\eqref{alphadeltaintervalsineg} instead of \eqref{alphadeltaintervals},
it follows that
$g_{_{i,\theta}}\evalat{\I_{i, \theta}}$
is the affine map whose graph joins the points
$\bigl(m_{i}(\theta), M_{i+1}(R_\omega(\theta))\bigr)$ and
$\bigl(M_{i}(\theta), m_{i+1}(R_\omega(\theta))\bigr),$
$g_{i}$ is well defined and continuous on
$\setfibball[\delta]{i}{\ai},$
\begin{align*}
& g_{_{i,\theta}}(\gams{\ai}(\theta))
= \gams{\aii}(R_\omega(\theta))
\text{ for every $\theta \in \basintabs[\delta]{i}\setminus\istarset,$}\\
& \text{$G_i$ sends the interval $\setfibth{\basicbox{i}}$ affinely onto
$\setfibth{\basicbox{i+1}}$, and}\\
& G_i\bigl(\setfibth{\A_{\ai}}\bigr)
= \setfibpt{\A_{\aii}}{R_\omega(\theta)}\quad
\text{for every $\theta \in \basintabs[\delta]{i}$}.
\end{align*}
\inidemopartfree{\textbullet}
$
\setfib{\wbasicbox{i}}{\bigl(\basintabs{i} \setminus \obasintabs[\delta]{i}\bigr)} =
\setfib{\basicbox{i}}{\bigl(\basintabs{i} \setminus \obasintabs[\delta]{i}\bigr)}.
$\\[\medskipamount]
From \eqref{gammathetaproperties} we know that the maps
$\gams{\ai}$ and $\gams{\aii} \circ R_\omega$
are continuous on the domain
$\basintabs{i} \setminus \obasintabs[\delta]{i}.$
Hence, the continuity of $g_{i}$ follows from the continuity of the maps
$\kappa_i$, $m_i,\ M_i,\ m_{i+1} \circ R_\omega$ and $M_{i+1} \circ R_\omega.$
Notice that, from the definition of $g_{i}$ in this region we clearly have that
\[
g_{_{i,\theta}}(\gams{\ai}(\theta)) = \gams{\aii}(R_\omega(\theta)), \text{ and}
\]
$G_i\evalat{\setfibth{\wbasicbox{i}}} = g_i(\theta, \cdot)$ is affine.
\inidemopartfree{\textbullet}
$
\setfib{\wbasicbox{i}}{\bigl(\wbasint{i} \setminus \obasintabs{i}\bigr)}.
$\\[\medskipamount]
In this case we have $m_{i}(\theta) = \gams{\ai}(\theta) = M_{i}(\theta)$
by definition. Then, the map
$G_i\evalat{\setfibth{\wbasicbox{i}}} = g_i(\theta, \cdot)$ is affine because
it is constant, and $g_i$ is continuous because
$\gams{\ai}$ and $\gams{\aii} \circ R_\omega$
are continuous on the domain $\wbasint{i} \setminus \istarset$
by \eqref{gammathetaproperties}.
To end the proof of (a) we have to see that $G_i$ is well defined
and globally continuous.
This amounts to show that it is well defined on the fibres
\[
\setfibpt{\basicbox{i}}{(\istar \pm \delta_{\ai})}
\andq
\setfibpt{\basicbox{i}}{(\istar \pm \alpha_{\ai})}
\]
We start by showing that the two definitions of $g_{i}$ coincide on
the fibres
$\setfibth{\basicbox{i}}$ for $\theta \in \{\istar \pm \alpha_{\ai}\}.$
In this case we have $m_{i}(\theta) = \gams{\ai}(\theta) = M_{i}(\theta).$
Consequently, $\I_{i, \theta} = \{\gams{\ai}(\theta)\}$ and
\[
\frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}} \kappa_{i}(\theta)
\left(\gams{\ai}(\theta) - x\right)
+ \gams{\aii}(R_\omega(\theta)) =
\gams{\aii}(R_\omega(\theta))
\]
for $x \in \I_{i, \theta}.$
Next we consider
$\setfibth{\basicbox{i}} = \{\theta\} \times \I_{i,\theta}$ with
$\theta = \istar + \delta_{\ai}.$
We will show that the two definitions of $g_{i}$ coincide on this set.
The case $\theta = \istar - \delta_{\ai}$ follows analogously.
For simplicity we will denote
\begin{align*}
g^{\delta_{\ai}}_{_{i,\theta}}(x) &:= \frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}}
\left(\gams{\aii}(\istar) - x\right)
+ \gams{\abs{i+2}}(\sstar{i+1}),\text{ and}\\
\xi_{_{i,\theta}}(x) &:= \frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}}
\left(\gams{\ai}(\theta) - x\right)
+ \gams{\aii}(R_\omega(\theta)).
\end{align*}
Notice that $g^{\delta_{\ai}}_{_{i,\theta}}$ is the map
$g_{_{i,\theta}}$ as defined in the first region while
\[
\kappa_i(\theta)\left( \xi_{_{i,\theta}} - \gams{\aii}(R_\omega(\theta)) \right)
+ \gams{\aii}(R_\omega(\theta))
\]
is the map $g_{_{i,\theta}}$ as defined in the second region.
In a similar way to the previous lemma we have that
$
\bigl(\gams{\ai}(\theta),\gams{\aii}(R_\omega(\theta))\bigr)
\in
\Graph\left(g^{\delta_{\ai}}_{_{i,\theta}}\right).
$
Hence, since $g^{\delta_{\ai}}_{_{i,\theta}}$ is affine with slope
$-\tfrac{2^{n_{\ai}}}{2^{n_{_{\aii}}}},$
it follows that
$g^{\delta_{\ai}}_{_{i,\theta}} = \xi_{_{i,\theta}}.$
So, to end the proof of the lemma, we only have to see that
$
\kappa_{i}(\istar + \delta_{\ai}) =
\widetilde{\kappa}_{i}(\istar + \delta_{\ai}) = 1.
$
Since the points
$\bigl(m_{i}(\theta), M_{i+1}(R_\omega(\theta))\bigr)$ and
$\bigl(M_{i}(\theta), m_{i+1}(R_\omega(\theta))\bigr)$
also belong to
$
\Graph\left(g^{\delta_{\ai}}_{_{i,\theta}}\right) =
\Graph\left(\xi_{_{i,\theta}}\right),
$
it follows that
\begin{align*}
m_{i+1}(R_\omega(\theta))
&= \xi_{_{i,\theta}}(M_{i}(\theta))
= \frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}}
\left(\gams{\ai}(\theta) - M_{i}(\theta)\right)
+ \gams{\aii}(R_\omega(\theta)), \text{ and}\\
M_{i+1}(R_\omega(\theta))
&= \xi_{_{i,\theta}}(m_{i}(\theta))
= \frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}}
\left(\gams{\ai}(\theta) - m_{i}(\theta)\right)
+ \gams{\aii}(R_\omega(\theta)).
\end{align*}
This shows that
$
\widetilde{\kappa}_{i}(\istar + \delta_{\ai}) =
\widetilde{\kappa}_{i}(\theta) = 1
$
and ends the proof of (a).
Now we prove (b) according to the three regions in the definition.
From the part of the lemma already proven we already know that
$G_i\evalat{\setfibth{\wbasicbox{i}}}$ is affine, and
$
G_i\bigl(\setfibth{\wbasicbox{i}}\bigr) =
\setfibpt{\basicbox{i+1}}{R_\omega(\theta)}
$
for every $\theta \in \basintabs[\delta]{i}.$
So, to end the proof of (b) we have to see that
\begin{equation}\label{fibinclu}
g_{_{i,\theta}}(\I_{i, \theta}) \subset \I_{i+1, R_\omega(\theta)}
\end{equation}
for every $\theta \in \wbasint{i} \setminus \basintabs[\delta]{i}$
(by definition, since $i < 0,$ $\wbasint{i} = \BSG{i}{\aii};$
therefore,
$R_\omega(\theta) \in \basintabs{i+1}$ and
$\I_{i+1, R_\omega(\theta)} = \setfibpt{\basicbox{i+1}}{R_\omega(\theta)}$).
For $\theta \in \wbasint{i} \setminus \obasintabs{i},$
by \eqref{gammathetaproperties}, we have
\[
g_{_{i,\theta}}(\I_{i, \theta})
= \{\gams{\aii}(R_\omega(\theta))\}
\subset \I_{i+1, R_\omega(\theta)}.
\]
Now we consider
$\theta \in \obasintabs{i} \setminus \basintabs[\delta]{i}.$
Since
\[
\kappa_{i}(\theta) \le \widetilde{\kappa}_{i}(\theta) \le
\frac{ M_{i+1}(R_\omega(\theta))-\gams{\aii}(R_\omega(\theta))
}{ \tfrac{2^{n_{\ai}}}{2^{n_{_{\aii}}}} (\gams{\ai}(\theta)-m_{i}(\theta))},
\]
we have
\begin{align*}
g_{_{i,\theta}}(m_{i}(\theta))
&\le \frac{2^{n_{\ai}}}{2^{n_{_{\aii}}}}
\frac{M_{i+1}(R_\omega(\theta))-\gams{\aii}(R_\omega(\theta))
}{\tfrac{2^{n_{\ai}}}{2^{n_{_{\aii}}}} (\gams{\ai}(\theta)-m_{i}(\theta))}
\left(\gams{\ai}(\theta) - m_{i}(\theta)\right)
+ \gams{\aii}(R_\omega(\theta))\\
&= M_{i+1}(R_\omega(\theta)).
\end{align*}
An analogous computation shows that
$g_{_{i,\theta}}(M_{i}(\theta)) \ge m_{i+1}(R_\omega(\theta)).$
Hence, \eqref{fibinclu} holds because $g_{_{i,\theta}}$ is affine.
This ends the proof of (b).
Then, by Lemma~\ref{Propertiesvarphi}(b), Statement~(b) of the lemma
implies that $-1 \le g_{_{i,\theta}}(x) \le 1$ for every $x \in \I_{i,\theta}.$
By the part of the lemma already proved we know that
$
G_i\bigl(\setfibth{\A_{\ai}}\bigr)
= \setfibpt{\A_{\aii}}{R_\omega(\theta)}
$
for every $\theta \in \basintabs[\delta]{i}$.
On the other hand, as in the previous lemma,
from \eqref{alphadeltaintervalsineg} and \eqref{gammathetaproperties}
we get
\begin{align*}
G_i\bigl(\setfibth{\A_{\ai}}\bigr)
&= G_i\bigl(\{(\theta, \gams{\ai}(\theta))\}\bigr)
=\{(R_\omega(\theta), g_{_{i,\theta}}(\gams{\ai}(\theta)))\}\\
&= \{(R_\omega(\theta), \gams{\aii}(R_\omega(\theta)))\}
= \setfibpt{\A_{\aii}}{R_\omega(\theta)}
\end{align*}
for every $\theta \in \wbasint{i} \setminus \basintabs[\delta]{i}.$
So, (c) holds.
\end{proof}
Up to now we have defined the family of auxiliary functions
{\map{G_i}{\wbasicbox{i}}[\Omega]} with $i \in \Z.$
The next step before being able to define
the family $\{T_m\} \subset \cSO$ is to fix some
stratification in the set of boxes $\wbasicbox{i}.$
\section{A stratification in the set of boxes $\protect\wbasicbox{i}$}\label{stratification}
In this section we introduce a notion of \emph{depth}
in the set of arcs $\wbasint{i}$ defined earlier.
This notion introduce a stratification in the set of boxes
$\wbasicbox{i}$ that we study below.
\begin{definition}\label{depth}
For every $\ell\in\Z$ we define the \emph{depth of $\ell$},
which will be denoted by $\dep(\ell)$,
as the cardinality of the set (see Lemma~\ref{Propertiesvarphi}(g))
\begin{align*}
\set{i\in\Z}{\wbasint{\ell} \varsubsetneq \wbasint{i}} &=
\set{i\in\Z}{\wbasint{\ell} \cap \wbasint{i} \ne \emptyset} =\\
\set{i\in\Z}{\wbasicbox{\ell} \varsubsetneq \wbasicbox{i}} &=
\set{i\in\Z}{\wbasicbox{\ell} \cap \wbasicbox{i} \ne \emptyset}.
\end{align*}
Also, for every $m \in \Z^+,$ we denote
\begin{align*}
\DS &:= \set{\ell \in \Z}{\dep(\ell) = m},\\
\sstar{\DS} &:= \set{\istar}{i\in\DS},\text{ and}\\
\mu_m &:= \min\set{\ai}{i \in \DS}.
\end{align*}
\end{definition}
The next lemma studies the stratification on $\Z$ created
by the notion of \emph{depth}.
\begin{lemma}\label{Dsets}
The following statements hold:
\begin{enumerate}[(a)]
\item $\DS[m+1] \subset \set{\ell \in \Z}{\exists\; i \in \DS
\text{ such that }
\wbasint{\ell} \varsubsetneq \wbasint{i}}.$
\item For every $\ell, k \in \DS$ it follows that
$\wbasint{\ell} \cap \wbasint{k} = \emptyset.$
\end{enumerate}
\end{lemma}
\begin{proof}
Observe that if $\wbasint{\ell} \varsubsetneq \wbasint{i}$ then $\dep(\ell)\ge \dep(i)+1.$
Hence, (a) holds.
Statement (b) follows from Lemma~\ref{Propertiesvarphi}(g).
\end{proof}
In what follows, for every $m \in \Z^+$ we set
\[
\wIBD := \bigcup_{i \in \DS} \wbasint{i}\supset \sstar{\DS}.
\]
Note that, by Lemma~\ref{Dsets}(b), $\wIBD$ is a disjoint union of closed arcs.
Therefore, for every $\theta \in \wIBD,$ there exists a unique
$i \in \DS$ such that $\theta \in \wbasint{i}.$
We will denote such integer $i$ by $\bt{\theta} \in \DS.$
The next two lemmas study the properties of the winged boxes
$\wbasint{i}$ and $\wbasicbox{i}$ according to the depth stratification.
Lemma~\ref{QuePassaALesAles} is the real motivation to introduce the winged boxes.
\begin{lemma}\label{denso}
The following statements hold:
\begin{enumerate}[(a)]
\item The sequence $\{\mu_m\}_{m=0}^{\infty}$ is strictly increasing.
In particular $\lim_{m\to\infty} \mu_m = \infty.$
\item For every $m \in \Z^+,$ $\wIBD$ is dense in $\SI,$
$\wIBD[m+1]\subset \wIBD$
and $\sstar{\DS} \cap \wIBD[m+1] = \emptyset.$
\item $\Orbom \subset \wIBD[0],$ and
$\setfibth{\A} = \{(\theta,0)\}$
for every $\theta \in \SI\setminus \wIBD[0].$
\item Let $i\in\Z$ and $\theta \in \wbasint{i} \setminus \wIBD[\dep(i)+1].$
Then, $\theta \notin \Orbom$ unless $\theta = \istar,$ and
$\setfibth{\A_n} = \setfibth{\A_{\ai}}$
for every $n \ge \ai.$
In particular $\setfibth{\A} = \setfibth{\A_{\ai}}.$
\end{enumerate}
\end{lemma}
\begin{proof}
By Lemmas~\ref{Dsets}(a) and \ref{Propertiesvarphi}(g) it follows that
for every $m \in \Z+$ and $\ell \in \DS[m+1]$ there exists $i\in \DS$
such that $\wbasint{\ell} \varsubsetneq \wbasint{i}$ and $\ai < \all.$
Thus, $\wIBD[m+1]\subset \wIBD$ and $\mu_m < \mu_{m+1}.$
This proves (a) and the second statement of (b).
Next we will show that $\istar \notin\wIBD[m+1]$
for every $i\in \DS.$
Assume by way of contradiction that there exists
$i\in \DS$ such that $\istar \in\wIBD[m+1].$
Let $k = \bt[m+1]{\istar} \in \DS[m+1].$
Clearly, $i \ne k$ and $\istar \in \wbasint{k}.$
Then, by Lemma~\ref{Propertiesvarphi}(g),
$\ak < \ai$ and $\wbasint{i} \varsubsetneq \wbasint{k}.$
Thus,
\[
m = \dep(i) \ge \dep(k) + 1 = m+2;
\]
a contradiction.
Now we prove the first statement of (c).
From the definitions and the part of (b) already proven we have
\[
\Orbom \subset \bigcup_{i \in \Z} \wbasint{i}
\subset \bigcup_{m=0}^\infty \wIBD = \wIBD[0].
\]
To end the proof of (b) it remains to show the density of $\wIBD.$
We will do it by induction on $m.$
Clearly $\wIBD[0] \supset \Orbom$ is dense in $\SI$ because so is $\Orbom.$
Suppose that (b) holds for $\wIBD.$
We will show that (b) also holds for $\wIBD[m+1].$
Choose $\theta\in \wIBD$ and set $i = \bt{\theta}.$
Since $\Orbom$ is dense in $\SI,$ there exists a sequence
$\{s_{n}\}_{n=0}^{\infty} \subset \Z$ such that
$\sstar{s}_{n} \in \wobasint{i}$ and
$\lim_{n\to\infty} \sstar{s}_n = \theta.$
As above, we get that $\dep(s_{n}) \ge \dep(i) + 1 = m + 1.$
Moreover,
$\sstar{s}_n \in \wIBD[\dep(s_{n})] \subset \wIBD[m+1]$
for every $n.$
Consequently, $\wIBD \subset \overline{\wIBD[m+1]},$ and
the density of $\wIBD[m+1]$ follows from the density of $\wIBD.$
Next we prove the second statement of (c).
From above it follows that
\[
\bigcup_{i \in \Z} \basintabs{i} \subset
\bigcup_{i \in \Z} \wbasint{i} \subset \wIBD[0].
\]
Hence, by the definition of the maps $\gams{m}$ (Definition~\ref{PCgenerators})
it follows that $\gams{m}(\theta) = \gams{0}(\theta) = 0$
for every $\theta \notin \wIBD[0]$ and $m \in \Z^+.$
So, $\gamma(\theta) = \lim_{m\to\infty} \gams{m}(\theta) = 0,$
and
$\setfibth{\A} = \{(\theta,\gamma(\theta))\} = \{(\theta,0)\}$
by Lemma~\ref{propiedadesPCA}(c).
This ends the proof of (c).
\inidemopart{d}
If $\theta = \istar$ then the statement follows from
Lemmas~\ref{propiedadesA}(b) and \ref{propiedadesPCA}(b).
So, we assume that $\theta \ne \istar.$
By Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2)
we get that
$\theta \notin \Zstar_{\ai + 1}.$
Hence, if $\theta \in \Orbom,$ it follows that
$\theta = \kstar \in \wIBD[\dep(k)]$ with $\ak > \ai + 1$ and
$\wbasint{k} \cap \wbasint{i} \ne \emptyset.$
Thus, by Lemma~\ref{Propertiesvarphi}(g), $\dep(k) \ge \dep(i)+1$.
By (b), this implies that
$\theta = \kstar \in \wIBD[\dep(i)+1];$
a contradiction.
Therefore, $\theta \notin \Orbom.$
On the other hand,
$\theta \notin \wbasint{-i}$
by Definition~\ref{PCgenerators}(\tsfR.2).
If $\theta \notin \basintabs{k}$
for every $k \in \Z$ such that $\ak > \ai,$ then
$\gams{n}(\theta) = \gams{\ai}(\theta)$ and
$\setfibth{\A_n} = \setfibth{\A_{\ai}}$ for every $n \ge \ai,$
by Definition~\ref{PCgenerators} and Lemma~\ref{propiedadesA}(c).
Now assume that $\theta \in \basintabs{k}$ for some $k \in \Z$
such that $\ak > \ai$ and $\ak$ is minimal with these properties.
If $\theta \in \wobasint{k},$
as above we get that $\dep(k) \ge \dep(i)+1$ and
$\theta \in \wIBD[\dep(k)] \subset \wIBD[\dep(i)+1].$
Thus, $\theta \in \Bd(\wbasint{k}) = \Bd(\basintabs{k})$
and $k \ge 0.$
So, by Lemma~\ref{Propertiesvarphi}(c) and
the definition of the maps $\gams{j}$ (Definition~\ref{PCgenerators}),
$\gams{\ak}(\theta) = \gams{\ak - 1}(\theta).$
Moreover, by Lemma~\ref{Propertiesvarphi}(e),
$\gams{j}(\theta) = \gams{\ak}(\theta)$ for every $j > \ak.$
On the other hand, the minimality of $\ak$ implies that
$\theta \notin \basintabs{\ell}$ for every
$\ell \in \Z$ such that $\ak > \all > \ai.$
Hence, by the definition of the maps $\gams{j}$ (Definition~\ref{PCgenerators}),
$\gams{j}(\theta) = \gams{\ai}(\theta)$ for every $\ak > j > \ai.$
In short, we have proved that
$\gams{j}(\theta) = \gams{\ai}(\theta)$ for every $j \ge \ai.$
Thus, as above, $\setfibth{\A_n} = \setfibth{\A_{\ai}}$ for every $n \ge \ai.$
This ends the proof of the lemma.
\end{proof}
\begin{lemma}\label{QuePassaALesAles}
Assume that $\wbasint{i} \subset \wbasint{k}$
for some $i \in \DS,\ k \in \DS[m-1]$ and $m \in \N.$
Then, $\ak < \ai$ and $\akk < \aii$
unless $k \ge 0$ and $i = -(k+2)$ (whence $\akk = \aii$).
Moreover, the following statements hold:
\begin{enumerate}[(a)]
\item For every $\theta \in \wbasint{i},$
\[
\gams{\ak}(\theta) = \gams{\ak+1}(\theta) = \dots = \gams{\ai-1}(\theta) \in \I_{i, \theta}
\]
and, when $\akk < \aii,$
\[
\gams{\akk}\left(R_\omega(\theta)\right) =
\gams{\akk+1}\left(R_\omega(\theta)\right) = \dots =
\gams{\aii-1}\left(R_\omega(\theta)\right)
\]
\item For every $\theta \in \wbasint{i} \setminus \obasintabs{i},$
\[
\gams{\ai}(\theta) = \gams{\ai-1}(\theta)
\andq
\I_{i, \theta} = \{\gams{\ai}(\theta)\} = \{\gams{\ak}(\theta)\} \subset \I_{k, \theta}.
\]
\end{enumerate}
\end{lemma}
\begin{proof}
The fact that $\ak < \ai$ follows from Lemma~\ref{Propertiesvarphi}(g).
Therefore, either
$\akk < \aii$ or
$k \ge 0,$ $i = -(k+2)$ and $\akk = \aii$ or
$k \ge 0,$ $i = -(k+1)$ and $\akk > \aii.$
In the last case, $\wbasint{i} = \wbasint{-(k+1)}$ and $\wbasint{k}$
must be disjoint by Definition~\ref{PCgenerators}(\tsfR.2) (with $j = k$);
which is a contradiction.
Thus $\akk < \aii$ unless $k \ge 0$ and $i = -(k+2)$ ($\akk = \aii$).
By Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
$\wbasint{i} \cap \Zstar_{\ai-1} = \emptyset.$
Hence, from the definition of the maps $\gams{j}$ (Definition~\ref{PCgenerators}),
to prove that
\[
\gams{\ak}\evalat{\wbasint{i}} =
\gams{\ak + 1}\evalat{\wbasint{i}} = \dots =
\gams{\ai-2}\evalat{\wbasint{i}} =
\gams{\ai-1}\evalat{\wbasint{i}},
\]
it is enough to show that
$\basintabs{\ell} \cap \wbasint{i} = \emptyset$
for every $\ell$ such that $\ak < \all < \ai.$
Assume that $\basintabs{\ell} \cap \wbasint{i} \ne \emptyset$
for some $\ell$ such that $\ak < \all < \ai.$
Then,
\[
\emptyset \ne \basintabs{\ell} \cap \wbasint{i} \subset
\wbasint{\ell} \cap \wbasint{i} \subset
\wbasint{\ell} \cap \wbasint{k}
\]
and, by Lemma~\ref{Propertiesvarphi}(g),
\[
\wbasint{i} \varsubsetneq \wbasint{\ell} \varsubsetneq \wbasint{k}.
\]
So, in a similar way as before,
\[
m = \dep(i) \ge \dep(\ell) + 1 \ge \dep(k) + 2 = m+1;
\]
a contradiction.
This ends the proof of the first statement of (a).
Now we show that if $\akk < \aii-1,$ then
\[
\gams{\akk}\left(R_\omega(\theta)\right) =
\gams{\akk+1}\left(R_\omega(\theta)\right) = \dots =
\gams{\aii-1}\left(R_\omega(\theta)\right),
\]
and are well defined.
First we prove that
$\gams{\ell}\left(R_\omega(\theta)\right)$
is well defined for every $\ell = 0,1,\dots,\aii-1.$
For every $\theta \in \wbasint{i}$ we have
\[
R_\omega(\theta) \in R_\omega\left(\wbasint{i}\right) =
\begin{cases}
\BSG{i+1}{i} &\text{when $i \ge 0$, and}\\
\basintabs{i+1} \subset \wbasint{i+1} &\text{when $i < 0$}.
\end{cases}
\]
In any case, by Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2)
with $j=i$ when $i \ge 0$ and $\ell = -(j+1) = i+1$ when $i < 0,$
and Lemma~\ref{Propertiesvarphi}(a),
\[
R_\omega(\theta) \notin \begin{cases}
\Zstar_{i} &\text{when $i \ge 0$, and}\\
\Zstar_{\aii-1} &\text{when $i < 0$,}
\end{cases}
\]
and $\gams{\ell}\left(R_\omega(\theta)\right)$ is well defined
for $\ell = 0,1,\dots,\aii-1$
(recall that $\Zstar_{m} \subset \Zstar_{m+1}$ for every $m \ge 0$).
Now, assume by way of contradiction that
\[
\gams{\ell}\left(R_\omega(\theta)\right) \ne \gams{\ell-1}\left(R_\omega(\theta)\right)
\text{ for some }
\ell \in \{\akk+1, \akk+2, \dots, \aii-1\},
\]
and $\ell$ is minimal with this property (observe that $\ell \ge 1$).
By the definition of the map $\gams{\ell}$ (Definition~\ref{PCgenerators}),
\[ R_\omega(\theta) \in \OBG{q+1}{\ell} \andq[with] q \in \{\ell-1, -(\ell+1)\} \]
and, hence, $\theta \in \OBG{q}{\ell}.$
Since $\akk+1 \le \ell < \aii$, when $q = -(\ell+1) \le -2,$
\[
\akk+2 \le -q \le \aii
\text{ and }
\OBG{q}{\ell} = \wobasint{-(\ell+1)} = \wobasint{q}.
\]
Otherwise, when $q = \ell-1 \ge 0,$ $\akk \le q \le \aii-2$ and
\[
\OBG{q}{\ell} \subset \obasint{\ell-1} = \wobasint{\ell-1} = \wobasint{q},
\]
by Definition~\ref{PCgenerators}(\tsfR.1).
Next we want to use Lemma~\ref{Propertiesvarphi}(g) to show that
$
\wbasint{i} \varsubsetneq \wbasint{q} \varsubsetneq \wbasint{k}.
$
To this end we have to compare $\aq$ with $\ai$ and $\ak.$
Notice $\wbasint{q} \cap \wbasint{k} \ne \emptyset$ because
\[
\theta \in \wobasint{q} \cap \wbasint{i} \subset \wobasint{q} \cap \wbasint{k}.
\]
If $k \ge 0,\ \aq \ge \akk > \ak.$
When $k,q < 0,\ \aq \ge \akk + 2 = \ak + 1 > \ak.$
If $k < 0$ and $q \ge 0,\ \aq = q \ge \akk = \ak - 1.$
If $q = \ak - 1$ (that is, $k = -(q+1)$), as above,
by Definition~\ref{PCgenerators}(\tsfR.2) with $j = q$ we get
$\wbasint{k} \cap \wbasint{q} = \emptyset;$ a contradiction.
So, $\aq > \ak$ unless $\aq = \ak$ and $k < 0 \le q.$
Summarizing, we have shown that $\aq \ge \ak$ and $q \ne k.$
Then, from Lemma~\ref{Propertiesvarphi}(g) we get that $\aq > \ak$
and $\wbasint{q} \varsubsetneq \wbasint{k}.$
Now we will study the relation of $\wbasint{q}$ with the box $\wbasint{i}.$
From above we get that $\wbasint{q} \cap \wbasint{i} \ne \emptyset.$
If $i < 0,\ \aq \le \aii = \ai-1.$
When $q,i \ge 0,$ we have $\aq = q \le \aii-2 = \ai - 1.$
If $i \ge 0$ and $q < 0,\ \aq \le \aii = \ai + 1.$
Assume that $i \ge 0$ and $q = -(i+1) < 0.$
In this case, additionally, $q = -(\ell+1)$ and, thus, $i = \ell \ge 1.$
Then,
\begin{align*}
R_\omega(\theta) & \in R_\omega\left(\wbasint{i}\right)
= R_\omega\left(\basint{i}\right)
= \BSG{i+1}{i}, \text{ and}\\
R_\omega(\theta) & \in \OBG{q+1}{\ell} = \OBG{-i}{i} \subset \wobasint{-i},
\end{align*}
which is a contradiction by Definition~\ref{PCgenerators}(\tsfR.2).
Summarizing, $\aq < \ai$ unless $\aq = \ai$ and $q < 0 \le i$
(that is, $\aq \le \ai$ and $q \ne i$).
Then, again by Lemma~\ref{Propertiesvarphi}(g), $\aq < \ai$
and $\wbasint{i} \varsubsetneq \wbasint{q} \varsubsetneq \wbasint{k}.$
So, as before,
\[
m = \dep(i) \ge \dep(q) + 1 \ge \dep(k) + 2 = m+1;
\]
a contradiction.
This ends the proof of (a).
Now we assume that $\theta \in \wbasint{i} \setminus \obasintabs{i}.$
By Lemmas~\ref{Propertiesvarphi}(e) and \ref{voresdelescaixesalesvores}(d),
\[
\gams{\ai}(\theta) = \gams{\ai-1}(\theta)
\andq
\I_{i, \theta} = \{\gams{\ai}(\theta)\} =
\{\gams{\ai-1}(\theta)\} = \{\gams{\ak}(\theta)\}.
\]
On the other hand, by Lemma~\ref{denso}(b),
$\sstarplain{\DS[m-1]} \cap \wIBD = \emptyset$
which implies that $\theta \ne \kstar$
because $\kstar \in \sstarplain{\DS[m-1]}$ and
$\theta \in \wbasint{i} \setminus \obasintabs{i} \subset \wIBD.$
So, by \eqref{gammathetaproperties},
\[
\I_{i, \theta} = \{\gams{\ai}(\theta)\} = \{\gams{\ak}(\theta)\} \subset \I_{k, \theta}.
\]
Now we prove that
$\gams{\ai-1}(\theta) \in \I_{i, \theta}$
for every $\theta \in \wbasint{i}.$
From above, we have $\I_{i, \theta} = \{\gams{\ai-1}(\theta)\}$
for every $\theta \in \wbasint{i} \setminus \obasintabs{i}.$
Moreover, when $\theta \in \obasintabs{i}$
the statement follows directly from Lemma~\ref{Propertiesvarphi}(c).
Thus, (b) is proved.
\end{proof}
\section{Boxes in the wings}\label{BoxesintheWings}
To prove Theorem~\ref{MainTh} we will inductively construct a Cauchy sequence
$\{T_m\}_{m=0}^{\infty} \subset \cSO$
that gives the function $T$ from Theorem~\ref{MainTh} as a limit.
This section is devoted to study the points in the wings of
boxes in the circle and its interaction with boxes of higher depth.
The resulting technology is necessary to be able to construct the
sequence $\{T_m\}_{m=0}^{\infty}$ so that it is Cauchy sequence.
Unfortunately this will complicate even more the definition of the
functions $T_m$ and the proof of its continuity.
We start by introducing some more notation.
For every $m \in \Z^+$ we set
\begin{align*}
\IBD &:= \bigcup_{i \in \DS} \basintabs{i} \subset \wIBD,\text{ and}\\
\WDB &:= \Bigl\{\theta \in \wIBD\setminus\IBD \,\colon
\theta \in \IBD[j] \text{ for some $j > m$}
\Bigr\}.
\end{align*}
On the other hand, the smallest number $j$ from the above definition will be called the
\emph{least essential depth of $\theta$ below $m$,}
and will be denoted by $\led{\theta}.$
That is, $\led{\theta}$ denotes the positive integer larger than $m$
such that
\[
\theta \in \wIBD[j] \setminus \IBD[j]
\text{ for } j= m, m+1,\dots, \led{\theta}-1
\andq
\theta \in \IBD[\led{\theta}].
\]
The following simple lemmas are useful
to better understand and use the above definitions.
The next lemma establishes the relation between boxes in the wings of increasing depth.
\begin{lemma}\label{DepthintheWings}
Assume that $\theta \in \WDB$ for some $m \in \Z^+$
and set $\ell = \led{\theta}.$
Then, the following statements hold.
\begin{enumerate}[(a)]
\item For every $j = m, m+1,\dots, \ell$
the numbers $\is_j = \bt[j]{\theta} \in \DS[j]$ are well
defined and are all of them negative except, perhaps,
$\is_{\ell} = \bt[\led{\theta}]{\theta}.$
\item
\begin{align*}
& \hspace*{2em}\abs{\is_{m}} < \abs{\is_{m+1}} < \dots <
\abs{\is_{\ell-1}} < \abs{\is_{\ell}},\text {and}\\
\theta &\in \basintabs{\is_{\ell}}
\subset \wobasint{\is_{\ell-1}} \setminus \basintabs{\is_{\ell-1}}\\
&\subset \wobasint{\is_{\ell-2}} \setminus \basintabs{\is_{\ell-2}}
\subset \cdots
\subset \wobasint{\is_{m}} \setminus \basintabs{\is_{m}}.
\end{align*}
\item For every $j = m, m+1,\dots,\ell-1,$
$\basintabs{\is_{\ell}} \subset \WDB[j],$
$\led[j]{\nu} = \led{\theta}$ and
$\bt[\led[j]{\nu}]{\nu} = \bt[\led{\theta}]{\theta} = \is_{\ell}$
for every $\nu \in \basintabs{\is_{\ell}}.$
\item $\I_{\is_{m}, \nu} = \{\gams{\abs{\is_{m}}}(\nu)\} \subset \I_{\is_{\ell}, \nu}$
for every $\nu \in \obasintabs{\is_{\ell}}$ and
\[
\I_{\is_{m}, \nu} = \{\gams{\abs{\is_{m}}}(\nu)\} =
\{m_{\is_{\ell}}(\nu)\} = \{M_{\is_{\ell}}(\nu)\} =
\{\gams{\abs{\is_{\ell}}}(\nu)\} = \I_{\is_{\ell}, \nu}
\]
for every $\nu \in \Bd\left(\basintabs{\is_{\ell}}\right).$
\end{enumerate}
\end{lemma}
\begin{proof}
Since $\wbasint{i} = \basint{i}$ for every $i \ge 0,$
\begin{equation}\label{thewings}
\wIBD\setminus\IBD = \bigcup_{\substack{i \in \DS\\ i < 0}}
\left(\wbasint{i} \setminus \basintabs{i}\right)
\end{equation}
for every $m \in \Z^+.$
Statement (a) follows from Lemma~\ref{denso}(b) and \eqref{thewings}.
Then, (b) follows from Lemma~\ref{Propertiesvarphi}(g).
Statement (c) is an easy consequence of (b) and the definitions.
Now we prove (d) iteratively.
Fix $\nu \in \obasintabs{\is_{\ell}}.$
By (b)
\[
\nu \in \wobasint{\is_{m+1}} \setminus \basintabs{\is_{m+1}}
\subset \wobasint{\is_{m}} \setminus \basintabs{\is_{m}}
\]
provided that $\ell = \led{\theta} > m+1.$
Hence, by Lemmas~\ref{voresdelescaixesalesvores}(d) and \ref{QuePassaALesAles},
\begin{align*}
\gams{\abs{\is_{m}}}(\nu) &= \gams{\abs{\is_{m}}+1}(\nu) = \dots = \gams{\abs{\is_{m+1}}}(\nu),\text{ and}\\
\I_{\is_{m}, \nu} &= \{\gams{\abs{\is_{m}}}(\nu)\} = \{\gams{\abs{\is_{m+1}}}(\nu)\} = \I_{\is_{m+1}, \nu}.
\end{align*}
By iterating this argument we get,
\[
\gams{\abs{\is_{m}}}(\nu) = \gams{\abs{\is_{m}}+1}(\nu) = \dots = \gams{\abs{\is_{\ell-1}}}(\nu)
\andq
\I_{\is_{m}, \nu} = \I_{\is_{\ell-1}, \nu}.
\]
Again by (b) and Lemmas~\ref{voresdelescaixesalesvores}(d) and \ref{QuePassaALesAles},
\[
\gams{\abs{\is_{m}}}(\nu) = \gams{\abs{\is_{m}}+1}(\nu) = \dots = \gams{\abs{\is_{\ell}}}(\nu)
\andq
\I_{\is_{m}, \nu} = \I_{\is_{\ell}, \nu}
\]
when $\nu \in \Bd\left(\basintabs{\is_{\ell}}\right)$
and, otherwise,
\[
\gams{\abs{\is_{m}}}(\nu) = \gams{\abs{\is_{m}}+1}(\nu) = \dots = \gams{\abs{\is_{\ell}-1}}(\nu)
\andq
\I_{\is_{m}, \nu} \subset \I_{\is_{\ell}, \nu}.
\]
\end{proof}
Equipped with above results and definition we are going to define two maps,
analogous to the maps $m_i$ and $M_i,$ on the wings of the negative boxes.
\begin{definition}\label{curvesinthewings}
For every $m \in \Z^+$ we define
\begin{align*}
\WDS &:= \set{\bt[\led{\theta}]{\theta}}{\theta \in \WDB} \subset \Z,\\
\WIB &:= \Int(\WDB) = \LSleftlimits{\bigcup}{i \in \WDS} \obasintabs{i},\\
\WB &:= \bigcup_{\substack{i \in \DS\\ i < 0}} \left(\wbasint{i} \setminus \obasintabs{i}\right),\text{ and}\\
\wEIBD & := \bigcup_{i \in \DS} \Bd\left(\wbasint{i}\right) \subset \wIBD.
\end{align*}
By Lemma~\ref{DepthintheWings}(a,c), $\WDS$ is well defined and
\[
\WIB \subset \WDB \subset \wIBD\setminus\IBD \subset \WB.
\]
Consequently,
\[ \wIBD = \IBD \cup \WB. \]
Then, we can define functions
{\map{\tau_m}{\WB}[\I]}
and
{\map{\lambda_m}{\WB}[\I]}
as follows:
\begin{align*}
\tau_m(\theta) &:= \begin{cases}
M_{\bt[\led{\theta}]{\theta}}(\theta) & \text{if $\theta \in \WIB$,}\\
\gams{\abs{\bt{\theta}}}(\theta) & \text{otherwise,}
\end{cases}\\
\lambda_m(\theta) &:= \begin{cases}
m_{\bt[\led{\theta}]{\theta}}(\theta) & \text{if $\theta \in \WIB$,}\\
\gams{\abs{\bt{\theta}}}(\theta) & \text{otherwise.}
\end{cases}
\end{align*}
Clearly, by Lemmas~\ref{voresdelescaixesalesvores}(a) and \ref{Propertiesvarphi}(b),
\[
-1 \le \lambda_m(\theta) \le \tau_m(\theta) \le 1
\]
for every $\theta \in \WB.$
So, we can define
\[
\IW{\theta} := [\lambda_m(\theta), \tau_m(\theta)] \subset [0,1].
\]
\end{definition}
The next lemmas will help us in the definition and study of the maps $T_m$.
\begin{lemma}\label{VerticalIntervalsIntheWings}
The following statements hold for every $m \in \Z^+.$
\begin{enumerate}[(a)]
\item $\WIB \cap \IBD = \WIB \cap \wEIBD = \emptyset.$
\item Let $\theta \in \WB.$ Then,
$
\I_{\bt{\theta}, \theta} = \left\{\gams{\abs{\bt{\theta}}}(\theta)\right\},
$
\begin{align*}
\I_{\bt{\theta}, \theta} &= \IW{\theta} && \text{when $\theta \notin \WIB$, and}\\
\I_{\bt{\theta}, \theta} &\subset \IW{\theta} && \text{when $\theta \in \WIB$.}
\end{align*}
\item Assume that $m \in \N$ and let $U$
be a connected component of $\WB$ such that $U \subset \WB[m-1].$
Then,
$\WDB \cap U \subset \WDB[m-1],$
$\WIB \cap U = \WIB[m-1] \cap U$ and
$\IW{\theta} = \IW[m-1]{\theta}$ for every $\theta \in U.$
\end{enumerate}
\end{lemma}
\begin{proof}
\inidemopart{a}
By Lemma~\ref{DepthintheWings}(b),
\[
\theta \in \wobasint{\bt{\theta}} \setminus \basintabs{\bt{\theta}}
\]
and $\bt{\theta} < 0$ for every $\theta \in \WIB \subset \WDB.$
So, by Lemma~\ref{Dsets}(b), we get
$\theta \notin \IBD \cup \wEIBD.$
\inidemopart{b}
The fact that
$
\I_{\bt{\theta}, \theta} = \left\{\gams{\abs{\bt{\theta}}}(\theta)\right\}
$
follows from Lemma~\ref{voresdelescaixesalesvores}(d).
The other two statements follow from Definition~\ref{curvesinthewings}
and Lemma~\ref{DepthintheWings}(d).
\inidemopart{c}
The assumption that $U$ is a connected component of $\WB$
and $U \subset \WB[m-1]$ implies
by Lemmas~\ref{Dsets}(b) and \ref{Propertiesvarphi}(g)
that there exist $i\in \DS$ and $k \in \DS[m-1],$ $i,k < 0,$
such that $U$ is a connected component of
\[
\wbasint{i}\setminus\obasintabs{i} \subset
\wobasint{k}\setminus\basintabs{k} \subset \WB[m-1].
\]
Again by Lemma~\ref{Dsets}(b)
this implies that $U \subset \wIBD[m-1]\setminus\IBD[m-1].$
Moreover, by definition,
$\WDB \subset \wIBD\setminus\IBD.$
Consequently, $\WDB \cap U \subset \WDB[m-1].$
Let $\theta \in \WIB \cap U \subset \WDB \cap U \subset \WDB[m-1] \cap U.$
By Definition~\ref{curvesinthewings} and Lemma~\ref{DepthintheWings}(a,b),
$i = \bt{\theta}$ and
there exists $\ell = \bt[\led{\theta}]{\theta} \in \WDS$ such that
\[
\theta \in \obasintabs{\ell} \subset
\wbasint{i}\setminus\obasintabs{i} \subset
\wobasint{k}\setminus\basintabs{k}.
\]
Therefore, again by
Lemma~\ref{DepthintheWings}(a--c) and Definition~\ref{curvesinthewings},
$\led[m-1]{\theta} = \led{\theta},$
\[ \ell = \bt[\led{\theta}]{\theta} = \bt[\led[m-1]{\theta}]{\theta} \in \WDS[m-1] \]
and
$\theta \in \obasintabs{\ell} \subset \WIB[m-1].$
Hence, $\WIB \cap U \subset \WIB[m-1].$
Now assume that $\theta \in \WIB[m-1] \cap U.$
As above, there exist $r = \bt{\theta} \in \DS$
and $\ell = \bt[\led[m-1]{\theta}]{\theta} \in \WDS[m-1]$ such that
\[
\theta \in \obasintabs{\ell} \subset
\wobasint{r}\setminus\basintabs{r} \subset
\wobasint{k}\setminus\basintabs{k}.
\]
Since $\theta \in U \subset \wbasint{i}$,
Lemma~\ref{Dsets}(b) gives
$i = r$ and $\theta \in \obasintabs{\ell} \subset U.$
Moreover, by Lemma~\ref{DepthintheWings}(c),
$\ell = \bt[\led[m-1]{\theta}]{\theta} = \bt[\led{\theta}]{\theta}\in \WDS$
and, so, $\theta \in \obasintabs{\ell} \subset \WIB.$
Thus, $\WIB \cap U = \WIB[m-1] \cap U.$
To end the proof of the lemma we have to show that
$\IW{\theta} = \IW[m-1]{\theta}$ for every $\theta \in U.$
Assume first that
$\theta \in U \setminus \WIB \subset \WB \setminus \WIB.$
Then,
\[
\theta \in U \setminus \WIB =
U \setminus \WIB[m-1] \subset
\WB[m-1] \setminus \WIB[m-1]
\]
and, by (b) and Lemmas~\ref{voresdelescaixesalesvores}(d) and \ref{QuePassaALesAles},
\[
\IW{\theta} = \I_{i, \theta} = \{\gams{\ai}(\theta)\} =
\{\gams{\ak}(\theta)\} = \I_{k, \theta} = \IW[m-1]{\theta}.
\]
If $\theta \in U \cap \WIB = U \cap \WIB[m-1]$ then we get
\begin{align*}
\IW{\theta}
&= \left[m_{\bt[\led{\theta}]{\theta}}(\theta),
M_{\bt[\led{\theta}]{\theta}}(\theta)\right] \\
&= \left[m_{\bt[\led[m-1]{\theta}]{\theta}}(\theta),
M_{\bt[\led[m-1]{\theta}]{\theta}}(\theta)\right]
= \IW[m-1]{\theta}
\end{align*}
from Definition~\ref{curvesinthewings} and
Lemma~\ref{DepthintheWings}(c).
\end{proof}
\begin{lemma}\label{continuouscurvesinthewings}
Let $m \in \Z^+$ and let $U$ be a connected component of $\WB.$
Then, the functions $\lambda_m\evalat{U}$ and $\tau_m\evalat{U}$
are continuous.
\end{lemma}
\begin{proof}
We will prove only the continuity of $\lambda_m\evalat{U}.$
The proof of the continuity of $\tau_m\evalat{U}$ is analogous.
By Lemmas~\ref{DepthintheWings}(c) and \ref{voresdelescaixesalesvores}(b)
we get
\begin{equation}\label{contindepboxes}
\text{\parbox{0.9\textwidth}{
for every $\ell \in \WDS,$
$\ell = \bt[\led{\nu}]{\nu}$ for every $\nu \in \basintabs{\ell},$
and the function $m_{\ell}$ is continuous on $\basintabs{\ell}.$
}}\end{equation}
Let $\ell \in \WDS$ be such that $\obasintabs{\ell} \subset \WIB \cap U.$
Thus, by \eqref{contindepboxes}, the function
$\lambda_m = m_{\ell}$ is continuous on $\obasintabs{\ell}.$
So, we have to show that $\lambda_m$ is continuous at every
$\theta \in U \setminus \WIB.$ To show this we will use a simple usual
$\varepsilon$--$\delta$ game. Fix $\varepsilon > 0.$
By Lemma~\ref{Dsets}(b) it follows that $U$ is a connected component
of $\wbasint{i} \setminus \obasintabs{i}$ for some $i \in \DS,$ $i < 0,$
and
\begin{equation}\label{whichbox}
\bt{\nu} = i \andq[for every] \nu \in U.
\end{equation}
By Lemma~\ref{Propertiesvarphi}(a) and
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
the function $\gams{\ai}\evalat{U}$ is continuous.
So,
\begin{equation}\label{gamscont}
\text{\parbox{0.9\textwidth}{
there exists $\overline{\delta}_{\ai} = \overline{\delta}_{\ai}(\theta) > 0$ such that
$\abs{\gams{\ai}(\theta), \gams{\ai}(\nu)} < \varepsilon/2$
provided that $\dSI(\theta, \nu) < \overline{\delta}_{\ai}.$
}}\end{equation}
On the other hand, by \eqref{contindepboxes},
\begin{equation}\label{contindepboxesepsdel}
\text{\parbox{0.9\textwidth}{
for every $\ell \in \WDS,$ there exists $\delta_{\ell} > 0$ such that
$\abs{m_{\ell}(\widetilde{\theta}), m_{\ell}(\nu)} < \varepsilon/2$
for every
$\widetilde{\theta} \in \Bd\left(\basintabs{\ell}\right)$ and
$\nu \in \Bd\basintabs{\ell}$
such that $\dSI(\theta, \nu) < \delta_{\ell}.$
}}\end{equation}
Now we will define $\delta.$
Note that there exists $N \in \N$ such that $2^{-N} < \varepsilon/2.$
Then we set:
\[
\delta = \delta(\theta) := \min\left\{
\overline{\delta}_{\ai}(\theta),
\min \set{\delta_{\ell}}{\ell \in \WDS \text{ and } \all < N}
\right\}.
\]
Clearly, $\delta > 0$ because the set $\set{\ell \in \WDS}{\all < N}$ is finite.
To end the proof of the lemma we have to show that
\[
\abs{\lambda_m(\theta) - \lambda_m(\nu)} < \varepsilon
\]
whenever $\nu \in U$ and $\dSI(\theta, \nu) < \delta.$
Assume that $\nu \in U$ and $\dSI(\theta, \nu) < \delta$
(recall that we have the assumption that $\theta \notin \WIB$).
If $\nu \notin \WIB,$ then
$\dSI(\theta, \nu) < \delta \le \overline{\delta}_{\ai}(\theta)$
and, by \eqref{whichbox} and \eqref{gamscont},
\[
\abs{\lambda_m(\theta) - \lambda_m(\nu)} =
\abs{\gams{\ai}(\theta) - \gams{\ai}(\nu)} < \varepsilon/2 < \varepsilon.
\]
Now assume that there exists $\ell \in \WDS$
such that $\nu \in \obasintabs{\ell} \subset \WIB.$
Clearly, there exists
$\widetilde{\theta} \in \Bd\left(\basintabs{\ell}\right)$
such that
\begin{align*}
& \dSI(\theta, \widetilde{\theta}) < \dSI(\theta, \nu) < \delta \le \overline{\delta}_{\ai}(\theta)\text{ and}\\
& \dSI(\widetilde{\theta}, \nu) < \dSI(\theta, \nu) < \delta.
\end{align*}
Observe that, by Lemma~\ref{Dsets}(b), $\widetilde{\theta} \notin \WIB.$
Hence, by \eqref{whichbox} and Lemma~\ref{DepthintheWings}(c,d),
\[
\lambda_m(\widetilde{\theta}) = \gams{\ai}(\widetilde{\theta}) = m_{\ell}(\widetilde{\theta}).
\]
If $\all < N,$ then $\dSI(\widetilde{\theta}, \nu) < \delta \le \delta_{\ell}$ and,
by \eqref{contindepboxesepsdel},
$
\abs{m_{\ell}(\widetilde{\theta}) - m_{\ell}(\nu)} < \varepsilon/2.
$
Otherwise, by Lemma~\ref{Propertiesvarphi}(f),
\[
\abs{m_{\ell}(\widetilde{\theta}) - m_{\ell}(\nu)} <
\diam\left(\basicbox{\ell}\right)
\le 2^{-\all} \le 2^{-N} < \varepsilon/2.
\]
In any case,
$
\abs{m_{\ell}(\widetilde{\theta}) - m_{\ell}(\nu)} < \varepsilon/2.
$
Thus, again by \eqref{whichbox} and \eqref{gamscont},
\begin{align*}
\abs{\lambda_m(\theta) - \lambda_m(\nu)}
& \le \abs{\lambda_m(\theta) - \lambda_m(\widetilde{\theta})} +
\abs{\lambda_m(\widetilde{\theta}) - \lambda_m(\nu)}\\
& = \abs{\gams{\ai}(\theta) - \gams{\ai}(\widetilde{\theta})} +
\abs{m_{\ell}(\widetilde{\theta}) - m_{\ell}(\nu)} < \varepsilon.
\end{align*}
\end{proof}
\section{A Cauchy sequence of skew products. Proof of Theorem~\ref{MainTh}}\label{skew-product}
In this section prove Theorem~\ref{MainTh}.
To do this we inductively construct a Cauchy sequence
$\{T_m\}_{m=0}^{\infty} \subset \cSO$
that gives the function $T$ from Theorem~\ref{MainTh} as a limit.
The sequence $\{T_m\}_{m=0}^{\infty} \subset \cSO$ is defined so that
\[
T_m(\theta,x) = (R_\omega(\theta), f_m(\theta, x))
\]
and $\map{f_m}{\Omega}[\I]$ is continuous in both variables.
To build these functions we will use the auxiliary functions
{\map{G_i}{\basicbox{i}}[\Omega]} with $i \in \Z$
from Section~\ref{FunctionsGi}.
The maps $f_m(\theta, \cdot)$ will also be denoted as $f_{m,\theta},$
and will be defined non-increasing, and such that
$f_{m,\theta}(2) = -2$ and $f_{m,\theta}(-2) = 2$
for every $\theta \in \SI.$
To make more evident the strategy of the construction of this sequence of maps
we will separate several cases, and we will state without proofs
the results that study these maps.
After establishing all the definitions and results related to the
construction of the sequence $\{T_m\}_{m=0}^{\infty}$
without having been distracted by the technicalities involving
the proofs, we will proceed to provide the missing proofs.
More precisely, we will start by defining the map
$T_0$ and stating without proof the proposition that summarizes
the necessary properties of this map.
Next we will inductively define the maps
$\{T_m\}_{m=1}^{\infty} \subset \cSO$
and state without proof the proposition that establishes the
properties of the whole sequence
$\{T_m\}_{m=0}^{\infty}$.
Then, as we have said, we prove Theorem~\ref{MainTh}
and in the next three sections we will provide all pending proofs.
In what follows $\mathcal{C}(\I,\I)$
will denote the class of all continuous maps from $\I$ to itself.
We endow $\mathcal{C}(\I,\I)$ with the supremum metric denoted by
$\norm{\cdot}$ so that $(\mathcal{C}(\I,\I), \norm{\cdot})$
is a complete metric space.
Next we define the map $T_0$.
\begin{definition}[The map $T_0$]\label{T0mapDefi}
Assume first that $\theta \in \wIBD[0]$ and let $\is = \bt[0]{\theta}$
(that is $\theta \in \wbasint{\is}$).
In this case we set:
\[
f_{0,\theta}(x) = \begin{cases}
g_{_{\is,\theta}}(x)
& \text{if $x\in \I_{\is,\theta}$}, \\[0.75ex]
\frac{g_{_{\is,\theta}}\left(m_{\is}(\theta)\right) - 2}{m_{\is}(\theta) + 2} (x + 2) + 2
& \text{if $x \in [-2,m_{\is}(\theta)]$},\\[1ex]
\frac{g_{_{\is,\theta}}\left(M_{\is}(\theta)\right) + 2}{M_{\is}(\theta) - 2} (x - 2) - 2
& \text{if $x \in [M_{\is}(\theta),2]$}.
\end{cases}
\]
If $\theta \in \SI \setminus \wIBD[0]$ then we define
$f_{0,\theta}$ to be the unique piecewise affine map
with two affine pieces whose graph joins the point
$(-2,2)$ with $(0, \gamma(R_\omega(\theta))),$
and this with the point $(2, -2)$.
\end{definition}
Next we introduce some more notation to be able to define the maps
$\{T_m\}_{m=1}^{\infty}.$
For every $k \in \Z$ we set
\[
\wbasband{k} := \setsilift{\wbasint{k}} = \wbasint{k} \times \I
\]
and, for every $m \in \Z^+,$
\[
\wIVD := \setsilift{\wIBD} = \wIBD \times \I = \bigcup_{i \in \DS} \wbasband{i}.
\]
\begin{definition}[The maps $T_m$ with $m > 0$]\label{seqTmDefi}
Now we assume that we have defined the function
$T_{m-1}$ for some $m \ge 1$ and we define
\[
T_m(\theta,x) = (R_\omega(\theta), f_m(\theta, x))
\]
as follows.
By Lemma~\ref{Dsets}(b),
for every $(\theta, x) \in \wIVD,$ we have
\[
\theta \in \wbasint{\is} \subset \wIBD
\andq[with]
\is = \bt{\theta} \in \DS
\]
(and, of course, $x \in \I$).
Then we define:
\[
f_{m,\theta}(x) = \begin{cases}
f_{m-1,\theta}(x)
& \text{if $\theta \in \SI \setminus \wIBD;\ x\in \I$},\\
g_{_{\is,\theta}}(x)
& \text{if $\theta \in \IBD;\ x \in \I_{\is,\theta}$}, \\[0.75ex]
\frac{2 - g_{_{\is,\theta}}\left(m_{\is}(\theta)\right)\hfill}{
2 - f_{m-1,\theta}\left(m_{\is}(\theta)\right)
} (f_{m-1,\theta}(x) - 2) + 2
& \text{if $\theta \in \IBD;\ x \in [-2,m_{\is}(\theta)]$},\\[1ex]
\frac{2 + g_{_{\is,\theta}}\left(M_{\is}(\theta)\right)\hfill}{
2 + f_{m-1,\theta}\left(M_{\is}(\theta)\right)
} (f_{m-1,\theta}(x) + 2) - 2
& \text{if $\theta \in \IBD;\ x \in [M_{\is}(\theta),2]$},\\
\gams{\abs{\is+1}}\left(R_\omega(\theta)\right)
& \text{if $\theta \in \WB;\ x \in \IW{\theta}$}, \\[0.75ex]
\frac{2 - \gams{\abs{\is+1}}\left(R_\omega(\theta)\right)\hfill}{
2 - f_{m-1,\theta}\left(\lambda_m(\theta)\right)
} (f_{m-1,\theta}(x) - 2) + 2
& \text{if $\theta \in \WB;\ x \in [-2,\lambda_m(\theta)]$},\\[1ex]
\frac{2 + \gams{\abs{\is+1}}\left(R_\omega(\theta)\right)\hfill}{
2 + f_{m-1,\theta}\left(\tau_m(\theta)\right)
} (f_{m-1,\theta}(x) + 2) - 2
& \text{if $\theta \in \WB;\ x \in [\tau_m(\theta),2]$}.
\end{cases}
\]
Since $\wIVD \subset \wIVD[m-1],$ $f_{m-1,\theta}$ is defined on $\wIVD$.
Moreover, the above formula defines $f_{m,\theta}$ for every
$\theta \in \wIBD$ since, by Definition~\ref{curvesinthewings},
$\wIBD = \IBD \cup \WB.$
We also remark that $f_{m,\theta}$ formally is defined in two different
ways when $\theta \in \WB \cap \IBD.$ Later on we will show that
$f_{m,\theta}$ is well defined.
\end{definition}
The next proposition studies the maps $\{T_m\}_{m=0}^{\infty}$
and describes their properties.
\begin{proposition}\label{seqTmProperties}\label{T0mapProperties}
The following statements hold for every $m \in \Z^+.$
\begin{enumerate}[(a)]
\item The map $T_m$ is well defined, continuous
and belongs to $\cSO$.
\item For every $\theta\in\SI,$
$f_{m,\theta}$ is non-increasing, and
$f_{m,\theta}(2) = -2,$ $f_{m,\theta}(-2) = 2.$
Moreover,
$-1 \le f_{0,\theta}\left(M_{\bt{\theta}}(\theta)\right) \le
f_{0,\theta}\left(m_{\bt{\theta}}(\theta)\right) \le 1$
for every $\theta \in \wIBD.$
\item For every $i \in \DS,$
$T_m\evalat{\wbasicbox{i}} = G_i,$
$T_m\left(\setfibpt{\A_{\ai}}{\istar}\right) = \setfibpt{\A_{\aii}}{\sstar{i+1}},$
and\newline
$T_k\evalat{\istarset \times \I} = T_m\evalat{\istarset \times \I}$
(that is, $f_{k,\istar} = f_{m,\istar}$) for every $k > m.$
\end{enumerate}
\end{proposition}
The next result shows that the sequence $\{T_m\}_{m=0}^{\infty}$
has a limit in $\cSO$.
\begin{proposition}\label{distTmTm-1}
For every $m \ge 2$ and $\theta \in \SI,$
\begin{equation}\label{fitanorma}
\norm{f_{m,\theta} - f_{m-1,\theta}} \le 2 \cdot 2^{-\abs{\bt[m-1]{\theta}}}.
\end{equation}
Moreover, the sequence $\{T_m\}_{k = 0}^{\infty}$ is a Cauchy sequence.
\end{proposition}
Finally we are ready to prove the main result of the paper.
It follows from the next result which gives a more concrete version
of Theorem~\ref{MainTh}.
\begin{theorem}
There exists a map $T \in \cSO$ with
$f(\theta,\cdot)$ non-increasing for every $\theta \in \SI,$
such that $T$ permutes the upper and lower circles of $\Omega$
(thus having a periodic orbit of period two of curves),
and there exists a connected pseudo-curve $\A \subset \Omega$
which does not contain any arc of a curve
such that $T(\A) = \A$ and there does not exist
any $T$-invariant curve.
\end{theorem}
\begin{proof}
By Propositions~\ref{seqTmProperties} and \ref{distTmTm-1},
there exists a map
\[
T(\theta, x) = (R_\omega(\theta), f(\theta, x))
= (R_\omega(\theta), \lim_{m\to\infty} f_m(\theta, x)) \in \cSO
\]
with $f(\theta,\cdot)$ non-increasing for every $\theta \in \SI$
such that $T$ permutes the upper and lower circles of $\Omega$
(that is, $f(\theta,2) = -2$ and $f(\theta,-2) = 2$).
As the connected set $\A$ we take the one given by
Proposition~\ref{teoremacentral} (and Definition~\ref{PCAtLast}).
To end the proof of the theorem we need to show that $T(\A) = \A,$
since this already implies that there does not exist any $T$-invariant curve.
To see it, assume by way of contradiction that there exists an invariant curve
and denote its graph by $B.$
Since $B$ is the graph of a (continuous) curve, it is
compact and connected.
On the other hand, let $\Omega_{+}$ and $\Omega_{-}$ be the two
connected components of $\Omega\setminus \A$
from the proof of Proposition~\ref{teoremacentral}.
The facts that
$T(\A) = \A,$
$f(\theta,\cdot)$ is decreasing for every $\theta \in \SI,$ and
$T$ permutes the upper and lower circles of $\Omega$
imply that
$T(\Omega_{+}) = \Omega_{-}$
and
$T(\Omega_{-}) = \Omega_{+}.$
Hence, by the invariance of $B,$
$B \nsubseteq \Omega_{+}$ and $B \nsubseteq \Omega_{-}.$
The connectivity of $\A$ and $B$ imply that there exists
$(\theta,x) \in \A \cap B.$
Consequently,
\[
B = \overline{\set{T^n(\theta,x)}{n\in\Z^+}} \subset \A;
\]
a contradiction because $\A$ does not contain any arc of a curve.
So, only it remains to prove that $T(\A) = \A.$
By using Proposition~\ref{seqTmProperties}(c)
and Lemma~\ref{propiedadesPCA}(b) we get that
$T_m\left(\setfibpt{\A}{\istar}\right) = \setfibpt{\A}{\sstar{i+1}},$
and
$T_k\evalat{\setfibpt{\A}{\istar}} = T_m\evalat{\setfibpt{\A}{\istar}}$
for every $k,m \in \Z^+,\ k\ge m$ and $i \in \DS.$
Consequently, by the definition of the map $T$ we have,
$T(\setfibpt{\A}{\istar}) = \setfibpt{\A}{\sstar{i+1}}$
for every $i\in\Z$ or, equivalently,
$T\bigl(\setfib{\A}{\Orbom}\bigr) = \setfib{\A}{\Orbom}.$
Now we consider $\setfibth{\A}$ with
$\theta \in \SI\setminus\Orbom.$
Since $\Orbom$ is dense in $\SI,$ there exists a sequence
$\{(\theta_{n}, x_{n})\}_{n=0}^{\infty} \subset \setfib{\A}{\Orbom}$
such that
$
\lim_{n\to\infty} \theta_{n} = \theta.
$
By the compacity of $\A$ we can assume (by taking a convergent subsequence, if necessary)
that $\{(\theta_{n}, x_{n})\}_{n=0}^{\infty}$ is convergent to a point $(\theta,x) \in \A.$
By Lemma~\ref{propiedadesPCA}(c),
$\setfibth{\A} = (\theta,x)$ (and $x = \gamma(\theta)$).
On the other hand, by the part of the statement already proven,
$T(\theta_{n}, x_{n}) \in \A$ for every $n.$
Hence, by the continuity of $T$ and the compacity of $\A,$
\[
T(\theta,x) = (R_\omega(\theta), f(\theta,x))
= \lim_{n\to\infty} T(\theta_{n}, x_{n})
\in \setfibpt{\A}{R_\omega(\theta)}.
\]
Since $\theta \notin \Orbom$
we have that $R_\omega(\theta) \notin \Orbom$ and,
again by Lemma~\ref{propiedadesPCA}(c),
$\setfibpt{\A}{R_\omega(\theta)}$ consists of a unique point.
Hence, $T(\setfibth{\A}) = \setfibpt{\A}{R_\omega(\theta)}$
for every $\theta \in \SI\setminus\Orbom.$
Equivalently,
$
T\Bigl(\setfib{\A}{\bigl(\SI\setminus\Orbom\bigr)}\Bigr) =
\setfib{\A}{\bigl(\SI\setminus\Orbom\bigr)}.
$
This ends the proof of the theorem.
\end{proof}
\section{Proof of Proposition~\ref{T0mapProperties} in the case $m=0$}\label{proofofT0mapProperties}
This section is devoted to prove Proposition~\ref{T0mapProperties} for $m=0$;
that is, to study the map $T_0$.
It is the first technical counterpart of Section~\ref{skew-product}.
To prove Proposition~\ref{T0mapProperties} for $T_0$ we will need some
more notation and a technical lemma.
Given a skew product
$F(\theta,x) = (R_\omega(\theta), \zeta(\theta,x)$
from $\Omega = \SI \times \I$ to itself
we define the \emph{fibre map function of $F,$}
{\map{\mathsf{fib}(F)}{\SI}[\mathcal{C}(\I,\I)]}
by $\mathsf{fib}(F)(\theta) := \zeta(\theta, \cdot).$
A simple exercise shows that $F$ is continuous if and only if
$\zeta(\theta, \cdot)$ is continuous
for every $\theta \in \SI,$ and $\mathsf{fib}(F)$ is continuous.
\begin{lemma}\label{f0alesvores}
Let $\theta \in \Bd\left(\wbasint{i}\right)$ for some $i \in \DS[0].$
Then, $m_i(\theta) = M_i(\theta) = 0,$
$g_i(\theta,m_i(\theta)) = \gamma(R_\omega(\theta)),$
and $f_{0,\theta}$ is the unique piecewise affine map
with two affine pieces whose graph joins the point
$(-2,2)$ with $(0, \gamma(R_\omega(\theta))),$
and this with the point $(2, -2)$.
\end{lemma}
\begin{proof}
By Lemma~\ref{voresdelescaixesalesvores}(d) and Definition~\ref{T0mapDefi},
we have $m_i(\theta) = M_i(\theta).$
Hence, $f_{0,\theta}$ is the piecewise affine map with two affine pieces
whose graph joins the point $(-2,2)$ with
$\left(m_{i}(\theta), g_{_{i,\theta}}\left(m_{i}(\theta)\right)\right),$
and this with the point $(2, -2)$.
So, we need to show that $m_{i}(\theta) = 0,$ and
$g_{_{i,\theta}}\left(m_{i}(\theta)\right) = \gamma(R_\omega(\theta)).$
Lemma~\ref{Propertiesvarphi}(g) and the fact that $\dep{i} = 0,$
$\wbasint{i} \cap \wbasint{\ell} = \emptyset$ for every $\ell\in Z_{\ai},$
$i \ne \ell.$
Consequently, by Definition~\ref{PCgenerators}(\tsfR.6),
$m_i(\theta) = M_i(\theta) = a_i^- = 0.$
Now we show that
$g_{_{i,\theta}}(m_i(\theta)) = \gamma(R_\omega(\theta)).$
From the definition of the map $g_i$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa}),
Lemma~\ref{Propertiesvarphi}(e) and
Definitions~\ref{gammalimit} and \ref{PCgenerators}(\tsfR.1),
we get
\[
g_{_{i,\theta}}(m_i(\theta))
= \gams{\aii}(R_\omega(\theta))
= \gamma(R_\omega(\theta)).
\]
This ends the proof of the lemma.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{T0mapProperties} for $m=0$]
By Lemma~\ref{Propertiesvarphi}(b),
\[ -1 \le m_{\bt[0]{\theta}}(\theta) \le M_{\bt[0]{\theta}}(\theta) \le 1 \]
for every $\theta \in \wIBD[0].$
So, $T_0$ is well defined.
\inidemopart{b}
If $\theta \in \SI\setminus\wIBD[0],$ then the statement follows directly from
Definition~\ref{T0mapDefi}.
Now assume that $\theta \in \wIBD[0]$ and let $i = \bt[0]{\theta}.$
From the definition of the maps $g_{i,\theta}$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa})
and Definition~\ref{T0mapDefi}, it follows that
$f_{0, \theta}\evalat{\I_{i,\theta}}$ is piecewise affine and non-increasing.
On the other hand, again by Definition~\ref{T0mapDefi},
$f_{0, \theta}\evalat{[-2, m_i(\theta)]}$ and $f_{0, \theta}\evalat{[M_i(\theta), 2]}$
are affine with negative slope and $f_{0,\theta}(2) = -2$ and $f_{0,\theta}(-2) = 2.$
The fact that
\[
-1 \le f_{0,\theta}\left(M_{\bt[0]{\theta}}(\theta)\right) \le
f_{0,\theta}\left(m_{\bt[0]{\theta}}(\theta)\right) \le 1
\]
for every $\theta \in \wIBD[0]$ follows from Definition~\ref{T0mapDefi}
and Lemmas~\ref{gpositiva}(a) and \ref{gnegativa}(a).
This ends the proof of (b).
\inidemopart{c}
Recall that
\[
\wbasicbox{i} =
\LSleftlimits{\bigcup}{\theta \in \wbasint{i}} \{\theta\} \times \I_{i,\theta}.
\]
Hence, from Definition~\ref{T0mapDefi} and the definition of $G_i$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa})
it follows that
\[
T_m(\theta,x) = \bigl(R_\omega(\theta), f_m(\theta, x)\bigr)
= \bigl(R_\omega(\theta), g_{_{i,\theta}}(x)\bigr)
= G_i(\theta, x),
\]
for every $(\theta,x) \in \wbasicbox{i}.$ Thus,
$
T_0\left(\setfibpt{\A_{\ai}}{\istar}\right)
= \setfibpt{\A_{\aii}}{\sstar{i+1}}
$
from Lemmas~\ref{propiedadesA}(b), \ref{gpositiva}(c) and \ref{gnegativa}(c).
On the other hand, Lemma~\ref{denso}(b) implies that
$\istar \in \wIBD[0]$ but $\istar \notin \wIBD[k]$
for every $k \in \N.$
Then, we get $f_{k,\istar} = f_{0,\istar}$
from Definition~\ref{seqTmDefi}.
\inidemopart{a}
Since $T_0$ is a skew product with base $R_\omega$
we only have to prove that $f_0$ is continuous.
By Definition~\ref{T0mapDefi}, for every $\theta \in \SI$,
the map $f_{0,\theta}$ is continuous.
So we have to prove that the map
$\mathsf{fib}(T_0)$ (that is, the map $s \mapsto f_{0,s}$)
is continuous.
In the rest of the proof we will denote
\[
\IndSetWWings{IB}{0} := \bigcup_{i \in \DS[0]} \wobasint{i} \subset \wIBD[0].
\]
Clearly, since for every $i \in \Z,$
the maps $m_i$ and $M_i$ are continuous
on $\wbasint{i},$ it follows that the map
$s \mapsto f_{0,s}$ is continuous on $\IndSetWWings{IB}{0}.$
Thus, we have to see that the fibre map function
is continuous at every
$\theta \in \SI \setminus \IndSetWWings{IB}{0};$
that is,
$\lim_{j\to\infty} f_{0,\theta_j} = f_{0,\theta}$
for every
$\{\theta_j\}_{j=1}^{\infty} \subset \SI$ converging to $\theta.$
Given $\alpha > 0,$
we can consider four sets associated to such a sequence:
\begin{align*}
& \set{j\in \N}{\theta_j \in \SI \setminus \IndSetWWings{IB}{0}},\quad
\set{j\in \N}{\theta_j \in \IndSetWWings{IB}{0} \setminus \ball{\theta}{\alpha}},\\
& \set{j\in \N}{\theta_j \in (\theta, \theta+\alpha) \cap \IndSetWWings{IB}{0}} \andq
\set{j\in \N}{\theta_j \in (\theta-\alpha, \theta) \cap \IndSetWWings{IB}{0}}.
\end{align*}
Observe that the second set
$\set{j\in \N}{\theta_j \in \IndSetWWings{IB}{0} \setminus \ball{\theta}{\alpha}}$
is always finite and that any of the other three sets gives rise to
a subsequence of $\{\theta_j\}_{j=1}^{\infty}$ converging to $\theta,$
when it is infinite.
Consequently, the continuity of the fibre map function $s \mapsto f_{0,s}$
at $\theta$ is equivalent to the fact that
$\lim_{j\to\infty} f_{0,\theta_j} = f_{0,\theta}$
for every
$\{\theta_j\}_{j=1}^{\infty}$ converging to $\theta$
and such that, for some $\alpha > 0,$
$\{\theta_j\}_{j=1}^{\infty}$ is contained either in
$\SI\setminus\IndSetWWings{IB}{0},$ or
$(\theta, \theta+\alpha) \cap \IndSetWWings{IB}{0},$ or
$(\theta-\alpha, \theta) \cap \IndSetWWings{IB}{0}.$
We will only deal with the first two cases
since the proof in the last case (for $(\theta-\alpha, \theta)$)
can be done symmetrically.
\begin{case}{Case 1:}
$\lim_{j\to\infty} \theta_j = \theta$ and
$\{\theta_j\}_{j=1}^{\infty} \subset \SI\setminus\IndSetWWings{IB}{0}.$
\end{case}
By Definition~\ref{T0mapDefi} and Lemma~\ref{f0alesvores},
$f_{0,\theta_j}$ (respectively $f_{0,\theta}$)
is the unique piecewise affine map with two affine pieces
whose graph joins the point
$(-2,2)$ with $(0, \gamma(R_\omega(\theta_j)))$
(respectively $(0, \gamma(R_\omega(\theta)))$),
and this with the point $(2, -2)$.
By Lemma~\ref{denso}(c) and Definition~\ref{gammalimit}
the function $\gamma$ is continuous at
$R_\omega(\theta) \notin \Orbom.$
Hence,
$\lim_{j\to\infty} \gamma(R_\omega(\theta_j)) = \gamma(R_\omega(\theta))$
and, thus,
$\lim_{j\to\infty} f_{0,\theta_j} = f_{0,\theta}.$
\begin{case}{Case 2:}
$\lim_{j\to\infty} \theta_j = \theta$ and
$\{\theta_j\}_{j=1}^{\infty} \subset (\theta, \theta+\alpha) \cap \IndSetWWings{IB}{0}.$
\end{case}
If there exists $i \in \DS[0]$ such that
$\theta$ is the left endpoint of $\wbasint{i} \subset \wIBD[0]$
then the result follows from Definition~\ref{T0mapDefi}, the
continuity of the maps $m_i$ and $M_i$ and
the continuity of the maps $g_i$
(Lemmas~\ref{gpositiva}(a) and \ref{gnegativa}(a)).
Assume now that $\theta$ is not the left endpoint of $\wobasint{i}$
for every $i \in \DS[0].$
For every $j \in \N$ we set $\is_j := \bt[0]{\theta_j} \in \DS[0]$
(that is, $\theta_j \in \wobasint{\is_j}$).
We claim that $\lim_{j\to\infty} \abs{\is_j} = \infty$
and consequently, by Definition~\ref{PCgenerators}(\tsfR.1),
\begin{equation}\label{limit-ns-zero}
\lim_{j\to\infty} 2^{-n_{\abs{\is_j + 1}}} =
\lim_{j\to\infty} 2^{-n_{\abs{\is_j}}} = 0.
\end{equation}
To prove this claim, assume by way of contradiction that
there exists $L$ such that
for every $k \in \N$ there exists $j_k \ge k$
such that $\abs{\is_{j_k}} \le L.$
Then,
\[
\{\theta_{j_k}\}_{k=1}^{\infty} \subset
\bigcup_{k=1}^\infty \wobasint{\is_{j_k}}
\]
and, since $\set{\is_{j_k}}{k\in \N}$ is finite, it follows that
there exists $i \in \set{\is_{j_k}}{k\in \N} \subset \DS[0]$
and a subsequence of $\{\theta_{j_k}\}_{k=1}^{\infty},$
that by abuse of notation will also be called
$\{\theta_{j_k}\},$
such that
$\{\theta_{j_k}\}_{k=1}^{\infty} \subset \wobasint{i}.$
So,
\[
\theta = \lim_{k\to\infty} \theta_{j_k} \in \wbasint{i};
\]
a contradiction. So, the claim (and hence \eqref{limit-ns-zero}) holds.
Next we claim that the conditions
\begin{align}
& \lim_{j\to\infty} M_{\is_j}(\theta_j)
= \lim_{j\to\infty} m_{\is_j}(\theta_j) = 0,\text{ and}\label{ElqueCalDemostrar1}\\
&\text{\parbox{0.9\textwidth}{there exists a sequence $\{x_j\}_{j=1}^{\infty}$ with
$x_j \in \I_{\is_j,\theta_j} = [m_{\is_j}(\theta_j), M_{\is_j}(\theta_j)]$
for every $j,$ such that $\lim\limits_{j\to\infty} f_{0,\theta_j}(x_j) = \gamma(R_\omega(\theta))$}}\label{ElqueCalDemostrar2}
\end{align}
imply
\[ \lim_{j\to\infty} f_{0,\theta_j} = f_{0,\theta}. \]
To prove the claim notice that,
by Definition~\ref{T0mapDefi} and Lemma~\ref{f0alesvores},
$f_{0,\theta}$ is the unique piecewise affine map
with two affine pieces whose graph joins the point
$(-2,2)$ with $(0, \gamma(R_\omega(\theta))),$
and this with the point $(2, -2).$
On the other hand, for every $j,$
\begin{itemize}
\item $f_{0,\theta_j}\evalat{[-2, m_{\is_j}(\theta_j)]}$
is the affine map joining the point $(-2,2)$ with the point
$(m_{\is_j}(\theta_j), g_{\is_j}(\theta_j, m_{\is_j}(\theta_j))),$
and
\item $f_{0,\theta_j}\evalat{[M_{\is_j}(\theta_j),2]}$
is the affine map joining the point
$(M_{\is_j}(\theta_j), g_{\is_j}(\theta_j, M_{\is_j}(\theta_j)))$
with the point$(2,-2)$
\end{itemize}
(see Figure~\ref{fig:mapsF0}).
Moreover, from the part of the proposition already proven we know that
$f_{0,\theta_j}$ is non-increasing and continuous.
\begin{figure}
\begin{tikzpicture}[scale=1.5]
\draw (-2,-2) rectangle (2,2);
\foreach \c in {-2, 2} { \node[below] at (\c,-2) {$\c$}; \node[left] at (-2,\c) {$\c$}; }
\draw[dashed, color=blue] (0,-2) -- (0,2); \node[below, color=blue] at (0,-1.98) {\scriptsize$0$};
\draw[dashed, color=blue] (-2, 0.95) -- (2, 0.95);
\node[left, color=blue] at (-1.97,0.95) {\scriptsize$\gamma(R_\omega(\theta))$};
\draw[dashed, color=red] (-0.2,-2.2) -- (-0.2,2); \draw[dashed, color=red] (0.1,-2.2) -- (0.1,2);
\draw[very thick, color=red] (-0.2,-2.22) -- (0.1,-2.22);
\node[below, color=red] at (-0.35,-2.18) {\scriptsize$\I_{\is_j, \theta_j} = [m_{\is_j}(\theta_j),M_{\is_j}(\theta_j)]$};
\draw[dashed, color=red] (-2, 0.8) -- (2.05, 0.8); \draw[dashed, color=red] (-2, 0.7) -- (2.05, 0.7);
\draw[very thick, color=red] (2.05,0.8) -- (2.05, 0.7);
\node[right, color=red] at (2.05,0.7) {\scriptsize$\I_{\is_j+1,R_\omega(\theta_j)}$};
\draw[very thick, color=red] (-2,2) -- (-0.2, 0.8) -- (0.1, 0.7) -- (2,-2);
\node[color=red, left] at (-0.8,1.2) {$f_{0,\theta_j}$};
\draw[very thick, color=blue] (-2,2) -- (0, 0.95) -- (2,-2);
\node[color=blue, right] at (0.3,0.5) {$f_{0,\theta}$};
\end{tikzpicture}
\caption{A symbolic representation of the maps
$f_{0,\theta}$ and $f_{0,\theta_j}$ in Case~2 of
the proof of Proof of Proposition~\ref{T0mapProperties} for $m=0$.
The map \textcolor{blue}{$f_{0,\theta}$} and the points
\textcolor{blue}{$0$} and \textcolor{blue}{$\gamma(R_\omega(\theta))$}
are drawn in \textcolor{blue}{blue}.
The map \textcolor{red}{$f_{0,\theta_j}$} and the corresponding intervals
\textcolor{red}{$\I_{\is_j, \theta_j} $} and \textcolor{red}{$\I_{\is_j+1,R_\omega(\theta_j)}$}
are drawn in \textcolor{red}{red}.}\label{fig:mapsF0}
\end{figure}
Therefore, the claim holds provided that
\[
\lim_{j\to\infty} \diam\left(f_{0,\theta_j}\left(\I_{\is_j,\theta_j}\right)\right) = 0
\]
(see again Figure~\ref{fig:mapsF0}).
When $\theta_j \in \basint{\is_j} \setminus \OBG{\is_j}{\is_j+1}$ and $\is_j \ge 0,$
by Definitions~\ref{T0mapDefi} and \ref{defi-gi-positiva},
\[
\diam\left(f_{0,\theta_j}\left(\I_{\is_j,\theta_j}\right)\right) =
\diam\left(g_{_{\is_j,\theta_j}}\left(\I_{\is_j,\theta_j}\right)\right) =
\diam\left(\{\gams{\is_j+1}(R_\omega(\theta_j)\}\right) = 0.
\]
Otherwise, by Definition~\ref{T0mapDefi}, and
Lemmas~\ref{gpositiva}(b) and \ref{gnegativa}(b),
\begin{align*}
\{R_\omega(\theta_j)\} \times f_{0,\theta_j}\left(\I_{\is_j,\theta_j}\right)
& = \{R_\omega(\theta_j)\} \times g_{_{\is_j,\theta_j}}\left(\I_{\is_j,\theta_j}\right)
= G_{_{\is_j}}\bigl(\setfibpt{\basicbox{\is_j}}{\theta_j}\bigr) \\
& \subset \setfibpt{\basicbox{\is_j + 1}}{R_\omega(\theta_j)}.
\end{align*}
So, by Remark~\ref{propiedadesR}(2),
\[
\diam\left(f_{0,\theta_j}\left(\I_{\is_j,\theta_j}\right)\right) \le
\diam\left(\basicbox{\is_j + 1}\right) \le
2 \cdot 2^{-n_{\abs{\is_j + 1}}}.
\]
In any case,
\[
0 \le \diam\left(f_{0,\theta_j}\left(\I_{\is_j,\theta_j}\right)\right) \le
2 \cdot 2^{-n_{\abs{\is_j + 1}}}
\andq[for every] j \in \N
\]
and, by \eqref{limit-ns-zero},
$
\lim_{j\to\infty} \diam\left(f_{0,\theta_j}\left(\I_{\is_j,\theta_j}\right)\right) = 0.
$
This ends the proof of the claim.
By the last claim, to end the proof of the proposition
in the case $m = 0$ it is enough to show that \mbox{(\ref{ElqueCalDemostrar1}--\ref{ElqueCalDemostrar2})} hold.
We start by proving \eqref{ElqueCalDemostrar1}.
By Lemma~\ref{f0alesvores},
\[
m_{\is_j}(\Bd(\wbasint{\is_j})) = M_{\is_j}(\Bd(\wbasint{\is_j})) = 0,
\]
and from the definition of the maps $m_{\is_j}$ and $M_{\is_j},$
Definition~\ref{GenericBoxes} (or Lemma~\ref{voresdelescaixesalesvores})
and Remark~\ref{propiedadesR}(2),
for every $s \in \wobasint{\is_j}$ we get
\begin{equation}\label{claimfibres}
\begin{split}
& -1 \le m_{\is_j}(s) < 0 < M_{\is_j}(s) \le 1,\text{ and}\\
& M_{\is_j}(s) - m_{\is_j}(s)
= \diam\bigl(\I_{{\is_j},s}\bigr)
\le 2\cdot2^{-n_{\abs{\is_j}}}.
\end{split}
\end{equation}
So, \eqref{ElqueCalDemostrar1} holds by \eqref{limit-ns-zero}.
Now we prove \eqref{ElqueCalDemostrar2}.
By \eqref{gammathetaproperties},
\eqref{alphadeltaintervals} and
\eqref{alphadeltaintervalsineg},
it follows that
\begin{align*}
& m_{\is_j}(\theta_j)
< \gams{\abs{\is_j}}(\theta_j) < M_{\is_j}(\theta_j)
&& \text{if $\theta_j \ne \sstar{\is_j},$ and}\\
& m_{\is_j}(\theta_j)
< \gams{\abs{\is_j}-1}(\theta_j) = 0 < M_{\is_j}(\theta_j)
&& \text{if $\theta_j = \sstar{\is_j}.$}
\end{align*}
Also, from Definition~\ref{T0mapDefi},
the definitions of $G_i$ and $g_{_{i,\theta}}$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa}),
and Lemmas~\ref{gpositiva}(c) and \ref{gnegativa}(c)
we get
\begin{align*}
& f_{0,\theta_j}(\gams{\abs{\is_j}}(\theta_j))
= g_{_{\is_j,\theta_j}}(\gams{\abs{\is_j}}(\theta_j))
= \gams{\abs{\is_j+1}}(R_\omega(\theta_j))
&& \text{if $\theta_j \ne \sstar{\is_j},$}\\
& f_{0,\theta_j}(\gams{\is_j-1}(\theta_j))
= g_{_{\is_j,\theta_j}}(\gams{\is_j-1}(\theta_j))
= \gams{\is_j}(R_\omega(\theta_j))
&& \text{if $\theta_j = \sstar{\is_j}$ and $\is_j \ge 0,$ and}\\
& f_{0,\theta_j}(\gams{\abs{\is_j}-1}(\theta_j))
= g_{_{\is_j,\theta_j}}(\gams{\abs{\is_j+1}}(\theta_j))
= \gams{\abs{\is_j+2}}(R_\omega(\theta_j))
&& \text{if $\theta_j = \sstar{\is_j}$ and $\is_j < 0.$}
\end{align*}
Thus, to prove \eqref{ElqueCalDemostrar2}, we have to show that
\begin{equation}\label{thelimits}
\begin{cases}
\lim_{j\to\infty} \gams{\abs{\is_j+1}}(R_\omega(\theta_j))
= \gamma(R_\omega(\theta))
&\text{if $\theta_j \ne \sstar{\is_j},$}\\
\lim_{j\to\infty} \gams{\is_j}(R_\omega(\theta_j))
= \gamma(R_\omega(\theta))
&\text{if $\theta_j = \sstar{\is_j}$ and $\is_j \ge 0,$ and}\\
\lim_{j\to\infty} \gams{\abs{\is_j+2}}(R_\omega(\theta_j))
= \gamma(R_\omega(\theta))
&\text{if $\theta_j = \sstar{\is_j}$ and $\is_j < 0$}
\end{cases}
\end{equation}
(that is, we take
$x_j := \gams{\abs{\is_j}}(\theta_j)$ if $\theta_j \ne \sstar{\is_j},$
$x_j := \gams{\is_j-1}(\theta_j)$ if $\theta_j = \sstar{\is_j}$ and $\is_j \ge 0,$ and
$x_j := \gams{\abs{\is_j}-1}(\theta_j)$ if $\theta_j = \sstar{\is_j}$ and $\is_j < 0$).
Let $\varepsilon > 0.$
By Lemma~\ref{denso}(c) and Definition~\ref{PCgenerators}(\tsfR.1)
we have that $\theta \notin\Orbom$ and, hence, $R_\omega(\theta) \notin \Orbom.$
By the continuity of $\gamma$ on $\SI\setminus\Orbom$
and the fact that $\lim_{i\to\infty}\gams{i} = \gamma,$
there exist $\delta > 0$ and $L \in \N$
such that
\begin{align*}
& \abs{\gamma(R_\omega(\theta)) - \gamma(\widehat{\theta})} < \varepsilon/2
\quad\text{for every
$\widehat{\theta} \in \ball{R_\omega(\theta)}{\delta} \setminus \Orbom$,
and}\\
& d_{\infty}\left(\gamma,\gams{i}\right) < \varepsilon/2
\quad\text{for every $i \ge L.$}
\end{align*}
Then, since
$\lim_{j\to\infty} \theta_j = \theta$ and
$\lim_{j\to\infty} \abs{\is_j} = \infty,$
there exists $N \in \N$ such that
$\abs{\theta - \theta_j} < \delta/2,$
and $\abs{\is_j} \ge L+2$
for every $j \ge N.$
First we will show that
\[
\abs{\gamma(R_\omega(\theta)) - \gams{\abs{\is_j+1}}(R_\omega(\theta_j))}
\le \varepsilon
\]
for every $j \ge N$ such that $\theta_j \ne \sstar{\is_j}.$
To see it observe that,
by Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
$\theta_j, R_\omega(\theta_j) \notin \Zstar_{\abs{\is_j+1}}$
whenever $\theta_j \ne \sstar{\is_j}.$
Thus,
$\gams{\abs{\is_j+1}}$ is continuous at $R_\omega(\theta_j)$
by Lemma~\ref{Propertiesvarphi}(a).
Also, there exists a sequence
$
\{\widehat{\theta}_{j_{\ell}}\}_{\ell=1}^\infty \subset
\bigl(\ball{\theta_j}{\delta/2} \cap \wobasint{\is_j}\bigr) \setminus \Orbom
$
converging to $\theta_j,$
because $\SI \setminus \Orbom$ is dense in $\SI.$
Clearly, for every $j \ge N,$ we have
$
\{R_\omega(\widehat{\theta}_{j_{\ell}})\}_{\ell=1}^\infty \subset
\ball{R_\omega(\theta)}{\delta} \setminus \Orbom
$
and
$
\lim_{\ell\to\infty} R_\omega(\widehat{\theta}_{j_{\ell}})
= R_\omega(\theta_j).
$
Moreover, since
$
\{R_\omega(\widehat{\theta}_{j_{\ell}})\}_{\ell=1}^\infty
\subset \SI \setminus \Orbom \subset \SI \setminus \Zstar_{\abs{\is_j+1}},\
\gams{\abs{\is_j+1}}
$
is defined for every $R_\omega(\widehat{\theta}_{j_{\ell}}).$
Then, for every $j \ge N$ and $\ell \in \N$, we have
\begin{align*}
\abs{\gamma(R_\omega(\theta)) -
\gams{\abs{\is_j+1}}(R_\omega(\widehat{\theta}_{j_{\ell}}))
}\ \le\
& \abs{\gamma(R_\omega(\theta)) -
\gamma(R_\omega(\widehat{\theta}_{j_{\ell}}))
} + \\
& \hspace*{5em}
\abs{\gamma(R_\omega(\widehat{\theta}_{j_{\ell}})) -
\gams{\abs{\is_j+1}}(R_\omega(\widehat{\theta}_{j_{\ell}}))
}\\
<\ & \tfrac{\varepsilon}{2} + d_{\infty}\Bigl(\gamma,\gams{\abs{\is_j+1}}\Bigr)
< \varepsilon.
\end{align*}
Consequently,
\[
\abs{\gamma(R_\omega(\theta)) - \gams{\abs{\is_j+1}}(R_\omega(\theta_j))}
= \lim_{\ell\to\infty}
\abs{\gamma(R_\omega(\theta)) -
\gams{\abs{\is_j+1}}(R_\omega(\widehat{\theta}_{j_{\ell}}))
} \le \varepsilon
\]
This ends the proof of the first equality of \eqref{thelimits}.
The second and third equalities of \eqref{thelimits} follow as above
by replacing $\gams{\abs{\is_j+1}}$ by
$\gams{\is_j}$ (respectively $\gams{\abs{\is_j+2}}$),
and noting that
\[
R_\omega(\theta_j) = R_\omega(\sstar{\is_j}) =
\begin{cases}
\sstar{(\is_j+1)} \notin \Zstar_{\is_j}
& \text{ if $\is_j \ge 0,$ and}\\
\sstar{(-(\abs{\is_j}-1))} \notin \Zstar_{\abs{\is_j}-2}
& \text{ if $\is_j < 0.$}
\end{cases}
\]
This ends the proof of the continuity of $T_0,$
and the proposition for the case $m=0.$
\end{proof}
\section{Proof of Proposition~\ref{seqTmProperties} for $m > 0$}\label{proofofseqTmProperties}
This section is the second technical counterpart of Section~\ref{skew-product}
and is devoted to prove Proposition~\ref{seqTmProperties} for every map $T_m$ with $m > 0.$
To do this we will need some more technical results.
Also we will use the notion of fibre map function introduced
in the previous section.
The next two lemmas establish some basic properties
of the maps $T_m\evalat{\wIVD}$ and clarify some aspects
of Definition~\ref{seqTmDefi}.
\begin{lemma}\label{seqTmDefirevisited}
For every $m \in \N$ and for every $\theta \in \wIBD,$
\[
f_{m,\theta}\evalat{\I_{\is,\theta}} =
g_{_{\is,\theta}}\evalat{\I_{\is,\theta}},
\]
where $\is = \bt{\theta}.$
Moreover, assume that
$
\theta \in \WB \setminus \WIB.
$
Then,
\[
f_{m,\theta}(x) = \begin{cases}
g_{_{\is,\theta}}(x)
& \text{if $x \in \I_{\is,\theta}$}, \\[0.75ex]
\frac{2 - g_{_{\is,\theta}}\left(m_{\is}(\theta)\right)\hfill}{
2 - f_{m-1,\theta}\left(m_{\is}(\theta)\right)
} (f_{m-1,\theta}(x) - 2) + 2
& \text{if $x \in [-2,m_{\is}(\theta)]$},\\[1ex]
\frac{2 + g_{_{\is,\theta}}\left(M_{\is}(\theta)\right)\hfill}{
2 + f_{m-1,\theta}\left(M_{\is}(\theta)\right)
} (f_{m-1,\theta}(x) + 2) - 2
& \text{if $x \in [M_{\is}(\theta),2]$}.
\end{cases}
\]
\end{lemma}
\begin{proof}
We start by proving the first statement.
When $\theta\in\IBD$ there is nothing to prove.
So, assume that $\theta \in \wIBD \setminus \IBD.$
By Definition~\ref{curvesinthewings}, $\theta \in \WB,$
$\is < 0$ and $\theta \in \wbasint{\is} \setminus \obasintabs{\is}.$
By Lemma~\ref{VerticalIntervalsIntheWings}(b),
\[
\I_{\is, \theta} = \{\gams{\abs{\is}}(\theta)\} \subset \IW{\theta}.
\]
Consequently, by Definition~\ref{seqTmDefi} and the definition of
the maps $g_{_{\is,\theta}}$ for $\is < 0$
(Definition~\ref{defi-gi-negativa} ---
notice that $\I_{\is, \theta} \subset \wbasicbox{\is}$ by definition),
\[
f_{m,\theta}\left(\gams{\abs{\is}}(\theta)\right) =
\gams{\abs{\is+1}}\left(R_\omega(\theta)\right) =
g_{_{\is,\theta}}\left(\gams{\abs{\is}}(\theta)\right).
\]
So, the first statement holds.
Now we prove the second one.
By Lemma~\ref{VerticalIntervalsIntheWings}(b),
\[
\I_{\is, \theta} =
\{m_{\is}(\theta)\} = \{M_{\is}(\theta)\} =
\{\gams{\abs{\is}}(\theta)\} =
\{\lambda_m(\theta)\} = \{\tau_m(\theta)\} =
\IW{\theta}.
\]
Thus, by the part already proven, the formulas
\begin{align*}
&\begin{cases}
g_{_{\is,\theta}}(x)
& \text{if $x \in \I_{\is,\theta}$}, \\[0.75ex]
\frac{2 - g_{_{\is,\theta}}\left(m_{\is}(\theta)\right)\hfill}{
2 - f_{m-1,\theta}\left(m_{\is}(\theta)\right)
} (f_{m-1,\theta}(x) - 2) + 2
& \text{if $x \in [-2,m_{\is}(\theta)]$},\\[1ex]
\frac{2 + g_{_{\is,\theta}}\left(M_{\is}(\theta)\right)\hfill}{
2 + f_{m-1,\theta}\left(M_{\is}(\theta)\right)
} (f_{m-1,\theta}(x) + 2) - 2
& \text{if $x \in [M_{\is}(\theta),2]$},
\end{cases}\\
\intertext{and}\\
&\begin{cases}
\gams{\abs{\is+1}}\left(R_\omega(\theta)\right)
& \text{if $x \in \IW{\theta}$}, \\[0.75ex]
\frac{2 - \gams{\abs{\is+1}}\left(R_\omega(\theta)\right)\hfill}{
2 - f_{m-1,\theta}\left(\lambda_m(\theta)\right)
} (f_{m-1,\theta}(x) - 2) + 2
& \text{if $x \in [-2,\lambda_m(\theta)]$},\\[1ex]
\frac{2 + \gams{\abs{\is+1}}\left(R_\omega(\theta)\right)\hfill}{
2 + f_{m-1,\theta}\left(\tau_m(\theta)\right)
} (f_{m-1,\theta}(x) + 2) - 2
& \text{if $x \in [\tau_m(\theta),2]$},
\end{cases}
\end{align*}
coincide.
\end{proof}
\begin{lemma}\label{seqTmPropsInBasicintervals}
The following statements hold for every $m \in \N$ and $i \in \DS:$
\begin{enumerate}[(a)]
\item The map $T_m\evalat{\wbasband{i}}$ is well defined and continuous.
\item For every $\theta\in\wbasint{i},$
\begin{enumerate}[(b.i)]
\item $f_{m,\theta}(2) = -2$ and $f_{m,\theta}(-2) = 2,$
\item $f_{m,\theta}$ is piecewise affine and non-increasing, and
\item $-1 \le f_{m,\theta}\left(M_{i}(\theta)\right) \le
f_{m,\theta}\left(m_{i}(\theta)\right) \le 1.$
\end{enumerate}
\item $T_m\evalat{\wbasicbox{i}} = G_i$ and
$T_m\left(\setfibpt{\A_{\ai}}{\istar}\right) = \setfibpt{\A_{\aii}}{\sstar{i+1}}.$
\end{enumerate}
\end{lemma}
\begin{proof}
Clearly, $T_m\evalat{\wbasband{i}}$ is well defined and continuous if and only if
so is $f_m\evalat{\wbasband{i}}.$
We will prove by induction on $m \in \Z^+$ that,
(a), (b) and
\begin{enumerate}[(b.i)]\setcounter{enumi}{3}
\item $f_{m, \theta}\evalat{[-2, -1]}$ and $f_{m, \theta}\evalat{[1, 2]}$
are affine, $f_{m, \theta}(-1) < 2$ and $f_{m, \theta}(1) > -2$
\end{enumerate}
hold for every $\theta\in\wbasint{i}.$
First we will show that (a), (b) and (b.iv) hold for $m=0$ and $i \in \DS[0]$
(we are including the map $f_0$ studied earlier to correctly
start the induction process).
By Proposition~\ref{T0mapProperties}(a,b) for $m=0$ we have that
$T_{0}\evalat{\wbasband{i}}$ is well defined and continuous and
(b) holds. By Definition~\ref{T0mapDefi},
we also know that
$f_{m, \theta}\evalat{[-2, m_{i}(\theta)]}$
and
$f_{m, \theta}\evalat{[M_{i}(\theta), 2]}$
are affine. Then, (b.iv) follows from
$
-1 \le m_{i}(\theta) \le M_{i}(\theta) \le 1
$
(see Lemma~\ref{voresdelescaixesalesvores}(a)) and (b.iii).
Assume now that (a), (b) and (b.iv) hold for some
$m-1 \in \Z +$ and prove it for $m$ and $i \in \DS.$
By Lemma~\ref{Dsets}(a),
$\theta \in \wbasint{i} \varsubsetneq \wbasint{k}$
for some $k \in \DS[m-1].$
Consequently, $\wbasband{i} \subset \wbasband{k}$
and $f_{m-1}\evalat{\wbasband{i}}$ is well defined and continuous.
By Lemma~\ref{voresdelescaixesalesvores}(a) and Definition~\ref{curvesinthewings},
\begin{equation}\label{colocacioms}
\begin{split}
-1 \le m_{i}(\theta) \le M_{i}(\theta) \le 1&
\qquad\text{for $\theta\in\wbasint{i}$, and}\\
-1 \le \lambda_m(\theta) \le \tau_m(\theta) \le 1&
\qquad\text{for $\theta\in\wbasint{i}\setminus\obasintabs{i} \subset \WB$ ($i < 0$).}
\end{split}
\end{equation}
Consequently, by (b.ii) and (b.iv) for $m-1,$
\[
-2 < f_{m-1,\theta}(1) \le f_{m-1,\theta}\left(M_{i}(\theta)\right)
\le f_{m-1,\theta}\left(m_{i}(\theta)\right) \le f_{m,\theta}(-1) < 2
\]
for every $\theta\in\wbasint{i}$, and
\[
-2 < f_{m-1,\theta}(1) \le f_{m-1,\theta}\left(\tau_m(\theta)\right)
\le f_{m-1,\theta}\left(\lambda_m(\theta)\right) \le f_{m,\theta}(-1) < 2
\]
for $\theta\in\wbasint{i}\setminus\obasintabs{i} \subset \WB$ when $i < 0$.
On the other hand, as it was observed in Definition~\ref{seqTmDefi},
$f_{m,\theta}$ is defined in two different ways when
$\theta \in \WB \cap \IBD.$
In such a case,
by Lemmas~\ref{VerticalIntervalsIntheWings}(a,b)
and \ref{seqTmDefirevisited}, $\theta \notin \WIB$ and
both definitions for $f_{m,\theta}$ coincide.
Hence, $f_{m}\evalat{\wbasband{i}}$ is well defined.
Now we prove that $f_{m}\evalat{\wbasband{i}}$ is continuous
by using the continuity of $f_{m-1}\evalat{\wbasband{i}}.$
Since $\basintabs{i} \subset \IBD,$
by Definition~\ref{seqTmDefi},
the continuity of the maps $m_i$ and $M_i$ (see Lemma~\ref{voresdelescaixesalesvores}(b)),
and the continuity of the maps $g_i$ (Lemmas~\ref{gpositiva}(a) and \ref{gnegativa} (a)),
$f_m\evalat{\setsilift{\basintabs{i}}}$ is continuous.
Now we assume that $i < 0$ and we study the continuity of
$f_m\evalat{\setsilift{U}}$ on a connected component $U$ of
$\wbasint{i}\setminus\obasintabs{i}.$
Observe that, by Definition~\ref{curvesinthewings} and Lemma~\ref{Dsets}(b),
$U$ is a connected component of $\WB.$
Then, again by Definition~\ref{seqTmDefi},
the continuity of the maps $\lambda_m\evalat{U}$ and $\tau_m\evalat{U}$
(Lemma~\ref{continuouscurvesinthewings}),
and the continuity of the map $\gams{\ai}\evalat{U}$
(Lemma~\ref{Propertiesvarphi}(a) and
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2)),
$f_m\evalat{\setsilift{U}}$ is continuous.
Therefore, $f_{m}\evalat{\wbasband{i}}$ is continuous because it is well defined
on $\setsilift{\left(\left(\wbasint{i}\setminus\obasintabs{i}\right) \cap \basintabs{i}\right)}.$
Let $\theta \in \basintabs{i} \subset \IBD$.
By Definition~\ref{seqTmDefi},
and the definition of the maps $g_{i,\theta}$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa}),
$f_{m, \theta}\evalat{\I_{i,\theta}}$
is piecewise affine and non-increasing.
So, by Lemma~\ref{seqTmDefirevisited} for $m-1$ and Definition~\ref{seqTmDefi},
$f_{m,\theta}(2) = -2,$ $f_{m,\theta}(-2) = 2,$ and
$f_{m, \theta}\evalat{[-2, m_i(\theta)]}$ and
$f_{m, \theta}\evalat{[M_i(\theta), 2]}$
are affine transformations of the map $f_{m-1, \theta}$
with positive slope.
Hence, (b.i,ii) hold for $f_{m, \theta}$ in this case.
Moreover, (b.iv) is verified
by \eqref{colocacioms} and (b.iv) for $m-1.$
Consider $\theta \in \wbasint{i}\setminus\obasintabs{i} \subset \WB$.
Again by Definition~\ref{seqTmDefi},
$f_{m, \theta}\evalat{\IW{\theta}}$ is constant.
Then, (b.i,ii) and (b.iv) hold for $f_{m, \theta}$ as above
by replacing $m_i(\theta)$ and $M_i(\theta)$ by
$\lambda_m(\theta)$ and $\tau_m(\theta),$ respectively.
By (b.ii) and \eqref{colocacioms} we have
$
f_{m,\theta}\left(M_{i}(\theta)\right) \le
f_{m,\theta}\left(m_{i}(\theta)\right).
$
Hence, (b.iii) follows from Lemma~\ref{seqTmDefirevisited},
Definition~\ref{seqTmDefi},
Lemmas~\ref{gpositiva}(b) and \ref{propiedadesA}(c),
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
Lemma~\ref{gnegativa}(b) and Lemma~\ref{Propertiesvarphi}(b).
\inidemopart{c}
In a similar way to the proof of Proposition~\ref{T0mapProperties} for the case $m = 0$,
\[
\wbasicbox{i} =
\LSleftlimits{\bigcup}{\theta \in \wbasint{i}} \{\theta\} \times \I_{i,\theta}
\subset \wbasband{i} \subset \wIVD
\]
and, by Definition~\ref{seqTmDefi}, Lemma~\ref{seqTmDefirevisited}
and the definition of $G_i$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa})
it follows that
\[
T_m(\theta,x) = \bigl(R_\omega(\theta), f_m(\theta, x)\bigr)
= \bigl(R_\omega(\theta), g_{_{i,\theta}}(x)\bigr)
= G_i(\theta, x),
\]
for every $(\theta,x) \in \wbasicbox{i}.$
Thus,
$
T_m\left(\setfibpt{\A_{\ai}}{\istar}\right)
= \setfibpt{\A_{\aii}}{\sstar{i+1}}
$
from Lemmas~\ref{propiedadesA}(b), \ref{gpositiva}(c) and \ref{gnegativa}(c).
\end{proof}
The next technical lemma compares the images of
$f_{m,\theta}$ and $f_{m-1,\theta}$ on a point.
It is an extension of Lemma~\ref{QuePassaALesAles}.
\begin{lemma}\label{fmfm-1alesales}
Assume that $\wbasint{i} \subset \wbasint{k}$
for some $i \in \DS,\ k \in \DS[m-1]$ and $m \in \N.$
Then, for every $\theta \in \wbasint{i} \setminus \obasintabs{i},$
$m_i(\theta) = M_i(\theta) = \gams{i}(\theta)$ and
\begin{align*}
f_{m,\theta}\left(m_i(\theta)\right) & = g_{_{i, \theta}}\left(m_i(\theta)\right) =
\gams{\aii}\left(R_\omega(\theta)\right),\text{ and}\\
f_{m-1,\theta}\left(m_i(\theta)\right) & = g_{_{k, \theta}}\left(m_i(\theta)\right) =
\gams{\akk}\left(R_\omega(\theta)\right).
\end{align*}
\end{lemma}
\begin{proof}
The fact that $m_i(\theta) = M_i(\theta) = \gams{i}(\theta)$ follows directly from the definitions.
The first equation follows from Lemma~\ref{seqTmDefirevisited},
and the definition of the map $g_{_{i, \theta}}$
(Definitions~\ref{defi-gi-positiva} and \ref{defi-gi-negativa}).
By Lemma~\ref{QuePassaALesAles},
$
\I_{i, \theta} = \{m_i(\theta)\} = \{\gams{\ak}(\theta)\} \subset \I_{k, \theta}.
$
Moreover, as in the proof of Lemma~\ref{QuePassaALesAles},
$\theta \ne \kstar.$
Consequently, by Definition~\ref{T0mapDefi}, Lemma~\ref{seqTmDefirevisited},
Lemmas~\ref{gpositiva}(c) and \ref{gnegativa}(c) and \eqref{gammathetaproperties}
(alternatively, for the last equality check directly the proofs
of the Lemmas~\ref{gpositiva}(c) and \ref{gnegativa}(c)),
\[
f_{m-1,\theta}\left(m_i(\theta)\right) =
g_{_{k, \theta}}\left(m_i(\theta)\right) =
g_{_{k, \theta}}\left(\gams{\ak}(\theta)\right) =
\gams{\akk}\left(R_\omega(\theta)\right).
\]
\end{proof}
The following lemma is the analogue of Lemma~\ref{f0alesvores} for $m \ge 1.$
To state it we will use the set
\[
\setsilift{\wEIBD} = \wEIBD \times \I \subset \wIVD.
\]
\begin{lemma}\label{voresiguals}
$
T_m\evalat{\setsilift{\wEIBD}} =
T_{m-1}\evalat{\setsilift{\wEIBD}}
$
for every $m \in \N.$
Equivalently, $f_{m,\theta} = f_{m-1,\theta}$
for every $m \in \N$ and $\theta \in \wEIBD.$
\end{lemma}
\begin{proof}
Fix $m \in \N$ and $\theta \in \wEIBD \subset \wIBD.$
By Lemma~\ref{Dsets}(a,b), there exist $i \in \DS$ and $k \in \DS[m-1]$ such that
$\theta \in \Bd\left(\wbasint{i}\right) \subset \wbasint{i} \varsubsetneq \wbasint{k}.$
So, we are in the assumptions of
Lemmas~\ref{QuePassaALesAles} and \ref{fmfm-1alesales}
and, hence,
\begin{align*}
\I_{i, \theta} & =\{m_i(\theta)\} = \{\gams{\ai}(\theta)\} = \{\gams{\ak}(\theta)\} \subset \I_{k, \theta},\\
f_{m,\theta}\left(m_i(\theta)\right) & = g_{_{i, \theta}}\left(m_i(\theta)\right) =
\gams{\aii}\left(R_\omega(\theta)\right),\text{ and}\\
f_{m-1,\theta}\left(m_i(\theta)\right) & = g_{_{k, \theta}}\left(m_i(\theta)\right) =
\gams{\akk}\left(R_\omega(\theta)\right).
\end{align*}
Thus, if $i \ge 0,$ $\theta \in \IBD$ and,
by Definition~\ref{seqTmDefi} and
Lemma~\ref{seqTmPropsInBasicintervals}(a),
to prove that $f_{m,\theta} = f_{m-1,\theta}$
we only have to show that
\[
g_{_{i, \theta}}\left(m_i(\theta)\right) =
\gams{\aii}\left(R_\omega(\theta)\right) =
\gams{\akk}\left(R_\omega(\theta)\right) =
f_{m-1,\theta}\left(m_i(\theta)\right).
\]
When $i < 0,$ $\theta \in \WB \cap \wEIBD$
and, by Lemma~\ref{VerticalIntervalsIntheWings}(a),
$\theta \notin \WIB.$
Then, by Lemma~\ref{seqTmDefirevisited},
we get again that
\[
g_{_{i, \theta}}\left(m_i(\theta)\right) =
\gams{\aii}\left(R_\omega(\theta)\right) =
\gams{\akk}\left(R_\omega(\theta)\right) =
f_{m-1,\theta}\left(m_i(\theta)\right).
\]
implies $f_{m,\theta} = f_{m-1,\theta}.$
If $\akk = \aii$ there is nothing to prove.
So, by Lemma~\ref{QuePassaALesAles}, we can assume that $\akk < \aii$
and we have
\[
\gams{\akk}\left(R_\omega(\theta)\right) =
\gams{\akk+1}\left(R_\omega(\theta)\right) = \dots =
\gams{\aii-1}\left(R_\omega(\theta)\right).
\]
Hence, we have to show that
$
\gams{\aii}\left(R_\omega(\theta)\right) = \gams{\aii-1}\left(R_\omega(\theta)\right).
$
If $i \ge 0$ we get
\[
\gams{\aii}\left(R_\omega(\theta)\right) =
\gams{i+1}\left(R_\omega(\theta)\right) =
\gams{i}\left(R_\omega(\theta)\right) =
\gams{\aii-1}\left(R_\omega(\theta)\right)
\]
by Lemma~\ref{Propertiesvarphi}(e).
Otherwise we have $i < 0,$
$\theta \in \Bd\left(\wbasint{i}\right) = \Bd\left(\BSG{i}{\aii}\right)$
and, consequently, $R_\omega(\theta) \in \Bd\left(\BSG{i+1}{\aii}\right).$
Again by Lemma~\ref{Propertiesvarphi}(e) for $j = \aii$,
\[
\gams{\aii}\left(R_\omega(\theta)\right) =
\gams{\aii-1}\left(R_\omega(\theta)\right).
\]
This ends the proof of the lemma.
\end{proof}
Now we aim at computing two different kind of
upper bounds for
$\norm{f_{m,\theta} - f_{m-1,\theta}}$
(Lemma~\ref{distTmTm-1smallboxes} and Proposition~\ref{distTmTm-1}).
This will be a key tool in the proof of
Propositions~\ref{seqTmProperties} for $m > 0$ and \ref{distTmTm-1}.
The next two lemmas and remark will be useful to automate
and simplify the proofs of these two results.
\begin{lemma}\label{normainterna}
\[
\norm{f_{m,\theta} - f_{m-1,\theta}} =
\begin{cases}
\norm{f_{m,\theta}\evalat{\I_{\bt{\theta}, \theta}} - f_{m-1,\theta}\evalat{\I_{\bt{\theta}, \theta}}}
& \text{when $\theta \in \wIBD\setminus\WIB$, and}\\[10pt]
\norm{f_{m,\theta}\evalat{\IW{\theta}} - f_{m-1,\theta}\evalat{ \IW{\theta}}}
& \text{when $\theta \in \WIB$,}
\end{cases}
\]
for every $m \ge 2$ and $\theta \in \wIBD.$
\end{lemma}
\begin{proof}
Set $i = \bt{\theta} \in \DS$, so that $\theta \in \wbasint{i}$.
When $\theta \in \wIBD\setminus\WIB = \IBD \cup \WB \setminus \WIB,$
by Definition~\ref{seqTmDefi} and Lemma~\ref{seqTmDefirevisited},
it is enough to show that
\[
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} \le
\abs{f_{m,\theta}(m_i(\theta)) - f_{m-1,\theta}(m_i(\theta))}
\]
for every $x \in [-2, m_i(\theta)]$, and
\[
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} \le
\abs{f_{m,\theta}(M_i(\theta)) - f_{m-1,\theta}(M_i(\theta))}
\]
for every $x \in [M_i(\theta),2]$.
We will prove the first statement. The second one follows similarly.
Definition~\ref{seqTmDefi} and Lemma~\ref{seqTmDefirevisited} give
\begin{align*}
f_{m,\theta}(x) - f_{m-1,\theta}(x) &=
\frac{2 - g_{_{i,\theta}}\left(m_{i}(\theta)\right)\hfill}{2 - f_{m-1,\theta}\left(m_{i}(\theta)\right)}
(f_{m-1,\theta}(x) - 2) + 2 - f_{m-1,\theta}(x)\\
&= \frac{2 - f_{m,\theta}\left(m_{i}(\theta)\right)\hfill}{2 - f_{m-1,\theta}\left(m_{i}(\theta)\right)}
(f_{m-1,\theta}(x) - 2) - (f_{m-1,\theta}(x) - 2)\\
&=(f_{m-1,\theta}(x) - 2)\left(
\frac{2 - f_{m,\theta}\left(m_{i}(\theta)\right)\hfill}{2 - f_{m-1,\theta}\left(m_{i}(\theta)\right)}
- 1\right)\\
&=(2 - f_{m-1,\theta}(x)) \frac{f_{m,\theta}\left(m_{i}(\theta)\right) - f_{m-1,\theta}\left(m_{i}(\theta)\right)}{2 - f_{m-1,\theta}\left(m_{i}(\theta)\right)}.
\end{align*}
By Lemma~\ref{seqTmPropsInBasicintervals}(b),
$2 \ge f_{m-1,\theta}(x) \ge f_{m-1,\theta}\left(m_{i}(\theta)\right)$
and $1 \ge f_{m-1,\theta}\left(m_{i}(\theta)\right).$ Hence,
\begin{align*}
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} &=
(2 - f_{m-1,\theta}(x)) \frac{\abs{f_{m,\theta}\left(m_{i}(\theta)\right) - f_{m-1,\theta}\left(m_{i}(\theta)\right)}}{2 - f_{m-1,\theta}\left(m_{i}(\theta)\right)}\\
&\le \abs{f_{m,\theta}\left(m_{i}(\theta)\right) - f_{m-1,\theta}\left(m_{i}(\theta)\right)}.
\end{align*}
Now assume that $\theta \in \WIB \subset \WB.$
By Definition~\ref{seqTmDefi} it is enough to show that
\[
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} \le
\abs{f_{m,\theta}(\lambda_m(\theta)) - f_{m-1,\theta}(\lambda_m(\theta))}
\]
for every $x \in [-2, \lambda_m(\theta)]$, and
\[
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} \le
\abs{f_{m,\theta}(\tau_m(\theta)) - f_{m-1,\theta}(\tau_m(\theta))}
\]
for every $x \in [\tau_m(\theta),2]$.
As before, we will prove the first statement. The second one follows similarly.
We have
\[
f_{m,\theta}(x) - f_{m-1,\theta}(x) =
(2 - f_{m-1,\theta}(x)) \frac{f_{m,\theta}\left(\lambda_m(\theta)\right) - f_{m-1,\theta}\left(\lambda_m(\theta)\right)}{2 - f_{m-1,\theta}\left(\lambda_m(\theta)\right)}.
\]
By Lemma~\ref{seqTmPropsInBasicintervals}(b),
$2 \ge f_{m-1,\theta}(x) \ge f_{m-1,\theta}\left(\lambda_m(\theta)\right)$
and hence,
\[
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}
\le \abs{f_{m,\theta}\left(m_{i}(\theta)\right) - f_{m-1,\theta}\left(m_{i}(\theta)\right)}
\]
provided that $2 - f_{m-1,\theta}\left(\lambda_m(\theta)\right) \ne 0.$
Assume by way of contradiction that we have
$f_{m-1,\theta}\left(\lambda_m(\theta)\right) = 2.$
Then, by Definition~\ref{curvesinthewings} and
Lemma~\ref{seqTmPropsInBasicintervals}(b),
$-1 \le \lambda_m(\theta)$ and
\[
2 \ge f_{m-1,\theta}(-1) \ge f_{m-1,\theta}\left(\lambda_m(\theta)\right) = 2;
\]
which contradicts statement (b.iv) from the proof of
Lemma~\ref{seqTmPropsInBasicintervals}.
\end{proof}
Next we compute an upper bound for
$\norm{f_{m,\theta} - f_{m-1,\theta}}$
for every $\theta \in \wbasint{i}$ and
$i \in \DS$ such that $\diam(\wbasint{i})$ is small enough.
\begin{lemma}\label{distTmTm-1smallboxes}
Assume that $T_{m-1}$ is continuous for some $m \ge 2$
and let $\varepsilon$ be positive. Then, there exist
$\varrho_m(\varepsilon) \in \N$ such that
\[
\norm{f_{m,\theta} - f_{m-1,\theta}} \le \varepsilon
\]
for every $\theta \in \wbasint{i}$ and
$i \in \DS$ (that is, $\wbasint{i} \subset \wIBD$)
such that $\ai \ge \varrho_m(\varepsilon).$
\end{lemma}
\begin{proof}
Since $T_{m-1}$ is uniformly continuous, there exists
$\delta_{m-1} = \delta_{m-1}(\varepsilon) > 0$ such that
$\dom(T_{m-1}(\theta,x), T_{m-1}(\nu, y)) < \varepsilon$
provided that $\dom((\theta,x), (\nu, y)) < \delta_{m-1}.$
We choose $\varrho_m = \varrho_m(\varepsilon) \in \N$ such that
\[
3 \cdot 2^{-\varrho_m} < \min\{ \delta_{m-1}(\varepsilon/2), \varepsilon/2 \}.
\]
Assume that $i \in \DS$ verifies $\ai \ge \varrho_m(\varepsilon)$
and let $(\theta,x) \in \wbasband{i} = \wbasint{i} \times \I.$
When $\theta \in \wbasint{i} \setminus \WIB$
we can use Lemma~\ref{normainterna} with $\I_{i, \theta}$
to compute $\norm{f_{m,\theta} - f_{m-1,\theta}}.$
We have to show that
$\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} < \varepsilon$
for every $x \in \I_{i, \theta}.$
Let $\nu \in \Bd\left(\wbasint{i}\right) \subset \wEIBD$.
We have $(\theta,x),(\nu, m_i(\nu)) \in \wbasicbox{i}$ and,
by Lemmas~\ref{seqTmPropsInBasicintervals}(c) and \ref{Propertiesvarphi}(f),
\begin{align*}
\dom(T_m(\theta, x),T_m(\nu, m_i(\nu)))
& = \dom(G_i(\theta, x),G_i(\nu, m_i(\nu)))\\
& \le \diam\left(G_i\left(\wbasicbox{i}\right)\right),\text{ and}\\
\dom((\theta, x),(\nu, m_i(\nu))
& \le \diam\left(\wbasicbox{i}\right)
\le 2 \cdot 2^{-\ai} < 3 \cdot 2^{-\varrho_m} < \delta_{m-1}(\varepsilon/2).
\end{align*}
Thus,
\[ \dom(T_{m-1}(\theta, x),T_{m-1}(\nu, m_i(\nu)) < \varepsilon/2. \]
Consequently, by Lemma~\ref{voresiguals},
\begin{align*}
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}
& = \dom(T_m(\theta, x),T_{m-1}(\theta, x))\\
& \le \dom(T_m(\theta, x),T_{m-1}(\nu, m_i(\nu)))\ +\\
&\hspace*{4em}\dom(T_{m-1}(\nu, m_i(\nu)), T_{m-1}(\theta, x))\\
& < \dom(T_m(\theta, x),T_{m}(\nu, m_i(\nu))) + \varepsilon/2\\
& < \diam\left(G_i\left(\wbasicbox{i}\right)\right) + \varepsilon/2.
\end{align*}
Now we look at the size of $G_i\left(\wbasicbox{i}\right).$
When $i < 0$, from Lemmas~\ref{gnegativa}(b) and \ref{Propertiesvarphi}(f),
we obtain
\begin{equation}\label{diamineg}
\diam\left(G_i\left(\wbasicbox{i}\right)\right)
\le \diam\left(\basicbox{i+1}\right) \le 2^{-(\ai-1)} < 2 \cdot 2^{-\ai}.
\end{equation}
When $i \ge 0,$ from Lemma~\ref{gpositiva}(b) we get
\[
G_i\left(\wbasicbox{i}\right) = G_i\left(\basicbox{i}\right) \subset
\basicbox{i+1} \cup \setfib{\A_{i+1}}{\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right)}.
\]
Moreover, as in the proof of Lemma~\ref{Propertiesvarphi}(f) for $\ell < 0,$
the set
\[
\basicbox{i+1} \cup \setfib{\A_{i+1}}{\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right)}
\]
is connected. So, by Lemma~\ref{Propertiesvarphi}(f),
\begin{align*}
\diam\left(G_i\left(\wbasicbox{i}\right)\right)
&\le \diam\left(\basicbox{i+1} \cup \setfib{\A_{i+1}}{\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right)}\right)\\
&\le \diam\left(\basicbox{i+1}\right) + \diam\left(\setfib{\A_{i+1}}{\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right)}\right)\\
&\le 2^{-(i+1)} + \diam\left(\setfib{\A_{i+1}}{\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right)}\right).
\end{align*}
As noticed earlier,
$\BSG{i+1}{i} \setminus \obasint{i+1}$ is disjoint from
\[ \obasint{i+1} \cup \wbasint{-(i+1)} \cup \Zstar_{i+1} \]
by Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2).
So, by Lemma~\ref{propiedadesA}(c),
Definition~\ref{PCgenerators} and Lemma~\ref{Propertiesvarphi}(a),
\begin{align*}
\setfibpt{\A_{i+1}}{\nu}
&= \{(\nu, \gams{i+1}(\nu)\} = \{(\nu, \gams{i}(\nu)\}\\
&\in \{\nu\} \times \left[\gams{i}(\sstar{i+1})-2^{-n_{i}}, \gams{i}(\sstar{i+1}) + 2^{-n_{i}}\right].
\end{align*}
for every $\nu \in \BSG{i+1}{i} \setminus \obasint{i+1}.$
On the other hand,
$
\gams{i}(\sstar{i+1}) \in \I_{i+1,\sstar{i+1}}
$
by Lemma~\ref{Propertiesvarphi}(c).
Hence, by Remark~\ref{propiedadesR}(2),
Definition~\ref{PCgenerators}(\tsfR.1) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.1),
\begin{align*}
\diam&\left(\setfib{\A_{i+1}}{\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right)}\right)\\
&\le \max\left\{
\diam\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right),
2\cdot (2^{-n_{i}}+2^{-n_{i+1}})
\right\}\\
&\le 2\cdot \max\left\{ \alpha_i, 2^{-n_{i}}+2^{-n_{i+1}}
\right\} = 2\cdot (2^{-n_{i}}+2^{-n_{i+1}})\\
&< 4\cdot 2^{-n_{i}} \le 2 \cdot 2^{-i}.
\end{align*}
Summarizing, when $i \ge 0$,
\[
\diam\left(G_i\left(\wbasicbox{i}\right)\right)
\le 2^{-(i+1)} + 2 \cdot 2^{-i} < 3 \cdot 2^{-i}
\]
and, from \eqref{diamineg},
\[
\diam\left(G_i\left(\wbasicbox{i}\right)\right)
< 3 \cdot 2^{-\ai} \le 3 \cdot 2^{-\varrho_m} < \varepsilon/2
\]
for every $i \in \Z^+$.
Thus, for every $x \in \I_{i, \theta}$,
\[
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}
< \diam\left(G_i\left(\wbasicbox{i}\right)\right) + \varepsilon/2 < \varepsilon.
\]
Now assume that $\theta \in \wbasint{i} \cap \WIB.$
We can use Lemma~\ref{normainterna} with $\IW{\theta}$
to compute $\norm{f_{m,\theta} - f_{m-1,\theta}}.$
We have to show that
$\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} < \varepsilon$
for every $x \in \IW{\theta}.$
Since $\theta \in \WIB,$
by Definition~\ref{curvesinthewings} and Lemma~\ref{VerticalIntervalsIntheWings}(b),
$i < 0$, $\theta \in \WB$ and
\[
\I_{i, \theta} = \left\{\gams{\ai}(\theta)\right\}
\subset \IW{\theta} = \I_{\ell, \theta} \ni x
\]
with $\ell = \bt[\led{\theta}]{\theta} \in \WDS.$
In this case we will consider the points
$(\theta,x) \in \basicbox{\ell}$ and
$(\nu, m_i(\nu)),(\theta,\gams{\ai}(\theta)) \in \wbasicbox{i}$ with
$\nu \in \Bd\left(\wbasint{i}\right) \subset \wEIBD$.
By Lemma~\ref{DepthintheWings}(b), Remark~\ref{propiedadesR}(2)
and Lemma~\ref{Propertiesvarphi}(f), $\ai < \all$ and
\begin{align*}
\dom((\theta, x),(\nu, m_i(\nu))
& \le \dom((\theta, x),(\theta,\gams{\ai}(\theta)) +
\dom((\theta,\gams{\ai}(\theta)),(\nu, m_i(\nu))\\
& \le \abs{x - \gams{\ai}(\theta)} +
\diam\left(\wbasicbox{i}\right)\\
& \le \diam\left(\basicbox{\ell}\right) +
\diam\left(\wbasicbox{i}\right)\\
& \le 2^{-\all} + 2 \cdot 2^{-\ai} < 3 \cdot 2^{-\ai}
\le 3 \cdot 2^{-\varrho_m} < \delta_{m-1}(\varepsilon/2).
\end{align*}
Thus,
\[ \dom(T_{m-1}(\theta, x),T_{m-1}(\nu, m_i(\nu)) < \varepsilon/2. \]
On the other hand,
by Lemma~\ref{seqTmPropsInBasicintervals}(c), Definition~\ref{seqTmDefi}
and \eqref{diamineg},
\begin{align*}
\dom(T_m(\theta&, x),T_m(\nu, m_i(\nu)))\\
& \le \dom(T_m(\theta, x),T_m(\theta,\gams{\ai}(\theta))) +
\dom(T_m(\theta,\gams{\ai}(\theta)),T_m(\nu, m_i(\nu)))\\
& \le \abs{f_{m, \theta}(x) - f_{m, \theta}(\gams{\ai}(\theta))} +
\dom(G_i(\theta,\gams{\ai}(\theta)),G_i(\nu, m_i(\nu)))\\
& = \dom(G_i(\theta, x),G_i(\nu, m_i(\nu)))
\le \diam\left(G_i\left(\wbasicbox{i}\right)\right) < 2 \cdot 2^{-\ai}\\
& \le 3 \cdot 2^{-\varrho_m} < \varepsilon/2.
\end{align*}
So, in a similar way as before, Lemma~\ref{voresiguals} gives
\begin{align*}
\abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}
& = \dom(T_m(\theta, x),T_{m-1}(\theta, x))\\
& \le \dom(T_m(\theta, x),T_{m-1}(\nu, m_i(\nu)))\ +\\
&\hspace*{4em}\dom(T_{m-1}(\nu, m_i(\nu)), T_{m-1}(\theta, x))\\
& < \varepsilon.
\end{align*}
\end{proof}
\begin{proof}[Proof of Proposition~\ref{seqTmProperties} for $m > 0$]
\inidemopart{a}
We start by proving by induction on $m$ that $T_m$ is continuous for every $m \in \Z^+.$
By Proposition~\ref{T0mapProperties}(a) for $m=0$, $T_0$ is continuous.
So, we may assume that $T_{m-1}$ is continuous for some $m\in \N$
and prove that $T_m$ is continuous.
Let $\varepsilon > 0$ be fixed but arbitrary, and
let $(\theta,x), (\nu, y) \in \Omega.$
We have to show that there exists $\delta(\varepsilon) > 0$
such that
\[ \dom(T_m(\theta,x), T_m(\nu, y)) < \varepsilon
\andq[when]
\dom((\theta,x), (\nu, y)) < \delta.
\]
We start by defining $\delta(\varepsilon)$. To this end we need to
introduce some more notation and establish some facts
about the maps $T_m$ and $T_{m-1}.$
Since $T_{m-1}$ is uniformly continuous, we know that
\begin{equation}\label{Tmm1UC}
\text{\parbox{0.9\textwidth}{
there exists $\delta_{m-1} = \delta_{m-1}(\varepsilon) > 0$ such that
$\dom(T_{m-1}(\theta,x), T_{m-1}(\nu, y)) < \varepsilon$
provided that $\dom((\theta,x), (\nu, y)) < \delta_{m-1}.$
}}\end{equation}
On the other hand, Lemma~\ref{seqTmPropsInBasicintervals}(a)
tells us that
$T_m\evalat{\wbasband{i}}$ is uniformly continuous
for every $i \in \DS.$
So, for every $i \in \DS,$
\begin{equation}\label{TmUCalesbandes}
\text{\parbox{0.9\textwidth}{
there exists $\delta_{m,i} = \delta_{m,i}(\varepsilon) > 0$
such that
$\dom(T_m(\theta,x), T_m(\nu, y)) < \varepsilon$
for every $(\theta,x), (\nu, y) \in \wbasband{i} \subset \wIVD$
verifying $\dom((\theta,x), (\nu, y)) < \delta_{m,i}(\varepsilon).$
}}\end{equation}
Then, by using the numbers
$\delta_{m-1}(\varepsilon/7)$ given by \eqref{Tmm1UC},
$\delta_{m,i}(\varepsilon/7)$ given by \eqref{TmUCalesbandes} and
$\varrho_m(\varepsilon/7)$ given by Lemma~\ref{distTmTm-1smallboxes},
we set
\[
\delta = \delta(\varepsilon) := \min\left\{
\delta_{m-1}(\varepsilon/7),
\min \set{\delta_{m,i}(\varepsilon/7)}{i \in \DS \cap Z_{\varrho_m(\varepsilon/7)}}
\right\}.
\]
Clearly, $\delta > 0$ because the set $\DS \cap Z_{\varrho_m(\varepsilon/7)}$ is finite.
Now we will show that if $\dom((\theta,x), (\nu, y)) < \delta,$
then $\dom(T_m(\theta,x), T_m(\nu, y)) < \varepsilon.$
Assume first that
$(\theta,x),(\nu,y) \in \wbasband{\ell}$
for some $\ell \in \DS \cap Z_{\varrho_m(\varepsilon/7)}.$
We have
\[
\dom((\theta,x), (\nu, y)) < \delta
\le \min \set{\delta_{m,i}(\varepsilon/7)}{i \in \DS \cap Z_{\varrho_m(\varepsilon/7)}}
\le \delta_{m,\ell}(\varepsilon/7).
\]
Hence, by \eqref{TmUCalesbandes},
\[
\dom(T_m(\theta,x), T_m(\nu, y)) < \varepsilon/7 < \varepsilon.
\]
Next we assume that
$(\theta,x),(\nu,y) \in \wbasband{\ell}$
for some $\ell \in \DS$ such that $\all > \varrho_m(\varepsilon/7)$
(in particular, $\theta,\nu \in \wbasint{\ell}$).
In this situation we have
\[
\dom((\theta,x), (\nu, y)) < \delta \le \delta_{m-1}(\varepsilon/7)
\]
and, by \eqref{Tmm1UC} and Lemma~\ref{distTmTm-1smallboxes},
\begin{align*}
\dom(T_m(\theta,x), T_m(\nu, y))
&\le \dom(T_m(\theta,x), T_{m-1}(\theta,x)) +
\dom(T_{m-1}(\theta,x), T_{m-1}(\nu, y))\ +\\
& \hspace*{4em}\dom(T_{m-1}(\nu, y), T_m(\nu, y)) \\
&= \abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} +
\dom(T_{m-1}(\theta,x), T_{m-1}(\nu, y))\ +\\
& \hspace*{4em}\abs{f_{m,\nu}(y) - f_{m-1,\nu}(y)}\\
&\le \norm{f_{m,\theta} - f_{m-1,\theta}} +
\dom(T_{m-1}(\theta,x), T_{m-1}(\nu, y))\ +\\
& \hspace*{4em}\norm{f_{m,\nu} - f_{m-1,\nu}}\\
&< \tfrac{3}{7} \varepsilon < \varepsilon .
\end{align*}
In summary, we have proved that
\[ \dom(T_m(\theta,x), T_m(\nu, y)) < \tfrac{3}{7} \varepsilon \]
when $\dom((\theta,x), (\nu, y)) < \delta$
and $(\theta,x),(\nu,y) \in \wbasband{\ell}$ for some $\ell \in \DS.$
Next we assume that $(\theta,x),(\nu,y) \in \wIVD$ but
$(\theta,x),(\nu,y) \notin \wbasband{\ell}$ for every $\ell \in \DS.$
By Lemma~\ref{Dsets}(a,b), there exist
$i = \bt{\theta},k=\bt{\nu} \in \DS,$ $i \ne k,$ such that
$\theta \in \wbasint{i},$
$(\theta,x) \in \wbasband{i},$
$\nu \in \wbasint{k}$ and
$(\nu,y) \in \wbasband{k}.$
Then, there exist
\[
\widetilde{\theta} \in A \cap \Bd\left(\wbasint{i}\right) \subset \wEIBD
\andq
\widetilde{\nu} \in A \cap \Bd\left(\wbasint{k}\right) \subset \wEIBD,
\]
where $A$ denotes the closed arc of $\SI$ such that
\[ \diam(A) = \dSI(\theta,\nu) \andq \Bd(A) = \{\theta, \nu\}. \]
Clearly we have,
$(\theta,x), \bigl(\widetilde{\theta}, x\bigr) \in \wbasband{i},$
$(\nu, y), \bigl(\widetilde{\nu}, y\bigr) \in \wbasband{k}$ and,
by the previous case,
\begin{align*}
\dom\left((\theta,x), \bigl(\widetilde{\theta}, x\bigr)\right)
&= \dSI\bigl(\theta,\widetilde{\theta}\bigr)
\le \dSI(\theta,\nu) \le \dom((\theta,x), (\nu, y)) < \delta,\\
& \dom\left(T_m(\theta,x), T_m\bigl(\widetilde{\theta}, x\bigr)\right) < \tfrac{3}{7} \varepsilon\\
\dom\left((\nu,y), \bigl(\widetilde{\nu}, y\bigr)\right)
&= \dSI\bigl(\nu,\widetilde{\nu}\bigr)
\le \dSI(\theta,\nu) \le \dom((\theta,x), (\nu, y)) < \delta,\text{ and}\\
& \dom\left(T_m(\nu, y), T_m\bigl(\widetilde{\nu}, y\bigr)\right) < \tfrac{3}{7} \varepsilon.
\end{align*}
On the other hand,
$
\bigl(\widetilde{\theta}, x\bigr),
\bigl(\widetilde{\nu}, y\bigr) \in \setsilift{\wEIBD}
\subset \wIVD \subset \wIVD[m-1]
$
and, by Lemma~\ref{voresiguals} and \eqref{Tmm1UC},
\begin{align*}
\dom\left(\bigl(\widetilde{\theta},x\bigr), \bigl(\widetilde{\nu}, y\bigr)\right)
&= \max\left\{\dSI\bigl(\widetilde{\theta},\widetilde{\nu}\bigr), \abs{x-y}\right\}
\le \max\left\{\dSI(\theta,\nu), \abs{x-y}\right\}\\
&= \dom((\theta,x), (\nu, y)) < \delta \le \delta_{m,i}(\varepsilon/7),\text{ and}\\
\dom(T_m(\theta,x), T_m(\nu, y))
&\le \dom\left(T_m(\theta,x), T_m\bigl(\widetilde{\theta}, x\bigr)\right) +
\dom\left(T_m\bigl(\widetilde{\theta}, x\bigr), T_m\bigl(\widetilde{\nu}, y\bigr)\right)\ +\\
& \hspace*{4em}\dom\left(T_m\bigl(\widetilde{\nu}, y\bigr), T_m(\nu, y)\right)\\
&< \tfrac{3}{7} \varepsilon +
\dom\left(T_{m-1}\bigl(\widetilde{\theta}, x\bigr), T_{m-1}\bigl(\widetilde{\nu}, y\bigr)\right) +
\tfrac{3}{7} \varepsilon = \varepsilon.
\end{align*}
If $(\theta,x),(\nu,y) \notin \wIVD$ then,
by Definition~\ref{seqTmDefi} and \eqref{Tmm1UC} ,
\[
\dom(T_m(\theta,x), T_m(\nu, y)) = \dom(T_{m-1}(\theta,x), T_{m-1}(\nu, y))
< \varepsilon/7 < \varepsilon
\]
because
$
\dom((\theta,x), (\nu, y)) < \delta \le \delta_{m-1}(\varepsilon/7).
$
Lastly, assume that $(\nu,y) \notin \wIVD$ but
$(\theta,x) \in \wbasband{i} \subset \wIVD,$
for some $i\in\DS$ (that is, $\theta \in \wbasint{i}$).
In this situation, as before, there exists
$
\widetilde{\theta} \in \Bd\left(\wbasint{i}\right) \subset \wEIBD
$
such that, by Lemma~\ref{voresiguals} and Definition~\ref{seqTmDefi}
($\bigl(\widetilde{\theta}, x\bigr) \in \setsilift{\wEIBD} \subset \wIVD \subset \wIVD[m-1]$), and \eqref{Tmm1UC},
\begin{align*}
\dom\left((\theta,x), \bigl(\widetilde{\theta}, x\bigr)\right) & < \delta,\\
\dom\left(\bigl(\widetilde{\theta},x\bigr), (\nu, y\bigr)\right) & < \delta \le \delta_{m-1}(\varepsilon/7),\\
\dom\left(T_m(\theta,x), T_m\bigl(\widetilde{\theta}, x\bigr)\right) & < \tfrac{3}{7} \varepsilon,\text{ and}\\
\dom(T_m(\theta,x), T_m(\nu, y))
&\le \dom\left(T_m(\theta,x), T_m\bigl(\widetilde{\theta}, x\bigr)\right) +
\dom\left(T_m\bigl(\widetilde{\theta}, x\bigr), T_m(\nu, y)\right)\\
&< \tfrac{3}{7} \varepsilon +
\dom\left(T_{m-1}\bigl(\widetilde{\theta}, x\bigr), T_{m-1}(\nu, y)\right) < \varepsilon.
\end{align*}
This ends the proof of the continuity of $T_m$
and, hence, of (a).
\inidemopart{b}
When $\theta \in \wIBD$ the statement follows from Lemma~\ref{seqTmPropsInBasicintervals}(b).
When $\theta \in \SI\setminus\wIBD,$
it follows from the part already proven and
the continuity of $T_m.$
\inidemopart{c}
The first two statements follow from
Lemma~\ref{seqTmPropsInBasicintervals}(c) and
statement (a). On the other hand,
as in the proof of Proposition~\ref{T0mapProperties}(c) for $m=0$,
Lemma~\ref{denso}(b) implies that
$\istar \in \wIBD$ but $\istar \notin \wIBD[k]$
for every $k > m.$
Then, we get $f_{k,\istar} = f_{m,\istar}$
from Definition~\ref{seqTmDefi}.
\end{proof}
\section{Proof of Proposition~\ref{distTmTm-1}}\label{proofofdistTmTm-1}
This section is devoted to prove
Proposition~\ref{distTmTm-1}.
It is the third technical counterpart of Section~\ref{skew-product}.
In contrast to Lemma~\ref{distTmTm-1smallboxes}
the bound given by Proposition~\ref{distTmTm-1}.
is valid for every $\theta \in \wIBD$.
Before starting the proof of this proposition we will state and
prove a number of very simple lemmas that will help in
automating the proof of Proposition~\ref{distTmTm-1}.
\begin{lemma}\label{pointnormbound}
Assume that $\wbasint{i} \subset \wbasint{k}$
for some $i \in \DS,\ k \in \DS[m-1]$ and $m \ge 2,$
and assume that either
\[
i < 0 \text{ and } \theta \in \wbasint{i} \setminus \istarset
\text{ or }
i \ge 0 \text{ and } \theta \in \basint{i} \setminus \OBG{i}{i+1}.
\]
Then,
\[
\abs{\gams{\aii}\left(R_\omega(\theta)\right) - \gams{\akk}\left(R_\omega(\theta)\right)}
\le 2^{-\ak}.
\]
\end{lemma}
\begin{proof}
The lemma holds trivially when $\akk = \aii.$
Thus, we may assume that $\akk \ne \aii.$
Then by Lemma~\ref{QuePassaALesAles},
$\ak < \ai,$ $\akk < \aii$ and
\[
\gams{\akk}\left(R_\omega(\theta)\right) =
\gams{\akk+1}\left(R_\omega(\theta)\right) = \dots =
\gams{\aii-1}\left(R_\omega(\theta)\right).
\]
By assumption we have
\[
\theta \in \begin{cases}
\basint{i} \setminus \OBG{i}{i+1} &\text{when $i \ge 0$, and}\\
\wbasint{i} \setminus \istarset =
\BSG{i}{\aii}\setminus \istarset
&\text{when $i < 0,$}
\end{cases}
\]
and, hence,
\[
R_\omega(\theta) \in \begin{cases}
\BSG{i+1}{i} \setminus \obasint{i+1} &\text{when $i \ge 0$, and}\\
\basintabs{i+1} \setminus \sstarset{i+1} &\text{when $i < 0$}.
\end{cases}
\]
Thus, in the case $i \ge 0$ we have
\[ R_\omega(\theta) \notin \obasint{i+1} \cup \wbasint{-(i+1)} \cup \Zstar_{i+1} \]
by Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2).
So, by Definition~\ref{PCgenerators},
\[
\gams{i+1}\left(R_\omega(\theta)\right)
= \gams{i}\left(R_\omega(\theta)\right)
= \gams{\akk}\left(R_\omega(\theta)\right).
\]
This ends the proof of the lemma in this case.
Assume now that $i < 0.$
By Lemma~\ref{Propertiesvarphi}(c,d,f) and
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
\begin{align*}
\abs{\gams{\aii}\left(R_\omega(\theta)\right) - \gams{\akk}\left(R_\omega(\theta)\right)}
&= \abs{\gams{\aii}\left(R_\omega(\theta)\right) - \gams{\aii-1}\left(R_\omega(\theta)\right)}\\
&\le \diam\left(\basicbox{i+1}\right) \le 2^{-\aii} \le 2^{-\ak}
\end{align*}
(observe that $\aii > \akk \ge \ak - 1$).
\end{proof}
\begin{lemma}\label{intervalsnormbound}
Let $s, t \in \Z,$ $s \ne t$ be such that
$\theta \in \wobasint{s} \setminus \obasintabs{s},$
and either $t < 0$ and $\theta \in \obasintabs{t}$
or $t \ge 0$ and $\theta \in \OBG{t}{t+1}.$
Then, the following statements hold:
\begin{enumerate}[(a)]
\item $R_\omega(\theta) \in \obasintabs{s+1} \cap \obasintabs{t+1}.$
\item Let $u,v \in \Z$ be such that $\{u,v\} = \{ s, t\}$
and $\abs{u + 1} \le \abs{v + 1}$. \\
Then, $\I_{v+1, R_\omega(\theta)} \subset \I_{u+1, R_\omega(\theta)}.$
\item \[
\abs{x - y} \le 2\cdot 2^{-\abs{u}}
\]
for every $x \in \I_{t+1, R_\omega(\theta)}$ and $y \in \I_{s+1, R_\omega(\theta)}.$
\end{enumerate}
\end{lemma}
\begin{proof}
By assumption we have
\[
\theta \in \begin{cases}
\OBG{t}{t+1} &\text{when $t \ge 0$, and}\\
\obasintabs{t} \subset \wobasint{t} = \OBG{t}{\abs{t+1}}
&\text{when $t < 0.$}
\end{cases}
\]
Hence,
$
R_\omega(\theta) \in \obasintabs{t+1}.
$
Moreover, as in the proof of Lemma~\ref{pointnormbound},
$s < 0$ and
$
R_\omega(\theta) \in \obasintabs{s+1}.
$
This proves (a).
Now we prove (b). From (a) we have
\begin{align*}
R_\omega(\theta) &\in \obasintabs{u+1} \cap \obasintabs{v+1}\\
&\subset \obasintabs{u+1} \cap \wbasint{v+1}.
\end{align*}
Moreover,
$s \ne t$ implies $u+1 \ne v+1$ and we have $\abs{u+1} \le \abs{v+1}$
by assumption. Consequently, by Lemma~\ref{Propertiesvarphi}(g,d) and
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
$\abs{u+1} < \abs{v+1}$ and
\[
\basicbox{v+1} \subset \Int\left(\basicbox{u+1} \setminus \setsilift{\sstarset{u+1}}\right)
\]
which implies (b).
Thus, $x,y \in \I_{u+1, R_\omega(\theta)}$ and, by Lemma~\ref{Propertiesvarphi}(f),
\[
\abs{x - y} \le \diam\left(\basicbox{u+1}\right)
\le 2^{-\abs{u+1}} \le 2^{-(\abs{u}-1)} = 2\cdot 2^{-\abs{u}}.
\]
\end{proof}
Now we are ready to start the proof of Proposition~\ref{distTmTm-1}.
\begin{proof}[Proof of Proposition~\ref{distTmTm-1}]
We start by showing that
$\{T_m\}_{k = 0}^{\infty}$ is a Cauchy sequence,
assuming that the bound \eqref{fitanorma}
holds for every $m \ge 2$ and $\theta \in \SI.$
We start by estimating $\dinf(T_m, T_{m+1})$ for every $m \in \N.$
From \eqref{fitanorma} and the definition of $\mu_m$
\[
\dinf(T_m, T_{m+1})
= \sup_{\theta \in \SI} \norm{f_{m,\theta} - f_{m+1,\theta}}
\le 2\cdot \sup_{\theta \in \SI} 2^{-\abs{\bt{\theta}}}
\le 2\cdot 2^{-\mu_m}.
\]
By Lemma~\ref{denso}(a) $\{\mu_m\}_{m=0}^\infty$
is strictly increasing (and $\lim_{m\to\infty} \mu_m = \infty$).
Therefore, for every $\varepsilon > 0,$
there exists $N \ge 2,$ such that
$4\cdot 2^{-\mu_m} < \varepsilon$ for every $m \ge N.$
Hence,
\begin{align*}
\dinf(T_m, T_{m+i})
&\le \sum_{\ell=m}^{m+i-1} \dinf(T_{\ell}, T_{\ell + 1})
\le 2\cdot \sum_{\ell=m}^{m+i-1} 2^{-\mu_{\ell}}\\
&\le 2\cdot 2^{-\mu_m} \sum_{\ell=0}^{\infty} 2^{-\ell}
= 4\cdot 2^{-\mu_m} \le 4\cdot 2^{-\mu_N} < \varepsilon
\end{align*}
for every $m \ge N$ and $i \in \N.$
So, $\{T_m\}_{k = 0}^{\infty}$ is a Cauchy sequence.
Now we prove \eqref{fitanorma}. That is,
\[
\norm{f_{m,\theta} - f_{m-1,\theta}} \le 2 \cdot 2^{-\abs{\bt[m-1]{\theta}}}
\]
for every $m \ge 2$ and $\theta \in \SI.$
From Definition~\ref{seqTmDefi} and Lemma~\ref{voresiguals}
we know that $f_{m,\theta} = f_{m-1,\theta}$ for every
$\theta \in \left(\SI \setminus \wIBD\right) \cup \wEIBD.$
Then, \eqref{fitanorma} holds in this case.
In the rest of the involved proof we assume that
$\theta \in \wIBD \setminus \wEIBD.$
Thus, by Lemmas~\ref{Dsets}(a,b), \ref{Propertiesvarphi}(g) and \ref{QuePassaALesAles},
\begin{quotation}\itshape
$
\theta \in \wobasint{i} \subset
\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \kstarset \right)
$
where\newline
$i = \bt{\theta} \in \DS,$ $k = \bt[m-1]{\theta} \in \DS[m-1],$\newline
$\ak < \ai,$ and $\akk \le \aii.$
\end{quotation}
Moreover, $\wbasband{i} \subset \wbasband{k} \subset \wIVD[m-1].$
Consequently, by Lemma~\ref{seqTmPropsInBasicintervals}(a,b),
the maps $f_{m, \theta}$ and $f_{m-1, \theta}$ are well defined, continuous,
piecewise affine and non-increasing, and
$f_{m,\theta}(2) = f_{m-1,\theta}(2) = -2$ and
$f_{m,\theta}(-2) = f_{m-1,\theta}(-2) = 2$
(see Figures~\ref{fig:Cas1.3}, \ref{fig:Cas2} and \ref{fig:Cas3.1}
for some examples in generic cases).
We split the proof into three cases according to whether $\theta$
belongs to
\[
\wobasint{i} \setminus \obasintabs{i},\
\obasintabs{i} \subset \wobasint{k} \setminus \basintabs{k}
\text{ or }
\obasintabs{i} \subset \obasintabs{k}.
\]
\begin{autocase}{1}
$\theta \in \wobasint{i} \setminus \obasintabs{i}.$
\end{autocase}
We have $i < 0$ because $\wobasint{i} = \obasint{i}$ for $i \ge 0.$
Moreover, by Definition~\ref{curvesinthewings}, $\theta \in \WB.$
To deal with this case we consider three subcases.
\begin{autocase}[Subcase]{1.1}
$\theta \in \left(\wobasint{i} \setminus \obasintabs{i}\right) \setminus \WIB.$
\end{autocase}
By Lemmas~\ref{QuePassaALesAles}, \ref{fmfm-1alesales},
\ref{normainterna} and \ref{pointnormbound},
\begin{align*}
\I_{i, \theta} & = \{m_i(\theta)\} = \{\gams{\ai}(\theta)\} = \{\gams{\ak}(\theta)\} \subset \I_{k, \theta},\\
f_{m,\theta}\left(m_i(\theta)\right) & = \gams{\aii}\left(R_\omega(\theta)\right),\\
f_{m-1,\theta}\left(m_i(\theta)\right) & = \gams{\akk}\left(R_\omega(\theta)\right), \text{ and}\\
\norm{f_{m,\theta} - f_{m-1,\theta}} &=
\norm{f_{m,\theta}\evalat{\I_{i, \theta}} - f_{m-1,\theta}\evalat{\I_{i, \theta}}} =
\abs{f_{m,\theta}\left(m_{i}(\theta)\right) - f_{m-1,\theta}\left(m_{i}(\theta)\right)}\\
&= \abs{\gams{\aii}\left(R_\omega(\theta)\right) - \gams{\akk}\left(R_\omega(\theta)\right)}
\le 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
\begin{autocase}[Subcase]{1.2}
$\theta \in \left(\wobasint{i} \setminus \obasintabs{i}\right) \cap \WIB$
and $\wobasint{i} \subset \wobasint{k} \setminus \basintabs{k}.$
\end{autocase}
In this subcase, by Definition~\ref{curvesinthewings} we have
\[
\theta \in \wobasint{k} \setminus \basintabs{k} \subset \WB[m-1]
\]
(recall that $i < 0$).
Then, by Lemmas~\ref{QuePassaALesAles} and \ref{VerticalIntervalsIntheWings}(b,c),
Definition~\ref{seqTmDefi} and Lemmas~\ref{normainterna} and \ref{pointnormbound},
\begin{align*}
\I_{i, \theta} & = \{\gams{\ai}(\theta)\} = \{\gams{\ak}(\theta)\} \subset \IW{\theta} = \IW[m-1]{\theta},\\
f_{m,\theta}(x) & = \gams{\aii}\left(R_\omega(\theta)\right) \text{ for every $x \in \IW{\theta}$,}\\
f_{m-1,\theta}(x) & = \gams{\akk}\left(R_\omega(\theta)\right) \text{ for every $x \in \IW[m-1]{\theta}$, and}\\
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \norm{f_{m,\theta}\evalat{\IW{\theta}} - f_{m-1,\theta}\evalat{\IW{\theta}}} \\
&= \abs{\gams{\aii}\left(R_\omega(\theta)\right) - \gams{\akk}\left(R_\omega(\theta)\right)}
\le 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
Observe that since $\wobasint{i}$ is connected and
\[
\wobasint{i} \subset
\wobasint{k} \setminus \left( \Bd\left(\basintabs{k}\right) \cup \kstarset \right),
\]
$\wobasint{i} \not\subset \wobasint{k} \setminus \basintabs{k}$
implies $\wobasint{i} \subset \obasintabs{k} \setminus \kstarset.$
\begin{autocase}[Subcase]{1.3}
$\theta \in \left(\wobasint{i} \setminus \obasintabs{i}\right) \cap \WIB$
and $\wobasint{i} \subset \obasintabs{k} \setminus \kstarset$
\upshape (see Figure~\ref{fig:Cas1.3} for a symbolic representation of this case).
\end{autocase}
\begin{figure}
\begin{tikzpicture}[scale=2]
\draw (-2,-2) rectangle (2,2);
\foreach \c in {-2, 2} { \node[below] at (\c,-2) {$\c$}; \node[left] at (-2,\c) {$\c$}; }
\draw[dashed, color=blue] (-1,-2.3) -- (-1,2); \draw[dashed, color=blue] (1,-2.3) -- (1,2);
\draw[decorate, very thick, decoration={brace,amplitude=5pt, mirror, raise=2pt}, color=blue] (-1,-2.3) -- (1,-2.3);
\node[below, color=blue] at (0,-2.4) {\scriptsize$\I_{k,\theta}$};
\node[below, color=blue] at (-1,-2.4) {\scriptsize$m_k(\theta)$};
\node[below, color=blue] at (1,-2.4) {\scriptsize$M_k(\theta)$};
\draw[dashed, color=blue] (-2, 0.3) -- (2, 0.3); \draw[dashed, color=blue] (-2, 1) -- (2, 1);
\draw[decorate, very thick, decoration={brace,amplitude=5pt,raise=2pt}, color=blue] (-2, 0.3) -- (-2, 1);
\node[left, color=blue] at (-2.1,0.65) {\scriptsize$\I_{k+1,R_\omega(\theta)}$};
\node[left, color=blue] at (-2.05, 0.3) {\scriptsize$m_{k+1}\bigl(R_\omega(\theta)\bigr)$};
\node[left, color=blue] at (-2.05, 1) {\scriptsize$M_{k+1}\bigl(R_\omega(\theta)\bigr)$};
\draw[very thick, color=blue] (-2,2) -- (-1.6, 1.85) -- (-1.1, 1.4) --(-1, 1)
-- (1, 0.3) -- (1.2, -1) -- (1.4, -1.1) -- (1.7, -1.8) -- (2,-2);
\node[color=blue, above right] at (1.15,-1) {$f_{m-1,\theta}$};
\node[color=blue] at (-0.5,0.95) {$g_{_{k, \theta}}$};
\draw[dashed, color=red] (-0.8,-2) -- (-0.8,2); \draw[dashed, color=red] (-0.1,-2) -- (-0.1,2);
\draw[decorate, very thick, decoration={brace,amplitude=5pt, mirror, raise=2pt}, color=red] (-0.8,-2) -- (-0.1,-2);
\node[below, color=red] at (-0.45,-2.1) {\scriptsize $\IW{\theta}$};
\node[below left, color=red] at (-0.75,-2.01) {\scriptsize$\lambda_m(\theta)$};
\node[below right, color=red] at (-0.15,-2.01) {\scriptsize$\tau_m(\theta)$};
\draw[dashed, color=red] (-2, 0.4) -- (2, 0.4);
\node[right, color=red] at (2, 0.4) {\scriptsize$\gams{\aii}\left(R_\omega(\theta)\right)$};
\draw[very thick, color=red] (-2,2) -- (-1.6, 1.8) -- (-1.1, 1.2266) --(-1, 0.711)
-- (-0.8, 0.4) -- (-0.1, 0.4) -- (1, -0.04848)
-- (1.2, -1.1515) -- (1.4, -1.2363) -- (1.7, -1.8303) -- (2,-2);
\node[color=red, left] at (-1,0.711) {$f_{m,\theta}$};
\end{tikzpicture}
\caption{A symbolic representation of the maps
$f_{m,\theta}$ and $f_{m-1,\theta}$ in
Subcase~1.3 of Proposition~\ref{distTmTm-1}
($\theta \in \left(\protect\wobasint{i} \setminus \protect\obasintabs{i} \right) \cap \WIB$
and $\protect\wobasint{i} \subset \protect\obasintabs{k} \setminus \protect\kstarset$).
The map \textcolor{blue}{$f_{m-1,\theta}$} and the corresponding intervals
\textcolor{blue}{$\I_{k,\theta}$} and \textcolor{blue}{$\I_{k+1,R_\omega(\theta)}$}
are drawn in \textcolor{blue}{blue}.
The map \textcolor{red}{$f_{m,\theta}$}, the interval
\textcolor{red}{$\IW{\theta}$} and the point \textcolor{red}{$\gams{\aii}\left(R_\omega(\theta)\right)$}
are drawn in \textcolor{red}{red}.}\label{fig:Cas1.3}
\end{figure}
By Lemmas~\ref{QuePassaALesAles} and \ref{VerticalIntervalsIntheWings}(b)
and Definition~\ref{seqTmDefi},
\begin{align*}
\I_{i, \theta} & = \{\gams{\ai}(\theta)\} = \{\gams{\ak}(\theta)\} \subset \IW{\theta},\text{ and}\\
f_{m,\theta}(x) & = \gams{\aii}\left(R_\omega(\theta)\right)\text{ for every $x \in \IW{\theta}$.}
\end{align*}
On the other hand,
by Definition~\ref{curvesinthewings} and Lemma~\ref{DepthintheWings}(a,b),
$\theta \in \WIB \subset \WDB,$ and
\[
\theta \in \basintabs{\ell} \subset
\wobasint{i} \setminus \basintabs{i} \subset
\obasintabs{k} \setminus \kstarset
\]
with $\ell = \bt[\led{\theta}]{\theta} \in \WDS$ and $\all > \ai > \ak.$
Then, by Lemma~\ref{Propertiesvarphi}(g) and Definition~\ref{curvesinthewings},
$\basicbox{\ell} \subset \Int\left(\basicbox{k} \setminus \setsilift{\kstar}\right)$
and
\[
\IW{\theta} = \I_{\ell, \theta} \subset \I_{k, \theta}.
\]
Moreover, since $\theta \in \obasintabs{k} \subset \IBD,$
Definition~\ref{seqTmDefi},
Lemmas~\ref{gpositiva}(b) and \ref{gnegativa}(b),
and the definition of the maps $g_{_{i, \theta}}$ for $i \ge 0$
(Definition~\ref{defi-gi-positiva}) give
\begin{align*}
f_{m-1,\theta}\left(\IW{\theta}\right)
&\subset f_{m-1,\theta}\left(\I_{k, \theta}\right)\\
&\subset \begin{cases}
\I_{k+1, R_\omega(\theta)} & \text{if $k < 0$ or $k \ge 0$ and $\theta \in \OBG{k}{k+1},$}\\
\{\gams{k+1}\left(R_\omega(\theta)\right)\} & \text{if $k \ge 0$ and $\theta \in \basint{k} \setminus \OBG{k}{k+1}$.}
\end{cases}
\end{align*}
Now, as before, we will use Lemma~\ref{normainterna} to bound
$\norm{f_{m,\theta} - f_{m-1,\theta}}.$
We start with the simplest case:
$k \ge 0$ and $\theta \in \basint{k} \setminus \OBG{k}{k+1}.$
By Lemma~\ref{pointnormbound},
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \norm{f_{m,\theta}\evalat{\IW{\theta}} - f_{m-1,\theta}\evalat{\IW{\theta}}} \\
&= \abs{\gams{\aii}\left(R_\omega(\theta)\right) - \gams{\akk}\left(R_\omega(\theta)\right)}
\le 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
Now we assume that $k < 0$ or $k \ge 0$ and $\theta \in \OBG{k}{k+1}.$
In this case Lemma~\ref{intervalsnormbound} applies.
By Lemmas~\ref{intervalsnormbound}, \ref{Propertiesvarphi}(d) and
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
and Lemma~\ref{normainterna} we have
\begin{align*}
\gams{\aii}\left(R_\omega(\theta)\right)
& \in \I_{i+1, R_\omega(\theta)} \subset \I_{k+1, R_\omega(\theta)},\\
f_{m-1,\theta}(x) &\in \I_{k+1, R_\omega(\theta)}\quad\text{for every $x \in \IW{\theta}$.}
\end{align*}
and
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \LSleftlimits{\sup}{x \in \IW{\theta}} \abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} \\
&= \LSleftlimits{\sup}{x \in \IW{\theta}} \abs{\gams{\aii}\left(R_\omega(\theta)\right) - f_{m-1,\theta}(x)}\\
&\le 2\cdot 2^{-\ak} = 2\cdot 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
This ends the proof of the proposition in this case.
\begin{autocase}{2}
$\theta \in \obasintabs{i} \subset \wobasint{k} \setminus \basintabs{k}$
\upshape (see Figure~\ref{fig:Cas2} for a symbolic representation of this case).
\end{autocase}
\begin{figure}
\begin{tikzpicture}[scale=2]
\draw (-2,-2) rectangle (2,2);
\foreach \c in {-2, 2} { \node[below] at (\c,-2) {$\c$}; \node[left] at (-2,\c) {$\c$}; }
\draw[dashed, color=red] (-1,-2.3) -- (-1,2); \draw[dashed, color=red] (1,-2.3) -- (1,2);
\draw[decorate, very thick, decoration={brace,amplitude=5pt, mirror, raise=2pt}, color=red] (-1,-2.3) -- (1,-2.3);
\node[below] at (0,-2.4) {\scriptsize\textcolor{red}{$\I_{i,\theta}$} = \textcolor{blue}{$\IW[m-1]{\theta}$}};
\node[below left, color=red] at (-0.6,-2.4) {\scriptsize\textcolor{red}{$m_i(\theta)$} = \textcolor{blue}{$\lambda_{m-1}(\theta)$}};
\node[below right, color=red] at (0.6,-2.4) {\scriptsize\textcolor{red}{$M_i(\theta)$} = \textcolor{blue}{$\tau_{m-1}(\theta)$}};
\draw[dashed, color=red] (-2, 1) -- (2, 1); \draw[dashed, color=red] (-2, 0.3) -- (2, 0.3);
\draw[decorate, very thick, decoration={brace,amplitude=2pt,raise=2pt}, color=red] (-2, 0.3) -- (-2, 1);
\node[left, color=red] at (-2.1,0.65) {\scriptsize$\I_{i+1,R_\omega(\theta)}$};
\node[left, color=red] at (-2.05, 0.3) {\scriptsize$m_{i+1}\bigl(R_\omega(\theta)\bigr)$};
\node[left, color=red] at (-2.05, 1) {\scriptsize$M_{i+1}\bigl(R_\omega(\theta)\bigr)$};
\draw[very thick, color=red] (-2,2) -- (-1.6, 1.85) -- (-1.1, 1.4) --(-1, 1)
-- (1, 0.3) -- (1.2, -1) -- (1.4, -1.1) -- (1.7, -1.8) -- (2,-2);
\node[color=red, above right] at (1.15,-1) {$f_{m,\theta}$};
\node[color=red] at (-0.5,0.95) {$g_{_{i, \theta}}$};
\draw[dashed, color=blue] (-2, 0.2) -- (2.2, 0.2);
\node[right, color=blue] at (2.1, 0.2) {\scriptsize$\gams{\akk}\left(R_\omega(\theta)\right)$};
\draw[dashed, color=blue] (-2, 0.05) -- (2, 0.05); \draw[dashed, color=blue] (-2, 1.15) -- (2, 1.15);
\draw[decorate, very thick, decoration={brace, mirror, amplitude=5pt,raise=2pt}, color=blue] (2, 0.05) -- (2, 1.15);
\node[right, color=blue] at (2.1,0.6) {\scriptsize$\I_{k+1,R_\omega(\theta)}$};
\draw[very thick, color=blue] (-2,2) -- (-1.6, 1.6) -- (-1.1, 0.9) --(-1, 0.2)
-- (1, 0.2) -- (1.2, -1.05) -- (1.4, -1.15) -- (1.7, -1.8303) -- (2,-2);
\node[color=blue, left] at (-1.05,0.711) {$f_{m-1,\theta}$};
\end{tikzpicture}
\caption{A symbolic representation of the maps
$f_{m,\theta}$ and $f_{m-1,\theta}$ in Case~2
($\theta \in \protect\obasintabs{i} \subset \protect\wobasint{k} \setminus \protect\basintabs{k}$)
of Proposition~\ref{distTmTm-1}.
The map \textcolor{blue}{$f_{m-1,\theta}$} and the corresponding intervals
\textcolor{blue}{$\IW[m-1]{\theta}$} and \textcolor{blue}{$\I_{k+1,R_\omega(\theta)}$}
are drawn in \textcolor{blue}{blue}.
The map \textcolor{red}{$f_{m,\theta}$} and the corresponding intervals
\textcolor{red}{$\I_{i,\theta}$} = \textcolor{blue}{$\IW[m-1]{\theta}$} and \textcolor{red}{$\I_{i+1,R_\omega(\theta)}$}
are drawn in \textcolor{red}{red}.}\label{fig:Cas2}
\end{figure}
In this case we will use Lemma~\ref{normainterna} with $\I_{i, \theta}.$
Thus, we need to compare the maps $f_{m,\theta}\evalat{\I_{i, \theta}}$
and $f_{m-1,\theta}\evalat{\I_{i, \theta}}.$
Directly from the definitions we get $k < 0,$
$\basintabs{i} \subset \IBD$ and
$\basintabs{k} \subset \IBD[m-1].$
Consequently, by Lemma~\ref{Dsets}(b) and Definition~\ref{curvesinthewings},
\[
\theta \in \IBD
\andq
\theta \in \wIBD[m-1] \setminus \IBD[m-1] \subset \WDB[m-1] \subset \WB[m-1].
\]
Moreover, $\led[m-1]{\theta} = m,$
$i = \bt{\theta} = \bt[\led[m-1]{\theta}]{\theta} \in \WDS[m-1]$
and, by Definition~\ref{curvesinthewings},
$\theta \in \WIB[m-1],$ and
\[
\IW[m-1]{\theta} = \I_{i, \theta}.
\]
Furthermore, since $k < 0,$ as in the proof of Lemma~\ref{pointnormbound},
$
R_\omega(\theta) \in \obasintabs{k+1}.
$
Thus, Definition~\ref{seqTmDefi},
Lemma~\ref{Propertiesvarphi}(d) and
Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
give
\[
f_{m-1,\theta}(x) = \gams{\akk}\left(R_\omega(\theta)\right) \in \I_{k+1, R_\omega(\theta)}
\]
for every $x \in \I_{i, \theta} = \IW[m-1]{\theta}.$
Now we will use Lemma~\ref{normainterna}
to bound the norm $\norm{f_{m,\theta} - f_{m-1,\theta}}.$
By Definition~\ref{curvesinthewings} and Lemma~\ref{normainterna},
$\theta \in \IBD \subset \wIBD \setminus \WIB,$ and
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)} \\
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - \gams{\akk}\left(R_\omega(\theta)\right)}.
\end{align*}
Next we will compute $f_{m,\theta}(\I_{i, \theta}).$
We start with the simplest case:
$i \ge 0$ and $\theta \in \obasint{i} \setminus \OBG{i}{i+1}.$
By Definition~\ref{seqTmDefi},
the definition of the maps $g_{_{i, \theta}}$ for $i \ge 0$
(Definition~\ref{defi-gi-positiva}) and Lemma~\ref{pointnormbound},
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - \gams{\akk}\left(R_\omega(\theta)\right)}\\
&= \abs{\gams{i+1}\left(R_\omega(\theta)\right) - \gams{\akk}\left(R_\omega(\theta)\right)}
\le 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
Assume that $i < 0$ or $i \ge 0$ and $\theta \in \OBG{i}{i+1}.$
Then, again by Definition~\ref{seqTmDefi} and
Lemmas~\ref{gpositiva}(b), \ref{gnegativa}(b) and \ref{intervalsnormbound},
\[
f_{m,\theta}(x) \in \I_{i+1, R_\omega(\theta)} \subset \I_{k+1, R_\omega(\theta)}
\andq[for every]
x \in \I_{i, \theta},
\]
and
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - \gams{\akk}\left(R_\omega(\theta)\right)}\\
&\le 2\cdot 2^{-\ak} = 2\cdot 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
This ends the proof of the proposition in Case~2.
\begin{autocase}{3}
$\theta \in \obasintabs{i} \subset \obasintabs{k}.$
\end{autocase}
\begin{figure}
\begin{tikzpicture}[scale=2]
\draw (-2,-2) rectangle (2,2);
\foreach \c in {-2, 2} { \node[below] at (\c,-2) {$\c$}; \node[left] at (-2,\c) {$\c$}; }
\draw[dashed, color=blue] (-1,-2.3) -- (-1,2); \draw[dashed, color=blue] (1,-2.3) -- (1,2);
\draw[decorate, very thick, decoration={brace,amplitude=5pt, mirror, raise=2pt}, color=blue] (-1,-2.3) -- (1,-2.3);
\node[below, color=blue] at (0,-2.4) {\scriptsize$\I_{k,\theta}$};
\node[below, color=blue] at (-1,-2.4) {\scriptsize$m_k(\theta)$};
\node[below, color=blue] at (1,-2.4) {\scriptsize$M_k(\theta)$};
\draw[dashed, color=blue] (-2, 0) -- (2, 0); \draw[dashed, color=blue] (-2, 1.3) -- (2, 1.3);
\draw[decorate, very thick, decoration={brace,amplitude=5pt,raise=2pt}, color=blue] (-2, 0) -- (-2, 1.3);
\node[left, color=blue] at (-2.1,0.65) {\scriptsize$\I_{k+1,R_\omega(\theta)}$};
\node[left, color=blue] at (-2.05, 0) {\scriptsize$m_{k+1}\bigl(R_\omega(\theta)\bigr)$};
\node[left, color=blue] at (-2.05, 1.3) {\scriptsize$M_{k+1}\bigl(R_\omega(\theta)\bigr)$};
\draw[very thick, color=blue] (-2,2) -- (-1.6, 1.85) -- (-1.1, 1.4) --(-1, 1)
-- (1, 0.3) -- (1.2, -1) -- (1.4, -1.1) -- (1.7, -1.8) -- (2,-2);
\node[color=blue, above right] at (1.15,-1) {$f_{m-1,\theta}$};
\node[color=blue] at (-0.5,0.95) {$g_{_{k, \theta}}$};
\draw[dashed, color=red] (-0.2,-2) -- (-0.2,2); \draw[dashed, color=red] (0.5,-2) -- (0.5,2);
\draw[decorate, very thick, decoration={brace,amplitude=5pt, mirror, raise=2pt}, color=red] (-0.2,-2) -- (0.5,-2);
\node[below, color=red] at (0.15,-2.1) {\scriptsize$\I_{i,\theta}$};
\node[below left, color=red] at (-0.1,-2.01) {\scriptsize$m_i(\theta)$};
\node[below right, color=red] at (0.4,-2.01) {\scriptsize$M_i(\theta)$};
\draw[dashed, color=red] (-2, 0.1) -- (2, 0.1); \draw[dashed, color=red] (-2, 0.35) -- (2, 0.35);
\draw[decorate, very thick, decoration={brace, mirror, amplitude=2pt,raise=2pt}, color=red] (2, 0.1) -- (2, 0.35);
\node[right, color=red] at (2.05,0.2) {\scriptsize$\I_{i+1,R_\omega(\theta)}$};
\node[below right, color=red] at (2, 0.15) {\scriptsize$m_{i+1}\bigl(R_\omega(\theta)\bigr)$};
\node[above right, color=red] at (2, 0.25) {\scriptsize$M_{i+1}\bigl(R_\omega(\theta)\bigr)$};
\draw[very thick, color=red] (-2,2) -- (-1.6, 1.8) -- (-1.1, 1.2266) --(-1, 0.711)
-- (-0.2, 0.35) -- (0.5, 0.1) -- (1, -0.04848)
-- (1.2, -1.1515) -- (1.4, -1.2363) -- (1.7, -1.8303) -- (2,-2);
\node[color=red, left] at (-1,0.711) {$f_{m,\theta}$};
\node[color=red, above right] at (-0.25,0.05) {$g_{_{i, \theta}}$};
\end{tikzpicture}
\caption{A symbolic representation of the maps
$f_{m,\theta}$ and $f_{m-1,\theta}$ in
Subcase~3.1 from the proof of Proposition~\ref{distTmTm-1}
($\theta \in \protect\obasintabs{i}$ and $\I_{i, \theta} \subset \I_{k, \theta}$
and either $k < 0$ or $k\ge 0$ and $\istar \in \BSG{k}{k+1}$).
The map \textcolor{blue}{$f_{m-1,\theta}$} and the corresponding intervals
\textcolor{blue}{$\I_{k,\theta}$} and \textcolor{blue}{$\I_{k+1,R_\omega(\theta)}$}
are drawn in \textcolor{blue}{blue}.
The map \textcolor{red}{$f_{m,\theta}$} and the corresponding intervals
\textcolor{red}{$\I_{i,\theta}$} and \textcolor{red}{$\I_{i+1,R_\omega(\theta)}$}
are drawn in \textcolor{red}{red}.}\label{fig:Cas3.1}
\end{figure}
In this case we have $\obasintabs{i} \subset \IBD$ and
$\obasintabs{k} \subset \IBD[m-1]$
so that, $\theta \in \IBD \cap \IBD[m-1].$
Moreover, by Lemma~\ref{Propertiesvarphi}(g),
$\basicbox{i} \subset \Int\left(\basicbox{k} \setminus \setsilift{\kstar}\right)$
and, hence,
\[
\I_{i, \theta} \subset \I_{k, \theta}.
\]
Since $\theta \in \IBD,$ by Definition~\ref{curvesinthewings} and
Lemma~\ref{normainterna},
$\theta \in \wIBD \setminus \WIB,$ and
\[
\norm{f_{m,\theta} - f_{m-1,\theta}}
= \norm{f_{m,\theta}\evalat{\I_{i, \theta}} - f_{m-1,\theta}\evalat{\I_{i, \theta}}}
= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}.
\]
Thus, we need to compare the maps $f_{m,\theta}\evalat{\I_{i, \theta}}$
and $f_{m-1,\theta}\evalat{\I_{i, \theta}}.$
To do this we consider two subcases.
\begin{autocase}[Subcase]{3.1} Either $k < 0$ or $k\ge 0$ and $\theta \in \OBG{k}{k+1}$\\
\upshape (see Figure~\ref{fig:Cas3.1} for a symbolic representation of this case).
\end{autocase}
In this situation we aim at proving that
\[
f_{m-1,\theta}\left(\I_{i, \theta}\right),
f_{m,\theta}\left(\I_{i, \theta}\right) \subset
\I_{k+1, R_\omega(\theta)}.
\]
We start with $f_{m-1,\theta}\left(\I_{i, \theta}\right).$
By Definition~\ref{seqTmDefi} and
Lemmas~\ref{gpositiva}(b) and \ref{gnegativa}(b)
we obtain
\[
f_{m-1,\theta}\left(\I_{i, \theta}\right)
\subset f_{m-1,\theta}\left(\I_{k, \theta}\right) =
g_{_{k, \theta}}\left(\I_{k, \theta}\right) \subset \I_{k+1, R_\omega(\theta)}.
\]
Next we show that
$f_{m,\theta}\left(\I_{i, \theta}\right) \subset \I_{k+1, R_\omega(\theta)}.$
Since $k < 0$ or $k\ge 0$ and $\theta \in \OBG{k}{k+1},$
by Definition~\ref{PCgenerators}(\tsfR.1) we obtain
\begin{equation}\label{eq:RotTheta}
R_\omega(\theta) \in
\begin{cases}
R_\omega\left(\obasintabs{k}\right) = \OBG{k+1}{\ak} \subset \obasintabs{k+1}\ \text{if $k < 0,$}\\
R_\omega\left(\OBG{k}{k+1}\right) = \obasint{k+1}\ \text{if $k\ge 0$ and $\theta \in \OBG{k}{k+1}$.}
\end{cases}
\end{equation}
Assume that $i < 0$ or $i \ge 0$ and $\theta \in \OBG{i}{i+1}.$
By \eqref{eq:RotTheta} with $k$ replaced by $i$,
\[
R_\omega(\theta) \in \obasintabs{i+1} \cap \obasint{k+1}
\subset \wbasint{i+1} \cap \wbasint{k+1}.
\]
Therefore, since $\akk \le \aii$ and $k+1 \ne i+1,$
from Lemma~\ref{Propertiesvarphi}(g) we obtain $\akk < \aii,$
\begin{align*}
\basintabs{i+1} &\subset \obasintabs{k+1} \setminus \sstarset{k+1},\text{ and}\\
\basicbox{i+1} &\subset \Int\left(\basicbox{k+1} \setminus \setsilift{\sstar{k+1}}\right).
\end{align*}
Thus, by Definition~\ref{seqTmDefi} and
Lemmas~\ref{gpositiva}(b) and \ref{gnegativa}(b),
\[
f_{m,\theta}\left(\I_{i, \theta}\right)
= g_{_{i, \theta}}\left(\I_{i, \theta}\right)
\subset \I_{i+1, R_\omega(\theta)} \subset \I_{k+1, R_\omega(\theta)}.
\]
Now we will consider the case
$i \ge 0$ and $\theta \in \obasint{i} \setminus \OBG{i}{i+1}.$
The fact that $\ak < \ai=i$ implies $\akk \le \ak+1 \le i.$
We claim that
\[
\OBG{i+1}{i} \subset \obasintabs{k+1} \setminus \sstarset{k+1}.
\]
To prove the claim note that, by \eqref{eq:RotTheta},
\[
R_\omega(\theta) \in R_\omega\left(\obasint{i}\right) \cap \obasintabs{k+1}
\subset \OBG{i+1}{i} \cap \wbasint{k+1}.
\]
Moreover, the interval $\OBG{i+1}{i}$
is disjoint from $\wbasint{i}$ and $\wbasint{-i}$
by Definition~\ref{PCgenerators}(\tsfR.2).
Thus, $i \ne k+1, -(k+1)$ and, hence, $\akk < i$
(that is, $k+1 \in Z_{i-1}$).
So, there exists $q\in Z_{i-1}$ such that
$\BSG{i+1}{i} \cap \wbasint{q} \ne \emptyset$ and
$ \aq \ge \akk$ is maximal verifying these conditions.
By Definition~\ref{PCgenerators}(\tsfR.4),
\[
\OBG{i+1}{i} \subset \wobasint{q} \setminus
\left(\Bd\left(\basintabs{q}\right) \cup \qstarset\right).
\]
So, the claim holds when $q = k+1.$
Assume that $q \ne k+1.$ Then,
\[
R_\omega(\theta) \in \OBG{i+1}{i} \cap \obasintabs{k+1} \subset \wobasint{q} \cap \obasintabs{k+1}.
\]
Hence, by Lemma~\ref{Propertiesvarphi}(g), $ \aq > \akk$ and
\[
\OBG{i+1}{i} \subset \wbasint{q} \subset \obasintabs{k+1} \setminus \sstarset{k+1}.
\]
This ends the proof of the claim.
On the other hand,
by Definition~\ref{PCgenerators}(\tsfR.2) and
Remark~\ref{PCgeneratorsExplicitConsequences}(\tsfR.2),
\[
\left(\BSG{i+1}{i} \setminus \obasint{i+1}\right) \cap Z_{i+1} = \emptyset.
\]
Thus, by the claim,
\begin{align*}
R_\omega(\theta)
&\in R_\omega\left(\obasint{i} \setminus \OBG{i}{i+1}\right)
= \OBG{i+1}{i} \setminus \obasint{i+1}\\
&\subset \obasintabs{k+1} \setminus Z_{i+1}.
\end{align*}
By Definition~\ref{seqTmDefi},
the definition of the maps $g_{_{i, \theta}}$ for $i \ge 0$
(Definition~\ref{defi-gi-positiva})
and Lemma~\ref{Propertiesvarphi}(d) (with $\ell = k+1$ and $n = i+1$),
\[
f_{m,\theta}\left(\I_{i, \theta}\right) =
g_{_{i, \theta}}\left(\I_{i, \theta}\right) =
\left\{\gams{i+1}\left(R_\omega(\theta)\right)\right\}
\subset \I_{k+1, R_\omega(\theta)}.
\]
Summarizing, we have proved that
\[
f_{m-1,\theta}\left(\I_{i, \theta}\right),
f_{m,\theta}\left(\I_{i, \theta}\right) \subset
\I_{k+1, R_\omega(\theta)}.
\]
So, by Lemma~\ref{Propertiesvarphi}(f) (and the fact that $\akk \ge \ak -1$),
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}
\le \diam\left(\I_{k+1, R_\omega(\theta)}\right)\\
&\le \diam\left(\basicbox{k+1}\right) \le 2^{-\akk}
\le 2 \cdot 2^{-\ak} = 2 \cdot 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
This ends the proof of the proposition in this subcase.
\begin{autocase}[Subcase]{3.2} $k \ge 0$ and $\theta \in \obasint{k} \setminus \OBG{k}{k+1}.$ \end{autocase}
We start by computing $f_{m-1,\theta}\left(\I_{i, \theta}\right).$
By Definition~\ref{seqTmDefi} and
the definition of the maps $g_{_{k, \theta}}$ for $k \ge 0$
(Definition~\ref{defi-gi-positiva}),
\[
f_{m-1,\theta}\left(\I_{i, \theta}\right)
\subset f_{m-1,\theta}\left(\I_{k, \theta}\right)
= g_{_{k, \theta}}\left(\I_{k, \theta}\right)
= \{\gams{k+1}\left(R_\omega(\theta)\right)\}.
\]
Analogously,
if $i \ge 0$ and $\theta \in \obasint{i} \setminus \OBG{i}{i+1},$
\[
f_{m,\theta}\left(\I_{i, \theta}\right)
= g_{_{i, \theta}}\left(\I_{i, \theta}\right)
= \{\gams{i+1}\left(R_\omega(\theta)\right)\}.
\]
Then, by Lemma~\ref{pointnormbound},
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \norm{f_{m,\theta}\evalat{\I_{i, \theta}} - f_{m-1,\theta}\evalat{\I_{i, \theta}}}\\
&= \abs{\gams{i+1}\left(R_\omega(\theta)\right) - \gams{k+1}\left(R_\omega(\theta)\right)}
\le 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
Assume now that $i < 0$ or $i \ge 0$ and $\theta \in \OBG{i}{i+1}.$
By \eqref{eq:RotTheta},
Definition~\ref{seqTmDefi} and Lemmas~\ref{gpositiva}(b) and
\ref{gnegativa}(b)
\begin{align*}
R_\omega(\theta) & \in \obasintabs{i+1},\text{ and}\\
f_{m,\theta}\left(\I_{i, \theta}\right)
& = g_{_{i, \theta}}\left(\I_{i, \theta}\right)
\subset \I_{i+1, R_\omega(\theta)}.
\end{align*}
Moreover, if $k+1 < \aii,$
by Lemmas~\ref{QuePassaALesAles}(a) and \ref{Propertiesvarphi}(c),
we have
\[
f_{m-1,\theta}\left(\I_{i, \theta}\right)
= \left\{\gams{k+1}\left(R_\omega(\theta)\right)\right\}
= \left\{\gams{\aii-1}\left(R_\omega(\theta)\right)\right\}
\subset \I_{i+1, R_\omega(\theta)}.
\]
Therefore, by Lemma~\ref{Propertiesvarphi}(f),
\begin{align*}
\norm{f_{m,\theta} - f_{m-1,\theta}}
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - f_{m-1,\theta}(x)}\\
&= \LSleftlimits{\sup}{x \in \I_{i, \theta}} \abs{f_{m,\theta}(x) - \gams{\aii-1}\left(R_\omega(\theta)\right)}\\
&\le \diam\left(\I_{i+1, R_\omega(\theta)}\right)
\le \diam\left(\basicbox{i+1}\right) \le 2^{-\aii}\\
&< 2^{-(k+1)} < 2^{-\abs{\bt[m-1]{\theta}}}.
\end{align*}
So, to end the proof of the proposition we have to show that,
in this subcase, $k+1 < \aii.$
To prove this, notice that when $i \ge 0,$
$k + 1 = \ak +1 < \ai + 1 = \aii.$
So, assume by way of contradiction that
$i < 0$ and $k+1 = \aii$ (recall that $k+1 \le \aii$).
Then, $k+1 = - (i+1)$ and, hence,
\begin{align*}
R_\omega(\theta)
&\in R_\omega\left(\obasint{k}\right) = \OBG{k+1}{k}, \text{ and}\\
R_\omega(\theta)
&\in \obasintabs{i+1} = \OBG{-(k+1)}{k+1} \subset \wobasint{-(k+1)},
\end{align*}
which is a contradiction by Definition~\ref{PCgenerators}(\tsfR.2).
\end{proof}
\def\cprime{$'$}
|
\section{Introduction}
The basic principles for a consistent time-independent quantum mechanical
treatment of quasi/pseudo-Hermitian and/or $\mathcal{PT}$-symmetric
Hamiltonians are well understood theoretically and by now widely accepted,
see e.g. \cite{Bender:1998ke,Benderrev,Alirev,BFGJ} for an overview. In
addition, many experiments have been carried out to confirm the key findings
of this approach and to make further predictions, see e.g. \cit
{Muss,MatMakris,Guo}.
A central element in such considerations is the construction of the
so-called Dyson map \cite{Dyson} that adjointly transforms a non-Hermitian
Hamiltonian to a Hermitian isospectral counterpart. Subsequently this map
can be used to manufacture a new metric operator for the physical Hilbert
space. This programme is conceptually straightforward, but it remains a
technical challenge even for simple examples \cite{Bender:2004sa,Mosta}.
Nonetheless, it has been carried out successfully for many concrete models
\cite{sinha2002iso,fringfaria,ACIso,MGH,PEGAAF2}. Alternatively, one may
also attempt to transform a model's more basic constituents, such as Lie
algebraic \cite{PEGAAF2} or bosonic \cite{Swanson,PEGAAF2} building blocks
by different types of transformations. By viewing the adjoint map as a
generalized Bogoliubov transformation we pursue here a combined approach to
achieve this goal. In \cite{FabAnd} we found that when the constraint on the
target Hamiltonian to be Hermitian is relaxed, the generalized Bogoliubov
transformations still lead to systems with $\mathcal{D}$-pseudo-bosons, see
e.g. \cite{adjbook} for an overview, as their central constituents. We will
also consider such a scenario here by maintaining the structure of a
two-fold Bogoliubov transformations, where one of them is making up the
pseudo-bosons and the other is taken to be equivalent to an adjoint map. We
will also study the situation in which the constraint on the target objects
is relaxed.
Our manuscript is organized as follows: In section 2 we define our doubly
Bogoliubov transformed pseudo-bosons. In section 3 we study a Hamiltonian
built from the pseudo-bosonic number operator in various constraint
settings. We state our conclusions and an outlook in section 4.
\section{Adjointly transformed pseudo-bosons}
We consider here systems whose basic constituents are pseudo-bosonic
creation and destruction operators $c$ and $d$, respectively. These
operators satisfy the standard canonical commutation relations $[d,c]
\mathbb{I}$, but they are not mutually Hermitian, i.e. $d^{\dagger }\neq c$.
In general Hamiltonian systems comprised out of these operators will
therefore be non-Hermitian. Motivated by the success of
pseudo/quasi-Hermitian system we address here the question of whether and
how these operators can be mapped adjointly into a pair of almost mutually
Hermitian canonical operators, as such a map could be utilized to restore
the Hermiticity of the entire Hamiltonian system. Hence we seek to solve
\begin{equation}
\eta \left(
\begin{array}{l}
d \\
\end{array
\right) \eta ^{-1}=\left(
\begin{array}{l}
e \\
\end{array
\right) ,\qquad \text{with }e^{\dagger }=\varkappa f,\varkappa \in \mathbb{R
, \label{phpb}
\end{equation
for $\eta $. This general problem may be tackled in various generic manners
depending on the type of pseudo-bosons considered. Here we choose a specific
realization by taking the pseudo-bosons to be related to the standard
canonical creation and annihilation operators, $a$ and $a^{\dagger }$ with
[a,a^{\dagger }]=\mathbb{I}$, by means of a generalized Bogoliubov
transformation $T(\alpha ,\beta ,\gamma ,\delta )$, see \cit
{Bogo,GenBog,Swanson,PEGAAF2,FabAnd}
\begin{equation}
\left(
\begin{array}{l}
d \\
\end{array
\right) =T(\alpha ,\beta ,\gamma ,\delta )\left(
\begin{array}{l}
a \\
a^{\dagger
\end{array
\right) .\quad \label{pbosons}
\end{equation
Notice that other possibilities exist and see \cite{FabAnd} for a detailed
discussion on what the choice (\ref{pbosons}) entails. For the matrix $T$ we
assume the for
\begin{equation}
T(\alpha ,\beta ,\gamma ,\delta )=\left(
\begin{array}{cc}
\beta & -\delta \\
-\alpha & \gamm
\end{array
\right) ,\qquad \det T(\alpha ,\beta ,\gamma ,\delta )=1,\qquad \alpha
,\beta ,\gamma ,\delta \in \mathbb{C}, \label{detT}
\end{equation
such that $d=\beta a-\delta a^{\dagger }$ and $c=-\alpha a+\gamma a^{\dagger
}$. Whereas $a$ and $a^{\dagger }$ are mutually Hermitian, $\left( a\right)
^{\dagger }=$ $a^{\dagger }$, this is obviously not the case for the
pseudo-bosons, unless $\beta =\gamma ^{\ast }$ and $\alpha =\delta ^{\ast }$.
Next we assume that the adjoint action on the standard bosons can also be
realized by a generalized Bogoliubov transformatio
\begin{equation}
\eta \left(
\begin{array}{l}
a \\
a^{\dagger
\end{array
\right) \eta ^{-1}=T(\hat{\alpha},\hat{\beta},\hat{\gamma},\hat{\delta
)\left(
\begin{array}{l}
a \\
a^{\dagger
\end{array
\right) \mathbf{.} \label{B2}
\end{equation
This is indeed possible, taking for instance $\eta $ to be the positive
Hermitian operato
\begin{equation}
\eta =\exp (\varepsilon a^{\dagger }a+\nu aa+\nu ^{\ast }a^{\dagger
}a^{\dagger }),\qquad \text{with }\varepsilon ^{2}\geq 4\left\vert \nu
\right\vert ^{2},
\end{equation
we find that the parameter $\hat{\alpha},\hat{\beta},\hat{\gamma},\hat{\delt
}$ and $\varepsilon ,\nu $ are related as
\begin{equation}
\hat{\alpha}=-2\frac{\nu }{\theta }\sinh \theta ,\quad \hat{\beta}=\cosh
\theta -\frac{\varepsilon }{\theta }\sinh \theta ,\quad \hat{\gamma}=\cosh
\theta +\frac{\varepsilon }{\theta }\sinh \theta ,\quad \hat{\delta}=2\frac
\nu ^{\ast }}{\theta }\sinh \theta , \label{sol}
\end{equation
where $\theta :=\sqrt{\varepsilon ^{2}-4\left\vert \nu \right\vert ^{2}}$.
The assumption $\det T(\hat{\alpha},\hat{\beta},\hat{\gamma},\hat{\delta})=1$
holds without any further constraint.
We may now solve (\ref{phpb}) by computin
\begin{equation}
T(\alpha ,\beta ,\gamma ,\delta )\cdot T(\hat{\alpha},\hat{\beta},\hat{\gamm
},\hat{\delta})\left(
\begin{array}{l}
a \\
a^{\dagger
\end{array
\right) \mathbf{=}\eta T(\alpha ,\beta ,\gamma ,\delta )\left(
\begin{array}{l}
a \\
a^{\dagger
\end{array
\right) \eta ^{-1}=\eta \left(
\begin{array}{l}
d \\
\end{array
\right) \eta ^{-1}, \label{TT}
\end{equation
where we used that evidently $[\eta ,T]=0$ and (\ref{pbosons}). From the
matrix multiplication on the left hand side and (\ref{phpb}) we obtai
\begin{equation}
T(\tilde{\alpha},\tilde{\beta},\tilde{\gamma},\tilde{\delta})\left(
\begin{array}{l}
a \\
a^{\dagger
\end{array
\right) =\left(
\begin{array}{l}
e \\
\end{array
\right) , \label{Ttil}
\end{equation
with
\begin{equation}
\tilde{\alpha}=\alpha \hat{\beta}+\gamma \hat{\alpha},\quad \tilde{\beta
=\beta \hat{\beta}+\delta \hat{\alpha},\quad \tilde{\gamma}=\alpha \hat
\delta}+\gamma \hat{\gamma},\quad \tilde{\delta}=\beta \hat{\delta}+\delta
\hat{\gamma}.
\end{equation
Since the determinants of $T(\alpha ,\beta ,\gamma ,\delta )$ and $T(\hat
\alpha},\hat{\beta},\hat{\gamma},\hat{\delta})$ are $1$, we also have $\det
T(\tilde{\alpha},\tilde{\beta},\tilde{\gamma},\tilde{\delta})=1$. Depending
now on the constraints imposed on the Bogoliubov transformation parameters
\alpha ,\beta ,\gamma ,\delta $ and those entering from the adjoint action
\varepsilon ,\nu $ we obtain different types of scenarios, which we now
investigate for a concrete model.
\section{Pseudo-bosonic Hamiltonians}
We consider here a system described by a Hamiltonian consisting of the
pseudo-bosonic number operator $N=cd$ of the type studied in \cite{FabAnd}
\begin{equation}
\mathcal{H}(d,c)=\hbar \omega \left( cd+\frac{1}{2}\right) , \label{1}
\end{equation
where in comparison we re-introduced the standard angular frequency $\omega
\in \mathbb{R}$ and the reduced Planck constant $\hbar $. Assuming here that
the pseudo-bosons are generated by a generalized Bogoliubov transformation
as specified in (\ref{pbosons}) the Hamiltonian in (\ref{1}) acquires the
form of a Swanson Hamiltonian \cite{Swanson}
\begin{equation}
\mathcal{H}(a,a^{\dagger })=\hbar \omega \left( \alpha \delta aa^{\dagger
}+\beta \gamma a^{\dagger }a-\alpha \beta aa-\gamma \delta a^{\dagger
}a^{\dagger }+\frac{1}{2}\right) . \label{HH}
\end{equation
Evidently the Hamiltonian $\mathcal{H}$ is non-Hermitian whenever $\alpha
\delta \neq \alpha ^{\ast }\delta ^{\ast }$ or $\beta \gamma \neq \beta
^{\ast }\gamma ^{\ast }$ or $\alpha \beta \neq \gamma ^{\ast }\delta ^{\ast
} $. We note that the transformation $\alpha \leftrightarrow \delta ^{\ast }
, $\beta \leftrightarrow \gamma ^{\ast }$ maps $\mathcal{H\rightarrow H
^{\dagger }$, which implies that our Hamiltonian becomes Hermitian when this
transformation becomes a symmetry. Let us now consider various cases for
possible constraints on the parameters involved.
\subsection{Hermitian constraint}
It is instructive to consider at first the simplest special scenario with
\alpha ,\beta ,\gamma ,\delta \in \mathbb{R}$ and the additional constraint
\alpha \beta =\gamma \delta $. In this case the Hamiltonian $\mathcal{H}$ in
(\ref{HH}) evidently becomes Hermitian, so that we do not require the
similarity transformation (\ref{phpb}) to achieve this, such that $\eta
\mathbb{I}$. We may solve the constraint together with the restrictions on
the determinant of the Bogoliubov transformation in (\ref{detT}) for two of
the constants, e.g
\begin{equation}
\beta =\frac{\gamma }{\gamma ^{2}-\alpha ^{2}},\qquad \text{and\qquad
\delta =\frac{\alpha }{\gamma ^{2}-\alpha ^{2}}. \label{bd}
\end{equation
Notice that although $\mathcal{H}=\mathcal{H}^{\dagger }$, we still maintain
the pseudo-bosonic property $d^{\dagger }\neq c$ wit
\begin{equation}
d^{\dagger }=\frac{1}{\gamma ^{2}-\alpha ^{2}}c.
\end{equation
Using now the common representation for the canonical creation and
annihilation operator
\begin{equation}
a(m,\omega ):=\frac{1}{\sqrt{2\hbar }}\left( \sqrt{m\omega }x+\frac{i}{\sqrt
m\omega }}p\right) ,\qquad a^{\dagger }(m,\omega ):=\frac{1}{\sqrt{2\hbar }
\left( \sqrt{m\omega }x-\frac{i}{\sqrt{m\omega }}p\right) , \label{cc}
\end{equation
in terms of the canonical coordinate and momentum operator $x$ and $p$,
obeying $[x,p]=i\hbar $, the Hamiltonian (\ref{1}) acquires the form of the
standard harmonic oscillator Hamiltonia
\begin{equation}
\mathcal{H}=h=\hbar \omega \left[ a^{\dagger }(M,\omega )a(M,\omega )+\frac{
}{2}\right] =\frac{1}{2M}p^{2}+\frac{M\omega ^{2}}{2}x^{2}, \label{Harm}
\end{equation
albeit with a modified mas
\begin{equation}
M=m\frac{\gamma -\alpha }{\gamma +\alpha }=m\frac{\beta -\delta }{\beta
+\delta }.
\end{equation
In order to ensure the mass to be physical, that is positive, we require
\gamma >\alpha >0$ and $\beta >\delta >0$, which together with $\alpha \beta
=\gamma \delta $ are precisely the constraints encountered in \cite{FabAnd}
for this situation as a requirement for the eigenfunctions of $\mathcal{H}$
to be square-integrable. Given the well-known form for the normalized
eigenfunctions for the Hamiltonian (\ref{Harm}) in terms of Hermite
polynomials $H_{n}$
\begin{equation}
\phi _{n}(x)=\frac{1}{\sqrt{n!2^{n}}}\left( \frac{M\omega }{\pi }\right)
^{1/4}e^{-\omega Mx^{2}/2}H_{n}(M\omega x),
\end{equation
these two requirements are therefore the same. In other words, square
integrability and positivity of the mass become synonymous, depending both
explicitly on the model parameters.
\subsection{Pseudo-Hermitian constraint}
Let us now relax the constraint $\alpha \beta =\gamma \delta $ and only
assume $\alpha ,\beta ,\gamma ,\delta \in \mathbb{R}$. We carry out an
analysis using standard techniques developed in the context of
pseudo/quasi-Hermitian and/or $\mathcal{PT}$-symmetric quantum mechanics as
outlined in \cite{Urubu,Bender:1998ke,Benderrev,Alirev}. To establish our
notation we briefly recall the key formulae.
Given two isospectral Hamiltonians one of which is Hermitian $h=h^{\dagger }$
and the other is not $\mathcal{H}\neq \mathcal{H}^{\dagger }$, related by a
similarity transformatio
\begin{equation}
h=\eta \mathcal{H}\eta ^{-1}, \label{hH}
\end{equation
the corresponding eigenstates for the eigenvalue equations
\begin{equation}
\mathcal{H}\left\vert \psi \right\rangle =E\left\vert \psi \right\rangle
\qquad \text{and\qquad }h\left\vert \phi \right\rangle =E\left\vert \phi
\right\rangle ,
\end{equation
are related a
\begin{equation}
\left\vert \psi \right\rangle =\eta ^{-1}\left\vert \phi \right\rangle .
\end{equation
The expectation values for any observable $\mathcal{O}$ and $\mathit{o}$, in
a non-Hermitian and corresponding Hermitian counterpart, respectively, are
related as $\left\langle \psi \right\vert \mathcal{O}\left\vert \psi
\right\rangle _{\eta }=$ $\left\langle \phi \right\vert \mathit{o}\left\vert
\phi \right\rangle $. Here the inner product is defined as $\left\langle
\psi \right\vert \left. \psi \right\rangle _{\eta }:=\left\langle \psi
\right\vert \eta ^{\dagger }\eta \left\vert \psi \right\rangle $ where the
positive operator $\eta ^{\dagger }\eta $ plays the role of the metric.
Let us now use the above formulae to carry out the similarity transformation
for the non-Hermitian Hamiltonian (\ref{1}) to a Hermitian isospectral
counterpart
\begin{equation}
h=\eta \mathcal{H}(d,c)\eta ^{-1}=\hbar \omega \left( \eta c\eta ^{-1}\eta
d\eta ^{-1}+\frac{1}{2}\right) =\hbar \omega \left( fe+\frac{1}{2}\right) .
\label{h}
\end{equation
The transformed Hamiltonian is Hermitian when $e^{\dagger }=\varkappa f$
with $\varkappa \in \mathbb{R}$, which according to (\ref{Ttil}) can be
achieved by imposing the two additional constraint
\begin{equation}
\tilde{\delta}=\varkappa \tilde{\alpha},\quad \text{and\quad }\tilde{\beta
=\varkappa \tilde{\gamma}, \label{xi}
\end{equation
with $\nu \in \mathbb{R}$. Eliminating $\varkappa $ by using the explicit
expressions in (\ref{sol}), these two equations reduce to
\begin{equation}
\frac{\tanh 2\theta }{\theta }=\frac{\alpha \beta -\gamma \delta }
\varepsilon (\alpha \beta +\gamma \delta )+2\nu (\beta \gamma +\alpha \delta
)}. \label{const}
\end{equation
Assuming the parameters $\alpha ,\beta ,\gamma ,\delta $ to be model
specific, and therefore fixed, the two additional parameters $\varepsilon $
and $\nu $ that entered through the second generalized Bogoliubov
transformation are constrained by (\ref{const}). Thus there is only one free
parameter left, which reflects the typical ambiguity present in
pseudo-Hermitian systems, see \cite{Urubu} for the Swanson model at hand.
Parameterizing one of the free parameters in terms of the other as $\nu
=\lambda \varepsilon /2$, we obtain $\varepsilon $ as a function of the new
variable $\lambda $
\begin{equation}
\varepsilon (\lambda )=\frac{1}{2\sqrt{1-\lambda ^{2}}}\func{arctanh}\left[
\frac{(\alpha \beta -\gamma \delta )\sqrt{1-\lambda ^{2}}}{(\alpha \beta
+\gamma \delta )+\lambda (\beta \gamma +\alpha \delta )}\right] ,\qquad
\lambda :=\frac{2\nu }{\varepsilon }. \label{eps}
\end{equation
The restrictions of the interval in which $\lambda $ is taken results from
demanding $\theta \in \mathbb{R}$. Recalling that $\varepsilon $ is real, we
restrict the argument of the $\func{arctanh}$ to be bounded by $\pm 1$.
Choosing for definiteness as specific ordering $0<\alpha /\gamma <\delta
/\beta <1$, we need to restrict $\lambda $ further to be in the disconnected
interval
\begin{equation}
-\infty <\lambda <-\frac{2\beta \delta }{\beta ^{2}+\delta ^{2}},\qquad
\text{or\qquad }-\frac{2\alpha \gamma }{\alpha ^{2}+\gamma ^{2}}<\lambda
<\infty , \label{lint}
\end{equation
for $\varepsilon $ to remain real. Notice that the restrictions on the model
parameters are just needed to ensure that the constraints (\ref{lint})
become unique. The intervals are only connected for $\alpha \beta =\gamma
\delta $, which corresponds to the Hermitian case discussed in the previous
subsection.
Given the above transformations we can of course express our isospectral
Hamiltonian (\ref{h}) also in terms of the standard bosonic operator
\begin{equation}
h(m,\omega )=\hbar \omega \left[ \mu _{a^{\dagger }a}a^{\dagger }(m,\omega
)a(m,\omega )-\mu _{aa}a(m,\omega )a(m,\omega )-\mu _{a^{\dagger }a^{\dagger
}}a^{\dagger }(m,\omega )a^{\dagger }(m,\omega )+\mu _{0}\right] ,
\end{equation
with coefficient
\begin{equation}
\mu _{a^{\dagger }a}=\tilde{\alpha}\tilde{\delta}+\tilde{\beta}\tilde{\gamma
,\quad \mu _{aa}=\tilde{\alpha}\tilde{\beta},\quad \mu _{a^{\dagger
}a^{\dagger }}=\tilde{\gamma}\tilde{\delta},\quad \text{and}\quad \mu _{0}
\frac{1}{2}+\tilde{\alpha}\tilde{\delta}.
\end{equation
The constraint (\ref{xi}) guarantees that $\mu _{aa}=$\ $\mu _{a^{\dagger
}a^{\dagger }}$, such that together with $\mu _{0},\mu _{a^{\dagger }a}\in
\mathbb{R}$ the Hamiltonian $h$ becomes Hermitian. In fact these constraints
are familiar from a general treatment of the Swanson model \cite{MGH}, which
is a special case of the more general model studied in \cite{PEGAAF2} and
precisely agrees when matching the constants appropriately. However, whereas
in \cite{MGH,PEGAAF2} the constraints resulted from an analysis of the
Hamiltonian, they emerge here as the combination of two constraints on their
more basic pseudo-bosonic constituents.
Just as the Hermitian case in the previous subsection, when implementing the
constraints the Hamiltonian $h$ can be brought into the form of a harmonic
oscillator
\begin{equation}
h(\hat{M},\omega )=\hbar \omega \left[ a^{\dagger }(\hat{M},\omega )a(\hat{M
,\omega )+\frac{1}{2}\right] =\frac{1}{2\hat{M}}p^{2}+\frac{\hat{M}\omega
^{2}}{2}x^{2}, \label{HHH}
\end{equation
again with modified mass
\begin{equation}
\hat{M}=\frac{m}{(\tilde{\alpha}+\tilde{\gamma})(\tilde{\beta}+\tilde{\delta
)}.
\end{equation
Notice that using the constraint (\ref{xi}) we only need to take $m\varkappa
>0$ in order to ensure that $\hat{M}$ is positive, when $\tilde{\alpha}
\tilde{\beta},\tilde{\gamma},\tilde{\delta}\in \mathbb{R}.$ Viewing $\alpha
,\beta ,\gamma ,\delta $ as model defining parameters, it follows from (\re
{sol}), (\ref{xi}) and (\ref{eps}), that for a specific model we may regard
\hat{M}$ and $\omega $ as functions of the single parameter $\lambda $. For
some specific choices we show in figure 1 the modified mass as a function of
$\lambda $. We observe that it is positive and according to (\ref{lint}) not
defined in the specified interval for $\lambda $.
We recall that $\lambda $ is not a model dependent parameter as it simply
entered through the adjoint map labeling infinitely many pseudo-Hermitian
counterparts to $\mathcal{H}(d,c)$. Any of the theories respecting the
constraint (\ref{lint}) is well defined. As noted in \cite{MGH}, the
theories for $\lambda =0,\pm 1$ are somewhat special as then some of the
auxiliary variables can be interpreted directly as the number operator, the
coordinate or the momentum. In these cases we fin
\begin{equation}
\hat{M}(\lambda =0)=\frac{m}{\left( \sqrt{\alpha \delta }+\sqrt{\beta \gamma
}\right) ^{2}},\quad \text{and\quad }\hat{M}(\lambda =\pm 1)=\frac{m}{\left[
\left( \gamma \pm \alpha \right) \left( \beta \pm \delta \right) \right]
^{\pm 1}}, \label{MM}
\end{equation
which are all well-defined for the ordering considered here. We may confirm
these expression for the values used in our examples in figure 1. For
instance, for the choice of parameters corresponding to the solid black line
we compute\ $\hat{M}(\lambda =0)=0.672$, $\hat{M}(\lambda =1)=0.518$ and
\hat{M}(\lambda =-1)=0.48$ numerically and also obtain the same values from
the explicit analytical expression (\ref{MM}).
\FIGURE{\epsfig{file=Mass.eps,height=6.5cm}[h]
\caption{Modified mass $M$ as a function of $\lambda$ for $m=1$ with $\lambda \in \mathbb{R}\backslash (-0.9665,-0.0998)$, $\lambda \in \mathbb{R}\backslash
(-0.7534,-0.3846)$, $\lambda \in \mathbb{R}\backslash (-0.9182,-0.8)$, $\lambda \in \mathbb{R}\backslash (-0.80709,-0.8)$ and $\lambda \in \mathbb{R}\backslash (-0.1978,-0.0200)$.}
\label{MassL}}
We expect to recover the Hermitian case by demanding either the position and
the momentum, the position and the number operator or the momentum and the
number operator to be observables. Indeed by equating any two of these
masses, i.e. $\hat{M}(0)=\hat{M}(\pm 1)$ or $\hat{M}(1)=\hat{M}(-1)$,
together with the constraint on the determinant of $T$ leads to the values
in (\ref{bd}) for two of the parameters.
\subsection{Adjoint constraint}
Let us now also relax the constraint (\ref{xi}), such that $e^{\dagger }\neq
\varkappa f$ and in addition allow $\alpha ,\beta ,\gamma ,\delta \in
\mathbb{C}$. Whereas in the previous subsections the construction of the
eigenfunctions follows trivially from the harmonic oscillator realization in
terms of modified masses and frequencies, this is less obvious for this
setting. We therefore present the construction commencing by expressing the
pseudo-bosons $e$ and $f$ in position space. From (\ref{Ttil}) and (\ref{cc
) simply follows
\begin{equation}
e=\frac{1}{\sqrt{2}}\left[ \left( \tilde{\beta}-\tilde{\delta}\right)
x+\left( \tilde{\beta}+\tilde{\delta}\right) \frac{d}{dx}\right] ,\quad f
\frac{1}{\sqrt{2}}\left[ \left( \tilde{\gamma}-\tilde{\alpha}\right)
x-\left( \tilde{\alpha}+\tilde{\gamma}\right) \frac{d}{dx}\right] ,
\end{equation
As we no longer need modified values for the mass and frequency we have set
here $m=\omega =\hbar =1$ for simplicity. Hence, following \cite{FabAnd},
the vacua of $e$ and $f^{\dagger }$ are
\begin{equation}
\varphi _{0}(x)=N_{\varphi }e^{-\frac{1}{2}\,x^{2}\,\frac{\tilde{\beta}
\tilde{\delta}}{\tilde{\beta}+\tilde{\delta}}},\quad \quad \text{and\quad
\quad }\Psi _{0}(x)=N_{\Psi }e^{-\frac{1}{2}\,x^{2}\,\frac{\tilde{\gamma
^{\ast }-\tilde{\alpha}^{\ast }}{\tilde{\gamma}^{\ast }+\tilde{\alpha}^{\ast
}}}, \label{f03}
\end{equation
respectively, where $N_{\varphi }$ and $N_{\Psi }$ are suitable
normalization constants to be specified further below. Naturally we require
\begin{equation}
\func{Re}\left( \frac{\tilde{\beta}-\tilde{\delta}}{\tilde{\beta}+\tilde
\delta}}\right) >0,\quad \quad \text{and\quad \quad }\func{Re}\left( \frac
\tilde{\gamma}-\tilde{\alpha}}{\tilde{\gamma}+\tilde{\alpha}}\right) >0,
\label{f04}
\end{equation
to ensure the square-integrability for both of these functions. In complete
analogy to \cite{FabAnd} we further construct the functions
\begin{eqnarray}
\varphi _{n}(x) &=&\frac{f^{n}}{\sqrt{n!}}\varphi _{0}(x)=\frac{N_{\varphi
}{\sqrt{n!2^{n}}}\left( \frac{\tilde{\alpha}+\tilde{\gamma}}{\tilde{\beta}
\tilde{\delta}}\right) ^{n/2}H_{n}\left[ \frac{x}{\sqrt{(\tilde{\alpha}
\tilde{\gamma})(\tilde{\beta}+\tilde{\delta})}}\right] \,e^{-\frac{1}{2
\,x^{2}\,\frac{\tilde{\beta}-\tilde{\delta}}{\tilde{\beta}+\tilde{\delta}}},
\\
\Psi _{n}(x) &=&\frac{{e^{\dagger }}^{n}}{\sqrt{n!}}\Psi _{0}(x)=\frac
N_{\Psi }}{\sqrt{n!2^{n}}}\left( \frac{\tilde{\beta}^{\ast }+\tilde{\delta
^{\ast }}{\tilde{\alpha}^{\ast }+\tilde{\gamma}^{\ast }}\right) ^{n/2}H_{n
\left[ \frac{x}{\sqrt{(\tilde{\alpha}^{\ast }+\tilde{\gamma}^{\ast })(\tilde
\beta}^{\ast }+\tilde{\delta}^{\ast })}}\right] \,e^{-\frac{1}{2}\,x^{2}\
\frac{\tilde{\gamma}^{\ast }-\tilde{\alpha}^{\ast }}{\tilde{\gamma}^{\ast }
\tilde{\alpha}^{\ast }}}. \notag
\end{eqnarray
from a repeated action of $f$ and ${e^{\dagger }}$ on the corresponding
ground states in (\ref{f03}) for $n\geq 0$. When the constraint (\ref{f04})
holds these functions are square-integrable and can be used to define the
sets $\mathcal{F}_{\varphi }:=\{\varphi _{n},n\geq 0\}$ and $\mathcal{F
_{\Psi }:=\{\Psi _{n},n\geq 0\}$. Applying here what was proven in \cit
{FabAnd}, we deduce that $\mathcal{F}_{\varphi }$ and $\mathcal{F}_{\Psi }$
form biorthogonal bases for $\mathcal{H}$ when $\tilde{\alpha}\,\tilde{\beta
=\tilde{\gamma}\,\tilde{\delta}$, i.e. we have
\begin{equation}
\chi (x)=\sum_{n=0}^{\infty }\left\langle \Psi _{n},\chi \right\rangle
\varphi _{n}(x)=\sum_{n=0}^{\infty }\left\langle \varphi _{n},\chi
\right\rangle \Psi _{n}(x),
\end{equation
for all $\chi (x)\in \mathcal{L}^{2}(\mathbb{R)}$. Notice that $\tilde{\alph
}\,\tilde{\beta}=\tilde{\gamma}\,\tilde{\delta}$ is simply the constraint
\ref{xi}) when eliminating $\varkappa $, such that $h$ becomes self-adjoint.
When this constraint is relaxed, i.e. $\tilde{\alpha}\,\tilde{\beta}\neq
\tilde{\gamma}\,\tilde{\delta}$, the two sets still form ${\mathcal{D}}
-quasi bases, i.e. we have
\begin{equation}
\left\langle \chi ,\xi \right\rangle =\sum_{n=0}^{\infty }\left\langle \chi
,\Psi _{n}\right\rangle \left\langle \varphi _{n},\xi \right\rangle
=\sum_{n=0}^{\infty }\left\langle \chi ,\varphi _{n}\right\rangle
\left\langle \Psi _{n},\xi \right\rangle ,
\end{equation
for all $\chi ,\xi \in {\mathcal{D}}$, a dense subset of $\mathcal{L}^{2}
\mathbb{R)}$ defined as follows:
\begin{equation}
{\mathcal{D}}=\left\{ \chi (x)\in \mathcal{L}^{2}(\mathbb{R)}:\,e^{\frac{1}{
}x^{2}|\tilde{\alpha}\tilde{\beta}-\tilde{\gamma}\tilde{\delta}|}\chi (x
\mathbb{\in }\mathcal{L}^{2}(\mathbb{R)}\right\} .
\end{equation}
It is possible to verify that each $\Psi _{n}(x)$ is in the domain of $\eta $
while each $\varphi _{n}(x)$ is in the domain of $\eta ^{-1}$, that is $\Psi
_{n}(x)\in D(\eta )$ and $\varphi _{n}(x)\in D(\eta ^{-1})$. We prove this
as follows: By the definition of $\varphi _{n}$ we know that $\varphi _{n+1}
\frac{1}{\sqrt{n+1}}\,f\varphi _{n}$, for all $n\geq 0$. Since $\varphi
_{k}\in \mathcal{L}^{2}(\mathbb{R)}$ for all $k$, we conclude that $\varphi
_{n}\in D(f)$. Recalling now that $f=\eta c\eta ^{-1}$, it follows that
D(f)\subseteq D(\eta ^{-1})$. Hence, since $\varphi _{n}\in D(f)$ this means
$\varphi _{n}\in D(\eta ^{-1})$. Similarly we can check that each $\Psi
_{n}\in D(\eta )$. These facts are important, since they imply that both
\eta $ and $\eta ^{-1}$ are densely defined. In fact, $\eta ^{-1}$ is
defined on $\mathcal{L}_{\varphi }$, the linear span of the $\varphi _{n}
's, which is dense in $\mathcal{H}$ since $\mathcal{F}_{\varphi }$ is
complete (or, if $\tilde{\alpha}\,\tilde{\beta}=\tilde{\gamma}\,\tilde{\delt
}$, is even a basis). Similarly, $\eta $ is defined on $\mathcal{L}_{\Psi }
, the linear span of the $\Psi _{n}$'s, which is dense in $\mathcal{H}$
since $\mathcal{F}_{\Psi }$ is complete (or, again, if $\tilde{\alpha}\
\tilde{\beta}=\tilde{\gamma}\,\tilde{\delta}$, is a basis).
More detail on this case may be found in \cite{FabAnd}.
\section{Conclusions}
We further investigated a particular type of pseudo-bosons that are
obtainable from generalized Bogoliubov transformations \cite{FabAnd}. We
apply on them a second generalized Bogoliubov transformation satisfying
certain properties and demand that it equals a particular adjoint map acting
on these operators. We employ these doubly transformed operators to build a
simple Hamiltonian consisting of these special pseudo-bosonic number
operators.
We impose constraints on the model parameters, which we gradually relax.
Choosing at first the model parameters in such a way that the Hamiltonian
becomes Hermitian we found that one requires further simple constraints on
the ordering of the model parameters in order to obtain a positive mass.
This requirement turned out to be the same as demanding square integrability
of the wave functions. As the next case we demand the adjoint map to be
equivalent to the Dyson map that achieves pseudo-Hermiticity. In this
setting, we obtain the typical scenario in pseudo-Hermitian systems namely a
whole ray of equivalent Hermitian Hamiltonians (\ref{HHH}) parametrized by a
non-model dependent quantity, $\lambda $ in our case, entering through the
similarity transformation. We found that $\lambda $ is always defined on two
disjoint intervals on the real line. In the excluded parameter regime the
mass becomes complex as a consequence of $\varepsilon (\lambda )$ in (\re
{eps}) becoming complex. Finally when relaxing all constraints we loose the
proper that $\mathcal{F}_{\varphi }$ and $\mathcal{F}_{\Psi }$ form
biorthogonal bases as in the Hermitian scenario, but we still obtain $
\mathcal{D}}$-quasi bases.
The main virtue of our construction lies in the reduction of the relevant
transformations to the more basic bosonic ingredients. Even though the
doubly Bogoliubov transformed objects are more restrictive when compared to
the most general treatment, they always select out a set of feasible models.
Naturally, they may be employed in other more complicated models, involving
for instance cubic \cite{PEGAAF} or higher order terms in its defining
Hamiltonian.\medskip
\noindent \textbf{Acknowledgments:} FB gratefully acknowledges financial
support from City University London, from the Universit\`{a} di Palermo, via
CORI 2014, Action D and from G.N.F.M.
|
\section{Introduction}
Testing whether two ordinal random variables are independent given a sample ${(x_i, y_i)}_{i=1}^n$ is a classic problem in statistics. Early efforts such as Pearson's correlation and Kendall's $\tau$ focused on testing against linear or monotone relationships. The first test for any type of independence was provided by \cite{Hoeffding}. This test is based on multiple partitions of the $(X,Y)$ plane into four quadrants where the number of points in a quadrant is compared to what it would be under independence. This score can be calculated in $O(n \log n)$ by counting the inversions of the permutation from the ranks of ${(x_i)}_{i=1}^n$ to ${(y_i)}_{i=1}^n$ as described in \cite{Hellerb}. Additional rank based scores were suggested by others, typically based on finer partitions of the plane (see \cite{Hellera} for review, algorithms and a powerful method that takes into account all possible partitions).
\cite{Bergsma} present another measure of independence which in the case of no ties is, also based on partitions of the plane into four quadrants. This method is in some sense a generalization of Kendall's $\tau$. Using the notation of \cite{Weihs} (in their equation 1) the statistic is defined as:
\begin{align}
\label{eq:t_star}
t^* &:= \frac{(n-4)!}{n!}\sum_{\substack{1\leq i,j,k,l \leq n \\ i,j,k,l\ \text{distinct}}} a(x_i,x_j,x_k,x_l)a(y_i,y_j,y_k,y_l)
\end{align}
where
\begin{align}
a(z_1,z_2,z_3,z_4) := sign(|z_1-z_2| + |z_3-z_4| - |z_1-z_3| - |z_2-z_4|).
\end{align}
This definition clearly shows that the statistic can be naively calculated in $O(n^4)$ since one can simply go over all quadruples of points. However, \cite{Weihs} show that the statistic can be calculated in $O(n^2logn)$ using red black trees. In this note we will show that the statistic can in fact be calculated in only $O(n^2)$ using methods described in \cite{Hellera}.
\section{The algorithms}
As \cite{Bergsma} show, the score is actually based on the number of concordant quadruples vs. the number of discordant quadruples. In a manner analogous to Kendall's $\tau$ concordance and discordance are defined for a quadruple ${(x_i, y_i)}_{i=1}^4$ as follows:
\begin{definition}
A quadruple is called \textbf{concordant} if either ($\max(x_1,x_2)<\min(x_3,x_4)$ and $\max(y_1,y_2)<\min(y_3,y_4)$) or ($\max(x_1,x_2)<\min(x_3,x_4)$ and $\min(y_1,y_2)>\max(y_3,y_4)$)
\end{definition}
\begin{definition}
A quadruple is called \textbf{discordant} if $\max(x_1,x_2)<\min(x_3,x_4)$ and $\max(y_1,y_2)>\min(y_3,y_4)$ and $\min(y_1,y_2)<\max(y_3,y_4)$
\end{definition}
We start with the simple case without ties.
\subsection{The algorithm without ties}
It is easy to see that in this case the statistic reduces to (equation 4 in \cite{Weihs})
\begin{align*}
t^* = \frac{(n-4)!}{n!}(24\cdot N_c) - \frac{1}{3}, \label{eq:simple_untied_t_star}
\end{align*}
where $N_c$ is the number of concordant quadruples. Therefore all we need to do is calculate $N_c$. Again as in \cite{Weihs} et. al clearly:
\begin{align*}
N_c = \sum_{3\leq k\leq n-1} \sum_{k<l\leq n}{M_<(k,l) \choose 2} + {M_>(k,l) \choose 2}
\end{align*}
where they define
\begin{align*}
M_<(k,l) &:= |\{i:\ x_i < \min(x_k, x_l),\ y_i < \min(y_k, y_l)\}|, \\
M_>(k,l) &:= |\{i:\ x_i < \min(x_k, x_l),\ y_i > \max(y_k, y_l)\}|.
\end{align*}
However contrary to their algorithm we show that $M_<(k,l)$ and $M_>(k,l)$ can be calculated in $O(1)$ and not in $O(\log n)$ after a preprocessing step which takes $O(n^2)$. This can be done by the methods described in \cite{Hellera} as follows:
We first note that the statistic is only based on ranks so we transform every pair $(x_i,y_i)$ to its respective ranks $(r_i,s_i)$ where $r_i,s_i\in\{1,...,n\}$, this can of course be done in $O(n \log n)$.
We can now calculate the empirical cumulative distribution
\begin{equation}\label{eq-agg-sum-ind-A}
A(r, s) = \sum_{i=1}^nI(r_i\leq r \ \textrm{and} \ s_i\leq s), \quad (r,s) \in \{0,1,\ldots,n \}^2
\end{equation}
(where $A(0,s) =0, A(r,0)=0$) in $O(n^2)$ time and space:
First, let $B$ be the $(n+1)\times (n+1)$ zero matrix, and initialize to one $B(r_i, s_i)$ for each observation $i=1,\ldots,n$. Next, go over the grid in $s$-major order, i.e., for every $s$ go over all values of $r$, and compute:
\begin{enumerate}
\item $A(r, s) = B(r , s - 1) + B(r-1, s ) - B(r - 1, s - 1) + B(r, s)$, and
\item $B(r,s) = A(r,s)$.
\end{enumerate}
It is easy to see that $M_<(k,l)=A[rank(min(x_k,x_l))-1,rank(min(y_k,y_l))-1]$ and similarly that $M_>(k,l)=rank(min(x_k,x_l))-A[rank(min(x_k,x_l)),rank(max(y_k,y_l))]$ and therefore for each $k,l$ $M_<(k,l)$ and $M_>(k,l)$ can be calculated in O(1), resulting in a total of $O(n^2)$ as desired.
\subsection{The algorithm for data with ties}
First, for ease of notation we order the samples such that $x_1 \leq x_2...\leq x_n$.
By Lemma 1 in \cite{Weihs} in this case the score reduces to
\begin{align*}
t^* &= \frac{(n-4)!}{n!}(16\cdot N_c - 8\cdot N_d) \\
\end{align*}
Therefore, greater care must be taken in this case as it requires calculating also $N_d$ (where $N_d$ is the number of discordant quadruples), which with ties is a little more subtle since a quadruple can be neither concordant nor discordant. We will use the following ranking scheme - $n$ observations with $m$ unique values will be transformed to $n$ ranks in the range $1...m$ (so for example ${2,2,3.5,4,4,4}$ will be ranked as ${1,1,2,3,3,3}$). We first note that calculating the number of concordant pairs can be done in the same way as in the section above without ties, except that when we calculate the empirical cumulative distribution $B(r, s)$ will be initialized to the \textbf{number} of observations with ranks $(r,s)$, which can be greater than one.
We now turn to computing $N_d$. Define $N_d(k,l)=|\{i<j<k : i,j,k,l \text { are discordant}\}|$. Clearly $N_d= \sum_{3\leq k\leq n-1} \sum_{k<l\leq n}{N_d(k,l)}$. Following \cite{Weihs} for any pair of samples $(x_k,y_k)$ and $(x_l,y_l)$ such that $k<l$ we define:
\begin{allowdisplaybreaks}
\begin{align*}
top(k,l) &= |\{i: x_i<x_k\ \text{and}\ y_i> \max(y_k,y_l)\} | , \\
mid(k,l) &= |\{i: x_i<x_k\ \text{and} \ \min(y_k,y_l) < y_i< \max(y_k,y_l)\} | ,\\
bot(k,l) &= |\{i: x_i<x_k\ \text{and}\ y_i< \min(y_k,y_l)\} | ,\\
\mathit{eqMin}(k,l) &= |\{i: x_i<x_k\ \text{and}\ y_i = \min(y_k,y_l)\} |,\\
\mathit{eqMax}(k,l) &= |\{i: x_i<x_k\ \text{and}\ y_i = \max(y_k,y_l)\} |.
\end{align*}
\end{allowdisplaybreaks}
Quoting equations 11 and 12 in \cite{Weihs}
if $y_k=y_l$ then
\begin{align*}
N_d(k,l) = 0,
\end{align*}
and if $y_k\not =y_l$ then
\begin{equation} \label{eq:N_d}
\begin{aligned}
N_d(k,l) &= top(k,l)\cdot \left(mid(k,l) + \mathit{eqMin}(k,l)+ bot(k,l)\right) \\
&+ bot(k,l)\cdot \left(mid(k,l) + \mathit{eqMax}(k,l)\right) \\
&+ \mathit{eqMin}(k,l)\cdot (mid(k,l) + \mathit{eqMax}(k,l)) \\
&+ \mathit{eqMax}(k,l)\cdot mid(k,l) \\
& + {mid(k,l) \choose 2} - \sum_{y\in unique(k,l)}{|\{1\leq i< k: x_k\not= x_i\ \text{and}\ y_i=y\}| \choose 2}
\end{aligned}
\end{equation}
where
\begin{align*}
unique(k,l) := \{y_i : 1\leq i< k\ \text{and}\ x_i\not=x_k\ \text{and}\ \min(y_k,y_l) <y_i < \max(y_k,y_l)\}.
\end{align*}
Clearly, $top(k,l)$, $mid(k,l)$, $bot(k,l)$, $eqMin(k,l)$,$eqMax(k,l)$ can be calculated using the empirical cumulative distribution in $O(1)$ as described in the previous section (e.g. $mid(k,l)= A[rank(x_k)-1,rank(max(y_k,y_l))-1]-A[rank(x_k)-1,rank(min(y_k,y_l))]$ and $eqMin(k,l)=A[rank(x_k)-1,rank(min(y_k,y_l))]-A[rank(x_k)-1,rank(min(y_k,y_l))-1]$).
We will now show how to calculate the last element in equation \ref{eq:N_d}. This will be done with a procedure similar to the one used to calculate the empirical cumulative distribution in the previous section. Our first goal will be to calculate in $O(n^2)$
\begin{align*}
A(r,s)=\sum_{y\ s.t. \ rank(y)<=s}{|\{i: rank(x_i)<r \ \text{and}\ y_i=y\}| \choose 2}
\end{align*}
We initialize $A(0,s)=A(r,0)=0$. We further set $B(r,s)=|\{i:rank(x_i)=r \text{ and } rank(y_i)=s\}|$
We now compute the cumulative row sum $R(r,s)=R(r-1,s)+B(r,s)$ and then we compute row by row $A(r,s)=A(r,s-1)+{R(r,s) \choose 2}$.
Once we have $A(r,s)$ we can easily calculate the last element in \ref{eq:N_d} in O(1).
\begin{align*}
\sum_{y\in unique(k,l)}{|\{1\leq i< k: x_k\not= x_i\ \text{and}\ y_i=y\}| \choose 2}=A[rank(x_k)-1,rank(max(y_k,y_l))-1]- \\
A[rank(x_k)-1,rank(min(y_k,y_l))]
\end{align*}
Thus completing the computation in $O(n^2)$ as required.
\section{Conclusion}
We have shown how to calculate the Bergsma Dassios association measure in $O(n^2)$. However, the question of the power of this method remains open. It would be interesting to compare its power to the power of methods based on finer partitions as in \cite{Hellera}.
|
\section{Introduction}
New imaging capabilities are driving revolutionary advances in science and technology, making it possible to image biological systems in three dimensions with stunning detail. For materials systems that cannot be labeled using fluorescent markers, and that are opaque to visible light, shorter wavelength light has advantages. As nanosystems relevant to the semiconductor industry become increasingly complex, there is a critical need for microscopy techniques that can distinguish between different materials. This capability is realized using spectromicroscopy techniques in the extreme ultraviolet (EUV) and X-ray range, which take advantage of element-specific absorption contrast. However, established techniques such as near edge X-ray absorption fine structure (NEXAFS) \cite{Ade1993}, X-ray absorption near edge structure (XANES) \cite{Ade1990}, and X-ray fluorescence \cite{Horowitz1972} spectromicroscopies can only record images at one wavelength at a time: either a monochromator or spectrometer must be scanned in order to collect a full series of images. In contrast, ptychographical information multiplexing \cite{Batey2014} (PIM), which is an extension to ptychographical coherent diffractive imaging (CDI), enables the collection of images at multiple wavelengths simultaneously. When combined with EUV and X-ray wavelengths, this technique can be used to produce images with elemental contrast without the need for wavelength scanning.
Ptychographic CDI \cite{Faulkner2004, Rodenburg2007, Thibault2008, Seaberg2014, Zhang2015} has achieved great success due to its high fidelity and robustness in situations where other CDI methods fail. In a ptychographical microscope, the object is scanned area-by-area to record a diffraction pattern at each scan position, ensuring overlap between adjacent scan positions. The diffraction patterns and scan locations are then fed into an iterative algorithm, to retrieve the phase of the diffracted fields. By back-propagating the diffraction field to the sample position, complex-valued images of the object are obtained with quantitative amplitude and phase information. Scanning with overlap is a simple, yet elegant way of introducing redundant information into the diffraction-based measurements, distinguishing ptychography from other computational imaging techniques.
In addition to removing experimental restrictions required for other types of CDI, such as sample isolation, recent breakthroughs demonstrate that the information redundancy in ptychography also enables the reconstruction of state mixtures, i.e. the decomposition of mutually incoherent modes \cite{Batey2014, Thibault2013, Karl2015}. Thibault and Menzel \cite{Thibault2013} demonstrated the reconstruction of five dominant spatial modes in a partially coherent X-ray beam used for ptychography illumination. Shortly thereafter, Batey et al. \cite{Batey2014} simultaneously illuminated a sample with three separate single-color visible lasers with blue, green and red wavelengths, and successfully recovered the sample’s response for each of these three colors using the PIM technique.
CDI requires light sources with a high degree of spatial coherence. To address this challenge at short wavelengths, coherent light sources based on tabletop high harmonic generation (HHG) and large-scale free electron lasers (FELs) are under rapid development. High harmonic generation \cite{Rundquist1998, Bartels2000, Chen2010, Popmintchev2012, Popmintchev2015, Fan2015} is an extreme nonlinear process that coherently upconverts infrared, visible, or UV light to extreme ultraviolet (EUV) and X-ray wavelengths. HHG is unique as a light source because the generation conditions can be adjusted so that the light emerges either as a coherent supercontinuum (corresponding to an isolated attosecond burst in time), or as a comb of harmonics, with periodicity both in space and time. Under the right conditions, the harmonics are spectrally narrow ($\Delta\lambda/\lambda < 1 \%$) \cite{Bartels2000, Bartels2002OL} and also tunable by changing the driving laser wavelength or chirp. Since its demonstration in 2007, CDI using HHG sources has become a successful imaging modality, enabling high resolution image reconstruction both in transmission \cite{Sandberg2007, Seaberg2011, Zhang2013} and more recently in reflection \cite{Seaberg2014, Zhang2015, Gardner2012}. Despite these successes, HHG CDI contains untapped potential because the spectral structure of HHG has been underutilized. Indeed, recent work used broadband HHG beams as a way to make more efficient use of available photons \cite{Chen2009, Abbey2011}, by assuming that the object looks identical at all constituent wavelengths \cite{Chen2009, Abbey2011, Witte2014}.
In this work, we extend PIM \cite{Batey2014} to utilize multiple high harmonics as the illumination source, in an important spectral range offering element-specific contrast due to adjacent absorption edges. By relaxing the previous assumption of a uniform spectral response, we retrieve a wealth of information about the sample, including the spectrally-dependent amplitude and phase that encode the sample composition and topography. To achieve this, we illuminate the sample with several harmonic orders simultaneously, and employ the PIM algorithm to reconstruct independent images at each wavelength. Thus, in addition to high-contrast, high-spatial-resolution imaging systems, we can now achieve intrinsic element-specific contrast at multiple wavelengths simultaneously. Information redundancy in ptychography not only allows the sample, but also the illumination to be reconstructed. Here we aslo demonstrate in situ imaging of a multicolor EUV beam. Finally, we note that the combination of a comb of coherent harmonics and the PIM algorithm is the most efficient use of HHG EUV radiation for imaging to date because there is no energy loss from any multilayer mirrors or monochromatizing optics.
\section{Experiment}
The experimental setup is shown in Fig. 1. Phase matched high harmonics near 30 nm were generated by focusing ultrashort pulses from a Ti:Sapphire amplifier (KMLabs Dragon, 790 nm central wavelength, 5 kHz repetition rate, 1.5 mJ pulse energy, 22 fs pulse width) into a hollow waveguide (150 {\micro\meter} inner diameter, 5 cm long) filled with argon at 35 Torr. In contrast to previous experiments, no EUV multilayer mirrors were used to select a single harmonic. Instead, an ellipsoidal mirror refocused the all the harmonics onto the sample, with an estimated beam diameter of 10 {\micro\meter} based on a knife-edge measurement.
Due to geometrical constraints of our imaging chamber, we added a pair of gold mirrors placed at an angle of 45{\degree} to steer the beam.
The sample (titanium features patterned on silicon \cite{Seaberg2014}) was placed at a 50.5{\degree} angle of incidence. Diffraction from the sample, which is an incoherent superposition of scattered light from all four harmonics, was measured by an EUV-sensitive CCD (Andor) detector placed 55 mm away from the sample, and was positioned normal to the specular reflection of the beam.
The ptychographical scan consisted of a 28 $\times$ 11 rectangular grid (with added random offsets up to 20\% of the step size to prevent periodic artifacts in the reconstruction).
We chose a scanning step size of 1 {\micro\meter} to ensure enough information redundancy for the ptychographical dataset. The equivalent numerical aperture (NA) of the collected diffraction is 0.086, providing a half-pitch resolution near 170 nm.
The exposure time was set to 1.5 seconds to avoid saturation of the CCD. We accumulated three exposures at each position to increase the signal-to-noise ratio, resulting in a total exposure time of 23 minutes. In a chamber designed specifically for this experiment, the gold steering mirrors, which have a combined reflectivity of 2\%, could be removed and the required exposure time would then be reduced to 30 seconds using the HHG source described here.
\begin{figure}
\includegraphics[width=\textwidth]{hyperspectral_pty_experiment_cmpr.eps}
\caption{Hyperspectral imaging by combining multiple EUV harmonics and ptychographical information multiplexing. An EUV HHG beam consisting of four harmonic orders is focused onto the sample by a grazing-incidence ellipsoidal mirror. The sample is scanned transversely to enable ptychographical imaging. At each scan position a pixel-array detector records the diffraction pattern, which is an incoherent superposition of the four different wavelengths. The PIM algorithm can decompose the diffraction pattern into its coherent components, to obtain the spectral response of the sample at each wavelength simultaneously.}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{spectrum_measurement.eps}
\caption{Spectrum of the multicolor EUV HHG source. (a) Diffraction intensity measured on the detector. (b) Diffraction amplitude obtained by performing tilted plane correction on (a). The axis is the spatial frequency normalized by $1/\lambda$, or effective NA (scale bar 0.02). (c) Spectral intensity. Dashed-line: line cut along the magenta line in (b). Solid line: higher-resolution estimate of the spectrum by deconvolving the modulus square of the remapped diffraction in (b).}
\end{figure}
To precisely determine the wavelengths of the harmonics, we placed a two-dimensional (2D), $\Lambda = 300$ nm period nano-pillar grating adjacent to the sample in the same plane. This grating acted as a low-resolution spectrometer. Diffraction from this grating illuminated by the EUV harmonic comb is shown in Fig.~2(a). We can clearly see the first-order diffraction peaks corresponding to spatial frequency $1/\Lambda$ for different harmonics. We remapped \cite{Gardner2012} the diffraction pattern to a uniform grid of spatial frequencies normalized by $1/\lambda$, with the result shown in Fig.~2(b). The spectral intensity, calculated as the modulus square of the diffraction amplitude along the magenta line in Fig.~2(b), is shown as a dashed line in Fig.~2(c), where the wavelength axis is calculated as the product of and the normalized spatial frequency. The peaks are located at 29.1, 31.5, 34.6, and 38.1 nm, and correspond to harmonic orders 27, 25, 23, and 21 of the laser driving wavelength centered at 790 nm. The spectrum obtained in this way are blurred due to the fact that the HHG beam was not focused onto the detector (see dashed line in Fig.~2(c)). To obtain spectral intensities more representative of the actual HHG spectrum, we performed a 2D deconvolution of the remapped diffraction intensity using the Lucy-Richardson algorithm, with the zero-order diffraction taken to be the point spread function. The results of this analysis give an upper bound to the harmonic widths of $\Delta \lambda/\lambda \le 2\%$ (see solid line in Fig.~2(c)). We ignore the weak harmonic at 41 nm. The four-harmonic spectrum spans a width $\Delta \lambda/\lambda$ of 27\%.
\section{Reconstruction}
We use a modified version of the PIM algorithm \cite{Batey2014} to reconstruct the response of the sample to the different HHG wavelengths present in the illumination. The modification is that we do not constrain the spectral weights during iterations, and let the relative scale of the probe and the sample float freely, while their product is invariant. This way, the algorithm will also work for applications where spectral weight information is missing, with the understanding that for each color, there is an undetermined relative scale factor between the reconstructed probe and the reconstructed object. The initial guess for the probe beams is obtained from a knife-edge focus measurement \cite{Zhang2015}. The entire reconstruction process consisted of $\approx$ 1300 iterations. During the final $\approx$ 100 iterations, we implemented the position refinement algorithm \cite{Zhang2013F} for each color.
\begin{figure}
\includegraphics[width=\textwidth]{decompose_multicolor_diffraction.eps}
\caption{Decomposition of a multicolor diffraction pattern. (a) Measured diffraction intensity after remapping around the position marked by magenta crosshair in the inset. (b-e) The coherent components of (a) at 29.1 nm, 31.5 nm, 34.6 nm, and 38.1 nm respectively. All the diffraction intensities in (a-e) are shown to the quarter power, and share the scale bar with (a), which has a length of 0.02 NA. (f) Comparison of the intensity profile of the incoherent superposition in (a) and its components in (b-e) along the dashed-line.}
\end{figure}
An illustration of how multicolor diffraction patterns are decomposed using PIM is shown in Fig.~3. We select the scan position centered around the crosshair mark in the inset of Fig.~3(a), and display the measured diffraction intensity (after remapping \cite{Gardner2012}). Figure~3(b-e) show the four quasi-monochromatic components, as determined by the PIM algorithm, that sum incoherently to form the pattern shown in Fig.~3(a). Figure~3(f) shows the intensity profiles of Fig.~3(a-e) along the dashed line. The lower contrast of the diffraction fringes in the mixed diffraction pattern in Fig.~3(a) relative to that of its coherent components is clear. The different colors have different scattering angles for the same spatial frequencies and different spatial distribution of illumination fields, causing blurring in the mixed diffraction pattern.
\begin{figure}
\includegraphics[width=\textwidth]{reconstruction_hyperspectral.eps}
\caption{Reconstructed sample reflection coefficients for the four harmonics in the illumination. The first row (a-d) shows the amplitudes and the second row (e-h) shows the phases. The third row (i-l) shows the normalized amplitude (left) and the phase (right) of the probe for each color. Each color has the amplitude normalized so that its maximum is one. (m) For comparison, a scanning electron microscopy image of the sample is shown. (n) The combination of probes at four different wavelengths to show the spatial dispersion of the HHG beam. Scale bar in (a) and (n), 10 {\micro\meter}. (b-m) share the same scale bar as in (a). The white box in (a) marks a region of interest for later quantitative analysis.}
\end{figure}
We reconstructed the complex amplitude of the reflection from the sample, with the moduli shown in Fig.~4(a-d) and phases shown in Fig.~4(e-h). As a reference, an SEM image of the sample is shown in Fig.~4(m). Excluding the reconstruction for the 21st harmonic, the main features of the sample are reconstructed well. There are several types of artifacts present in the PIM reconstructions, such as double-lining near edges and noise on the silicon substrate. Compared with our previous reconstruction with single color illumination \cite{Zhang2015}, the multicolor reconstruction has degraded quality. The probable reason is that while the signal-to-noise ratio for the sum of all four colors is about the same as for the single-color case, after decomposition into four colors, each color has much lower signal-to-noise ratio leading to lower image quality. This is mainly a technical limitation due to the limited dynamic range of the detector used in this experiment, and should not be considered to be a fundamental limitation to the technique. The image reconstruction quality is especially low for the 21st harmonic, which not only shows a low spectral weight in the measured spectrum in Fig.~2(c) but also has a lower reflectivity from titanium: the reflectivity at 38.1 nm is only 11\% of that at 29.1 nm \cite{Henke1993}. Ignoring other weak harmonics present in the spectrum may also lead to reconstruction errors.
Reconstructed moduli and phases of the HHG beam at four different wavelengths are shown in Fig.~4(i-l), with Fig.~4(n) showing the different color beams on the same plane to provide a direct view of the entire HHG beam. The outline of each beam in Fig.~4(n) corresponds to the $1/e^2$ intensity level. In terms of the HHG probe beam reconstruction, we can clearly see astigmatism in Fig.~4(i-l), with the wavefront converging horizontally and diverging vertically. This extracted wavefront curvature agrees with our observations during the experiment: using a knife-edge measurement we found that the x-focus is after the y-focus by roughly 400 {\micro\meter} along the beam axis, and we positioned the sample at the midpoint of the x- and y-foci.
\begin{figure}
\includegraphics[width=\textwidth]{compare_modulus_phase_RKedit_parenthesis.eps}
\caption{Spectral contrast analysis. For the area enclosed by the white box in Fig.~4(a), a histogram of the modulus and the phase of the reflection from the sample is shown in (a) and (b) respectively. For convenience, in (a) the reflection coefficient moduli $|r|$ are normalized such that $|r|=1$ for silicon for all wavelengths; while in (b) the phases for silicon are set to zero for all wavelengths. (c) and (d) show the comparison of modulus $|r_{\mathrm Ti} |/|r_{\mathrm Si} |$ and phase difference $\phi_{\mathrm Ti}-\phi_{\mathrm Si}$ from theory, the single color experiment in previous work, and the multicolor experiment in this work.}
\end{figure}
\section{Spectral contrast analysis}
To verify the spectral contrast obtained in the PIM experiments, we performed a comparison against both the previous single-color work \cite{Seaberg2014} as well as theoretical calculations. For the region marked by a white rectangle in Fig.~4(a), we performed histogram analysis of the modulus ratio of the titanium region versus that of the silicon region $|r_{\mathrm Ti} |/|r_{\mathrm Si}|$, and also the phase difference between the two regions, $\phi_{\mathrm Ti}-\phi_{\mathrm Si}$, as shown in Fig.~5(a) and Fig.~5(b). To obtain the theoretical values of these two quantities, we used the results from a previous characterization of the same sample \cite{Seaberg2014}: the thickness of the oxidation layers on top of the titanium and the silicon regions are assumed to be 2.92 nm and 3.01 nm thick, respectively. From an atomic force microscopy measurement, we characterized the roughness of the titanium region and the silicon region to be 0.91 nm and 0.23 nm respectively, and the height difference $h_{\mathrm Ti} -h_{\mathrm Si}$ to be 32.7 nm. We found evidence of up to 1 nm of carbon buildup in samples maintained in the same environment as the one used in this work, using Auger electron spectroscopy. Here we assume a thickness between 0 and 1 nm of the carbon contamination, which is a source of uncertainty in the theoretical reflectivities of the sample surfaces. With this model of the sample, and tabulated values of indices of refractions and reflectivities \cite{Henke1993}, we calculated $|r_{\mathrm Ti} |/|r_{\mathrm Si} |$ and $\phi_{\mathrm Ti}-\phi_{\mathrm Si}$ (the total phase results from both the complex Fresnel reflection coefficient and the height difference \cite{Seaberg2014, Zhang2015}) as a function of wavelength. The comparison between these calculations and the results from the previous single color experiment \cite{Seaberg2014} along with the current multicolor experiment is shown in Fig.~5(c) and Fig.~5(d). Both of these comparisons demonstrate good agreement between the multicolor and single-color experimental results and theoretical simulations.
\section{Discussions and conclusions}
As illustrated above, the combination of high harmonic combs and ptychographic information multiplexed diffraction imaging allows the amplitude and phase response of a sample to be recovered at multiple wavelengths simultaneously. Each of the four illuminating probes, one for each phase-matched harmonic, is reconstructed separately, over a spectral range corresponding to $\Delta\lambda/\lambda \approx 27\%$. No wavelength-scanning or separation hardware was used, making the experimental setup simpler than other techniques while at the same time enabling more efficient use of available photons. The combination of a comb of coherent harmonics with the PIM algorithm is the most efficient use of HHG EUV radiation for imaging to date because there is no energy loss from any multilayer mirrors or monochromatizing optics.
The limitations to this new spectromicroscopy technique are not currently well understood, including how many colors, $N_{\mathrm c}$, can be reconstructed simultaneously, or what spectral resolving power, $\lambda/\Delta\lambda$, can be achieved. The limit for $N_{\mathrm c}$ can be estimated by simply considering the ratio between the number of knowns and unknowns in the phase retrieval problem. If a point on the sample is illuminated with $N_{\mathrm p}$ overlapping ptychography scan positions, i.e. it is independently measured $N_{\mathrm p}$ times, and the diffraction pattern is sampled at an oversampling \cite{Bates1982} ratio $\rho$, then there is $N_{\mathrm p} \rho$-fold information redundancy for that point. Beyond $\rho=2$, no further information is gained, so for $\rho>2$ the information redundancy is simply $2N_{\mathrm p}$. Each color needs two-fold redundancy for phase retrieval, so an upper bound for $N_{\mathrm c}$ is $N_{\mathrm p} \rho/2$ for $\rho\le2$, and $N_\mathrm{p}$ for $\rho>2$. It is reasonable to believe that there is a point beyond which increasing $N_{\mathrm p}$ (increasing overlap between scan positions) fails to result in increased information redundancy i.e. adjacent scans are no longer independent. Regarding spectral resolution, the detection of scattered light at high numerical aperture can help separate adjacent colors via diffraction. In object space, this translates to the fact that in order to discriminate between adjacent colors, the respective fields of view must differ in image size (number of pixels).
With this in mind the spectral resolution is limited to approximately $\lambda/\Delta\lambda \ge 2 \mathrm{NA} L/\lambda$, where $L$ is the maximum distance between scan positions, and NA is the numerical aperture. Thus an increase in either the NA or the maximum distance between scan positions can result in increased spectral resolution. For the experiment described here, the estimated upper bound for $N_c$ is $\approx$100, much greater than the 4 colors reconstructed here. The spectral resolving power is estimated to be 80, which is higher than the necessary resolving power of 15 necessary to distinguish adjacent harmonics.
Looking forward, the spatial resolution can easily reach below 50 nm by placing the detector closer to the sample for a higher numerical aperture \cite{Zhang2015}. Various element-specific absorption edges in the EUV/X-ray spectrum range at higher photon energies than studied here provide opportunities for high-contrast spectral imaging. Broadband and narrowband shorter wavelength HHG sources are being developed that are bright (i.e. phase matched) and that can operate in the water window range \cite{Chen2010, Popmintchev2015} for biological imaging, reaching photon energies up to 1.6 keV \cite{Popmintchev2012}. It is also possible to extend this technique to multicolor sources other than HHG, such as synchrotrons and free electron lasers. The HHG source used in this work has a favorable spectral structure with discrete spectral lines, each of which has a narrow bandwidth. The question of how to address more continuous, broadband spectra with this technique will be addressed in future work. Each harmonic has a transform limited pulse duration on the order of $\approx 5$ femtoseconds (30 nm wavelength with a spectral bandwidth of $\Delta\lambda/\lambda = 1\%$ and a Gaussian profile corresponds to pulse duration of 5 femtoseconds). By making use of the ultrashort pulse duration of the HHG source in a pump-probe geometry, this technique can be used for spectral imaging of ultrafast charge, spin and phonon dynamics in functioning nanosystems \cite{Hoogeboom-Pot2015, Turgut2013}.
\section*{Acknowledgements}
This work was performed at JILA. We thank Dr. Ming-Chang Chen for his help on the ellipsoidal mirror. We also gratefully acknowledge support from the Defense Advanced Research Projects Agency (DARPA) PULSE program, the NSF Engineering Center in EUV Science and Technology, the Semiconductor Research Corporation (2013-OJ-2443), the National Science Foundation COSI IGERT 0801680, the National Science Foundation Graduate Research Fellowship Program, Ford Foundation fellowship program, and the Katherine Burr Blodgett fellowship program. The current address of B.Z. is KLA-Tencor Corporation, One Technology Drive, Milpitas, CA 95035, USA. The current address of M.H.S. is SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA.
\end{document}
|
\section{Introduction}\label{sec1}
\beginsec
Rate controlled queueing systems commonly arise from applications in communication systems, see e.g.~\cite{Fendick1994,Kushner2001,Ata2005,Budhiraja2011} and references therein.
Recently, they have also been considered in modeling limit order books, see e.g. \cite{MR2838580,MR3274927,MR3358591,Bowe,Blair,Lachapelle2015}.
A common approach to the study of such rate control problems when the system is in heavy traffic is through diffusion approximations. In a problem setting where there is interaction between servers/queues in that the rates or costs
associated with a particular queue and server can depend on the states of the other queues, this approach leads to a stochastic control problem for $n$-dimensional reflected diffusions, where $n$ is the number of queues in the system. When $n$ is large such control problems are computationally intractable and in general this `curse of dimensionality' is unavoidable. However, when there are certain symmetries present and the interaction between
queues is weak, in that each queue has ${\cal O}(1/n)$ affect on any other queue in the system, a natural approach is to consider, in addition to heavy traffic, another asymptotic regime where the number of queues $n$ approaches $\infty$ as well. Such model settings arise in many applications, e.g., cloud computing, live streaming, limit order books, customer service systems, etc. In many of these contexts the servers are strategic, for example, in customer service networks, servers respond to workload incentives (see \cite{Gopalakrishnan2014}), and in the context of limit order books buyers and sellers place their orders in a strategic manner and interact weakly through their impact on the price distribution.
\vspace{\baselineskip}
\noindent We study one of the simplest forms of queuing networks, namely a collection of $n$ single server queues.
The system is assumed to be \emph{critically loaded}. For this, we fix an arbitrary sequence of scaling parameters $\{e_n\}_n$ that satisfy $e_n\to\iy$ when $n\to\iy$. We assume that, for each queue, the traffic intensity, which is the ratio between the rate of arrivals to the rate of service is $1-{\cal O}(1/\sqrt{e_n})$ (see \cite{Chen2001}). The arrival and service rates for each queue are ${\cal O}(e_n)$.
In absence of control and state dependence of rates, the analysis simply reduces to that of a single $M/M/1$ queue which in the heavy traffic can be approximated by a one dimensional reflected Brownian motion.
In the setting we consider, every server can exercise control on the arrival and service rates associated with its own queue. This control which is ${\cal O}(\sqrt{e_n})$ is of lower order compared to the overall rate but it can have significant impact on performance in the asymptotic regime we consider.
In the heavy traffic regime with a fixed number of servers, performance improvement using an ${\cal O}(\sqrt{e_n})$ control has been well studied (see for example \cite[Chapter 9]{Kushner2001}). In a setting (that is quite different from the one considered here) where the number of nodes/servers approach $\infty$, numerical results that show performance improvement under ${\cal O}(\sqrt{e_n})$ controls can be found in \cite{BudFri2016}.
In the model we consider, the rates can depend on the state of the individual queue, furthermore a particular queue's state is influenced by the remaining queue states through their empirical measure. The control action for each server can use information on the history of queue lengths, arrival and processing times and control actions for all the queues in the system. We consider a rate control problem where each server aims to minimize its individual cost. Although many different types of cost criteria can be considered, here for simplicity we consider a cost function over a finite time horizon. This cost function may depend
on the individual queue lengths, the control action, and the empirical measure of all the queue states. A natural way to formulate optimality for such $n$-player games is through the notion of a (near) Nash equilibrium. Computing Nash equilibria, even if one considers the simplified heavy traffic approximation in terms of reflected diffusions, is computationally a very challenging problem when $n$ is large. The goal of this work is to analyze an asymptotic formulation where simultaneously each queue approaches heavy traffic {\em and} the number of queues approaches $\infty$. There is extensive literature on heavy traffic limits of rate controlled queuing systems with a fixed number of queues \cite{Kushner2001,Ata2005,Budhiraja2011}, however to the best of our knowledge this is the first work to study the asymptotics for controlled queues where in addition to heavy traffic, the number of queues approach infinity as well. We will show that the asymptotics in this regime is governed by a Lasry-Lions type mean field game (MFG) for reflected diffusions. We will use the solution of the MFG to construct an asymptotic Nash equilibrium for the $n$-player queueing system. This equilibrium has a simple and appealing decentralized structure: each server bases his decision only upon the length of his own queue and a deterministic measure valued function obtained from the solution of the MFG. We also prove that the value
of the above near Nash equilibrium converges to the value function of the MFG as $n\to \infty$.
In general closed form solutions for MFG of the form that arise here are not available and thus one needs numerical approximations. In \cite{BBC2017} we study one such procedure that uses the Markov chain approximation method (\cite{Kushner1992}) and establish convergence of the scheme over a small time interval. We refer the reader to \cite{Achdou2010,Lachapelle2010, Gueant2012,Achdou2013, Carlini2014, Carlini2015,Benamou2015, Achdou2016, Chassagneux2017} for some recent results on numerical methods for mean field games.
\vspace{\baselineskip}
\noindent The theory of mean field games was initiated a decade ago in the seminal work of Lasry and Lions \cite{Lasry2006,Lasry2006b,Lasry2007}, and Huang, Malham{\'e}, and Caines \cite{Huang2006,Huang2007}. In recent years there has been a growing interest in this field. For recent theoretical developments and applications of this theory see \cite{Cardaliaguet2013,Gueant2013,Gomes2013,Carmona2013,Lacker2015,Carmona2015,Lacker2015general,Fischer2016} and references therein. Mean field approximations for weakly interacting stochastic particles have a long history starting from the works of Boltzmann, McKean, Kac and others (see \cite{sznit} and references therein). Even in the context of queuing systems and communication networks, there have been many works \cite{gibbens1990,Baccelli1992,Hunt1994,Vvedenskaya1997,Borovkov1998,graham2000, Benaim2008,Bobbio2008,graham2009,antunes2008, BudFri2016}.
Another related branch is agent based models with mean-field interaction (but without strategic agents), see e.g.~\cite{MR1856682,BayraktarHorst1,BayraktarHorst2}. MFGs have also been used by queuing theorists as a tool in recent years see e.g.~\cite{Manjrekar2014,Wiecek2015,Li2015,Lachapelle2015}. In contrast to these works we consider an MFG associated with queuing systems in heavy traffic.
\vspace{\baselineskip}
\noindent We now make some comments on proof techniques. Roughly speaking, solution of an MFG considered here can be viewed as the solution of a fixed point problem on the space of probability measures on certain path spaces (see Section \ref{sec3}).
In order to characterize such solutions, in a setting where each agent's state evolution is described through a diffusion
process, Lasry and Lions \cite{Lasry2006,Lasry2006b,Lasry2007} studied wellposedness of two coupled nonlinear partial differential equations; one is an equation of Hamilton-Jacobi-Bellman (HJB) type while the second takes the form of a Kolmogorov forward equation. A somewhat different approach is taken in recent works of Carmona, Delarue, and Lacker \cite{Carmona2013, Carmona2015}. Using probabilistic methods, authors characterize the MFG solution as a solution to certain forward backward stochastic differential equations.
\noindent In all the above papers the $n$-player system is described through a collection of stochastic differential equations and the mean field game gives the asymptotic behavior of the system as $n\to \infty$.
In contrast, in the current work, for a fixed value of $n$ the state process is given through a collection of controlled jump-Markov processes with jump sizes that approach $0$ as $n\to \infty$.
Thus here we have two forms of asymptotic behavior, one corresponds to the large agent limit while the other corresponds to diffusion approximations of pure jump processes with vanishing jump sizes. Analyzing this simultaneous limit behavior, which is key to the proof of Theorem \ref{thm1} that identifies an asymptotic Nash equilibrium for the $n$-player game as $n\to \infty$, requires new techniques. A key ingredient
is to prove suitable tightness properties and to analyze and characterize the weak limit points of certain stochastic processes and random measures. Here we make use of a result from \cite{Kotelenez2010} (see Lemma 4.2 therein) which can be viewed as an extension of de Finetti's theorem for sequences of random measures on certain path spaces. In proving suitable tightness properties of control processes we consider a relaxed control formulation. Our assumption on the uniqueness of the minimizer (Assumption \ref{assumption1}(b)) ensures that extending the class of controls in this manner does not lower the cost. Such relaxed control formulations have been widely used in stochastic control theory, see e.g.~\cite{Fleming1976, El1987, Borkar1989, Kushner1992}. More recently, they have also been invoked in the study of mean field games \cite{Lacker2015, Lacker2015general, Fischer2016}.
Other key ingredients needed for the proof are (a) regularity properties of the optimal feedback controls for certain
stochastic control problems for reflected diffusions; (b) weak formulation of stochastic control problems for reflected diffusions. We find that the approach based on HJB characterizations of value functions is particularly well suited for our problem setting. In particular, regularity results from \cite{Lieberman1996} for quasilinear PDE with Neumann boundary conditions give us the required estimates for obtaining the desired properties of the optimal feedback controls.
Use of HJB theory in our analysis is a feature that is common to the approach initiated by Lasry and Lions, however a
point of departure is that instead of using Kolmogorov forward equations we characterize probability distributions as unique weak
solutions of suitable reflected diffusions. In this sense our approach is closer to \cite{Carmona2015} although, unlike \cite{Carmona2015}, we do not make use of forward backward stochastic differential equations in our work.
\vspace{\baselineskip}
\noindent In summary our main contributions are as follows. We
\begin{itemize} \itemsep0em
\item consider a rate control problem for large symmetric queuing systems in heavy traffic with strategic servers;
\item introduce an MFG for controlled reflected diffusions, and in Theorem \ref{thm_fixed} establish its solvability under Assumption \ref{assumption1}, and prove unique solvability assuming in addition Assumption \ref{assumptionU};
\item use the solution of a diffusion MFG to construct, under Assumptions \ref{assumption1} and \ref{assumption4}, an asymptotically optimal Nash equilibrium for the $n$ player countable state
game using techniques that combine heavy traffic analysis with large agent asymptotics (Theorem \ref{thm1}).
\end{itemize}
Thus Assumption \ref{assumption1} is a basic assumption for all our results. This assumption imposes Lipschitz continuity of the various functions in the model. Assumption \ref{assumptionU}, that is introduced for the uniqueness of the solution of the MFG, is common in the literature of MFG. It says that the drift is independent of the mean-field term and that the running cost and terminal cost satisfy a certain monotonicity property, see Remark \ref{rem_uniqueness} for a discussion of the assumption. In order to obtain an asymptotic Nash-equilibrium, we require, in addition to Assumption \ref{assumption1},
Assumption \ref{assumption4} which as the first part of Assumption \ref{assumptionU} says that the drift is free of the mean-field term and in addition requires a basic convergence property for the initial conditions. For this result we do not require the monotonicity property, in particular the uniqueness of the solution of the MFG is not needed.
\noindent The paper is organized as follows. In Section \ref{sec2} we introduce the queueing model, the scaling regime and the cost criterion. Next in Section \ref{sec3} we introduce the MFG and present our main result on its solvability (Theorem \ref{thm_fixed}). Section \ref{sec4} constructs an asymptotic Nash equilibrium for the $n$-player game and the main result of this section is Theorem \ref{thm1} which proves asymptotic optimality.
\subsection{Notation} \label{sec11}
We use the following notation.
For metric spaces ${\cal S}_1, {\cal S}_2$ denote by ${\cal C}({\cal S}_1:{\cal S}_2)$ the space of ${\cal S}_2$ valued continuous functions on ${\cal S}_1$. When ${\cal S}_2 = \mathbb{R}$, we abbreviate the notation to simply ${\cal C}({\cal S}_1)$. The space ${\cal C}([0,T]: {\cal S})$ for a Polish space ${\cal S}$ will be equipped with the
uniform topology. We will denote by ${\cal D}([0,T]: {\cal S})$ the space of $S$ valued functions that are right continuous and have left limits (RCLL) defined on $[0,T]$. This space is equipped with the usual Skorohod topology.
For $f \in {\cal D}([0,T]:\mathbb{R}^d)$ and $0\le t\le T$, $\|f\|_t \doteq \sup_{0\le s \le t}\|f(s)\|$.
In case that $d=1$, we often use $|f|_t$.
We will denote by $\text{Lip}_1({\cal S})$ the space of real Lipschitz functions on ${\cal S}$ whose Lipschitz norm is bounded by $1$, namely the class of functions $f:{\cal S} \to \mathbb{R}$ with
$$|f(x) - f(y)| \le d(x,y),\; x,y \in {\cal S}$$
where $d$ denotes the metric on ${\cal S}$.
Denote by ${\cal P}({\cal S})$ the space of probability measures on ${\cal S}$. We endow ${\cal P}({\cal S})$ with the topology of weak convergence of measures. Convergence in distribution of $S$ valued random variable $X_n$ to $X$ will be denoted as
$X_n \Rightarrow X$. For $T, L \in (0,\infty)$, the space ${\cal P}({\cal C}([0,T]:[0,L]))$ will be denoted as ${\cal P}_{T,L}$.
The Wasserstein distance of order $1$ on ${\cal P}({\cal S})$, where ${\cal S}$ is a compact metric space,
is defined as
\begin{align}\notag
W_1(\eta',\eta)=\inf\left\{\left[\int_{{\cal S}}d(x,y)d\pi(x,y)\right] : \pi\in{\cal P}({\cal S}\times {\cal S})\;\;\text{ with marginals $\eta'$ and $\eta$}\right\},
\end{align}
where $\eta,\eta'\in{\cal P}({\cal S})$.
We denote by ${\cal C}^{1,2}([0,T]\times [0,L])$ the space of
functions $f:(0,T)\times (0,L)\to \mathbb{R}$ that are continuously differentiable (resp., twice continuously differentiable) with respect to (w.r.t.) the first (resp., second) variable and are such that the function and the derivatives extend continuously to $[0,T]\times [0,L]$. For $\phi \in {\cal C}^{1,2}([0,T]\times [0,L])$, $D_t\phi, D\phi, D^2\phi$ will denote the time derivative and the first two space derivatives of $\phi$, respectively.
For $x \in {\cal S}$, $\delta_x \in {\cal P}({\cal S})$ denotes the
Dirac measure at $x$.
\section{The $n$-server queuing control problem}\label{sec2}\beginsec
We consider a symmetric $n$-server stochastic queueing system. Each server $i \in \{1, \dots, n\}$
is associated with a queue with a finite capacity and is able to control the rates of the service and arrivals of jobs to its queue.
The rates can also depend on the state of the queue and on the empirical measure of the states of all the $n$-queues. We will consider a regime where the arrival and processing rates are approximately the same and are of the same order as the number of queues in the system. We now describe our precise scaling regime and introduce the various processes that determine the evolution of the state of the system.
\subsection{Diffusion scaling}\label{sec2a}
Fix $T,L>0$ and $n\in\mathbb{N}$. Here $T$ denotes the terminal time of our finite time horizon and $[0,L]$ will be the state space of the scaled queue length process.
Let $U$ be a compact subset of $\mathbb{R}$. We will denote by $\lambda^{n,i}$ and $\mu^{n,i}$ the (controlled) arrival and service rates associated with queue $i$.
Fix an arbitrary sequence of scaling parameters $\{e_n\}_n$ that satisfy $e_n\to\iy$ when $n\to\iy$. The rates will be ${\cal O}(\sqrt{e_n})$ perturbations of certain nominal (uncontrolled) ${\cal O}(e_n)$ overall rates.
More precisely, we assume that there exist $\hat\lambda,\hat\mu>0$, and bounded and measurable functions $\lambda,\mu: [0,T]\times{\cal P}([0,L])\times[0,L]\times U \to \mathbb{R}$ such that
\begin{align}\nota
\lambda^{n,i}(t)&=\hat\lambda e_n+\lambda(t,\tilde\nu^{n}(t), \tilde{Q}^{n,i}(t), \alpha^{n,i}(t))\sqrt{e_n}+o(\sqrt{e_n}),\\
\mu^{n,i}(t)&=\hat\mu e_n+\mu(t,\tilde\nu^{n}(t),\tilde{Q}^{n,i}(t), \alpha^{n,i}(t))\sqrt{e_n}+o(\sqrt{e_n}), \label{2}
\end{align}
where $\tilde{Q}^{n,i}(t) = \tfrac{1}{\sqrt{e_n}} Q^{n,i}(t)$, $Q^{n,i}(t)$ is the size of the $i$-th queue at time $t$,
$$\tilde\nu^{n}(t)\doteq\frac{1}{n}\sum_{i=1}^n\delta_{\tilde{Q}^{n,i}(t)}$$
is the empirical distribution of the scaled queue lengths in the $n$-th system at time $t$, and $\alpha^{n,i}(t)$ is the control that server $i$ exercises at time $t$.
The term $o(\sqrt{e_n})$ represents an expression of the form $r_n(t,\tilde\nu^{n}(t), \tilde{Q}^{n,i}(t), \alpha^{n,i}(t))$
where $r_n:[0,T]\times{\cal P}([0,L])\times[0,L]\times U \to \mathbb{R}$ are measurable functions such that
$r_n/\sqrt{e_n}$ converges uniformly to $0$ as $n\to \infty$.
Additional conditions on $\lambda$ and $\mu$ will be introduced in later sections.
Each server in the $n$-th system has a finite buffer of size
$L^n=\sqrt{e_n}L$ and arriving jobs to a full buffer are lost. We assume that the system is in heavy traffic, namely
$$\hat\lambda=\hat\mu.$$
We now give a precise description of controlled stochastic processes of interest.
For each fixed $n$, let $(\Omega',{\cal F}',{\mathbb P}')$ be a probability space on which are given unit rate independent Poisson processes $N^{n,i}$ and $M^{n,i}$, $i = 1,\ldots,n$. Roughly speaking, $N^{n,i}$ will correspond to the stream of jobs entering the $i$-the queue and $M^{n,i}$ to the jobs that leave the system after service. The evolution of the $i$-th controlled queue is given as follows.
\begin{align}\label{3}
Q^{n,i}(t)=Q^{n,i}(0)+A^{n,i}(t)-D^{n,i}(t), \quad i = 1,\ldots,n,\quad t\in[0,T],
\end{align}
where $Q^{n,i}(0)\ge 0$ is the initial size of the $i$-th queue,
\begin{equation}
\label{eq:eq928}A^{n,i}(t)=N^{n,i}\left(\int_0^t 1_{\{Q^{n,i}(s)<L\}}\lambda^{n,i}(s)ds\right),\; D^{n,i}(t)=M^{n,i}\left(\int_0^t 1_{\{Q^{n,i}(s)>0\}}\mu^{n,i}(s)ds\right)\end{equation}
are the arrival and departure processes respectively, where $\lambda^{n,i}$ and $\mu^{n,i}$ are defined as in \eqref{2} in terms of control processes $\{\alpha^{n,i}\}$.
We will assume that $\{Q^{n,i}(0)\}_{i=1}^n$ are exchangeable for all $n$. We require that the control processes are suitably non-anticipative. Specifically, we assume that for some filtration $\{{\cal F}_t\}$ on $(\Omega',{\cal F}',{\mathbb P}')$, $\alpha^{n,i}$ is $\{{\cal F}_t\}$-progressively measurable, $Q^{n,i}$, $n\ge i\ge 1$ are $\{{\cal F}_t\}$-adapted, and
\begin{align}\nota
\tilde A^{n,i}(t)\doteq\frac{A^{n,i}(t)-\int_0^t 1_{\{Q^{n,i}(s)<L\}}\lambda^{n,i}(s)ds}{\sqrt{e_n}},\;
\tilde D^{n,i}(t)\doteq\frac{D^{n,i}(t)-\int_0^t 1_{\{Q^{n,i}(s)>0\}}\mu^{n,i}(s)ds}{\sqrt{e_n}}
\end{align}
are $\{{\cal F}_t\}$ martingales with quadratic variations
\begin{equation}\label{eq: eq415}
\langle\tilde A^{n,i}, \tilde A^{n,j} \rangle(t) =\delta_{ij}\frac{1}{e_n}\int_0^t 1_{\{Q^{n,i}(s)<L\}}\lambda^{n,i}(s)ds,\; \langle \tilde D^{n,i}, \tilde D^{n,j} \rangle(t) =
\delta_{ij}\frac{1}{e_n}\int_0^t 1_{\{Q^{n,i}(s)>0\}}\mu^{n,i}(s)ds,
\end{equation}
and $\langle \tilde A^{n,i}, \tilde D^{n,j} \rangle(t) = 0$, $t\in [0,T]$, $i,j = 1, \ldots, n$ where $\delta_{ij}=1$ if $i=j$ and $0$ otherwise.
The process $\alpha^{n} = \{\alpha^{n,i}\}_{i=1}^n$ will be referred to as an {\it admissible} control and we denote the collection of all such controls by ${{\cal U}}^n$. With an abuse of terminology, for $\alpha^{n} = \{\alpha^{n,i}\}_{i=1}^n$ as above, we will refer to
$\alpha^{n,i}$ as an admissible control for the $i$-th player.
From \eqref{3} we have the following evolution equation for the scaled queue length processes. For $t\in [0,T]$
\begin{align}\label{7}
\tilde Q^{n,i}(t)=\tilde Q^{n,i}(0)+\tilde A^{n,i}(t)-\tilde D^{n,i}(t)+\int_0^t\tilde b^{n,i}(s)ds+\tilde Y^{n,i}(t)-\tilde R^{n,i}(t) + o(1),
\end{align}
where
\begin{align}\label{8}
\tilde Y^{n,i}(t)\doteq\frac{\int_0^t 1_{\{\tilde Q^{n,i}(s)=0\}}\mu^{n,i}(s)ds}{\sqrt{e_n}},\;
\tilde R^{n,i}(t)\doteq\frac{\int_0^t 1_{\{\tilde Q^{n,i}(s)=L\}}\lambda^{n,i}(s)ds}{\sqrt{e_n}},
\end{align}
\begin{align}\nota
\tilde b^{n,i}(t)\doteq b(t,\tilde\nu^{n}(t),\tilde Q^{n,i}(t), \alpha^{n,i}(t)), \quad b\doteq\lambda-\mu,
\end{align}
and $o(1)$ represents a stochastic process $\eta^{n,i}$ satisfying for each $i$
$\sup_{0\le t \le T}|\eta^{n,i}(t)|\to 0$, in probability, as $n \to \infty$.
The above dynamics can equivalently be described in terms of a Skorohod map as we do below.
Let ${\cal D}_0[0,T]$ be the subset of ${\cal D}([0,T]:\mathbb{R})$ consisting of all $\psi$ such that $\psi(0) \in [0,L]$.
\begin{definition}\label{def_Skorohod}
Given $\psi\in{\cal D}_0[0,T]$ we say the triplet of functions $(\varphi,\zeta_1,\zeta_2)\in{\cal D}([0,T]: \mathbb{R}^3)$ solves the Skorohod problem for $\psi$ if the following properties are satisfied:
\noindent
(i) For every $t\in[0,T], \;\varphi(t)=\psi(t)+\zeta_1(t)-\zeta_2(t)\in [0,L]$.
\noindent
(ii) $\zeta_i$ are nonnegative and nondecreasing, $\zeta_1(0)=\zeta_2(0)=0$, and
\begin{align}\nota
\int_{[0,T]}1_{(0,L]}(\varphi(s))d\zeta_1(s)=\int_{[0,T]}1_{[0,L)}(\varphi(s))d\zeta_2(s)=0.
\end{align}
We denote by $\Gamma(\psi)=(\Gamma_1,\Gamma_2,\Gamma_3)(\psi)=(\varphi,\zeta_1,\zeta_2)$ and refer to $\Gamma$ as the Skorohod map.
\end{definition}
It is known that there is a unique solution to the Skorohod problem for every $\psi \in{\cal D}_0[0,T]$ and so the Skorohod map in Definition \ref{def_Skorohod} is well defined.
The Skorohod map has the following Lipschitz property (see \cite{Kruk2007}).
\begin{lemma}\label{lem_Skorohod}
There exists $c_S \in (0,\infty)$ such that
for all $\omega,\tilde\omega\in{\cal D}_0([0, T])$,
\begin{equation}\nota
\sum_{i=1}^3\|\mathnormal{\Gamma}_i(\omega) - \mathnormal{\Gamma}_i(\tilde\omega)\|_{T}
\le c_S\|\omega-\tilde\omega\|_T.
\end{equation}
\end{lemma}
The dynamics in \eqref{7} can be described in terms of the Skorohod map as follows
\begin{align}\label{eq_23}
(\tilde Q^{n,i},\tilde Y^{n,i},\tilde R^{n,i})(t)=\Gamma\left(\tilde Q^{n,i}(0)+\int_0^\cdot\tilde b^{n,i}(s)ds+\tilde A^{n,i}(\cdot)-\tilde D^{n,i}(\cdot) + o(1)\right)(t),\quad t\in[0,T].
\end{align}
\subsection{The control problem}
\label{sec:sec-cont}
We now introduce the cost function and the control problem studied in this work.
The total expected cost for server $i$ associated with the initial condition $\\$ $\tilde{Q}^n(0)=(\tilde{Q}^{n,1}(0),\ldots,\tilde{Q}^{n,n}(0))$ and control $\alpha^{n}=(\alpha^{n,1},\ldots,\alpha^{n,n})\in {\cal U}^n$ is given by
\begin{align}\label{12}
J^{n,i}(\tilde Q^{n}(0);\alpha^{n}) &\doteq
\mathbb{E}\Big[\int_0^T f(t,\tilde\nu^{n}(t), \tilde Q^{n,i}(t), \alpha^{n,i}(t))dt + g(\tilde\nu^{n}(T),\tilde Q^{n,i}(T))\\\notag
&\quad+\int_0^Ty(t,\tilde\nu^{n}(t))d\tilde Y^{n,i}(t)+\int_0^Tr(t,\tilde\nu^{n}(t))d\tilde R^{n,i}(t)\Big]
\end{align}
where $f:[0,T]\times {\cal P}([0,L])\times[0,L] \times U\to\mathbb{R}$ is the running cost, $g: {\cal P}(\mathbb{R})\times [0,L]\to\mathbb{R}$ is the terminal cost, and $r,y:[0,T]\times {\cal P}([0,L])\to\mathbb{R}_+$ are the costs associated with rejection
of jobs and empty buffers, respectively. Here $f,g,y$, and $r$ are bounded measurable functions that will be required to satisfy additional conditions that will be introduced in Section \ref{sec3} (see Assumption \ref{assumption1}).
Each player (server) seeks to minimize its cost. A natural formulation of optimality for such $n$-player games is given in terms of a (near) Nash equilibrium. Computing near Nash equilibria for such complex and large multi-player games is
in general intractable and thus we instead consider an asymptotic formulation of the problem.
\begin{definition}
A sequence of admissible controls $\{\tilde{\alpha}^{n,i}: 1\le i\le n\}_{n \in \mathbb{N}}$ is called an {\it asymptotic Nash equilibrium} if for every player $j$, and every sequence of admissible controls $\{\beta^n\}_{n=1}^\iy$ for that player, one has
\begin{align}\nota
\underset{n\to\iy}{\lim\;\sup\;} J^{n,j}(\tilde Q^{n}(0);\tilde \alpha^{n,1},\dots,\tilde \alpha^{n,n})\le\underset{n\to\iy}{\lim\;\inf\;} J^{n,j}(\tilde Q^{n}(0);\tilde \alpha^{n,1},\dots,\tilde\alpha^{n,j-1},\beta^n,\tilde\alpha^{n,j+1},\ldots,\tilde \alpha^{n,n}).
\end{align}
\end{definition}
Objective of this work is to show that, under conditions, an asymptotic Nash equilibrium exists which can be approximated by analyzing a related MFG.
The main results are Theorem \ref{thm_fixed} (solvability of MFG) and Theorem \ref{thm1} (asymptotic optimality)
\section{The MFG}\label{sec3} \beginsec
A natural approach for constructing asymptotic near Nash equilibria for the above $n$-player game has emerged from
the works of \cite{Cardaliaguet2013,Carmona2013,Carmona2015}. Starting point in this approach is to consider an MFG that formally corresponds to the limit of the above $n$-player games as $n\to \infty$. In this section we give a precise description of this MFG in the current context and give our main results on existence and uniqueness of solutions.
\subsection{Description of the MFG}\label{sec3a}
The basic idea in the formulation of the MFG is to approximate the scaled queue length process for a typical queue by a suitable drift-controlled reflected Brownian motion. We next introduce this controlled process.
Let $(\Omega,{\cal F},\{{\cal F}_t\},{\mathbb P})$ be a filtered probability space on which is given a one dimensional standard ${\cal F}_t$-Brownian motion $B$. We will refer to the collection $(\Omega,{\cal F},\{{\cal F}_t\},{\mathbb P}, B)$
as a system and denote it by $\Xi$.
Given $x \in [0,L]$, $t \in [0,T]$, and $\nu \in {\cal P}_{T,L}$, we denote by ${\cal A}(\Xi, t,x,\nu)$ the collection of
all pairs $(\alpha, Z)$ where $\alpha = \{\alpha(s)\}_{0\le s \le T-t}$ is a $U$-valued
${\cal F}_s$-progressively measurable process, $Z = \{Z(s)\}_{0\le s \le T-t}$ is a $[0,L]\times \mathbb{R}_+\times \mathbb{R}_+$ valued ${\cal F}_s$-adapted continuous process such that, $Z=(X,Y,R)$ and
\begin{align}\label{13}
Z(s) = (X,Y,R)(s)=\Gamma\left(x+\int_0^\cdot\bar b(u)du+\sigma B(\cdot)\right)(s),\quad s\in[0,T-t],
\end{align}
where
\begin{align}\nota
\bar b(u)\doteq b(t+u,\nu(t+u),X(u), \alpha(u)),\; u \in [0, T-t],
\end{align}
$\nu(s)$ is the marginal of $\nu$ at time instant $s$ and $\sigma=\sqrt{2\hat\lambda}$.
We now introduce the cost function in the MFG. Given $\nu \in {\cal P}_{T,L}$, $x \in [0,L]$, $t \in [0,T]$, and a system
$\Xi$ as above, let $(\alpha,Z) \in {\cal A}(\Xi, t,x,\nu)$. Define
\begin{align}\notag
J_\nu(t,x,\alpha,Z)&\doteq \mathbb{E}\Big[\int_0^{T-t} f(s+t,\nu(s+t),X(s),\alpha(s) )ds + g(\nu(T),X(T-t))\\
&\quad+\int_0^{T-t}y(s+t,\nu(s+t))dY(s)+\int_0^{T-t}r(s+t,\nu(s+t))dR(s)\Big].\label{eq:eq234}
\end{align}
We define the value function associated with the above cost as:
\begin{align}\label{15}
V_\nu(t,x)=\inf_{\Xi}\inf_{(\alpha,Z)\in {\cal A}(\Xi,t,x,\nu)}J_\nu(t,x,\alpha,Z).
\end{align}
\begin{definition}
\label{def:mfg729}
A {\bf solution to the MFG} with initial condition $x \in [0,L]$ is defined to be a $\nu \in {\cal P}_{T,L}$ such that
there exist a system $\Xi$ and an $(\alpha,Z) \in {\cal A}(\Xi,0,x,\nu)$ such that $Z=(X,Y,R)$
satisfies ${\mathbb P} \circ X^{-1} = \nu$ and
\begin{equation}\label{eq:eq716}
V_\nu(0,x) = J_{\nu}(0, x, \alpha,Z).
\end{equation}
If there exists a unique such $\nu$, we refer to $V_\nu(0,x)$ as the {\bf value of the MFG} with initial condition
$x$.
\end{definition}
To find a solution of the MFG one usually follows the following iterative procedure:
\begin{itemize}
\item[(i)] For a fixed $\nu \in {\cal P}_{T,L}$ solve the stochastic control problem \eqref{15}(with $t=0$), namely find a system $\Xi$ and $(\alpha,Z)\in {\cal A}(\Xi,0,x,\nu)$
such that \eqref{eq:eq716} holds. Denote by $\bar \nu$ the law of $X$ where $Z=(X,Y,R)$ and write
$\bar \nu = \Phi(\nu)$ (this is not precise since in general there may be more than one solution of the stochastic control problem in \eqref{15}).
\item[(ii)] Find the fixed point of the map $\Phi$, namely a $\bar \nu \in {\cal P}_{T,L}$ for which
$\bar \nu = \Phi(\bar \nu)$. Note that by definition, such a $\bar \nu$ will be a solution of the MFG.
\end{itemize}
We now analyze the MFG by following the above steps. The main result is Theorem \ref{thm_fixed} which gives existence of solutions of the MFG under suitable conditions and proves uniqueness
of solutions under stronger conditions.
\subsection{Solving the stochastic control problem \eqref{15}}\label{sec3b1}
For $c \in(0,\infty)$, let ${\cal M}_c$ be the collection of all $\nu \in {\cal P}_{T,L}$
such that
\begin{align}\notag
\sup_{0\le s<t\le T}\frac{W_1(\nu(t),\nu(s))}{(t-s)^{1/2}}\le c
\end{align}
and let
\begin{align}\notag
{\cal M}_0 = \cup_{c>0} {\cal M}_c.
\end{align}
Fix a measure $\nu\in{\cal M}_0$.
For $\nu \in {\cal P}_{T,L}$, the function $t \mapsto \nu(t)$ is a continuous function from $[0,T]$ to
${\cal P}([0,L])$ and with an abuse of notation we denote this continuous function once more as $\nu$.
As one might expect, the value function $V_{\nu}(t,x)$ corresponds to the Hamilton-Jacobi-Bellman (HJB) equation:
\begin{align}\label{HJB1}
-D_t \phi-H(t,\nu(t),x,D \phi)-\frac{1}{2}\sigma^2 D^2\phi=0,\qquad (t,x)\in[0,T]\times[0,L],
\end{align}
with the boundary conditions (BC)
\begin{align}\label{HJB2}
\phi(T,x)=g(\nu(T),x),\; D\phi(t,0)=-y(t,\nu(t)),\text{ and } D\phi(t,L)=r(t,\nu(t)),\; t\in[0,T],
\end{align}
where $H$ is the Hamiltonian given as
$$H(t,\eta,x,p)=\inf_{u\in U} h (t,\eta,x,u,p)$$
and $h:[0,T]\times{\cal P}([0,L])\times [0,L]\times U\times \mathbb{R} \to \mathbb{R}$ is defined as
\begin{align} \label{19}
h(t,\eta,x,u,p)=f(t,\eta,x,u)+b(t,\eta,x,u)p.
\end{align}
We now introduce a key condition under which the above HJB equation characterizes the value function $V_{\nu}(t,x)$.
\begin{assumption}\label{assumption1}$\,$
\begin{enumerate}
\item[(a)] There exists $c_L \in (0,\infty)$ such that for every $t,t'\in[0,T]$, $\eta,\eta'\in{\cal P}([0,L])$, $x,x'\in[0,L]$, and $\alpha,\alpha'\in U$,
\begin{align}
&|f(t,\eta,x,\alpha)-f(t',\eta',x',\alpha')| + |g(\eta,x) - g(\eta',x')|+|b(t,\eta,x,\alpha)-b(t',\eta',x',\alpha')|\nonumber\\
&\quad +|y(t,\eta)-y(t',\eta')|+|r(t,\eta)-r(t',\eta')|\nonumber\\
&\quad\quad\le c_L (|t-t'|+W_1(\eta,\eta')+|x-x'|+|\alpha-\alpha'|).\label{19b}
\end{align}
\item[(b)]
For every $(t,\eta, x,p) \in [0,T]\times {\cal P}([0,L])\times [0,L]\times \mathbb{R}$, there is a unique
$\hat \alpha(t,\eta,x,p) \in U$ such that
\begin{align}\label{19a}
\hat \alpha(t,\eta,x,p)=\underset{u\in U}{\arg\min}\;h(t,\eta,x,u,p) .
\end{align}
\end{enumerate}
\end{assumption}
From Berge's maximum theorem (see \cite[Theorem 17.31]{Aliprantis2006}) and part (b) of the above assumption
it follows that
$\hat \alpha$ is a continuous function on $[0,T]\times {\cal P}([0,L])\times [0,L]\times \mathbb{R}$.
Also note that \eqref{19b} implies that $b,f,g,y,r$ are bounded functions, in particular,
\begin{equation}\sup_{(\eta,x,u)\in [0,T]\times {\cal P}([0,L])\times [0,L]\times U}|b(t,\eta,x,u)| \doteq c_B < \infty.
\label{eq:eq429}
\end{equation}
The first part of the condition implies H\"{o}lder-1/2 continuity in $t$ when $\eta$ is replaced by $\nu(t)$
for some $\nu \in {\cal M}_0$.
For example, if $\nu \in {\cal M}_c$, for $t,t' \in [0,T]$
\begin{align}\label{eq_20}
|y(t,\nu(t))-y(t',\nu(t'))|&\le c_L(c+\sqrt{T})|t-t'|^{1/2}.
\end{align}
Similar estimates hold for $b,f,r$ and $g$.
Assumption \ref{assumption1} will guarantee that the value function is the unique classical solution of the HJB equation \eqref{HJB1} with BC \eqref{HJB2} (see Proposition \ref{lem2}).
The assumption will also be used in the fixed point argument (Section \ref{sec3b2}) and in the asymptotic analysis of the $n$-player game (Section \ref{sec4}).
Using Girsanov's theorem it is easily checked that given a measurable function
$\gamma: [0,T]\times [0,L] \to U$, $\nu \in {\cal P}_{T,L}$, and $(t,x) \in [0,T]\times [0,L]$
there is a unique weak solution $Z$ to
\begin{equation}
\label{eq:feedback}
Z(s) = (X,Y,R)(s)=\Gamma\left(x+\int_0^\cdot b(t+t', \nu(t+t'), X(t'), \gamma(t', X(t')))dt'+\sigma B(\cdot)\right)(s),
\end{equation}
$s\in[0,T-t]$, namely there is a system $\Xi = (\Omega,{\cal F},\{{\cal F}_t\},{\mathbb P}, B)$ on which is given a ${\cal F}_s$-adapted
continuous process $Z = (X,Y,R)$ solving the above equation and if $\Xi' = (\Omega',{\cal F}',\{{\cal F}'_t\},{\mathbb P}', B')$ is another system on which is given a ${\cal F}'_s$-adapted
continuous process $Z' = (X',Y',R')$ solving the above equation with $(Z,X, B)$ replaced by
$(Z',X', B')$ then ${\mathbb P}\circ Z^{-1} = {\mathbb P}'\circ (Z')^{-1}$. Note also that with $\alpha(s) = \gamma(s, X(s))$,
$(\alpha,Z)\in {\cal A}(\Xi,t,x,\nu)$. We refer to the function $\gamma$ as a feedback control and we call it an {\em optimal} feedback control for the stochastic control problem in \eqref{15} if
$$V_{\nu}(t,x) = J_{\nu}(t,x,\alpha, Z).$$
The following result which is a consequence of Theorem 13.24 of \cite{Lieberman1996} says that
the HJB equation \eqref{HJB1} with BC \eqref{HJB2} admits a unique classical solution which can be characterized as the value function $V_\nu$. Moreover, there exists an optimal feedback control with certain regularity properties.
We follow the notation from \cite{Lieberman1996}. Let for $\delta \in (0,1]$, $H_{\delta}$ be the collection of
maps $\psi:(0,T)\times (0,L) \to \mathbb{R}$ such that
$$\sup_{0<t <t'<T, 0< x<x'<L} \frac{|\psi(t,x) - \psi(t',x')|}{|t-t'|^{\delta/2}+ |x-x'|^{\delta}} < \infty.$$
Note that such a function can be continuously extended to $[0,T]\times [0,L]$ and we denote the extension by the same symbol.
Also, let $H_{2+\frac{1}{2}}$ be the collection of continuous real functions $\psi$ on $[0,T]\times [0,L]$
such that
$x\mapsto \psi(t,x)$ is twice continuously differentiable on $(0,L)$ for all $t \in (0,T)$,
$t\mapsto \psi(t,x)$ is continuously differentiable on $(0,T)$ for all $x \in (0,L)$,
the functions
$$\psi,\; D_t \psi,\; D \psi\; D^2 \psi,$$
are bounded on $(0,T)\times (0,L)$ and the functions
$D_t \psi$, $D^2 \psi$ are in $H_{1/2}$.
\begin{proposition}\label{lem2}
Fix $\nu\in {\cal M}_0$ and suppose that Assumption \ref{assumption1} holds. Then $V_\nu\in H_{2+\frac{1}{2}}$ and
it is the unique solution of \eqref{HJB1}--\eqref{HJB2}.
Furthermore, with $\hat \alpha$ as introduced in Assumption \ref{assumption1},
the map $(s,x')\mapsto \hat \alpha(s,\nu(s),x',D V_\nu(s,x'))$ is continuous and the
feedback control $\hat\gamma(u, x')\doteq \hat \alpha(u+t,\nu(u+t),x', D V_\nu(u+t,x'))$ is an optimal feedback control for \eqref{15} for every $t \in (0,T)$. Moreover, any optimal control $\alpha$ for \eqref{15} satisfies $\alpha(u,\omega)=\hat\gamma(u,X(u,\omega))$, $\lambda_T^t\otimes{\mathbb P}$ almost surely (a.s.), where $\lambda_T^t$ denotes the Lebesgue measure on $[0,T-t]$.
\end{proposition}
{\bf Proof:}
From \cite[Theorem 13.24]{Lieberman1996} and the paragraph following its statement it follows that \eqref{HJB1}--\eqref{HJB2} admits a solution in $H_{2+\tfrac{1}{2}}$. We remark that the key conditions needed to appeal to this theorem are that
$$\sup_{0< t <t' <T} \frac{|y(t,\nu(t)) - y(t',\nu(t'))| + |r(t,\nu(t)) - r(t',\nu(t'))|}{|t-t'|^{1/2}} < \infty,$$
and for each $M\in (0,\iy)$
$$
\sup_{\substack{t,t' \in [0,T],\, x,x' \in [0,L],\, p,p' \in [-M,M],\\ t\neq t', x\neq x', p\neq p'}}
\frac{|H(t,\nu(t), x,p) - H(t',\nu(t'), x', p')| }{|t-t'|^{1/2} + |x-x'| + |p-p'|} < \infty.$$
Both these conditions are easily seen to hold on using Assumption \ref{assumption1} and the property that
$\nu \in {\cal M}_c$ for some $c<\infty$ (see e.g.~\eqref{eq_20}).
We now argue uniqueness and characterize the solution as the value function $V_{\nu}$ in \eqref{15}.
Let $\psi\in H_{2+ \frac{1}{2}}$ be a solution of \eqref{HJB1}--\eqref{HJB2}.
Fix $(t,x) \in [0,T]\times [0,L]$, a system $\Xi$ and $(\alpha, Z) \in {\cal A}(\Xi, t,x,\nu)$.
By an application of It\^{o}'s lemma and using \eqref{HJB1}--\eqref{HJB2} we get
\begin{align}
\mathbb{E}[g(\nu(T),X(T-t))] &= \mathbb{E}[\psi(T,X(T-t))]\nonumber\\
& =\psi(t,x) + \mathbb{E}\left[\int_0^{T-t} D\psi(s+t,0)dY(s)-\int_0^{T-t} D\psi(s+t,L)dR(s)\right]\nonumber\\
&\quad +\mathbb{E}\int_0^{T-t}\Big[D_s \psi(s+t,X(s))+b(s+t,\nu(s+t),X(s),\alpha(s))D \psi(s+t,X(s))\nonumber\\
&\quad +\frac{1}{2}\sigma^2D^2\psi(s+t,X(s))\Big]ds \nonumber\\
&\ge \psi(t,x)+\mathbb{E}\Big[-\int_0^{T-t} f(s+t,\nu(s+t),X(s),\alpha(s))ds\nonumber\\
&\quad-\int_0^{T-t}y(s+t,\nu(s+t))dY(s) -\int_0^{T-t}r(s+t,\nu(s+t))dR(s)\Big].
\label{eq:eq624}
\end{align}
Hence, with $Z=(X,Y,R)$,
\begin{align}\label{unique1}
J_\nu(t,x,\alpha, Z)&=\mathbb{E}\left[\int_0^{T-t}f(s+t,\nu(s+t),X(s),\alpha(s))ds +g(\nu(T),X(T-t)) \right.\\\notag
&\qquad\qquad\left.
+\int_0^{T-t}y(s+t,\nu(s+t))dY(s) +\int_0^{T-t}r(s+t,\nu(s+t))dR(s)\right]\\\notag
&\ge \psi(t,x).
\end{align}
Since $\Xi, \alpha, Z$ are arbitrary, we get that
$V_\nu(t,x)\ge\psi(t,x)$ for all $(t,x)\in [0,T]\times [0,L]$.
Let $\hat \gamma$ be as in the statement of the proposition with $V_{\nu}$ replaced by $\psi$.
Let $Z = (X,Y,R)$ be a solution of \eqref{eq:feedback} with $\gamma$ replaced by $\hat \gamma$ given on some system
$\Xi$.
Then a similar calculation using It\^{o}'s formula but with $\alpha(u) = \hat\gamma(u, X(u))$ shows that
$J_{\nu}(t,x,\alpha, Z) = \psi(t,x)$ for all $(t,x) \in [0,T]\times [0,L]$. This shows that $\psi = V_{\nu}$
and so $V_\nu$ is the unique $H_{2+\frac{1}{2}}$ solution of \eqref{HJB1}--\eqref{HJB2}. Also, we have that $\hat \gamma$
as in the statement of the lemma is an optimal feedback control.
Since $V_{\nu}\in H_{2 + \frac{1}{2}}$, the continuity of the map $(t,x)\mapsto \hat \alpha(t,\nu(t),x,D V_\nu(t,x))$
is immediate from the continuity of the map $\hat\alpha$ that was noted below Assumption \ref{assumption1}.
We now show that $\hat\gamma$ is the unique optimal control for \eqref{15}. Fix an optimal control $(\alpha, Z) \in {\cal A}(\Xi, t, x, \nu)$ given on some system $\Xi$. We claim that
$\alpha(u,\omega)=\hat\gamma(u, X(u,\omega))$, $\lambda_T^t\times {\mathbb P}$-a.s. Indeed, suppose that there is a set with a positive $\lambda_T^t\times{\mathbb P}$-measure on which
the equality fails. Then by Assumption \ref{assumption1}.(b) together with \eqref{HJB1}--\eqref{HJB2}, it follows that \eqref{eq:eq624} holds with a strict inequality, which in turn implies that \eqref{unique1} holds with a strict inequality. Recalling that $V_\nu=\psi$, we arrive at a contradiction.
\hfill$\Box$
\subsection{Solving the MFG}\label{sec3b2}
We now turn to step (ii) in solving the fixed point problem. Although for existence of a fixed point Assumption \ref{assumption1} will suffice, in order to get uniqueness, we will need the following additional assumption.
Similar assumption has been used to argue uniqueness of fixed points in previous works on MFG (see e.g.~\cite[Theorem 2.4]{Lasry2007}, \cite[Section 3]{Cardaliaguet2013}, \cite[equation (17)]{Gomes2013}, \cite[Assumption (U)]{Carmona2015}).
Fix $\eta_0 \in {\cal P}([0,L])$.
\begin{assumption}\label{assumptionU}
For every $(t,\eta,x,u)\in[0,T]\times{\cal P}([0,L])\times[0,L]\times U$,
\begin{align}\label{19q}
b(t,\eta,x,u)&=b(t,\eta_0,x,u),\quad
f(t,\eta,x,u)=f_0(t,\eta,x)+f_1(t,x,u),\\\label{19pp}
y(t,\eta)&=y(t,\eta_0),\quad
r(t,\eta)=r(t,\eta_0).
\end{align}
Moreover, for every $t\in[0,T]$ and $\eta,\eta'\in{\cal P}([0,L])$, $f_0$ and $g$ satisfy the following monotonicity property
\begin{align}\nota
\int_0^L [f_0(t,\eta,x)-f_0(t,\eta',x))]d(\eta-\eta')(x)&\ge 0,\\\nota
\int_0^L (g(\eta,x)-g(\eta',x))d(\eta-\eta')(x)&\ge 0
\end{align}
\end{assumption}
Abusing notation, when Assumption \ref{assumptionU} holds, we will write $b(t,x,u) =b(t,\eta_0,x,u)$, $y(t)=y(t,\eta_0)$, and $r(t)=r(t,\eta_0)$.
\begin{remark}\label{rem_uniqueness} Examples satisfying Assumption \ref{assumptionU} are given in \cite[page 8]{Cardaliaguet2013} and \cite[page 6]{bayzha2016}. Another natural example that satisfies this assumption is a cost function that is linear in the mean-field term. That is,
\begin{align}\notag
f_0(t,\eta,x)=a_1(t)(c_1+\psi_1(x))\int_0^L\psi_1(y)d\eta(y),\qquad g(\eta,x)=(c_2+\psi_2(x))\int_0^L\psi_2(y)d\eta(y),
\end{align}
where $a_1: [0,T]\to\mathbb{R}_+$, $\psi_1,\psi_2:[0,L]\to\mathbb{R}$, and $c_1,c_2\in\mathbb{R}$.
From a modeling perspective, by choosing positive and nondecreasing $\psi_i$'s and a positive $a_1$, the system planner penalizes all servers collectively when the empirical measure has high
$\psi_i$-moments and in addition it penalizes individual servers for long queues.
\end{remark}
\begin{theorem}\label{thm_fixed}
Under Assumption \ref{assumption1}, there exists a solution of the MFG. If in addition Assumption \ref{assumptionU} holds then there is a unique MFG solution.
\end{theorem}
The proof will appeal to Schauder's fixed point theorem. Since Schauder's original work (cf.~\cite{Schauder1930}), there have been several versions of this result. We now quote the version (\cite[Theorem 4.1.1]{Smart1974}) that will be used here.
\begin{lemma}\label{lem_Schauder}
Let ${\cal M}$ be a non-empty convex subset of a normed space ${\cal B}$. Let $\Phi$ be a continuous mapping of ${\cal M}$ into a compact set ${\cal K}\subset{\cal M}$. Then $\Phi$ has a fixed point, namely there exists $x \in {\cal K}$ such that
$\Phi(x)=x$.
\end{lemma}
\noindent {\bf Proof of Theorem \ref{thm_fixed}:}
We will apply Lemma \ref{lem_Schauder} to the space ${\cal B}$ of finite signed measures on ${\cal C}([0,T]:[0,L])$ which is equipped with the Kantorovich-Rubinstein norm
\begin{align}\notag
\|\nu\|_{KR}=\sup\left\{\left|\int_{{\cal C}([0,T]:[0,L])}F(\omega)d\nu(\omega)\right|\;:\;F\in\text{Lip}_1({\cal C}([0,T]:[0,L]))\right\},\; \nu\in{\cal B} .
\end{align}
The distance driven by the norm, restricted to ${\calP_{T,L}}$, coincides with the Wasserstein's distance of order 1 (see \cite[Remark 6.5]{Villani2009}), which due to compactness of $[0,L]$ generates the same topology
on ${\calP_{T,L}}$ as that for weak convergence.
We now introduce a mapping $\Phi$ on the non-empty and convex set ${\cal M}_0$ that satisfies the conditions stated in Lemma \ref{lem_Schauder}.
\noindent {\bf Definition of $\Phi$.}
For $\nu\in {\cal M}_0$, let $\alpha_{\nu}$ denote the optimal feedback control $\hat \gamma$ for \eqref{15} (with $t=0$) given through Proposition \ref{lem2}.
Let $Z^{\nu} = (X^{\nu}, Y^{\nu}, R^{\nu})$ denote the unique weak solution of \eqref{eq:feedback} with $\gamma$ replaced with $\alpha_{\nu}$ given on some system
$\Xi =(\Omega,{\cal F},\{{\cal F}_t\},{\mathbb P}, B)$. Define
$\Phi(\nu) = {\mathbb P} \circ (X^{\nu})^{-1}$.
\noindent {\bf Invariance of ${\cal M}_0$.}
For $\nu\in {\cal M}_0$ and $0\le s \le t <T$
\begin{align}\label{50}
W_1(\Phi(\nu)(t),\Phi(\nu)(s))&\le \mathbb{E}|X^{\nu}(t)-X^{\nu}(s)|\\\notag&\le c_S\max\{c_B,\sigma\}\left((t-s)+\mathbb{E} \sup_{s \le u \le t}|B(u)-B(s)|\right)\\\notag&\le 4\hat{C} (t-s)^{1/2},
\end{align}
where $\hat C \doteq c_S\max\{c_B,\sigma\}$, and the last inequality uses Doob's maximal inequality. This shows that $\Phi(\nu) \in {\cal M}_0$ for all $\nu\in {\cal M}_0$.
We now show
that $\Phi({\cal M}_0)$ is contained in a compact set in ${\calP_{T,L}}$, i.e.~$\Phi({\cal M}_0)$ is relatively compact.
\noindent {\bf Relative compactness of $\Phi({\cal M}_0)$.} For $f \in {\cal C}([0,T]:\mathbb{R})$ and $\delta >0$, let
$$\varpi_f(\delta) = \sup_{s,t \in [0,T]:|t-s|\le \delta}|f(t)-f(s)|.$$
Then for $\varepsilon, \delta,\varrho >0$ and $\nu \in {\cal M}_0$, similar to the estimate in \eqref{50},
\begin{align*}
\Phi(\nu) (X^{\nu}: \varpi_{X^{\nu}}(\delta) \ge \varrho) &\le \frac{1}{\varrho}\mathbb{E} \sup_{0\le s \le t \le s+\delta \le T} |X^{\nu}(t) - X^{\nu}(s)|\\
&\le \frac{\hat C}{\varrho} \left(\delta + \mathbb{E} \sup_{0\le s \le t \le s+\delta \le T} \sup_{s\le u \le t}|B(u)-B(s)| \right)\\
&\le \frac{\hat C}{\varrho} (\delta + \mathbb{E} \varpi_{B}(\delta)).
\end{align*}
Since $\mathbb{E} \varpi_{B}(\delta) \to 0$ as $\delta \to 0$, we have from \cite[Theorem 7.3]{Billingsley1999} that $\{\Phi(\nu): \nu \in {\cal M}_0\}$ is relatively
compact in ${\calP_{T,L}}$.
\noindent {\bf Continuity of $\Phi$.}
We now argue that $\Phi$ is a continuous map on ${\cal M}_0$. Consider a system $\Xi =(\Omega,{\cal F},\{{\cal F}_t\},{\mathbb P}, B)$ and let $Z^0 = (X^0, Y^0, R^0)$ be given as
\begin{equation}\label{eq:eq433}
Z^0(s) = (X^0(s), Y^0(s), R^0(s)) = \Gamma(x + \sigma B(\cdot))(s), \; 0 \le s \le T.\end{equation}
Define for $\nu' \in {\cal M}_0$
\begin{align}\label{up1}
\alpha^0_{\nu'}(t) = \hat \alpha(t,\nu'(t),X^0(t),D V_{\nu'}(t,X^0(t))), \; t \in [0,T].
\end{align}
Let $\nu^n, \nu \in {\cal M}_0$ be such that $\nu^n \to \nu$.
Since $\sup_{0\le t\le T}W_1(\nu^n(t),\nu(t))\le W_1(\nu^n,\nu)$, the above convergence implies that
\begin{align}\label{eq_21}
\lim_{n\to\iy} \sup_{0\le t\le T}W_1(\nu^n(t),\nu(t))=0.
\end{align}
We now show that $\alpha_{\nu^n}^0\to\alpha_\nu^0$ in $\lambda_T\times\cal{\mathbb P}$-measure.
Recall that $\hat \alpha$ is a continuous map.
Hence, in view of \eqref{eq_21}, for the desired convergence, it is sufficient to show that
\begin{align}\label{new79}
\lim_{n\to\iy}\mathbb{E}\left[\int_0^T|D V_{\nu^n}(t,X^0(t)) - D V_{\nu}(t,X^0(t))|dt\right]=0.
\end{align}
The proof is a modification of the proof of Theorem 2.1 in \cite{Hu1997}.
Applying It\^{o}'s lemma to $V_{\nu'}(t,X^0(t))$ for fixed $\nu'\in{\cal M}_0$ and using from Proposition \ref{lem2} the fact that $V_{\nu'}$ solves \eqref{HJB1}--\eqref{HJB2}, we have for every $t\in [0,T]$,
\begin{align}\label{eq_40}
&V_{\nu'}(t,X^0(t)) - V_{\nu'}(T,X^0(T))
\\\notag
&\quad= \int_t^TH(s,\nu'(s),X^0(s),D V_{\nu'}(s,X^0(s)))ds-\sigma\int_t^T DV_{\nu'}(s,X^0(s))dB(s)\\\notag
&\quad+\int_t^T y(s,\nu'(s))dY^0(s)+\int_t^T r(s,\nu'(s))dR^0(s).
\end{align}
Let,
\begin{align}
\Delta V^n(t)&= V_{\nu^n}(t,X^0(t))-V_{\nu}(t,X^0(t)),\; \Delta DV^n(t)= DV_{\nu^n}(t,X^0(t))-DV_{\nu}(t,X^0(t)) \notag\\
\Delta g^n(T)&= \Delta V^n(T) = g(\nu^n(T),X^0(T))-g(\nu(T),X^0(T)),\notag\\
\Delta y^n(t)&= y(t,\nu^n(t))-y(t,\nu(t)),\;
\Delta r^n(t) = r(t,\nu^n(t))-r(t,\nu(t)), \notag\\
\Delta \psi^n(t)&= \sup_{\alpha\in U}|\psi(t,\nu^n(t),X^0(t),\alpha)-\psi(t,\nu(t),X^0(t),\alpha)|, \quad \psi \in \{f, b\}.
\label{eq_41}
\end{align}
Then, \eqref{eq_40} and \eqref{eq_41} imply,
\begin{align}\notag
&\Delta V^n(t)+ \sigma\int_t^T\Delta D V^n(s)dB(s)\\\notag
&\quad=\Delta g^n(T) + \int_t^T\Delta y^n(s)dY^0(s)+\int_t^T\Delta r^n(s)dR^0(s)\\\notag
&\qquad+\int_t^T[H(s,\nu^n(s),X^0(s),DV_{\nu^n}(s,X^0(s)))-H(s,\nu^n(s),X^0(s),DV_{\nu}(s,X^0(s)))]ds\\\notag
&\qquad+\int_t^T[H(s,\nu^n(s),X^0(s),DV_{\nu}(s,X^0(s)))-H(s,\nu(s),X^0(s),DV_{\nu}(s,X^0(s)))]ds.
\end{align}
Squaring both sides and then taking expectations gives
\begin{align}\label{new200}
&\mathbb{E}[(\Delta V^n(t))^2]+ \sigma^2\mathbb{E}\left[\int_t^T(\Delta D V^n(s))^2ds\right]\\\notag
&\quad\le 2(T-t)\mathbb{E}\Big[\int_t^T[H(s,\nu^n(s),X^0(s),DV_{\nu^n}(s,X^0(s)))-H(s,\nu^n(s),X^0(s),DV_{\nu}(s,X^0(s)))]^2ds\Big]\\\notag
&\qquad +2C^n(t)\\\notag
&\quad\le 2c_B^2(T-t)\mathbb{E}\Big[\int_t^T(\Delta DV^n(s))^2ds\Big]+ 2C^n(t),
\end{align}
where
\begin{align}\notag
C^n(t)&=
\mathbb{E}\Big[ \Delta g^n(T) + \int_t^T\Delta y^n(s)dY^0(s)+\int_t^T\Delta r^n(s)dR^0(s)\\\notag
&\qquad\quad+\int_t^T[H(s,\nu^n(s),X^0(s),DV_{\nu}(s,X^0(s)))-H(s,\nu(s),X^0(s),DV_{\nu}(s,X^0(s)))]ds\Big]^2.
\end{align}
Letting $\delta=\sigma^2/(4c_B^2)$, we get from \eqref{new200} that for every $t\in[T-\delta,T]$,
\begin{align}\label{new201}
\mathbb{E}[(\Delta V^n(t))^2]+\frac{\sigma^2}{2}\mathbb{E}\left[\int_t^T(\Delta DV^n(s))^2ds\right]\le 4C^n(t).
\end{align}
We now show that $\limsup_{n\to \infty} \sup_{0\le t \le T} C^n(t)=0$.
Note that there exists $C_1\in (0, \infty)$ such that the following inequality holds for all $n \in \mathbb{N}$.
\begin{align}\notag
\frac{1}{C_1} \sup_{0\le t \le T} C^n(t)&\le\mathbb{E}\left[(\Delta g^n(T))^2\right]+\mathbb{E}\left[(Y^0(T))^2\right]\sup_{0\le t\le T}(\Delta y^n(t))^2+\mathbb{E}\left[(R^0(T))^2\right]\sup_{0\le t\le T}(\Delta r^n(t))^2\\
&\quad+\mathbb{E}\left[
\int_0^T(\Delta f^n(s))^2ds\right]+\mathbb{E}\left[\int_0^T(\Delta b^n(s))^2ds \int_0^T (DV_\nu(s,X^0(s)))^2ds\right].\label{eq:eq1259}
\end{align}
Using the properties of the Skorohod map (Lemma \ref{lem_Skorohod}) it follows that
$$ \mathbb{E}\left[(Y^0(T))^2 + (R^0(T))^2\right]<\iy .$$
The convergence of the right side of \eqref{eq:eq1259} to $0$ is now immediate from Assumption \ref{assumption1}(a), the boundedness of $DV_{\nu}$, and \eqref{eq_21}.
Thus from \eqref{new201} we have that
\begin{align}\label{new89}
\lim_{n\to\iy}\mathbb{E}\left[\int_{T-\delta}^T|D V_{\nu^n}(t,X^0(t)) - D V_{\nu}(t,X^0(t))|dt\right]=0,\; \lim_{n\to\iy}\mathbb{E}[(\Delta V^n(T-\delta))^2] =0.
\end{align}
Using the second convergence in \eqref{new89} and repeating the above argument for $t \in [T-\delta, T]$ to the interval $[T-2\delta, T-\delta]$, we see that
\eqref{new89} holds with $T$ replaced with $T-\delta$. Proceeding recursively in this manner we have \eqref{new79}. Hence we have shown that
\begin{equation}
\label{eq:eq215}
\alpha_{\nu^n}^0\to\alpha_\nu^0 \mbox{ in } \lambda_T\times\cal{\mathbb P} \mbox{ - measure}.
\end{equation}
Using the above property we will now argue that $\Phi(\nu^n) \to \Phi(\nu)$ as $n\to \infty$, completing the proof of continuity of $\Phi$.
Let for $\nu' \in {\cal M}_0$,
$$\gamma_{\nu'}(t,x) = \hat \alpha(t, \nu'(t), x, DV_{\nu'}(t,x)), \;\; \hat b_{\nu'}(t,x) = b(t, \nu'(t), x, \gamma_{\nu'}(t,x)),\; (t,x) \in [0,T]\times [0,L]$$
and let ${\mathbb P}^{\nu'}$ be a probability measure on $(\Omega, {\cal F})$ defined as
\begin{align}\label{up2}
d{\mathbb P}^{\nu'} = \exp \left[ \frac{1}{\sigma} \int_0^T \hat b_{\nu'}(t, X^0(t)) dB(t) - \frac{1}{2\sigma^2} \int_0^T \hat b^2_{\nu'}(t, X^0(t)) dt\right] d{\mathbb P}.
\end{align}
By Girsanov's theorem ${\mathbb P}^{\nu'} \circ (X^0)^{-1} = \Phi(\nu')$. Thus to show $\Phi(\nu^n) \to \Phi(\nu)$ it suffices to argue that ${\mathbb P}^{\nu^n} \to {\mathbb P}^{\nu}$. We will in fact show that
$R({\mathbb P}^{\nu}\|{\mathbb P}^{\nu^n}) \to 0$ as $n\to \infty$, where
$$
R({\mathbb P}^{\nu}\|{\mathbb P}^{\nu^n}) = \mathbb{E}^{\nu} \left(\log \frac{d{\mathbb P}^{\nu}}{d{\mathbb P}^{\nu^n}}\right) = \mathbb{E} \left( \frac{d{\mathbb P}^{\nu}}{d{\mathbb P}}\log \frac{d{\mathbb P}^{\nu}}{d{\mathbb P}^{\nu^n}}\right)$$
is the relative entropy of ${\mathbb P}^{\nu}$ with respect to ${\mathbb P}^{\nu^n}$, which due to Pinsker's inequality (see \cite[Page 132]{Tsybakov2009}) gives the convergence of ${\mathbb P}^{\nu^n}$ to ${\mathbb P}^{\nu}$.
Let
$$
\Delta \hat b_n(t,x) = \hat b_{\nu}(t,x) - \hat b_{\nu^n}(t,x), \;\; \Delta \hat b_n^2(t,x) = \hat b_{\nu}^2(t,x) - \hat b_{\nu^n}^2(t,x), \; (t,x) \in [0,T]\times [0,L].
$$
With this notation,
$$
\log \frac{d{\mathbb P}^{\nu}}{d{\mathbb P}^{\nu^n}} = \frac{1}{\sigma} \int_0^T \Delta \hat b_{n}(t, X^0(t)) dB(t) - \frac{1}{2\sigma^2} \int_0^T \Delta \hat b^2_{n}(t, X^0(t)) dt .$$
Also, noting that since $b$ is bounded $\mathbb{E} (\frac{d{\mathbb P}^{\nu}}{d{\mathbb P}})^2 \doteq \kappa < \infty$, we have from Cauchy-Schwarz inequality
\begin{equation}\label{eq:eq317}
R({\mathbb P}^{\nu}\|{\mathbb P}^{\nu^n}) \le \sqrt{\kappa}\left(\mathbb{E} \left[\frac{1}{\sigma} \int_0^T \Delta \hat b_{n}(t, X^0(t)) dB(t) - \frac{1}{2\sigma^2} \int_0^T \Delta \hat b^2_{n}(t, X^0(t)) dt\right]^2\right)^{1/2} .\end{equation}
Next note that
\begin{align*}
&\mathbb{E} \left[\frac{1}{\sigma} \int_0^T \Delta \hat b_{n}(t, X^0(t)) dB(t)\right]^2 \\
&\quad = \frac{1}{\sigma^2} \int_0^T \mathbb{E}[\Delta \hat b_{n}(t, X^0(t))]^2 dt\\
&\quad = \frac{1}{\sigma^2} \int_0^T \mathbb{E} \left[b(t, \nu(t), X^0(t), \alpha^0_{\nu}(t, X^0(t))) - b(t, \nu^n(t), X^0(t), \alpha^0_{\nu^n}(t, X^0(t))) \right]^2 dt.
\end{align*}
The last term converges to $0$ from the boundedness and continuity of $b$, \eqref{eq_21} and \eqref{eq:eq215}. Similarly
$$
\mathbb{E} \left[\frac{1}{2\sigma^2} \int_0^T \Delta \hat b^2_{n}(t, X^0(t)) dt\right]^2 \to 0$$
as $n \to \infty$. Using the above two observations in \eqref{eq:eq317} we have $R({\mathbb P}^{\nu}\|{\mathbb P}^{\nu^n}) \to 0$ and thus,
as argued earlier, the proof of continuity of $\Phi$ is complete. Thus we have shown that $\Phi$ is a continuous map on ${\cal M}_0$, which is a non-empty convex subset of the normed space ${\cal B}$, into a compact set ${\cal K}\subset{\cal M}_0$. Thus by the fixed point theorem in Lemma \ref{lem_Schauder}, $\Phi$ has a fixed point.\\
The first results on unique solvability of a MFG go back to \cite{Lasry2007}. Since then uniqueness has been argued in various settings (see e.g.~\cite{Cardaliaguet2013,Gomes2013,Carmona2015}). The proof given below uses arguments similar to those in \cite[Section 7.3]{Carmona2015}, however for the sake of completeness we give the details.
Consider as before a system $\Xi =(\Omega,{\cal F},\{{\cal F}_t\},{\mathbb P}, B)$ and let $Z^0 = (X^0, Y^0, R^0)$ be given through \eqref{eq:eq433}.
Let $\nu_1,\nu_2\in{\cal M}_0$. For $i=1,2$, let $\alpha^0_{\nu_i}$ and ${\mathbb P}^{\nu_i}$ be given by \eqref{up1} and \eqref{up2}, respectively, with $\nu_i$ replacing $\nu'$.
Applying It\^{o}'s lemma to $V_{\nu_1}(t,X^0(t))$ and recalling that $V_{\nu_1}$ solves \eqref{HJB1}--\eqref{HJB2} with $\nu$ replaced with $\nu_1$, we get
\begin{align*}
&V_{\nu_1}(T,X^0(T))-V_{\nu_1}(0,x) \\
&\quad = - \int_0^T H(t, \nu_1(t), X^0(t), DV_{\nu_1}(t,X^0(t))) dt
+ \sigma \int_0^T DV_{\nu_1}(t,X^0(t)) dB(t) + \zeta_T\\
&= - \int_0^T h(t, \nu_1(t), X^0(t), \alpha^0_{\nu_1}(t), DV_{\nu_1}(t,X^0(t))) dt
+ \sigma \int_0^T DV_{\nu_1}(t,X^0(t)) dB(t) + \zeta_T,\\
\end{align*}
where recalling the form of $y$ and $r$ from Assumption \ref{assumptionU}
\begin{align*}
\zeta_T &= \int_{[0,T]} DV_{\nu_1}(t,0) dY^0(t) + \int_{[0,T]} DV_{\nu_1}(t,L) dR^0(t)\\
&= \int_{[0,T]} y(t) dY^0(t) + \int_{[0,T]} r(t) dR^0(t).
\end{align*}
Observing that for $t \in [0,T]$
$$
h(t, \nu_1(t), X^0(t),\alpha^0_{\nu_1}(t), DV_{\nu_1}(t,X^0(t)))
= h(t, \nu_1(t), X^0(t),\alpha^0_{\nu_1}(t), DV_{\nu_2}(t,X^0(t))) + {\cal R}_t,$$
where
$${\cal R}_t = [DV_{\nu_1}(t,X^0(t)) - DV_{\nu_2}(t,X^0(t))] \hat b_{\nu_1}(t, X^0(t)),$$
we have
\begin{align}\label{asaf1}
V_{\nu_1}(0,x) &= g(\nu_1(T),X^0(T)) + \int_0^T h(t, \nu_1(t), X^0(t), \alpha^0_{\nu_1}(t), DV_{\nu_2}(t,X^0(t))) dt\\\notag
&\quad + \int_0^T {\cal R}_t dt - \sigma \int_0^T DV_{\nu_1}(t,X^0(t)) dB(t) - \zeta_T .
\end{align}
Similarly, applying It\^{o}'s lemma to $V_{\nu_2}(t,X^0(t))$
\begin{align}\label{asaf2}
V_{\nu_2}(0,x) &= g(\nu_2(T),X^0(T)) + \int_0^T H(t, \nu_2(t), X^0(t), DV_{\nu_2}(t,X^0(t))) dt\\\notag
&\quad - \sigma \int_0^T DV_{\nu_2}(t,X^0(t)) dB(t) - \zeta_T.
\end{align}
Substracting \eqref{asaf2} from \eqref{asaf1}
\begin{align*}
&V_{\nu_1}(0,x) - V_{\nu_2}(0,x) \\
&\quad = g(\nu_1(T),X^0(T)) - g(\nu_2(T),X^0(T))
- \sigma \int_0^T [DV_{\nu_1}(t,X^0(t))- DV_{\nu_2}(t,X^0(t))] dB^{\nu_1}(t)\\
&\quad + \int_0^T \left[h(t, \nu_1(t), X^0(t), \alpha^0_{\nu_1}(t), DV_{\nu_2}(t,X^0(t)))
- H(t, \nu_2(t), X^0(t), DV_{\nu_2}(t,X^0(t)))\right] dt,
\end{align*}
where for $i=1,2$, $B^{\nu_i}(t)\doteq B(t)-\frac{1}{\sigma}\int_0^t\hat b_{\nu_i}(s,X^0(s))ds$, $t\in [0,T]$. Since under ${\mathbb P}^{\nu_i}$, $B^{\nu_i}$ is a standard Brownian motion, taking expectation under the measure ${\mathbb P}^{\nu_1}$
\begin{align}\notag
&V_{\nu_1}(0,x)-V_{\nu_2}(0,x)\\
&\quad=\mathbb{E}^{\nu_1}\Big[\int_0^T\left[
h(t,\nu_1(t),X^0(t),\alpha^0_{\nu_1}(t),DV_{\nu_2}(t,X^0(t)))-H(t,\nu_2(t),X^0(t),DV_{\nu_2}(t,X^0(t)))\right]dt\notag\\
&\qquad\qquad\qquad+g(\nu_1(T),X^0(T))-g(\nu_2(T),X^0(T))
\Big].\label{up3}
\end{align}
A similar calculation shows
\begin{align}\notag
&V_{\nu_1}(0,x)-V_{\nu_2}(0,x)\\
&\quad=\mathbb{E}^{\nu_2}\Big[\int_0^T\left[
H(t,\nu_1(t),X^0(t),DV_{\nu_1}(t,X^0(t)))-h(t,\nu_2(t),X^0(t),\alpha^0_{\nu_2}(t),DV_{\nu_1}(t,X^0(t)))\right]dt \notag\\
&\qquad\qquad\qquad+g(\nu_1(T),X^0(T))-g(\nu_2(T),X^0(T))
\Big].\label{up3b}
\end{align}
By the definition of the $H$ and the form of $f$ in Assumption \ref{assumptionU} we get,
\begin{align}\notag
&h(s,\nu_1(s),X^0(s),\alpha^0_{\nu_1}(s),DV_{\nu_2}(s,X^0(s)))\\\notag
&\quad\ge H(s,\nu_1(s),X^0(s),DV_{\nu_2}(s,X^0(s)))\\\notag
&\quad=f_0(s,\nu_1(s),X^0(s))+f_1(s,X^0(s),\alpha^0_{\nu_2}(s))+DV_{\nu_2}(s,X^0(s))\hat b_{\nu_2}(s)
\end{align}
where the last equality uses the observation that since
$\hat \alpha(s, \eta,x,p)$ does not depend on $\eta$,
$\hat \alpha(s, \nu_1(s), X^0(s), DV_{\nu_2}(s,X^0(s))) = \alpha^0_{\nu_2}(s)$.
Therefore for all $s\in [0,T]$,
\begin{align}
&h(s,\nu_1(s),X^0(s),\alpha^0_{\nu_1}(s),DV_{\nu_2}(s,X^0(s)))-H(s,\nu_2(s),X^0(s),DV_{\nu_2}(s,X^0(s))) \notag\\
&\quad\ge f_0(s,\nu_1(s),X^0(s))-f_0(s,\nu_2(s),X^0(s)), \label{up5}
\end{align}
Similarly for all $s\in [0,T]$,
\begin{align}
&H(s,\nu_1(s),X^0(s),DV_{\nu_1}(s,X^0(s)))-h(s,\nu_2(s),X^0(s),\alpha^0_{\nu_2}(s),DV_{\nu_1}(s,X^0(s)))\notag\\
&\quad\le f_0(s,\nu_1(s),X^0(s))-f_0(s,\nu_2(s),X^0(s)). \label{up6}
\end{align}
Applying the last two inequalities to \eqref{up3} and \eqref{up3b}, we get
\begin{align}\label{newnew1}
0\le [\mathbb{E}^{\nu_2}-\mathbb{E}^{\nu_1}]\Big[&g(\nu_1(T),X^0(T))-g(\nu_2(T),X^0(T))\\\notag
&\quad+\int_0^T[f_0(s,\nu_1(s),X^0(s))-f_0(s,\nu_2(s),X^0(s))]ds\Big].
\end{align}
Until now $\nu_1$ and $\nu_2$ were arbitrary measures in ${\cal M}_0$. Suppose now that $\nu_i$, $i=1,2$, are fixed points
of $\Phi$. Then, for $i=1,2$, ${\mathbb P}^{\nu_i} \circ (X^0)^{-1} = \nu_i$ and so for all $s \in [0,T]$,
${\mathbb P}^{\nu_i} \circ (X^0(s))^{-1} = \nu_i(s)$.
In this case, using the inequalities in Assumption \ref{assumptionU} we get that the inequality \eqref{newnew1} can be replaced with equality. We claim that $\alpha^0_{\nu_1}=\alpha^0_{\nu_2}$, $\lambda_T^0\times{\mathbb P}$-a.s. Indeed, suppose that there is a set with positive $\lambda_T^0\times{\mathbb P}$-measure on which $\alpha^0_{\nu_1}\neq\alpha^0_{\nu_2}$. Then on this set \eqref{up5} and \eqref{up6} will hold with strict inequalities by Assumption \ref{assumption1} (b). Since the measures ${\mathbb P}$, ${\mathbb P}^{\nu_1}$, and ${\mathbb P}^{\nu_2}$ are equivalent, this will say that \eqref{newnew1} holds with a strict inequality as well, which contradicts the equality that was established above. This proves the claim. Since $b(t,\eta,x,u)$ does not depend on $\eta$ we conclude from the equality of $\alpha^0_{\nu_1}$ and $\alpha^0_{\nu_2}$ that $\hat b_{\nu_1}(t, X^0(t,\omega)) = \hat b_{\nu_2}(t, X^0(t,\omega))$, $\lambda_T^0\times{\mathbb P}$-a.s. and thus by
\eqref{up2}, ${\mathbb P}^{\nu_1}={\mathbb P}^{\nu_2}$.
Combining this with the fact that $\nu_i$ are fixed points of $\Phi$, we now have
$\nu_1={\mathbb P}^{\nu_1}\circ (X^0)^{-1}={\mathbb P}^{\nu_2}\circ (X^0)^{-1}=\nu_2$. \hfill$\Box$
\section{Asymptotic Nash equilibrium}\label{sec4}
\beginsec
The main result of this section is Theorem \ref{thm1}. The main idea in the proof is to use a solution $\bar \nu$ to the MFG (which from Theorem \ref{thm_fixed} exists under Assumption \ref{assumption1}) and the associated feedback control given by Proposition \ref{lem2} in order to construct an admissible control
$\tilde \alpha^n = \{\tilde \alpha^{n,i}\}_{i=1}^n$ for the $n$-player game. Specifically, the control
will be given in a feedback form through the following relation
\begin{align}
\tilde \alpha^{n,i}(t) \doteq \hat \alpha(t,\bar\nu(t),\tilde Q^{n,i}(t), D V_{\bar \nu}(t,\tilde Q^{n,i}(t))),
\label{eq:eq424}
\end{align}
where $\tilde Q^{n,i}$ is the corresponding scaled queue length under the feedback control. Note that the only information each of the players uses is its own state. Therefore, the problem is decentralized in the sense that players do not need to observe each others' states.
\color{black}
Our main condition, in addition to Assumption \ref{assumption1}, for $\{\tilde \alpha^n\}$ to be an asymptotic Nash equilibrium is the following. It in particular says that the drift function does not depend on the mean-field term.
Fix $\eta_0 \in {\cal P}([0,L])$.
\begin{assumption}\label{assumption4}
$\,$
\begin{enumerate}
\item[(a)]
For every $t\in[0,T]$, $\eta\in{\cal P}([0,L])$, $x\in[0,L]$, and $\alpha\in U$, one has,
\begin{align}\notag
b(t,\eta,x,\alpha)&=b(t,\eta_0,x,\alpha);
\end{align}
\item[(b)] There exists $x\in[0,L]$ such that for every $i\in\mathbb{N}$,
\begin{align}
\lim_{n\to\iy}\tilde Q^{n,i}(0)=x.
\end{align}
\end{enumerate}
\end{assumption}
As before, with an abuse of notation, we will write $b(t,\eta,x,\alpha)$ as $b(t,x,\alpha)$ when Assumption \ref{assumption4} holds.
As discussed in Remark \ref{rem_uniqueness}, part (a) of the assumption means that the empirical measure affects the drift only through the control, which in turn is affected by the empirical measure through the running cost.
Recall the probability space $(\Omega', {\cal F}', {\mathbb P}')$ from Section \ref{sec2a}.
Let for $n \in \mathbb{N}$, $t \in [0,T]$ and $i = 1, \cdots ,n$, $\beta^n(t): \Omega' \to U$ be such that
$$\tilde \alpha^n_{-i} =\{ \tilde \alpha^{n,1}, \ldots , \tilde \alpha^{n, i-1}, \beta^n, \tilde \alpha^{n, i+1}, \ldots , \tilde \alpha^{n,n}\}$$
is an admissible control (i.e.~$\tilde \alpha^n_{-i} \in {\cal U}^n$).
The following is the main result of the section.
\begin{theorem}\label{thm1}
Suppose Assumptions \ref{assumption1} and \ref{assumption4} hold.
Let $\bar \nu$, $\tilde \alpha^n$ and $\tilde \alpha^n_{-i}$, $i=1, \ldots, n$, $n \in \mathbb{N}$ be as introduced above.
Then
\begin{align}\label{20aa}
\limsup_{n\to \infty} J^{n,1}(\tilde Q^{n}(0);\tilde \alpha^{n})= V_{\bar\nu}(0,x)\le\liminf_{n\to \infty} J^{n,1}(\tilde Q^{n}(0);
\tilde \alpha^{n}_{-i})
\end{align}
\end{theorem}
Theorem \ref{thm1} in particular says that for every $\varepsilon>0$, there is $n\in\mathbb{N}$ sufficiently large such that $\tilde \alpha^n$ forms an $\varepsilon$-Nash equilibrium in the $n$-player game.
The proof is given in the next three sections. First in Section \ref{4z} (Proposition \ref{prop:tightemp}) we will prove
the convergence of empirical measures of the scaled queue length processes under controls
$\tilde \alpha^{n}$ and $\tilde \alpha^{n}_{-i}$ to $\bar \nu$. Next, in Section \ref{sec4a} (Proposition \ref{prop2}) we will prove the first equality in \eqref{20aa} and finally Proposition \ref{prop3} in Section \ref{sec4b} will prove the
inequality in \eqref{20aa}.
\subsection{Convergence of empirical measures}\label{4z}
Let for $i\in \mathbb{N}$, $\tilde\alpha^n_{-i} \in {\cal U}^n$ be as defined below Assumption \ref{assumption4}.
Let
$$\tilde \nu^n_{-i} = \frac{1}{n} \sum_{j=1}^n \delta_{\tilde Q^{n,j}},$$
where $\tilde Q^{n,j}$ is the controlled queue length process defined by \eqref{7} with $\alpha^n$ replaced with
$\tilde \alpha^n_{-i}$. The following result gives the convergence of $\tilde \nu^n_{-i}$ to $\bar \nu$.
\begin{proposition}\label{prop:tightemp}
Suppose Assumptions \ref{assumption1} and \ref{assumption4} hold. Then for every $i \in \mathbb{N}$,
$\tilde \nu^n_{-i}$ converges in probability, in ${\cal P}({\cal D}([0,T]:[0,L]))$, to $\bar \nu$ as $n \to \infty$.
\end{proposition}
{\bf Proof:}
Without loss of generality we assume that $i=1$. Recall that for $j \in \mathbb{N}$, $\tilde Q^{n,j}$ is defined by \eqref{7}
with $\alpha^n$ replaced with $\tilde \alpha^n_{-1}$. Define
$$\tilde\zeta^{n,i}(t) = \tilde Q^{n,i}(0) + \tilde A^{n,i}(t) - \tilde D^{n,i}(t) + \int_0^t\tilde b^{n,i}(s)ds, \; t \in [0,T],$$
where $\tilde A^{n,i}, \tilde D^{n,i}$ are as in \eqref{eq:eq928}. Define for $i = 1, \ldots, n$,
$$\tilde G^{n,i} \doteq (\zeta^{n,i}, \tilde Q^{n,i}, \tilde Y^{n,i}, \tilde R^{n,i})$$
and let
$$\Xi^n \doteq \frac{1}{n-1} \sum_{i=2}^n \delta_{\tilde G^{n,i}}.$$
Note that since by assumption $\{\tilde Q^{n,i}(0)\}_{i=1}^n$ are exchangeable and the controls $\tilde \alpha^{n,i}$
are given in terms of the same feedback function $\hat \alpha$ for each $i= 2, \ldots , n$, the processes
$\{\tilde G^{n,i}\}_{i=2}^n$ are exchangeable. Defining $\tilde G^{n,i}$ to be the zero process for $i > n$
we can regard, $\tilde G^n \doteq \{\tilde G^{n,i}\}_{i=2}^{\infty}$ as a random variable with values in
${\cal D}([0,T]: (\mathbb{R}^4)^{\otimes\infty})$. We now argue the tightness of the sequence $\{\tilde G^n\}$.
It suffices to show for each $i$, the tightness of $\{\tilde G^{n,i}\}_{n\in\mathbb{N}}$ in ${\cal D}([0,T]: \mathbb{R}^4)$.
Since
\begin{equation}
\sup_{n,i,\omega} \sup_{t\in [0,T]} \frac{\lambda^{n,i}(t,\omega)+ \mu^{n,i}(t,\omega)}{e_n} \doteq C_0 < \infty, \label{eq:eq858}
\end{equation}
the following two conditions are satisfied with $X^n$ equal to $\langle\tilde A^{n,i}, \tilde A^{n,j} \rangle$ and $\langle\tilde D^{n,i}, \tilde D^{n,j} \rangle$ for all $i,j$.
\begin{description}
\item{[A]} For each $\varepsilon>0,\eta>0$ there exists a $\delta>0$ and $n_0\in \mathbb{N}$ with the property that for every family of stopping times $\{\tau_n\}_{n\in\mathbb{N}}$ ($\tau_n$ being an ${\cal F}_t$-stopping time on $(\Omega', {\cal F}', {\mathbb P}')$) with $\tau_n\leq T-\delta$,
\begin{align*}
\sup_{n\geq n_0}\sup_{\theta\leq \delta}{\mathbb P}\{\|X^n(\tau_n)-X^n(\tau_n+\theta)\|\geq \eta\}\leq \varepsilon.
\end{align*}
\item{[T$_1$]} For every $t$ in some dense subset of $[0,T]$, $\{X^n(t)\}_{n\in\mathbb{N}}$ is a tight sequence of $\mathbb{R}$ valued random variables.
\end{description}
Then by Rebolledo's theorem (see \cite[Theorem 2.3.2]{joffe1986weak})
$\{\tilde A^{n,i}\}_{n\ge 1}$ and $\{\tilde D^{n,i}\}_{n\ge 1}$ are tight in ${\cal D}([0,T]:\mathbb{R})$ for each $i$.
Also since the jumps of $\tilde A^{n,i}$ and $\tilde D^{n,i}$ are of size $1/\sqrt{e_n}$, these processes are
${\cal C}$-tight (namely all weak limit points are continuous a.s.). From boundedness of $b$ we see that
$\{\int_0^\cdot\tilde b^{n,i}(s)ds\}_{n\in \mathbb{N}}$ is tight in ${\cal C}([0,T]:\mathbb{R})$.
Combining this with Assumption \ref{assumption4}(b) we see that $\{\zeta^{n,i}\}_{n\in \mathbb{N}}$ is ${\cal C}$-tight
in ${\cal D}([0,T]:\mathbb{R})$. Using now the continuity of the Skorohod map (Lemma \ref{lem_Skorohod}) we have the desired tightness of
$\{\tilde G^{n,i}\}_{n\in\mathbb{N}}$.
Suppose now that, along some subsequence, $\tilde G^n$ converges to $\tilde G \doteq (\zeta^{i}, \tilde Q^{i}, \tilde Y^{i}, \tilde R^{i})$, in distribution, in
${\cal D}([0,T]: (\mathbb{R}^4)^{\otimes\infty})$. Then $\tilde G \in {\cal C}([0,T]: (\mathbb{R}^4)^{\otimes\infty})$ a.s. and from \cite[Lemma 4.2]{Kotelenez2010} and the exchangeability of $\{\tilde G^{n,i}\}_{i=2}^n$ it follows that
$\{\tilde G^{i}\}_{i=2}^{\infty}$ is exchangeable and(along the subsequence),
\begin{equation}
\label{eq:eq453}
(\tilde G^n, \Xi^n) \Rightarrow (\tilde G, \Xi)
\end{equation}
in ${\cal D}([0,T]: (\mathbb{R}^4)^{\otimes\infty})\times {\cal P}({\cal D}([0,T]:\mathbb{R}^4))$ where
$\Xi \doteq \lim_{m\to \infty} \frac{1}{m-1}\sum_{i=2}^m \delta_{\tilde G^i}$.
We will now characterize the distribution of $\{\tilde Q^i\}$. From tightness of $\{\tilde Y^{n,i}\}_{n\in N}$
and $\{\tilde R^{n,i}\}_{n\in N}$ argued above and \eqref{8} it follows that
$$ \frac{1}{e_n}\int_0^t 1_{\{\tilde Q^{n,i}(s)=0\}}\mu^{n,i}(s)ds \to 0,\;\mbox{ and }
\frac{1}{e_n}\int_0^t 1_{\{\tilde Q^{n,i}(s)=L\}}\lambda^{n,i}(s)ds \to 0,$$
uniformly on $[0,T]$, in probability.
Also, from \eqref{2} it follows that
$$\sup_{0\le t \le T} \left[ \left| \frac{\lambda^{n,i}(t)}{e_n} - \hat \lambda\right| +
\left| \frac{\mu^{n,i}(t)}{e_n} - \hat \mu\right|\right] \to 0 \mbox{ a.s. }$$
as $n\to \infty$.
Thus from \eqref{eq: eq415} (and the relation $\hat \lambda = \hat \mu$), for all $i,j$,
$$
\langle\tilde A^{n,i}, \tilde A^{n,j} \rangle(t) \to \delta_{ij}\hat \lambda,\; \langle \tilde D^{n,i}, \tilde D^{n,j} \rangle(t) \to
\delta_{ij}\hat \lambda,\; \langle\tilde A^{n,i}, \tilde D^{n,j} \rangle(t) \to 0$$
in probability, uniformly on $[0,T]$, as $n\to \infty$.
By standard martingale techniques it now follows that
$$\{\tilde A^{n,i} - \tilde D^{n,i}\}_{i\ge 1} \Rightarrow \{\sigma B^i\}_{i\ge 1},$$
in $D([0,T]:\mathbb{R}^{\infty})$, where $\{B^i\}$ are mutually independent standard Brownian motions.
Also, since for $i\ge 2$
$$
(\tilde Q^{n,i},\tilde Y^{n,i},\tilde R^{n,i})(t)=\Gamma\left(\tilde Q^{n,i}(0)+\int_0^\cdot\tilde b^{n,i}(s)ds+\tilde A^{n,i}(\cdot)-\tilde D^{n,i}(\cdot) + o(1)\right)(t), \; t \in [0,T],$$
where
$$\tilde b^{n,i}(t) = b(t,\tilde Q^{n,i}(t), \hat \alpha(t,\bar\nu(t),\tilde Q^{n,i}(t), D V_{\bar \nu}(t,\tilde Q^{n,i}(t)))),$$
we have from the continuity of $b$ (Assumption \ref{assumption1}), $\hat \alpha$, and $D V_{\bar \nu}$, for $i\ge 2$,
\begin{equation}\label{eq:eq637}
(\tilde Q^i,\tilde Y^i,\tilde R^i)(t)=\Gamma\left(x+\int_0^\cdot
b(t,\tilde Q^i(t), \hat \alpha(t,\bar\nu(t),\tilde Q^{i}(t), D V_{\bar \nu}(t,\tilde Q^{i}(t))))dt +\sigma B^i(\cdot) \right)(t).
\end{equation}
Once again using standard martingale arguments it follows that for $0\le s \le t \le T$,
$B^i(t)-B^i(s)$ is independent of $\sigma\{(\tilde Q^i(u), \tilde R^i(u), \tilde Y^i(u), B^i(u)): u \le s\}$.
From weak uniqueness property noted in Section \ref{sec3b1} and the fact that $\bar \nu$ is a fixed point of $\Phi$ we now have that $\tilde Q^i$ has distribution $\bar \nu$ for $i=2,3, \cdots$.
Using the fact that $\{B^i\}$ are mutually independent, a simple argument based on Girsanov's theorem shows that $\{\tilde Q^i\}$ are mutually independent as well. This characterize the distribution
of $\{\tilde Q^i\}_{i\ge 2}$ as $\bar \nu^{\otimes \infty}$. We now have from \eqref{eq:eq453}, the definition of $\Xi$, and the law of large numbers that
$$\lim_{n\to \infty }\tilde \nu^n_{-i} = \lim_{n\to \infty}\frac{1}{n} \sum_{i=1}^n \delta_{\tilde Q^{n,i}}
= \lim_{n\to \infty}\frac{1}{n-1} \sum_{i=2}^n \delta_{\tilde Q^{n,i}} = \lim_{m\to \infty} \frac{1}{m-1}\sum_{i=2}^m \delta_{\tilde Q^i} = \bar \nu.$$
The result follows. \hfill \hfill $\Box$
\begin{remark}
\label{rem:rem640}
The above proof also shows that if $\tilde \alpha^n_{-1} = \tilde \alpha^n$, namely $\beta^n = \tilde \alpha^{n,1}$, then \eqref{eq:eq637}
holds for all $i\ge 1$ and the law of $\{\tilde Q^i\}_{i\ge 1}$ is $\bar \nu^{\otimes \infty}$.
\end{remark}
\subsection{Same strategy for all players}\label{sec4a}
In this section we prove the
equality in \eqref{20aa}.
\begin{proposition}\label{prop2}
Suppose Assumptions \ref{assumption1} and \ref{assumption4} are satisfied. Let $\tilde \alpha^n = \{\tilde \alpha^{n,i}\}_{i=1}^{\infty}$ be as in \eqref{eq:eq424}. Then for all $i\ge 1$
\begin{align}\label{20}
\lim_{n\to\iy} J^{n,i}(\tilde Q^{n}(0);\tilde \alpha^{n})= V_{\bar\nu}(0,x)
\end{align}
\end{proposition}
\noindent
{\bf Proof:}
Without loss of generality we assume $i=1$. From the proof of Proposition \ref{prop:tightemp} (see Remark \ref{rem:rem640})
$$(\tilde Q^{n,1}, \tilde Y^{n,1}, \tilde R^{n,1}) \Rightarrow (\tilde Q^{1}, \tilde Y^{1}, \tilde R^{1})$$
where the processes on the right side are given through \eqref{eq:eq637} with $i=1$.
Let
$$\hat \alpha(t, \bar \nu(t), \tilde Q^1(t) , D V_{\bar \nu}(t,\tilde Q^{1}(t))) \doteq \gamma(t, \tilde Q^1(t)), \; t \in [0,T].$$
Recall that $f$ and $g$ are bounded continuous functions and from Proposition \ref{prop:tightemp} we have that, for every $t \in [0,T]$,
$$
(\tilde \nu^n(t), \tilde Q^{n,1}(t), \alpha^{n,1}(t)) \Rightarrow (\bar \nu(t), \tilde Q^{1}(t), \gamma(t, \tilde Q^1(t))).$$
This shows that
\begin{align}
&\mathbb{E}\left[\int_0^T f(t,\tilde\nu^{n}(t), \tilde Q^{n,1}(t), \alpha^{n,1}(t))dt + g(\tilde\nu^{n}(T),\tilde Q^{n,1}(T))\right]\nonumber\\
&\quad \to \mathbb{E}\left[\int_0^T f(t,\bar \nu(t), \tilde Q^1(t), \gamma(t, \tilde Q^1(t)))dt + g(\bar \nu(T),\tilde Q^1(T))\right].\label{eq:eq656}
\end{align}
Also by continuity of $y$ and $r$
$$
\left( y(\cdot, \tilde \nu^n(\cdot)), r(\cdot, \tilde \nu^n(\cdot)), \tilde Y^{n,1}(\cdot), \tilde R^{n,1}(\cdot)\right)
\Rightarrow \left( y(\cdot, \bar \nu(\cdot)), r(\cdot, \bar \nu(\cdot)), \tilde Y^{1}(\cdot), \tilde R^{1}(\cdot)\right)$$
in $D([0,T]:\mathbb{R}^4)$.
It then follows (cf.~\cite[Lemma 2.4]{dai1996existence})
\begin{equation}
\label{eq:eq853}
\left ( \int_0^Ty(t,\tilde\nu^{n}(t))d\tilde Y^{n,1}(t), \int_0^Tr(t,\tilde\nu^{n}(t))d\tilde R^{n,1}(t)\right)
\Rightarrow \left ( \int_0^Ty(t,\bar \nu(t))d\tilde Y^1(t), \int_0^Tr(t,\bar \nu(t))d\tilde R^1(t)\right)\end{equation}
as $n\to \infty$.
Also from Lemma \ref{lem_Skorohod},
\begin{equation}\label{eq:eq905}\mathbb{E}\left [ (\tilde Y^{n,1}(T))^2 + (\tilde R^{n,1}(T))^2 \right] \le c_S^2 \mathbb{E}\left [\sup_{0\le s \le T} |\tilde \zeta^{n,1}(s)|^2\right].\end{equation}
Next note that
$$\mathbb{E} \sup_{0\le t \le T} (\tilde A^{n,1}(t))^2 \le 4 \frac{1}{e_n} \int_0^T \lambda^{n,1}(s) ds \le 4C_0T,$$
where $C_0$ is as in \eqref{eq:eq858}.
Similarly, $\mathbb{E} \sup_{0\le t \le T} (\tilde D^{n,1}(t))^2 \le 4C_0T$.
Combining these estimates
$$\sup_{n}\mathbb{E} \left[\sup_{0\le t \le T}|\tilde \zeta^{n,1}(s)|^2\right] < \infty$$
which combined with \eqref{eq:eq853} and the boundedness of $y,r$ implies
$$\sup_{n}\mathbb{E}\left[\int_0^Ty(t,\tilde\nu^{n}(t))d\tilde Y^{n,1}(t) + \int_0^Tr(t,\tilde\nu^{n}(t))d\tilde R^{n,1}(t) \right]^2 < \infty.$$
Combining this with the weak convergence in \eqref{eq:eq853} we have
\begin{align}
&\mathbb{E}\left[\int_0^Ty(t,\tilde\nu^{n}(t))d\tilde Y^{n,1}(t) + \int_0^Tr(t,\tilde\nu^{n}(t))d\tilde R^{n,1}(t) \right]\nonumber \\
&\quad
\to \mathbb{E}\left[\int_0^Ty(t,\bar \nu(t))d\tilde Y^1(t)+ \int_0^Tr(t,\bar \nu(t))d\tilde R^1(t)\right].\label{eq:eq908}
\end{align}
Combining \eqref{eq:eq656} and \eqref{eq:eq908} and recalling from the optimality of $\hat \alpha$ that
\begin{align}\label{new5}
V_{\bar\nu}(0,x)=\mathbb{E}\Big[&\int_0^T f(t,\bar\nu(t),\tilde Q^1(t),\hat\alpha(t,\bar\nu(t),\tilde Q^1(t), DV_{\bar\nu}(t,\tilde Q^1)))dt + g(\bar\nu(T),\tilde Q^1(T))\\\notag
&\quad+\int_0^Ty(t,\bar\nu(t))d\tilde Y^1(t)+\int_0^Tr(t,\bar\nu(t))d\tilde R^1(t)\Big],
\end{align}
we have the desired convergence $\lim_{n\to\iy} J^{n,1}(\tilde Q^{n}(0);\tilde \alpha^{n})= V_{\bar\nu}(0,x)$. \hfill \hfill $\Box$
\subsection{Deviation of Player 1}\label{sec4b}
In this section we prove the inequality on the right side of \eqref{20aa}.
\begin{proposition}\label{prop3}
Suppose Assumptions \ref{assumption1} and \ref{assumption4} hold.
Let $\bar \nu, \beta^n, \tilde \alpha^n_{-i}$ be as introduced at the beginning of Section \ref{sec4}. Then for each $i\ge 1$
\begin{align}\label{500}
\underset{n\to\iy}{\lim\;\inf\;} J^{n,1}(\tilde Q^{n}(0);\tilde \alpha^{n}_{-i})\ge V_{\bar\nu}(0,x)
\end{align}
\end{proposition}
{\bf Proof:}
As before, we assume without loss of generality that $i=1$. We will need to argue the tightness of the control sequence $\{\beta^n\}$ in an appropriate space.
For this it will be convenient to consider a relaxed control formulation. Consider the relaxation of the stochastic control problem in \eqref{13}--\eqref{15} where the control space $U$ is replaced by
${\cal P}(U)$, the drift function $b$ is replaced by the function $b_{{\cal R}}:[0,T]\times [0,L]\times {\cal P}(U) \to \mathbb{R}$
defined as
$$b_{{\cal R}}(t,x, r) \doteq \int_U b(t, x, u) r(du), \; (t, x, r)\in [0,T]\times [0,L]\times {\cal P}(U),$$
and the running cost function $f$ is replaced by $f_{{\cal R}}: [0,T]\times {\cal P}([0,L])\times [0,L]\times {\cal P}(U) \to \mathbb{R}$, defined as
$$f_{{\cal R}}(t,\eta, x, r) \doteq \int_U f(t, \eta, x, u) r(du), \; (t, \eta, x, r)\in [0,T]\times {\cal P}([0,L])\times [0,L]\times {\cal P}(U).$$
Also, the class of admissible controls ${\cal A}(\Xi, t,x, \bar \nu)$ is replaced by ${\cal A}_{{\cal R}}(\Xi, t,x, \bar \nu)$ of pairs $(\alpha_{{\cal R}}, Z)$ that are similar to pairs
$(\alpha, Z)$ introduced above \eqref{13} except that $\alpha_{{\cal R}}$ is ${\cal P}(U)$ valued rather than $U$ valued and \eqref{13} holds with
$\bar b(u) = b(u, X(u), \alpha(u))$ replaced with $b_{{\cal R}}(u, X(u), \alpha_{{\cal R}}(u))$. The corresponding cost function
$J_{\bar \nu, {\cal R}}$ is defined by \eqref{eq:eq234} with $f$ replaced by $f_{{\cal R}}$. The value function in this relaxed formulation, denoted as $V_{\bar \nu, {\cal R}}$,
is given by \eqref{15} with ${\cal A}$ replaced by ${\cal A}_{{\cal R}}$. Define the function $h_{{\cal R}}$ by \eqref{19}, replacing $(f,b)$ with $(f_{{\cal R}},b_{{\cal R}})$.
Then, from Assumption \ref{assumption1}(b),
$$H(t,\eta,x,p) = \inf_{u \in U} h(t,\eta, x, u,p) = \inf_{r \in {\cal P}(U)} h_{{\cal R}}(t,\eta, x, r,p).$$
This shows that $V_{\nu}$ and $V_{\nu, {\cal R}}$ are both solutions of the PDE \eqref{HJB1} - \eqref{HJB2}. In view of the uniqueness result from Proposition \ref{lem2}, $V_{\nu} = V_{\nu, {\cal R}}$.
Let $\beta^n_{{\cal R}}(t) \doteq \delta_{\beta^n(t)}$, $t \in [0,T]$ and define $\bar \beta^n_{{\cal R}} \in {\cal M}(U \times [0,T])$ as
$$\bar \beta^n_{{\cal R}} (du\, dt) \doteq \beta^n_{{\cal R}}(t)(du) dt,$$
where ${\cal M}(U \times [0,T])$ is the space of finite measures on $U \times [0,T]$ equipped with the topology of weak convergence.
Then we can rewrite
\begin{align}\label{eq:eq300}
J^{n,1}(\tilde Q^n(0), \tilde \alpha^n_{-1})\doteq &
\mathbb{E}\Big[\int_{U\times [0,T]} f(t,\tilde\nu^{n}(t), \tilde Q^{n,1}(t), u) \bar \beta^n_{{\cal R}}(du\,dt) + g(\tilde\nu^{n}(T),\tilde Q^{n,1}(T))\\\notag
&\quad+\int_0^Ty(t,\tilde\nu^{n}(t))d\tilde Y^{n,1}(t)+\int_0^Tr(t,\tilde\nu^{n}(t))d\tilde R^{n,1}(t)\Big
\end{align}
and
\begin{align}\label{eq:eq303}
(\tilde Q^{n,1},\tilde Y^{n,1},\tilde R^{n,1})=\Gamma\left(\tilde Q^{n,1}(0)+
\int_{U\times [0,\cdot]}b(s, \tilde Q^{n,1}(s), u) \bar \beta^n_{{\cal R}}(du\,ds) +\tilde A^{n,i}(\cdot)-\tilde D^{n,i}(\cdot) + o(1)\right).
\end{align}
From Proposition \ref{prop:tightemp}, $\tilde\nu^{n} \Rightarrow \bar \nu$. Also, the arguments of the same proposition show that
\begin{equation}
\label{eq:eq426}\left\{\tilde Q^{n,1}(\cdot), \tilde A^{n,1}(\cdot) - \tilde D^{n,1}(\cdot), \int_{U\times [0,\cdot]} b(s, \tilde Q^{n,1}(s), u) \bar \beta^n_{{\cal R}}(du\,ds) \right\}_{n\in \mathbb{N}}
\end{equation}
are ${\cal C}$-tight in ${\cal D}([0,T]: \mathbb{R}^3)$. Furthermore, since $U\times [0,T]$ is compact and $\bar \beta^n_{{\cal R}}(U\times [0,T]) = T$, the sequence
$\{\bar \beta^n_{{\cal R}}\}_{n\in \mathbb{N}}$ is tight in ${\cal M}(U\times [0,T])$.
Suppose now that along a subsequence (labeled once more as $\{n\}$) the sequence in \eqref{eq:eq426} along with $\{\bar \beta^n_{{\cal R}}\}$ converges in distribution to
$(\tilde Q^1, \sigma B^1, \vartheta, \bar \beta_{{\cal R}}).$
Then from the Lipschitz property of $b$ (Assumption \ref{assumption1}(a)) we have, for $t\in [0,T]$,
\begin{align*}
\vartheta(t) &= \int_{U\times [0,t]} b(s, \tilde Q^1(s), u) \bar \beta_{{\cal R}}(du ds)\\
&= \int_{0}^t b_{{\cal R}}(s, \tilde Q^1(s), \beta_{{\cal R}}(s)) ds,
\end{align*}
where $\beta_{{\cal R}}(s)$ is obtained by disintegrating $\bar \beta$, i.e.~$\bar \beta_{{\cal R}}(du\, ds) = \beta_{{\cal R}}(s)(du) ds$.
Also as in the proof of Proposition \ref{prop:tightemp} it can be argued that $B^1$ is a standard Brownian motion and thus we can conclude as in the proof of \eqref{eq:eq637}
\begin{equation}\label{eq:eq637b}
\tilde Z(t)\equiv (\tilde Q^1,\tilde Y^1,\tilde R^1)(t)=\Gamma\left(x+\int_0^\cdot
b_{{\cal R}}(t,\tilde Q^1(t), \beta_{{\cal R}}(t))dt +\sigma B^1(\cdot) \right)(t),\; t \in [0,T].
\end{equation}
Once again, by a standard martingale argument, one can argue that for $0 \le s \le t \le T$, $B^{1}(t) - B^1(s)$ is independent of
$$\tilde {\cal F}_s \doteq \sigma \left\{ \tilde Q^1(s'), \tilde Y^1(s'), \tilde R^1(s'), \bar\beta_{{\cal R}}(A \times [0,s']): 0\le s'\le s,\, A \in {\cal B}(U)\right\}.$$
Thus denoting by $(\tilde \Omega, \tilde {\cal F}, \tilde {\mathbb P})$ the probability space on which the limit processes are defined,
$\Xi = (\tilde \Omega, \tilde {\cal F}, \{\tilde {\cal F}_t\},\tilde {\mathbb P}, B^1)$ is a system and
$(\bar\beta_{{\cal R}}, \tilde Z) \in {\cal A}_{{\cal R}}(\Xi, 0, x, \bar \nu)$.
Exactly as in the proof of Proposition \ref{prop2} we see that the convergence in \eqref{eq:eq908} holds. Also using the weak convergence
$$(\tilde \nu^n, \tilde Q^{n,1}, \bar \beta^n_{{\cal R}}) \Rightarrow (\bar \nu, \tilde Q^1, \bar \beta_{{\cal R}})$$
and the Lipschitz property of $f$ in Assumption \ref{assumption1}, we have
\begin{align}
&\mathbb{E}\left[\int_{U\times [0,T]} f(t,\tilde\nu^{n}(t), \tilde Q^{n,1}(t), u)\bar \beta^n_{{\cal R}}(du\, dt) + g(\tilde\nu^{n}(T),\tilde Q^{n,1}(T))\right]\nonumber\\
&\quad \to \mathbb{E}\left[\int_0^T f(t,\bar \nu(t), \tilde Q^1(t), u)\bar \beta_{{\cal R}}(du\, dt) + g(\bar \nu(T),\tilde Q^1(T))\right].\label{eq:eq504}
\end{align}
Combining the above convergence properties, we have as $n \to \infty$,
$$J^{n,1}(\tilde Q^n(0), \tilde \alpha^n_{-1}) \to J_{\bar \nu, {\cal R}}(0, x, \bar \beta_{{\cal R}}, \tilde Z) \ge V_{\bar \nu, {\cal R}}(0,x) = V_{\bar \nu}(0,x).$$
Since the above holds for an arbitrary convergent subsequence of processes in \eqref{eq:eq426} and the sequence $\{\bar \beta^n_{{\cal R}}\}$, the result follows. \hfill \hfill $\Box$
\vspace{\baselineskip}
\noindent
{\bf Acknowledgment.}
We are grateful to two anonymous referees for their suggestions that improved the presentation of the paper. We will also like to thank the associate editor for the careful reading of the manuscript, valuable comments, and pointing out a key error in the convergence proof of the numerical scheme presented in a previous version of the paper. The corrected version of the statement and proof of the convergence result is given in \cite{BBC2017}.
\small{
\bibliographystyle{abbrv}
|
\section{Introduction}
The last decade has witnessed the emergence of a new scheme to understand the structure of disc galaxies (e.g.
\citealt{erwin05a,pohlen06}). Following this picture, disc galaxies can be classified according to the different slopes they
present in their inner and outer regions. Galaxies which are best described by purely exponential surface-brightness
profiles \citep[e.g.][]{bland-hawthorn05, weiner01} are classified as Type I. Type II galaxies are those where the
surface-brightness profile is characterised by a broken exponential with a steep outer and a shallower inner region
\citep[e.g.][]{pohlen02}. Finally, Type III are disc galaxies the surface-brightness profiles of which are anti-truncated, i.e. where the exponential decline of surface brightness is steeper in the inner
region and shallower at larger galactocentric distances \citep[e.g.][]{erwin05a}.\\
The emergence of Type II galaxies has been studied by a number of authors using numerical simulations (e.g. \citealt{li06, bournaud07, foyle08, roskar08, martinez-serrano09, sanchez-blazquez09}). To date, the general perception is that truncations that are not related to Lindblad-resonances (see e.g. \citealt{debattista06}) are the consequence of stellar migration and a radial star-formation threshold as the gas disc becomes thinner at large galactocentric radii. However, it is still being debated whether the other surface-brightness profile types also formed through internal mechanisms or are a consequence of external galaxy evolutionary processes. Theoretical work by e.g. \citet{yoshii89} and \citet{elmegreen06} shows that in principle, single-exponential and antitruncated discs can form ab initio, albeit under very specific conditions. However, it is worth noting that \citet{younger07} have shown that antitruncations can be the consequence of external processes like tidal interactions and minor galaxy mergers. Furthermore, in S0 galaxies, antitruncated profiles might also be related to an extended bulge component (e.g. \citealt{erwin05, maltby15}).\\
The first attempt to quantify the frequency of each profile type was done by \citet{pohlen06} (hereafter PT06). Using $\sim$90 nearby late-type (Sb-Sdm) spiral galaxies, they found that only $\sim$10\% are single-exponentials. The rest are $\sim$60\% Type II and $\sim$30\% anti-truncated galaxies. This study was conducted using disc galaxies of different environments. \citet{erwin08}, using a sample of 66 barred S0-Sb disc galaxies from different environments, found a distribution of 27\%, 42\% and 24\% of Type I, Type II and Type III profiles (with 6\% combinations of Type II and III). In a follow-up to this work (\citealt{gutierrez11}), the authors investigated a sample covering the full morphological range of disc galaxies (i.e. S0-Sm) and reported a correlation of morphology and surface-brightness profile type; Type I and Type III galaxies were found to be most frequent in early-type discs while the fraction of Type II profiles was found to increase with Hubble type, i.e. to be higher in late-type discs. Overall, they reported a distribution of 21\%, 50\% and 38\% of Type I, Type II and Type III profiles, with 8\% combined Type II and III galaxies that were counted twice.\\
A potential
change on the structural break properties of the disc galaxies is expected when comparing the field with the cluster
environment. There are many physical mechanisms in high density regions that should be particularly relevant for affecting
the outermost (weakly bound) zones of galaxies; tidal galaxy-galaxy interactions, tidal interactions between galaxies in high-density cluster regions and with the gravitational potential of the cluster (i.e. galaxy harassment) as well as hydrodynamical interactions between galaxies and the intra-cluster or intra-group medium such as ram-pressure stripping (see e.g. \citealt{toomre72, gunn72, moore96, lopes14, head14, hiemer14}).\\
The effect of the environment on the structure of galaxy discs is still a debated question and recent analyses have led to somewhat incongruent conclusions. For a sample of spiral galaxies, \citet{maltby15}
explored whether the frequency of each profile type changes in a cluster environment, finding $\sim$10\% Type I, $\sim$50\% Type II and $\sim$40\% Type III galaxies in their cluster and field samples. The authors also investigated a sample of S0 cluster and field galaxies, and found $\sim$25\% Type I, $<$5\% Type II and $\sim$50\% Type III galaxies with $\sim$20\% of the profiles exhibiting general curvature and hence remaining unclassified. Comparing their field and cluster galaxies, the authors concluded that the stellar distribution in the outer regions of disc galaxies is not significantly affected by the galaxy environment. However, \citet{erwin12} reported that truncated (Type II) S0 galaxies are nonexistent in Virgo Cluster while they account for roughly one third of S0s in the local field. The difference in the cluster was found to be almost entirely compensated by Type I galaxies. The authors reasoned different mechanisms driving the structural evolution of galaxies in the cluster and field environment. Other works have also pointed to environment-mediated mechanisms and their effect on the structural
break properties of disc galaxies \citep[e.g.][]{roediger12,laine14,head14}.\\
\indent In this paper we want to investigate in detail whether the environment plays a role on the frequency of each profile
(break) type. To isolate as much as possible the effects produced by the cluster environment, we select our sample of
field and cluster galaxies to have the same redshifts and stellar mass ranges. We choose a narrow mass range since it has been shown that the structural parameters of disc galaxies change with mass (see PT06) and we want to minimise this effect in our analysis. We will show that the main effect of the cluster
environment is to redden by around $\sim$0.2 mag the (g-r) colour and to decrease the global size (as parameterised by the
effective radius $R_e$) of the discs by $\sim$15\%. These two global changes are accompanied by an increase (by a factor of
$\sim$2.5) in the fraction of Type I (pure exponential) disc galaxies in the cluster regions and an increase in the
(outer) scale lengths of Type I ($\sim$8\%) and Type III ($\sim$16\%) profiles.\\
In the subsequent section (Section \ref{sec:data}) we present a description of the data including a discussion of our selection criteria, the sample compilation and the background subtraction techniques. Section \ref{sec:fitting} describes the methods used for structural galaxy fitting and for colour determination. The results are presented in Section \ref{sec:results} and discussed in Section \ref{sec:discussion}. We summarise the results and conclusions in Section \ref{sec:summary}. Finally, we present prototypical profiles, further details of the analysis of the field sample and comprehensive data tables in an appendix to this paper.
\noindent Throughout this paper we assume $H_{0}$=70 km/s/Mpc, $\Omega_{m}$=0.3 and $\Omega_{\Lambda}$=0.7.
\begin{figure}
\vspace{-0.5cm}
\centering
\includegraphics[width=0.95\columnwidth]{hist_r200.pdf}
\caption{Distribution of the $r_{200}$ values for all 130 host clusters.}
\label{fig:hist_r200}
\end{figure}
\begin{figure*}
\vspace{0.25cm}
\centering
\includegraphics[width=0.75\textwidth]{adjust_hist_giant_fin.pdf}
\caption{\textit{Left panel:} normalised redshift distribution of the total cluster and total field sample. \textit{right panel:} normalised distribution of stellar mass for the total cluster and total field sample.}
\label{fig:1}
\end{figure*}
\section{The data}
\label{sec:data}
To conduct our work, we used the NYU Value-Added Galaxy Catalogue (NYU-VAGC, \citealt{blanton05}) based on the SDSS-DR7
\citep{abazajian09} as the basis for our parent sample. This catalogue provides spectroscopic redshifts and photometry, global S\'{e}rsic-indices and corresponding effective radii (see \citealt{blanton05a} for a detailed description of the morphological fitting technique) as well as stellar masses (\citealt{blanton07}) calculated on the basis of a \citet{chabrier03} IMF and a population synthesis model
from \citet{bruzual03}. Our parent sample is analogue to the one of \citet{cebrian14}. First, we selected only the galaxies contained within the region of the survey described by \citet{varela12}, in order to avoid problems with the borders of the survey. Then, we only considered the objects above the mass-completeness limit presented in \citet{cebrian14} (see Eq. 1 from that paper). This ensures that our sample is complete in stellar mass, avoiding biases due to the magnitude limit inherent to the SDSS catalogue. To assure that our initial sample of galaxies are predominantly disc-dominated and within a
narrow range of stellar masses, we took only objects with S\'{e}rsic-index n $<$ 2.5 and 0.8 $\times$ 10$^{10}$ $M_{\odot}$ $<$ $M_{\star}$ $<$ 4 $\times$ 10$^{10}$ $M_{\odot}$ The lower mass limit corresponds to our completeness limit while the higher mass limit was chosen to minimise the influence of stellar mass on the results of our study. \\
We also estimated, for all the galaxies, the 3D spatial (X, Y, Z) position within
the survey using the information provided by the R.A., Dec and redshift of each object. To do this, we used the following set of
equations provided by \citet{varela12}:
\begin{equation} \label{eq:cartesian}
\begin{array}{l}
X=D(z)\cos\delta\cos\alpha \\
Y=D(z)\cos\delta\sin\alpha \\
Z=D(z)\sin\delta \\
\end{array}
\end{equation}
\noindent with $\alpha$, $\delta$ and $D(z)$ being the equatorial coordinates and the comoving radial distance, respectively. The spatial distribution of the galaxies is used to explore the environment inhabited by the objects.
\subsection{The cluster sample}
Following \citet{cebrian14}, we compiled a large sample of galaxy clusters in our explored volume using a number of catalogues: the
Abell catalogue \citep{abell89}, a catalogue extracted from SDSS-DR6 \citep{szabo11}, three catalogues from SDSS III \citep{einasto12, tempel12, wen12}; the GMBCG cluster catalogue \citep{hao10}, and the XMMi-SDSS
galaxy cluster survey \citep{takey11}. This is a total of 1877 galaxy clusters.\\
To build the sample of galaxies in clusters, we took only those galaxies from the parent sample that are located at a clustercentric distance
less than 1 Mpc to the nearest galaxy cluster centre. Since the goal of this work is to conduct a detailed analysis of the structural properties of
the disc of the galaxies, we selected the objects with the lowest redshifts. These selection criteria left us with 246 catalogue galaxies. To minimise
the influence of dust and to ensure the reliability of morphological information, we followed PT06 in selecting face-on to
intermediately inclined galaxies and we discarded close pairs and obvious galaxy mergers after a visual inspection. After doing this,
we were left with 175 galaxies in 130 different clusters (25 clusters from \citealt{abell89}, 18 clusters from \citealt{szabo11}, 72
clusters from \citealt{tempel12} and 15 clusters from \citealt{wen12}). The sample of 175 galaxies will henceforth be referred to as the total cluster sample. The selection of a reasonable number of cluster galaxies finaly resulted in a redshift range of 0.021$<$z$<$0.063. All galaxies in the total cluster sample are listed in the appendix. The $r_{200}$\footnote{Values adopted from the cluster catalogues, except for Abell-clusters for which values were taken from the literature.}-distribution of the 130 host clusters is shown in Fig. \ref{fig:hist_r200}.
\subsection{The field sample}
In order to compare the properties of disc galaxies in the cluster environment to those of disc galaxies in the low-density to
intermediate-density field environment, we created a sample of field galaxies drawn from the same parent galaxy catalogue as the
initial cluster sample. In a first step, we confined the parent catalogue to galaxies with a 3D spatial distance greater than 3.5 Mpc
to the nearest cluster centre to ensure our objects are beyond the virial radius of the nearest cluster. In order to have a field sample with stellar mass and redshift distribution similar to
those distributions in the cluster sample, we generated random samples of 246 field galaxies (the same initial number as for the cluster galaxies),
and selected three samples closely resembling the initial cluster sample. As we did for the
cluster sample, we selected face-on to intermediately inclined galaxies and discarded close pairs and galaxy mergers after a by-eye
inspection. We were left with 172, 172 and 177 galaxies in the three field samples (no overlap). They will henceforth be referred to as field sample 1, field sample 2 and field sample 3. By joining the three field samples we constructed a large field sample of 521 galaxies which will henceforth be referred to as the total field sample. All galaxies in the total field sample are listed in the appendix. The redshift and stellar mass distributions of the total cluster and field samples are compared in Fig. \ref{fig:1}. The corresponding illustrations for the individual field samples are shown in Fig. \ref{fig:1a}.\\
We explored whether the global size of our field galaxies is representative of the general field population. With the redshift and
stellar mass limits of our cluster sample, we had 14868 field galaxies. Their median size (as parametrised by $R_{e}$) is larger than the median size of the cluster sample by
$\sim$13\%. Our field samples have median global sizes larger than the median size of our total cluster sample by $\sim$15\%.
This difference in global size is in compliance with \citet{cebrian14}.
All analyses presented in this paper were carried out on each of the initially selected samples (four in total). We use the three field samples to test for effects possibly introduced by random sampling and to estimate the robustness of the observed differences between the cluster and the field population.
\begin{figure*}
\vspace{0.25cm}
\centering
\includegraphics[width=\textwidth]{fig2_fin2_ID.pdf}
\caption{Prototypical examples for each galaxy type in both environments. \textit{Left column, top to bottom:} Type I, Type II (double break) and Type III surface-brightness profiles and corresponding r-band cluster galaxy images. \textit{Right column, top to bottom:} Type I, Type II and Type III surface-brightness profiles and corresponding r-band field galaxy images. The over-plotted short-dashed lines in the profile plots indicate the exponential disc regions resulting from the fitting routine. The short vertical lines represent the break radii in the different bands. The horizontal dashed line indicates our conservative confidence threshold. For scale reference, the purple ellipses in the images mark the mean radius corresponding to this threshold, indicated by a solid vertical line in the profile plots. The galaxies' NYU-VAGC-ID is inserted in the images.
}
\label{fig:2}
\end{figure*}
\subsection{Sky subtraction and limiting surface brightness}
As this work aims to explore the outer regions of disc galaxies to characterise their properties, it is necessary to have an
accurate sky subtraction and an estimation of the limiting surface brightness. Using the same SDSS imaging, PT06 showed that surface
brightness profiles in the g- and r-band can be reliably traced down to 27 mag/arcsec$^2$ (see also \citealt{trujillo16}). In order to
reach such accuracy, we followed the same strategy as PT06 to obtain an estimate of the sky from as close as possible to the galaxies under study. First, we
measured the mean sky after three 3-$\sigma$ clippings within 5 large rectangular sky boxes. Chosen sky boxes were placed as close as possible to the respective galaxy with the aim of including as little source flux as possible. This necessitates a variability in box size according to secondary source conditions unique to each pointing (typically in the range of 300-500 pixels on a side), with the total area fixed at 160k pixels per box. Any remaining extended sources within a sky box were masked out. For further details of this methodology, see PT06. As a second method, we investigated the radial
profile of the galaxy image pixel counts (with masked adjacent sources) in 120 angular directions, separated by 3 degrees. We controlled whether the sky value
measured with the first method was in compliance with the averaged value at which the radial profiles were visually found to flatten out. This
was true within $\pm$0.16 counts for all galaxies in our total field and cluster samples, both for g- and r-band data. For the photometric calibration we followed again PT06. This error in the sky determination corresponds to $\sim$16\% of the sky and to a surface brightness of $>$27 mag/arcsec$^{2}$. Since we do not attempt to extend our analyses beyond this value but limit ourselves to a conservative threshold of 26 mag/arcsec$^{2}$, this uncertainty has no effect on the results presented in this work.
\section{Structural galaxy fitting}
\label{sec:fitting}
In order to characterise the structural (break) properties of our sample of 696 disc galaxies (175 cluster and 521
field), we used the surface-brightness
fitting code \texttt{IMFIT} (\citealt{erwin15}). We applied this code to both, the g- and r-band data. The images provided to \texttt{IMFIT}
were masked and sky-subtracted. The masks were created using \texttt{SExtractor} (\citealt{bertin96}) in a hot and cold configuration
mode. To account for the effect of the Point Spread Function (PSF), we built a PSF for each SDSS image frame using \texttt{PSFExtractor}
\citep{bertin11} to estimate the FWHM of the PSF. Then, we used the image generator of the \texttt{IMFIT} package, \texttt{MAKEIMAGE}, to generate Moffat-PSF-images (SDSS standard $\beta$-values of 3.1 and 2.9 for the r- and g-band; see e.g. \citealt{trujillo01, erwin15} for more details) of
51$\times$51 pixels ($\sim$20\arcsec$\times$20\arcsec). The surface-brightness distributions of the galaxies in our sample were modelled
using a two-dimensional S\'ersic-bulge and a broken exponential (\citealt{erwin08}) for the disc component. These models were convolved with the
PSF.\\
\noindent
All 696 galaxies of the total cluster and field sample were successfully modelled by \texttt{IMFIT} in the g- and r-band. Each of the 1392 fits was set up individually, i.e. for each galaxy image we
created an individual configuration file with initial parameter values based on visual estimates after a by-eye inspection of the
azimuthally averaged 2D projected surface-brightness profile. The majority of the fits ran without any problems. For $\sim$30\% of the galaxies
it was necessary to further refine the initial fitting parameters. Approximately 40\% of these were to ensure that the innermost region of the galaxy (i.e. the bulge) was properly modelled by the
S\'{e}rsic part, while the disc was to be (dominantly) modelled by the broken exponential function to exclude potential degeneracies. For the remaining 60\%, the
profiles showed more than one break in the galaxy's disc region. In these cases, we manually forced \texttt{IMFIT} to model the
\textit{outermost} break above our limiting surface brightness. The profile types of the galaxies where the fit initially failed were Type II ($\sim$76\%) and III ($\sim$24\%).\\
For 56 galaxies (27 in the total cluster sample, 10, 10 and 9 in the field samples), the best fitting result was obtained for $h_{1}$=$h_{2}$
(i.e. the scale lengths in and outside the break have the same value). These galaxies were classified as Type I, following PT06. Another
199 galaxies (50 in the total cluster sample, 47, 50 and 52 in the field samples) showed an ``up-bending" or ``anti-truncated" profile, i.e. $h_{1}<h_{2}$, and were
classified as Type III. The remaining 441 galaxies (98 in the total cluster sample, 115, 112 and 116 in the field samples) exhibited a
"down-bending" profile, i.e. $h_{1}>h_{2}$, and thus were classified as Type II. The profile type number counts are the same in both bands. Depending on the analysed band, we notate the
structural parameters as follows: $h_{1,g}$, $h_{2,g}$, $R_{b,g}$ and $h_{1,r}$, $h_{2,r}$, $R_{b,r}$, where $R_{b,g}$ and $R_{b,r}$ denote the break radii. Profile type fractions and average parameter values are shown in Table \ref{table:results}. In Fig. \ref{fig:2} we show prototypical examples for the surface-brightness profiles and the fitted scale lengths for all three types in both investigated environments. \\
In PT06, the authors applied several sub-classifications to Type II and Type III galaxies, based on their Hubble-type (barred vs. non-barred
galaxies) and on the radial position of the break. Since the galaxy sample analysed for the present work are at a higher redshift
(0.021$\leq$z$\leq$0.063) than the sample analysed by PT06 (z$\leq$0.01), it was impossible to robustly determine the Hubble-type for all
galaxies in our study. Thus, we did not apply any further sub-classification for Type II and Type III galaxies.
\subsection{Galaxy colour}
\label{sec:colour}
For all 696 galaxies in the combined total cluster and field sample, we measured (g-r) restframe colour, $(g-r)_{restframe}$, both
outside ($(g-r)_{o}$) and inside ($(g-r)_{i}$) the break radius $\overline{R_{b}}$=$(R_{b,g}+R_{b,r})/2$. The k-correction was conducted following \citet{chilingarian10}. In addition, to explore the effect of the bulge on the colour properties
of the inner regions of the galaxies, we repeated the above colour measurements but this time masking the inner $R_{c1}$=0.5$R_{e}$ (obtaining $(g-r)_{i,1}$) and
$R_{c2}$=0.75$R_{e}$ (obtaining $(g-r)_{i,2}$). For all investigated galaxies $R_b > R_{c1}$ and $R_b > R_{c2}$. The masking of the bulge regions results in a de-reddening of galaxy colours which is in general more pronounced in field galaxies and appears to be strongest in field Type II objects. This was to be expected, since Type II galaxies are predominantly late-type objects, which exhibit comparatively strong star-formation activity in their discs. The median colour values for the total cluster and total field sample are shown in Table \ref{table:results}, along with the other disc properties studied here.
\begin{table*}
\vspace{0.25cm}
\caption{Listing of the median values of the structural properties of the galaxies analysed in this work. The total cluster and total field sample
are abbreviated by TCS and TFS, respectively. The ratios of the inner and outer scale lengths in the g- and r-band surface-brightness
profiles are listed as $ratio_{g}$ and $ratio_{r}$, respectively. Median colour values determined between $\overline{R_{b}}$ and
$R_{c1}$ are given as $(g-r)_{in,1}$, those determined between $\overline{R_{b}}$ and $R_{c2}$ as $(g-r)_{in,2}$ (see Section \ref{sec:colour}). Type I galaxies were not considered for the calculation of the median $(g-r)_{out}$ colour for the total samples. We also list the
number of galaxies in each sub-sample, N, and its percental equivalent (along with the corresponding \citealt{wilson27}-confidence intervals), as well as the median global effective radius $R_{e}$, median S\'{e}rsic-index $n$ and median stellar mass $M_{\star}$. All errors were estimated through 1000 1-$\sigma$ bootstrap
iterations. Field values with an asterisk indicate significant differences between the cluster and the field as estimated by a standard KS-test (P-value $<$0.05).}
\label{table:results}
\centering
\begin{tabular}{c|ccc|c||ccc|c}
\hline
\hline
\multicolumn{1}{c|}{} &
\multicolumn{4}{c||}{Cluster} & \multicolumn{4}{c}{Field}\\
\hline
Sample & Type I & Type II & Type III & TCS & Type I & Type II & Type III & TFS \\
\hline
N & 27 & 98 & 50 & 175 & 29 & 343 & 149 & 521 \\
\% & 15$_{-4}^{+7}$ & 56$_{-7}^{+7}$ & 29$_{-7}^{+7}$ & 100 & 6$_{-2}^{+2}$ & 66$_{-4}^{+4}$ & 29$_{-4}^{+4}$ & 100 \\
\hline
$R_{e}$ & 2.20$\pm$0.13 & 2.79$\pm$0.09 & 2.02$\pm$0.08 & 2.44$\pm$0.07 & 2.41$\pm$0.14 & 3.20$\pm$0.06* & 2.22$\pm$0.07 & 2.88$\pm$0.05* \\
$n$ & 2.07$\pm$0.04 & 1.83$\pm$0.04 & 2.17$\pm$0.03 & 1.99$\pm$0.03 & 1.60$\pm$0.06* & 1.63$\pm$0.02 & 2.00$\pm$0.03* & 1.73$\pm$0.02* \\
$M_{\star}$ [10$^{10}M_{\odot}$] & 1.52$\pm$0.10 & 1.53$\pm$0.05 & 1.89$\pm$0.11 & 1.63$\pm$0.05 & 1.55$\pm$0.14 & 1.60$\pm$0.03 & 1.55$\pm$0.06 & 1.59$\pm$0.03 \\
\hline
$(g-r)_{in}$ & 0.504$\pm$0.022 & 0.363$\pm$0.015 & 0.534$\pm$0.014 & 0.455$\pm$0.015 & 0.238$\pm$0.014* & 0.266$\pm$0.007* & 0.392$\pm$0.012* & 0.288$\pm$0.006* \\
$(g-r)_{in,1}$ & 0.504$\pm$0.022 & 0.334$\pm$0.015 & 0.506$\pm$0.014 & 0.434$\pm$0.011 & 0.222$\pm$0.012* & 0.235$\pm$0.007* & 0.362$\pm$0.012* & 0.265$\pm$0.006* \\
$(g-r)_{in,2}$ & 0.504$\pm$0.023 & 0.315$\pm$0.016 & 0.500$\pm$0.014 & 0.426$\pm$0.011 & 0.208$\pm$0.013* & 0.217$\pm$0.007* & 0.354$\pm$0.012* & 0.247$\pm$0.006* \\
$(g-r)_{out}$ & - & 0.300$\pm$0.020 & 0.482$\pm$0.026 & 0.363$\pm$0.015 & - & 0.201$\pm$0.010* & 0.288$\pm$0.015* & 0.231$\pm$0.008* \\
\hline
$h_{1,g}$ [kpc] & 1.55$\pm$0.08 & 2.75$\pm$0.14 & 1.38$\pm$0.06 & - & 1.47$\pm$0.05 & 2.85$\pm$0.07 & 1.37$\pm$0.04 & - \\
$h_{2,g}$ [kpc] & -"- & 1.37$\pm$0.05 & 2.18$\pm$0.11 & - & -"- & 1.46$\pm$0.03 & 1.83$\pm$0.05* & - \\
$ratio_{g}$ & 1 & 1.91$\pm$0.09 & 0.64$\pm$0.02 & - & 1 & 1.87$\pm$0.04 & 0.74$\pm$0.01* & - \\
$R_{b,g}$ [kpc] & - & 5.21$\pm$0.19 & 4.68$\pm$0.20 & - & - & 5.33$\pm$0.10 & 4.41$\pm$0.15 & - \\
$h_{1,r}$ [kpc] & 1.57$\pm$0.09 & 2.66$\pm$0.12 & 1.37$\pm$0.06 & - & 1.45$\pm$0.05 & 2.66$\pm$0.06 & 1.32$\pm$0.04 & - \\
$h_{2,r}$ [kpc] & -"- & 1.37$\pm$0.04 & 2.12$\pm$0.08 & - & -"- & 1.38$\pm$0.02 & 1.87$\pm$0.05* & - \\
$ratio_{r}$ & 1 & 1.71$\pm$0.10 & 0.63$\pm$0.02 & - & 1 & 1.69$\pm$0.04 & 0.71$\pm$0.01* & - \\
$R_{b,r}$ [kpc] & - & 5.72$\pm$0.19 & 4.74$\pm$0.20 & - & - & 5.53$\pm$0.10 & 5.22$\pm$0.15 & - \\
\hline
\hline
\end{tabular}
\vspace{0.5cm}
\end{table*}
\begin{figure}
\vspace{-0.5cm}
\centering
\includegraphics[width=0.95\columnwidth]{dist_norm.pdf}
\caption{Normalised radial distributions of the different profile types in the total cluster sample (green solid, blue dashed and red short-dashed lines for Type I, II and III, respectively). The short vertical lines indicate the corresponding median values.}
\label{fig:dist_norm}
\end{figure}
\section{Galaxy properties as a function of environment}
\label{sec:results}
\subsection{Global properties}
There are several global results to highlight in this section. In relation to the global (break) structure: a) the number of Type I (single-exponential) galaxies in the cluster environment
is significantly higher (by a factor of 2.5) than in the field; b) the number of Type II (truncated) galaxies is lower in the cluster environment by 10 percentage points, matching the increase of Type I objects; c) the number of Type III (anti-truncated) galaxies, however, is quite similar in both environments. It is worth noting that the average radial distribution of the different surface-brightness profile types within the clusters is
different, with Type III galaxies residing at a median clustercentric distance (normalised to $r_{200}$) of 0.58$\pm$0.05, while the other types preferentially occupy regions at comparatively larger median normalised distances of 0.98$\pm$0.09 and 0.86$\pm$0.04 for Type I and II, respectively (errors estimated via 1000 1-$\sigma$ bootstrapping iterations). The normalised radial distribution of the different profile types is illustrated in Fig. \ref{fig:dist_norm}.
The global size value, as mentioned before, measured by $R_{e}$, is $\sim$15\% larger in the field than in the cluster. Finally, also the global
S\'{e}rsic-index (as provided by the NYU catalogue) is significantly higher in the cluster than in the field (by $\sim$15\%). By construction of our
subsamples, there is no global difference between cluster and field environment in average total stellar mass. The global structural and colour differences between the cluster and the field environment are illustrated in Fig. \ref{fig:firsthist}.\\
\noindent Independent of the surface-brightness profile type, in all the regions, the median $(g-r)_{restframe}$ colour is significantly ($\sim$0.2
mag) redder in the cluster than in the field (most significant for Type I and III). For Type II and III field galaxies, with the bulge suitably masked, we find the inner disc to be notably redder than the regions at galactocentric distances greater than the break radius. In the cluster sample no such difference is seen within the errors. This finding indicates that the reddening in the cluster is stronger in the outer parts of the disc, which is in compliance with the scenario of disc-fading (see e.g. \citealt{christlein04}), i.e. the fading/aging of a galaxy's (outer) disc consequent to the stripping of gas by ram-pressure or to the consumption of gas through star formation (i.e. ``strangulation").
Since we have selected our samples to minimise the influence of dust (see Section \ref{sec:data}), we can assume that the observed reddening is truly caused by older stellar populations. Even though our data is not good enough to robustly determine the morphological type of the galaxies, we can further assume that the reddening is not solely related to a different morphological mix (i.e. a higher fraction of S0s) in the cluster environment; it has been shown that there exist hardly any Type II S0 galaxies in clusters (\citealt{erwin12,maltby15}), however, the reddening we detect is similarly seen in Type II and Type III galaxies. Moreover, given our selection limit in S\'{e}rsic-index and stellar mass, our sample is very likely biased towards late-type galaxies (see e.g. \citealt{ravindranath04}) and against S0s, which are among the most massive disc objects. The reddening observed in Type II and III cluster galaxies is slightly stronger in the outer disc regions, which is in compliance with galaxy evolutionary scenarios invoking ram-pressure stripping (see e.g. \citealt{steinhauser16}) or stellar migration (see e.g. \citealt{roskar08}). A summary of the global properties discussed here is illustrated in Fig. \ref{fig:3}.
\begin{figure*}
\centering
\vspace{0.5cm}
\includegraphics[width=0.725\textwidth]{firsthist.pdf}
\caption{Normalised distributions of colour, $R_{e}$ and S\'ersic-index for the total cluster and field samples (field galaxies in the background in blue colour, cluster galaxies in the foreground in light-red colour, overlapping regions of the histograms in purple). The vertical lines indicate the median values (dash-dotted for field and dashed for cluster galaxies). The KS-test P-value is inserted in each panel. Type I galaxies are not included in the bottom-left panel.}
\label{fig:firsthist}
\end{figure*}
\begin{figure*}
\centering
\vspace{0.5cm}
\includegraphics[width=0.725\textwidth]{hist1norm_p.pdf}
\caption{Normalised distributions of the measured disc properties for Type I galaxies (field galaxies in blue, cluster galaxies in red). The vertical lines indicate the median values (dash-dotted for field and dashed for cluster galaxies). The KS-test P-value is shown in the upper right corner of each panel.}
\label{fig:hist1}
\end{figure*}
\begin{figure*}
\centering
\vspace{0.5cm}
\includegraphics[width=\textwidth]{hist2norm_p.pdf}
\caption{Same as Fig. \ref{fig:hist1} for Type II galaxies.}
\label{fig:hist2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{hist3norm_p.pdf}
\caption{Same as Fig. \ref{fig:hist1} for Type III galaxies.}
\label{fig:hist3}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{fig3_nacho3new2.pdf}
\caption{Median galaxy properties for both environments and all three surface-brightness profile types. Errors are 1-sigma bootstrap estimates (1000 iterations). Open symbols and thin error bars represent the individual field samples (1, 2 and 3 from left to right, respectively). \textit{Top row:} effective radius, $R_{e}$, and $(g-r)_{restframe}$ colour. For Type II and Type III galaxies, $(g-r)_{in,1}$ and $(g-r)_{out}$ are plotted. For Type I galaxies, we plot only $(g-r)_{in,1}$, which in this case is the colour measured in the entire disc region (bulge masked at 0.5 $R_{e}$). The horizontal lines in the first panel indicate the median $R_{e}$ (and bootstrap errors) for the total cluster and total field sample, respectively. \textit{Middle row:} inner ($h_{1}$) and outer ($h_{2}$) exponential scale length for both measured bands. \textit{Bottom row:} Scale length ratio ($h_{1}$/$h_{2}$) and break radius ($R_{b}$) for both measured bands. For Type I galaxies, the break radius is not defined and the scale length ratio is by definition equal to unity.}
\label{fig:3}
\end{figure*}
\subsection{The different profile types}
In total, the fractions of Type I, II and III galaxies in the total cluster and field sample are 15$_{-4}^{+7}$\%, 56$_{-7}^{+7}$\% and 29$_{-7}^{+7}$\% and 6$_{-2}^{+2}$\%, 66$_{-4}^{+4}$\% and 29$_{-4}^{+4}$\%, respectively (errors correspond to \citet{wilson27}-confidence intervals). In Table \ref{table:results} we list these fractions along with the median values of the structural parameters for the galaxies in both environments (numbers with an asterisk indicate significant differences between cluster and field, according to a standard KS-test P-value $<$0.05). We plot the distributions of the measured parameters for the different profile types in Figs. \ref{fig:hist1} to \ref{fig:hist3} and illustrate their median values in Fig. \ref{fig:3}. The results for the three individual field samples can be found in Table \ref{table:resultsfc} for comparison. Note that all trends reported in this paper are confirmed by the evaluation of the individual field samples.
\subsubsection{Type I galaxies}
The most interesting result in relation to this type of galaxies is the potentially larger scale length (a factor of $\sim$1.08) in the clusters environment compared to the field . However, this result is not significant according to standard KS-testing (P-value: 0.481 and 0.361 for r- and g-band, respectively). Moreover, Type I cluster galaxies are as red as Type III cluster galaxies while in the field Type I colours are comparable to Type II field objects (see Figs. \ref{fig:hist1} and \ref{fig:3}). Note that for Type I galaxies we measure the colour in the entire disc region.
\subsubsection{Type II galaxies}
The inner and outer scale length, $h_1$ and $h_2$, (in both bands) are similar in clusters and in the field. In relation to the position of the break, $R_b$, there is not an obvious trend
depending on environment. The most remarkable issue is that the location of the break position is independent of the band for the field galaxies (within the errors) but rather different in the clusters. However, the observed difference is not significant according to standard KS-testing. The decrease in $R_e$ (field to cluster) is strongest for Type II cluster galaxies. For an illustration of these results, (see Figs. \ref{fig:hist2} and \ref{fig:3}).
\subsubsection{Type III galaxies}
Following the trend found in Type I galaxies, the effect of the cluster environment in Type III galaxies is to significantly enlarge
the value of the outer scale length, $h_2$ (KS-test P-values 0.007 and 0.001 for r- and g-band, respectively). For Type III galaxies, the increase is by a factor of $\sim$1.16. We also find that the ratio of the inner-to-outer scale lengths is smaller in the cluster environment by $\sim$15\% (for both bands). These findings could be interpreted as tidal effects consequent to galaxy harassment (see \citealt{moore96} for a detailed definition), as an increase of the rate of minor mergers building up the outer disc or as an increased contribution of an extended bulge component at larger galactocentric radii. However, the latter scenario is unlikely since it has been found to occur preferentially in S0 galaxies (see \citealt{maltby12b}) and our sample is biased against those.
While merger events could also explain why Type III galaxies in the cluster are more massive than those in the field (see \ref{table:results}), it has to be noted that the merger rate is assumed to be very small in inner cluster regions (where we find Type III galaxies to reside; see Fig. \ref{fig:dist_norm}) due to high relative velocities. However, interactions of Type III galaxies with the intra-cluster medium and/or with other cluster galaxies (and the gravitational potential of the cluster) could have taken place in differently dense cluster regions over their comparatively long infall-time and thus might have been efficiently changing the galaxies' properties. The position of the break does not seem to be connected with the observed band for Type III cluster galaxies. This is different in the field, where $R_{b}$ is smaller in the g-band.
However, the data on the break radii show a large scatter (resulting in large error bars), the observed differences in $R_{b}$ between the environments are not significant according to KS-testing and hence the results on $R_{b}$ have to be interpreted with caution.
Illustrations of the results on Type III galaxies are given in (Figs. \ref{fig:hist3} and \ref{fig:3}).\\
The comparison of the three individual field samples shows that for some measured properties there are notable differences (e.g. in $R_{e}$ for Type I; in $h_{1,g}$, $h_{2,g}$ and $R_{b}$ for Type II). However, the KS-test results (comparing each of the field samples to the corresponding cluster sample) are consistent for all of these quantities, except for one sole outlier, namely $h_{2,g}$ for Type II galaxies in field sample 2.
\section{Discussion}
\label{sec:discussion}
Comparing two different environments like the field (low to intermediate density) and the cluster (high density), we are in position to explore how the physical
processes associated with high overdensity regions affect the peripheral parts of galaxy discs. There are three significant
differences between our field and cluster samples:
\begin{itemize}
\item The global size of the galaxies, as parameterised by the effective radius, is smaller by $\sim$15\% in the cluster
environment than in the field, while the global S\'ersic-index is higher in the cluster by $\sim$15\%.
\item The global (g-r) colour of the galaxies is redder by around 0.2 magnitudes in the clusters than in the field.
\item We find $\sim$2.5 times more Type I (pure exponential disc) galaxies in the clusters than in the field (15\% vs. 5\%). This difference is compensated by the lack of the corresponding percentage of Type II cluster galaxies (56\% vs. 66\%).
\end{itemize}
\noindent Our work allows us to probe, in detail, how and where the ageing and global size transformations have taken place. Figure
\ref{fig:3} shows that the global size difference between the discs in the field and in the clusters, even though largest in Type II galaxies, seems to hold for all the profile types when explored separately. When taking a detailed look
at the reddening of the individual profile types, Fig. \ref{fig:3} indicates that all the classes have undergone a
similar reddening from the field to the cluster environments. It is worth noting that both Type I and II have a similar colour in the field, the Type III objects being notably redder. However, in the cluster environment this colour
similarity is reversed so that Type I and III are the ones which share similar colours. It is remarkable that Type III field galaxies are redder than the other profile types even after masking the bulge. A possible explanation is that Type III field galaxies are formed by tidal interactions and minor mergers in the course of group pre-processing (see e.g. \citealt{younger07, lopes14}) which might have led to the observed reddening. The fact that Type II cluster galaxies are bluer than the other types in the same environment could indicate that the Type II feature is erased relatively quickly upon the cluster infall, allowing only for a limited reddening before the break vanishes. Note, however, that these are tentative interpretations.\\
\noindent As the inner scale length of the disc galaxies, $h_1$, barely changes when moving from the field to the clusters, to
understand the ultimate reason of the change in the global size of the galaxies we need to focus our attention on the outer
scale length, $h_2$. For Type I and Type III discs, we find that $h_2$ is larger (by a factor of $\sim$1.08 and $\sim$1.16, respectively) in
the clusters than in the field. If this was the only difference in the structure of the field versus the cluster
galaxies, the global size of the cluster galaxies (at fixed stellar mass) should be larger than in the field. To
understand why the cluster galaxies are yet more compact, we have to account also for the reddening of the objects. At
decreasing the brightness of the galaxies' discs, the bulges of all these objects become more prominent. This effect
moves the effective radius towards the inner regions of the objects.\\
\noindent Both the increment of $h_2$ and the global reddening of the discs in the cluster regions are suggestive of physical
process connected to the cluster environment.
The rise of $h_2$ in Type I and more significantly in Type III cluster galaxies could be understood as the result of an increased contribution of a prominent bulge component. Since we already accounted for the bulge component when fitting the surface-brightness profiles, this hypothesis is not sufficient to explain our findings. Other processes that might play an important role include tidal effects consequent to galaxy harassment (\citealt{moore96}) and minor merger events building up the outer disc. The global reddening of the cluster galaxies is potentially connected to the decrease
of star formation in the cluster associated with the exhaustion of gas by ram-pressure stripping (see e.g.
\citealt{boesch13, pranger13, steinhauser16}). It is worth noting that the reddening of the discs seems to be stronger outside the break radius of the galaxies. This is consistent with the view that gas removal by ram-pressure stripping should be more efficient in the outer disc regions (see e.g. \citealt{steinhauser12}).\\
\indent Another interesting aspect to discuss is the dramatic increase (by a factor of $\sim$2.5) in the number of Type I discs in
the clusters. This large variation in the frequency of Type I galaxies is accompanied by a substantial change in their
colours. Whereas Type II and III have become redder by around 0.1-0.2 magnitudes, in the case of Type I this change is
$\sim$0.3 mag. The large increase in the number of Type I galaxies is compensated by a corresponding decrease of the percentage of Type II objects. This, together with our results on galaxy colour, indicates a transformation from Type II to Type I as the galaxies fall onto the cluster.
This hypothesis is in compliance with \citet{maltby15} who find indications for a transformation from spiral to S0 as the break is erased in Type II galaxies that fall onto a cluster, however, \citet{maltby15} (and \citet{maltby12b}) conclude that the structural disc parameters are not significantly influenced by environment. This disparity could be explained by the higher mass and redshift range of the galaxies selected for their study that might have made it more difficult to detect environmental trends. Furthermore, our suggested scenario is in agreement with \citet{erwin12} who analysed S0 galaxies in the cluster and in the field and concluded that the lack of Type II cluster S0s indicates a transformation from Type II to Type I. A suppression of the survival of Type II objects in the cluster environment has also been reported by \citet{roediger12}, analysing Virgo disc galaxies and, recently, by \citet{clarke17} using N-body SPH simulations. Supporting this view, it is worth noting that the scale length of the Type I discs has a value which is
intermediate between the corresponding Type II and Type III values (both for the inner and outer scalengths). This has also been found by
\citealt{munoz-mateos13}. Also their comparatively large scatter in $R_{e}$ can be interpreted as a hint for Type I galaxies experiencing a structural transformation. The unchanging frequency of Type III galaxies from field to cluster remains unexplained by our analysis and interpretation. We will address this point in our follow-up investigations.\\
Our results also have to be interpreted with respect to galaxy morphology. The quality of our data is not high enough to robustly assign all objects to morphological classes along the Hubble sequence, however, invoking the morphology-density relation (\citealt{dressler80}) we can assume a higher fraction of early-type disc galaxies in the cluster environment. Referring to \citet{gutierrez11} who find that Type I and III galaxies are more common in early-type discs while Type II galaxies are dominant in late-type discs, this would indicate that the trends we illustrate in Fig. \ref{fig:3} could - at least to some extent - be caused by morphological segregation between the two environments. On the other hand, as explained in Section \ref{sec:results}, our sample is very likely biased against S0 galaxies due to our selection criteria in stellar mass and S\'ersic-index. Moreover, the reddening for cluster Type II and III profiles is approximately the same. Since Type II galaxies are not found in cluster S0s (\citealt{erwin12, maltby15}), this indicates that the trends we find are not driven by S0 galaxies. In general, we conclude that our results are not significantly influenced by morphological segregation.
\section{Summary}
\label{sec:summary}
We have selected four samples of disc galaxies within the same narrow stellar mass range (0.8$<$ $M_{\star}$ $<$ 4) $\times$
10$^{10}$ $M_{\odot}$. While one of these samples consists of galaxies residing in galaxy clusters, the other three consist of galaxies outside of galaxy clusters, i.e. from the field. Each of the four samples holds around 175 galaxies which have been classified according to their disc
structure (Type I
$\equiv$ single-exponential, Type II $\equiv$ truncated, Type III $\equiv$ anti-truncated). We
find the following results:
\begin{itemize}
\item Disc galaxies are $\sim$15\% more compact in clusters than in the
field. They are also $\sim$0.2 mag redder in (g-r) colour and show higher S\'ersic-indices (by $\sim$15\%) in the cluster region. The reddening is observed both inside and outside the break radius.\\
\item We find $\sim$2.5 times more Type I (pure exponential disc) galaxies in the clusters than in the field. This difference is compensated by the lack of the corresponding percentage of Type II cluster galaxies. The fraction of Type III galaxies is the same in both environments.\\
\item Type III cluster galaxies reside significantly closer to the cluster centre than the other break types.\\
\item The structural parameter that changes
most significantly, at comparing cluster versus field, is the outer scalelength of Type III discs, increasing by $\sim$16\% from field to cluster. Consequent to this change, the ratio of the inner and outer scale length of Type III galaxies changes by $\sim$15\%.\\
\item We suggest that Type I galaxies form from Type II galaxies, consequent to the physical mechanisms acting on the Type II population, produced/enhanced by environment. \\
\end{itemize}
\noindent In a follow-up to this work we will investigate the structural parameters of disc galaxies as functions of clustercentric distance. To ensure adequate statistics, our analyses will be carried out on an extended sample of galaxies, including a representative fraction of galaxies that reside in the transition region between cluster and field. For a consistent follow-up we will apply the same methods as presented and described in this paper.\\
\noindent A selection of prototypical surface-brightness profiles, the evaluation of the field control samples and comprehensive lists of the measured and analysed galaxy data can be found in the appendix.
\section*{Acknowledgements}
We would like to thank the anonymous referee for detailed comments that contributed a lot to improving the quality of the present article. IT and MC acknowledge support from grant AYA2013-48226-C3-1-P and programme SEV-2011-0187 from the Spanish Ministry of Economy and Competitiveness (MINECO). This work is part of the research activities at the National Astronomical Research Institute of Thailand (Public Organization).
\bibliographystyle{mnras}
|
\section{Introduction}\label{sc:intro}
This paper is the third and last in a series of papers devoted to the structure theory of \emph{cubespaces} and \emph{nilspaces},
the previous two parts of the project being \cite{GMV1} and \cite{GMV2}.
A cubespaces is a structure consisting of a compact metric space $X$, together with a closed collection
of ``cubes'' $C^k(X)\subseteq X^{2^k}$ for each integer $k\ge 0$, satisfying certain axioms that we will recall later.
The structure $(X,C^k(X))$ is further called a nilspace
if it also satisfies certain extra rigidity conditions.
The notion of nilspaces
has its origins in the work of Host and Kra \cite{HK08}, where these objects appeared under the name of ``parallelepiped structures''. The study of these objects was furthered by Antol\'\i n Camarena and Szegedy \cite{CS12}, who in the same work formulated a strong structure theorem for nilspaces, subject to certain further hypotheses.
The papers of Candela \cites{Can1,Can2} expand on \cite{CS12}, providing more detailed proofs. He also includes
several additional results implicit in \cite{CS12}, particularly about continuous systems of measures.
The purpose of our project is to provide a new exposition of this theory, and also
to derive new applications to topological dynamics.
Although we rely heavily on the ideas contained in earlier work \cites{HK05,HM07,CS12,HK08,HKM10,GT10}, our proofs differ from the existing literature in many respects
and we also obtain new results.
The study of nilspaces might be motivated in three different ways. First, it can be a useful tool in the area of higher order Fourier analysis,
and in particular, forms the basis of Szegedy's approach \cite{S12} to proving an
inverse theorem for the Gowers norms (another approach being due to Green, Tao and Ziegler \cite{GTZ12}).
Second, nilspaces can be used in topological dynamics.
For example, we use them in this paper to generalize a result of Host, Kra and Maass
\cite{HKM10} characterizing the largest pronilfactor of a minimal group action.
Third, nilspaces can be used in the context of ergodic theory, in particular to give a new and more general proof of the structure theorem for \emph{characteristic factors}, introduced by Host and Kra \cite{HK05} and by Ziegler \cite{Z07} (using a different
framework). For details, see \cite{G15}.
Given the close
connection of the subject to the study of \textit{nilsequences} arising in additive number theory and ergodic theory \cites{BHK05,GT12}, we might expect further applications to arise.
In this paper, we explain the notion of nilspaces, and outline our project, from the point of view of topological dynamics.
The reader whose main interest lies in combinatorics or higher order Fourier analysis may wish
to consult the paper \cite{GMV1}, where an outline is given from that perspective.
She may then continue directly to Section \ref{sec:cubeinvlim} of this paper, where the main result of this paper
is presented in a fashion that does not require familiarity with the dynamical material introduced
in Section \ref{sc:intro}.
\subsection{Regional proximality}\label{sc:RP}
Let $(T,X)$ be a {\bf topological dynamical system}, which for our purposes means that
$X$ is a compact metric space and $T$ is a homeomorphism on $X$.
We begin by recalling some definitions.
We say that a pair of points $(x,y)\in X^2$ is {\bf proximal}, if there is a sequence
of integers $\{n_i\}$ such that $\lim \dist(T^{n_i}x,T^{n_i}y)=0$.
We say that $(x,y)\in X^2$ is {\bf regionally proximal}, if there are sequences of points
$\{x_i\},\{y_i\}\subseteq X$ and a sequence of integers $\{n_i\}$ such that
$\lim x_i=x$, $\lim y_i=y$ and $\lim \dist(T^{n_i}x_i,T^{n_i}y_i)=0$.
We write $(x,y)\in \Pop_T(X)$ if $(x,y)$ is proximal and $(x,y)\in \RP_T(X)$ if $(x,y)$ is
regionally proximal.
We say that the system $(T,X)$ is {\bf distal} if the proximality relation is trivial; that is, if
$(x,y)\in \Pop_T(X)$ if and only if $x=y$.
We say that the system $(T,X)$ is {\bf minimal} if the orbit of each point is dense.
It turns out that $\RP_T(X)$ is an equivalence relation, and the quotient $X/\RP_T(X)$
has the following property.
\begin{thm}[Ellis and Gottschalk \cite{EG60}*{Theorem 2}]
Let $(T,X)$ be a minimal
topological dynamical system.
Then $\RP_T(X)$ is a closed equivalence relation and $(T,X/\RP_T(X))$ is the maximal
equicontinuous factor of $(T,X)$.
In other words, there is a compact abelian group $K$ and an element $t\in K$ that generates
a dense subgroup of $K$, such that the system $(T,X/\RP_T(X))$ is isomorphic to $(t,K)$.
Moreover, $(T,X/\RP_T(X))$ is the maximal factor with this property.
\end{thm}
Motivated by this result in part, Host, Kra and Maass \cite{HKM10} introduced the notion
of the higher order regional proximal relation,
which can be used analogously to identify the maximal \emph{pronilfactor} of a topological dynamical system.
We recall some definitions.
We call a system $(T,X)$ a {\bf nilsystem} of degree $s$ if there is an $s$-step nilpotent
Lie group $G$, a discrete cocompact subgroup $\Ga \le G$, and an
element $t\in G$, such that $(T,X)$ is isomorphic as a topological dynamical system to $(t,G/\Ga)$.
Here, by a {\bf Lie group} we mean a Hausdorff, second countable
topological group $G$ equipped with a differentiable structure, such that the map
$G^2\to G:(g,h)\mapsto gh^{-1}$
is differentiable. We do not assume that Lie groups are connected, and so in particular any countable discrete group is Lie.
Some authors require the Lie
group $G$ in the definition of a nilsystem to be connected (which we do not).
We note that a connected nilmanifold $G/\Ga$ can always be realized in such a way that
$G$ is connected:
indeed, if $G^\circ$ denotes the connected component of the identity in $G$ then we may identify $G/\Ga$ with $G^\circ/(\Ga\cap G^\circ)$.
On the other hand, it may be that a connected nilsystem $(t,G/\Ga)$ cannot be
represented as a nilsystem in which $G$ is connected, if $t \notin G^\circ$.
We write $(T,X_\infty)=\invlim (T,X_i)$ if
the system $(T,X_\infty)$ is the {\bf inverse limit} of the systems $\{(T,X_i)\}$;
i.e., if there is a
family of morphisms (in the category of topological dynamical systems, i.e.~continuous $T$-equivariant maps)
\[\{\f_{i,j}:(T,X_j)\to (T,X_i)\}_{i<j\le\infty}\]
such that $\f_{i,j}\circ\f_{j,l}=\f_{i,l}$ for all triplets of indices $i<j<l\le\infty$,
and such that for any two points $x\neq y\in X_\infty$ there is $i<\infty$ such that $\f_{i,\infty}(x)\neq\f_{i,\infty}(y)$.
A collection of maps $\{\f_{i,j}\}$ with this property is called an {\bf inverse system}.
We say that a system $(T,X)$ is {\bf pronil} of degree $s$, if it is the inverse limit of nilsystems
of degree $s$. We call a factor $X \to Y$ of $X$ a {\bf pronilfactor} of degree $s$ if $Y$ is pronil of degree $s$,
and say $Y$ is the {\bf maximal pronilfactor} if every other pronilfactor of $X$ factors through $Y$.
The {\bf discrete cube of dimension} $d$ is the set $\{0,1\}^d$.
We write $\vec 0=(0,\ldots,0)\in \{0,1\}^d$ and use the notation $\vec 1$ in a similar manner.
For a vertex $\o=(\o_1,\ldots,\o_d)\in\{0,1\}^d$ and a vector $n=(n_1,\ldots, n_d)\in\Z^d$, we write
$\langle\o, n\rangle=\sum_{i=1}^d\o_i n_i$.
If $G$ is a group and $g=(g_1,\ldots,g_d)\in G^d$, then we write $g^\o=g_1^{\o_1}\cdots g_d^{\o_d}$.
Let $(T,X)$ be a system.
Following Host, Kra and Maass, we say that the pair of points $x,y\in X$
is {\bf regionally proximal of order $s$} if
there are sequences of points $\{x_i\},\{y_i\}\subseteq X$ and a sequence of integer vectors
$\{n_i\}\subseteq\Z^{s}$ such that $\lim x_i = x$, $\lim y_i =y$ and
\[
\lim \dist(T^{\langle\o ,n_i\rangle}x_i,T^{\langle\o ,n_i\rangle} y_i)=0
\]
for all $\o\in\{0,1\}^{s}\sm \{\vec 0\}$.
We denote this relation by $\RP^s_T(X)$.
For a nilsystem $(T,X)$ of degree $s$, $\RP^s_T(X)$ is trivial.
This fact is non-trivial, we return to it in Section \ref{sc:canonical}
after introducing some additional concepts.
The following theorem of Host, Kra and Maass characterizes the maximal pronilfactor of $(T,X)$ using the higher order regional
proximal relation. \footnote{The case $s=2$ had been proven previously in \cite{HM07}*{Theorem 2}.}
\begin{thm}[\cite{HKM10}*{Theorem 1.3}]\label{th:HKM}
Let $(T,X)$ be a minimal distal system.
Then $\RP^s_T(X)$ is a closed equivalence relation and
$(T,X/\RP^s_T(X))$ is the maximal pronilfactor of $(T,X)$ of degree $s$.
\end{thm}
Host, Kra and Maass also proved that a minimal system $(T,X)$ is distal provided the relation
$\RP_T^s(X)$ is trivial.
Shao and Ye \cite{SY12}*{Theorem 3.5} proved that $\RP^s_T(X)$ is a closed equivalence relation for minimal
(but not necessarily distal)
$(T,X)$ and they combined this with \cite{HKM10} to deduce that Theorem \ref{th:HKM}
holds without the assumption on distality.
The definition of $\RP_T^s$, and those of nilsystems and pronilsystems, generalize straightforwardly to the context of a dynamical system $(H, X)$ where $H$ is any abelian group (that is, $H$ acts continuously on the space $X$), the previous discussion corresponding to the case $H = \Z$.
In \cite{GGY}, Gutman, Glasner and Ye further generalize the definition of the regionally proximal relation to the case of an action by an arbitrary, possibly non-abelian group.
For abelian group actions, in particular for $\mathbb{Z}$-actions, the new definition coincides with the old one.
We will review this new definition and discuss it in detail in the next subsection, as well as stating a generalization of Theorem \ref{th:HKM} to these more general group actions, which is one of the
main goals of this paper.
\subsection{Regional proximality for non-abelian actions}\label{sc:non-Ab}
We introduce some more definitions.
Let $X$ be a compact metric space and let $H$ be a metrizable topological group acting
continuously on $X$.
We denote the action by $g.x$ for $g\in H$ and $x\in X$ and call the pair
$(H,X)$ a {\bf topological dynamical system} or simply a system.
The topology of $H$ does not play any significant role.
We assume that it is induced by the {\bf maximum displacement metric}
\[
\dist(h_1,h_2)=\max\{\dist(h_1x,h_2x):x\in X\}.
\]
We denote the set of maps $\{0,1\}^\ell\to X$ by $X^{\{0,1\}^\ell}$
and call its elements $\ell$-{\bf configurations}.
Given a configuration $c\in X^{\{0,1\}^\ell}$, we call the points $\{c(\o)\}_{\o\in\{0,1\}^\ell}$
the {\bf vertices} of $c$.
We call a configuration {\bf constant} if all its vertices are equal.
A set of the form $\{\o\in \{0,1\}^\ell:\o_i=\a\}$ for some $1\le i\le d$ and $\a\in\{0,1\}$ is
called a {\bf hyperface} of the discrete cube.
For a hyperface $F$ and an $h\in H$ we denote by $[h]_F$ the element of $H^{\{0,1\}^\ell}$
defined as $[h]_F(\o)=h$ if $\o\in F$ and $[h]_F(\o)=e$ otherwise.
Here and everywhere below $e$ denotes the identity element of the group.
We call the subgroup of $G^{\{0,1\}^\ell}$ generated by
\[
\{[h]_F : h\in H \text{ and $F$ is a hyperface of $\{0,1\}^\ell$}\}
\]
the Host--Kra cube group and denote it by $\HK^\ell(H)$.
These groups originate in \cite{HK05}*{Section 5} and coincide with the \textit{parallelepiped groups} of
\cite{HKM10}*{Definition 3.1} introduced for abelian actions. The terminology is due to
\cite{GT10}*{Definition E.3} where it is employed in the context of filtered
Lie groups.
The Host--Kra cube group acts naturally on the space of $\ell$-configurations on $X$, via
$\g.c(\o)=\g(\o).c(\o)$ for $\g\in \HK^\ell(H)$ and $c\in X^{\{0,1\}^\ell}$.
Following Host, Kra and Maass \cite{HKM10} we call the orbit closure of constant configurations
the set of {\bf dynamical cubes}\footnote{
In fact, Host, Kra and Maass call these configurations \emph{parallelepipeds},
but we use the term \emph{cubes} in order to conform with \cite{CS12}
and for the sake of brevity.}, denoted
\begin{equation}\label{eq:def of Cn H}
C^\ell_H(X)=\overline{\{\g.x^{\{0,1\}^\ell}:\g\in \HK^\ell(H), x\in X\}} \ .
\end{equation}
If $x,y\in X$ are two points, we write $\llcorner^k(x;y)$ for the configuration
\[
\o \mapsto \begin{cases} x &\colon \o \ne \vec1 \\ y &\colon \o = \vec1 \end{cases} \ .
\]
We return to the setting of $\Z$-systems as in the previous section, i.e.~we take $H=\{T^n\}_{n\in\Z}$ for a homeomorphism $T$ of $X$.
Host, Kra and Maass \cite{HKM10}*{Corollary 4.3} gave the following
characterization of the regional proximal relation of order $s$
if the system $(H,X)=(\{T^n\},X)$ is minimal and distal:
we have $(x,y)\in \RP^s_T(X)$ if and only if $\llcorner^{s+1}(x;y)\in C^{s+1}_H(X)$.
Shao and Ye \cite{SY12} proved that this holds for
general abelian actions, even without the assumption that the system is distal.
Motivated by this, in \cite{GGY}, the following definition for the regional
proximal relation for general group actions is introduced.
\begin{dfn}
Let $(H,X)$ be a topological dynamical system.
We say that a pair of points, $x,y\in X$ is {\bf regionally proximal of order} $s$ and write
$(x,y)\in \RP_H^s(X)$, if and only if $\llcorner^{s+1}(x;y)\in C^{s+1}_H(X)$.
\end{dfn}
It is shown in \cite{GGY} that, perhaps surprisingly, the newly introduced relation is an
equivalence relation for any minimal action (without any restriction on the group).
Moreover the proof of this more general fact is simpler than the one given in \cite{SY12}.
The nature of $\RP_H^s$ can vary significantly, and we now give two extreme examples.
For the first, note that if $H_{s+1}$ denotes the $(s+1)$-th element of the lower central series of $H$,
then $(x,hx)\in \RP_H^s(X)$ for any $h\in H_{s+1}$.
(For a proof of this fact, see Section \ref{sc:HK}).
Hence if $H$ is perfect, i.e.~$H=[H,H]$, and the action is minimal, then $\RP_H^s(X)=X^2$; in other words, every pair of points is regionally proximal to all orders.
For the second example, let $(H,G/\Ga)$ be a nilsystem in the following generalized sense: let $G$ be an $s$-step nilpotent
Lie group, let $\Ga$ be a discrete cocompact subgroup, and let $H$ act on $G/\Ga$ via a continuous group homomorphism $\phi:H\to G$, i.e.~$(h, x \Ga) \mapsto \phi(h) x \Ga$. Then it turns out that $\RP_H^t(G/\Ga)$ is the trivial relation for all $t \ge s$ (we will see a proof of this in Section \ref{sc:canonical}).
We are now ready to state a significant generalization of Theorem \ref{th:HKM} to general $H$-actions.
\begin{thm}\label{th:dynamics}
Let $(H,X)$ be a minimal system.
Suppose further that $H$ has a dense subgroup generated by a compact set.
Then $\RP_H^s(X)$ is a closed $H$-invariant equivalence relation, and $(H,X/\RP_H^s(X))$
is the maximal pronilfactor of $(H,X)$ of degree at most $s$.
\end{thm}
The proof of Theorem \ref{th:HKM} (which is a special case of Theorem \ref{th:dynamics}) by Host, Kra and Maass relies
on an ergodic theoretic analogue obtained previously by Host and Kra \cite{HK05}.
Our proof, however, works entirely within the topological category.
We believe that this feature makes our approach of interest even in the case of $\Z$-actions.
We note that in our proof we use the result from \cite{GGY} that $\RP^s_H(X)$
is an equivalence relation for an arbitrary minimal topological dynamical system $(H,X)$.
\subsection{Cubespaces and nilspaces}\label{sc:cubes}
In the preceding discussion we introduced the notion of dynamical cubes.
In this section we discuss similar structures in a much more abstract setting,
following Host and Kra \cite{HK08} and Antol\'\i n Camarena and Szegedy \cite{CS12}.
Everything in this section is taken from \cite{CS12}, although our
terminology and notation differs from that paper.
We first formalize the notion of a cubespace. Given the previous discussion, the following definitions should seem fairly reasonable; the reader may also consult \cite{GMV1}*{Sections 1--3} for further discussion.
\begin{dfn}
A map $\f=(\f_1,\ldots,\f_k):\{0,1\}^\ell\to \{0,1\}^k$ is termed a {\bf morphism of discrete cubes }
if each coordinate function $\f_j(\o_1,\ldots,\o_l)$ is equal to either $0$, $1$, $\o_i$ or $1-\o_i$
for some $1\le i \le m$.
\end{dfn}
\begin{dfn}
Let $X$ be a metric space and for each integer $\ell\ge 0$ let $C^\ell\subseteq X^{\{0,1\}^\ell}$
be a closed set. We say that $(X,C^\ell)$ is a {\bf cubespace} if $C^0=X$ and
$c\circ\f\in C^\ell$ for any $c \in C^k$ and any morphism of discrete cubes
$\f:\{0,1\}^\ell \to \{0,1\}^k$.
We refer to this property as {\bf cube invariance}.
\end{dfn}
We call the elements of $C^\ell$ the $\ell$-{\bf cubes} of $X$.
Given $c \in C^\ell$, we call the elements $c(\o)$ for $\o\in \{0,1\}^\ell$ the {\bf vertices} of $c$.
To simplify our notation, we simply write $X$ to refer to the full cubespace structure $(X, C^\ell)$,
and we write $C^\ell(X)$ to refer unambiguously to the cubes $C^\ell$ associated to $X$.
The cube-invariance property encodes certain fairly natural operations that produce new cubes from old ones. For instance:
\begin{itemize}
\item cube-invariance for $\f(\o_1,\ldots, \o_{\ell-1})=(\o_1,\ldots, \o_{\ell-1},0)$ encodes the fact that a face of an $\ell$-cube of dimension $(\ell-1)$ is again an $(\ell-1)$-cube;
\item similarly, cube-invariance for $\f(\o_1,\ldots, \o_{\ell})=(1-\o_1,\ldots, \o_{\ell})$ states that reflecting an $\ell$-cube in one of the coordinate axes yields another $\ell$-cube;
\item using $f(\o_1, \dots, \o_\ell) = (\o_1, \dots, \o_{\ell-1})$ shows that ``duplicating'' an $(\ell-1)$-cube creates an $\ell$-cube; and
\item applying $\f(\o_1,\o_2,\o_3)=(\o_2,\o_1,\o_3)$ shows that permuting the coordinates of a cube yields another cube.
\end{itemize}
As well as considering cubespaces, we will also need to discuss maps between them that respect the cubespace structure. The natural definition is as follows.
\begin{dfn}
Let $X$ and $Y$ be two cubespaces and let $\f:X\to Y$ be a continuous map.
Then we say that $\f$ is a {\bf cubespace morphism}, if
\[
\{\f\circ c:c\in C^k(X)\}\subseteq C^k(Y).
\]
\end{dfn}
Cubespaces also admit a natural notion of sub-objects.
\begin{dfn}
If $X$ and $Y$ are two cubespaces,
we say that $Y$ is a {\bf subcubespace} of $X$ if $Y\subseteq X$ and $C^\ell(Y)\subseteq C^\ell(X)$ for all $\ell$.
If $Z\subseteq X$ is a closed set, the subcubespace of $X$ {\bf induced} by $Z$ is defined by
$C^\ell(Z)=C^\ell(X)\cap Z^{\{0,1\}^\ell}$ for all $\ell$.
\end{dfn}
We now introduce a further technical definition.
\begin{dfn}
We say a cubespace $X$ is {\bf ergodic} if $C^1(X)=X^{\{0,1\}}$, i.e.~if any pair of points
form a $1$-cube. More generally, we say that $X$ is $s$-{\bf ergodic} if $C^s(X)=X^{\{0,1\}^s}$.
\end{dfn}
Observe that $s$-ergodicity implies $t$-ergodicity for all $t\le s$, as a consequence of cube-invariance.
It turns out that non-ergodic cubespaces are fairly uninteresting, insofar as -- under reasonable extra hypotheses -- they decompose into a number of essentially non-interacting ergodic pieces; hence, we will almost always work with ergodic cubespaces as a further sanity condition.
We now turn to the notion of \emph{corner-completion}. This is both fundamental and rather difficult to motivate from our previous discussion. Perhaps the best one can say is that this property arises naturally in many of the examples of cubespaces arising in applications -- e.g.~in dynamics, the study of nilmanifolds, etc.~-- and is an essential assumption if one hopes to prove any kind of structure theory. Ultimately, though, this definition is thoroughly non-obvious, and constitutes one of the key insights of \cite{HK08}.
\begin{dfn}
Let $X$ be a cubespace and let $\l:X\to\{0,1\}^\ell \sm \{\vec1\}$ be a map.
We call $\l$ an $\ell$-{\bf corner} if $\l|_{\o_i=0}$ is an $(\ell-1)$-cube for all $1\le i\le \ell$.
We say that the cubespace $X$ has {\bf $s$-completion}
if any $s$-corner can be completed to an $s$-cube;
in other words, if for any such $\l$ there exists a cube $c\in C^{s}(X)$ such that $c|_{\{0,1\}^{s}\sm \{\vec1\}}=\l$.
We say that $X$ is {\bf fibrant} if it has $s$-completion for all $s\ge0$.
\end{dfn}
It is also imperative to know when this completion process is unique. The dimension at which this occurs determines the ``degree'' or ``step'' of the space.
\begin{dfn}
We say that a cubespace $X$ has $s$-{\bf uniqueness},
if $c_1|_{\{0,1\}^{s}\backslash \vec1}=c_2|_{\{0,1\}^{s}\backslash \vec1}$
implies $c_1=c_2$ for any two $s$-cubes $c_1,c_2\in C^{s}(X)$.
\end{dfn}
\begin{dfn}
We say that a cubespace $X$ is a {\bf nilspace} of degree $s$ if it is fibrant and $s\ge 0$ is the
smallest integer such that $X$ has $(s+1)$-uniqueness.
\end{dfn}
In \cite{GMV1}, the following relative analogue of the corner-completion property is introduced, applying to a map between two cubespaces.
\begin{dfn}
Let $X$ and $Y$ be two cubespaces and let $\f:X\to Y$ be a continuous map.
We say that $\f$ is a {\bf fibration}
if it is a cubespace morphism, and if furthermore the following holds for every $\ell$.
Let $\l: X\to \{0,1\}^\ell \sm \{\vec 1\}$ be an $\ell$-corner and $c\in C^k(Y)$ a compatible cube, in the sense that
$\f\circ\l=c|_{\{0,1\}^\ell\sm\{\vec 1\}}$.
Then there is a completion $c'$ of $\l$ compatible with $c$; that is, there exists $c'\in C^\ell(X)$ such that $c'|_{\{0,1\}^\ell \sm \{\vec1\}} = \l$ and $\f\circ c'=c$.
\end{dfn}
It is easy to see that composition of fibrations is a fibration. We also recall
that fibrations have the following ``universal property''.
\begin{lem}[\cite{GMV1}*{Lemma 7.8}]\label{lem:universal1}
Let $\f_{YX}:X\to Y$ and $\f_{ZX}:X\to Z$ be fibrations between compact cubespaces.
Suppose that for every $y\in Y$ there is $z\in Z$ such that $\f_{YX}^{-1}(y)\subseteq\f_{ZX}^{-1}(z)$.
Then there is a unique fibration $\f_{ZY}: Y\to Z$ such that $\f_{ZX}=\f_{ZY}\circ\f_{YX}$.
Equivalently, the following holds.
Let $\f:X\to Y$ be a fibration and $\psi: Y\to Z$ be an arbitrary continuous map between two compact cubespaces.
If $\f\circ\psi$ is a fibration then so is $\psi$.
\end{lem}
Observe that a cubespace $X$ is fibrant if and only if the map from $X$ to the one-point cubespace $\{\ast\}$ is a fibration.
If we set $Z = \{\ast\}$ in the above lemma, we see that the image of a fibrant
cubespace under a fibration is also fibrant.
An almost equivalent condition to fibrations appears in \cite{CS12}*{Section 2.8} under the name ``fiber-surjective morphism''. There, a cubespace morphism $\f \colon X \to Y$ between nilspaces $X$ and $Y$ is called fiber-surjective if the image of any $\sim_k$ class in $X$ is a $\sim_k$ class in $Y$, for any $k \ge 0$. (Here $\sim_k$ are the canonical equivalence relations, which we will introduce in Section \ref{sc:canonical}.)
If $X$ and $Y$ are nilspaces and $\f \colon X \to Y$ a cubespace morphism, it is not hard to see that $\f$ is a fibration if and only if it is fiber-surjective. At some points in the project, we have reason to consider fibrations between cubespaces that are not nilspaces, and in these cases the definitions are inequivalent and the notion of a fibration appears to be the correct one to use. Moreover, in general the authors have also found it a more natural and convenient starting point on which to build the associated theory.
It is proved in \cite{GMV1}*{Remark 7.9} that the image of a nilspace under a fibration is a nilspace.
Given two $\ell$-configurations $c_{0},c_{1}:\{0,1\}^{\ell}\to X$,
the {\bf concatenation} of $c_{0}$ and $c_{1}$ is
the $(l+1)$-configuration $[c_{0},c_{1}]:\{0,1\}^{\ell+1}\to X$ given
by $[c_{0},c_{1}](\o,0)=c_{0}(\o)$ and $[c_{0},c_{1}](\o,1)=c_{1}(\o)$
for all $\o\in\{0,1\}^{\ell}$.
\begin{dfn}
We say that a cubespace $X$ has the {\bf glueing property} if the following holds for all $\ell\ge 0$:
for all $c_1,c_2,c_3\in C^{\ell}(X)$, $[c_1,c_2],[c_2,c_3]\in C^{\ell+1}(X)$
implies that $[c_1,c_3]\in C^{\ell+1}(X)$.
\end{dfn}
We recall from \cite{GMV1}*{Proposition 6.2} that fibrant cubespaces always have the glueing property.
This property is useful for the following reason: if $Y$ is an induced subcubespace of $X$, and $X$ has the glueing property, then so does $Y$; however, if $X$ is fibrant this does not necessarily imply that $Y$ is fibrant.
We note that an ergodic nilspace of degree $0$ is necessarily the $1$ point space.
More interesting examples of cubespaces are the dynamical cubespaces $(X,C^k_H(X))$ introduced
in Section \ref{sc:non-Ab}.
Indeed, we recall the following fact from \cite{GGY}*{Theorems 3.6 and 7.13}.
\begin{thm}\label{th:dyn-nilspace}
Let $(H,X)$ be a minimal topological dynamical system.
Then $\RP_H^\ell(X)$ is a closed $H$-invariant equivalence relation,
and $(X/\RP_H^\ell,C_H^k)$ is an ergodic nilspace of degree at most $\ell$, for each $\ell\in\N$.
\end{thm}
Finally, we give another example of a nilspace, which will play a special role in the theory.
Let $A$ be a compact abelian group, with the group operation denoted additively.
We write $\mathcal{D}_s(A)$ for the cubespace defined by requiring that
$c\in C^\ell(\mathcal{D}_s(A))$ if and only if
\[
\sum_{\o\in\{0,1\}^{s+1}}(-1)^{|\o|}c(\f(\o))=0
\]
holds for any morphism of discrete cubes $\f:\{0,1\}^{s+1}\to \{0,1\}^\ell$,
where we write $|\o|=\sum_{1\le i\le s+1}\o_i$ for $\o\in\{0,1\}^{s+1}$.
We will consider this object again in the next section, as it turns out to be a
special case of a construction discussed there.
We will see that it follows from general results that $\mathcal{D}^s(A)$ is
a nilspace of degree $s$, and leave it to the reader to verify that it is also $s$-ergodic.
\subsection{Host--Kra nilspaces}\label{sc:HK}
In this section, we discuss a variant of the dynamical cubespace construction considered above.
A more detailed exposition, with examples, is available in \cite{GMV1}*{Section 2 and Appendix A}.
Let $G$ be a topological group.
We call a chain of closed subgroups
\[
G=G_0\supseteq G_1\supseteq G_2\supseteq\ldots \supseteq G_{s+1}=\{1\}
\]
a {\bf filtration of degree} $s$ if $[G_i,G_j]\subseteq G_{i+j}$ for all $i,j \ge 0$, adopting the convention that $G_i = \{1\}$ for all $i \ge s+1$.
Note that a filtration is always a central series, but a central series may not be
a filtration.
E.g. if $G$ is a nilpotent Lie group of degree $2$, then $G_0=G_1=G_2=G$, $G_3=[G,G]$,
$G_4=\{1\}$ is a central series, but it is not a filtration, because $[G_2,G_2]\not\subseteq G_4$.
On the other hand, we note that the lower central series is always a filtration (see \cite{MKS66}*{Theorem 5.3}).
We call the filtration {\bf proper} if $G_0=G_1$.
Note that if a group admits a proper filtration of degree $s$ then it must be nilpotent of nilpotency class at most $s$.
In this paper, we always assume that filtrations are proper even if we do not state this explicitly.
We call a group a {\bf filtered} group if we want to emphasize that it is equipped with a particular
filtration.
We write $G_\bullet$ as a shorthand to denote a group $G$ equipped with a filtration $\{G_i\}$.
We now consider a generalization of the notion of Host--Kra cubegroups introduced in Section \ref{sc:non-Ab}.
A subset $F\subseteq\{0,1\}^{\ell}$ of the vertices of the discrete cube is called a {\bf face
of co-dimension} $d$ if there are indices $1\le i_1<\ldots<i_d\le \ell$ and $\a\in\{0,1\}^d$
such that
\[
F=\{\o\in\{0,1\}^\ell:\o_{i_j}=\a_j\;\text{for all $1\le j\le d$}\}.
\]
Let $G_\bullet$ be a filtered topological group of degree $s$.
If $F\subseteq\{0,1\}^\ell$ and $g\in G$, we write $[g]_F$ for the element of $G^{\{0,1\}^\ell}$
given by $[g]_F(\o)=g$ if $\o\in F$ and $[g]_F(\o)=e$ otherwise.
We define the {\bf Host--Kra cubegroup} $\HK^\ell(G_\bullet)$ for each $\ell$ to be the subgroup of $G^{\{0,1\}^{\ell}}$
generated by the elements of the form
$[g]_F$, where $F\subseteq\{0,1\}^\ell$ is a face of codimension $i$ for some $1\le i\le s+1$, and $g\in G_i$.
We note that the definition of $\HK^\ell(G_{\bullet})$ in \cite{GMV1}*{Definition 2.2} differs in that
only ``upper'' faces $F$ are used to construct generators.
However, this gives rise to the same group, as noted in \cite{GMV1}*{Remark 2.3}.
If $F$ is a face of codimension $d$, then for any positive integers $d_1,d_2$ with $d_1+d_2=d$
we can find faces $F_1$ and $F_2$ of codimension $d_1$ and $d_2$, respectively, such that
$F_1\cap F_2=F$.
Then, the identity $[[g_1]_{F_1},[g_2]_{F_2}]=[[g_1,g_2]]_{F}$ holds (where, confusingly, some of the square brackets denote commutators and others do not).
Using these observations, it is easy to see that the Host--Kra cubegroup $\HK^\ell(G)$ defined in Section
\ref{sc:non-Ab} agrees with the above construction applied to the lower central series filtration.
For a different filtration, however,
the Host--Kra cubegroup may be larger.
We digress to justify a claim we made in Section \ref{sc:non-Ab}.
Let $g\in G_{s+1}$, the $(s+1)$-th element of the lower central series filtration.
Then $[g]_F\in \HK^{s+1}(G)$ for any face of co-dimension $(s+1)$ in $\{0,1\}^{s+1}$,
i.e.~for any single vertex of $\{0,1\}^{s+1}$.
Thus, $[g]_F.x^{\{0,1\}^{s+1}}\in C^{s+1}_G(X)$,
which implies that $(x,gx)\in \RP_G^{s+1}(X)$ as claimed in Section \ref{sc:non-Ab}.
Let $G_\bullet$ be a degree $s$ filtered Lie group.\footnote{Note that, by Cartan's closed subgroup theorem, the groups $G_i$ appearing in the filtration of a Lie group are automatically themselves Lie groups.}
We say that a discrete co-compact subgroup $\Gamma$ of $G$ is {\bf compatible}
with the filtration if $\Gamma\cap G_i$
is a (discrete) co-compact subgroup of $G_i$ for all $i$.
The group $\HK^\ell(G_\bullet)$ naturally acts on the space
$(G/\Gamma)^{\{0,1\}^\ell}$.
It turns out that
the stabilizers are discrete cocompact subgroups in $\HK^\ell(G_\bullet)$,
provided $\Gamma$ is compatible with the filtration, and hence the orbits of the action
are compact and therefore closed.
(For a proof of this fact, see \cite{GT10}*{Lemma E.10}, where the only fact that is used (implicitly) is that $\Gamma$ is compatible with the filtration.)
We define the {\bf Host--Kra nilspace} $\HK(G_\bullet)/\Gamma$ associated to $G_\bullet$ and $\Gamma$ as follows.
The base topological space is $X = G/\Gamma$, and the cubes are defined as
\[
C^\ell(G/\Gamma) := \left\{\omega \mapsto g(\omega).x \colon g\in \HK^\ell(G_\bullet), x\in X \right\}.
\]
In \cite{GMV1} we defined Host--Kra nilspaces slightly differently:
we considered $(G,\HK^\ell(G_\bullet))$ as a cubespace, and defined the cubespace $G/\Gamma$
as the quotient of this under the map $G\to G/\Gamma$; so the cubes of $G/\Gamma$ are denoted as $\HK^\ell(G_\bullet)/\Gamma$.
However, this is completely equivalent to the above definition, and the discrepancy made deliberately for consistency
with the dynamical viewpoint adopted in this paper.
It is proven in \cite{GMV1}*{Proposition 2.6}
that $\HK(G_\bullet)/\Gamma$ is an ergodic nilspace of degree $s$.
Recall the definition of $\mathcal{D}_s(A)$ from the previous section.
It turns out that $C^\ell(\mathcal{D}_s(A))=\HK^\ell(A_\bullet)$, where
$A$ is considered with the filtration $A_0=A_1=\ldots = A_s=A$ and $A_{s+1}=\{0\}$
of degree $s$.
This equivalence is proved in \cite{GMV1}*{Propostion 5.1}.
When $A$ is a compact Lie group, this is a Host--Kra nilspace.
Host--Kra nilspaces are relatively easy to understand thanks to the well-developed
theory of nilmanifolds.
The main aim of our project following Antol\'\i n Camarena and Szegedy \cite{CS12}
is that we want to approximate general nilspaces by Host--Kra nilspaces.
We outline our program to achieve this goal in the next sections.
\subsection{Canonical factors}\label{sc:canonical}
The first stage of our program (following \cite{CS12}) is to realize an ergodic nilspace of degree $t$ as a tower of extensions
\[
X\to\pi_{t-1}(X)\to\pi_{t-2}(X)\to\ldots\to\pi_0(X)=\{\ast\},
\]
where the degree of the nilspace is reduced by one each time we move along the sequence.
In the setting of Host--Kra nilspaces, this corresponds to taking quotients by the normal subgroups
$G_s$ that form the filtration; i.e.~we expect that $\pi_s(G/\Gamma)$ should be $\HK((G_\bullet/G_{s+1}) / (\Gamma / (\Gamma \cap G_{s+1})))$. The challenge is to simulate this construction in the setting of general nilspaces.
It is clear that we will need to construct $\pi_s(X)$ as quotients of $X$ in a suitable sense, and we first verify that this makes sense.
\begin{dfn}
Let $X$ be a cubespace and let $\sim$ be a closed equivalence relation on $X$.
Write $\pi:X\to X/\sim$ for the corresponding quotient map.
Then we define a cubespace structure on $X/\sim$, the {\bf quotient cubespace}, by declaring a configuration $c\in (X/\sim)^{\{0,1\}^\ell}$
to be a cube if and only if there is a cube $c'\in C^{\ell}(X)$ such that $\pi(c')=c$.
\end{dfn}
It is easy to verify that $X/\sim = \pi(X)$ as constructed is indeed a cubespace.
Resuming the above discussion: the key feature of $\pi_s(G/\Gamma)$ for a Host--Kra nilspace is that it has degree $s$, and is in some sense the largest quotient with this property. In cubespace language, this states that $\pi_s(G/\Gamma)$ has $(s+1)$-uniqueness.
So, for $X$ a general cubespace,
our task is to find an equivalence relation $\sim_s$ on $X$ such that the quotient $X/\sim_s$ has $(s+1)$-uniqueness,
and it is the smallest relation with this property. The correct definition is as follows.
\begin{dfn}
Given a cubespace $X$ and $s \ge 0$, define a relation $\sim_s$ on $X$ as follows: $x\sim_i y$ if and only if there are two cubes $c_1,c_2\in C^{i+1}(X)$
such that $c_1(\o)=c_2(\o)$ for $\o\neq\vec 1$, $c_1(\vec 1)=x$ and $c_2(\vec 1)=y$.
We call $\sim_s$ it the $s$-th {\bf canonical equivalence relation} on $X$.
\end{dfn}
This may be compared with \cite{CS12}*{Definition 2.3}.
It is clear that if $X/\sim_s$ is to have $(s+1)$-uniqueness, then $\sim_s$ must contain at least these pairs of points.
What is much less obvious is that this definition does indeed give rise to a closed equivalence relation, and that the corresponding quotient is a nilspace whenever $X$ is.
However, all this is proved in \cite{GMV1}*{Proposition 6.3},
following \cite{CS12}*{Section 2.4} and \cite{HK08}*{Section 3.3} closely.
Note that $\sim_s$ is the trivial relation if and only if $X$ already has $(s+1)$-uniqueness.
The canonical equivalence relation $\sim_s$ has the following alternative definition, whose equivalence is
proven in \cite{GMV1}*{Lemma 6.6}
(following \cite{CS12}*{Lemma 2.3} and \cite{HK08}*{Proposition 3}).
Recall from Section \ref{sc:non-Ab} that we denote by $\llcorner^s(x;y)$ the configuration $\o\mapsto x$
for $\o\neq\vec1$ and $\vec1\mapsto y$.
\begin{lem}\label{lem:alter-canonical}
Let $X$ be a fibrant cubespace.
Then $x\sim_s y$ if and only if $\llcorner^{s+1}(x;y)$ is a cube.
\end{lem}
This alternative characterization of $\sim_s$ shows that
if $X$ is a fibrant dynamical cubespace induced by a group $G$ acting on $X$, then $\sim_s=\RP_G^s$.
Now we can explain why $\RP_G^s$ is trivial on a degree $s$ nilsystem, as we claimed
at the end of Section \ref{sc:RP} for $\mathbb{Z}$-systems and in Section \ref{sc:non-Ab} for general $G$-actions.
In this case, the dynamical cubes form a Host--Kra cubespace in the above sense, which is a
nilspace of degree at most $s$ by \cite{GMV1}*{Proposition 2.6}.
Hence, $\sim_s$ is trivial (by the above remarks) and it follows that $\RP_G^s$ is, since they agree.
The canonical equivalence relation has the following {\bf universal replacement}
property proved in \cite{GMV1}*{Proposition 6.3}
(see also \cite{CS12}*{Lemma 2.5} and \cite{HK08}*{Proposition 3}).
\begin{lem} \label{lem:replacement}
Let $X$ be a fibrant cubespace.
Let $c\in C^{\ell}(X)$ for some $\ell\le s+1$, and
suppose that $c'\in X^{\{0,1\}^\ell}$ is a configuration such that $c(\o)\sim_s c'(\o)$
for all $\o\in\{0,1\}^\ell$.
Then $c'\in C^\ell(X)$.
\end{lem}
Finally, we summarize our notation for these constructions.
If $X$ is a fibrant cubespace,
we call $X/\sim_s$ the $s$-th {\bf canonical factor} of $X$.
The quotient map is denoted $\pi_s \colon X \to X/\sim_s$, and we also use the notation $\pi_s(X)$ to denote the quotient space.
\subsection{Structure groups}\label{sc:structure-groups}
We state the first structure theorem for nilspaces in this section, which
is proved in \cite{GMV1}*{Theorem 5.4}
(see also \cite{CS12}*{Theorem 1}, and also \cite{HK08}*{Section 5} for a closely related discussion).
\begin{thm}[Weak Structure Theorem]\label{th:weak-structure}
Let $X$ be a compact ergodic nilspace of degree $s$.
Then there is a compact abelian group $A = A_s(X)$ (notated additively) acting continuously and freely on $X$, such that the following hold.
\begin{enumerate}
\item The orbits of $A$ coincide with the fibres of $\pi_{s-1}$, the $(s-1)$-th canonical projection.
\item Let $c_1,c_2\in C^{\ell}(X)$ be two cubes such that $\pi_{s-1}(c_1)=\pi_{s-1}(c_2)$.
Denote by $a:\{0,1\}^{\ell}\to A$ the unique configuration in $A$ such that $a(\o).c_1(\o)=c_2(\o)$ for all $\o \in \{0,1\}^\ell$.
Then $c_2\in C^\ell(X)$ if and only if $a\in C^{\ell}(\mathcal{D}_s(A))$.
\end{enumerate}
\end{thm}
It is worth spelling out the meaning of item $2$ in a few special cases.
If $\ell=s+1$, then the condition $a\in C^{\ell}(\mathcal{D}_s(A))$ is equivalent to
\[
\sum_{\o\in\{0,1\}^{s+1}}(-1)^{|\o|}a_\o=0.
\]
If $l<s+1$, then any configuration $c:\{0,1\}^\ell\to X$ is a cube provided $\pi(c)\in C^\ell(\pi(X))$, because
$\mathcal{D}_s(A)$ is $s$-ergodic. Note this is consistent with Lemma \ref{lem:replacement}.
If we consider cubes contained in a single fibre of $\pi_{s-1}$, we see that they admit a free and transitive action by $C^\ell(\mathcal{D}_s(A))$. Equivalently, the subcubspaces of $X$ induced by the fibres of $\pi_{s-1}$ are all isomorphic copies of $\mathcal{D}_s(A)$.
Using the characterization of $\mathcal{D}_s(A)$ in terms of Host--Kra cubegroups, we obtain the
following.
If $c\in C^{\ell}(X)$ is a cube, and $F$ is a face of codimension at most $s$, then
$[a]_F.c$ is also a cube for all $a\in A$.
This holds in particular when $\ell=s+1$ and $F$ is an edge, i.e.~a face of dimension $1$.
The weak structure theory constitutes progress towards a structure theorem for nilspaces, for the following reasons.
\begin{itemize}
\item By part (1), we have that $X$ is an $A$-principal bundle over $\pi_{s-1}(X)$. Iterating this procedure on $\pi_{s-1}(X)$, we realize $X$ as a tower of extensions by compact abelian groups, terminating in $\pi_0(X)$, the $1$-point space, which constitutes fairly strong information about the structure of $X$.
\item In order to recover the cubes of $X$ given knowledge of $\pi_{s-1}(X)$, it suffices to exhibit a single $(s+1)$-cube lying above each $(s+1)$-cube of $\pi_{s-1}(X)$, since part (2) then gives us all such cubes. Again, we can iterate this on $\pi_{s-1}(X)$ to obtain a full description of the cubespace.
\end{itemize}
We call the group $A = A_s(X)$ the {\bf top structure group} of $X$. Also, we define $A_t(X) := A_t(\pi_t(X))$, the top structure group of the canonical factor $\pi_{t}(X)$ for $0 \le t \le s$, and call it the $t${\bf -th structure group} of $X$.
The proof of Theorem \ref{th:weak-structure} is given in the paper \cite{GMV1}, we only recall here how the group $A$ is constructed.
Recall that we denote by $\llcorner^\ell(x;y)$ the $\ell$-configuration all of whose vertices are $x$ except for the one labelled
by $\vec 1$, which is $y$.
Recall also the notation $[c_1,c_2]$ from Section \ref{sc:cubes},
which denotes the concatenation of the configurations $c_1$ and $c_2$.
We consider the set
\[
Y=\{(x,y)\in X\times X:x\sim_{s-1} y\}
\]
and introduce a relation $\approx$ on $Y$, given by
setting $(x,y)\approx(x',y')$ if and only if $[\llcorner^{s}(x;y),\llcorner^{s}(x';y')]\in C^{s+1}(X)$.
It is shown in \cite{GMV1} that $\approx$ is a closed equivalence relation; hence we define
$A$ to be the quotient $Y/\approx$.
One can argue using $(s+1)$-uniqueness that, given $x \in X$, each equivalence class of $\approx$ has
an unique representative of the form $(x,y)$ for some $y$.
Hence each element of $A$ can be identified with the graph of a transformation on $X$, and we
use this identification to define simultaneously the group law on $A$ and its action on $X$.
\subsection{Lie-fibered nilspaces}
We say that a nilspace $X$ is {\bf Lie-fibered} if the structure groups $A_i(X)$ of $X$ defined in
the previous section are all Lie groups.
We recall the main result of the paper \cite{GMV2} below, which classifies Lie-fibered nilspaces under the additional technical assumption that $C^k(X)$
is connected for each $k$.
We call nilspaces satisfying this latter property {\bf strongly connected}.
We say that a homeomorphism $f$ of $X$ is an $i$-{\bf translation} if
$[f]_F.c\in C^{\ell}(X)$ for each $c\in C^\ell(X)$
and any face $F\subseteq \{0,1\}^\ell$ of the discrete cube of codimension $i$. It follows in particular that $f$ is a cubespace morphism $X \to X$.
It is clear that the set of $i$-translations, endowed with the maximum displacement
metric, forms a topological group, which we denote by $\Aut_i(X)$.
The notion of translations originate from the work of Host and Kra \cite{HK08}*{Definition 6}
and they play a prominent role in the program of
Antol\'\i n Camarena and Szegedy \cite{CS12}.
It is easy to verify from the definitions that the groups $\Aut_i(X)$ are nested, and form a (proper) filtration of the group $\Aut_1(X)$, by
\[
\Aut_1(X) \supseteq \Aut_1(X) \supseteq \Aut_2(X) \supseteq \dots \ .
\]
We denote this filtered group by $\Aut_\bullet(X)$.
Let $G_\bullet$ be a filtered Lie group with a compatible discrete cocompact subgroup $\Gamma$.
It is a direct consequence of the definitions that
the elements of $G_i$ are $i$-translations on the Host--Kra nilspace $\HK(G_\bullet)/\Gamma$.
Therefore, if it is possible to represent a nilspace $X$ as a Host--Kra nilspace, then it must be possible to
locate the filtered group $G_\bullet$ inside the filtered group $\Aut_\bullet(X)$.
The next result confirms that, in the case of Lie-fibered strongly connected nilspaces,
it is sufficient simply to take $G_i = \Aut_i^\circ(X)$, the connected component of the identity in $\Aut_i(X)$.
For a proof, see \cite{GMV2}*{Theorem 2.18}
(and see also \cite{CS12}*{Theorem 7}).
\begin{thm}\label{th:toral}
Let $X$ be a compact ergodic Lie-fibered strongly connected nilspace of degree $s$.
Fix a point $x_0\in X$.
Then $G=\Aut^\circ_1(X)$ is a Lie group which admits a filtration
\[
G_\bullet:\,G=G_0=\Aut^\circ_1(X)\supseteq \Aut^\circ_2(X)\supseteq\ldots \supseteq \Aut^\circ_{s+1}(X)=\{1\}
\]
of degree $s$ and a discrete subgroup
$\Gamma=\Stab(x_0)\subseteq G$ compatible with the filtration, such that the map
\begin{align*}
G/\Gamma &\to X \\
f\cdot \Ga &\mapsto f(x_0)
\end{align*}
is an isomorphism of cubespaces between $\HK(G_\bullet)/\Gamma$
and $X$.
\end{thm}
In \cite{CS12}*{Theorem 7} the same conclusion is shown to hold
under the assumption that the structure groups of $X$ are all connected, i.e.~are all tori
of various dimensions.
It is easy to see that this condition implies that $X$ is strongly connected.
The other implication -- that strong connectivity implies that the structure groups are tori --
also holds, but it is less obvious.
Indeed, the only proof of which we are aware makes use of the full force of our
structure theorem:
Theorem \ref{th:toral} implies that a strongly connected nilspace with Lie structure
groups is isomorphic to a Host--Kra nilspace of a connected nilpotent Lie group $G$ endowed with a filtration
$\{G_i\}$ of connected subgroups, and it follows from this that the structure groups $(G_i/G_{i+1}) / ((G_i \cap \Gamma) / (G_{i+1} \cap \Gamma))$ are tori.
There is a third possible formulation of Theorem \ref{th:toral}.
One can replace the condition that $X$ is Lie-fibered with suitable topological conditions, e.g.~requiring
that $X$ is locally connected and has finite Lebesgue covering dimension. These conditions certainly hold whenever $X$ is a topological manifold, which is clearly a necessary condition for $X$ to be isomorphic to a nilmanifold. See Theorem \ref{th:top-conditions} for further details.
In the setting of nilspaces constructed from a topological dynamical system, as in Section \ref{sc:non-Ab},
we have the following variant.
Recall that a nilsystem is a topological dynamical system $(H,X)$ such that there is a nilpotent
Lie group $G$ and a discrete cocompact subgroup
$\Gamma$ of $G$, such that $(H,X)$ is isomorphic to $(H,G/\Gamma)$, where the action of $H$
on $G/\Gamma$ is induced from a continuous homomorphism $\alpha \colon H\to G$.
\begin{thm}\label{th:Lie-dynamical}
Let $(H,X)$ be a minimal topological dynamical system.
Assume that $X$ is locally connected and has finite Lebesgue covering dimension, and also that $\RP_H^s(X)$, the regional proximal relation of order $s$, is trivial for some $s$.
Then $(H,X)$ is a nilsystem.
\end{thm}
The proof of this variant of this result is discussed in
Section \ref{sc:proof-dyn-alg-struc} of the appendix. A
similar result is proved in \cite{GMV2}*{Corollary 2.20},
where as in Theorem \ref{th:toral} the topological conditions are
replaced by cubespace-theoretic conditions on the dynamical nilspace constructed from the action of $H$.
We recall the main ideas of the proofs of these theorems from the paper \cite{GMV2}.
The most difficult part of the proof of Theorem \ref{th:toral}
is in verifying that the group $\Aut_1^\circ(X)$ acts transitively
on $X$. This is proved by induction on the degree of $X$.
It can be seen easily that the canonical projection $\pi_{s-1}$ induces a homomorphism
$\pi_{s-1}^*:\Aut_1^\circ(X)\to\Aut_1^\circ(\pi_{s-1}(X))$.
Furthermore, it can be seen that the kernel of $\pi_{s-1}^*$ contains the connected
component of the top structure group $A$ of $X$,
which acts transitively on the connected components of the fibres of
$\pi_{s-1}$ as we discussed in the previous section.
Therefore, it remains to prove that $\pi_{s-1}^*$ is surjective, from which transitivity follows by these observations and inductive hypothesis on $\pi_{s-1}(X)$.
Fix a small parameter $\ve>0$ and let $f\in\Aut_1^\circ(\pi_{s-1}(X))$ be a translation such that $\dist(x,f(x))<\ve$
for all $x\in\pi_{s-1}(X)$.
We want to show that there is a translation $\wt f\in\Aut_1^\circ(X)$ such that $\pi_{s-1}^*(\wt f)=f$.
To this end, we first find a homeomorphism $g$ of $X$ -- not necessarily a translation or even a cubespace morphism --
such that $\pi_{s-1}(g(x))=f(\pi_{s-1}(x))$; and moreover such that $g$ commutes with the action of the top structure group $A_s(X)$, i.e.~$g(a.x)=a.g(x)$ for all $a\in A$ and $x\in X$.
In the next step, we attempt to correct $g$ so as to make it a genuine translation on $X$. Specifically, we look for a map $\a:X\to A$ such that the transformation
\[
\wt f: x\mapsto \a(x).g(x)
\]
is a translation.
For $c\in C^{s}(X)$, we write $D(c)$ for the unique element of $A$ such that
$[c,[D(c)]_{\{\vec 0\}}.g\circ c]\in C^{s+1}(X)$.
The function $c\mapsto D(c)$ encodes the amount $g$ deviates from being a translation at $c$.
It can be seen from the weak structure theorem that the transformation
$\wt f$ defined above is a translation if and only if the functional equation
\begin{equation}\label{eq:funct}
\sum_{\o\in\{0,1\}^s} (-1)^{|\o|}\a(c(\o))=D(c)
\end{equation}
holds for all $c \in C^s(X)$.
The equation \eqref{eq:funct} appears in \cite{CS12} (although not in exactly this way) and plays a prominent role in the whole theory.
We recall some definitions and then a result about the solutions of \eqref{eq:funct} from \cite{GMV2}.
\begin{dfn}
Let $A$ be a compact abelian Lie group, let $X$ be a cubespace, and let $\ell \in \N$ be given.
A continuous function $\sigma:C^{\ell}(X)\to A$
is called a {\bf cocycle} if
\[
\rho([c_1,c_3])=\rho([c_1,c_2])+\rho([c_2,c_3])
\]
holds for any $c_{1},c_{2},c_3\in C^{\ell-1}(X)$ such that all three concatenations appearing in the equation are cubes.
We call this property \emph{additivity}.
\end{dfn}
We stress that our rather vague notation allows the concatenation operation $[-,-]$
on any coordinate $\{1,\ldots,\ell\}$, not just the first one; hence there are strictly speaking $\ell$ additivity conditions, one per coordinate.
We note the following two consequences of additivity:
\begin{itemize}
\item (degenerate cubes) if $c=[c_0,c_0]$ then $\rho(c)=0$;
\item (reflections) we have $\rho([c_0,c_1])=-\rho([c_1,c_0])$.
\end{itemize}
Again let $X$ be a cubespace, $A$ an abelian group and $\ell\in\N$,
and let $f:X\to A$ be a function.
Then we define the function $\partial^\ell f:C^\ell(X)\to A$ by
\[
\partial^\ell f(c)=\sum_{\o\in\{0,1\}^\ell} (-1)^{|\o|}f(c(\o)).
\]
The key step in the proof of Theorem \ref{th:toral} is the following result, which
guarantees the existence of solutions to \eqref{eq:funct} under certain hypotheses.
In fact, we state the theorem in a slightly more general form than is necessary for the purposes of
Theorem \ref{th:toral}, but we will need the full power of it in the proof of the results
that we state in the next section.
\begin{thm}
\label{th:functional}
Let $A$ be a compact abelian Lie group and let $s \ge 0$, $\ell \ge 1$ be given. Then there exists $\delta = \delta(s, \ell, A) > 0$ such that the following holds.
Let $\f \colon X \to Y$ be any fibration of degree $s$ between compact ergodic cubespaces $X$ and $Y$ that obey the glueing axiom. Let $\rho$ be an $\ell$-cocycle on $X$ with values in $A$, let $0 < \delta' \le \delta$ be given and suppose that $\dist(\rho(c), \rho(c')) \le \delta'$ whenever $\f(c) = \f(c')$.
Then there is a continuous map $f:X\to A$ and a cocycle $\wt\rho: C^\ell(Y)\to A$ such that
\begin{equation}\label{eq:functional}
\rho=\partial^\ell f+(\wt\rho\circ\f)
\end{equation}
and $\dist(f(x), f(y)) \lesssim_{s,\ell} \delta'$ (that is, there exists a constant $c=c(s,\ell)>0$ such that $\dist(f(x), f(y))\leq c\delta'$) whenever $\f(x) = \f(y)$.
Moreover, the function $f$ is unique in the following sense: if $f,f'$ are two continuous solutions of \eqref{eq:functional}; equivalently, if $f$ is and $\partial^\ell(f - f')$ is constant on fibers of $\f$; and if moreover $\dist(f(x),f'(x))\le\d$
for all $x\in X$, then $f-f'$ is constant on the fibres of $\f$.
\end{thm}
This result in this form is proved in \cite{GMV2}*{Theorem 5.1 and Corollary 5.3},
but it is very closely modelled on \cite{CS12}*{Lemma 3.19}.
The main difference is that \cite{CS12} considers only the special case where $Y$ is the one point
space (which is all that is required in our proof of Theorem \ref{th:toral}).
The general case requires, among other things, the ``relative'' weak structure theory developed in \cite{GMV1}.
The proof of Theorem \ref{th:toral} uses the connectedness hypothesis in two
crucial ways.
First, we can find continuous lifts only for translations of $\pi_{s-1}(X)$ with small
displacement.
Second, we can solve the functional equation \eqref{eq:funct} only for small cocycles.
In the proof of Theorem \ref{th:Lie-dynamical} we can get around this problem using that the acting group $H$
immerses into the group of translations and it acts transitively on the space of connected components.
We can then realize $X$ as a homogeneous space of $G=\langle \Aut_1^\circ(X), H\rangle$.
Observe that $G$ need not be connected
even if $X$ is.
The approach in \cite{CS12} to the proof of Theorem \ref{th:toral}
is both closely related to ours and in other ways somewhat different.
Both proofs use at their core the triviality of certain cocycles, in the sense of Theorem \ref{th:functional},
but the way these arise, and the method of constructing small translations, vary.
The approach in \cite{CS12} can be summarized as follows.
The authors develop a kind of cohomology theory for nilspaces, whereby an extension of a nilspace by an abelian group may be characterized
up to isomorphism by a measurable cocycle (up to ``coboundaries'').
In this picture, cocycles which are ``trivial'', or equal to coboundaries (i.e.~of the from $\rho = \partial^\ell g$) are shown to correspond to split or direct product extensions.
Armed with these tools, and given an element
$f\in\Aut_1^\circ(\pi_{s-1}(X))$ with small displacement, they construct an extension of a certain nilspace by a compact abelian group, such that the extension splits if and only if $f$ has a lift in $\Aut_1^\circ(X)$. Hence, the problem is reduced to showing triviality of a measurable cocycle associated to this extension.
\subsection{Inverse limits}
\label{subsec:invlim}
We turn to the structure theory of (general) nilspaces.
It turns out that these are not all Lie-fibered nilspaces (as can be seen by considering $\mathcal{D}_s(A)$ where $A$ is a compact abelian group but not a Lie group); but they can be approximated
by Lie-fibered nilspaces in some sense, as the following result (identical to \cite{CS12}*{Theorem 4}) shows.
\begin{thm}[Inverse Limit Theorem]\label{th:invlim}
Let $X$ be a compact ergodic nilspace.
Then there is a sequence of compact ergodic Lie-fibered nilspaces $\{X_n\}$,
and an inverse system of fibrations $\{\f_{m,n}:X_n\to X_m\}_{m<n}$
such that $X=\invlim X_n$.
\end{thm}
The proof of this result is given in Section \ref{sec:cubeinvlim}; we now outline the main ideas.
The first step is to note that compact abelian groups are inverse limits of
compact abelian Lie groups, and to apply this fact to the structure groups of $X$.
This is enough to deduce the degree $1$ case; we prove
the theorem for higher degree nilspaces by an inductive argument.
We may identify the top structure group of $X$ with an inverse limit of Lie groups,
and thereby write $X=\invlim X_\infty^{(m)}$
where each $X_\infty^{(m)}$ is a quotient of $X$ under the action of a subgroup
of the top structure group, and the top structure group of $X_\infty^{(m)}$ is a Lie group.
The degree $(s-1)$ factors are unaffected, i.e.~$\pi_{s-1}(X)=\pi_{s-1}(X_\infty^{(m)})$.
Furthermore, we may construct the sequence so that we eventually quotient by the whole top structure group, i.e.~$X_\infty^{(0)}=\pi_{s-1}(X)$.
Next, we apply the induction hypothesis to $X_\infty^{(0)}$ to write
$X_\infty^{(0)}=\invlim X_n^{(0)}$, where $X_n^{(0)}$ is Lie-fibered.
So far, we have some degree $(s-1)$ Lie-fibered nilspaces $X_n^{(0)}$ approximating $\pi_{s-1}(X)$, and some spaces $X_\infty^{(m)}$ approximating $X$ whose top structure groups are Lie, but whose degree $(s-1)$ quotients are still huge.
Our remaining task is to fill in the diagonal: we want to build a space $X_n^{(m)}$ for enough pairs $n$ and $m$, whose degree $(s-1)$ factor is the Lie-fibered space $X_n^{(0)}$ and whose top structure group is the same as that of $X_\infty^{(m)}$.
More precisely, we want to construct a nilspace $X_n^{(m)}$ for each $m$ and all sufficiently large $n \ge n_0(m)$ (depending on $m$),
such that $\pi_{s-1}(X_n^{(m)})=X_n^{(0)}$ and the fibration
$X_\infty^{(0)}\to X_n^{(0)}$ can be lifted to a fibration $X_\infty^{(m)}\to X_n^{(m)}$.
Once this is done, we can write $X_\infty^{(\infty)}$ as the inverse limit of a sequence of the form $\{X_{n_m}^{(m)}\}_{m\in\N}$ for some sequence $n_m \rightarrow \infty$.
In other words, we wish to construct the following commuting diagram of fibrations, and then take an inverse limit up the diagonal.
\begin{tikzpicture}
\matrix(m) [matrix of math nodes, row sep=3em, column sep=3em]
{
X_\infty^{(\infty)}=X & & & \\
X_\infty^{(m+1)} & X_{n_{m+1}}^{(m+1)} & & \\
X_\infty^{(m)} & X_{n_{m+1}}^{(m)} & X_{n_{m}}^{(m)} & \\
X_\infty^{(0)}=\pi_{s-1}(X) & X_{n_{m+1}}^{(0)}
& X_{n_{m}}^{(0)} & X_0^{(0)}=\{\bullet\}\\};
\path[-stealth]
(m-1-1) edge (m-2-1)edge [dashed] (m-2-2)
(m-2-1) edge (m-3-1)edge (m-2-2)
(m-3-1) edge (m-4-1)edge (m-3-2)
(m-4-1) edge (m-4-2)
(m-2-2) edge (m-3-2)edge [dashed] (m-3-3)
(m-3-2) edge (m-4-2) edge (m-3-3)
(m-4-2) edge (m-4-3)
(m-3-3) edge (m-4-3)edge [dashed] (m-4-4)
(m-4-3) edge (m-4-4)
;
\end{tikzpicture}
The main difficulty of the approach lies in the construction of this nilspace $X_n^{(m)}$. To reiterate this isolated problem: we are given a fibration $X_\infty^{(0)} \to X_n^{(0)}$, and also that the nilspace $X_\infty^{(0)}$ is the quotient of $X_\infty^{(m)}$ under the free action of a compact abelian group $A = A_s(X_\infty^{(m)})$. We wish to construct $X_n^{(m)}$ and a fibration $X_\infty^{(m)} \to X_n^{(m)}$, such that $X_n^{(0)}$ is the
quotient of $X_n^{(m)}$ under the free action of the same group $A$, and such that the following diagram commutes:
\begin{center}
\begin{tikzpicture}
\matrix(m) [matrix of math nodes, row sep=3em, column sep=3em]
{
X_\infty^{(m)} & X_{n}^{(m)} & & \\
X_\infty^{(0)} & X_{n}^{(0)} & & \\
};
\path[-stealth]
(m-1-1) edge (m-1-2) edge node [left] {$\pi_{s-1}$} (m-2-1)
(m-1-2) edge node [right] {$\pi_{s-1}$} (m-2-2)
(m-2-1) edge (m-2-2)
;
\end{tikzpicture}
\end{center}
It turns out that this is not a natural or categorical construction, and indeed it is not possible in general to construct $X_n^{(m)}$ with these properties. To do so, we will need some topological input; in particular, we will have to make use of the fact that $A$ is Lie and that $X_n^{(0)}$ is ``sufficiently close'' to $X_\infty^{(0)}$ if $n$ is large enough, in the sense that the fibers of the map $X_\infty^{(0)} \to X_n^{(0)}$ have small diameter.
The (only reasonable) way to construct the nilspace $X_n^{(m)}$ is as a quotient of $X_\infty^{(m)}$ by a
closed equivalence relation, which we denote by $\sim_n^m$.
For the quotient to have the required properties, we need the following hold for every equivalence class $D\subseteq X_\infty^{(m)}$ of $\sim_n^m$:
\begin{itemize}
\item the image $\pi_{s-1}(D) \subseteq X_\infty^{(0)}$ is equal to the inverse image of a single point under the fibration $X_\infty^{(0)}\to X_n^{(0)}$; and
\item the restriction
\[
\pi_{s-1}|_D \colon D \to \pi_{s-1}(D)
\]
of the canonical projection $\pi_{s-1}$ to the subcubespace induced by $D$, is a cubespace isomorphism.
\end{itemize}
We call a set $D$ satisfying these two properties a {\bf straight class}.
So, to construct $\sim_n^m$, we need to show that if $n$ is sufficiently large depending on $m$, then
given any point $x \in X_\infty^{(m)}$ we can find a canonical straight class containing $x$.
The proof of this fact proceeds as follows. First, we need to invoke with Gleason's theorem on the existence of local sections
for the bundle $X_\infty^{(m)}\to X_\infty^{(0)}$.
I.e.~this states that any point $x\in X_\infty^{(0)}$ has a neighborhood $U$ that admits a continuous
map $\s:U\to X_\infty^{(m)}$ such that $\pi_{s-1}\circ \s=\Id_U$.
Next, we want to adjust this section $\s$ so that it sends fibers of $X_\infty^{(0)} \to X_n^{(0)}$ to straight classes in $X_\infty^{(m)}$.
In other words, we need to choose a suitable map
$f:U\to A$ (where as above $A = A_s(X_\infty^{(m)})$ is the top structure group of $X_\infty^{(m)}$) and set $\s'(x)=f(x).\s(x)$ for $x\in U$.
The condition on $f$ that asserts that $\s'$ maps inverse images of points in $X_n^{(0)}$ into
straight classes, is very similar to \eqref{eq:funct}, and so we are able to find
such an $f$ using Theorem \ref{th:functional}.
The approach of \cite{CS12} to proving the inverse limit theorem is different.
As we mentioned previously, Antol\'\i n Camarena and Szegedy develop a cohomology theory for nilspaces.
By an intricate argument measurable cocycles are shown to correspond to extensions by compact groups.
This principle is then shown to be valid in a relative setting.
Indeed, the nilspace $X_{n}^{(m)}$ is constructed from a measurable
cocycle arising from a section for $X_{\infty}^{(m)}\to X_{\infty}^{(0)}$
which is compatible with $X_{n}^{(0)}$.
\subsection{Equivariance of fibrations for translation groups}
We have seen in the previous section that strongly connected Lie-fibered nilspaces
can be endowed with the structure of a Host--Kra nilspace.
It is natural to ask whether this structure is respected by the maps in the inverse system
realizing a strongly connected nilspace as the inverse limit of Host--Kra nilspaces.
The answer turns out to be positive, as confirmed by the following theorem that
will be obtained in Section \ref{sc:functor} as a byproduct of the proof of Theorem \ref{th:invlim}.
The results stated in this section are new.
\begin{thm}\label{th:functoriality}
Let $\f: X\to Y$ be a fibration between two compact ergodic Lie-fibered nilspaces.
Then $\f$ induces a surjective continuous homomorphism $\f_*:\Aut_i^\circ(X)\to\Aut_i^\circ(Y)$
such that
\[
\f_*f.\f(x)=\f(f.x)
\]
for all $x\in X$ and $f\in\Aut_i^\circ(X)$.
\end{thm}
Combining this with Theorems \ref{th:toral} and \ref{th:invlim}, we deduce the following.
\begin{thm} \label{th:alg-struc}
Let $X$ be a compact ergodic strongly connected nilspace of degree $s$.
Then:
\begin{itemize}
\item there exists a sequence of connected Lie groups $G^{(n)}$
equipped with filtrations $G^{(n)}_\bullet$ of degree at most $s$ (with $G^{(n)}_i$ also connected for each $i$);
\item for each $n$ there is a discrete co-compact subgroup $\Gamma^{(n)}$ of $G^{(n)}$ compatible with the filtration; and
\item for each $\infty>n\ge m$, there are surjective group homomorphisms
$\psi_{m,n} : G^{(n)} \to G^{(m)}$;
\end{itemize}
such that, letting $X_n:=G^{(n)}/\Gamma^{(n)}$ be the Host--Kra nilspace
associated to $G^{(n)}_\bullet$ and $\Gamma$, the following hold:
\begin{itemize}
\item for each $n \ge m$, we have $\psi_{m,n}\left(G^{(n)}_i\right) = G^{(m)}_i$ for each $i \ge 0$;
\item again for each $n \ge m$, we have that $\psi_{m,n}\left(\Gamma^{(n)}\right) \subseteq \Gamma^{(m)}$;
\item the map
\begin{align*}
\f_{m,n} \colon X_n &\to X_m \\
g\cdot\Gamma^{(n)} &\mapsto \psi_{m,n}(g)\cdot\Gamma^{(m)}
\end{align*}
induced by $\psi_{m,n}$, is a fibration; and
\item $X$ is homeomorphic and isomorphic as a nilspace to the inverse limit $\varprojlim X_n$
given by the inverse system $\f_{m,n}$.
\end{itemize}
\end{thm}
This result should be thought of as a strengthening of the statement that $X$ is an inverse limit of nilmanifolds: indeed, it states precisely this, but giving much more information concerning the nature of the connecting maps.
As a consequence of these results, it is possible to endow a compact ergodic strongly connected nilspace $X$
with an action of the inverse limit of the groups $\Aut_1^\circ(X_n)$. One might hope to use this information to represent $X$ itself as a homogeneous space of this inverse limit group.
Unfortunately, this fails, because the action need not be transitive in general.
Moreover, there are examples due to Rudolph \cite{R95}, in which $X$ cannot be represented as a homogeneous space
of any nilpotent group.
Nevertheless, it is possible to make some direct statements about the nilspace structure of $X$ on the strength of these results.
We believe there are interesting issues here; however, we will not pursue them presently, though we may return to them in future work.
Let $(H,X)$ be a minimal topological dynamical system as in the setting of Theorem \ref{th:dynamics}.
As we discussed in Section \ref{sc:cubes}, we know that $X/\RP_H^k(X)$ is an ergodic
nilspace of degree $k$.
Using Theorem \ref{th:invlim}, we can write $X/\RP_H^k(X)$ as the inverse limit
of Lie-fibered nilspaces $X_n$.
Notice that $H\subseteq\Aut_1(X/\RP_H^k(X))$, by the definition of dynamical cubes.
In light of Theorem \ref{th:Lie-dynamical}, we only need to show that the action of $H$
descends to $X_n$ for $n$ sufficiently large,
to conclude that
$X/\RP_H^k(X)$ is a pronilsystem, as claimed in Theorem \ref{th:dynamics}.
This is proved in Section \ref{sc:dyninverse}
using the same circle of ideas as is discussed above.
Moreover, we will obtain the following slightly stronger form of Theorem \ref{th:dynamics}.
\begin{thm} \label{th:dyn-alg-struc}
Let $(H,X)$ be a minimal topological dynamical system.
Suppose that $\RP_H^s(X)$ is trivial for some $s$, and that $H$
has a dense subgroup generated by a compact set.
Then:
\begin{itemize}
\item there exists a sequence of nilpotent Lie groups $G^{(n)}$ of degree at most $s$;
\item for each $n$, there is a continuous homomorphism $\a_n:H\to G^{(n)}$;
\item for each $n$, there is a discrete co-compact subgroup $\Gamma^{(n)}\subseteq G^{(n)}$; and
\item for each $n>m$, there is a continuous homomorphism $\psi_{m,n}:G^{(n)}\to G^{(m)}$;
\end{itemize}
such that
\begin{itemize}
\item $\psi_{m,n}(\Gamma_n)\subseteq\Gamma_m$;
\item $\a_m=\f_{m,n}\circ \a_n$; and
\item the system $(H,X)$ is isomorphic, as a topological dynamical system, to the inverse limit of
$(H,G^{(n)}/\Gamma^{(n)})$ along the inverse system of maps induced by $\f_{m,n}$, where $H$
acts on $G^{(n)}/\Gamma^{(n)}$ via $\a_n$.
\end{itemize}
\end{thm}
Finally, we note that Theorem \ref{th:dyn-alg-struc} is also valid in the following slightly more
general setting.
The condition that $\RP_H^s(X)$ is trivial could be replaced by the assumption that $X$
is a nilspace of degree at most $s$ and $H$ acts on $X$ via a continuous homomorphism
$H\to \Aut_1(X)$.
This variant allows the cubespace structure on $X$ to have more cubes than the dynamical cubespace $(X,C^k_H(X))$, provided it is still assumed to be a nilspace of degree $s$.
\subsection{Acknowledgments}
First and foremost we owe gratitude to Bernard Host who introduced us to the subject and to Omar Antol\'\i n Camarena and Bal\'azs Szegedy
whose groundbreaking work \cite{CS12} was a constant inspiration
for us.
We would like to thank Emmanuel Breuillard, J\'er\^ome Buzzi,
Yves de Cornulier, Sylvain Crovisier, Eli Glasner, Ben Green, Bernard Host, Micha\pol{} Rams, Bal\'azs Szegedy, Anatoly Vershik and Benjamin
Weiss for helpful discussions. We are grateful to Pablo Candela and Bryna Kra for a careful reading of a preliminary version.
\section{Inverse limits in the category of cubespaces}\label{sec:cubeinvlim}
The purpose of this section is to prove Theorem \ref{th:invlim}, the inverse limit theorem. We recommend the reader review the outline of the proof contained in Section \ref{subsec:invlim} before proceeding with the details.
\subsection{Preliminaries and definitions}
For each metrizable topological space we fix a metric that we always denote by $\dist(\cdot,\cdot)$.
The choices of these metrics may be arbitrary.
The following fact is the starting point of the proof.
\begin{lem}
\label{lem:groupinvlim}
A compact abelian group is the inverse limit of compact abelian Lie groups.
\end{lem}
This allows us to deduce that the structure groups of a nilspace are inverse limits of Lie groups.
In particular, the degree $1$ case of the theorem could be easily deduced from this alone.
\begin{proof}[Proof of Lemma \ref{lem:groupinvlim}]
This is straightforward given Pontryagin duality. Indeed, it follows easily from the fact that a compact abelian group is a Lie group if and only if its dual group is finitely generated.
See also \cite{Sep07}*{Theorem 5.2(a)}.
\end{proof}
Let $X_\infty^{(\infty)}$ be a compact ergodic nilspace of degree $s$.
(The reason for the double index will become clear in due course, or may be deduced from Section \ref{subsec:invlim}.)
We prove that $X_\infty^{(\infty)}$ is an inverse limit of Lie-fibered nilspaces --
the statement of Theorem \ref{th:invlim} -- by induction on $s$.
If $s=0$, then $X_\infty^{(\infty)}$ is the one point space and the theorem is trivial.
We fix $s\ge 1$, and assume that the theorem holds for nilspaces of degree $(s-1)$.
We denote by $B_\infty=\pi(X_\infty^{(\infty)})$ the $(s-1)$-th canonical factor of $X_\infty^{(\infty)}$.
Here, and throughout the remainder of this section, we abbreviate $\pi_{s-1}$ to $\pi$, as
we will not need to use any of the other canonical projections.
By the induction hypothesis, we may write $B_\infty=\invlim B_m$
for a sequence $\{B_m\}_{m\in\N}$ of Lie-fibered nilspaces, along fibrations $\psi_m \colon B \to B_m$.
We denote the top structure group of $X_\infty^{(\infty)}$ by $A^{(\infty)}$.
Applying Lemma \ref{lem:groupinvlim}, we write $A^{(\infty)}=\invlim A^{(n)}$, where
$\{A^{(n)}\}_{n\in \N}$ is a sequence of Lie groups.
Let $K_n$ denote the kernel of the surjective homomorphism $A^{(\infty)}\to A^{(n)}$.
We define
$X_\infty^{(n)}$ to be the quotient of $X_\infty^{(\infty)}$ under the action of $K_n$
and write $\a^{(n)}:X_\infty^{(\infty)}\to X_\infty^{(n)}$ for the quotient map.
It is not hard to show that $\a^{(n)}$
is a fibration, from which it follows that $X_\infty^{(n)}$ is a nilspace (by \cite{GMV1}*{Remark 7.9}).
Indeed, let $\l: X_\infty^{(\infty)}\to \{0,1\}^\ell\sm\{\vec 1\}$ be an $\ell$-corner and $c\in C^\ell(X_\infty^{(n)})$
a cube which is compatible with $\l$ in the sense that
$\a^{(n)}\circ\l=c|_{\{0,1\}^\ell\sm\{\vec 1\}}$.
Let $\wt c\in C^{\ell}(X_\infty^{(\infty)})$ be an arbitrary cube such that $\a^{(n)}(\wt c) = c$ (which exists by definition of the quotient cubespace).
Let $f:\{0,1\}^{\ell}\backslash\{\vec 1\}\to K_{n}$ be the unique configuration such that $f.\wt c|_{\{0,1\}^\ell\sm\{\vec 1\}}=\l$.
It follows from the weak structure theorem that $f$ is a corner in $\mathcal{D}_s(K_n)$,
hence it can be completed to a cube $f'$.
Thus $f'.\wt c$ is a completion of $\l$ that is compatible with $c$.
We will shortly state a technical proposition, which claims the existence of several nilspaces and
maps in between them.
We will see that Theorem \ref{th:invlim} follows from it directly.
In order to formulate the statement more easily, we first introduce some terminology.
Note that these definitions depend on the value of $s$ that has been fixed above.
Let $\f:X\to Y$ be a fibration between two compact ergodic nilspaces of degree $s$.
We claim that there is a unique fibration $\psi \colon \pi(X) \to \pi(Y)$ such that the diagram
\begin{center}
\begin{tikzpicture}
\matrix(m) [matrix of math nodes, row sep=3em, column sep=3em]
{
X & Y & & \\
\pi(X) & \pi(Y) & & \\
};
\path[-stealth]
(m-1-1) edge node [above] {$\f$} (m-1-2) edge node [left] {$\pi$} (m-2-1)
(m-1-2) edge node [right] {$\pi$} (m-2-2)
(m-2-1) edge node [above] {$\psi$} (m-2-2)
;
\end{tikzpicture}
\end{center}
commutes.
Indeed, by the universal property of the canonical factor (see \cite{GMV1}*{Remark 6.8}),
the composite $\pi \circ \f \colon X \to \pi(Y)$ must factor uniquely through $\pi(X)$, which gives us $\psi$; and by the universal property of fibrations (see Lemma \ref{lem:universal1}), since $\pi$ and $\psi \circ \pi$ are fibrations, so is $\psi$.
\begin{dfn}
With this set-up, we call $\psi$ the {\bf shadow} of $\f$.
\end{dfn}
We recall from \cite{GMV1}*{Definition 7.10}
the notion of $k$-uniqueness for a fibration $\f \colon X \to Y$: this states that if $c_1, c_2 \in C^k(X)$ are such that $\f(c_1) = \f(c_2)$ and $c_1(\o) = c_2(\o)$ for all $\o \ne \vec{1}$, then in fact $c_1 = c_2$.
We introduce the following terminology for the special case that $X$ and $Y$ are nilspaces of degree $s$ and $k=s$.
Recall again that we write $\llcorner^s(x;y)$ to denote the configuration given by $\vec1 \mapsto y$ and $\o \mapsto x$ for all $\o \ne \vec1$.
\begin{dfn}
\label{dfn:horizontal}
We say that a fibration $\f \colon X \to Y$ between two nilspaces of degree $s$ is {\bf horizontal} if one of the following two equivalent conditions holds:
\begin{enumerate}
\item we have $\f(x_1)\neq\f(x_2)$ for any two points $x_1,x_2\in X$ with $\pi(x_1)=\pi(x_2)$ and $x_1\neq x_2$;
\item for any $x \in X$ with $\f(x) = y$, we have that $\f$ restricts to a bijection $\pi^{-1}(x) \to \pi^{-1}(y)$.
\end{enumerate}
\end{dfn}
\begin{rmk}
Using the notation and terminology introduced in \cite{GMV1}*{Section 7}, the notion of horizontal fibrations
can also be characterized by one of the following equivalent conditions:
\begin{enumerate}
\item[(3)] $\f$ has (relative) $s$-uniqueness;
\item[(4)] the canonical equivalence relation $\sim_{f,s-1}$ is trivial (see \cite{GMV1}*{Section 7.2} for a definition);
\item[(5)] for all $x_1, x_2 \in X$ such that $\f(x_1) = \f(x_2)$, the configuration $\llcorner^{s}(x_1;x_2)$ is a cube only if $x_1 = x_2$.
\end{enumerate}
We will use the equivalence of (3)--(5) and the definition only later in Section \ref{sc:functor}.
In particular, the proof of Theorem \ref{th:invlim} does not require this.
Hence the reader may safely choose to ignore this remark for the moment.
The equivalence of (3), (4) and (5) is dealt with in \cite{GMV1}*{Section 7.2}.
Logically, (1) is the same as saying that if $\pi(x_1)=\pi(x_2)$ and $\f(x_1)=\f(x_2)$ then $x_1=x_2$; and by Lemma \ref{lem:alter-canonical}, this is the same as (5).
\end{rmk}
We now state the promised technical proposition.
\begin{prp}\label{prp:invlimlift}
Let $X_\infty^{(n)}$, $\alpha^{(n)}$, $B_m$, $\psi_m$, etc.~be as above.
Then there is an increasing sequence $\{M_n\}$ of positive integers such that the following holds.
For all $n\in\N$ and $m\ge M_n$, there is a compact ergodic nilspace $X^{(n)}_m$ of degree $s$
such that its $(s-1)$-th canonical factor is $\pi(X^{(n)}_m)= B_m$, and its top structure group is $A_n$.
Furthermore, there is a horizontal fibration $\f_{m}^{(n)}:X_\infty^{(n)}\to X_m^{(n)}$,
whose shadow is $\psi_m$.
In addition, we have the following property.
If $m_1\le m_2$ and $n_1\le n_2$ are such that $X_{m_1}^{(n_1)}$, $X_{m_2}^{(n_2)}$ are both
defined, then the fibres of
\[
\f_{m_2}^{(n_2)}\circ\a^{(n_2)}:X_\infty^{(\infty)}\to X_{m_2}^{(n_2)}
\]
partition the fibres of
\[
\f_{m_1}^{(n_1)}\circ\a^{(n_1)}:X_\infty^{(\infty)}\to X_{m_1}^{(n_1)};
\]
more precisely,
for each $x_1\in X_{m_1}^{(n_1)}$ there is $x_2\in X_{m_2}^{(n_2)}$
such that $(\f_{m_2}^{(n_2)}\circ\a^{(n_2)})^{-1}(x_2)\subseteq (\f_{m_1}^{(n_1)}\circ\a^{(n_1)})^{-1}(x_1)$.
\end{prp}
Note that the last claim about the fibres of $\f_m^{(n)}\circ\a^{(n)}$ is used to construct fibrations from
$X_{m_2}^{(n_2)}$ to $X_{m_1}^{(n_1)}$ with the help of the universal property of fibrations. Specifically, it asserts the existence of a map of sets $X_{m_2}^{(n_2)} \to X_{m_1}^{(n_1)}$ that makes everything commute; and by these universal properties this will turn out to be continuous and a fibration.
Assuming this proposition, whose proof we will return to below, we now complete the proof of Theorem \ref{th:invlim}. The deduction is more or less a case of formalizing the previous remark, and we advise against taking the formal details too seriously.
\begin{proof}[Proof of Theorem \ref{th:invlim} assuming Proposition \ref{prp:invlimlift}]
We assume without loss of generality that $M_n\to \infty$, and
set $X_n= X_{M_n}^{(n)}$ and $X_\infty=X_\infty^{(\infty)}$.
It is clear by construction that $X_n$ are Lie-fibered nilspaces, and so it remains only to construct the
inverse system of fibrations $\{\f_{n,l}:X_l\to X_n\}_{\infty\ge l>n}$ and thereby verify that $X_\infty = \varprojlim X_n$ as required.
We set $\f_{n,\infty}=\f_{M_n}^{(n)}\circ\a_n$ and construct the fibrations $\f_{n,l}$ for $l<\infty$
by applying Lemma \ref{lem:universal1} (the universal property of fibrations)
with $X=X_\infty=X^{(\infty)}_\infty$,
$Y=X_l=X_{M_l}^{(l)}$ and $Z=X_n=X_{M_n}^{(n)}$.
We verify that $\f_{n,l}\circ\f_{l,o}=\f_{n,o}$.
If $o=\infty$, this already follows from Lemma \ref{lem:universal1}.
If $o<\infty$ then we can write
\[
(\f_{n,l}\circ\f_{l,o})\circ\f_{o,\infty}=\f_{n,l}\circ(\f_{l,o}\circ\f_{o,\infty})=
\f_{n,l}\circ\f_{l,\infty}=\f_{n,\infty}.
\]
Hence $\f_{n,l}\circ\f_{l,o}=\f_{n,o}$
by the uniqueness part of Lemma \ref{lem:universal1} applied with $X=X_\infty$,
$Y=X_o$ and $Z=X_n$.
Finally, now that we have an inverse system, we need to verify that it separates points of $X_\infty$. Let $x,y\in X_\infty=X_\infty^{(\infty)}$ be distinct points; we need to show that $\f_n(x)\neq\f_n(y)$
if $n$ is sufficiently large.
If $\pi(x)\neq\pi(y)$, then $\psi_{M_n}(\pi(x))\neq\psi_{M_n}(\pi(y))$ if $n$ is sufficiently large.
Since the shadow of $\f_{M_n}^{(n)}$ is $\psi_{M_n}$, we have
\[
\f_{n,\infty}(x)=\f_{M_n}^{(n)}(\a^{(n)}(x))\neq\f_{M_n}^{(n)}(\a^{(n)}(y))=\f_{n,\infty}(y).
\]
If $\pi(x)=\pi(y)$, then there is a unique $a\in A^{(\infty)}$ such that $a.x=y$.
If $n$ is sufficiently large then $a\neq K_n$, hence $\a_n(x)\neq\a_n(y)$.
The claim follows again since $\f_{M_n}^{(n)}$ is horizontal hence injective on fibres of $\pi$.
\end{proof}
\subsection{Straight classes and sections}\label{sec:straight}
We now turn to the proof of Proposition \ref{prp:invlimlift}.
As discussed in Section \ref{subsec:invlim}, we will construct the nilspace $ X_{m}^{(n)}$ as a quotient of $X_{\infty}^{(n)}$ (and indeed, there is essentially no other choice).
To do this, we need to specify a procedure for identifying the fibers $\pi^{-1}(x), \pi^{-1}(y)$ in $X_{\infty}^{(n)}$
for every pair of points $x,y \in B_\infty$ such that $\psi_{m}(x) = \psi_m(y)$. See Figure \ref{fig:straight} for a pictorial representation.
Moreover, this identification needs to respect the cubespace structure.
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.5]
\draw (0,10) node [above left] {$X_\infty^{(n)}$} -- (5,10) -- (5,5) -- (0, 5) -- cycle;
\draw (0,0) node [below left] {$B_\infty$} -- (5,0) -- (5,1) -- (0,1) -- cycle;
\draw (10,0) node [below right] {$B_m$} -- (9,0) -- (9,1) -- (10,1) -- cycle;
\draw (10,10) node [above right] {$X_m^{(n)}$} -- (9,10) -- (9,5) -- (10,5) -- cycle;
\draw[->] (2.5,4) -- (2.5, 2);
\node [right] at (2.5,3) {$\pi$};
\draw[->,dashed] (9.5,4) -- (9.5, 2);
\node [right] at (9.5,3) {$\pi$};
\draw[->] (6,0.5) -- (8,0.5);
\node [above] at (7,0.5) {$\psi_m$};
\draw[->,dashed] (6,7.5) -- (8,7.5);
\node [above] at (7,7.5) {$\f_m^{(n)}$};
\fill (9.42,0.42) rectangle (9.58,0.58);
\fill (9.42,6.42) rectangle (9.58,6.58);
\fill (9.42,7.42) rectangle (9.58,7.58);
\fill (9.42,8.42) rectangle (9.58,8.58);
\fill (2,0.42) rectangle (3,0.58);
\fill (2,6.42) rectangle (3,6.58);
\fill (2,7.42) rectangle (3,7.58);
\fill (2,8.42) rectangle (3,8.58);
\draw[thick,dotted] (2.2,5) -- (2.2,10);
\draw[thick,dotted] (2.8,5) -- (2.8,10);
\draw[thick,dotted] (9.5,5) -- (9.5,10);
\end{tikzpicture}
\end{center}
\caption{A schematic of the process of constructing $X_m^{(n)}$. The map $\f_m^{(n)}$ is supposed to be a horizontal fibration; i.e.~map vertical fibers (left dotted lines) bijectively to vertical fibers (right dotted line). The three horizontal black lines in $X_\infty^{(n)}$ are three of the desired equivalence classes of the quotient. In particular, we need to construct an identification between the two left-hand fibers.}
\label{fig:straight}
\end{figure}
In particular, suppose that $\f_{m}^{(n)}:X_\infty^{(n)}\to X_{m}^{(n)}$ is indeed
a horizontal fibration whose shadow is $\psi_m$.
Then we can make the following observation.
Let $c\in C^{s+1}(B_\infty)$ be a cube all of whose vertices are mapped by $\psi_{m}$ to the same point $b\in B_m$.
Let $x\in X_m^{(n)}$ be a point above $b$, that is $b=\pi(x)$.
Denote by $c_x: \{0,1\}^{s+1}\to X_\infty^{(n)}$ the unique configuration such that $\pi(c_x)=c$ and each
vertex of $\f_{m}^{(n)}(c_x)$ is $x$.
In other words, for the vertices of $c_x$ we pick the unique points above the vertices of $c$ that
are mapped to $x$ by $\f^{(n)}_{m}$.
We claim that $c_x$ is a cube in $X_\infty^{(n)}$.
Indeed, it follows from the weak structure theorem (Theorem \ref{th:weak-structure})
that there is a cube $c'\in C^{s+1}(X_\infty^{(n)})$ with $\pi(c')=c$ and with all
but possibly one of its vertices matching the corresponding vertices of $c_x$; say $c'(\vec 1)$ is the exception.
If we remove the exceptional vertex from $c'$, we obtain a corner all of whose vertices are mapped to $x$
by $\f_{m}^{(n)}$.
Since $\f_{m}^{(n)}$ is a fibration,
this corner can be completed to a cube $c''$ such that $\f_{m}^{(n)}(c'')(\vec 1)=x$.
On the other hand, $\pi(c'')(\vec 1)=c(\vec 1)=\pi(c_x)(\vec 1)$, since $B_m$ is of degree $(s-1)$.
So, we must have that $c''=c_x$.
Motivated by the above observation, we make the following definition.
\begin{dfn}
Let $X$ be a compact ergodic nilspace, and let $\psi:B_1:=\pi(X)\to B_2$ be a fibration
onto a nilspace $B_2$.
We call a set $D\subseteq X$ a {\bf straight $\psi$-class}, if there is a point $b'\in B_2$ such that $D$
contains exactly one point on the fiber $\pi^{-1}(b)$ for every $b\in\psi^{-1}(b')$,
and if a configuration $c:\{0,1\}^{s+1}\to D$ is a cube if and only if $\pi(c)$ is a cube.
\end{dfn}
Note that the ``only if'' part of the above condition holds for any set $D$, since $\pi$ is a cubespace morphism. Also, the reader may wish to verify that this definition is a restatement of the one in Section \ref{subsec:invlim}.
We can summarize the above discussion in the following lemma.
\begin{lem}\label{lem:horizontal-fib}
If $\f:X\to Y$ is a horizontal fibration between two compact ergodic nilspaces, then the
inverse images of points under $\f$ are straight $\psi$-classes, where $\psi$ is the shadow of $\f$.
\end{lem}
A key fact, which we will formalize in Propositions \ref{prp:section-exist} and \ref{prp:section-unique} below, is the following.
\begin{fact}
If $m$ is sufficiently large (in terms of $n$), then each point $x\in X^{(n)}_\infty$ is contained
in a straight $\psi_m$-class. Moreover, this class is unique if we also insist that its diameter be small.
\end{fact}
This allows us to define $X^{(n)}_m$ as a quotient of $X_\infty^{(n)}$ by identifying straight classes; i.e., the equivalence classes of the relation $\sim_n^m$ defining $X_m^{(n)}$ are precisely the straight classes given by this fact.
We are not yet done: the condition of Lemma \ref{lem:horizontal-fib} is necessary but not sufficient. We have further work to do to show that this equivalence relation is closed and respects the cubespace structure in the appropriate way to give a horizontal fibration. However, by Lemma \ref{lem:horizontal-fib} and the above fact, we do know that this quotient is essentially the \emph{only} possible candidate, which is reassuring.
Indeed, the main problem is that the definition of straight classes only carries information about those cubes of $X_\infty^{(n)}$ that
collapse down to a single point in $B_m$ under the map $\psi_m\circ\pi$. It does not directly tell us anything about general cubes.
Fortunately, our proof of the existence of straight classes in fact yields
further information along these lines.
In order to formalize this, we now make some further definitions.
Let $X$ be a compact ergodic nilspace and denote by $A$ its top structure group.
For a vertex $\o\in \{0,1\}^{s+1}$ of the discrete cube,
and an element $a\in A$, we write $[a]_\o:\{0,1\}^{s+1}\to A$
for the map which assigns $a$ to $\o$ and the identity element $0$
to all other vertices.
Let $c: \{0,1\}^{s+1}\to X$ be a configuration such that $\pi(c)$ is a cube.
By the weak structure theorem, there is a unique element $a\in A$
such that $[a]_{\vec 0}.c$ is a cube in $X$.
We call this element $a$ the {\bf discrepancy} of $c$ and denote it by $D(c):=a$.
Let $X$ be a compact ergodic nilspace and let $\psi:B_1:=\pi(X)\to B_2$ be a fibration.
Let $U'\subseteq B_2$ be an open set and put $U=\psi^{-1}(U')$.
We say that a continuous map $\s:U\to X$ is a {\bf straight section}
if $\pi\circ\sigma=\Id_U$, and the following holds.
If $c_1,c_2:\{0,1\}^{s+1}\to U$ are two $(s+1)$-cubes of $B_1$ such that $\psi(c_1)=\psi(c_2)$,
then $D(\s(c_1))=D(\s(c_2))$.
We remark that the straightness of a section $\s$ implies that it maps the fibres of $\psi$
onto straight classes.
Indeed, let $c\in C^{s+1}(B_1)$ be a cube contained in a single fibre of $\psi$
and let $c_2$ be a constant cube in the same fibre.
Then $D(\s(c))=D(\s(c_2))=0$, since $\s\circ c_2$ is also a constant cube.
This proves that $\s(c)$ is a cube.
We also note that discrepancy is a continuous function.
Indeed, let $\{c_n\}\subseteq X^{\{0,1\}^{s+1}}$ be a sequence of configurations
converging to a limit $c\in X^{\{0,1\}^{s+1}}$, such that $\pi(c_n)$ is a cube for all $n$ (and hence so is $\pi(c)$).
We show that $\lim D(c_n)=D(c)$.
We assume as we may that $\lim D(c_n)=a$ exists.
Since $C^{s+1}(X)$ is closed, we see that
\[
\lim [D(c_n)]_{\vec 0}.c_n=[a]_{\vec 0}.c
\]
is a cube.
Hence $D(c)=a$ by definition.
We now state two technical propositions, extending the fact about existence and uniqueness of straight classes stated above.
\begin{prp}\label{prp:section-exist}
Let $X$ be an ergodic compact nilspace of degree $s$ and suppose that its top structure group is
a Lie group. Let $\ve>0$ be given.
Then there is a number $\d>0$ such that the following holds.
Let $\psi: B_1:=\pi(X)\to B_2$ be a fibration onto a nilspace $B_2$
such that $\diam(\psi^{-1}(b_2))\le\d$ for all $b_2\in B_2$.
Then for every $c\in C^{s+1}(B_1)$, there is an open set $U_2\subseteq B_2$ such that
$U_1:=\psi^{-1}(U_2)$ contains the vertices of $c$, and such that there exists a straight $\psi$-section $\s: U_1\to X$
such that $\diam(\s(\psi^{-1}(b)))\le\ve$ for all $b\in U_2$.
In particular, each point in $X$ is contained in a straight $\psi$-class
of diameter at most $\ve$.
\end{prp}
\begin{prp}\label{prp:section-unique}
Let $X$ be an ergodic compact nilspace of degree $s$ and suppose that $A$, the top structure group of $X$,
is a Lie group.
Then there is a number $\d>0$ such that the following holds.
Let $\psi: B_1:=\pi(X)\to B_2$ be a fibration onto a nilspace $B_2$.
Let $D_1,D_2\subseteq X$
be two straight $\psi$-classes with $\pi(D_1)=\pi(D_2)$ and
$\diam((D_1\cup D_2)\cap \pi^{-1}(b))\le\d$
for all $b\in\pi(D_1)$.
Then $D_1=a.D_2$ for some $a\in A$.
In particular, $D_1$ and $D_2$ are either equal or disjoint.
\end{prp}
The rest of the section is devoted
to the proofs of Propositions \ref{prp:section-exist} and \ref{prp:section-unique}.
The fact that (local) sections $s: U\to X$ exist even without any further properties
is already non-trivial, and was established by Gleason in his work on Hilbert's 5th problem.
\begin{thm}[\cite{G50}*{Theorem 3.3}]\label{thm:Gleason}
Suppose that a compact Lie group $A$ acts freely and continuously on a
completely regular topological space $X$.
Denote by $\pi:X\to X/A$ the quotient map under the action of $A$.
Then every point $x\in X/A$ has a neighborhood $U$ such that
there is a local section $\s:U\to X$; that is, a continuous map
satisfying $\pi\circ \s=\Id_U$.
\end{thm}
Although this is a very deep result, we note that the proof in the abelian case
is significantly simpler, which is all that we are using.
See the notes of Tao \cite{T14} for a self-contained treatment.
To prove Proposition \ref{prp:section-exist}, we start with a local section
$\s$ that we obtain from Gleason's theorem.
Then we ``straighten'' it using the action of the structure group $A$;
that is, we choose a suitable continuous function $f:U\to A$
and verify that $x\mapsto f(x).\s(x)$ is a straight $\psi$-section.
To this end, we work out what condition on $f$ implies the straightness of the section $f.s$,
and record it in the following lemma.
We recall the following definition.
Let $f:X\to A$ be a continuous function from a cubespace to an abelian group.
We call the function $\partial^{s+1} f: C^{s+1}(X)\to A$ defined by
\[
\partial^{s+1} f (c):=\sum_{\o\in\{0,1\}^{s+1}}(-1)^{|\o|}f(c(\o))
\]
the $(s+1)$-th {\bf derivative} of $f$.
\begin{lem}\label{lem:straightness}
Let $X$ be an ergodic compact nilspace of degree $s$, and let $\psi:B_1:=\pi(X)\to B_2$ be a fibration
onto a nilspace $B_2$.
Let $U_2\subseteq B_2$ be an open set and put $U_1=\psi^{-1}(U_2)$.
Denote by $A$ the top structure group of $X$.
Let $\s: U_1\to X$ be a section and $f: U_1\to A$ be a continuous map.
Then the section $x\mapsto f(x).\s(x)$ is straight if and only if
\[
\partial^{k+1} f(c_1)-D(\s(c_1))=\partial^{k+1} f(c_2)-D(\s(c_2))
\]
for any two $c_1,c_2\in C^{s+1}(B_1)$ such that $c_1,c_2\subseteq U_1$ and $\psi(c_1)=\psi(c_2)$.
\end{lem}
The above lemma is an immediate corollary of the following.
\begin{lem}\label{lem:discrepancy}
Let $X$ be an ergodic compact nilspace of degree $s$ and denote by $A$ its top structure group.
Consider the configurations $c:\{0,1\}^{s+1}\to X$ and $f:\{0,1\}^{s+1}\to A$.
Then $D(f.c)=D(c)-\partial^{s+1} f$.
\end{lem}
\begin{proof}
By definition $[D(c)]_{\vec 0}.c$ is a cube.
By the weak structure theorem and the identity
\[
\partial^{s+1}(f-[\partial^{s+1}f]_{\vec 0})(c)=0,
\]
the configuration
\[
(f-[\partial^{s+1}f]_{\vec 0}+[D(c)]_{\vec 0}).c=[D(c)-\partial^{s+1} f]_{\vec 0}.(f.c)
\]
is also a cube.
This proves the claim.
\end{proof}
At this point, we have reduced the proof of Proposition \ref{prp:section-exist} to solving a cocycle equation in the sense of Theorem \ref{th:functional}, and therefore the remaining technical core of the proof will be an appeal to that theorem.
However, we caution that we do really need the full technical power of Theorem \ref{th:functional} to make this work, and in particular will apply it to some slightly odd spaces $X$ and $Y$ that are not in general nilspaces.
The remaining work is therefore devoted to setting up these spaces in detail, and to a lot of technical epsilon management.
\begin{proof}[Proof of Proposition \ref{prp:section-exist}]
We will prove only the existence of the straight section, as the claim about straight classes
follows immediately from this.
In fact, we prove the following formally weaker version.
We show that for every cube $c_0\in C^{s+1}(B_1)$,
there is an open set $U_2\subseteq B_2$ such that
$U_1:=\psi^{-1}(U_2)$ contains the vertices of $c_0$,
and there exists a $\d=\d(U_2) > 0$ depending on $U_2$, $X$, $\ve$ such that the
claim of the proposition holds for $U_2$ with this $\d$.
By compactness, this implies the proposition: we may take the minimum of $\d(U_2)$ over a finite collection of open sets whose preimages under $\psi$ cover $C^{s+1}(B_1)$.
Fix a cube $c_0\in C^{s+1}(B_1)$.
By Gleason's theorem, there is an open set $U_{-1}\subseteq B_1$ containing $c_0$ and a
continuous section $\s:U_{-1}\to X$.
Indeed, if this were to fail, then let $V\subseteq B_1$ be an open set that admits a continuous section
and that contains the maximal possible number of vertices of $c_0$.
Let $b$ be a vertex of $c_0$ not covered by $V$.
We can assume without loss of generality that $b\notin \overline V$.
Then we can find an open neighborhood $V'$ of $b$ disjoint from $V$ that admits a continuous
section using Gleason's lemma.
Then $V\cup V'$ violates the assumption we made on $V$.
We choose a smaller open set $U_{0}$ containing $c_0$ such that $\overline U_{0}\subseteq U_{-1}$.
We also fix a number $\tau$ to be set later, depending only on $\ve$ and the number denoted by $\d$
in Theorem \ref{th:functional} applied with the top structure group of $X$ in the role of $A$, and with $\ell = (s+1)$.
We consider $\overline U_{0}$ with the induced cubespace structure, i.e.~
we denote by $C^{s+1}(\overline U_{0})$ the set of $(s+1)$-cubes all of whose vertices lie in $\overline U_{0}$.
We define the function $\rho: C^{s+1}(\overline U_{0})\to A$ by $\rho(c)=D(\s(c))$,
where $D$ is the discrepancy function defined above.
Since $\s$ and $D$ are continuous, $\rho$ is also continuous.
By uniform continuity, there is a number $\d_0>0$ such that $\dist(\rho(c_1),\rho(c_2))<\tau$
whenever $\dist(c_1,c_2)=\max_\omega(c_1(\omega),c_2(\omega))<\d_0$,
and $\dist(\s(x_1),\s(x_2))\le\tau/C$ whenever $\dist(x_1,x_2)\le4\d_0$.
Here $C$ is the implicit constant from Theorem \ref{th:functional}.
We now set $\d$ to be small enough such that $\d\le\d_0$ and
such that the ball of radius $\d$ around each vertex
of $c_0$ is contained inside $U_{0}$.
We construct open sets $U_1\subseteq B_1$ and $U_2\subseteq B_2$ such that $U_1=\psi^{-1}(U_2)$
and $c_0\subseteq U_1\subseteq U_{0}$.
To this end, let $Z$ be the complement of the union of the $\d$-balls around the vertices of $c_0$.
Then $\psi(Z)$ is a compact set, which does not contain any vertex of $\psi(c_0)$, since
$\diam(\psi^{-1}(\psi(c_0(\o))))< \d$ for all $\o\in\{0,1\}^{k+1}$.
We let $U_2$ be the complement of $\psi(Z)$ in $B_2$.
Then $U_1=\psi^{-1}(U_2)$ contains $c_0$ and it is disjoint form $Z$
hence it is contained in the $\d$ neighborhood of the vertices of $c_0$.
This implies $U_1\subseteq U_{0}$.
Moreover, $\diam(\s(\psi^{-1}(b)))\le\tau/C$ for all $b\in \overline{U}_2$ by the assumptions on $\psi$, $\d_0$ and $\d$.
We check that $\rho$ restricted to $C^{s+1}(\overline{U_1})$ is a cocycle.
Let $c_1,c_2,c_3\in C^{s}(U_1)$ be two cubes such that $[c_1,c_2]$, $[c_1,c_3]$ and $[c_2,c_3]$
are all cubes.
Then we know that $[\rho([c_i,c_j])]_{\vec0}.\s([c_i,c_j])$ are cubes in $X$ for $i,j=1,2,3$.
It follows then that
\begin{equation}\label{eq:cube}
([\rho([c_1,c_2])+\rho([c_2,c_3])]_{\vec0}+[\rho([c_2,c_3])]_{(0,\ldots,0,1)}).\s([c_1,c_2])
\end{equation}
is also a cube.
Indeed, this configuration is obtained from $[\rho([c_1,c_2])]_{\vec0}.\s([c_1,c_2])$ by acting
on two adjacent vertices with the same element of the top structure group.
Using that nilspaces have the gluing property (\cite{GMV1}*{Proposition 6.2})
for the cube \eqref{eq:cube} and $[\rho([c_2,c_3])]_{\vec0}.\s([c_2,c_3])$, we obtain
that
\[
[\rho([c_1,c_2])+\rho([c_2,c_3])]_{\vec0}.\s([c_1,c_3])
\]
is also a cube.
By the definition of discrepancy this
proves $\rho([c_1,c_3])=\rho([c_1,c_2])+\rho([c_2,c_3])$.
We check that Theorem \ref{th:functional} applies to the fibration $\psi:\overline U_1\to \overline U_2$.
Both $\overline U_1$ and $\overline U_2$ are considered with the induced cubespace structure
and they are compact, ergodic and have the glueing property.
Moreover as $B_1$ is an ergodic nilspace of degree $(s-1)$, $\psi$ is a fibration of degree $(s-1)$.
For two cubes $c_1, c_2\in C^{s+1}(\overline{U_1})$ with $\psi(c_1)=\psi(c_2)$, we have $\dist(c_1,c_2)<\d\le\d_0$,
hence $\dist(\rho(c_1),\rho(c_2))<\tau$ that is assumed to be smaller than the number denoted by $\d$ in
Theorem \ref{th:functional}.
We obtain a function $f:U_1\to A$ from Theorem \ref{th:functional} such that
$\rho=\partial^{s+1}f+\wt\rho\circ\psi$ and $\dist(f(x), f(y))\leq C\cdot\tau/C=\tau$
whenever $\psi(x) = \psi(y)$.
This implies that for any two cubes $c_1,c_2\subseteq U_1$ with $\psi(c_1)=\psi(c_2)$ we have
\[
\partial^{s+1}f(c_1)-D(\s(c_1))=-\wt\rho(\psi(c_1))=-\wt\rho(\psi(c_2))=\partial^{s+1}f(c_2)-D(\s(c_2)).
\]
Thus $b \mapsto f(b).\s(b)$ is a straight section on $U_1$ by Lemma \ref{lem:straightness}.
We set $\tau$ to be small enough so that $\tau\le \ve/2$ and such that for any $x\in X$ and $a_1,a_2\in A$
with $\dist(a_1,a_2)<\tau$ we have $\dist(a_1.x,a_2.x)<\ve/4$. As $\diam(\s(\psi^{-1}(b)))<\tau<\ve/2 $,
then $\diam((f.\s)(\psi^{-1}(b)))\le\ve/2+2\ve/4=\ve$ for all $b\in B_2$, as required.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prp:section-unique}]
Write $B_1'=\pi(D_1)=\pi(D_2)$.
Since $D_1$ and $D_2$ are $\psi$-classes, $B_1'$ is the $\psi$ inverse image of
a point $x\in B_2$, hence $B_1'$ is a compact ergodic nilspace and has the gluing property.
We write $B_2'$ for the nilspace whose only point is $x$.
Denote by $\s_i:B_1'\to D_i$ the inverses of $\pi$ restricted to $D_i$ for $i=1,2$.
Let $f:X\to A$ be the function such that $\s_2(x)=f(x).\s_1(x)$ for all $x\in B_1'$.
Fix $c\in C^{s+1}(B_1')$.
Since $D_1$ and $D_2$ are straight classes, $\s_1(c)$ and $\s_2(c)$ are both cubes.
This and the weak structure theorem shows that $\partial^{k+1}f(c)=0$.
On the other hand, we have $\dist(f(x),0)\le\d$ by assumption.
By Theorem \ref{th:functional} applied to $\psi:B_1'\to B_2'$ and the cocycle $0$, $f$ must be constant.
This proves the proposition.
\end{proof}
\subsection{Quotienting by straight classes}\label{sec:quotient by straight class}
Let $X$ be a compact ergodic nilspace of degree $s$ and suppose that $A$, its top
structure group, is a Lie group.
Let $\psi:B_1=\pi(X)\to B_2$ be a fibration onto a nilspace.
If the fibers of $\psi$ have sufficiently small diameter depending on $X$, then
by Propositions \ref{prp:section-exist} and \ref{prp:section-unique},
we know that each point of $X$ is contained in a unique straight
$\psi$-class of small diameter.
We define the equivalence relation $\approx$ on $X$
by requiring that each equivalence class is this unique straight $\psi$-class.
Or purpose in this section is to prove the following proposition.
\begin{prp}\label{prp:quotient by straight class}
Let $X$ be an ergodic compact nilspace of degree $s$ and suppose that $A$, its top
structure group is a Lie group.
There is a number $\d>0$ depending only on $X$ such that the following holds.
Let $\psi:B_1=\pi(X)\to B_2$ be a fibration onto a nilspace.
Suppose that $\diam(\psi^{-1}(b_2))\le\d$ for all $b_2\in B_2$.
Then the equivalence relation $\approx$ defined above
is closed.
The quotient cubespace $Y:=X/\approx$ is a nilspace,
whose $(s-1)$-th canonical factor is $B_2$.
The quotient map $\f:X\to Y$
is a horizontal fibration with shadow $\psi$.
\end{prp}
The next lemma will be used to establish that the quotient map $\f$ is a fibration.
(The lemma is equivalent to $Y$ having $(s+1)$-uniqueness.)
\begin{lem}\label{lem:quotient by straight class}
In the setting of Proposition \ref{prp:quotient by straight class}, let $c_1,c_2\in C^{\ell}(X)$
for some $\ell\ge s+1$ and suppose that $c_1(\o)\approx c_2(\o)$ for all vertices
$\o\in\{0,1\}^{\ell}\backslash \{\vec 1\}$.
Then we have $c_1\approx c_2$.
\end{lem}
\begin{proof}
It is enough to prove the lemma for $\ell=s+1$, for if $\ell$ is larger we can use the case $\ell=s+1$
for suitable faces of $c_1$ and $c_2$.
Note that $\psi(\pi(c_1))$ and $\psi(\pi(c_2))$ are cubes in $B_2$,
which is a nilspace of degree $(s-1)$ (as fibrations do not increase the degree),
and we have $\psi(\pi(c_1(\o)))=\psi(\pi(c_2(\o)))$ for $\o\in\{0,1\}^{s+1}\sm\{\vec 1\}$.
Then $\psi(\pi(c_1(\vec1)))=\psi(\pi( c_2(\vec1)))$ also.
Hence there is an element $a\in A$
such that $a.c_1(\vec1)\approx c_2(\vec1)$ and hence
$[a]_{\vec1}.c_1\approx c_2$.
We need to show that $a=0$.
We apply Proposition \ref{prp:section-exist} for the cube $\pi(c_1)$.
Let $U_1\subseteq B_1$ and $U_2\subseteq B_2$ be open sets such that
$\pi(c_1)\subseteq U_1=\psi^{-1}(U_2)$ and let $\s:U_1\to X$ be a
straight section.
Since $\psi(\pi(c_1))=\psi(\pi(c_2))$, this implies that $\pi(c_2)\subseteq U_1$.
Let $f:\{0,1\}^{s+1}\to A$ be such that $f.\s(\pi(c_1))=c_1$.
For each $\o\in\{0,1\}^{s+1}$ the set
\[
([a]_{\vec1}+f)(\o).\s(\psi^{-1}(\psi(\pi(c_2(\o)))))
\]
is a straight class that contains $[a]_{\vec 1}.c_1(\o)$ hence $c_2(\o)$ also.
Thus
$([a]_{\vec1}+f).\s(\pi(c_2))=c_2$.
By Lemma \ref{lem:discrepancy}, we have
\[
0=D(c_1)=D(\s(\pi(c_1)))-\partial^{s+1} f
\]
and
\[
0=D(c_2)=D(\s(\pi(c_2)))-\partial^{s+1} f + (-1)^{s}a.
\]
By the definition of straightness, we have $D(\s(\pi(c_1)))=D(\s(\pi(c_2)))$, hence $a=0$
proving the claim.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prp:quotient by straight class}]
We first show that $\approx$ is closed.
Let $\lim x_i=x$ and $\lim y_i=y$ be two convergent sequences of points in $X$
such that $x_i\approx y_i$ for all $i$.
By the continuity of $\pi$ and $\psi$, we have
$\psi(\pi(x))=\psi(\pi(y))$.
By Proposition \ref{prp:section-exist}, if $\d$ is sufficiently small, then there is an open set $U\subseteq B_1$
containing both $\pi(x)$ and $\pi(y)$ and a straight section $\s:U\to X$.
We assume without loss of generality that $\pi(x_i),\pi(y_i)\in U$ for all $i$.
Using Proposition \ref{prp:section-unique} we conclude that there are $a_i\in A$ such that $x_i=a_i.\s(\pi(x_i))$
and $y_i=a_i.\s(\pi(y_i))$.
We can assume without loss of generality that $a_i$ converges to an element $a\in A$.
Then by the continuity of $\s$ and the action of $A$, we have $x=a.\s(\pi(x))$ and $y=a.\s(\pi(y))$.
Then $x$ and $y$ are in the same straight class, since $\s(\pi(x))$ and $\s(\pi(y))$ are.
This shows that $x\approx y$ proving closedness.
We denote the projection map $X\to Y$ by $\f$.
We denote by $\pi':Y\to B_2$ the unique map that satisfies $\psi\circ\pi=\pi'\circ \f$,
and we show that it is a cubespace morphism.
Indeed, if $c$ is a cube in $Y$, then there is a cube $c'$ such that $\f(c')=c$,
hence $\pi'(c)=\psi(\pi(c'))$ is a cube, as well.
We show that $\f$ is a fibration.
It is a cubespace morphism by the definition of the cubespace structure on
the quotient cubespace $Y$, so it is left to prove the relative completion
property.
Let $\l$ be an $\ell$-corner in $X$ and let $c$ be a completion
of $\f(\l)$.
We show that there is a completion $c_1$ of $\l$ such that $\f(c_1)=c$.
We first consider the case $\ell\ge s+1$.
In this case, $\l$ has a unique completion $c_1$.
Let $c_2\in C^{\ell}(X)$ be such that $c=\f(c_2)$.
Then $c_1(\o)\approx c_2(\o)$ for all $\o\in\{0,1\}^\ell\backslash \{\vec 1\}$ hence
$c_1\approx c_2$ by Lemma \ref{lem:quotient by straight class}.
Thus $\f(c_1)=\f(c_2)=c$ as required.
Second, we assume that $\ell\le s$.
We use that $\psi$ is a fibration and find a cube $c_0\in C^\ell(B_1)$
that is a completion of $\pi(\l)$ such that $\psi(c_0)=\pi'(c)$.
(Here we used that $\pi'$ is a morphism, a fact that we proved above.)
We set $c_1$ to be the unique configuration such that $\pi(c_1)=c_0$ and $\f(c_1)=c$.
Since the dimension of $c_0$ is at most $s$, any $\pi$ lift of it is a cube, so $c_1$ is a cube in particular.
The fact that $\f$ is a fibration, implies by \cite{GMV1}*{Remark 7.9}
that $Y$ is a nilspace
and by the universal property that $\pi'$ is a fibration.
We leave it to the reader to verify that $\pi'$ is the $s$-th canonical projection,
$\f$ is horizontal and $\psi$ is its shadow.
\end{proof}
\subsection{Proof of Proposition \ref{prp:invlimlift}}
We take $M_n$ sufficiently large so that $\diam(\psi_m^{-1}(x))<\d$ for all $m\ge M_n$
and $x\in B_m$, where $\d$ is sufficiently small such that
Proposition \ref{prp:quotient by straight class} and Proposition \ref{prp:section-unique} can
both be applied to $X_\infty^{(n)}$.
Then the existence of the nilspace $X_m^{(n)}$ and the fibration $\f_m^{(n)}$ follows
from Proposition \ref{prp:quotient by straight class} and it remains to verify
the claim about the inverse images of points under the maps $\f_m^{(n)}\circ \a^{(n)}$.
We take indices $n_2\ge n_1$ and $m_2\ge m_1\ge M_{n_1}$ such that $m_2\ge M_{n_2}$.
Let $x\in X^{(n_2)}_{m_2}$ be a point.
Let $\wt x\in (\f_{m_2}^{(n_2)}\circ\a^{(n_2)})^{-1}(x)\in X_{\infty}^{(\infty)}$ be an arbitrary point and take
$y=\f_{m_1}^{(n_1)}(\a^{(n_1)}(\wt x))$.
We set out to prove that
\[
Z_2:=(\f_{m_2}^{(n_2)}\circ\a^{(n_2)})^{-1}(x)\subseteq Z_1:=(\f_{m_1}^{(n_1)}\circ\a^{(n_1)})^{-1}(y).
\]
We first show that $\a^{(n_1)}(Z_2)\subseteq X^{(n_1)}_\infty$ is a straight $\psi_{m_2}$-class.
Let $b\in D:=\psi_{m_2}^{-1}(\pi(x))$ be an arbitrary point.
We show that $\a^{(n_1)}(Z_2)$ contains a unique point in the fibre of $\pi$ above $b$.
Since $(\f_{m_2}^{(n_2)})^{-1}(x)$ is a straight $\psi_{m_2}$-class, it follows that
the points of $Z_2$ in the fibre of $\pi$ above $b$ is a single $K_{n_2}$ orbit.
This projects to a single point under $\a^{(n_1)}$ as $K_{n_2}\subseteq K_{n_1}$.
Let $c:\{0,1\}^{s+1}\to D$ be a cube.
We show that there is a cube $\wt c:\{0,1\}^{s+1}\to\a^{(n_1)}(Z_2)$ with
$\pi(\wt c)=c$.
Since $(\f_{m_2}^{(n_2)})^{-1}(x)$ is a straight $\psi_{m_2}$-class, we can find a
cube $c_1:\{0,1\}^{s+1}\to(\f_{m_2}^{(n_2)})^{-1}(x)$ with
$\pi(c_1)=c$.
Since $\a^{(n_2)}$ is a fibration, there is a cube $c_2:\{0,1\}^{s+1}\to Z_2$
with $\a^{(n_2)}(c_2)=c_1$ and hence $\pi(c_2)=c$.
Thus $\wt c:=\a^{(n_1)}(c_2)$ satisfies the requirements.
This shows that $\a^{(n_1)}(Z_2)$ is indeed a straight $\psi_{m_2}$-class.
We note that $\a^{(n_1)}(Z_1)$ is a straight $\psi_{m_1}$-class.
Then
\[
\a^{(n_1)}(Z_1)\cap\pi^{-1}(D)
\]
is a straight $\psi_{m_2}$-class and it contains the point $\a^{(n_1)}(\wt x)$, which
is also contained in $\a^{(n_1)}(Z_2)$.
By Proposition \ref{prp:section-unique} we have hence
\[
\a^{(n_1)}(Z_2)\subseteq \a^{(n_1)}(Z_1)
\]
and then $Z_2\subseteq Z_1$.
This completes the proof of Proposition \ref{prp:invlimlift}.
\section{Equivariance under translations}\label{sc:functor}
We recall that we denote by $\Aut_i(X)$ the group of $i$-translations of a nilspace $X$.
We endow it with the maximum displacement metric
\[
\dist(f,g)=\max_{x\in X}\{\dist(f(x),g(x))\}.
\]
We denote by $\Aut_i^\ve(X)$ the $\ve$-neighbourhood of the identity in this metric.
If $X$ is a Lie-fibered nilspace then $\Aut_i(X)$ is a Lie group (see \cite{GMV2}*{Theorem 2.18}), hence
\[
\Aut_i^{\circ}(X)=\langle \Aut_i^\ve(X)\rangle
\]
if $\ve$ is sufficiently small.
The purpose of this section is the proof of Theorem \ref{th:functoriality}, which is
an immediate consequence of the following.
\begin{prp}\label{prp:small-translations}
Let $\f:X\to Y$ be a fibration between two compact ergodic Lie-fibered nilspaces
and let $\ve>0$ be given.
Then there is a $\d>0$ depending only on $X$, $Y$, $\f$ and $\ve$
such that the following holds.
For every $f\in \Aut_i^\d(X)$ there is an $f'\in \Aut_i^\ve(Y)$, and respectively for every
$f'\in \Aut_i^\d(Y)$ there is an $f\in \Aut_i^\ve(X)$, such that
$f'\circ \f=\f\circ f$.
\end{prp}
Note there are really two distinct statements here: a ``pushing forward'' result and a ``pulling back'' one. The proofs of these will be handled separately and have different flavours. The ``pulling back'' part is really an existence fact, and will be implied fairly easily by results concerning the existence of translations from \cite{GMV2}. The ``pushing forward'' part is about proving properties of small translations -- namely, that they are compatible with the fibration in some sense -- and will require a new argument.
In both cases, the proof of this proposition is by induction on $s$, the degree of $X$.
For $s=0$ the claim is trivial, so
we fix $s\ge1$ and
assume that the proposition holds for nilspaces of degree $(s-1)$.
Recall that in Definition \ref{dfn:horizontal} we coined the notion of a \emph{horizontal fibration}. In Section \ref{sec:verhor} we introduce the complementary notion of \emph{vertical} fibrations, and show that
an arbitrary fibration may be decomposed as a composition
of a vertical and a horizontal one.
The ``pushing forward'' part of Proposition \ref{prp:small-translations} is reasonably straightforward for vertical fibrations, and in fact holds without any smallness assumption. We handle this in Section \ref{sec:vertical}. We thereby reduce to the case where $\f$ is horizontal.
We consider this case in Section \ref{sec:horizontal}. The crucial step is to show that small translations map sets of the form $\f^{-1}(y)$
onto each other.
The key observations are to note that such sets are straight classes, and that, in general, translations map
straight classes onto straight classes.
Hence, the results of the previous section can be exploited to give what we want in the case of small translations.
Finally, we prove the ``pulling back'' result in Section \ref{sec:lifting}.
\subsection{A decomposition of fibrations}\label{sec:verhor}
We recall from Section \ref{sec:straight} that the shadow of a fibration $\f:X\to Y$ between
compact ergodic nilspaces of degree $s$ is the unique fibration $\psi:\pi(X)\to\pi(Y)$ that satisfies $\pi\circ\f=\psi\circ\pi$.
We continue to use our convention that $\pi$ abbreviates $\pi_{s-1}$ as we do not use the
other canonical projections.
We also recall that a fibration $\f$ is called \emph{horizontal} if it has relative $s$-uniqueness; or equivalently, if it is injective on fibres of $\pi$ (see Definition \ref{dfn:horizontal} and the remarks that follow).
Finally, we recall from \cite{GMV1}*{Definition 7.18}
that a fibration $\f \colon X \to Y$ is called \emph{relatively $k$-ergodic} if whenever $c \colon \{0,1\}^k \to X$ is a configuration such that $\f \circ c \in C^k(Y)$ then $c \in C^k(X)$; i.e.~if all $k$-configurations in $X$ are cubes provided they map to cubes of $Y$.
The complementary notion to a horizontal fibration is as follows.
\begin{dfn}
We say that a fibration $\f:X\to Y$ between ergodic compact nilspaces of degree $s$ is {\bf vertical} if any of the following equivalent conditions holds:
\begin{enumerate}
\item given $x_1, x_2 \in X$ such that $\pi(\f(x_1)) = \pi(\f(x_2))$, we must have $\pi(x_1) = \pi(x_2)$;
\item the shadow of $\f$ is an isomorphism $\pi(X) \xrightarrow{\sim} \pi(Y)$;
\item $\f$ is relatively $s$-ergodic;
\item we have that $x_1 \sim_{\f,s-1} x_2$ for any $x_1, x_2 \in X$ such that $\f(x_1) = \f(x_2)$ (again, see \cite{GMV1}*{Section 7.2} for a definition);
\item for any $x_1,x_2 \in X$ such that $\f(x_1) = \f(x_2)$, the configuration $\llcorner^s(x_1;x_2)$ is a cube.
\end{enumerate}
\end{dfn}
The equivalence of (3), (4) and (5) is covered in \cite{GMV1}*{Section 7.2}.
Now, (1) states precisely that the shadow of $\f$ is injective, and any fibration is an isomorphism if and only if it is injective; so (1) and (2) say the same. Clearly, (1) implies (5); and (3) implies (1), since (by Lemma \ref{lem:alter-canonical}) $\pi(\f(x_1)) = \pi(\f(x_2))$ if and only if $\llcorner^s(\f(x_1); \f(x_2))$ is a cube, which holds if and only if $\llcorner^s(x_1; x_2)$ is (assuming (3)), which implies $\pi(x_1) = \pi(x_2)$.
Examples of such fibrations are quotient maps by subgroups of the top structure group.
In fact, it turns out that these are the only examples.
The main result of this section is the following decomposition result.
\begin{prp}\label{prp:verhor}
Let $\f:X\to Y$ be a fibration between two compact ergodic nilspaces of degree $s$.
Then there is a compact ergodic nilspace $Z$, a vertical fibration $\f_v: X\to Z$
and a horizontal fibration $\f_h:Z\to Y$ such that $\f=\f_h\circ\f_v$.
\end{prp}
\begin{proof}
This is immediate from \cite{GMV1}*{Proposition 7.12}.
This states that there is a decomposition
\[
\f \colon X \xrightarrow{\pi_{\f,s-1}} X / \sim_{\f,s-1} \xrightarrow{g} Y
\]
where $\sim_{\f,s-1}$ is the canonical equivalence relation attached to the fibration $\f$, and that both of these maps are fibrations. It is immediate from the definition of $\sim_{\f,s-1}$ that the quotient map is relatively $s$-ergodic; and it follows from \cite{GMV1}*{Remark 7.9}
that $Z = X / \sim_{\f,s-1}$ is a nilspace (and also trivially compact and ergodic).
\end{proof}
We make a final remark before proceeding. We note that the relation $\sim_{\f,s-1}$ is finer than $\sim_{s-1}$ on $X$ (which is immediate from the definitions), and hence $\pi$ factors as
\[
\pi \colon X \xrightarrow{\pi_{\f,s-1}} \pi_{\f,s-1}(X) \to \pi(X).
\]
Also, the relative structure theorem \cite{GMV1}*{Theorem 7.19}
states that $X$ admits a free continuous action by a compact abelian group $A_s(\f)$ whose orbits are the fibers of $\pi_{\f,s-1}$.
It is not quite immediate from this, but is nonetheless true and not hard to argue, that when $X$ is a nilspace of degree $s$ this group $A_s(\f)$ may be identified with a closed subgroup of the top structure group $A_s$. Hence, any vertical fibration is the quotient of $X$ by a subgroup of the top structure group. However, we will not explicitly need such a result.
\subsection{The case of vertical fibrations}\label{sec:vertical}
In this section, we prove the following.
\begin{prp}\label{prp:translation-vertical}
Let $\f:X\to Y$ be a vertical fibration between two compact ergodic nilspaces.
Then there is a continuous homomorphism $\psi: \Aut_i(X)\to\Aut_i(Y)$
such that $\psi(f)(\f(x))=\f(f(x))$
for all $x\in X$ and $f\in \Aut_i(X)$.
\end{prp}
We begin with a simple lemma which gives a condition for a
translation to descend to a factor through a fibration.
\begin{lem}\label{lem:universal2}
Let $\f:X\to Y$ be a fibration between two compact nilspaces
and let $f\in \Aut_i(X)$ be a translation.
Suppose that for every $y_1\in Y$ there is a $y_2\in Y$ such that
$f(\f^{-1}(y_1))=\f^{-1}(y_2)$.
Then there is a unique translation $f'\in \Aut_i(Y)$ such that $f'\circ\f=\f\circ f$.
\end{lem}
\begin{proof}
We apply Lemma \ref{lem:universal1} for the fibrations $\f: X\to Y$
and $\f\circ f:X\to Y$ and deduce that a unique fibration $f':Y\to Y$ exists
satisfying $f'\circ\f=\f\circ f$.
We show that $f'\in \Aut_i(Y)$.
To this end, we fix a cube $c\in C^{\ell}(Y)$ and let $\wt c$ be a $\f$-preimage of $c$ in $C^{\ell}(X)$.
Let $F\subseteq\{0,1\}^\ell$ be a face of codimension $i$.
Then $[f']_F.c=\f([f]_F.\wt c)$ is a cube, showing that $f'$ is indeed a translation.
\end{proof}
It is possible to verify the condition of Lemma \ref{lem:universal2} directly in the case of vertical fibrations. Many approaches are possible here; ours is fairly direct, using minimal structure theory.
Recall that we write $\llcorner^k(x;y)$ to denote the $k$-configuration given by $\vec1 \mapsto y$ and $\o \mapsto x$ for all $\o \ne \vec1$. We also introduce the notation $\square^k(x)$ to denote the constant $k$-cube $\omega \mapsto x$.
\begin{lem}\label{lem:vertical-respects-fibers}
Suppose $\f \colon X \to Y$ is a horizontal fibration between compact ergodic nilspaces of degree $s$, and suppose $f \in \Aut_1(X)$ is an $1$-translation.
Let $x,x' \in X$ be such that $\f(x) = \f(x')$. Then $\f(f(x)) = \f(f(x'))$.
\end{lem}
\begin{proof}
By relative $s$-uniqueness of $\f$, we have that $\llcorner^s(x;x')$ is an $s$-cube, and hence $c = [\llcorner^s(x;x'), \llcorner^s(f(x), f(x'))]$ is an $(s+1)$-cube, since $f$ is a $1$-translation.
Now let $\tilde{c} = [\square^s(\f(x)), \square^s(\f(f(x)))]$. This is a cube of $Y$ (by ergodicity); and moreover, $c|_{\{0,1\}^{s+1} \sm \{\vec1\}}$ and $\tilde{c}$ form a compatible $(s+1)$-corner for $\f$. Since $\f$ is a fibration, we may complete $c|_{\{0,1\}^{s+1} \sm \{\vec1\}}$ to a cube $c'$ such that $\f(c'(\vec1)) = \f(f(x))$.
But $X$ has $(s+1)$-uniqueness, and hence $f(x') = c(\vec1) = c'(\vec1)$, which gives the result.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prp:translation-vertical}]
Combining Lemma \ref{lem:universal2} and Lemma \ref{lem:vertical-respects-fibers}, we have shown that for all $f \in \Aut_i(X)$ there exists an unique $\psi(f) \in \Aut_i(Y)$ such that $\psi(f) \circ \f = \f \circ f$. It is routine to verify that $\psi$ must define a continuous group homomorphism.
\end{proof}
\subsection{Horizontal fibrations}\label{sec:horizontal}
In this section, we prove the following.
\begin{prp}\label{prp:small-translations-horizontal}
Let $\f:X\to Y$ be a horizontal fibration between two compact ergodic Lie-fibered nilspaces
of degree $s$,
and let $\ve>0$ be given.
Then there is a $\d>0$ depending only on $X$, $Y$, $\f$ and $\ve$
such that the following holds.
For every $f\in \Aut_i^\d(X)$ there is $f'\in \Aut_i^\ve(Y)$ such that
$f'\circ \f=\f\circ f$.
\end{prp}
Recall our induction hypothesis,
that Proposition \ref{prp:small-translations} holds for nilspaces of degree at most $(s-1)$.
We fix a horizontal fibration $\f: X\to Y$ between two compact ergodic Lie-fibered nilspaces of degree $s$
and a parameter $\ve_0>0$.
Denote by $\psi:\pi(X)\to\pi(Y)$ the shadow of $\f$.
We choose a sufficiently small number $\d_0$ such that
Proposition \ref{prp:small-translations} holds for
$\psi,\pi(X),\pi(Y),\ve_0,\d_0$.
Proposition \ref{prp:small-translations-horizontal} is an immediate consequence of
Lemma \ref{lem:universal2} and the following lemma:
\begin{lem}
If $\d$ is sufficiently small, then
for any point $y\in Y$ there is a point $z\in Y$ such that we have $f(\f^{-1}(y))=\f^{-1}(z)$.
\end{lem}
\begin{proof}
Let $f\in \Aut_i^\d(X)$.
Write $g\in \Aut_i(\pi(X))$ for the shadow of $f$.
If $\d$ is sufficiently small, then $ g\in \Aut_i^{\d_0}(\pi(X))$.
Hence there is a translation $g'\in \Aut_i^{\ve_0}(\pi(Y))$
such that $g'(\psi(x))=\psi(g(x))$
for all $x\in \pi(X)$.
Thus for every $y_1\in \pi(Y)$ there is $z_1\in \pi(Y)$ such that
$g(\psi^{-1}( y_1))=\psi^{-1}(z_1)$.
Recall from Lemma \ref{lem:horizontal-fib} that the inverse images of points under $\f$ are
straight $\psi$-classes in $X$.
We use that $f$ is a cubespace automorphism and the conclusion of the previous paragraph
to deduce that $D_1:=f(\f^{-1}(y))$ is also a straight $\psi$-class.
Now fix a point $x_0\in\f^{-1}(y)$, let $z=\f(f(x_0))$ and let $D_2=\f^{-1}(z)$.
Then $D_1$ and $D_2$ are both straight $\psi$-classes, and $f(x_0)\in D_1\cap D_2$.
Recall that Proposition \ref{prp:section-unique} gives conditions under which two straight classes are always either identical or disjoint; if these conditions hold, we have that $D_1 = D_2$ and hence $f(\f^{-1}(y)) = \f^{-1}(z)$ as required.
So, its suffices to check the hypotheses of Proposition \ref{prp:section-unique}.
We fix a parameter $\kappa>0$ to be specified later.
Since $\f$ is continuous, it follows that for any $x\in X$ we have $\dist(\f(f(x)),\f(x))\le\kappa$
provided $\d$ is sufficiently small.
Hence for all $x\in \f^{-1}(y)$, we have $\dist(\f(f(x)),y)\le \kappa$.
In particular, $\dist(z,y)\le \kappa$, hence $\dist(\f(f(x)),z)\le 2\kappa$ for all $x\in\f^{-1}(y)$.
Let now $x_1\in D_1$, $x_2\in D_2$ with $\pi(x_1)=\pi(x_2)$.
Denote by $a\in A$ the element of the top structure group such that $a.x_1=x_2$.
Since $\f$ is a horizontal fibration, it follows that $a.\f(x_1)=\f(x_2)=z$.
Since $x_1=f(x)$ for some $x\in \f^{-1}(y)$ it follows that $\dist(\f(x_1),\f(x_2))<2\kappa$.
Denote by $\d_1$ the number $\d$ from Proposition \ref{prp:section-unique}.
There is a number $\kappa'>0$ such that $\dist(b.x,x)<\d_1$ for all $x\in X$
whenever $\dist(b,0)<\kappa'$.
Furthermore, we can set $\kappa>0$ in such a way that $\dist(b,0)<\kappa'$ for all $b\in A$
if there is $y'\in Y$ such that $\dist(b.y',y')<2\kappa$.
This choice of $\kappa$ implies that $\dist(a,0)<\kappa'$ and hence $\dist(x_1,x_2)<\d_1$, as required.
\end{proof}
\subsection{Pulling back translations}\label{sec:lifting}
All that now remains is the ``pulling back'' component of Proposition \ref{prp:small-translations}, which we now recall.
\begin{prp}\label{prp:lifting}
Let $\f:X\to Y$ be a fibration between two compact ergodic Lie-fibered nilspaces
and let $\ve>0$ be number.
Then there is a $\d>0$ depending only on $X$, $Y$, $\f$ and $\ve$
such that the following holds.
For every $f'\in \Aut^\d_i(Y)$ there is $f\in \Aut^\ve_i(X)$ such that
$f'\circ \f=\f\circ f$.
\end{prp}
We recall the following result form \cite{GMV2}*{Lemmas 3.10 and 3.4}
that will be used in the proof.
\begin{thm}
Let $X$ be a compact ergodic Lie-fibered nilspace.
Then $\Aut_i^\circ(X)$ acts transitively on each connected component of the equivalence classes
of the $(i-1)$-th canonical equivalence relation $\sim_{i-1}$.
In fact, the following strengthening is true. For every $\ve>0$ there is $\d>0$ such that
the following holds:
for every two points $x_1,x_2\in X$ satisfying $x_1\sim_{i-1} x_2$ and $\dist(x_1,x_2)<\d$,
there is a translation $f\in \Aut_i^\ve(X)$ such that $f(x_1)=x_2$.
Also, the stabilizer $\Stab_x(\Aut_1(X))$ is discrete for any point $x\in X$.
\end{thm}
Essentially this states that, under some topological assumptions, $i$-translations are uniquely characterized by where they send a single point, and we have almost total freedom to choose that point. So, if $f'$ sends $y$ to $y'$, we can choose $f$ by insisting it maps $x$ to $x'$ for some $x \in \f^{-1}(y)$, $x' \in \f^{-1}(y')$ that we choose; and if we are sufficiently careful, this $f$ will have the required properties.
We now turn to the details.
\begin{proof}[Proof of Proposition \ref{prp:lifting}]
Fix a point $y_1\in Y$.
Let $\ve>0$ be sufficiently small such that $\Stab_{y_1}(\Aut_i^{2\ve}(Y))=\{e\}$.
Let $\d_1>0$ be sufficiently small so that for each $f\in \Aut_i^{\d_1}(X)$ there
is $f'\in \Aut_i^{\ve}(Y)$ such that $f'\circ \f=\f\circ f$.
Let $\d_2$ be sufficiently small so that for every two points
$x_1,x_2\in X$ satisfying $\dist(x_1,x_2)<\d_2$ and $x_1\sim_i x_2$,
there is a translation $f\in \Aut_i^{\d_1}(X)$ such that $f(x_1)=x_2$.
Let $\d_3$ be sufficiently small
so that for the above fixed $y_1\in Y$ and any point $y_2\in Y$
satisfying $\dist(y_1,y_2)\le \d_3$,
there are $x_1\in\f^{-1}(y_1)$ and $x_2\in\f^{-1}(y_2)$ so that
$\dist(x_1,x_2)\le\d_2$.
We show that for every $f_0\in \Aut_i^{\d_3}(Y)$, there is $f\in \Aut_i^{\d_1}(X)$
such that $f_0\circ \f=\f\circ f$.
Write $y_2=f_0(y_1)$.
Then $\dist(y_1,y_2)\le \d_3$.
Let $x_1,x_2\in X$ be such that $\f(x_i)=y_i$ for $i=1,2$, and $\dist(x_1,x_2)\le \d_2$.
Let $f\in \Aut_i^{\d_1}(X)$ be such that $f(x_1)=x_2$.
Let $f'\in \Aut_i^{\ve}(Y)$ be such that $\f\circ f=f'\circ \f$.
Then $f'(y_1)=\f(f(x_1))=y_2=f_0(y_1)$.
Hence $f_0^{-1}f'\in \Stab_{y_1}(\Aut_i^{2\ve}(Y))$ and thus $f_0=f'$.
\end{proof}
\section{Inverse limits in the dynamical category}\label{sc:dyninverse}
The purpose of this section is to prove Theorems \ref{th:dynamics} and \ref{th:dyn-alg-struc}.
We begin with the proof of the former.
Let $(H,X)$ be a minimal topological dynamical system such that $H$ has a dense subgroup generated by a compact
set $K$.
By Theorem \ref{th:dyn-nilspace},
$\RP_H^s(X)$ is a closed $H$-invariant equivalence relation.
We show that $(H, X/\RP_H^s(X))$ is a pronilsystem.
We recall from Theorem \ref{th:dyn-nilspace} that $X/\RP_H^s(X)$ equipped with its dynamical cubes $C_H^k(X/\RP_H^s)$, is an ergodic nilspace of degree
at most $s$.
We invoke Theorem \ref{th:invlim} and obtain a sequence of Lie-fibred nilspaces $\{X_n\}$ together
with an inverse system of fibrations $\{\f_{m,n}:X_n\to X_m\}$ such that $X/\RP_H^s(X)=X_\infty=\invlim X_n$.
We want to show that the action of $H$ descends to $X_n$ through the fibration $\f_{n,\infty}$ and
hence
\[
(H,X/\RP_H^s(X))=\invlim(H,X_n)
\]
holds in the dynamical category.
This follows from the following result, which we prove in Section \ref{sc:H-descends}.
\begin{thm}\label{th:dyninvlim}
Let $X$ be a compact ergodic nilspace of degree $k$ and let $H<\Aut_1(X)$ be a group of translations,
which has a dense subgroup generated by a compact set.
Then there is a number $\d>0$ depending on $X$ and $H$ such that the following holds.
Let $\f:X\to Y$ be a fibration to a nilspace $Y$ such that
$\diam(\f^{-1}(y))<\d$ for all $y\in Y$.
Then there is a continuous homomorphism $\psi:H\to \Aut_1(Y)$ such that $\psi(f)\circ\f=\f\circ f$ for all
$f\in H$.
\end{thm}
We note that $H$ immerses into $\Aut_1(X/\RP_H^s(X),C^k_H)$ by the very definition of dynamical cubespaces,
and hence Theorem \ref{th:dyninvlim} applies, at least for $n$ large enough.
Now we recall a result from \cite{GMV2}*{Corollary 2.20},
a variant of Theorem \ref{th:Lie-dynamical}, which
implies that $(H,X_n)$ is a nilsystem of degree at most $s$, thereby completing the proof that $X/\RP_H^s$ is indeed
a pronilfactor of degree at most $s$.
\begin{thm}\label{th:dyn-Lie2}
Let $(H,X)$ be a minimal topological dynamical system, where $X$ is a compact ergodic Lie-fibred
nilspace of degree $s$ and $H$ acts on $X$ through a continuous group homomorphism $\a:H\to\Aut_1(X)$.
Then $(H,X)$ is a nilsystem.
More specifically, the group $G=\langle \Aut_1^\circ(X), \a(H)\rangle$ is a nilpotent Lie group that acts
transitively on $X$.
Hence $(H,X)$ is isomorphic to $(H,G/\Gamma)$, where $\Gamma$ is the stabilizer in $G$ of an arbitrary point
and $H$ acts through the homomorphism $\a$.
\end{thm}
We prove Theorem \ref{th:dyninvlim} in Section \ref{sc:H-descends}, which completes the proof of the part
of Theorem \ref{th:dynamics} that claims that $X/\RP_H^s$ is a pronilsystem.
Then we prove that it is also the largest pronil factor in Section \ref{sc:largest}.
Finally, we prove the remaining parts of Theorem \ref{th:dyn-alg-struc} in Section \ref{sc:dyn-alg-struc}.
\subsection{Proof of Theorem \ref{th:dyninvlim}}\label{sc:H-descends}
We proceed by induction on $s$, the degree of the nilspace.
For $s=0$, the claim is trivial, hence we assume that $s\ge 1$ and the statement holds for nilspaces
of degree $(s-1)$.
We apply Proposition \ref{prp:verhor} to the fibration $\f$, hence we can write it as a composition of a
vertical fibration and a horizontal one.
Since Proposition \ref{prp:translation-vertical} does not require an assumption on connectedness or on
Lie-fiberedness,
it implies Theorem \ref{th:dyninvlim} for vertical fibrations.
Therefore, it is enough to prove Theorem \ref{th:dyninvlim} for horizontal fibrations, which is
what we do next.
Denote by $\bar\f:\pi(X)\to \pi(Y)$ the shadow of $\f$ and by $\bar H<\Aut_1(\pi(X))$ the group
of translations consisting of the shadows of the elements of $H$.
By the induction hypothesis, we know that there is a continuous homomorphism $\bar\psi:\bar H\to \Aut_1(\pi(Y))$
such that $\bar\psi(f)\circ\bar\f=\bar\f\circ f$ for all
$f\in \bar H$ provided the number $\d$ in the statement of the theorem is sufficiently small.
We assume that $\d$ is sufficiently small so that this holds.
We fix a compact set $K\subseteq H$ that generate a dense subgroup of $H$.
We first show that these translations factor through $\f$.
\begin{lem}\label{lem:compactset}
With the above notation and assumptions, for every $f_1\in K$ there is a unique $f_2\in \Aut_1(Y)$ such that
$f_2\circ\f=\f\circ f_1$, provided $\d$ is sufficiently small depending on $K$.
\end{lem}
\begin{proof}
In light of Lemma \ref{lem:universal2}, it is enough to show that for any $f\in K$ and $y_1\in Y$
there is $y_2\in Y$ such that $f(\f^{-1}(y_1))=\f^{-1}(y_2)$.
To prove this, we recall that $\f^{-1}(y)$ is a straight
$\bar\f$-class for all $y\in Y$ (see Lemma \ref{lem:horizontal-fib}).
By Proposition \ref{prp:section-unique}, we know that
there is a number $\d_1>0$ such that each point $x\in X$ is contained in
at most one straight $\bar\f$-class of diameter at most $\d_1$.
We assume as we may that $\d$ is so small that $\diam(f(D))<\d_1$
for any set $D\subseteq X$ with $\diam(D)\le\d$ and $f\in K$.
This is possible, because $K$ is compact and the action is continuous.
We also assume that $\d<\d_1$.
We fix a point $y_1\in Y$, then pick an arbitrary element $x_2\in f(\f^{-1}(y_1))$ and let $y_2=\f(x_2)$.
We observe that $\f^{-1}(y_2)$ and $f(\f^{-1}(y_1))$ are both straight $\bar \f$-classes of diameter at most
$\d_1$.
Moreover, they both contain the point $x_2$, hence they must be equal by Proposition \ref{prp:section-unique},
as we noted above.
This completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:dyninvlim} for horizontal fibrations]
For any translation $f_1\in \Aut_1(X)$, there is at most one translation $f_2\in \Aut_1(Y)$ such that
\be\label{eq:factorsthrough}
f_2\circ\f=\f\circ f_1.
\ee
For $f_1\in K$ the existence of $f_2$ satisfying \eqref{eq:factorsthrough} follows from
Lemma \ref{lem:compactset}.
If $f_1^{(1)},f_1^{(2)}\in \Aut_1(X)$ and $f_2^{(1)},f_2^{(2)}\in\Aut_1(Y)$ satisfy
the analogue of \eqref{eq:factorsthrough}, then $f_1^{(1)}\circ f_1^{(2)}$ and $f_2^{(1)}\circ f_2^{(2)}$
satisfy it, as well.
In addition, if the analogue of \eqref{eq:factorsthrough} holds for
$\{f_1^{(i)}\}_{i\in \N}\subseteq \Aut_1(X)$ and $\{f_2^{(i)}\}_{i\in \N}\subseteq\Aut_1(Y)$ for all $i\in\N$
and $f_1^{(i)}$ uniformly converges to a translation $f_1$, then $f_2^{(i)}$ also converges to
a translation and we have \eqref{eq:factorsthrough}.
It follows from the above observations that a map $\psi: H\to \Aut_1(Y)$ exists such that
$\psi(f)\circ\f=\f\circ f$ for all $f\in H$.
We leave it to the reader to verify that this is also a continuous group homomorphism.
\end{proof}
\subsection{}\label{sc:largest}
Let $Y=X/{\sim}$ be a pronilfactor of degree $s$.
We show that $\RP_H^{s}(X)\subseteq\sim$.
We denote by $\f:X\to Y$ the quotient map.
It follows directly from the definition that $\f(c)\in C_H^{s+1}(Y)$
for all $c\in C_H^{s+1}(X)$ and hence $(\f(x),\f(y))\in \RP_H^s(Y)$ for
any pair of points $(x,y)\in\RP_H^s(X)$.
Thus the claim follows from the following lemma.
\begin{lem}
Let $(H,Y)$ be a pronil system of degree $s$.
Then $\RP_H^{s}(Y)$ is trivial.
\end{lem}
\begin{proof}
It is enough to show that $\RP_H^{s}(Y)$ is trivial for a nilsystem $(H,Y)$
of degree $s$.
In this case, the dynamical cubes form a Host--Kra nilspace, which is a nilspace of degree at most $s$, as shown in \cite{GMV1}*{Proposition 2.6}.
Hence $\RP_H^{s}(Y)=\sim_s$ is indeed trivial.
\end{proof}
\subsection{Proof of Theorem \ref{th:dyn-alg-struc}}\label{sc:dyn-alg-struc}
Let $(H,X)$ be a minimal system such that $\RP_H^s(X)$ is trivial and suppose that $H$ has
a dense subgroup generated by a compact set.
This means that we are in the setting of the proof of Theorem \ref{th:dynamics}.
We have, therefore, a sequence of Lie-fibred nilspaces $\{X_n\}$ together
with an inverse system of fibrations $\{\f_{m,n}:X_n\to X_m\}$ such that $X=X_\infty=\invlim X_n$.
We have already proved that the action of $H$ factors through the fibrations $\f_{n,\infty}$, i.e.~we
have continuous homomorphisms $\a_n:H\to \Aut_1(X_n)$ such that $\a_m=\f_{m,n}\circ\a_n$ for $m<n$.
Moreover, by Theorem \ref{th:dyn-Lie2}, the group $G_n=\langle \Aut_1(X_n),\a_n(H)\rangle$
act transitively on $X_n$.
We saw in the proof of Theorem \ref{th:alg-struc} that there is a continuous homomorphism
$\psi_{m,n}:\Aut_{1}^\circ(X_n)\to\Aut_1^\circ(X_m)$ such that
\[
\psi_{m,n}(g).\f_{m,n}(x)=\f_{m,n}(g.x)
\]
for all $g\in\Aut_1^\circ(X_n)$ and $x\in X_n$.
We can extend $\psi_{m,n}$ to a homomorphism $G_n\to G_m$ by taking
$\psi_{m,n}(\a_n(h))=\a_m(h)$.
We leave it to the reader to verify that this extension satisfies the properties claimed in Theorem \ref{th:dyn-alg-struc}.
Finally, we note that the proof is valid in the slightly more general setting, when $X$ is an arbitrary
compact ergodic nilspace and $H$ acts on $X$ via a continuous homomorphism $H\to \Aut_1(X)$.
(The action has to be continuous and minimal, and $H$ has to contain a compactly generated dense subgroup.)
That is to say, the proof does not require that the cubespace structure on $X$ is defined using the dynamical
construction; it may be larger.
|
\section{Introduction}
The hysteretic constitutive relation between the magnetization and the magnetic field in ferromagnets remains one of the main difficulties in
electromagnetic modeling. The Preisach model \cite{Mgoyz}, providing for, probably, the most accurate macroscopic description for ferromagnetic hysteresis at present, is a black-box-type method for
storing, and using for interpolation, a vast amount of experimental data necessary for the
implementation of this model.
A simpler and very popular Jiles--Atherton model \cite{JA} needs a patch to avoid a nonphysical behaviour \cite{Carp,MZ}; the physical arguments used for the derivation of this model have been criticized in \cite{Zirka}. Both the Preisach and Jiles--Atherton models are scalar and, although there exist numerous vector modifications, these also lack a true physical justification. Furthermore, in a general situation, the use of these models to predict the evolving magnetization does not make computing the accompanying energy loss straightforward (see, e.g., \cite{FrMgoyz}).
In a seminal work \cite{B97}, Bergqvist proposed a new quasi-static magnetic hysteresis model, phenomenological but having a consistent and genuine energy interpretation, intrinsically vectorial, and ensuring a direct calculation of the stored magnetic energy and the dissipated energy at any moment in time, and not only after the completion of a closed hysteresis loop as is usually the case. This model differs significantly from the previous ones but, like the Jiles--Atherton model, regards the pinning of domain walls as the cause of hysteresis and presents, similarly to the Preisach model, the complex hysteretic behavior as a superposition of reactions of simple hysteretic elements, ``pseudoparticles". Later, the Bergqvist model \cite{B97} and models, closely related to it, have been considered in a series of works;
see \cite{B97b,KrahB04,HenrNicHam2006,HenrHam2006,SEH12,FLavHenr2013,HSHG2014,B14} and the references therein.
To make the magnetization update at each time step explicit Bergqvist \cite{B97} employed an approximation, turning his vectorial energy-based model into a vector play hysteron model. Such an approximation was used also in almost all following works: the only exception that we know is \cite{FLavHenr2013}, where an optimization problem is solved to find the new value of the magnetization. Although in the scalar case this approximation does not introduce any error at all, in the general vectorial case it leads to an error that does not disappear as the time steps (external field increments) tend to zero.
In this work we avoid such an approximation and propose a more efficient numerical method than in \cite{FLavHenr2013}.
We start with the derivation, and a discussion, of a simplified variational hysteresis model in order to clarify its mathematical structure, then make the model more realistic. We identify the parameters of this model for nonoriented electrical steel using a set of experimental first order reversal curves. Finally, we implement the model as a constitutive relation in a finite element simulation taking into account both the quasi-static hysteretic magnetization and the eddy current.
\section{Energy balance and dry-friction like model of magnetization}
The magnetostatic field energy in a magnetic material can be presented as a sum of the empty space energy, depending on the magnetic field $\bm{h}$, and the internal energy determined by the material magnetization $\bm{m}$. The energy density, \begin{equation} W=\frac{1}{2}\mu_0h^2+U(\bm{m}),\label{En}\end{equation} changes as
\begin{equation} \dot{W}=\bm{h}\cdot\dot{\bm{b}}-|r\dot{\bm{m}}|,\label{En_t}\end{equation}
where $\bm{b}=\mu_0(\bm{h}+\bm{m})$ is the magnetic induction, $\mu_0$ is the permeability of vacuum, $\bm{h}\cdot\dot{\bm{b}}$ is the rate of the magnetic field work, and $|r\dot{\bm{m}}|$ is the rate of dissipation caused by the irreversible movement of the domain walls accompanying the changes in magnetization \cite{B97}. For an isotropic material the ``friction coefficient" $r$ is a positive scalar; otherwise it is a symmetric positive definite matrix.
Here and below the time derivative of $\bm{u}$ is denoted as $\dot{\bm{u}}$ and, if $\bm{u}$ is a vector, $u$ means $|\bm{u}|$.
Equations (\ref{En}) and (\ref{En_t}) yield
$\mu_0\bm{h}\cdot\dot{\bm{h}}+\bm{\nabla} U(\bm{m})\cdot\dot{\bm{m}} =\mu_0\bm{h}\cdot(\dot{\bm{h}}+\dot{\bm{m}})-|r\dot{\bm{m}}|$
or
\begin{equation} (\bm{h}-\bm{f}(\bm{m}))\cdot\dot{\bm{m}}=|k\dot{\bm{m}}|,\label{k}\end{equation}
where $\bm{f}(\bm{m})=\frac{1}{\mu_0}\bm{\nabla} U(\bm{m})$ and $k=\frac{1}{\mu_0}r$.
Unlike the Jiles--Atherton model, where the magnetization $\bm{m}$ is assumed to be a sum of its reversible and irreversible parts, the Bergqvist model of hysteresis uses a similar representation for the magnetic field; this difference is crucial. The field $\bm{h}_r=\bm{f}(\bm{m})$ is called reversible because the magnetic work it delivers is fully converted into internal energy; the remaining field $\bm{h}_i=\bm{h}-\bm{h}_r$ is called irreversible. Equation (\ref{k}) then takes the form \begin{equation} \bm{h}_i\cdot\dot{\bm{m}}=|k\dot{\bm{m}}|.\label{ki}\end{equation}
For an isotropic material, (\ref{ki}) is satisfied if the following ``dry-friction" constitutive relation is postulated:
\begin{equation}
\begin{array}{l} |\bm{h}_i|\leq k;\\ \mbox{if}\ |\bm{h}_i|< k\ \mbox{then}\ \dot{\bm{m}}={\bm{0}};\\ \mbox{if}
\ \dot{\bm{m}}\neq{\bm{0}}\ \mbox{it has the direction of}\ \bm{h}_i.\end{array} \label{cr}
\end{equation}
We note that this multivalued relation is similar to the relation between the rate of plastic deformation and stress in an elasto-plastic material with the yield strength $k$.
To obtain a more convenient formulation of (\ref{cr}) we note that $\bm{h}_i\in \widetilde{K}:=\left\{\bm{u}\in \mathbb{R}^3\ :\ |\bm{u}|\leq k\right\}$ and recall the notion of a subdifferential from convex analysis. Let $f:\ \mathbb{R}^n\rightarrow \mathbb{R}\bigcup\{+\infty\}$ be a convex function which may take also the $+\infty$ values. The set \begin{eqnarray*}\partial f(\bm{x}):=\{\bm{p}\in \mathbb{R}^n\ :\ f(\bm{y})\geq \ f(\bm{x})+\bm{p}\cdot(\bm{y}-\bm{x})\\ \mbox{for all}\ \bm{y}\in \mathbb{R}^n\}\end{eqnarray*} is called the subdifferential of $f$ at the point $\bm{x}$; its elements $\bm{p}\in \partial f(\bm{x})$ are subgradients of $f$ at $\bm{x}$. If $f$ is differentiable at $\bm{x}$ then
$\partial f(\bm{x})=\{\nabla f(\bm{x})\}$ and, if $f(\bm{x})=+\infty$, $\partial f(\bm{x})$ is an empty set.
In addition, if $\bm{0}\in \partial f(\bm{x})$ then $f(\bm{x})\leq f(\bm{y})$ for all $\bm{y}$.
It is not difficult to find the subdifferential of the indicator function of the set $\widetilde{K}$,
$$I_{\widetilde{K}}(\bm{x})=\left\{\begin{array}{lr}0&\bm{x}{\in}\widetilde{K},\\ \infty& \bm{x}{\not \in} \widetilde{K}.\end{array}\right.$$
For $\bm{x}\in \widetilde{K}$ we obtain $\bm{p}\in \partial I_{\widetilde{K}}(\bm{x})$ if $\bm{p}\cdot(\bm{y}-\bm{x})\leq 0$ for any $\bm{y}\in \widetilde{K}$. Clearly, if $|\bm{x}|<k$ this condition holds only for $\bm{p}=\bm{0}$ and, if $|\bm{x}|=k$, $\bm{p}$ can be any vector of the same direction as $\bm{x}$. Hence, the
conditions in (\ref{cr}) can be written as \begin{equation} \dot{\bm{m}}\in \partial I_{\widetilde{K}}(\bm{h}_i).\label{isoI}\end{equation}
It follows from the definition of a subdifferential that while $\bm{h}_i$ belongs to the interior of the set $\widetilde{K}$, i.e. $|\bm{h}_i|<k$, the magnetization does not change: $\dot{\bm{m}}=\bm{0}$. Whereas, if $|\bm{h}_i|=k$ then (\ref{isoI}) determines the unique direction of $\dot{\bm{m}}$, since $\partial I_{\widetilde{K}}(\bm{h}_i)=\{\bm{u}\in \mathbb{R}^3\ :\ \bm{u}=\lambda\bm{h}_i,\ \lambda\geq 0\}$.
Until now, the ``dry friction law" was not defined precisely. Now we explain our choice of (\ref{cr}), which does not follow from (\ref{ki}) since it is not the only constitutive relation for which (\ref{ki}) holds. According to a general definition by Moreau (\cite{Moreau}, p. 64), to set a dry friction relation between the irreversible field $\bm{h}_i$ (the ``friction force") and the magnetization velocity $\dot{\bm{m}}$, it is required to define a closed convex set of admissible irreversible fields, $\widetilde{K}$, and postulate the maximal dissipation principle: for a given $\dot{\bm{m}}$ the field $\bm{h}_i$ should
maximize the dissipation power $\mu_0\bm{h}_i\cdot\dot{\bm{m}}$ in the set $\widetilde{K}$. Such a relation between $\dot{\bm{m}}$ and $\bm{h}_i$ is equivalent to (\ref{isoI}), which is equivalent to (\ref{cr}).
In the anisotropic case
we also postulate that $\dot{\bm{m}}\in \partial I_{\widetilde{K}}(\bm{h}_i)$, where now, since $k$ is a symmetric positive definite matrix, $\widetilde{K}:=\left\{\bm{u}\in \mathbb{R}^3\ :\ |k^{-1}\bm{u}|\leq 1\right\}.$
In this case, rewriting (\ref{ki}) as $k^{-1}\bm{h}_i\cdot k\dot{\bm{m}}=|k\dot{\bm{m}}|$ we see that this equality holds, since
$\dot{\bm{m}}\in \partial I_{\widetilde{K}}(\bm{h}_i)$ means that
\begin{equation} \begin{array}{c}\bm{h}_i\in \widetilde{K}\ \mbox{is such that}\\ \dot{\bm{m}}\cdot(\bm{u}-\bm{h}_i)\leq 0\ \mbox{for any}\ \bm{u}\in\widetilde{K},\end{array}\label{in0} \end{equation}
which is equivalent to the multivalued constitutive relation
$$\begin{array}{l}|k^{-1}\bm{h}_i|\leq 1;\\ \mbox{if}\ |k^{-1}\bm{h}_i|< 1\ \mbox{then}\ \dot{\bm{m}}=\bm{0};\\ \mbox{if}\ \dot{\bm{m}}\neq \bm{0}\ \mbox{then}\ k\dot{\bm{m}}\ \mbox{has the direction of}\ k^{-1}\bm{h}_i.\end{array}$$
Note that if $\bm{h}_i\in\widetilde{K}$ then the reversible field $\bm{h}_r=\bm{h}(t)-\bm{h}_i$ belongs to the set $$K(t):=\{\bm{u}\in\mathbb{R}^3\ :\ |k^{-1}(\bm{h}(t)-\bm{u})|\leq 1\}$$ and the inequality (\ref{in0}) can be rewritten for $\bm{h}_r$:
\begin{equation} \begin{array}{c}\bm{h}_r\in K(t)\ \mbox{is such that}\\ \dot{\bm{m}}\cdot(\bm{u}-\bm{h}_r)\geq 0\ \mbox{for any}\ \bm{u}\in K(t).\end{array} \label{in}\end{equation}
Inverting the dependence $\bm{h}_r=\bm{f}(\bm{m})$ we obtain that $\bm{m}=\bm{f}^{-1}(\bm{h}_r)$ and, as in \cite{B97}, assume further that the vectors $\bm{h}_r$ and $\bm{m}$ are parallel, i.e.
$\bm{m}=M_{an}(h_r)\frac{\bm{h}_r}{h_r},$ where the anhysteretic function $M_{an}$
is non-decreasing and $M_{an}(0)=0$.
Let $S(\bm{u})=\int_0^uM_{an}(s)ds$. Then
\begin{equation}\bm{m}=\bm{\nabla} S(\bm{h}_r).\label{nS}\end{equation}
To solve (\ref{in})--(\ref{nS}) numerically, we substitute (\ref{nS}) into the discretized version of (\ref{in}), $$\begin{array}{c}\bm{h}_r\in K(t)\ \mbox{is such that}\\ (\bm{m}-\check{\bm{m}})\cdot(\bm{u}-\bm{h}_r)\geq 0\ \ \mbox{for any}\ \bm{u}\in K(t),\end{array}$$
where ``$\check{\ }$" means the value from the previous time level. This yields the variational inequality
\begin{equation}\begin{array}{c} \mbox{find}\ \bm{h}_r\in K(t)\ \mbox{such that}\\ (\bm{\nabla} S(\bm{h}_r)-\check{\bm{m}})\cdot(\bm{u}-\bm{h}_r)\geq 0\\ \mbox{for any}\ \bm{u}\in K(t),\end{array}\label{vi_hr}\end{equation}
which is equivalent to an optimization problem: $\bm{h}_r(t)$ is a solution of
\begin{equation} \min_{\bm{u}\in K(t)}\{S(\bm{u})-\check{\bm{m}}\cdot\bm{u}\}.\label{opt}\end{equation}
It can be shown that if the derivative $M'_{an}>0$ then $S$ is a strictly convex function and, since the set $K(t)$ is convex, (\ref{opt}) has a unique solution.
The unconstrained minimum of $S(\bm{u})-\check{\bm{m}}\cdot\bm{u}$ is at a point $\bm{u}$ where $\bm{\nabla} S(\bm{u})-\check{\bm{m}}=0$;
in this case $\bm{u}=\check{\bm{h}}_r$. Hence, if $|k^{-1}(\bm{h}(t)-\check{\bm{h}}_r)|\leq 1$, this is a solution also to the constrained problem (\ref{opt}) and $\bm{m}=\check{\bm{m}}$. Otherwise, the equality constraint
$|k^{-1}(\bm{h}(t)-\bm{h}_r(t))|= 1$ holds. We use this observation to solve the optimization problem (\ref{opt}) numerically as follows.
At each time level, the solution to (\ref{opt}) is $\bm{h}_r(t)=\check{\bm{h}}_r$ if $|k^{-1}(\bm{h}(t)-\check{\bm{h}}_r)|\leq 1$. In 2d problems, if this inequality is not true, $\bm{h}_r=\bm{h}(t)+k\bm{i}_{\phi}$,
where $\bm{i}_{\phi}=(\cos\phi,\,\sin\phi)$ is a unit vector and, therefore, one is required to solve an unconstrained 1d minimization problem
\begin{equation} \min_{\phi}\{S(\bm{h}(t)+k\bm{i}_{\phi})-\check{\bm{m}}\cdot(\bm{h}(t)+k\bm{i}_{\phi})\}.\label{2dopt}\end{equation}
Although there can be several local minima, a good initial approximation to the optimal direction $\phi$ is the direction of the vector $k^{-1}(\check{\bm{h}}_r-\bm{h}(t))$. Starting from this approximation, we solved the problem $g'(\phi)=0$, where $g(\phi)=S(\bm{h}(t)+k\bm{i}_{\phi})-\check{\bm{m}}\cdot(\bm{h}(t)+k\bm{i}_{\phi})$, efficiently using Newton's method (see Appendix A). Usually, two or three iterations of this method have been sufficient to find the solution with high accuracy. Although we have only solved 2d problems, this method should be efficient also in 3d problems, where the optimal direction is determined by two angles.
We note that in \cite{B97,B97b,KrahB04,HenrNicHam2006,HenrHam2006,SEH12,HSHG2014,B14} the vector
\begin{equation} \bm{h}_r=\bm{h}+k\frac{\check{\bm{h}}_r-\bm{h}(t)}{|\check{\bm{h}}_r-\bm{h}(t)|}\label{Appr1}\end{equation}
is chosen (in the isotropic case) as the new value of $\bm{h}_r$ if $|\bm{h}(t)-\check{\bm{h}}_r|> k$. This is equivalent to using our initial approximation for $\phi$ without any further correction and turns (\ref{in})--(\ref{nS}) into a vector play model. In the vectorial case such an approach can introduce an error that does not disappear as the increments of $\bm{h}$ tend to zero (see below).
The hysteresis model (\ref{in})--(\ref{nS}) is oversimplified but it will be used as a building block for a more realistic model (Section III). First, it seems instructive to illustrate the behaviour of this model by several examples. Let us assume, as in \cite{B97}, that
\begin{equation} M_{an}(h_r)=\frac{2 m_s}{\pi}\arctan\left(\frac{h_r}{A}\right),\label{Man}\end{equation} where $m_s$ is the saturation magnetization and the parameter $A$ determines the steepness of the curve. Another popular representation of the anhysteretic curve (see, e.g., \cite{HenrNicHam2006,HenrHam2006,dARTV}) is the Langevin function $M_{an}(h_r)=m_s[\coth(h_r/B)-B/h_r]$, which is very well approximated by (\ref{Man}) if $A=1.7B$. In general, the curve can be approximated by a spline (and we will use a spline representation of $M_{an}$ to model nonoriented electrical steel in Section \ref{Sec4}). We found that for $m_s=1.23\cdot 10^6$ A/m, $A=38$ A/m and $k=71$ A/m the model (\ref{in})--(\ref{nS}) with (\ref{Man}) describes well the major hysteresis loop shown in figure 5 of \cite{B14} (here $k$ determines the loop width which is almost constant except close to saturation, where it quickly drops to zero).
Let $\bm{m}(0)=\bm{0}$. First, we set $\bm{h}=(H_m\sin t,0)$. This example is one-dimensional, the approximation (\ref{Appr1}) does not introduce any error and, furthermore, the problem (\ref{2dopt}) can be solved analytically. We used it to check our optimization procedure. Here, and throughout this section and the next, the time step is chosen sufficiently small, about
200--400 time steps per cycle, so that the shown figures are independent of the time step.
The simulation results for two values of the amplitude $H_m$ (Fig. \ref{Fig1}) show that, although the model's prediction of the major hysteresis loop is correct, the minor loop and the initial magnetization curve are unrealistic.
\begin{figure}[!h]
\centering
\includegraphics[width=8cm,height=7.3cm]{Fig3_1d_sin.pdf}\caption{Two hysteresis loops, simplified model; $\bm{h}=(H_m\sin t,0)$.\label{Fig1}}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=8cm]{Fig4_spir_bx1.pdf}
\caption{Model solution $\bm{m}=(m_x,m_y)$ (solid line) and the explicit approximation (\ref{Appr1}) (dashed line);
$\bm{h}=H_m(t)(\cos t,\sin t)$, where $H_m(t)=110\min(t/6\pi,1)$ A/m, $k=71$ A/m.}\label{Fig2}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=8cm]{Fig5_spir_bx3.pdf}\\ \includegraphics[width=8cm]{Fig5_spir_bx1aniso.pdf}
\caption{Similar to Fig. \ref{Fig2} except: top -- for $\bm{h}=H_m(t)(3\cos t,\sin t)$; bottom -- for an anisotropic material characterised by the diagonal matrix $k=\mbox{diag}(71,35.5)$ A/m.}\label{Fig3}
\end{figure}
In our next example (Fig. \ref{Fig2}) we assume the magnetic field rotates, $\bm{h}=H_m(t)(\cos t,\sin t)$, with the amplitude $H_m(t)=110\min(t/6\pi,1)$ growing with time until its maximal value 110 A/m is reached. This is a non-scalar situation and we compare the accurate numerical solution of (\ref{2dopt}), equivalent to the time discretized version of (\ref{in}), to the explicit, at each time step, discretized vector play model based on the approximation (\ref{Appr1}). Although the solutions are different in the transient regime, the difference is small and disappears soon after the amplitude of the rotating magnetic field becomes constant. However, the approximation (\ref{Appr1}) is less accurate if the amplitudes of the magnetic field components $h_x$ and $h_y$ are different (Fig. \ref{Fig3}, top) or the material is anisotropic (Fig. \ref{Fig3}, bottom).
\section{A more realistic composite model of a ferromagnetic material}
Prior to using the described energy-based dry-friction like model for modeling hysteresis in real ferromagnets,
this model should be made more realistic. The main modifications, at least partially implemented in all works where such a model has been used, have been suggested already in \cite{B97,B97b}; their analogues can be found also in some previous models of hysteresis.
First, instead of a single value of the ``friction coefficient" $r$, the material can be characterized by a distribution of $r$ values with the volume density $\omega(r)$; this approach corresponds better to the statistical distribution of the pinning center strengths in the ferromagnetic microstructure. The total magnetization is $\bm{m}=\int \bm{m}^{r}\omega(r)dr$, where each moment $\bm{m}^{r}(t)$ obeys the dry friction model with its own value of $r$. This can improve the description of the initial magnetization curve and the minor loops.
For numerical simulations we approximate the distribution by a mixture of $N$ types of pseudoparticles with volume fractions $\omega^l>0$, satisfying $\sum_{l=1}^N\omega^l=1$. Each type is characterised by its own $r=r^l$ and, to account for partial reversibility of the material response \cite{B97,SEH12} we assign $r=0$ to one of the pseudoparticle types. Overall, we assume
$$\begin{array}{c}W=\frac{1}{2}\mu_0h^2+\sum_{l=1}^N \omega^lU(\bm{m}^l),\\ \bm{b}=\mu_0(\bm{h}+\sum_{l=1}^N \omega^l\bm{m}^l),\\ \dot{W}=\bm{h}\cdot\dot{\bm{b}}-\sum_{l=1}^N \omega^l|r^l\dot{\bm{m}}^l|\end{array}$$
and arrive at an analogue of (\ref{k}),
\begin{equation} \sum_{l=1}^N\omega^l\left\{(\bm{h}-\bm{f}(\bm{m}^l))\cdot \dot{\bm{m}}^l-|k^l\dot{\bm{m}}^l|\right\}=0,
\label{kj}\end{equation}
where $k^l=\frac{1}{\mu_0}r^l$ and, as before, $\bm{f}=\frac{1}{\mu_0}\bm{\nabla} U$. We set $\bm{h}_r^l=\bm{f}(\bm{m}^l)$, $\bm{h}_i^l=\bm{h}-\bm{h}_r^l$ and, similarly to what was done above, satisfy (\ref{kj}) by postulating the constitutive relations
$$\dot{\bm{m}}^l\in \partial I_{\widetilde{K}^l}(\bm{h}_i^l),$$ where $$\widetilde{K}^l:=\{\bm{u}
\in {\mathbb R}^3
\ : \ |(k^l)^{-1}\bm{u} |\leq 1\};$$ if $k^l=0$ we assume $\widetilde{K}^l:=\{\bm{0}\}$. As before, we reformulate these conditions as variational inequalities, similar to (\ref{in})--(\ref{nS}). After discretization in time these inequalities become equivalent to optimization problems similar to (\ref{opt}): we find $\bm{h}_r^l$ on a new time level as a solution to
\begin{equation} \min_{\bm{u}\in K^l(t)}\{S(\bm{u})-\check{\bm{m}}^l\cdot\bm{u}\},\label{optj}\end{equation}
where $$K^l(t):=\{\bm{u} \in {\mathbb R}^3\ :\ |(k^l)^{-1}(\bm{h}(t)-\bm{u})|\leq 1\}$$ except for $k^l=0$: in that case $K^l(t):=\{\bm{h}(t)\}$.
Finally, we compute $\bm{m}^l=\bm{f}^{-1}(\bm{h}_r^l)=M_{an}(h_r^l)\frac{\bm{h}_r^l}{h_r^l}$ and $\bm{m}=\sum_{l=1}^n\omega^l\bm{m}^l$.
As an example, we simulated several hysteresis loops (Fig. \ref{Fig4}) for a material characterized by the anhysteretic function (\ref{Man}) with $m_s=1.23\cdot 10^6$ A/m, $A=50$ A/m, and represented by $N=20$ pseudoparticle types with $k^l=140(l-1)/(N-1)$ A/m, each having the same volume fraction $\omega^l=1/N$.
\begin{figure}[h!]
\begin{center}\includegraphics[width=7cm,height=6cm]{multi_k_example.pdf}
\caption{Simulation using the composite model.}\label{Fig4}
\end{center}
\end{figure}
As was noted in \cite{B97}, it may be better to assume the magnetization of a pseudoparticle does not evolve independently but is influenced by the other particles. Hence, as the second essential modification of his model, Bergqvist replaced the ``driving force" of this evolution, $\bm{h}(t)$, by the ``effective" field $\bm{h}(t)+\alpha \bm{m}(t)$, where $\alpha$ is a material-dependent parameter. Such effective fields are often employed also in other models of hysteresis (see, e.g., \cite{Mgoyz,JA,LPA}); in \cite{Torre} Della Torre presented an explanation of the interaction term $\alpha \bm{m}(t)$ (see Ch. 4).
With this modification the convex time-dependent sets $K^l(t)$ should be replaced by
$$\begin{array}{c}K^l(t,\bm{m}):=\{\bm{u} \in {\mathbb R}^3\ :\\ |(k^l)^{-1}(\bm{u}-\bm{h}(t)-\alpha\bm{m})|\leq 1\}\end{array}$$
if $k^l\neq 0$ and $K^l(t,\bm{m}):=\left\{\bm{h}(t)+\alpha\bm{m}\right\}$ otherwise. The internal variables $\bm{h}_r^l$ are now solutions of the optimization problems
\begin{equation} \min_{\bm{u}\in K^l(t,\bm{m})}\{S(\bm{u})-\check{\bm{m}}^l\cdot\bm{u}\},\label{optjm}\end{equation}
in which the constraints depend on the unknown solution itself, since $$\bm{m}=\sum_{l=1}^N
\omega^l\bm{m}^l=\sum_{l=1}^N
\omega^lM_{an}(h_r^l)\frac{\bm{h}_r^l}{h_r^l}.$$
The implicit constraints in (\ref{optjm}) complicate the determination of the magnetizations $\bm{m}^l$ (such problems are equivalent to quasivariational inequalities). Nevertheless, an efficient iterative method can be proposed (see Section V).
Further modification of the model is needed to account for the known phenomenon of zero hysteresis loss in a rotational field at saturation \cite{Torre,Appino}. To describe this ``saturation property", it was suggested \cite{B97,B14,LZC15} to replace the constant intrinsic coercivity of each pseudoparticle, $k^l$, by a decreasing function $k^l(h_r^l)$ attaining zero at a saturation value $h_r^l=h_s$. In \cite{B14,LZC15}, however, only a ``monoparticle" model with the approximate update rule (\ref{Appr1}) has been considered. We suppose that in a composite model the lossless coherent rotation of all magnetization vectors $\bm{m}^l$ is, probably, better described if, for all $l$, $k^l=k^l(m)$ which simultaneously drop to zero as the total magnetization $m$ reaches saturation. In this work, however, we consider for simplicity only constant $k^l$ values.
\section{Identification of the parameters in the model \label{Sec4}}
The practical implementation of the phenomenological model described above needs identification of the anhysteretic curve $M_{an}$, the material parameter $\alpha$, the coefficients $k^l$ (scalar or matrices, constant or dependent on $m$) and the corresponding weights $\omega^l$, $l=1,...,N$. Of course, the identification of parameters is needed also for other models of hysteresis and a variety of approaches have been proposed; the identification procedure depends on the model and the experimental data available.
Here we present a consistent algorithm for the identification of the parameters in the model for a nonoriented steel using the experimental major hysteresis loop and eleven first order reversal curves (FORCs) (the 1.8\% Si steel N3 in \cite{ZVis}, Fig. \ref{FigData}; see also \cite{ZInv}). Because of the symmetry, only the ascending part of the major loop was used; this curve was regarded as an additional FORC that starts at a strong negative magnetic field -800 A/m. We assume that the material is isotropic (in the anisotropic case similar measurements along the main material axes should be used for the identification) and, for simplicity, seek constant scalar values for $k^l$, $l=1,...,N$. Postulating that $N=41$ possible $k^l$ values are uniformly distributed from zero to 800 A/m with the step 20 A/m, we need to find only the appropriate weights $\omega^l$ to specify the distribution of the pseudoparticles.
No specific formulae was assumed for the anhysteretic function $M_{an}(|\bm{u}|)$: this function was approximated by a cubic spline with the ``not-a-knot'' end conditions as follows. First, we determined the interval in which this function needs to be defined. Although the measured magnetic field did not exceed 800 A/m, the reversible fields $\bm{h}_r^l$ driven by the effective field $\bm{h}_{eff}=\bm{h}+\alpha \bm{m}$ can be out of this range. After several tests needed to estimate, at least crudely at first, the value of $\alpha$ (see below), we chose to approximate this function in the interval [0, 1750] A/m. The anhysteretic function was then sought as a cubic spline with a fixed set of equidistant knots: $250i$ A/m, $i=0,1,...,7$. Since $M_{an}(0)=0$ is fixed, the spline is determined by its \textit{a priori} unknown values $M_i$ at the seven other knots. Note that in this experiment the fields can be regarded as scalar. In such a case let $h$ denote the field itself (which may be negative) and not its magnitude $|h|$ as above.
Let the anhysteretic function $M_{an}$, the particle fractions $\omega^l$, and the parameter $\alpha$ be given. At any point $\{h_0,m_0\}$ on the descending branch of the major hysteresis loop, we know the internal state $h_r^l$ of all pseudoparticle types: $h_r^l(h_0)=h_0+\alpha m_0+k^l$. If at this point the magnetic field starts to grow, the particle state evolution along the FORC $\{h,m(h)\}$ thus created can be described as $h_r^l(h)=\max\{h_r^l(h_0),h+\alpha m(h) -k^l\}$. The model prediction for the total magnetization along this curve, ${\cal M}(h)=\sum_{l=1}^N \omega^lM_{an}(h_r^l(h))$, can be compared to the measured $m^*(h)$ values for the same FORC. Our identification procedure finds the parameters $M_i$ of $M_{an}$, the weights $\omega^l$ and $\alpha$ minimizing the least-squares method residual, the sum $L$ of $[{\cal M}(h)-m^*(h)]^2$ over all measured points on all experimental FORCs. Efficient standard functions of Matlab were used in our three-level optimization algorithm as follows.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth,height=7.3cm]{Z_data.pdf}
\caption{Hysteresis curves, experimental.}\label{FigData}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=\linewidth,height=7.3cm]{Z_data_appr.pdf}
\caption{The least-squares fit of the FORCs.}\label{FigAppr}
\end{figure}
\begin{figure}[!t]\centering
\includegraphics[width=7.5cm,height=6cm]{Man.pdf}
\caption{Anhysteretic function $M_{an}$.}\label{FigMan}
\end{figure}
\begin{figure}[!h]\centering
\includegraphics[width=7.5cm,height=6cm]{Z_weights.pdf}
\caption{Weights $\omega(k)$; identification result.}\label{FigW}
\end{figure}
On the lowest level, the program determines the weights $\omega^l$ satisfying $\omega^l\geq 0,\ \sum_{l=1}^N\omega^l=1$ and minimizing the least-squares residual $L$ for given values of $M_i$ and $\alpha$. This is a quadratic programming problem solved by the \verb"lsqlin" function from the Matlab Optimization toolbox 7.2. On the next level, the Matlab function \verb"fminunc" realizes the unconstrained minimization over the set of $M_i$ for a given value of $\alpha$. Plotting the resulting function $L=L(\alpha)$ helped us to crudely estimate the position of the minimum and, on the upper level, this information was used by the Matlab function \verb"fminbnd" to find the optimal value of $\alpha$; the corresponding $M_i,\ \omega^l$ were thereby also obtained. Minimization of this function on a standard PC takes about 10 minutes.
Although we did not take into account any possible dependence of $k^l$ on $\bm{m}$, a very good fit of the experimental curves has been obtained for $\alpha=8.8\cdot 10^{-4}$ (Fig. \ref{FigAppr}); the anhysteretic function $M_{an}$ and the weights $\omega^l$ are shown in Figs. \ref{FigMan} and \ref{FigW}, respectively.
We note that for only 17 of the 41 predefined values of $k^l$ did the identified volume fractions $\omega^l$ exceed $10^{-3}$. Only these pseudoparticles are taken into account in our finite element computations below; their total fraction is 0.9991. The results obtained show that for this material the reversible magnetization dominates (see Fig. \ref{FigW}) and there is a very significant reciprocal influence of pseudoparticles via the $\alpha \bm{m}$ term in $\bm{h}_{eff}$ (see Figs. \ref{FigData} and \ref{FigMan}): without this term the $M_{an}(h)$ curve would be a much steeper line between the ascending and descending branches of the main hysteresis loop.
\section{A numerical scheme for a 2d hysteresis and eddy current problem}
Let us consider a long ferromagnetic cylinder, parallel to the $z$-axis and having a cross-section $\Omega$, carrying a transport current $I(t)$ and placed into a perpendicular uniform external field $\bm{h}_e(t)$. The electric field $\bm{e}(x,y,t)$ and the current density $\bm{j}(x,y,t)$ are parallel to the $z$-axis; we can also choose the vector magnetic potential $\bm{a}(x,y,t)$ parallel to the $z$-axis (so that $\nabla\cdot\bm{a}=0$). Hence, these variables can be regarded as scalar and we will use the scalar notation $e,\ j$ and $a$ (which should not be confused with the absolute values of these vectors).
The vector fields $\bm{h}(x,y,t),\bm{b}(x,y,t)$ and $\bm{m}(x,y,t)$ are parallel to the $xy$ plane.
We will use a 2d eddy current and magnetization problem formulation, similar to that proposed for 3d problems with hysteresis in \cite{dARTV}, but employ the Bergqvist model for magnetization discussed above. This model should be incorporated as a local constitutive relation
\begin{equation} \bm{m}=\bm{M}[\bm{h}_{eff}],\label{constM}\end{equation}
where $\bm{h}_{eff}=\bm{h}+\alpha\bm{m}$ and $\bm{h}=\frac{1}{\mu_0}\bm{b}-\bm{m}=\frac{1}{\mu_0}{\bm{\nabla}}\times a-\bm{m}$. Here (and below) square brackets are used for the operator argument and ${\bm{\nabla}}\times a=(\partial_ya,-\partial_xa)$. The value $\bm{M}[\bm{h}_{eff}]$ at a point depends on the history of $\bm{h}_{eff}$ at this point: implicitly, the local operator $\bm{M}$ keeps track of the state of the internal variables $\bm{h}^l_r$ and $\bm{m}^l$.
The density of the power loss due to magnetization is $W_H=\mu_0\sum_{l=1}^N\omega^l|k^l\dot{\bm{m}}^l|$. Ohm's law $e= \rho j$, where $\rho$ is the resistivity, is another constitutive relation characterizing the material. The density of the eddy and transport current loss is $W_E=\rho j^2$.
We note that the so-called anomalous or excess loss, attributed to microscopic eddy currents caused by small, almost instantaneous jumps of domain walls (the Barkhausen noise), is not described by this model. As is shown in \cite{Zloss}, for the nonoriented steel considered in Section \ref{Sec4} this loss is small, at least for the magnetization in the rolling direction; for other magnetization directions the excess loss can, possibly, reach 10-20\% of the total loss, see \cite{Zloss1}.
For the specified geometry and gauge the electric field can be written as
$e=-\partial_t a+c(t),$ where the time-dependent constant $c(t)$ results from the parallel to the $z$-axis gradient of the scalar potential, $\nabla\Phi$ (see, e.g., \cite{P97}). This yields that
\begin{equation} \rho j=-\partial_t a+c(t),\end{equation}
where the unknown constant $c(t)$ is determined implicitly by the given transport current,
\begin{equation}\int_{\Omega}j\,d\bm{r} =I(t).\end{equation}
In our numerical simulations we will assume that $I(t)=0$.
The vector potential can be represented as a sum, $a=a_e+a_j+a_m$, of the potentials associated with the external field, current density, and magnetization, respectively. Here $a_e=\mu_0(yh_{e,x}-xh_{e,y})$ and, see \cite{Jackson},
\begin{eqnarray} a_j[j]=\mu_0\int_{\Omega}G(\bm{r}-\bm{r}')j(\bm{r}',t)\,d\bm{r}',\label{aj}\\
a_m[\bm{m}]=\mu_0\int_{\Omega}\nabla G(\bm{r}-\bm{r}')\times\bm{m}(\bm{r}',t)\,d\bm{r}',\label{am}\end{eqnarray}
where $\nabla$ is the gradient with respect to $\bm{r}=(x,y)$
and $G(\bm{r})=\frac{1}{2\pi}\ln\frac{1}{|\bm{r}|}$ is the Green's function.
Correspondingly, the magnetic field can be represented as the sum:
$\bm{h}=\bm{h}_e+\bm{h}_j+\bm{h}_m,$ where \begin{align*} \bm{h}_j[j]=&\int_{\Omega}\nabla G(\bm{r}-\bm{r}')\times j(\bm{r}',t)\,d\bm{r}'=\\&\left( \int_{\Omega}j(\bm{r}',t)\partial_{y}G(\bm{r}-\bm{r}')\,d\bm{r}',\right.\\& \left. -\int_{\Omega}j(\bm{r}',t)\partial_{x}G(\bm{r}-\bm{r}')\,d\bm{r}' \right)\end{align*}
and, see \cite{bofm},
\begin{align} \bm{h}_m[\bm{m}]=&\frac{1}{\mu_0}\bm{b}_m-\bm{m}=\nonumber\\ &\nabla\int_{\Omega}\nabla\cdot\left[G(\bm{r}-\bm{r}')\bm{m}(\bm{r}',t)\right]\,d\bm{r}'.
\label{bm}\end{align}
The main unknowns in this model are $\bm{m}(\bm{r},t)$ and $j(\bm{r},t)$: provided these variables are found, $\bm{h}$ and $\bm{b}$ can be also calculated; $c(t)$ is an auxiliary unknown. After discretization in time, the problem to be solved on each time level $n$ consists of three linear
equations,
\begin{align}\tau\rho j^{n}&+a_j[j^{n}]+a_m[\bm{m}^{n}]-\tau{c}^{n}=\nonumber\\&a_e^{n-1}-a_e^{n}+a_j[j^{n-1}]+
a_m[\bm{m}^{n-1}],\label{it1}\\
&\int_{\Omega}j^{n}\,d\bm{r}=I^n,\label{it2}\\
&\bm{h}^{n}_{eff}=\bm{h}_e^n+\bm{h}_j[j^{n}]+\bm{h}_m[\bm{m}^{n}]+\alpha\bm{m}^{n},\label{ith}
\end{align}
where $\tau$ is the time step, supplemented by the nonlinear relationship,
\begin{equation}\bm{m}^{n}=\bm{M}[\bm{h}^{n}_{eff}]\label{m_up}.\end{equation}
Our iterative scheme was based on the representation
\begin{equation} \bm{m}^{n,k+1}=\bm{M}[\bm{h}_{eff}^{n,k}]+D[\bm{h}_{eff}^{n,k}]\left(\bm{h}_{eff}^{n,k+1}-\bm{h}_{eff}^{n,k}\right),\label{it3}\end{equation}
where $k$ is the iteration number and
$D[\bm{u}]$ is, at each point of $\Omega$, the $2\times 2$ matrix of partial derivatives,
$$D[\bm{u}]=\left(\begin{array}{cc}
\partial_{u_x} M_x[\bm{u}]&\partial_{u_y} M_x[\bm{u}]\\
\partial_{u_x} M_y[\bm{u}]&\partial_{u_y} M_y[\bm{u}]
\end{array}\right).$$
Since the analytical calculation of these derivatives is difficult, we used their numerical approximations:
replacing $u_x$ by $u_x\pm \Delta$ and keeping $u_y$ unchanged, we used the central difference to estimate $\partial_{u_x} \bm{M}[\bm{u}]$; similarly for $\partial_{u_y} \bm{M}[\bm{u}]$.
Substituting (\ref{it3}) into (\ref{it1})--(\ref{ith}) yields a linear system for $\bm{h}_{eff}^{n,k+1},\ j^{n,k+1},\ {c}^{n,k+1}$. These iterations should be repeated until convergence of $\bm{h}_{eff}^{n,k}$ with a given tolerance; the magnetization and magnetic field in the magnetic material are then found as
$\bm{m}^n=\bm{M}[\bm{h}_{eff}^n]$ and
$\bm{h}^{n}=\bm{h}^n_{eff}-\alpha\bm{m}^n$, respectively.
The described iterative procedure has been applied to the finite element approximation of this problem.
We triangulated the domain $\Omega$ and used piecewise constant approximations for all variables, $j,\ a,\ \bm{h}$ and $\bm{m}$.
The finite element
approximation of the integral operator equation (\ref{it1}) involves the computation of matrices with the entries, see (\ref{aj}), (\ref{am}),
\begin{eqnarray} L_{e,e'}=\int_e\int_{e'}G(\bm{r}-\bm{r}')d\bm{r}'\,d\bm{r},\label{M_int}\\ L^x_{e,e'}=\int_e\int_{e'}
\partial_xG(\bm{r}-\bm{r}')d\bm{r}'\,d\bm{r},\\ L^y_{e,e'}=\int_e\int_{e'}\partial_yG(\bm{r}-\bm{r}')d\bm{r}'\,d\bm{r}.\label{Mx_int}\end{eqnarray}
for each pair of triangles $e,\,e'$.
Only the matrices $L^x,\ L^y$ defined above are needed to find $\bm{h}_j[j]$ in (\ref{ith}). However, to compute
the $\bm{h}_m[\bm{m}]$ part of $\bm{h}_{eff}$ one needs, see (\ref{bm}), matrices with entries
\begin{eqnarray} L^{xx}_{e,e'}=\int_e\int_{e'}\partial^2_{xx}G(\bm{r}-\bm{r}')d\bm{r}'\,d\bm{r},\label{Mxx_int}\\
L^{yy}_{e,e'}=\int_e\int_{e'}\partial^2_{yy}G(\bm{r}-\bm{r}')d\bm{r}'\,d\bm{r},\\
L^{xy}_{e,e'}=L^{yx}_{e,e'}=\int_e\int_{e'}\partial^2_{xy}G(\bm{r}-\bm{r}')d\bm{r}'\,d\bm{r}.
\label{Myy_int}\end{eqnarray}
Some of the integrals in (\ref{M_int})--(\ref{Myy_int}) are singular; their computation is described in Appendix B. We note that the matrices $L,\ L^{xx},\ L^{xy}$ and $L^{yy}$
are symmetric, whereas the $L^x,\ L^y$ are antisymmetric.
To approximate the matrix $D$ in each element we used $\Delta =2\cdot 10^{-6}\max(1,h)$. Convergence of the iterations was stable and fast: three-five iterations per time level ensured convergence with the relative tolerance $10^{-6}$ (in the $L^1$ norm) in all our simulations.
\section{Simulation results}
In all numerical simulations below we assumed the material is isotropic and used the hysteresis model parameters identified for a nonoriented steel in Section \ref{Sec4}.
As our first example, we considered a cylinder with a circular cross-section $\Omega=\{\bm{r}\ :\ r<r_0\}$. We made the eddy current negligible by choosing a very high resistivity $\rho$. Then, for a material with constant magnetic permeability $\mu$, the magnetic field can be expressed via the scalar magnetic potential, $\bm{h}=-\nabla \psi$, satisfying the Laplace equation
$\frac{1}{r}\frac{\partial}{\partial r}\left(r\frac{\partial\psi}{\partial r}\right)+\frac{1}{r^2}\frac{\partial^2\psi}{\partial \theta^2}=0$
inside and outside the domain $\Omega$ with the interface conditions $$\psi|_{r_0-}=\psi|_{r_0+},\ \ \ \left.\mu\frac{\partial\psi}{\partial r}\right|_{r_0-}=\left.\mu_0\frac{\partial\psi}{\partial r}\right|_{r_0+}$$ and the boundary condition $-\nabla \psi\rightarrow\bm{h}_e$ as $r\rightarrow \infty$. Solving this problem by separation of variables, one finds that inside the cylinder the fields $\bm{h}$ and $\bm{m}$ are uniform:
\begin{equation}\bm{h}=\frac{2\bm{h}_e}{\mu_r +1},\ \ \bm{m}=\frac{2\bm{h}_e(\mu_r-1)}{\mu_r +1},\label{fields}\end{equation}
where $\mu_r=\mu/\mu_0$.
For a circular ferromagnetic cylinder we now assume the virgin initial state ($\bm{m}^l|_{t=0}=0$ for all $l$) and a uniform external field $\bm{h}_e(t)$ growing from zero monotonically in a fixed direction. In this case $\bm{h}$ and $\bm{m}$ are also uniform in $\Omega$ but obey (\ref{fields}) with an unknown relative permeability $\mu_r$ varying with $h_e$. For this unidirectional situation the model (\ref{constM}) employed predicts
\begin{equation} m=\sum_{l=1}^N\omega^lM_{an}([h+\alpha m-k^l]_+),\label{nl}\end{equation}
where $u_+=\max\{u,0\}$. Substituting relations (\ref{fields}) into (\ref{nl}) we arrive at a nonlinear algebraic equation for $\mu_r$, which is easy to solve numerically; this determines $h$ and $m$ for any given $h_e$.
This solution does not depend on $r_0$ and was used as a partial test for our finite element simulations.
We set $\bm{h}_{e}=(10^3t,\,0)$ A/m and used two finite element meshes (742 and 2436 triangles) to compute the solution at $t=100$ s. For the crude mesh the relative errors of our finite element solution (in the $L^1$ norm) did not exceed $\delta_h=0.08\%$ and $\delta_m=0.21\%$ for $\bm{h}$ and $\bm{m}$, respectively; for the finer mesh we obtained $\delta_h=0.05\%$ and $\delta_m=0.18\%$. The moderate accuracy gain for $\bm{m}$ for the finer mesh probably indicates that a non-negligible part of the error is induced by the use of numerical integration in (\ref{M_int})--(\ref{Myy_int}). We found that in this example the accuracy was practically independent of the constant time step $\tau$; such a phenomenon is often observed in modeling ``simple regimes" of rate-independent models (see, e.g., \cite{BP14}, p. 1010) and such is the employed model without the eddy current.
The average number of iterations per time level did not depend on the mesh and increased from 3.7 for $\tau=5$ s to 4.5 for $\tau=50$ s.
Our second example is a hollow ferromagnetic cylinder with the cross-section $r_1\leq r\leq r_2$; such a configuration can be employed for magnetic shielding. If the eddy current can be neglected and the magnetic permeability of the material is constant, the problem can be solved analytically \cite{Bess}. In this case the magnetic field inside the hole is uniform:
$$\bm{h}=\frac{4\mu_r\bm{h}_e}{(\mu_r+1)^2-(r_1/r_2)^2(\mu_r-1)^2}.$$
Assuming the steel resistivity $\rho=0.43\, \mu\Omega\cdot$m (see \cite{ZVis}) and taking $r_1=0.1$ m, $r_2=0.15$ m we solved the magnetization problem taking both the eddy current and hysteresis into account. We now triangulated a larger than the cross-section domain, the square $-0.2\leq x,y\leq 0.2$ m, see Fig. \ref{Fig_mesh}. The mesh contains 6424 triangles; 2352 of them belong to the ferromagnetic domain.
\begin{figure}[!h]
\centering
\includegraphics[width=8.5cm]{mesh.pdf}
\caption{Finite element mesh.}\label{Fig_mesh}
\end{figure}
Although our numerical algorithm needs only the latter elements, to find the magnetic induction also outside the ferromagnet we computed the elements of matrices (\ref{M_int})--(\ref{Myy_int}) for all triangle pairs $e,\,e'$, where at least one of the triangles belongs to the magnetic domain (the remaining elements of these matrices can be set to zero). We set $\bm{h}_e=(10^3t,0)$ A/m for the first 100 s and assumed that in the next 100 s the external field rotates $90^{\circ}$ counter clockwise with a constant angular velocity, its magnitude remaining at $10^5$ A/m. The time steps $\tau=10$ s and $\tau=2.5$ s were used, respectively, for these two time intervals; on average, convergence with the relative tolerance $10^{-6}$ was achieved in five iterations per time level. Numerical simulation results for $t = 100$ s and $t = 200$ s are shown in Fig. \ref{Fig_ring}.
\begin{figure*}[!t]
\centering
\includegraphics[width=7.6cm]{rng_b100.jpg}\includegraphics[width=7.6cm]{rng_b200.jpg}\\
\includegraphics[width=16.8cm]{rng_m.jpg}\\
\includegraphics[width=16.8cm]{rng_j.jpg}
\caption{Numerical simulation. Left: solution for $t=100$ s, right: for $t=200$ s. Top: magnetic induction; middle: magnetization $|\bm{m}|$; bottom: eddy current density. }\label{Fig_ring}
\end{figure*}
\clearpage
\noindent A further decrease of the time steps does not change these results.
The eddy current density is not negligible now but,
according to our computation, for the assumed external field variation rate the power of the eddy current loss per unit of length,
$p_j=\int_{\Omega}\rho j^2 d\bm{r}$, is smaller than the power of the magnetization
dissipation $p_m=\mu_0\int_{\Omega}\left\{\sum_l k^l|\dot{\bm{m}}^l|\right\} d\bm{r}$. Thus, in the end of the linear growth of $h_e$ ($t=100$ s) we obtained $p_j=2.4\cdot 10^{-3}$ W/m and $p_m=15\cdot 10^{-3}$ W/m; at $t=200$ s (rotating $\bm{h}_e$) we obtained $p_j=5.0\cdot 10^{-3}$ W/m and $p_m=20\cdot 10^{-3}$ W/m. The corresponding losses during the whole time interval $0<t<200$ s are 0.70 J/m and 2.3 J/m, respectively. It may be noted that not only the eddy current density, but also the related part of the total loss should increase with the ramping rate of the external field. We found also that the magnetic field inside the cylindrical hole does not exceed 0.14\% of $h_e$ (the shielding effect). Further simulations showed that the pronounced non-monotonicity of the magnetization $\bm{m}(\bm{r},t)$ in the radial directions (see Fig. \ref{Fig_ring}, middle) is caused by the eddy current. Suppressing this current by choosing a high resistivity value makes the solution monotonic radially; qualitatively, for $t\leq 100$ s the solution without the eddy current resembles and corresponds to a similar shielding as the analytical solution for an appropriate constant permeability, $\mu_r\approx 7\cdot 10^3$. However, since the magnetic field inside the magnet is not uniform, in the hysteretic (and even simply in the nonlinear) case obtaining an analytical solution (as we did in the previous example) is not possible. The numerical simulation showed that also the magnetic field inside the hole is not uniform. Finally, we note that, during the rotation of the external field, the magnetization of the inner ring layers lags behind (Fig. \ref{Fig_ring}, middle-right), which is the hysteresis effect.
It should be noted that the magnetization model employed here is complicated and cannot compete with much simpler models in terms of the computation time. Even after the matrices (\ref{M_int})--(\ref{Myy_int}) were computed, our Matlab program still needed several hours to solve examples like these on a standard PC. The main elements, determining the efficiency of our computations are: the integral vector potential formulation, the inner iterations for updating the magnetization of pseudoparticles in each finite element, and the
outer iterations for solving (\ref{it1})--(\ref{m_up}) using (\ref{it3}).
Integral formulations like this (we followed \cite{dARTV}) are often employed for solving electrical engineering problems in general, and eddy current problems in particular. Their main advantage is that all computations are confined to the area occupied by the conducting and/or magnetic materials; but a disadvantage is the linear systems with dense matrices. Overall, such formulations are competitive with other formulations used for finite element approximations.
The inner iterations converged in 2--3 iterations, much faster than in \cite{FLavHenr2013} and, as was shown above (see Fig. 3), the faster approach based on the explicit approximation (\ref{Appr1}) can be inexact.
As is well known, for nonlinear materials with high differential susceptibility
values it is difficult to obtain good convergence of iterations in finite element
simulations, especially, if the Newton method cannot be employed. This is the case for the model employed here: in the Newton-like method that we used the necessary derivatives could only be numerically approximated and, for steel in our simulations, the maximal susceptibility exceeded $10^5$. Nevertheless, we were able to reach convergence in five iterations per time level in the outer cycle of our scheme, which is a fast convergence.
\section{Conclusions}
Like most existing macroscopic models for ferromagnetic hysteresis, the quasi-static model proposed by Bergqvist in \cite{B97} is phenomenological. However, it is based upon consistent energy arguments and a clear albeit simplified physical picture of the dry-friction like pinning of the domain walls. This model is naturally vectorial, has a variational formulation convenient for numerical simulations, and can be incorporated into a finite element code as a local constitutive relation with memory. As an example we considered a problem, where both the magnetization and eddy current were taken into account.
In this work we tried to clarify the mathematical derivation of the variational formulation, extended it to the anisotropic case, proposed an efficient numerical method based on this formulation, and also demonstrated that the usually employed approximation, which turns the model into a play hysteron model, can be inaccurate. We showed that this approximation is not a possible version of the dry friction law, as is typically assumed, but a replacement by an alternative assumption, not related to dry friction, and determining a different direction of the system's evolution in the vectorial case.
The model has sufficient degrees of freedom to be fitted to hysteretic behavior of different materials; here we presented a method for the identification of the parameters in this model using a set of experimental FORCs. Another advantage of this model is its ability to predict both the stored and dissipated energies at any moment in time. These properties make the model highly attractive; its further comparison to experiments would be desirable.
|
\section{Introduction}
Various field theories suggest the existence of stringlike configurations,
which are the particle physics analogs of vortices or magnetic flux tubes in
condensed matter physics. These configurations can arise at scales ranging from
the fundamental distances in string theory to astrophysical distances, where in
the latter case they are often called {\it cosmic strings}. (See for example
the reviews~\cite{Copeland:2011dx,Hindmarsh:2011qj}.)\footnote{Arguments for
a closer connection between cosmic and fundamental
strings are given in Ref. \cite{Copeland:2009ga}.} A well--known
representative is the Nielsen--Olesen vortex~\cite{Nielsen:1973cs} in a model
with an Abelian gauge field coupled to a single Higgs field. This vortex
is classically stable, as are particular embeddings in non--Abelian models
with several Higgs fields~\cite{Hindmarsh:2016lhy}. In general, however,
non--Abelian string configurations are not classically stable. In this
context the $Z$--string, which typically involves the $Z$--boson field
in the standard model, is of particular interest~\cite{Achucarro:1999it}.
Though not classically stable, it is possible that
these strings are stabilized by quantum effects. The vacuum polarization
energy (\emph{VPE}), which is the regularized and renormalized sum of
all zero point energies of the quantum fluctuations in the classical
background, is central to these investigations. In field theory
quantum effects are typically estimated by Feynman diagram
techniques. However, stringlike configurations have a nontrivial
structure at spatial infinity which makes the formulation of a
Feynman perturbation expansion impossible without any further
adaptation. Even then, the convergence of the series is not guaranteed
as the relevant couplings are not necessarily small and the series is
only asymptotic. On top of that, the rich topological
structures~\cite{Kibble:2015twa} of theories with cosmic strings
require techniques beyond perturbative treatments. Not surprisingly,
the study of the \emph{VPE} of cosmic string configurations has a long
history of slow progress without a fully concluding answer.
Numerous publications have analyzed quantum fluctuations about cosmic
strings. Naculich~\cite{Naculich:1995cb} has discussed that in the
limit of weak coupling, fermion fluctuations tend to destabilize the string.
The quantum properties of $Z$--strings have also been connected to
nonperturbative anomalies~\cite{Klinkhamer:2003hz}. Furthermore, the
emergence or absence of exact neutrino zero modes in a $Z$--string background
and the possible consequences for the string topology were investigated
in Ref.~\cite{Stojkovic}. A first attempt at a full calculation of the
fermion quantum corrections to the $Z$--string energy was carried out
in Ref.~\cite{Groves:1999ks}. Those authors were only able to compare the
energies of two string configurations, rather than comparing a single
string to the vacuum, because of limitations arising from the nontrivial
behavior at spatial infinity. (We will discuss this issue in more
detail below.) The fermion vacuum polarization
energy of the Abelian Nielsen--Olesen vortex~\cite{Nielsen:1973cs} has
been estimated in Ref.~\cite{Bordag:2003at} with regularization limited
to the subtraction of the divergences in the heat--kernel expansion.
On the other hand, quantum energies of bosonic fluctuations in
string backgrounds were calculated in Ref.~\cite{Baacke:2008sq}.
However, these are suppressed compared to fermion fluctuations
when the number of internal degrees of freedom, {\it e.g.} color, is
large.
Using the spectral method~\cite{Graham:2009zz} the (one--loop) \emph{VPE} can be
computed from scattering data. An essential feature of this method is the
identification of elements from the Born expansion with Feynman diagrams.
These elements are added and subtracted to make contact with standard
renormalization techniques and conditions which prescribe certain Green
functions for particular values of transferred momenta. In a sequence of projects
we succeeded computing the fermion \emph{VPE} of cosmic strings after solving a number
of problems:\footnote{See Ref.~\cite{Weigel:2015lva} for a recent review.}
\begin{enumerate}
\item The string configuration does not have a well defined Born series to be
identified with the Feynman series of quantum field theory;
this can be overcome by a special local gauge transformation~\cite{Weigel:2010pf}.
\item A correction factor to the na{\"\i}ve Jost function is required to maintain
the analytic properties of scattering data~\cite{Weigel:2009wi,Graham:2011fw},
because the effective fermion mass depends on the distance from the string core.
\item Higher order Feynman diagrams are required which become exceedingly
difficult to evaluate numerically; this is solved by the
so-called fake boson approach~\cite{Farhi:2001kh}.
\end{enumerate}
\noindent
Formally the unregularized and unrenormalized fermion \emph{VPE} of the string
is the sum of energy eigenvalues from a Dirac Hamiltonian. These eigenvalues
are invariant for a particular path in the space of parameters which define the
(weak) isospin orientation of the string~\cite{Klinkhamer:1997hw}. Previous
calculations of the \emph{VPE}~\cite{Weigel:2010pf,Weigel:2009wi,Graham:2011fw} were
restricted to a simplifying submanifold in isospace that could not access this path.
The invariance of the single particle energies is, however, not sufficient to ensure
that the full fermion \emph{VPE} is also invariant in this calculation. The sum of the energy
eigenvalues is ultraviolet divergent and in the inevitable process of
regularization and renormalization divergent contributions emerge that are
manifestly variant. They are conjectured to cancel based on their formal equivalence
as expansions in powers of the string background. On the regularization side
terms from the Born expansion to scattering data are subtracted, which on the
renormalization side are added back in the form of Feynman diagrams.
An exact match of these quantities is not at all obvious.
For instance, Feynman diagrams allow us to distinguish between the divergences that
emerge from the quantum loops and the Fourier modes of the background
(this is e.g.~essential for understanding the Casimir effect~\cite{Graham:2003ib} in the
context of spectral methods). On the other hand scattering data, and thus the Born expansion
terms, do not distinguish between external and loop momenta.
Using dimensional regularization, the equivalence of the two schemes has been
verified for the leading (tadpole) divergence, both for boson~\cite{Farhi:2000ws} and
fermion\cite{Farhi:2000gz} fluctuations. At higher order the distinction between
loop and Fourier momenta is essential and so far no such proof has been provided.
The scattering data decouple into angular momentum channels. As we will explain in
Sec.~\ref{sec:vpe}, a channel by channel subtraction is mandatory for contributions
that can be related to the \emph{quadratic} ultraviolet divergences in the Feynman series.
The subleading logarithmic divergences require to include higher order Born/Feynman terms,
which are very cumbersome to simulate numerically. Fortunately, the set of divergences
terminates at this logarithmic level so that these divergences can be cavalierly treated
by simulating them in a simpler (typically bosonic) theory. This method brings into the
game an additional contribution that is not manifestly invariant under the particular
isospin transformation mentioned above. Furthermore, the simulation of divergences by
a boson model also requires the exchange of momentum integrals with orbital angular
momentum sums. which by itself demands care: for instance, swapping these operations
for momenta on the real axis gives erroneous results~\cite{Pasipoularides:2000gg}; instead,
an analytic continuation to imaginary momenta is required~\cite{Schroder:2007xk}.
In any event, the whole regularization procedure is not manifestly gauge invariant while
gauge invariance should, of course, be maintained by the final result in order for the
adopted calculational procedure to produce unambiguous results. A good example to
demonstrate the subtleties of gauge invariance in the context of the spectral approach
are the \emph{vacuum charges} induced by a nontrivial background configuration:
improper regularization may falsely predict anomalous vacuum charges~\cite{Farhi:2000gz}.
From these considerations, it is clear that consistency checks are indispensable to ensure
that the spectral method does not artificially break (gauge) symmetries leading to
erroneous results. In the present paper, we will explore such a test based on a global
isospin invariance. Because of the operation under item 1) above, this also probes
a local invariance.
\bigskip\noindent
We conclude this introduction with a brief description of our model.
The bosonic part is described by the Lagrangian
\begin{equation}
\mathcal{L}_{\phi,W}=-\frac{1}{2} \mbox{tr}
\left(G^{\mu\nu}G_{\mu\nu}\right) +
\frac{1}{2} \mbox{tr} \left(D^{\mu}\Phi \right)^{\dag} D_{\mu}\Phi
- \frac{\lambda}{2} \mbox{tr} \left(\Phi^{\dag} \Phi - v^2 \right)^2 \,,
\label{Lbosonic}
\end{equation}
where the Higgs doublet is written using the matrix representation
\begin{equation}
\Phi=\begin{pmatrix}
\phi_0^* & \phi_+ \cr -\phi_+^* & \phi_0 \end{pmatrix} \,.
\label{higgsmatrix}
\end{equation}
The gauge coupling constant $g$ enters through both the covariant derivative
$D_\mu = \partial_\mu - i \,g W_\mu$ and the $SU(2)$ field strength tensor
\begin{equation}
G_{\mu\nu} = \partial_\mu\,W_\nu - \partial_\nu W_\mu - i \,g\,[ \,W_\mu\,,\,
W_\nu\,]\,.
\label{fieldtensor}
\end{equation}
The classical potential has been chosen such that the Higgs field
acquires a vacuum expectation value (\emph{VEV}) $v$, where
$\langle {\rm det}(\Phi)\rangle = v^2 \neq 0$. As a consequence,
all bosons become massive: $m_W = g v / \sqrt{2}$ and $m_H = 2 v \,\sqrt{\lambda}$.
The interaction of the (classical) string with the left--handed fermions is described
by
\begin{equation}
\mathcal{L}_\Psi=i\overline{\Psi}
\left(P_L D \hskip -0.6em / + P_R \partial \hskip -0.6em / \right) \Psi
-f\,\overline{\Psi}\left(\Phi P_R+\Phi^\dagger P_L\right)\Psi\,.
\label{gaugelag}
\end{equation}
Here, $P_{R,L}=\frac{1}{2}\left(1\pm\gamma_5\right)$ are projection
operators on left/right--handed components, respectively, and the strength
of the Higgs-fermion interaction is parametrized by the Yukawa coupling $f$,
which gives rise to the fermion mass, $m = f v$.
This short report is organized as follows. In Sec.~II we discuss the particular
form of the cosmic string configuration and describe the
path in weak isospace along which the Dirac eigenvalues are unchanged. In section III
we explain how spectral methods are utilized to compute the fermion contribution to the
\emph{VPE}, including the subtleties needed to make the approach feasible.
We present numerical results for the \emph{VPE} in section IV and show that
this particular invariance is indeed reproduced within our numerical accuracy.
We conclude with a brief summary in Sec.~V and leave some technical
details to appendixes.
\section{Cosmic String Configuration}
The starting point to parametrize cosmic string configurations is the four
dimensional unit vector~\cite{Klinkhamer:1994uy,Graham:2006qt}
\begin{equation}
\hat{\Vek{n}}(\xi_1,\xi_2,\varphi)=\begin{pmatrix}
{\rm sin}\xi_1\,{\rm sin}\xi_2\, {\rm cos}\varphi\cr
{\rm cos}\xi_1 \cr {\rm sin}\xi_1 \,{\rm cos}\xi_2\cr
{\rm sin}\xi_1 \,{\rm sin}\xi_2\, {\rm sin}\varphi
\end{pmatrix}\,,
\label{eq:nhat}
\end{equation}
where $\xi_1$ and $\xi_2$ describe the isospin orientation of the string
and $\varphi$ is the azimuthal angle in coordinate space.\footnote{The string
configuration will be infinitely extended along the $3$-direction in
coordinate space.} For simplicity, we will always consider unit winding of
the string; generalizations to winding number $n$ merely require the replacement
${\rm cos}\varphi\to{\rm cos}(n\varphi)$ and ${\rm sin}\varphi\to{\rm sin}(n\varphi)$.
In what follows we also employ the abbreviations
\begin{equation}
s_i={\rm sin}\xi_i \qquad {\rm and}\qquad
c_i={\rm cos}\xi_i
\label{eq:abbr}
\end{equation}
for the trigonometrical functions of the isospin angles $\xi_1$
and $\xi_2$. A global rotation within the plane of the second
and third components by an angle
$\alpha$ with ${\rm tan}\alpha=s_1c_2/c_1$
transforms the unit vector $\hat{\Vek{n}}$ into
\begin{equation}
\widetilde{\Vek{n}}(\xi_1,\xi_2,\varphi)=\begin{pmatrix}
s_1s_2 {\rm cos}\varphi\cr \sqrt{1-s_1^2s_2^2} \cr
0 \cr s_1s_2 {\rm sin}\varphi
\end{pmatrix}\,.
\label{eq:nhat1}
\end{equation}
Hence observables (which are, by definition, gauge invariant) will not
depend on the two angles $\xi_1$ and $\xi_2$ individually but only on the
product $s_1s_2$. Stated otherwise, all observables must remain
invariant along paths of constant $s_1s_2$ in isospin space~\cite{Klinkhamer:1997hw}.
The unit vector $\hat{\Vek{n}} = (n_0,\Vek{n}) \in S^4$ defines the $SU(2)$ matrix
$U(\xi_1,\xi_2,\varphi)=n_0\mbox{{\sf 1}\zr{-0.16}\rule{0.04em}{1.55ex}\zr{0.1}}-i\Vek{n}\cdot\Vek{\tau}$, where
$\Vek{\tau}=(\tau^1, \tau^2,\tau^3)$ are the three Pauli matrices.
The Higgs and gauge fields of the string are then characterized
by two profile functions $f_H$ and $f_G$ that are functions of the
distance ($\rho$) from the string center:
\begin{align}
\begin{pmatrix} \phi_+(\rho,\varphi) \\[1mm]
\phi_0(\rho,\varphi) \end{pmatrix}=
f_H(\rho) U(\xi_1,\xi_2,\varphi)
\begin{pmatrix}0 \\[1mm] v \end{pmatrix}
\qquad {\rm and}\qquad
\Vek{W}(\rho,\varphi)= \frac{1}{g}\,
\frac{\hat{\Vek{\varphi}}}{\rho} f_G(\rho)\,
U(\xi_1,\xi_2,\varphi)\,\partial_\varphi
U^\dagger(\xi_1,\xi_2,\varphi)\,.
\label{eq:profiles}\end{align}
Here, the gauge connection $\Vek{W}$ is a vector in coordinate space and
a matrix in the adjoint representation of weak isospace. The profile functions
vanish at the core of the string ($\rho=0$) and approach unity at spatial infinity. From this
parametrization we find the classical mass of the string\footnote{Here and in the following,
the prime indicates a derivative with respect to the radial argument $\rho$, and we omit
the argument for simplicity if no confusion can occur.}
\begin{equation}
\frac{E_{\rm cl}}{m^2}=2\pi\int_0^\infty \rho\, d\rho\,\left\{
(s_1s_2)^2\,\biggl[\frac{2}{g^2}
\left(\frac{f_G^\prime}{\rho}\right)^2
+\frac{f_H^2}{f^2\rho^2}\,\left(1-f_G\right)^2\biggr]
+\frac{f_H^{\prime2}}{f^2}
+\frac{\mu_h^2}{4f^2}\left(1-f_H^2\right)^2\right\}\,,
\label{eq:ecl}
\end{equation}
where the dimensionless radial integration variable is related to the physical
radius by $\rho_{\rm phys}=\rho/m$ and we have introduced the mass ratio
$\mu_H\equiv m_H/m$. As expected, the classical mass only depends on the isospin
angles via the combination $s_1s_2$, which reflects gauge invariance.
Note that the configuration, Eq.~(\ref{eq:profiles}) approaches a local gauge
transformation of the constant vacuum configuration at spatial infinity. As a
consequence, this configuration is not appropriate for techniques that
require some kind of perturbative expansions which do not preserve
gauge invariance order by order. In particular, individual Fourier
transformations of the Higgs and gauge fields are ill defined. We therefore introduce
an additional radial function $\xi(\rho)$ with the boundary values $\xi(0)=0$ and
$\lim_{\rho\to\infty}\xi(\rho)=\xi_1$ to define the local $SU_L(2)$
gauge transformation
\begin{equation}
V={\rm exp}\left[-i\Vek{\tau}\cdot\Vek{\xi}(\rho,\varphi)\right]
\qquad {\rm with}\qquad
\Vek{\xi}(\rho,\varphi)=\xi(\rho)\,\begin{pmatrix}
s_2\, {\rm cos}\varphi \cr
-s_2\, {\rm sin}\varphi \cr c_2
\end{pmatrix}\,.
\label{eq:localgauge}
\end{equation}
Since $\xi(0)=0$ this gauge transformation does not introduce any
singularity at the origin; at spatial infinity it accounts for the above
mentioned gauge transformation of the constant vacuum. With the gauge
transformation, Eq.~(\ref{eq:localgauge}) applied, perturbative expansions
can be performed. Of course, this comes at the expense of an additional
radial function. By construction, observables are independent of its detailed form
as long as the boundary conditions described above are maintained. For the particular
case of $\xi_2=\frac{\pi}{2}$ this was verified in Ref.\cite{Weigel:2010pf}.
In the present study, we will also consider deviations from that particular
parameter value. We emphasize that the introduction of the gauge rotation,
Eq.~(\ref{eq:localgauge}) has effectively made our test isospin symmetry
\emph{local}, since $\xi_1$ has turned into a space dependent quantity.
To write down the Dirac Hamiltonian from which we compute the spectrum
of the fermion fluctuations we extract the Hamiltonian, $\mathcal{H}$
from the Lagrangian, Eq.~(\ref{gaugelag}) and then perform the
left--handed gauge transformation defined in Eq.~(\ref{eq:localgauge}):
$H=\left(P_R+VP_L\right)\mathcal{H}\left(P_R+VP_L\right)^\dagger$. To
simplify the presentation we define $\Delta(\rho)\equiv\xi_1-\xi(\rho)$
and separate the interaction part (again using dimensionless variables)
\begin{eqnarray}
H&=&-i\begin{pmatrix}0 & \Vek{\sigma}\cdot\hat{\Vek{\rho}} \cr
\Vek{\sigma}\cdot\hat{\Vek{\rho}} & 0\end{pmatrix} \partial_\rho
-\frac{i}{\rho}\begin{pmatrix}0 & \Vek{\sigma}\cdot\hat{\Vek{\varphi}}\cr
\Vek{\sigma}\cdot\hat{\Vek{\varphi}} & 0\end{pmatrix} \partial_\varphi
+\begin{pmatrix} 1 & 0 \cr 0 &-1\end{pmatrix}
+H_{\rm int}\,, \label{eqDirac0}\\[4mm]
H_{\rm int}&=&
\left[\left(f_H{\rm cos}(\Delta)-1\right)
\begin{pmatrix} 1 & 0 \cr 0 &-1\end{pmatrix}
+if_H\,{\rm sin}(\Delta)\begin{pmatrix}0 & 1 \cr -1 & 0\end{pmatrix}
I_H\right]
+\frac{1}{2}\frac{\partial \xi}{\partial \rho}
\begin{pmatrix}-\Vek{\sigma}\cdot\hat{\Vek{\rho}}
& \Vek{\sigma}\cdot\hat{\Vek{\rho}} \cr
\Vek{\sigma}\cdot\hat{\Vek{\rho}}
& -\Vek{\sigma}\cdot\hat{\Vek{\rho}}\end{pmatrix}I_H
\nonumber \\[3mm]
&&
+\frac{s_2}{2\rho}\, \begin{pmatrix}
-\Vek{\sigma}\cdot\hat{\Vek{\varphi}}
& \Vek{\sigma}\cdot\hat{\Vek{\varphi}} \cr
\Vek{\sigma}\cdot\hat{\Vek{\varphi}}
& -\Vek{\sigma}\cdot\hat{\Vek{\varphi}}\end{pmatrix}
\Big[f_G\,{\rm sin}(\Delta)\,I_G(\Delta)
+(f_G-1)\,{\rm sin}(\xi)\,I_G(-\xi)\Big]\,.
\label{eqDirac}
\end{eqnarray}
The isopsin matrices in this expression are
\begin{equation}
I_H=\begin{pmatrix}
c_2 & s_2 {\rm e}^{i\varphi} \cr
s_2 {\rm e}^{-i\varphi} & -c_2 \end{pmatrix}
\qquad {\rm and} \qquad
I_G(x)=\begin{pmatrix}
-s_2{\rm sin}(x) & [c_2{\rm sin}(x)-i{\rm cos}(x)]\,{\rm e}^{i\varphi} \cr
[c_2{\rm sin}(x)+i{\rm cos}(x)]\,{\rm e}^{-i\varphi} &
s_2{\rm sin}(x)\end{pmatrix}\,.
\label{eq:IG}
\end{equation}
Note that the latter appears with different arguments in Eq.~(\ref{eqDirac}).
Nothing from the invariance along the path with $s_1s_2={\rm const.}$ is
manifest in Eq.~(\ref{eqDirac}), neither is the gauge invariance from
Eq.~(\ref{eq:localgauge}).
\bigskip\noindent
To proceed, we diagonalize the Hamiltonian in a basis of wave functions
\begin{equation}
\Psi_{\ell}(\rho,\varphi) = \sum_{s,j = \pm \frac{1}{2}}
\Big(\langle \rho\,|\, \langle\,\varphi\,;\,S\,I\,| \Big)\,
|\epsilon \,\ell\,s\,j\,\rangle \,.
\label{eq:GSansatz}
\end{equation}
that decouple radial and angular coordinates in the upper
and lower components of the Dirac spinors
($\epsilon$ refers to the energy eigenvalue
defined in Eq.~(17) below)
\begin{equation}
\begin{array}{ll}
\langle \rho |\epsilon\,\ell++\rangle =
\begin{pmatrix}f_1(\rho)|\ell + +\rangle \cr
g_1(\rho)|\ell - +\rangle \end{pmatrix} & \qquad
\langle \rho |\epsilon\,\ell+-\rangle =
\begin{pmatrix}f_2(\rho)|\ell + -\rangle \cr
g_2(\rho)|\ell - -\rangle \end{pmatrix}
\cr \cr
\langle \rho |\epsilon\,\ell-+\rangle =
\begin{pmatrix}f_3(\rho)|\ell - +\rangle \cr
g_3(\rho)|\ell + +\rangle \end{pmatrix}& \qquad
\langle \rho |\epsilon\,\ell--\rangle =
\begin{pmatrix}f_4(\rho)|\ell - -\rangle \cr
g_4(\rho)|\ell + -\rangle \end{pmatrix}\,,
\end{array}
\label{eq:GSspinors}
\end{equation}
The notation is such that the signs denote the spin
and isospin projection quantum numbers. For instance,
\begin{equation}
\langle \varphi;SI|\ell + +\rangle=
{\rm e}^{i(\ell+1)\varphi}\,
\begin{pmatrix}1 \cr 0 \end{pmatrix}_S
\otimes\begin{pmatrix}1 \cr 0 \end{pmatrix}_I \,.
\label{eq:spinexample}
\end{equation}
Diagonalization means that we construct the eigenvalues
of the stationary Dirac equation
\begin{equation}
H\Psi=\epsilon\Psi\,,
\label{eqDirac1}
\end{equation}
with $|\epsilon|<1$. For $|\epsilon|>1$ we construct the full
scattering matrix as a function of momentum $k=\sqrt{\epsilon^2-1}$.
Since we employ four component Dirac spinors, we have
$\left\{H,\alpha_3\right\}=0$ and the spectrum is charge conjugation
invariant. Our final result of the scattering problem (described in
appendix A) is the Jost function $\nu(t)$ for imaginary momentum
$k=it$, as well as the first two terms of its Born expansion obtained by
iterating the interaction part $H_{\rm int}$.
\section{Vacuum Polarization Energy (VPE)}
\label{sec:vpe}
The main goal of the present investigation is to verify that our
treatment of the ultraviolet divergences does not produce any
dependence on the isospin angles $\xi_1$ and $\xi_2$ that cannot be
expressed as $s_1s_2$. Any change in the renormalization conditions is
described by finite counterterms. As was the case for the classical energy
Eq.~(\ref{eq:ecl}), the counterterms are manifestly functions of $s_1s_2$.
We are therefore free to employ the simplest renormalization scheme, which
is $\overline{\rm MS}$. For the profile functions we choose a specific
form and introduce dimensionless width parameters $w_G$, $w_H$ and $w_\xi$:
\begin{equation}
f_H(\rho)=1-\exp\left(-\frac{\rho}{w_H}\right)\,,\quad
f_G(\rho)=1-\exp\left(-\frac{\rho^2}{w_G^2}\right)
\quad \mbox{and} \quad
\xi(\rho)=\xi_1\left[1-\exp\left(-\frac{\rho^2}{w_\xi^2}\right)\right]\,.
\label{eqn:profile}
\end{equation}
Observable values for the width parameters are in units of $m^{-1}$ since
$\rho_{\rm phys}=\rho/m$. Recall again that $\xi(\rho)$ is just an auxiliary
profile describing the local gauge transformation, Eq.~(\ref{eq:localgauge}),
and that the \emph{VPE} should be independent of $w_\xi$.
With these conventions on the ansatz parameters, the \emph{VPE}
depends on the model parameters $g$, $f$ and $v$ only via the overall factor
$m^2=(vf)^2$, see also Eq.~(\ref{eq:ecl}). In this sense the dependence
on the model parameters is completely contained in the classical energy and
the counterterms, and thus requires little numerical effort.
\medskip\noindent
The spectral method \cite{Graham:2009zz} to compute the \emph{VPE} from scattering
data identifies the change of the density of states caused by a static background
as the derivative of the scattering phase shift (also known as the phase of the Jost
function for real momenta) via the Krein--Friedel--Lloyd formula,
cf.~Ref.~\cite{Faulkner:1977aa} and references therein. More precisely, we obtain the
phase shift as $(-i/2){\rm ln}({\rm det}S)$, where $S$ is the scattering matrix
of the multichannel problem. Integration over the momentum along the string then
yields the \emph{VPE} per unit length. However, that integral is only finite due to
particular sum rules among the scattering data \cite{Graham:2001iv}. Ultimately
this leads to the interface formalism \cite{Graham:2001dy} in which we only need
to integrate over the momentum $k$ of the scattering problem in the plane perpendicular
to the string. In this situation, it is prudent to use the analytic properties
of the scattering data to perform the final momentum integral over imaginary
momentum $t$ with $k=it$. This analytic continuation has several advantages: First,
it allows to interchange the momentum integral with the angular momentum
sum~\cite{Schroder:2007xk} and second, it implicitly collects the contributions
to the \emph{VPE} coming from the bound states. This is beneficial, as there is
generally a large number of such states, in particular for wide strings, and
identifying them numerically is cumbersome. To express the \emph{VPE} as an integral over
imaginary momenta it is essential that the scattering phase shift is an odd function
of the real momentum. Typically this property results from the Hamiltonian being
real~\cite{Chadan:1977pq,Newton:1982qc} which is, however, not the case here:
The gauge transformation, Eq.~(\ref{eq:localgauge}) turns the global isospin transformation
along the path $s_1 s_2={\rm const.}$ into a local one
and, consequently, there is
no global transformation on the basis states, Eq.~(\ref{eq:GSspinors}) which could result in
a real Hamiltonian.\footnote{The Hamiltonian is still Hermitian, of course, and the
single particle energies are real.} In Appendix~A we show that nevertheless the phase
shift is odd in the momentum.
After collecting all information the \emph{VPE} per unit length of the string is expressed as
\begin{equation}
E_{\rm vac}=\frac{m^2}{2\pi} \int_0^\infty d\tau\, \tau
\left\{\sum_{\ell} D_{\ell}\left[\nu(\tau,\ell)-\nu_1(\tau,\ell)-\nu_2(\tau,\ell)\right]
-\frac{c_F}{c_B}\sum_{\ell}\bar{D}_\ell\bar{\nu}_2(\tau,\ell)\right\}
+E_2+E_{\rm f.b.}\,,
\label{eq:master}
\end{equation}
where we performed a final change of variable $t\to\tau=\sqrt{t^2-1}$ to avoid the
integrable singularity at $t=m$. In Eq.~(\ref{eq:master}) $\nu$ is the full Jost function
with orbital angular momentum $\ell$ and degeneracy factor $D_\ell=2-\delta_{\ell,-1}$
on the imaginary momentum axis, while $\nu_1$ and $\nu_2$ are first two terms of its
Born expansion with respect to $H_{\rm int}$. These two subtractions are performed
before summing over angular momentum channels. This is indispensable in order to
identify and disentangle the subleading logarithmic divergence and the relevant finite
contributions from the two leading Born terms. In fact, the logarithmic divergence has
additional contributions from the third and fourth order Feynman diagrams, and their
total strength\footnote{Here, the term ``strength'' means that the Feynman diagrams
produce the singularity $\frac{c_F/2\pi}{4-D}$ in dimensional regularization.
In Ref.~\cite{Graham:2011fw} a factor 4 was omitted in the definition of both $c_F$
and $c_B$, so that the ratio remains unaffected.}
is $c_F$. The second order contribution of quantum corrections from a complex boson field
about a static background also produces a logarithmic divergence. Let $c_B$ be its
strength and $\bar{\nu}_2(\tau,\ell)$ the second order Born term of its Jost function for
imaginary momenta in the angular momentum channel $\ell$. Then the last term
in curly brackets of Eq.~(\ref{eq:master}) removes the logarithmic divergence
from the integral. Since there is no further (sub--subleading)
divergence, this subtraction can be made after summing over angular
momenta. In the last step all subtractions are added back in the form of
Feynman diagrams. They are computed by standard techniques using, {\it e.g.},
dimensional regularization. Their divergent parts are uniquely compensated
by counterterms in a definite renormalization scheme. All that remains
are the finite parts $E_2$ and $E_{\rm f.b.}$ of the second order fermion and
fake boson diagrams, which correspond to the finite parts of the subtractions
$\nu_{1,2}$ and $\bar{\nu}$, respectively. Equation~(\ref{eq:master}) is the master
formula to compute the \emph{VEV} of string configurations.
We stress that only the very first term under the integral in eq.~(\ref{eq:master})
remains unchanged when varying the string isospin orientation, provided that
${\rm sin}(\xi_1){\rm sin}(\xi_2)$ remains constant. All other contributions are
more general functions of $\xi_1$ and $\xi_2$ and thus vary along our particular isospin path of
constant ${\rm sin}(\xi_1){\rm sin}(\xi_2)$.
These terms should eventually cancel provided that the
identity of Born and Feynman series holds. However, individually they
represent ill defined ultraviolet divergent quantities that
undergo distinct regularization procedures and it is therefore unclear whether the
spectral approach and, in particular, the renormalization procedure spoil
gauge invariance. We will investigate this question numerically in the next section.
\section{Results}
The computation of the momentum integral and its integrand in Eq.~(\ref{eq:master})
is by far the most expensive part of the numerical procedure. To begin with,
the $\ell=-1$ and $\ell=0$ channels require particular consideration. They
involve Hankel functions of order zero whose irregular component diverges
logarithmically at small arguments rather than by an inverse power law. Thus
regular and irregular components are numerically difficult to separate. When
integrating the radial differential equation [Eqs.~(\ref{deqGFin2}) and~(\ref{defDZud})
for $k=it$] we take the lower boundary to be $\rho_{\rm min}\sim 10^{-50}$ for these
two channels, and from $\rho_{\rm min}$ we extrapolate to
$\rho=0$. In other channels a lower boundary of $\rho_{\rm min}\sim 10^{-12}$
is fully reliable. Angular momenta are typically summed up to $\ell_{\rm max}=600$
or $\ell_{\rm max}=700$ above which numerical stability for Hankel functions at
small arguments is lost. For background profiles with small or moderate widths this
gives sufficient accuracy. Once the angular momentum sum is completed, the analog
contribution from the fake boson (mimicking the logarithmic ultraviolet divergences
from third and fourth order Feynman diagrams) is subtracted and the large $\tau$
behavior of the integrand is treated by fitting a $1/\tau^3$ tail, {\it cf.}
the right panel of Fig.~\ref{fig:diffwx}. Finally,
for wider profiles an additional extrapolation of the angular momentum sum to
$\ell_{\rm max}\to\infty$ is necessary which typically adds about $1\ldots2 \%$
to the \emph{VPE}.
We start with a few examples, displayed in Table~\ref{tab:wx} and Fig.
\ref{fig:diffwx}, in order to verify the independence from the gauge
profile $\xi(\rho)$.
\begin{table}
\centerline{
\begin{tabular}{c|c|c|c||c||c}
$w_\xi$ & $E_{\delta}$ & $c_F$ & $E_{\rm FD}$ & $E_{\rm vac}$
& $|E_{\delta}|+|E_{\rm FD}|$\cr
\hline
2.0 & 0.3010 & -10.00 & -0.0108 &~~0.2902~~~& 0.3118 \cr
3.5 & 0.2974 & -11.59 & -0.0072 &~~0.2902~~~& 0.3046 \cr
5.0 & 0.2953 & -14.29 & -0.0047 &~~0.2905~~~& 0.3000 \cr
6.5 & 0.2915& -17.82 & -0.0015 &~~0.2901~~~& 0.2930
\end{tabular}}
\caption{\label{tab:wx}Example for the invariance with respect to
the local gauge transformation, Eq.~(\ref{eq:localgauge}) with
$E_{\rm FD}=E_2+E_{\rm f.b.}$. Listed are all ingredients from
Eq.~(\ref{eq:master}) that explicitly depend on
the width $w_\xi$ of the gauge profile $\xi(\rho)$.
Parameters are $w_G=w_H=4.82$, $\xi_1=0.3\pi$ and $\xi_2=0.25\pi$.
In all cases an identical fake boson profile was employed because
it affects $E_{\rm FD}$.}
\end{table}
\begin{figure}
\centerline{
\epsfig{file=diff2nd.png,width=8cm,height=7.0cm}\hspace{1cm}
\epsfig{file=diffall.png,width=8cm,height=7.0cm}}
\caption{\label{fig:diffwx}(Color online) Partial sums that enter the \emph{VPE},
Eq.~(\ref{eq:master}), for the string profiles with $w_G=w_H=4.82$, $\xi_1=0.3\pi$
and $\xi_2=0.25\pi$. The left panel shows the fermion part for four different
values of $w_\xi$. The right panel shows the total integrand for three values
of $w_\xi$ relative to $w_\xi=2.0$. The double--dashed lines that start
at $\tau=2$ in the right panel are simple power decays which serve to
guide the eye on the large $\tau$ behavior. Note the different
scales in the two panels.}
\end{figure}
The variation of the individual contributions to the \emph{VPE} is an order of
magnitude larger than that of the total result. The tiny variation of the
latter is due to errors from the numerical simulation. The cancellation
of the gauge variant parts for the \emph{VPE} is most obvious when adding them
as absolute values which contains spreads of up to 10\%. A large variation
appears in the fermion part of the momentum integral (\emph{i.e.}~the
contribution from the first term in curly brackets) in Eq.~(\ref{eq:master}),
as can be seen in Fig.~\ref{fig:diffwx}.
Even though we have just established that the \emph{VPE} does not vary
with the width of the gauge profile, it is prudent for numerical efficiency
and stability to choose that width similar to one in the profile
functions of the physical boson fields, because otherwise large angular momenta
play too significant a role.
\begin{table}
\centerline{
\begin{tabular}{c|c|c|c|c||c||c}
$\xi_1/\pi$ &$\xi_2/\pi$ & $E_{\delta}$ & $c_F$ & $E_{\rm FD}$ & $E_{\rm vac}$
& $|E_{\delta}|+|E_{\rm FD}|$\cr
\hline
0.1 & 0.4 & 0.1504 & -4.913 & 0.0014 &~~0.1518~~~& 0.1518\cr
0.4 & 0.1 & 0.1702 & -8.541 & -0.0180 &~~0.1521~~~& 0.1882 \cr
0.3 & 0.11834 & 0.1496 & -6.814 & 0.0021 &~~0.1517~~~& 0.1517 \cr
0.2 & 1/6 & 0.1639 & -5.615 & -0.0117 &~~0.1522~~~& 0.1758
\end{tabular}}
\caption{\label{tab:zz} Contributions to Eq.~(\ref{eq:master}) and their
variations with the isospin angles. In all cases we have $s_1s_2\approx0.29389$.
The width parameters of the boson profiles are $w_G=w_H=3.5$. The results were
obtained with various values for the widths of the gauge and fake boson
profiles.}
\end{table}
\begin{figure}
\centerline{
\epsfig{file=zz.png,width=9cm,height=6cm}}
\caption{\label{fig:zz}(Color online) The fermion part of the momentum integrand in
Eq.~(\ref{eq:master}) (similar to the left panel of Fig.~\ref{fig:diffwx}).
The selected width parameters are $w_H=w_G=3.5$.}
\end{figure}
In Fig.~\ref{fig:zz} we show the strongly varying fermion part of the integrand
for the \emph{VPE} for sets of isospin angles that produce identical products
$s_1s_2$. Despite the pronounced variation of this particular piece, the total
\emph{VPE} only differs at the order of the numerical accuracy as can clearly be seen
from the data in Table~\ref{tab:zz}. The comparison with the (incorrect) addition
of the absolute values of the gauge variant contributions further illustrates this
observation.
For $\xi_2=\frac{\pi}{2}$ the Hamiltonian is real. In this
simpler case the \emph{VPE} was computed for about 50 sets of width parameters
($w_H$, $w_G$, {\it cf.} Appendix B) and eight different values for
$\xi_1\in[0,\frac{\pi}{2}]$ in Ref.~\cite{Graham:2011fw}. These results\footnote{We
have reproduced these earlier results for $\xi_2=\frac{\pi}{2}$
using the more general numerical simulation for the complex Hamiltonian.} were then
used to establish stable charged cosmic strings for fermion masses only slightly
larger than that of the top quark. Here we consider the same sets of width parameters
for two pairs of isospin angles that yield the identical products $s_1s_2$. In the
first of the two pairs we simply swap the isospin angles as compared to the earlier
calculations \cite{Graham:2011fw}, and show the resulting
\emph{VPE} (in the $\overline{\rm MS}$ renormalization scheme) in Fig.~\ref{fig:x1x2A}.
\begin{figure}
\centerline{
\epsfig{file=x1x2.png,width=8cm,height=6.5cm}\hspace{1cm}
\epsfig{file=x1x2a.png,width=8cm,height=6.5cm}}
\caption{\label{fig:x1x2A}(Color online) The vacuum polarization energy for different
background profiles with the two isospin angles swapped. In the right panel we zoom
in by omitting narrow profiles that suffer from the Landau ghost
problem \cite{Ripka:1987ne,Hartmann:1994ai,Graham:2011fw}. Details of the profiles
are listed in appendix B.}
\end{figure}
Obviously the computed \emph{VPE}s agree within the numerical accuracy for the full
range of considered width parameters. However, merely swapping the isospin angles is
not sufficient to fully establish dependence on only the product $s_1s_2$. For
example, there could be gauge variant contributions involving ${\rm sin}(\xi_1+\xi_2)$.
To rule out such a dependence, we have made a second study and
compared the two sets $(\xi_1,\xi_2)=(0.1,0.4)\pi$ and $(\xi_1,\xi_2)=(0.3,0.11834)\pi$.
The resulting \emph{VPEs} are shown in Fig.~\ref{fig:x1x2B}.
\begin{figure}
\centerline{
\epsfig{file=y1y2.png,width=8cm,height=6.5cm}\hspace{1cm}
\epsfig{file=y1y2a.png,width=8cm,height=6.5cm}}
\caption{\label{fig:x1x2B}Same as Fig.~\ref{fig:x1x2A} for
a second pair of isospin angles.}
\end{figure}
Again we observe perfect agreement for the computed \emph{VPEs} as the tiny numerically
discrepancies are not resolved within Figs.~\ref{fig:x1x2A} and \ref{fig:x1x2B}.
So we conclude that the spectral methods to compute the \emph{VPE} of cosmic strings
indeed preserve gauge and isospin invariance even though some of its components do not.
The comparison of the results in Fig.~\ref{fig:x1x2A} with those in Fig.~\ref{fig:x1x2B}
suggests that the \emph{VPE} depends on the isospin orientation only mildly,
except for the very narrow configurations that suffer from the Landau ghost
problem~\cite{Ripka:1987ne,Hartmann:1994ai,Graham:2011fw}. This is not quite the case:
in the current study our goal is to compare the \emph{VPE} for configurations with equal
$s_1s_2$, as in either of Figs.~\ref{fig:x1x2A} or \ref{fig:x1x2B}.
To reveal the discussed invariance, the difference between the two angles
is usually chosen deliberately large, so that
one of the angles is always small and so is the product $s_1s_2$.
When we lift this restriction we find e.g.~with $w_G=w_H=6.0$ that
$E_{\rm vac}$ increases from $0.438m^2$ to $0.479m^2$ between
$s_1s_2=0$ and $s_1s_2=1$.
In a separate study we have implemented a boundary condition at large
separation from the string to construct discretized basis states
that serve to compute matrix elements of the Dirac Hamiltonian, Eq.~(\ref{eqDirac}).
These matrix elements form a complex Hermitian matrix that we have diagonalized
using \texttt{LAPACK}~\cite{laug1999}. Eigenvalues below threshold are identified
as bound state energies. We have verified that all energy eigenvalues of the Dirac
Hamiltonian remain unchanged when altering $\xi_1$ and $\xi_2$ such that $s_1s_2$
stays constant. This is expected for bound states that have no support in the
vicinity of the boundary. Scattering states, however, reach out to spatial
infinity and are thus sensitive to the discretizing boundary conditions which
are not manifestly gauge invariant; so the invariance of these states
comes as some surprise. In addition, this discretization approach
requires to impose a numerical cutoff on the energy to produce a finite
dimensional Hamiltonian matrix. The levels slightly below that cutoff
exhibit a soft variation along the path of invariance in isospace.
This reflects the fact that unitarity of the transformation is lost for a
finite dimensional Hilbert space. Similarly, such near-cutoff energies
do also vary with the gauge profile $\xi(\rho)$. Renormalized
\emph{VPE} calculations based on this or similar numerical discretization
approaches \cite{Diakonov:1993ru} will probably be erroneous.
In the spectral approach, we consequently use the discretization technique
only for the bound states, while scattering states are treated
in the continuum formulation.
Finally we note in passing that we have numerically verified the bound state energies
from the above discretization computation against the roots of the Jost function on
the imaginary axis and also ensured that the number of bound states satisfies
Levinson's theorem.\footnote{For the bound states the discretization
procedure is advantageous because root finding algorithms may fail to identify degenerate
bound states that appear in multi--channel scattering. Also identifying the roots very
close to threshold is numerically cumbersome.}
\section{Conclusion}
There are numerous obstacles in computing the \emph{VPE}
of string type configuration in gauge theories that are similar to the standard
model of particle physics. Within the so--called spectral approach, these obstacles
can be overcome by an interplay of techniques which individually are \emph{not}
gauge invariant. If the spectral approach is a meaningful tool in gauge theories,
it must ensure that the gauge variant contributions eventually cancel. To the best
of our knowledge there is no formal proof of this cancellation at the moment, and it
is also far from obvious because the gauge-variant contributions are related to
ultraviolet divergent quantities that undergo different methods of regularization.
Hence analytical or numerical verifications of gauge invariance in the
spectral approach are indispensable.
In the present study we have therefore comprehensively revisited the
computation of the \emph{VPE} for string type configurations arising from
fermion fluctuations, in order to justify and validate earlier computations
(carried out in a limited parameter space) that suggested novel solutions
in theories closely related to the standard model \cite{Weigel:2010zk}. Those
earlier studies were implicitly based on the assumption that the spectral method
would not spoil gauge invariance as the identification of Born and Feynman series
would hold even for (differently regularized) divergent contributions.
Here we have extended the parameter space for an independent numerical corroboration
of this assumption. It employs the invariance of the spectrum of the
Dirac Hamiltonian along a particular path in the enlarged parameter space.
This invariance must be reflected in the \emph{VPE}. However, this is not manifest
in the actual \emph{VPE} calculation, because regularization and renormalization
indeed require delicate operations on divergent contributions that vary under the
isospin transformation.
Our numerical simulations show that individual contributions that are not gauge invariant
but need to be included for regularization and renormalization may vary by 10\% or more
along the path of isospin invariance. But then, the contributions combine such that these
variations actually do cancel in the total result, leading to changes of the fermion
quantum energy of the cosmic string along the path of isospin invariance of the order of
only a fraction of a percent. Such variations are within the bounds of the numerical accuracy.
Thus we have verified numerically that the spectral method preserves gauge invariance and
is hence a valid tool to study quantum corrections to extended configurations, such as
cosmic strings in the standard model of particles.
\acknowledgments
H.~W.\ is supported in part by the NRF (South Africa) by Grant No.~77454.
N.~G.\ is supported in part by the NSF through Grant No.~PHY15-20293.
|
\subsection*{\centering Abstract}
{\em
In genome-wide association studies (GWASs) of common diseases/traits, we often analyze multiple GWASs with the same phenotype together to discover associated genetic variants with higher power. Since it is difficult to access data with detailed individual measurements, summary-statistics-based meta-analysis methods have become popular to jointly analyze data sets from multiple GWASs.
In this paper, we propose a novel summary-statistics-based joint analysis method based on controlling the joint local false discovery rate (Jlfdr). We prove that our method is the most powerful summary-statistics-based joint analysis method when controlling the false discovery rate at a certain level. In particular, the Jlfdr-based method achieves higher power than commonly used meta-analysis methods when analyzing heterogeneous data sets from multiple GWASs. Simulation experiments demonstrate the superior power of our method over meta-analysis methods. Also, our method discovers more associations than meta-analysis methods from empirical data sets of four phenotypes. The R-package is available at: \url{http://bioinformatics.ust.hk/Jlfdr.html}.
}
\section{Introduction}
Understanding genetic mechanisms of common diseases and traits is important in biological and medical research. The goal of genome-wide association studies (GWASs) is to discover the susceptibility of single nucleotide polymorphisms (SNPs) to common diseases/traits \citep{altshuler2008genetic}. Due to decreasing genotyping costs \citep{perkel2008snp}, constantly emerging successful stories \citep{klein2005complement, kraft2010gwas} and efforts of the GWAS consortiums \citep{burton2007genome, schizophrenia2014biological}, more and more GWASs have been conducted for common phenotypes \citep{welter2014nhgri}.
Analyses of GWAS results show that the identified associations can only explain a small part of the additive genetic variances. This is referred to as the ``missing heritability'' problem \citep{manolio2009finding}. The hints of hidden heritability \citep{gibson2010hints, yang2010common} and the estimated distribution of common SNPs' effect sizes \citep{park2010estimation} suggest that common diseases/traits are influenced by thousands of SNPs with small effects. To discover these genetic variants with small effects, we need to improve studies' power. Jointly analyzing data sets from multiple GWASs on the same diseases in the same population provide an opportunity to improve the power.
There are two kinds of joint analysis methods: individual-level joint analysis and summary-statistics-based joint analysis. Individual-level joint analysis uses individual-level genotype data from all studies. One such example is mega-analysis \citep{ripke2013mega}, which pools all data together. Summary-statistics-based joint analysis only uses summary statistics from different studies. Since individual-level genotype data is difficult to access, summary-statistics-based analysis is widely used in joint analysis. The most commonly used method of summary-statistics-based joint analysis is meta-analysis \citep{evangelou2013meta}, which derives a new statistic for each SNP using summary statistics from multiple studies.
Our focus in this paper is to study summary-statistics-based joint analysis methods. More specifically, we like to study which joint analysis method provides the highest power for a given false discovery rate level. Figure \ref{Figure1} illustrates our motivation.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.90\textwidth]{Figure1.pdf}
\caption{\textbf{Rejection boundaries determined by different summary-statistics-based joint analysis methods: the optimal method and the meta-analysis method.} Assume we jointly analyze data from two GWASs. For simplicity, we assume the tests are one-sided. We plot the test statistic pair $(z^{(1)}, z^{(2)})$ into the coordinate plane. A SNP at the upper right corner shows more significant association than a SNP at the bottom left corner. The true associated SNPs are plotted with blue circles, and the false associated SNPs are plotted with yellow triangles. For each rejection boundary, the SNPs in the upper right region are discovered. All three analysis methods have the same false discovery proportion (10\%). The optimal method has more empirical power (red solid line, 72\%) than the meta-analysis method (purple dashed line, 36\%).}\label{Figure1}
\end{figure}
Our major contribution in this paper is that we propose a novel summary-statistics-based joint analysis method based on controlling the joint local false discovery rate (Jlfdr). The Jlfdr generalizes the concept of the local false discovery rate \citep{efron2005local} from the analysis of single study to the joint analysis of multiple studies. We prove that our method is the most powerful summary-statistics-based joint analysis method for a given false discovery rate level. In particular, the Jlfdr-based method is more powerful than commonly used meta-analysis methods when analyzing heterogeneous data sets from multiple GWASs.
The rest of this paper is organized as follows. In Section \ref{methods}, we will first give the mathematical formulation of summary-statistics-based joint analysis methods. We will prove that the most powerful summary-statistics-based joint analysis method should control the Jlfdr. Then we will give implementation details of the Jlfdr-based method under the Gaussian mixture model. We will also discuss the relationship between the Jlfdr-based method and meta-analysis methods. In Section \ref{results}, we will use simulation experiments to demonstrate that Jlfdr-based method outperforms meta-analysis methods in terms of achieving higher power. Then we will show the empirical results using four different data sets.
In Section \ref{discussion}, we will discuss limitations of our current method. Section \ref{conclusion} concludes the paper.
\section{Method}\label{methods}
\subsection{Notations and criteria}
Our method deals with a multiple GWAS setting. For simplicity, we illustrate the concepts with a two-GWAS setting. We use parenthesized superscript ``$(j)$'' to denote the study index. For example, the sample sizes in study 1 and 2 are $n^{(1)}$ and $n^{(2)}$, respectively. We use subscript $i$ ($i=1,\dots, m$, $m$ is the total number of genotyped SNPs) to denote the SNP index.
To detect associations, we construct a null hypothesis for each SNP, in which association is assumed nonexistent. Assume we use a $z$-value scheme to detect associations between SNPs and the phenotype, i.e., the test statistics follow a standard normal distribution under a null hypothesis. We use $\widehat{\mu}^{(j)}$ to denote the observed effect size in study $j$. The asymptotically standard error of $\widehat{\mu}^{(j)}$ is $\sigma^{(j)}$. Correspondingly, the test statistic in study $j$ is $z^{(j)}=\widehat{\mu}^{(j)}/\sigma^{(j)}$. The underlying expected effect size is $\mu^{(j)}$. The expected effect size of the same SNP may vary in different studies due to heterogeneity. The test statistic $Z^{(j)}$ (uppercase letter indicates a random variable) follows an $N(\mu^{(j)}/\sigma^{(j)},1)$ distribution. We use $\pmb{z}$ to represent the vector of test statistics in all studies, i.e., $\pmb{z}=(z^{(1)},z^{(2)})^T$. Similarly, we use $\pmb{\mu}$ to represent the vector of expected effect sizes in all studies, i.e., $\pmb{\mu}=(\mu^{(1)}, \mu^{(2)})^T$.
We further assume $m_0$ SNPs have no association with the phenotype and $m_1$ SNPs have associations. Thus, the null proportion reads $\pi_0=m_0/m$ ($0\leq \pi_0\leq 1$). We use $\mathcal{H}_0$ and $\mathcal{H}_1$ to denote the null hypothesis and the alternative hypothesis, respectively.
In the joint analysis of summary statistics from multiple GWASs, we assume that $R$ of the $m$ hypotheses are rejected. There are $V$ false positives and $S$ true positives (i.e., $V+S=R$). Table \ref{contTable} summarizes the numbers of hypotheses in the different categories.
\begin{table}[!htbp]
\centering
\caption{\textbf{The status of all hypotheses in the joint analysis}. The letter in each cell denotes the count of the hypotheses in each category. }\label{contTable}
\begin{tabular}{c|cc|c}
& $\mathcal{H}_0$ is true & $\mathcal{H}_0$ is false & Total\\ \hline
$\mathcal{H}_0$ is rejected & $V$ & $S$ & $R$\\
$\mathcal{H}_0$ is not rejected & $U$ & $T$ & $m-R$\\ \hline
Total & $m_0$ & $m_1$ & $m$
\end{tabular}
\end{table}
When testing multiple hypotheses, it is very easy to have false positives by random chance. This problem is known as the ``multiplicity'' problem. Many criteria are proposed to address the multiplicity problem. We present an incomplete list of these criteria in Table \ref{criteria} (a). Let's define the false discovery proportion (Fdp) as $V/(R\vee 1)$ with ``$\vee$'' denoting the maximum operation. Fdp is an unknown quantity in real cases. The classical false discovery rate (FDR) is the expectation of the Fdp. Controlling the FDR is more powerful than controlling the family-wise error rate (FWER). The Bayesian false discovery rate (Fdr) is the expected value of the Fdp given $R>0$. Compared to FDR, Fdr is conditional on $R>0$ since we are only interested in controlling false positives when $R>0$. We adopt Fdr in this paper as the criterion to avoid a plethora of false positives.
\begin{table}[!htb]
\caption{Different criteria for evaluating a rejection region in multiple testing scenario. Here $\mathcal{R}$ is the rejection region in the analysis. $\pmb{\mu}$ denotes effect sizes. $V$ and $R$ as well as other notations are explained in Table \ref{contTable}. ``$\vee$'' denotes the maximum operation.}\label{criteria}
\begin{subtable}{\linewidth}
\centering
\caption{Different criteria for controlling false positives in multiple testing scenario.}
\begin{tabular}{c|c|c}
Criteria & Mathematical Definitions & References \\ \hline
Family-wise error rate (FWER) & $\text{FWER}(\mathcal{R})=P(V\geq 1)$ & \citet{Tukey1953}\\
False discovery rate (FDR) & $\text{FDR}(\mathcal{R})=E(V/(R\vee 1))$ & \citet{benjamini1995controlling}\\
Bayesian false discovery rate (Fdr) & $\text{Fdr}(\mathcal{R})=E(V/R\big| R>0)$ & \citet{storey2003positive}\\ \hline
\end{tabular}
\end{subtable}%
\\
\begin{subtable}{\linewidth}
\centering
\caption{Different criteria for measuring the amount of true positives in multiple testing scenario.}
\begin{tabular}{c|c|c}
Criteria & Mathematical Definitions & References \\ \hline
Power & $\beta(\mathcal{R},\pmb{\mu})=P(\pmb{z}\in \mathcal{R} \big| \mathcal{H}_1,\pmb{\mu})$ & \citet{Neyman1933}\\
Bayesian power & $\eta(\mathcal{R})=P(\pmb{z}\in \mathcal{R} \big| \mathcal{H}_1)$ & \citet{kruschke2010believe}\\\hline
\end{tabular}
\end{subtable}
\end{table}
In addition to controlling false positives, we also need a criterion to measure the amount of true positives when evaluating a rejection region. A direct concept is power. The classical definition of power is a function of a given effect size as shown in the first row of Table \ref{criteria}(b). Since effect sizes of associated SNPs are different and unobserved, the actual power values are unknown. The Bayesian power removes the dependence of power on effect size by taking the expectation of the empirical power, which is defined as $S/m_1$ ($m_1>0$). We list the definitions of the power and the Bayesian power in Table \ref{criteria} (b). In this paper, we use the Bayesian power as the criterion to measure the amount of true positives.
Both Fdr and Bayesian power are functions of the rejection region $\mathcal{R}$. For two different rejection regions with the same Fdr level, we prefer the region with higher Bayesian power because it can find more true associations without increasing the proportion of false positives in the findings. Thus, we propose a joint analysis method determining the optimal rejection region when controlling the Fdr at a certain threshold $q$, i.e.
\begin{eqnarray}
\max_{\mathcal{R}}&\ & \eta(\mathcal{R})\nonumber\\
s.t. &\ & \text{Fdr}(\mathcal{R})\leq q.\label{Optimization}
\end{eqnarray}
Here $\eta(\mathcal{R})$ denotes the Bayesian power. Actually, when controlling the Fdr at the same threshold, meta-analysis methods can also be regarded as the solutions to the above optimization problem with further constraint about the form of $\eta(\mathcal{R})$, i.e.
\begin{eqnarray}
\max_{\mathcal{R}_C}&\ & \eta(\mathcal{R}_C)\nonumber\\
s.t. &\ & \text{Fdr}(\mathcal{R}_C)\leq q \nonumber \\
&\ & \mathcal{R}_C= \{\pmb{z}\big| |g(\pmb{\alpha},\pmb{z})|\geq C\}.\label{MetaOptimization}
\end{eqnarray}
Here $\pmb{\alpha}=(\sqrt{n^{(1)}},\sqrt{n^{(2)}})^T$, and $g$ is a function which has different forms in different meta-analysis methods. We will give the explicit forms of the function $g$ in meta-analysis methods in subsection \ref{relation}. Also, we will discuss the relationship between our proposed method and meta-analysis methods in detail in that subsection. In the next subsection, we will present the solution to the optimization problem in Eq. (\ref{Optimization}).
\subsection{Jlfdr and optimal rejection region}
To derive the solution to the optimization problem in Eq. (\ref{Optimization}), we need to introduce the concept of joint local false discovery rate (Jlfdr) first. Jlfdr is a simple extension of the local false discovery rate \citep{efron2005local} from the analysis of single study to the joint analysis of multiple studies. It reads as
\begin{equation}
\text{Jlfdr}(\mathbf{z})=P(\mathcal{H}_0 \big| \mathbf{z}),
\end{equation}
which is the posterior probability of a null hypothesis, given the observed summary statistic vector $\mathbf{z}$.
The relationship between Jlfdr and Fdr is (see the Supplementary Note for details)
\begin{equation}
\text{Fdr}(\mathcal{R})=E(\text{Jlfdr}(\mathbf{z})\big| \mathbf{z}\in \mathcal{R}).\label{rel}
\end{equation}
In other words, Fdr is the expectation of Jlfdr, given that the test statistic vector is in the rejection region $\mathcal{R}$.
Let us define a rejection region $\mathcal{R_O}=\{\mathbf{z} \big| \text{Jlfdr}(\mathbf{z})\leq t(q)\}$, where $t(q)$ is a threshold such that $\text{Fdr}(\mathcal{R_O})=q$. We have the following theorem:
\begin{them}\label{them1}
For any rejection region $\mathcal{R}$ with $\text{Fdr}(\mathcal{R})\leq q$, we have $\eta(\mathcal{R}) \leq \eta(\mathcal{R_O})$.
\end{them}
We show the proof of Theorem \ref{them1} in the Supplementary Note. Theorem \ref{them1} shows that $\mathcal{R_O}$ is the most powerful rejection region when controlling Fdr at $q$. This gives us a clue that we can improve the power of summary-statistics-based joint analysis by controlling the Jlfdr. In the next section, we shall provide details of the implementation of the Jlfdr-based method under the Gaussian mixture model.
\subsection{Implementation of Jlfdr-based method under the Gaussian mixture model}
The hints of hidden heritability \citep{gibson2010hints, yang2010common} and the estimated distribution of common SNPs' effect sizes \citep{park2010estimation} suggest that thousands of common SNPs with small effect sizes are associated with complex diseases. A natural prior to depict this ``infinitesimal model'' \citep{gibson2012rare} is Gaussian distribution with mean $0$ and variance $\sigma_0^2$. We assume the effect sizes of associated SNPs have this prior distribution. Since we don't know which SNP is associated with diseases, we propose the following two-component mixture model to describe the prior distribution of effect sizes:
\begin{equation}
\mu\sim \pi_0 \delta_0+(1-\pi_0) N(0, \sigma_0^2),
\end{equation}
where $\delta_0$ is the unit point mass distribution at zero.
There may be some heterogeneity in different studies. The effect sizes of the same SNP may vary in different studies. We assume the effect sizes of the same associated SNP in different studies are normally distributed with mean $\mu$ and variance $\tau \sigma_0^2$, i.e., $\mu^{(j)}\big| \mathcal{H}_1\sim N(\mu, \tau\sigma_0^2)$. The distribution of the effect size vector $\pmb{\mu}=(\mu^{(1)},\mu^{(2)})^T$ is
\begin{equation}
\pmb{\mu} \sim \pi_0 \pmb{\delta}_0+(1-\pi_0) N_2\left(\pmb{0}, \left( \begin{array}{cc}
(\tau+1)\sigma_0^2 & \sigma_0^2 \\
\sigma_0^2 & (\tau+1)\sigma_0^2 \end{array} \right) \right),\label{prior}
\end{equation}
where $\pmb{\delta}_0$ is the bivariate unit point mass distribution at the origin, and $N_2(\pmb{\eta},\Sigma)$ denotes the bivariate Gaussian distribution with expectation $\pmb{\eta}$ and covariance matrix $\Sigma$.
Since the observed effect size $\widehat{\mu}^{(j)}$ asymptotically follows Gaussian distribution $N(\mu^{(j)}, (\sigma^{(j)})^2)$, the test statistics vector $\pmb{Z}_i=(Z^{(1)}_i,Z^{(2)}_i)^T$ with prior (\ref{prior}) follows two-component Gaussian mixture distribution:
\begin{equation}
\pmb{Z}_i \sim \pi_0 N_2(\mathbf{0}, I)+(1-\pi_0) N_2(\mathbf{0}, I+\Sigma_i), \text{ where } \Sigma_i=\left( \begin{array}{cc}
(\tau+1)(\frac{\sigma_0}{\sigma^{(1)}_i})^2 & \frac{\sigma_0^2}{\sigma^{(1)}_i\sigma^{(2)}_i} \\
\frac{\sigma_0^2}{\sigma^{(1)}_i \sigma^{(2)}_i} & (\tau+1)(\frac{\sigma_0}{\sigma^{(2)}_i})^2 \end{array} \right).
\end{equation}
Here, $I$ is the identity matrix.
In order to obtain the global behavior of all SNPs, we need to obtain the marginal distribution of the test statistic vectors of all SNPs. Overall, the test statistic vector $\pmb{Z}$ follows
\begin{equation}
\pmb{Z}\sim \pi_0 N_2(\mathbf{0}, I)+ \frac{1-\pi_0}{m_1} \sum_{i\in S_1} N_2(\mathbf{0}, I+\Sigma_i). \label{gmm0}
\end{equation}
Here $S_1$ is the index set of all associated SNPs, and $m_1$ is the corresponding cardinality of $S_1$. $S_1$ is normally unknown.
The above Gaussian mixture model is computationally difficult due to the large number of components ($m_1$ is normally in the range of hundred to thousand). To simplify the model, we use a $K$-component Gaussian mixture model to approximate the non-null components, i.e.,
\begin{equation}
\frac{1-\pi_0}{m} \sum_{i=1}^m N_2(\mathbf{0}, I+\Sigma_i)\approx \sum_{k=1}^K \pi_{1k} N_2(\mathbf{0}, I+\bar{\Sigma}_k), \text{ where } \sum_{k=1}^K \pi_{1k}=1-\pi_0.
\end{equation}
Then we reduce the distribution of $\pmb{Z}$ to a $(K+1)$-component Gaussian mixture model:
\begin{equation}
\pmb{Z}\sim \pi_0 N_2(\mathbf{0}, I)+\sum_{k=1}^K \pi_{1k} N_2(\mathbf{0}, I+\bar{\Sigma}_k).\label{gmm}
\end{equation}
There are some unknown parameters $\pmb{\pi}_1=(\pi_{11},\dots,\pi_{1K})^T$ and $\pmb{\bar{\Sigma}}=\{\bar{\Sigma}_1,\dots,\bar{\Sigma}_K\}$ in the above mixture model. \cite{dempster1977maximum} proposed an EM-algorithm to estimate parameters with unobserved latent variables. With the observed vectors of summary statistics $\pmb{z}_i$ ($i=1,\dots,m$), we use the EM-algorithm to estimate the parameters $\pmb{\pi}_1$ and $\pmb{\bar{\Sigma}}$ in the Gaussian mixture model (\ref{gmm}). Please note that $\pi_0$ is always much larger than any entry of $\pmb{\pi}_1$ in the GWAS setting. Hence, a Dirichlet$(\beta_0,\pmb{0}^T)$ prior is added for the proportions $(\pi_0,\pmb{\pi_1}^T)$. This is the same penalty strategy proposed by \cite{muralidharan2010empirical}. Our experiments show that the rejection regions are not sensitive to the penalization parameter $\beta_0$ and the number of mixture components in the associated SNPs $K$. In our default setting, $\beta_0=m/5$ and $K=2$.
Denote the probability density function (pdf) of bivariate normal distribution $N_2(\mathbf{0},I)$ as $f_0(x_1,x_2)$ and the pdf of $N_2(\mathbf{0}, I+\bar{\Sigma}_k)$ as $f_1(x_1,x_2|\bar{\Sigma}_k)$. The Jlfdr reads
\begin{equation}
\text{Jlfdr}(\mathbf{z})=\frac{\pi_0 f_0(z_1, z_2)}{\pi_0 f_0(z_1, z_2)+\sum_{k=1}^K \pi_{1k} f_1(z_1, z_2|\bar{\Sigma}_k)}\label{Jlfdr}.
\end{equation}
After calculating the Jlfdr, we approximate Fdr as
\begin{equation}
\text{Fdr}(\mathcal{R})=E(\text{Jlfdr}(\mathbf{z})\big| \mathbf{z}\in \mathcal{R})\approx \frac{1}{|\{\mathbf{z}\in \mathcal{R}\}|} \sum_{\mathbf{z}\in \mathcal{R}} \text{Jlfdr}(\mathbf{z}).
\end{equation}
We determine the optimal rejection region $\mathcal{R_O}$ by Jlfdr-thresholding, which determines the rejection region with $\text{Jlfdr}(\mathbf{z})$ smaller than the threshold $t(q)$. To determine the threshold $t(q)$, we sort the calculated Jlfdr values of each SNP in an ascending order first. Denote the $a$-th Jlfdr value as $\text{Jlfdr}_a$. We can approximate the Fdr of the region $\mathcal{R}_a=\{\mathbf{z} \big| \text{Jlfdr}(\mathbf{z})\leq \text{Jlfdr}_a\}$ as
\begin{equation}
\text{Fdr}(\mathcal{R}_a)\approx \frac{1}{a} \sum_{b=1}^{a} \text{Jlfdr}_{b}.\label{Fdr}
\end{equation}
We use $c$ to denote the largest $a$ such that $\text{Fdr}(\mathcal{R}_a)\leq q$, namely
\begin{equation}
c=\max \{a\big| \text{Fdr}(\mathcal{R}_a)\leq q\}.
\end{equation}
Then the Jlfdr threshold $t(q)$ is $\text{Jlfdr}_{c}$. We reject all SNPs with $\text{Jlfdr}(\pmb{z})\leq t(q)$.
We present the detailed steps of the Jlfdr-based method in Algorithm \ref{algorithm}.
\begin{algorithm}[!htbp]
\caption{Jlfdr-based method for summary-statistics-based joint analysis}\label{algorithm}
\centering
\begin{algorithmic}
\State\hskip-\ALG@thistlm \emph{Inference using the EM-algorithm:}
\State \textbf{Initialize} $\pmb{\pi}_1$ and $\pmb{\bar{\Sigma}}$
\Repeat
\State \textbf{E Step:}
\[\pi_0^t \gets 1-\sum_{k=1}^K \pi_{1k}^t\]
\[h_{i0} \gets \frac{\pi_0^t f_0(z^{(1)}_i, z^{(2)}_i)}{\pi_0^t f_0(z^{(1)}_i, z^{(2)}_i)+\sum_{k=1}^K \pi_{1k} f_1(z^{(1)}_i, z^{(2)}_i|\Sigma_{1k}^t)},\ i=1,\cdots,m \]
\[h_{il}\gets \frac{\pi_{1l}^t f_1(z^{(1)}_i, z^{(2)}_i|\Sigma_{1l}^t)}{\pi_0^t f_0(z^{(1)}_i, z^{(2)}_i)+\sum_{k=1}^K \pi_{1k} f_1(z^{(1)}_i, z^{(2)}_i|\Sigma_{1k}^t)},\ l=1,\cdots,K \]
\State \textbf{M Step:}
\[\pi_{1l}^{t+1} \gets \frac{\sum_{i=1}^m h_{il}}{m+\beta_0} \]
\[\Sigma_{1l}^{t+1} \gets \frac{\sum_{i=1}^m h_{il}\pmb{z}_i\pmb{z}_i^T}{\sum_{i=1}^m h_{il}}-I, l=1,\cdots,K\nonumber\]
\Until{$\pmb{\pi}_1$ and $\pmb{\bar{\Sigma}}$ converge}
\State\hskip-\ALG@thistlm \emph{Jlfdr-thresholding:}
\State \textbf{Initialize} $t(q) \gets 0$
\State Calculate Jlfdr for each SNP using Eq. (\ref{Jlfdr}) with inferred $\pmb{\pi}_1$ and $\pmb{\bar{\Sigma}}$.
\State Sort calculated Jlfdr in ascending order
\For{$a \gets 1$ to $m$}
\State Calculate $\text{Fdr}(\mathcal{R}_a)$ using Eq. (\ref{Fdr}).
\If{$\text{Fdr}(\mathcal{R}_a) > q$,}
\State $t(q) \gets \text{Jlfdr}_{a-1}$; break
\EndIf
\EndFor
\State Output: the SNPs with $\text{Jlfdr}\leq t(q)$
\end{algorithmic}
\end{algorithm}
\subsection{Relationship between Jlfdr-based method and meta-analysis methods}\label{relation}
We have the following theorem about the rejection region of Jlfdr-based method when using the Gaussian mixture model:
\begin{them}\label{them2}
In the Gaussian mixture model (\ref{gmm0}), the rejection region of the Jlfdr-based method is
\begin{equation}
\mathcal{R}_1=\{\mathbf{z}\big| \sum_{i\in S_1} exp(\mathbf{z}^T(I-(I+\Sigma_i)^{-1})\mathbf{z}) \geq C_1\},
\end{equation}
where $C_1$ is a constant determined by $\text{Fdr}(\mathcal{R_O}_1)=q$. If no heterogeneity exists between studies, the rejection region is
\begin{equation}
\mathcal{R}_2=\{\mathbf{z}\big| |\pmb{\alpha}^T \pmb{z}|\geq C_2\},
\end{equation}
where $\pmb{\alpha}=(\sqrt{n^{(1)}},\sqrt{n^{(2)}})^T$, and $C_2$ is a constant determined by $\text{Fdr}(\mathcal{R_O}_2)=q$.
\end{them}
We present the proof of the above theorem in the Supplementary Note.
Meta-analysis methods are the most commonly used summary-statistics-based joint analysis methods. In meta-analysis, we usually calculate the weighted average of effect sizes in different studies. Dividing the weighted average effect size by its standard error yields a new $z$-value-based test statistic. There are two kinds of models in the meta-analysis: fixed-effects model and random-effects model.
In the fixed-effects model, we assume that the underlying true effect sizes in different studies are identical. This corresponds to $\tau=0$ in the Gaussian mixture model \ref{gmm0}. The optimal weighting strategy is the inverse-variance weighting since it minimizes the variance of the weighted average. Each effect size is weighted by the inverse of its variance, i.e.,
\begin{equation}\label{metaMu}
\widehat{\mu}_w=\frac{w_1 \widehat{\mu}^{(1)}+w_2 \widehat{\mu}^{(2)}}{w_1+w_2}, \text{ with } w_j=\frac{1}{(\sigma^{(j)})^2}, j=1,2.
\end{equation}
Here $\widehat{\mu}_{w}$ is the weighted average effect size. Its standard error is
\begin{equation}\label{metaSE}
\sigma_w=\sqrt{\frac{1}{w_1+w_2}}.
\end{equation}
Dividing $\widehat{\mu}_w$ by $\sigma_w$ yields the new test statistic $z_w$. We have the following theorem about the rejection region of the fixed-effects meta-analysis method (See the Supplementary Note for detailed proof):
\begin{them}\label{them3}
The rejection region of the fixed-effects meta-analysis method is asymptotically
\begin{equation}
\mathcal{R}_3=\{\mathbf{z}\big| |\pmb{\alpha}^T \pmb{z}|\geq C_3\},
\end{equation}
where $\pmb{\alpha}=(\sqrt{n^{(1)}},\sqrt{n^{(2)}})^T$, and $C_3$ is a constant determined by $\text{Fdr}(\mathcal{R}_3)=q$. .
\end{them}
This kind of rejection region is illustrated in Figure \ref{Figure1}. The region coincides with the rejection region of the Jlfdr-based method when no heterogeneity exists between studies. Hence, the Jlfdr-based method and the fixed-effects meta-analysis method will have the same performance. In contrast, if heterogeneity exists between studies, the rejection regions determined by the Jlfdr-based method and the fixed-effects meta-analysis method are different. According to Theorem \ref{them1}, the rejection region determined by the Jlfdr-based method can achieve the highest power among all summary-statistics-based joint analysis methods when controlling the Fdr. In other words, the Jlfdr-based method is more powerful than the fixed-effects meta-analysis method.
In the random-effects model, we assume that the true effect sizes in different studies are not identical and follow a distribution. Then we adjust the weights by incorporating the between-study variance. The weighted average effect size is
\begin{gather}
\widehat{\mu}_{w}^*=\frac{w_1^* \widehat{\mu}^{(1)}+w_2^* \widehat{\mu}^{(2)}}{w_1^*+w_2^*},\nonumber\\
\text{ with } w_j^*=\frac{1}{(\sigma^{(j)})^2+\hat{\Delta}^2}, j=1,2,\text{ and } \hat{\Delta}^2=\max \left( 0, \frac{w_1(\widehat{\mu}^{(1)}-\widehat{\mu}_w)^2+w_2(\widehat{\mu}^{(2)}-\widehat{\mu}_w)^2-1}{(w_1+w_2)-(w_1^2+w_2^2)/(w_1+w_2)} \right).
\end{gather}
Its standard error is
\begin{equation}
\sigma_w^*=\sqrt{\frac{1}{w_1^*+w_2^*}}.
\end{equation}
Dividing $\widehat{\mu}_w^*$ by $\sigma_w^*$ yields the new test statistic $z_w^*$. We have the following theorem about the rejection region of the random-effects meta-analysis method (See the Supplementary Note for detailed proof):
\begin{them}\label{them4}
The rejection region of the random-effects meta-analysis method is asymptotically
\begin{gather}
\mathcal{R}_4=\{\mathbf{z}\big| \left|\frac{\pmb{\alpha}^T V \pmb{z}}{\pmb{\alpha}^T V \pmb{\alpha}}\right|\geq C_4\} \nonumber\\
\text{ with } V=\left( \begin{array}{cc}
1/(1+\alpha_1^2 s) & 0 \\
0 & 1/(1+\alpha_2^2 s) \end{array} \right) \text{ and } s=\max \left(0, \frac{||\pmb{\alpha}||_2^2 \left[||\pmb{z}||_2^2-(\pmb{\alpha}^T \pmb{z}/||\pmb{\alpha}||_2)^2-1 \right]}{(\pmb{\alpha}^T\pmb{\alpha})^2-||\pmb{\alpha}^{o 2}||_2^2}\right).
\end{gather}
Here $\pmb{\alpha}=(\sqrt{n^{(1)}},\sqrt{n^{(2)}})^T$, $\pmb{\alpha}^{o 2}=(n^{(1)},n^{(2)})^T$ which is the 2nd Hadamard power of $\pmb{\alpha}$, and $C_4$ is a constant determined by $\text{Fdr}(\mathcal{R}_4)=q$.
\end{them}
If no heterogeneity exists between studies, the random-effects meta-analysis method is less powerful than the fixed-effects meta-analysis method and the Jlfdr-based method. If heterogeneity exists between studies, we usually need a large number of studies to estimate the between-study variance precisely in the random-effects meta-analysis. Since we usually only have a few GWASs of the same diseases in the same population, the random-effects meta-analysis is not powerful enough. The Jlfdr-based method overcomes this problem by borrowing information from all genotyped SNPs. In any case, the rejection region determined by the Jlfdr-based method and the random-effects meta-analysis method are different. According to Theorem \ref{them1}, the Jlfdr-based method is more powerful than the random-effects meta-analysis method.
\section{Results}\label{results}
\subsection{Simulation experiments}
We use simulation experiments to demonstrate that the Jlfdr-based method is more powerful than the commonly used meta-analysis methods in analyzing summary statistics from multiple GWASs.
In our simulation experiments, we fix the sample size at 10000 in study 1. We conduct experiments with different sample sizes of 5000, 10000 and 15000 in study 2. The sample size ratios $n^{(2)}/n^{(1)}$ are 0.5, 1 and 1.5 correspondingly. The individual numbers in the control group and case group are the same in both studies, and the number of SNPs is $m=1\times 10^6$. We simulate the minor allele frequency of each SNP according to uniform distribution $U(0.05, 0.5)$. The proportion of the associated SNPs is $5\%$. For associated SNPs, the expected $log$-odds ratio $\mu^{(j)}$ in each study is simulated according to the following model:
\begin{eqnarray}
\mu^{(j)}\big| \mathcal{H}_1 &\sim& N(\mu, \tau \sigma_0^2) \nonumber \\
\mu \big| \mathcal{H}_1 &\sim& N(0,\sigma_0^2),
\end{eqnarray}
where $\sigma_0^2=0.04$. In the homogeneous setting, $\tau=0$. In the heterogeneous setting, $\tau=0.5$. For non-associated SNPs, the expected $log$-odds ratio $\mu^{(j)}$ is $0$. The prevalence of the disease is $1\%$. We use the $log$-odds ratio test to detect associations in our experiments.
We use the Jlfdr-based method, the fixed-effects meta-analysis method and the random-effects meta-analysis method to jointly analyze summary statistics from study 1 and study 2. The Fdr is controlled at $q=5\times 10^{-5}$. In the fixed-effects meta-analysis and the random-effects meta-analysis, we use the one-dimensional mixture method \cite{muralidharan2010empirical} to control the Fdr at $q$.
In the homogeneous setting ($\tau=0$), each SNP shares the same expected effect size between the two studies. Figure \ref{Figure2} presents the average empirical power and the average Fdp of 10 experimental runs using different methods. The average Fdp is well controlled in all methods. In this setting, the average empirical powers are almost the same in the Jlfdr-based method and the fixed-effects meta-analysis method. The subtle differences are due to random initial choices of the EM-algorithm and the Fdr approximations used in Eq. (\ref{Fdr}). This verifies the previous statement about the equivalence between the Jlfdr-based method and the fixed-effects meta-analysis method in the homogeneous setting.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.95\textwidth]{Figure2.pdf}
\caption{\textbf{(a) The average empirical power and (b) the average Fdp in the homogeneous setting ($\tau=0$) of the simulation experiment.} The experiments are repeated 10 times with different sample size ratios ($n^{(2)}/n^{(1)}=0.5$, $1$ and $1.5$). The average Fdp of the three methods (the Jlfdr-based method (Jlfdr), the fixed-effects meta-analysis method (MetaF) and the random-effects meta-analysis method (MetaR)) are about $5\times 10^{-5}$. When controlling Fdr at the same level, the Jlfdr-based method and the fixed-effects meta-analysis method have almost the same average empirical power. The subtle differences are due to random initial choices of the EM-algorithm and the Fdr approximations used in Eq. (\ref{Fdr}). }\label{Figure2}
\end{figure}
In the heterogeneous setting ($\tau=0.5$), the expected effect sizes of each SNP vary between studies. Figure \ref{Figure3} plots the discovered associations using the Jlfdr-based method and the fixed-effects meta-analysis method in one run when $n^{(2)}=10000$. Although the Jlfdr-based method missed some associations detected by the fixed-effects meta-analysis method, it identifies more associations than the meta-analysis method. We ran the simulation experiments 10 times for the sample size ratio $n^{(2)}/n^{(1)}=0.5, 1$ and $1.5$. Figure \ref{Figure4} shows the average empirical power and the average Fdp. The average Fdp using all three methods are about $q=5\times 10^{-5}$ in all sample size ratio settings. From the figure, we can see that the Jlfdr-based method can achieve higher power than the other methods when controlling Fdr at the same threshold.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.8\textwidth]{Figure3.pdf}
\caption{\textbf{The discovered associations in the heterogeneous setting ($\tau=0.5$) of the simulation experiment.} Both the first and second studies have $10000$ individuals. For each SNP, the pair of summary statistics $(z^{(1)}, z^{(2)})$ is plotted with transformation $(|z^{(1)}|, sgn(z^{(1)})z^{(2)})$. We use light grey circles to represent the associations discovered by both the Jlfdr-based method and fixed-effects meta-analysis method. We use black upward-pointing triangles and dark grey downward-pointing triangles to represent the associations only discovered by the Jlfdr-based method and the fixed-effects meta-analysis method, respectively. The rejection boundary in the Jlfdr-based method is plotted as the solid curve. The rejection boundary in the fixed-effects meta-analysis method is plotted as the dashed straight line. The Jlfdr-based method discovered more associations overall than the meta-analysis method, although it also misses some associations identified by the meta-analysis method.}\label{Figure3}
\end{figure}
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.95\textwidth]{Figure4.pdf}
\caption{\textbf{(a) The average empirical power and (b) the average Fdp in the heterogeneous setting ($\tau=0.5$) of the simulation experiment.} We ran experiments 20 times with different sample size ratios ($n^{(2)}/n^{(1)}=0.5$, $1$ and $1.5$). The average Fdp values in three methods are about $5\times 10^{-5}$. When controlling Fdr at the same level, our proposed Jlfdr-based method can achieve higher power than the other methods in every sample size ratio setting.}\label{Figure4}
\end{figure}
\subsection{Real data applications}
\subsubsection{SCZ data from PGC}
We jointly analyze the summary statistics from schizophrenia (SCZ) studies conducted by the Psychiatric Genomics Consortium (PGC). The summary statistics from two SCZ studies, Sweden+SCZ1 \citep{ripke2013genome} and SCZ2 \citep{schizophrenia2014biological}, are available from the PGC. Sweden+SCZ1 is a large-scale meta-analysis of Swedish and mixed-European ancestry individuals that comprises 13,833 schizophrenia cases and 18,310 controls in the analysis. We use it as Study 1. SCZ2 is a larger-scale meta-analysis that comprises 36,989 schizophrenia cases and 113,075 controls. The analysis includes the individuals which have been analyzed in Sweden+SCZ1. By using the following inverse meta-analysis formula, we obtain the summary statistics from the meta-analysis comprising the individuals only be analyzed in SCZ2. The formula is
\begin{equation}
z^{(2)}=\frac{z_w/\sigma_w-z^{(1)}/\sigma^{(1)}}{1/(\sigma_w)^2-1/(\sigma^{(1)})^2}.
\end{equation}
We use $z^{(2)}$ as the summary statistics of Study 2. We remove the SNPs with $p$-value$<0.01$ in the test of homogeneity. After that, there are $m=8,157,410$ SNPs remaining.
We use the Jlfdr-based method, the fixed-effects meta-analysis method and the random-effects meta-analysis method to jointly analyze the summary statistics from two studies. The Fdr is controlled at $q=5\times 10^{-5}$. We adopt the one-dimensional mixture method to control the Fdr at $q$ in meta-analysis methods.
Figure \ref{Figure5}(a) plots the discovered associations using the Jlfdr-based method and the fixed-effects meta-analysis method. The Jlfdr-based method identifies more associations. Table \ref{SCZtable} shows the numbers of discovered associations and the rejection criteria of the different analysis methods. Besides the loci discovered by meta-analysis methods, there are eight novel loci discovered by the Jlfdr-based method. Each locus is separated by at least 500 kilobases (kb) or a weak linkage disequilibrium ($r^2<0.1$). The SNPs showing the most significant association with SCZ in these novel loci are presented in Supplementary Table 1.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.9\textwidth]{Figure5.pdf}
\caption{\textbf{The rejection region determined in the empirical datasets: }(a) SCZ data from the PGC; (b) SLE data from dbGaP; (c) BMI data from the GIANT; (d) WHRadjBMI data from the GIANT. The descriptions of the three datasets are presented in the main text. For each SNP, the vector of summary statistics $(z^{(1)}, z^{(2)})$ is plotted with transformation $(|z^{(1)}|, sgn(z^{(1)})z^{(2)})$. We use light grey circles to represent the associations discovered by both the Jlfdr-based method and the fixed-effects meta-analysis method. We use black upward-pointing triangles and dark grey downward-pointing triangles to represent the associations only discovered by the Jlfdr-based method and the fixed-effects meta-analysis method, respectively. }\label{Figure5}
\end{figure}
\begin{table}[!htbp]
\centering
\caption{\textbf{The rejection criterion and the number of identified associations in SCZ data from the PGC.} $z_{MF}$ and $z_{MR}$ are the combined $z$-values in the fixed-effects meta-analysis and random-effects meta-analysis, respectively.}\label{SCZtable}
\begin{tabular}{c|c|c}
Method & Rejection Criterion & \#\{Identified SNPs\}\\ \hline
Jlfdr-based method & $\text{Jlfdr}(\pmb{z})\leq 3.206\times 10^{-4}$ & 13405\\
Fixed-effects meta-analysis & $|z_{MF}|\geq 5.273$ & 13014\\
Random-effects meta-analysis & $|z_{MR}|\geq 5.352$ & 8348\\\hline
\end{tabular}
\end{table}
\subsubsection{SLE data from dbGaP}
We conduct summary-statistics-based joint analysis in systemic lupus erythematosus (SLE) data from phs000122.v1.p1 and phs000216.v1.p1 in dbGaP \citep{mailman2007ncbi, tryka2014ncbi}. We use the study phs000122.v1.p1, in which there are 1,311 SLE cases and 3,340 controls, as Study 1, and we use the study phs000216.v1.p1, in which there are 706 cases and 353 controls, as Study 2. The individuals in the first study are all North Americans of European descent, and those in the second study are all females of European ancestry. We use the following quality control procedures for both studies:
\begin{enumerate}
\item Missing data control: The SNPs with a missing data rate larger than $1\%$ are discarded.
\item Minor allele frequency control: The SNPs with minor allele frequency less than $0.05$ in either case group or control group are discarded.
\item Hardy-Weinberg equilibrium control: In the Hardy-Weinberg equilibrium test, the SNPs with $p$-values less than $0.001$ in either case group or control group are discarded.
\item Homogeneity control: In the homogeneity test, SNPs with $p$-values less than $0.01$ are discarded.
\end{enumerate}
After the quality control steps, there are $m=195,318$ autosome SNPs remaining.
We use $q=5\times 10^{-5}$ as the Fdr threshold in all analyses. \ref{Figure5}(b) plots the associations discovered by the Jlfdr-based method and the fixed-effects meta-analysis method. The Jlfdr-based method discovers more associations than the meta-analysis methods. Table \ref{SLEtable} lists the numbers of the associations discovered using the different methods. Besides the loci discovered by meta-analysis methods, there are three novel loci discovered by the Jlfdr-based method. The loci are separated by at least 500kb or a weak linkage disequilibrium ($r^2<0.1$). The most significant associations in these novel loci can be seen in Supplementary Table 2.
\begin{table}[!htbp]
\centering
\caption{\textbf{The rejection criterion and the number of identified associations in SLE data from dbGaP.} $z_{MF}$ and $z_{MR}$ are the combined $z$-values in the fixed-effects meta-analysis and random-effects meta-analysis, respectively.}\label{SLEtable}
\begin{tabular}{c|c|c}
Method & Rejection Criterion & \#\{Identified SNPs\}\\ \hline
Jlfdr-based method & $\text{Jlfdr}(\pmb{z})\leq 5.543\times 10^{-4}$ & 106\\
Fixed-effects meta-analysis & $|z_{MF}|\geq 5.508$ & 94\\
Random-effects meta-analysis & $|z_{MR}|\geq 5.586$ & 54 \\\hline
\end{tabular}
\end{table}
\subsubsection{BMI data from GIANT}
We jointly analyze summary statistics from body mass index (BMI) studies conducted by the Genetic Investigation of ANthropometric Traits (GIANT) consortium \citep{locke2015genetic}. We use the joint GWAS and metabochip meta-analysis of 152,893 European men as Study 1, and we use the joint GWAS and metabochip meta-analysis of 171,977 European women as Study 2. There are $m=2,466,338$ autosome SNPs passing the homogeneity control ($p$-value$\geq 0.01$).
We use $q=5\times 10^{-5}$ as the Fdr threshold in all analyses. Figure \ref{Figure5}(c) plots the associations discovered by the Jlfdr-based method and the fixed-effects meta-analysis method. The Jlfdr-based method discovers more associations than meta-analysis methods. Table \ref{BMItable} shows the number of discovered associations and the corresponding rejection criterion of each method. There are six novel loci discovered by the Jlfdr-based method. The SNPs showing the most significant associations in these novel loci are listed in Supplementary Table 3.
\begin{table}[!htbp]
\centering
\caption{\textbf{The rejection criterion and the number of identified associations in BMI data from the GIANT.} $z_{MF}$ and $z_{MR}$ are the combined $z$-values in the fixed-effects meta-analysis and random-effects meta-analysis, respectively.}\label{BMItable}
\begin{tabular}{c|c|c}
Method & Rejection Criterion & \#\{Identified SNPs\}\\ \hline
Jlfdr-based method & $\text{Jlfdr}(\pmb{z})\leq 3.722\times 10^{-4}$ & 2717\\
Fixed-effects meta-analysis & $|z_{MF}|\geq 5.336$ & 2667\\
Random-effects meta-analysis & $|z_{MR}|\geq 5.383$ & 2186 \\\hline
\end{tabular}
\end{table}
\subsubsection{WHRadjBMI data from GIANT}
We conduct joint analysis in waist-to-hip ratio after adjusting for BMI (WHRadjBMI) studies from GIANT consortium \citep{shungin2015new}. We use the joint GWAS and metabochip meta-analysis of 93,480 European men as Study 1, and we use the joint GWAS and metabochip meta-analysis of 116,742 European women as Study 2. There are $m=2,127,324$ autosome SNPs passing the homogeneity control ($p$-value$\geq 0.01$).
Figure \ref{Figure5}(d) highlights the associations discovered by the Jlfdr-based method and the fixed-effects meta-analysis method. The Jlfdr-based method identifies more associations than meta-analysis methods when controlling Fdr at the same level $q=5\times 10^{-5}$. Table \ref{WHRadjBMItable} shows the number of the discovered associations and the corresponding rejection criterion of each method. Besides the loci discovered by meta-analysis methods, there are four novel loci discovered by the Jlfdr-based method. The details of the most significant SNPs in these loci are listed in Supplementary Table 4.
\begin{table}[!htbp]
\centering
\caption{\textbf{The rejection criterion and the number of identified associations in WHRadjBMI data from the GIANT.} $z_{MF}$ and $z_{MR}$ are the combined $z$-values in the fixed-effects meta-analysis and random-effects meta-analysis, respectively.}\label{WHRadjBMItable}
\begin{tabular}{c|c|c}
Method & Rejection Criterion & \#\{Identified SNPs\}\\ \hline
Jlfdr-based method & $\text{Jlfdr}(\pmb{z})\leq 5.750\times 10^{-4}$ & 452\\
Fixed-effects meta-analysis & $|z_{MF}|\geq 5.617$ & 420\\
Random-effects meta-analysis & $|z_{MR}|\geq 5.742$ & 192 \\\hline
\end{tabular}
\end{table}
\section{Discussion}\label{discussion}
Both the Jlfdr-based method and the meta-analysis methods jointly analyze summary statistics from multiple GWASs. Meta-analysis methods collapse the test statistics of all studies into a weighted average value for each SNP, which is simpler than the Jlfdr-based method. When no heterogeneity exists between studies, the Jlfdr-based method will degenerate to the fixed-effects meta-analysis method. This can be understood by the fact that there is no information loss during the collapsing when all studies are homogeneous. When heterogeneity exists between studies, however, the Jlfdr-based method can achieve higher power than the fixed-effects meta-analysis method. This is understandable as information about heterogeneity is lost during collapse when using the meta-analysis method. Since heterogeneity widely exists in most cases, we suggest to use the Jlfdr-based method instead of meta-analysis methods to jointly analyze summary statistics from multiple GWASs.
This paper proves that the Jlfdr-based method is the most powerful summary-statistics-based joint analysis method when the underlying distribution of the test statistics is known. In reality, we only know the theoretical distribution under a null hypothesis. The distribution under alternative hypotheses is usually unknown. Hence, in the implementation of the Jlfdr-based method, we assume test statistics follow the Gaussian mixture model. Then we use the EM-algorithm to infer parameters in the mixture model. Violation of the model assumptions and inaccuracy of parameters estimation will decrease the performance of the Jlfdr-based method.
We assume an independence between SNPs in the Gaussian mixture model. However, correlations between nearby SNPs often exist, which is known as linkage disequilibrium. We may further improve the Jlfdr-based method by taking advantage of the dependency information between SNPs.
\section{Conclusion}\label{conclusion}
Jointly analyzing data sets from multiple GWASs is a common strategy to discover associations. Since it is usually difficult to access individual-level genotyping data, summary-statistics-based joint analysis has become popular for jointly analyzing data sets from multiple GWASs. Among different summary-statistics-based joint analysis methods, we prefer the method with higher Bayesian power when Fdr is controlled at the same level, because it can discover more associations. With this criterion, we propose the Jlfdr-based method. It is the most powerful summary-statistics-based method. Simulation and empirical experiments demonstrate its superior performance over traditional meta-analysis methods.
\section*{Acknowledgement}
This paper was partially supported by a grant under the Theme-based Research Scheme (project T12-402/13N) of the Hong Kong Research Grant Council (RGC).
|
\section{Introduction}
\input{introduction}
\section{Multimodal Machine Translation}
\label{sec:multimt}
This task consists in translating an English sentence that describes an image into German, given the English sentence itself and the image that it describes.
\subsection{Phrase-based System}
\label{sec:pbsmt}
\input{pbsmt}
\subsection{Neural MT System}
\label{sec:nmt}
\input{nmt}
\subsubsection{Training}
\label{sec:task1_train}
\input{task1_train}
\subsection{Data}
\label{sec:task1_data}
\input{task1_data}
\subsection{Results and Analysis}
\label{sec:task1_results}
\input{task1_res}
\section{Multimodal Image Description Generation}
\label{sec:multicaption}
The objective of Task 2 is to produce German descriptions of images given the image itself and one or more English descriptions as input.
\subsection{Visual Data Representation}
\label{sec:visual_repr}
\input{visual_data}
\subsection{Multimodal NMT System}
\label{sec:mnmt}
\input{mnmt}
\subsubsection{Generation}
\label{sec:task2_train}
Since we are provided 5 source descriptions for each image in order to generate a single German description, we let the NMT generate a German description for each source and pick the one with the highest probability and preferably without an UNK token.
\subsection{Data}
\label{sec:task2_data}
\input{task2_data}
\subsection{Results and Analysis}
\label{sec:task2_results}
\input{task2_res}
\section{Human multimodal description}
\label{sec:human}
\input{human}
\section{Conclusion and Discussion}
\input{conclusion}
\section*{Acknowledgments}
This work was supported by the Chist-ERA project M2CR\footnote{\tt m2cr.univ-lemans.fr}.
We kindly thank KyungHyun Cho and Orhan Firat for providing the DL4MT tutorial as open source and Kelvin Xu for the arctic-captions\footnote{\tt github.com/kelvinxu/arctic-captions} system.
|
\section{Introduction}
Effective actions in general contain an infinite number of higher dimensional operators whose precise coefficients require detailed understanding of their ultra-violet (UV) completion. In particular, except for low energy global symmetries and some positivity constraints~\cite{Adams:2006sv}, these coefficients are in principle arbitrary. On the other hand for effective theories associated with spontaneous symmetry breaking, it has long been known that soft theorems associated with the broken symmetries can be exploited to constrain the S-matrix, and in turn the effective action. Famous examples include Adler's zero for single U(1) Goldstone boson (GB)~\cite{Adler:1964um}, as well as its non-abelian extension~\cite{Low:2014nga}. Recently it has been shown that a class of effective field theories, including non-linear sigma models, Dirac Born-Infeld (DBI) and a special Galileon, can be completely determined through the use soft theorems~\cite{Cheung:2015ota}.
When spacetime, or both spacetime and internal symmetries are spontaneously broken, the soft-limits of GB's in general will no-longer vanish and are proportional to lower point amplitudes.\footnote{Flat space DBI action has vanishing soft limits due to the vanishing of amplitudes with odd number of external legs.} This is due to the fact that there are multiple GB's that mix under the broken symmetries. That this is true can be understood from the Ward identity of the broken generator:
\begin{equation}\label{Master}
\partial_\mu\langle J^\mu(x)\phi(x_1)\cdots\phi(x_{n{-}1})\rangle=-\sum_{i=1}^{n{-}1} \delta(x-x_i)\langle \phi(x_1)\cdots\delta\phi(x_i)\cdots\phi(x_{n{-}1})\rangle\,.
\end{equation}
If $\delta\phi$ leads to a state in the physical spectrum, then the RHS can lead to a non-vanishing result upon LSZ reduction and thus a non-vanishing soft limit. The conventional vanishing soft-pion limits simply reflect the fact that pions shift under the broken symmetry, and hence $\delta\phi$ does not lead to a physical state under infinitesimal transformations.
For broken conformal symmetry, the Goldstone modes that arise from dilatation and conformal boost are not independent, leading to a single dilaton~\cite{Low:2001bw}. This implies that the soft-dilaton limit can be non-vanishing, as the broken symmetries relate the dilaton to itself. Indeed the plurality of broken generators is reflected in the universality of the single soft dilaton behaviour. In particular expanding the $n$-pt amplitude involving one dilaton in terms of its soft momentum leads to leading and sub-leading terms that are simply proportional to the $(n{-}1)$-point amplitude~\cite{Boels:2015pta, DiVecchia:2015jaq}. In the presence of other global symmetries, the broken generators can rotate the dilaton into the new GB's and {\it vice versa}. This is a common situation for super conformal field theories on the Coulomb or Higgs branch, where both conformal and R-symmetry are broken. Consider for example $D=4$, $\mathcal{N}=4$ super Yang-Mills ({S}YM) in the Coulomb branch, where the massless scalars comprise one dilaton and 5 GB's for R-symmetry breaking SO(6)$\rightarrow$SO(5). As the broken R-symmetry generators mix the GB's and the dilaton, we will find non-vanishing soft limits. In this perspective, the Coulomb branch effective action of maximal {S}CFTs not only enjoys maximal supersymmetry but also exposes ``maximal broken symmetry".
Note that these soft theorems must be respected both in the UV where massive degrees of freedom are present, and in the infrared (IR) where they are integrated out. In this paper we verify this perturbatively by computing the one-loop effective action of $\mathcal{N}=4$ {S}YM up to six fields. This is done by considering the one-loop amplitude of maximal SYM in higher dimensions with the extra component of loop momenta identified as the mass of the massive multiplet. Expanding the integrand around the large mass limit, the integral yields the matrix element of the effective action. For non-perturbative tests, we examine the amplitudes from the instanton effective action obtained in~\cite{Bianchi:2015cta}. We have verified the validity of the new soft theorems to order $s^5$ at six points and $s^{10}$ at five points for one-loop amplitudes, where $s$ generically denotes Mandelstam invariants $s_{ij}=2k_i{{\cdot}}k_j$. While for the amplitudes generated from the one-instanon effective action~\cite{Bianchi:2015cta} are always of order $s^4$ for the scalar sector, we have confirmed the soft theorems for pure-dilaton amplitudes to nine points and for dilaton and pion mixed amplitudes up to seven points. In~\cite{Huang:2015sla,Luo:2015tat} leading and sub-leading soft theorems have also been checked against the amplitudes generated by the dilaton effective action, related to the trace anomaly in the recent study of the $a$-theorem~\cite{Schwimmer:2010za,Komargodski:2011vj,Elvang:2012st,Elvang:2012yc,Dymarsky:2013pqa,Schwimmer:2013jma,Bobev:2013vta}.
Soft theorems provide additional information on the analytic structure of scattering amplitudes, which can be combined with factorization constraints to recursively construct higher multiplicity results. Armed with the dilaton soft theorems, one can show that the matrix elements of the pure dilaton effective action are fully determined by a subset of operators via on-shell recursion~\cite{Luo:2015tat}. In particular, at $2n$-derivative order, the S-matrix for any multiplicity, {\it i.e.} any number of dilaton insertions, is completely determined in terms of operators of the form ${\partial}^{2k}\varphi^k$ for $k\leq n$. For maximal susy, the dilaton effective action for arbitrary number of dilatons are fixed up to ten derivatives in terms of three parameters: the coefficients of four-point operators at orders $s^2, s^4$ and $s^5$. For $D=4$, $\mathcal{N}=4$, we find that the dilaton amplitude at $s^{2}$ and $s^3$ are one and two-loop exact respectively for arbitrary multiplicity. At orders $s^{4}$ and $s^5$, amplitudes with arbitrary multiplicity are completely determined in terms of the four-point coefficient. Beyond $s^5$ higher point coefficients are necessary to determine the $n$-point amplitude.
Dilaton soft theorem is separated in two pieces, reflecting the fact that there are two kinds of generators being broken, scale and conformal boost. A theory endowed with only scale invariance will satisfy the leading soft theorem but not the sub-leading one. Thus the question of scale vs conformal symmetry becomes to which extent sub-leading soft theorem follows from leading. We study this question beginning with five-point amplitudes to very high order in $s$ (until $s^{11}$), and show that amplitudes satisfying the leading soft theorems automatically satisfy sub-leading soft theorem. Similar statements hold if one considers the amplitudes determined by recursion relations using the leading soft behaviour alone, for which we have verified the statements with many non-trivial examples. This can be viewed as supporting evidence for the equivalence of scale and conformal symmetry.
This paper is organized as follows: in section \ref{Sec:soft}, we give a review of soft theorems for spontaneous symmetry breaking, and show that the mixing of GB modes under the broken symmetry can lead to non-vanishing soft limits, in contrast to the usual Adler's zero. Explicit tests for the new soft theorems were conducted in subsection \ref{Sec:OneLoop} on the one-loop and \ref{section:instantoncheck} for the instanton effective action. In section \ref{Sec:Recur}, we consider to which extent the matrix element of the dilaton effective action is fixed via soft and factorization constraints. In section \ref{Sec:SUSY}, we consider further constraints from maximal supersymmetry. In section \ref{section:scalevsconformal}, we study scale vs conformal symmetry in the context of soft-theorems.
We conclude in section \ref{section:conclusion}.
\section{Soft theorems}\label{Sec:soft}
Soft behaviour of amplitudes with massless particles are often dictated by Ward identities of the underlying symmetries. Here we follow the discussion in~\cite{Weinberg}, and clarify where one departs from the usual Adler's zero. Spontaneous broken symmetry implies that the current associated with the broken generators excite GBs from the vacuum:
\begin{equation}
\langle \pi^{a}(q)|J^{b\mu}(x)|0\rangle=i f_\pi q^\mu e^{iqx} \delta^{ab}
\end{equation}
where $a,b$ label the generators. Inserting the current between a set of incoming and out going asymptotic states ($\alpha,\beta$), one finds, with $q^\mu=p^\mu_\alpha-p_\beta^\mu$
\begin{equation}\label{soft1}
\langle \alpha| J^\mu(0)|\beta\rangle=\frac{q^\mu}{q^2}A(\pi,\alpha,\beta)+N^\mu
\end{equation}
where the RHS is understood as an expansion in $q$ and we've separated out the pole term for the emission of a GB, which corresponds to fig.\ref{Fig1}(a), and $A$ is the transition amplitude.
\begin{figure}
\begin{center}
\includegraphics[scale=0.55]{Dia}
\caption{Contributions to the soft limit}
\label{Fig1}
\end{center}
\end{figure}
Contracting $q_\mu$ on both sides of eq.(\ref{soft1}), the LHS vanishes since the current is conserved:
\begin{equation}
0=\langle \alpha| \partial_\mu J^\mu(x)|\beta\rangle=\langle \alpha| \partial_\mu e^{iqx}J^\mu(0)|\beta\rangle=e^{iqx}q_\mu\langle \alpha| J^\mu(0)|\beta\rangle\,.
\end{equation}
This implies that
\begin{equation}\label{Soft2}
A(\pi,\alpha,\beta)=-q^\mu N_\mu\,.
\end{equation}
Thus in the limit where $q\rightarrow0$, the soft limit of the amplitudes involving a GB would vanish unless $q^\mu N_\mu$ is finite. This requires non-vanishing contributions from diagrams associated with fig.\ref{Fig1}(b). Note that for the latter to yield non-trivial contribution, there must be more than one massless state in the spectrum that is charged under the current, and thus form the necessary three-point vertex.\footnote{A vector current cannot couple to two identical particles.} In other words, the broken symmetry must transform a physical state to another.
The explicit form of $q^\mu N_\mu$ can be directly read off from the Ward identity:
\begin{equation}
\partial_\mu \langle J^\mu(x) \phi(x_1)\cdots\phi(x_{n{-}1})\rangle=-\sum_{i=1}^{n{-}1} \delta(x-x_i)\langle \phi(x_1) \cdots\delta\phi(x_i)\cdots\phi(x_{n{-}1})\rangle\,.
\end{equation}
Fourier transform on both sides leads to
\begin{equation}
-q_\mu\langle \tilde{J}^\mu(q) \tilde\phi(p_1)\cdots\tilde\phi(p_{n{-}1})\rangle=-\sum_{i=1}^{n{-}1}\langle \tilde\phi(p_1) \cdots\delta\tilde\phi(p_i+q)\cdots\tilde\phi(p_{n{-}1})\rangle\,,
\end{equation}
where $\tilde{\phi}$ represents Fourier transformed field. We now perform LSZ reduction on legs $1,\cdots,n{-}1$ on both sides by multiplying $\prod_{i} p_i^2$ and taking the momenta on-shell. The RHS vanishes for generic $q$, due to one uncanceled inverse propagator from the reduction. Taking the limit $q\rightarrow0$, the RHS develops the requisite inverse propagator if $\delta \tilde\phi$ yields a physical state in the spectrum. At the same time, the LHS is simply the amplitude with one soft GB.
Thus we see that if $\delta \tilde\phi$ does not correspond to another particle in the spectrum, then the RHS will not survive the LSZ reduction and hence vanishes. This is the Adler's zero for soft pion emission~\cite{Adler:1964um}. Indeed in these classical examples, the Goldstone bosons transforms non-linearly under the broken symmetry, and hence its infinitesimal transformation (a shift) does not yield a particle in the spectrum. On the other hand, if $\delta \phi$ does produce a particle in the spectrum then the RHS is non-zero, and is given by the sum of Fourier transformed amplitude with the $i$-th field transformed under the generator of the broken generator. This would be $q\cdot N$.
For broken conformal symmetry, one has the latter case. The broken dilatation symmetry constrains the leading term whilst the conformal boost generators constrain the sub-leading term in the soft momentum expansion. Thus amplitudes with single soft dilaton ($\varphi$) satisfy the following universal soft theorem~\cite{Boels:2015pta, DiVecchia:2015jaq}:
\begin{eqnarray} \label{eq:softlimitCDBI}
v A_n{\big |}_{p_n \rightarrow 0} =
\left( \mathcal{S}^{(0)}_n + \mathcal{S}^{(1)}_n \right) A_{n{-}1}+ \mathcal{O}(p_n^2) \, ,
\end{eqnarray}
where the superscript indicates the degree in $p_n$ and $v$ is the vacuum expectation value of the dilaton field. The explicit form of $\mathcal{S}^{(0)}_n, \mathcal{S}^{(1)}_n$ are given by\footnote{Note that one should replace $p_{n{-}1}$ in the $n{-}1$ point amplitude by its solution to the momentum conservation $\overline{p}_{n-1} = - (\sum^{n-2}_{i=1} p_i)$.}
\begin{eqnarray}\label{eq:softnomassS1}
\mathcal{S}^{(0)}_n &=&- \sum^{n-1}_{i=1}
\left( p_i {\cdot} {\partial \over \partial p_i } + {D- 2 \over 2} \right) + D \, ,\cr
\mathcal{S}^{(1)}_n &=& - p^{\mu}_n \sum^{n-1}_{i=1}
\left[ p^{\nu}_i { \partial^2 \over \partial {p_i^{\nu}} \partial {p_i^{\mu}} }
-
{\frac{p_{i\mu}}{2}}{ \partial^2 \over \partial {{p_i}_{\nu}} \partial {p_i^{\nu}} }
+ {D-2 \over 2} {\partial \over \partial {p^{\mu}_i} } \right] \, .
\end{eqnarray}
where $D$ is the space-time dimension.
For spontaneously broken superconformal theories, the set of massless scalars comprise the dilaton as well as the GB's for the spontaneous breaking of R-symmetry. If the dilaton is identified with one of the scalars that transforms non-trivially under the broken R-symmetry generator, following the above discussion the soft limit of the R-symmetry GB is non-vanishing. For instance, in ${\cal N} =4$ SYM, the scalars form a $\textbf{6}$ of SO(6), any one of the scalars taking a vev (say $\phi^6$) breaks R-symmetry down to SO(5), with 5 GB's associated with the broken rotation generators $R^{6I}$ with $I=1,\cdots,5$. Under this broken generator, the GB's $\phi^I$ is rotated into $\phi^6\equiv\varphi$, while $\varphi$ is rotated into $\phi^I$ with a relative minus sign due to the antisymmetry of $R^{6I}$. Thus the soft limit of R-symmetry GB's are given as:
\begin{eqnarray}\label{eq:softnomassS2}
v \, A_n(\phi_1,{\cdots}, \phi_n^I){\big |}_{p_n\rightarrow 0}=
\sum_i A_{n{-}1}({\cdots},\delta_I\phi_i,{\cdots}) + \mathcal{O}(p_n^1) \, ,
\end{eqnarray}
where $\phi_i$ represents either a dilaton $\varphi$ or $\phi^I$, with $\delta^I\varphi=\phi^{I}$ and $\delta_I\phi^J=-\delta_I{}^J \varphi$. In the following subsections we will verify the soft theorems by explicitly computations of scattering amplitudes one-loop and one-instaton effective action of $\mathcal{N}=4$ SYM in the Coulomb branch.
We should add a comment at this point. In $\mathcal{N}=4$ SYM one can define a different dilaton $\hat\varphi = \sqrt{\sum_I \phi_I^2}$ that represents the radial direction in holographic contexts and coincides with the above $\varphi = \phi_6$ (up to a sign) if the other GB's are set to zero. Moreover, the orthogonal `angular' directions of $S^5 =$ SO$(6)$/SO$(5)$ would behave as {\it bona fide} pions and satisfy Adler's theorem, since they would transform non-linearly into one another and would not mix with the radial dilaton, that is a singlet of SO$(6)$. While this is not particularly useful in the $\mathcal{N}=4$ SYM context, since it would spoil the beautiful symmetry among the various scalars, for SCFT's with lower supersymmetry, such as theories holographically dual to D3-branes at Calabi-Yau singularities (CY cones), the reduced R-symmetry would not allow such a `linear' representation of the dilaton and pions as above but only the standard non-linear one, whereby the dilaton is an R-symmetry singlet (radial direction) and the pions are the angular directions of the Sasaki-Einstein base of the CY cone.
\subsection{The one-loop verification}\label{Sec:OneLoop}
As discussed in the introduction, soft theorems hold both in the presence of the massive states and in the low energy limit where the massive states are integrated away. To verify this, we construct the one-loop effective action of $\mathcal{N}=4$ {S}YM on the Coulomb branch.\footnote{The one-loop effective action has been constructed in the constant field strength limit~\cite{Fradkin:1982kf,Chepelev:1997av}. Here we consider terms involving derivatives.} Integrands for the Coulomb branch theory can be obtained by compactifying higher-dimensional {S}YM theory, with the extra components of momenta identified with mass induced by scalar vev $v$\footnote{Obtaining spontaneously-broken SYM via a dimensional compactification was recently also studied in~\cite{Chiodaroli:2015rdg}}. We rely both on the $D=10$ {S}YM integrand constructed in~\cite{Mafra:2014gja} as well as on six-dimensional generalized unitarity methods for $(1,1)$ SYM~\cite{Dennen:2009vk, Brandhuber:2010mm} as a cross-check. At four and five points, the one-loop amplitudes of $\mathcal{N}=4$ {S}YM on the Coulomb branch are relatively simple, and have been obtained in~\cite{Chen:2015hpa},\footnote{Hereon we exploit the spinor-helicity formalism, whereby $p^{\alpha \dot{\alpha}}_i = \lambda^{\alpha}_i \tilde{\lambda}^{\dot{\alpha}}_i \,,$
and scalar products read
\begin{equation}
\lambda_i^\alpha\lambda_j^\beta\epsilon_{\alpha\beta}=\langle ij\rangle \,, \quad \tilde\lambda_{i\dot\alpha}\tilde\lambda_{j\dot\beta}\epsilon^{\dot\alpha\dot\beta}=[ij] \,, \quad
s_{ij}=\langle ij\rangle[ji] \, .
\end{equation} }
\begin{eqnarray}
\mathcal{A}_4 &=& {g^4 N} \, \delta^{8}(Q) {[12]^2 \over \langle 34\rangle^2} \times \sum_{S_4/Z_4} I_4(1,2,3,4; m) \,,
\cr
\mathcal{A}_5 &=&v {g^4 N} \, \delta^{8}(Q) {m^{(1)}_{1,2,3} m^{(2)}_{1,2,3} +
m^{(3)}_{1,2,3} m^{(4)}_{1,2,3} \over \langle 45\rangle^2}
\times \sum_{S_5/Z_5} I_5(1,2,3,4,5; m) \, ,
\end{eqnarray}
with the super charge $Q^{\alpha A}= \sum_i \lambda^{\alpha}_i \eta^A_i$. Notice that the prefactors containing fermionic $\eta$'s in both four and five points are permutation symmetric. The integrals $I_4(1,2,3,4; m)$ and $I_5(1,2,3,4,5; m)$ are scalar one-loop box and pentagon integrals with massive propagators and we sum over non-cyclic permutations, and
\begin{eqnarray}
m^{(A)}_{i,j,k} = [i \, j] \eta^A_k + [j \, k] \eta^A_i+ [k \, i] \eta^A_j \, .
\end{eqnarray}
In the above formulae the breaking of SU(4) to Sp(4) is manifest in the choice of R-symmetry indices in $m^{(A)}_{1,2,3}$, which correspond to taking the anti-symmetric $4\times 4$ Sp(4) metric to be $\Omega^{12}=-\Omega^{21}=\Omega^{34}=-\Omega^{43}=1$. In this notation, the dilaton $\Omega_{AB} \phi^{AB}$ represents fluctuations around the vev $v=m/g =\Omega_{AB} v^{AB}$. With this choice the dilaton is $\varphi = \phi^{12} + \phi^{34}$ and the other five real scalars corresponding to the pions of R-symmetry breaking are
\begin{eqnarray} \label{eq:Rpions}
\{\phi^{1}, \phi^{2}, \phi^{3}, \phi^{4}, \phi^{5}\} = \{i(\phi^{12}{-}\phi^{34}), \phi^{13}{+}\phi^{24},
i(\phi^{13}{-}\phi^{24}),\phi^{14}{+}\phi^{23}, i(\phi^{14}{-}\phi^{23})\} \, .
\end{eqnarray}
One can straightforwardly verify that five and four-point amplitudes do satisfy the soft theorems. Six-point amplitudes are more involved, we utilize the integrand of 10D YM obtained in~\cite{Mafra:2014gja} (especially equation $(5.10)$ in the reference) and campactify to 4D. In particular, to distinguish the dilaton from other five scalars, we set $\ell \cdot e_i =m$ if $e_i$ is dilaton and $\ell \cdot e_i =0$ if $e_i$ is one of the R-symmetry pions, here $\ell$ denotes the loop momentum and $e_i$ is the 10D polarization vector which becomes a scalar after compactification. We computed six-point amplitudes up to the order $s^5$ from the integrands by performing the integrals in the large-mass expansion, and checked the six-point amplitudes also obey the soft theorems. We have done the same computation by obtaining the corresponding integrand for $(1,1)$ SYM using the generalized unitary cuts. Some of the results will be summarized in what follows in the form of the effective action.
Although the SU(4) R-symmetry is broken down to Sp(4) on the Coulomb branch, the effective action can be conveniently decomposed into SU(4) singlet and non-singlet sectors. The one-loop effective action up to six field strengths reads
\begin{eqnarray}\label{OneL}
\mathcal{L}^{singlet}_{\rm 1-loop} = {g^4 N \over 32 m^4\pi^2}\left(
\mathcal{O}_{F^4} +
{ \mathcal{O}_{D^4F^4} \over 2^{3}{\times}15 m^4}
-{2 \mathcal{O}_{D^2F^6} \over 15 m^{6}} +{ \mathcal{O}_{D^4F^6} \over 2^3{\times}21 m^{8}}
- {\mathcal{O}_{D^6 F^6} \over 2{\times}15^2 m^{10}} + \mathcal{O}(m^{-12})\right) \cr
\end{eqnarray}
where $\mathcal{O}_{D^mF^n}$ represents super-local operators that contain $D^mF^n$. In the Coulomb branch $D =\partial$. Including an overall $\delta^8(Q)$, the explicit form of the superfunctions reads
\begin{eqnarray}
\nonumber&&\mathcal{O}_{F^4}:\:\delta^8(Q)\frac{[12]^2}{\langle34\rangle^2},\quad \mathcal{O}_{D^4F^4}:\:\delta^8(Q)\frac{[12]^2}{2\langle34\rangle^2}(\sum_{i<j}s^2_{ij})\,,\quad\mathcal{O}_{D^2F^6}:\:\frac{-\delta^8(Q)}{8}\sum_{S_6 /S_3{\times}S_3}\Xi^2_{123}\Xi^2_{456}\\
\nonumber&&\mathcal{O}_{D^4F^6}:\:\delta^8(Q)\sum_{S_6/Z_6}\prod_i[ii{+}1],\quad\mathcal{O}_{D^6F^6}:\:\delta^8(Q)\sum_{S_6/Z_6}(\prod_i[ii+1])s_{24}\,.
\end{eqnarray}
The Grassmann odd parameters $\eta^A$ appear in the super-polynomials
\begin{eqnarray}
\Xi^2_{123}\Xi^2_{456}=\frac{\epsilon^{ABCD}m^{(A)}_{123}m^{(B)}_{123}m^{(C)}_{456}m^{(D)}_{456}}{4!} \, .
\end{eqnarray}
For the non-singlet part, we will only list the results of scalar operators which are relevant for the soft theorems we will discuss momentarily. Note that since the SO(5)$\sim$Sp(4) subgroup of R-symmetry is preserved, the $\phi^I$ pion fields must come with even multiplicity. In the following we list the result of one-loop effective action with mixed dilaton and pions,
\begin{eqnarray}
&&\mathcal{L}^{\rm Sp(4)}_{\rm 1-loop} ={g^4 N \over 4\pi^2m^4} \left[ {{\partial}^4 \varphi^4 \over 4 }
+{{\partial}^8 \varphi^4 \over 2^{10} \times 15 m^4} +
{{\partial}^{10} \varphi^4 \over 2^{5} \times 3^2 \times 35 m^6}
+ {{\partial}^{12} \varphi^4 \over 2^{13} \times 3^3 \times 35 m^{8}} -{{\partial}^4 \varphi^5 \over m^2}
\right.
\cr
&&\qquad -{{\partial}^8 \varphi^5 \over 2^7 \times 135 m^{6}}
-{5\, {\partial}^{10} \varphi^5 \over 2^6 \times 3^4 m^{8}}
- {{\partial}^{12} \varphi^5 \over 2^6 \times 3^5 \times 35 m^{10}}
+{5 {\partial}^4 \varphi^6 \over m^4} +{{\partial}^8 \varphi^6 \over 120 m^{8}}
+{5\,{\partial}^{10} \varphi^6 \over 2^{8} 3^5 m^{10}}
\cr
&& \qquad \left. +{{\partial}^{12} \varphi^6 \over 2^{9} 3^2 m^{12}}
+ {{\partial}^4 \varphi^2 \phi^2 \over 2} - {5 {\partial}^4 \varphi^2 \phi^4\over m^2} + {{\partial}^4 \varphi^4 \phi^2 \over m^2} \right] + \ldots
\end{eqnarray}
where the on-shell matrix elements corresponding to the higher-dimensional operators are given by
\begin{eqnarray}
{\partial}^4 \varphi^m &:\:& \sum_{ i<j } s^2_{ij} \, , \quad
{\partial}^8 \varphi^4 :\: \big( s^2_{12} + \mathcal{P}_4 \big)^2 \, , \quad
{\partial}^{10} \varphi^4 :\: \big( s^5_{12} + \mathcal{P}_4 \big) \, , \quad
{\partial}^{12} \varphi^4 :\: \big( s^2_{12} + \mathcal{P}_4 \big)^3 \, , \cr
{\partial}^8 \varphi^5 &:\:& \big(s^2_{12}+ \mathcal{P}_5 \big)^2 \, , \quad
{\partial}^{10} \varphi^5 :\: {a^{(5)}_1 \over 5} + {3 \, a^{(5)}_2 \over 7} \, , \quad
{\partial}^{12} \varphi^5 :\: {a^{(6)}_1 \over 96} + a^{(6)}_2 \, , \cr
{\partial}^8 \varphi^6 &:\:& -{b^{(4)}_{1} \over 6} + {5\,b^{(4)}_{2} \over 768}
- {3 \, b^{(4)}_{3} \over 2} + {b^{(4)}_{4}\over 36} \, , \cr
{\partial}^{10} \varphi^6 &:\:&
{114 \over 35} b^{(5)}_{1} + {60 \over 7} b^{(5)}_{2}
-{48\,b^{(5)}_{3} \over 7 }
+ {108 \over 7} b^{(5)}_{4}
+{36 \over 35} b^{(5)}_{5} \, , \cr
\cr
{\partial}^{12} \varphi^6 &:\:& {433 \over 1350 } b^{(6)}_{1}- {58 \over 2025} b^{(6)}_{2}
+ {20 \over 9} b^{(6)}_{3}+ {117 \over 35} b^{(6)}_{4} - {184 \over 945} b^{(6)}_{5} , \cr
\cr
&&-{74 \over 45 } b^{(6)}_{6}+ {334 \over 315} b^{(6)}_{7}
+{73 \over 35} b^{(6)}_{8}- {64 \over 63} b^{(6)}_{9} + {104 \over 105} b^{(6)}_{10} \cr
{\partial}^4 \varphi^2 \phi^2 &:\:& s_{12}^2 -s_{13}^2- s^2_{23} \, , \quad
{\partial}^4 \varphi^2 \phi^4 :\: b^{(2)}_{1,S_2 \times S_4} - b^{(2)}_{2,S_2 \times S_4} + b^{(2)}_{3,S_2 \times S_4} - {8 \over 5} b^{(2)}_{4,S_2 \times S_4} \, ,
\cr
{\partial}^4 \varphi^4 \phi^2 &:\:& b^{(2)}_{1,S_2 \times S_4} - b^{(2)}_{2,S_2 \times S_4}
+ b^{(2)}_{3,S_2 \times S_4} + {8 } b^{(2)}_{4,S_2 \times S_4}
\end{eqnarray}
and the $b$'s are independent symmetric polynomials, they are given by
\begin{eqnarray}
a^{(5)}_1 &=& s_{12}^5 + \mathcal{P}_5 \, , \quad
a^{(5)}_2 = s_{12}^2 s_{34}^3 + \mathcal{P}_5 \, ,\quad
a^{(6)}_1 = (s_{12}^2 + \mathcal{P}_5)^3 \, , \quad
a^{(6)}_2 = s_{12}^2 s_{34}^4 + \mathcal{P}_5 \, ,
\cr
b^{(4)}_{1} &=& s_{12}^4 + \mathcal{P}_6 \, , \quad
b^{(4)}_{2} = (s_{12}^2 + \mathcal{P}_6)^2 \, , \quad
b^{(4)}_{3} = s_{12}^2 s_{13}^2 + \mathcal{P}_6 \, , \quad
b^{(4)}_{4} = s_{123}^4 + \mathcal{P}_6 \, , \cr
b^{(5)}_{1} &=& s_{12}^5 + \mathcal{P}_6 \, , \quad
b^{(5)}_{2} = s_{12}^2 s_{123}^3 + \mathcal{P}_6 \, , \cr
b^{(5)}_{3} &=& s_{12}^2 s_{23}^3 + \mathcal{P}_6 \, , \quad
b^{(5)}_{4} = s_{12}^2 s_{34}^2 + \mathcal{P}_6 \, , \quad
b^{(5)}_{5} = s_{123}^5 + \mathcal{P}_6
\label{basis}\\
b^{(6)}_{1} &=& s_{12}^6 + \mathcal{P}_6 \, , \quad
b^{(6)}_{2} = s_{123}^6 + \mathcal{P}_6 \, , \quad
b^{(6)}_{3} = s_{12}^4 s_{13}^2 + \mathcal{P}_6 \, , \cr
b^{(6)}_{4}&=& s_{12}^4 s_{34}^2 + \mathcal{P}_6 \, , \quad
b^{(6)}_{5} = s_{12}^3 s_{13}^3 + \mathcal{P}_6,\quad
b^{(6)}_{6} = s_{12}^3 s_{34}^3 + \mathcal{P}_6 \, , \cr
b^{(6)}_{7} &=& s_{12}^2 s_{123}^4 + \mathcal{P}_6 \, , \quad
b^{(6)}_{8} = s_{14}^2 s_{123}^4 + \mathcal{P}_6 \, , \quad
b^{(6)}_{9} = s_{14}^4 s_{123}^2 + \mathcal{P}_6 \, , \cr
b^{(6)}_{10} &=& s_{123}^2 s_{124}^2 s_{135}^2 + \mathcal{P}_6,\quad
b^{(2)}_{1,S_2 \times S_4} = s_{12}^2+ \mathcal{P}_{\{12|3456\}} \, , \quad
b^{(2)}_{2,S_2 \times S_4} = s_{13}^2 + \mathcal{P}_{\{12|3456\}} \, , \cr
\!\!\!\!\!\!\!\!\! b^{(2)}_{3,S_2 \times S_4} &=& s_{34}^2 + \mathcal{P}_{\{12|3456\}} \, , \quad
b^{(2)}_{4,S_2 \times S_4} = s_{12}s_{13} + \mathcal{P}_{\{12|3456\}} \, , \nonumber
\nonumber
\end{eqnarray}
here $\mathcal{P}_n$ denotes summing over permutations of $n$ elements, while $\mathcal{P}_{\{n|m\}}$ denotes summing over permutations of $n$ and $m$ elements.
\subsection{Non-perturbative checks} \label{section:instantoncheck}
Relying on (unoriented) open strings and D-brane instantons\footnote{See e.g. \cite{Bianchi:2007ft, Bianchi:2009ij, Bianchi:2012ud} for recent reviews.}, the one-instanton corrections to the effective action of $\mathcal{N}=4$ SYM in the Coulomb branch have been computed in~\cite{Bianchi:2015cta}.
For Sp(2N) the integration over (super)moduli space can be performed, and the resulting effective action can be written in a very compact and elegant form
\begin{eqnarray}
S^{1-inst}_{\rm eff} = c' \, {g^4\over \pi^6} \, e^{2\pi {\rm i} \tau } \, \int{ { d^4 x\, d^8 \theta\, \sqrt{{\rm det}_{4N} \, 2\bar\Phi_{Au,Bv} } \over
\sqrt{{\rm det}_{2N} \left( \Phi^{AB}\bar\Phi_{AB} +{1\over g} {\bar{\mathcal{F}} } +\frac{1}{\sqrt{2} g} \bar{\Lambda}_{A} (\Phi^{-1})^{AB} \bar{\Lambda}_{B} \right)_{\dot\alpha u, \dot\beta v} } } }\, ,
\label{seff0}
\end{eqnarray}
where $\tau = (\vartheta/2\pi) + (4\pi i/g^2)$ is the complexified coupling and the ${\cal N} =4$ on-shell superfields can be expanded in terms of the component fields $\{\phi^{AB}, \lambda^A_\alpha, F^-_{\alpha\beta}\}$ and their conjugate according to
\begin{align}
\bar\Phi_{AB}&=\bar\phi_{AB}+\varepsilon_{ABCD}\theta^{C\alpha}\lambda^D_{\alpha}+\frac{1}{2}\varepsilon_{ABCD}\,\theta^{C\alpha} F^{-}_{\alpha\beta}\theta^{D\beta}\\
\bar\Lambda_{\dot\alpha A}&=\bar\lambda_{\dot\alpha A}+i\,\theta^{B\alpha}\partial_{\alpha\dot\alpha}\bar\phi_{AB}+\frac{i}{2}\varepsilon_{ABCD}\theta^{B\beta}\theta^{C\gamma}\partial_{\{ \beta\dot\alpha} \lambda^D_{\gamma\}}+\frac{i}{6}\varepsilon_{ABCD}\theta^{B\alpha}\theta^{C\beta}\theta^{D\gamma}\partial_{\{\alpha\dot\alpha}F^-_{\beta\gamma\}}\\
\bar{\mathcal{F}}_{\dot\alpha\dot\beta}&=F^+_{\dot\alpha\dot\beta}-i\,\theta^{A\alpha}\partial_{\alpha\{ \dot\alpha} \bar\lambda_{A\dot\beta\}}+\frac{1}{2}\theta^{A\alpha}\theta^{B\beta}\partial_{\alpha\dot\alpha}\partial_{\beta\dot\beta}\bar\phi_{AB}+\frac{1}{6}\varepsilon_{ABCD}\theta^{A\alpha}\theta^{B\beta}\theta^{C\gamma} \partial_{\alpha\dot\alpha}\partial_{\beta\dot\beta}\lambda^D_\gamma\nonumber\\
&-\frac{1}{24}\varepsilon_{ABCD}\theta^{A\alpha}\theta^{B\beta}\theta^{C\gamma}\theta^{D\delta}\partial_{\alpha\dot\alpha}\partial_{\beta\dot\beta}F^-_{\gamma\delta}.
\end{align}
For the study of soft-dilaton and soft-pion theorems, we will turn on just the scalar fields so that
\begin{equation}
\bar\Phi_{AB}=\bar\phi_{AB} \, , \quad \bar\Lambda_{A\dot\alpha}=i\,\theta^{B\alpha}\partial_{\alpha\dot\alpha}\bar\phi_{AB}\, , \quad \bar{\mathcal{F}}_{\dot\alpha\dot\beta}=\frac{1}{2}\theta^{A\alpha}\theta^{B\beta}\partial_{\alpha\dot\alpha}\partial_{\beta\dot\beta}\bar\phi_{AB} \, ,
\end{equation}
and\footnote{$\bar\phi_{AB} = {1\over 2} \varepsilon_{ABCD}\phi^{CD}$, $\phi^2 = \sum_i \phi_i^2 = {1\over 4}\bar\phi_{AB}\phi^{AB}$}
$(\bar\Phi^{-1})^{AB}={\phi^{AB} }/{ \phi^2}$.
As a result the one-instanton effective action drastically simplifies and takes the following form
\begin{equation}
S^{1-inst}_{\rm eff} = c' {g^4 \over \pi^6} e^{2\pi i \tau} \int {d}^4x\, {d}^8\theta\,\frac{1}{1- H_{\dot\alpha\dot\beta}H^{\dot\alpha\dot\beta}}= c' {g^4 \over \pi^6} e^{2\pi i \tau} \int {d}^4x\, {d}^8\theta\, (H_{\dot\alpha\dot\beta}H^{\dot\alpha\dot\beta})^2,
\end{equation}
where
\begin{equation}
H_{\dot\alpha\dot\beta}=\frac{1}{g \, \Phi^2}\left(\bar{\mathcal{F}}_{\dot\alpha\dot\beta}+{1\over\sqrt{2}}\bar\Lambda_{A\dot\alpha}(\bar\Phi^{-1})^{AB}\bar\Lambda_{B\dot\beta}\right) \, .
\end{equation}
In the last step we have expanded the denominator and only kept the term which is non-vanishing after Grassman integration if one takes into account that the super-field $H_{\dot\alpha\dot\beta}$ becomes
\begin{equation}
H_{\dot\alpha\dot\beta}=\frac{1}{4 \, g \, \phi^2}\left(\frac{1}{2} \partial_{\alpha\dot\alpha}\partial_{\beta\dot\beta}\phi_{AB} -\frac{\phi^{CD}\partial_{\alpha\dot\alpha}\phi_{A C}\partial_{\beta\dot\beta}\phi_{DB}}{\phi^2} \right) \theta^{A\alpha}\theta^{B\beta} = K_{\alpha\dot\alpha\beta\dot\beta, AB} \theta^{A\alpha}\theta^{B\beta}\, ,
\end{equation}
when only scalars are turned on as in the case of interest here.
Switching to 4-vector indices $K_{\alpha\dot\alpha\beta\dot\beta, AB}$ may be decomposed into a symmetric traceless tensor $S^{(\mu\nu)}_{[AB]}$ in the ${\bf 6}$ of SU(4)$\sim$SO(6) and an anti-symmetric
tensor $B^{[\mu\nu]}_{(AB)}$ in the ${\bf 10^*}$ of SU(4)$\sim$SO(6). For instance, for pure dilaton sector only symmetric tensor $S^{(\mu\nu)}_{[AB]}$ contributes, and after performing the fermionic integration the action is given as,
\begin{equation}
\mathcal{S}_{\rm dilaton}=\int d^4x\,\left[ (S_{\mu\nu}S^{\mu\nu})^2-S_{\mu\nu}S^{\nu\rho}S_{\rho\sigma}S^{\sigma\mu} \right].
\end{equation}
where
\begin{equation}
S_{\mu\nu}=\frac{\partial_\mu\partial_\nu\varphi}{\varphi^2}-2\frac{\partial_\mu\varphi\partial_\nu\varphi}{\varphi^3}-\frac{1}{4}\eta_{\mu\nu}\frac{\partial^2\varphi}{\varphi^2}+\frac{1}{2}\eta_{\mu\nu}\frac{\partial_\rho\varphi\partial^\rho\varphi }{\varphi^3} \, ,
\end{equation}
and the dilaton $\varphi$ has a non-vanishing vev $\varphi \rightarrow \varphi + v.$
With the one-instanton action at hand, we have computed amplitudes up to seven points for dilaton and pion mixed amplitudes and pure-dilaton amplitudes up to nine points. We find that they indeed satisfy all the soft theorems. Here we list a few pure dilaton amplitudes\footnote{The overall coupling dependence such as $e^{2 \pi i \tau}$ is understood.}, which are degree-four symmetric polynomials in $s_{ij}$,
\begin{align} \label{inst_amps}
v^8 \mathcal{A}_4^{inst}&=\frac{1}{32} \left( s_{12}^2 + \mathcal{P}_4 \right)^2 \, , \quad
v^9 \mathcal{A}_5^{inst}=-\frac{1}{36} \left( s_{12}^2 + \mathcal{P}_5 \right)^2 \, ,\\ \nonumber
v^{10} \mathcal{A}_6^{inst}&=-\frac{2}{3} b_1^{(4)} +\frac{5}{192} b_2^{(4)}
- 6 \, b_3^{(4)} +\frac{1}{9} b_4^{(4)} \, ,\\ \nonumber
v^{11} \mathcal{A}_7^{inst}&=4 \,b_{1,7}^{(4)} + 40 \,b_{2,7}^{(4)}
-\frac{5}{3} b_{3,7}^{(4)}
-25 \,b_{4,7}^{(4)} \, , \\ \nonumber
v^{12} \mathcal{A}_8^{inst} &= -{809\over 144} b_{1,8}^{(4)} - {395\over 8} b_{2,8}^{(4)} +
{1339\over 576} b_{3,8}^{(4)} + {595\over 32} b_{4,8}^{(4)} + {535\over 32} b_{5,8}^{(4)}
\, , \\ \nonumber
v^{13} \mathcal{A}_9^{inst} &= {3935\over 294} b_{1,9}^{(4)} +{846\over 7} b_{2,9}^{(4)} -
{475\over 126} b_{3,9}^{(4)} - {491\over 14} b_{4,9}^{(4)} - {535\over 14} b_{5,9}^{(4)} \, , \nonumber
\end{align}
where the six-point amplitude $\mathcal{A}_6^{inst}$ is expanded in the basis given by eq.(\ref{basis}), while for the higher-point amplitudes
\begin{eqnarray}
b^{(4)}_{1,7} &=& s_{12}^4 + \mathcal{P}_7 \, , \quad
b^{(4)}_{2,7} = s_{12}^2s_{23}^2 + \mathcal{P}_7 \, , \quad
b^{(4)}_{3,7} = s_{123}^4 + \mathcal{P}_7 \, , \quad
b^{(4)}_{4,7} = s_{123}^2s_{124}^2 + \mathcal{P}_7 \, ,\cr
b^{(4)}_{1,8} &=& s_{12}^4 + \mathcal{P}_8 \, , \quad
b^{(4)}_{2,8} = s_{12}^2s_{23}^2 + \mathcal{P}_8 \, , \quad
b^{(4)}_{3,8} = s_{123}^4 + \mathcal{P}_8 \, , \quad
b^{(4)}_{4,8} = s_{123}^2s_{124}^2 + \mathcal{P}_8\, , \cr
b^{(4)}_{5,8} &=& s_{123}^2s_{145}^2 + \mathcal{P}_8\, ,\cr
b^{(4)}_{1,9} &=& s_{12}^4 + \mathcal{P}_9 \, , \quad
b^{(4)}_{2,9} = s_{12}^2s_{23}^2 + \mathcal{P}_9 \, , \quad
b^{(4)}_{3,9} = s_{123}^4 + \mathcal{P}_9 \, , \quad
b^{(4)}_{4,9} = s_{123}^2s_{124}^2 + \mathcal{P}_9\, , \cr
b^{(4)}_{5,9} &=& s_{123}^2s_{145}^2 + \mathcal{P}_9.
\end{eqnarray}
In appendix \ref{Appendix:oneinstanton}, we have also listed the higher-dimensional vertices that generate the above amplitudes. As we mentioned we have verified that all these amplitudes indeed satisfy the soft theorems. In fact, as we will discuss in section \ref{section:N=4SUSY}, at four, five and six points, amplitudes (with both dilatons and pions) at order $s^4$ are fully fixed by $\mathcal{N}=4$ SUSY and the soft theorems. Furthermore, for the pure-dilaton amplitudes all the higher-point amplitudes at this order are fully determined by the soft-dilaton theorems from the knowledge of the five-point amplitude, as we will discuss in the next section. Thus consistency with the conformal symmetry and $\mathcal{N}=4$ SUSY (which fixes the form of the five-point amplitude), the pure-dilaton amplitudes in fact must take the unique form given in (\ref{inst_amps}), and the same holds true for higher-point ones.
\section{Constraining the effective actions by means of soft theorems}\label{Sec:Recur}
An immediate consequence of the dilaton soft theorem is its constraint on the effective action. A systematic way to explore soft constraints is the recently constructed on-shell recursion relations~\cite{Cheung:2015ota, Luo:2015tat}. On-shell recursive methods are constructed using the fact that under complex deformation of the momenta, the only allowed singularities are propagator singularities whose residues are determined by lower point data. Using the fact that S-matrix elements are analytic functions, we start with~\cite{Britto:2004ap, Britto:2005fq}:
\begin{equation}
A_n(0)= {1 \over 2 \pi i} \oint_{{\cal C}_0} dz \frac{A_n(z)}{z}\,,
\end{equation}
where the contour ${\cal C}_0$ encircles the origin, and $A_n(z)$ is the $n$-point amplitude with deformed momenta and $A_n(0)$ is the undeformed amplitude which we would like to compute. If $A_n(z)$ is meromorphic, via the residue theorem, we can recast the amplitude as a sum over residues at finite values in the complex plane plus the one at infinity. The poles at finite values in the complex plane are simply due to factorization and their residues are determined by lower-point amplitudes. The usefulness of the recursion then relies on whether one can avoid contributions from the point at infinity or one can determine that contribution {\it a priori}. Effective theories in general do receive contributions at infinity. In~\cite{Cheung:2015ota, Luo:2015tat}, it was shown that if it is known that the amplitude has universal behaviour in some kinematic regime, then one can instead consider
\begin{equation}\label{BCFW}
A_n(0)= {1 \over 2 \pi i} \oint_{{\cal C}_0} dz \frac{A_n(z)}{z F(z)} \, ,
\end{equation}
where $F(z)$ is a polynomial in $z$ with $F(0)=1$, and its zeroes correspond to the special kinematic configurations. At large $z$, $F(z) \sim z^d$ with some positive $d$. The function $F(z)$ introduces extra power of suppression at large $z$, allowing for non-vanishing boundary contributions from $A(z)$ of higher mass dimension. The amplitude $A(0)$ is then determined by the residues of the factorisation pole as well as the contributions from the poles in $1/F(z)$ which are given by the universal behaviour of the amplitudes.
For theories with universal soft theorems, one deforms the amplitude by shifting each momentum as $p_i\rightarrow (1-z a_i) p_i$, such that $z \rightarrow 1/a_i$ one approaches the soft limit. This leads to the choice of $F(z)$ given in~\cite{Cheung:2015ota, Luo:2015tat}
\begin{equation}
F_n(z)=\prod_{i=1}^n(1-z a_i)^{d_i}\,,
\end{equation}
where $\sum_{i}a_i p_i=0$ to ensure momentum conservation for $A(z)$, and the positive integer $d_i$ depends on the soft theorem for the particle species of external leg $i$: particle $i$ enjoys universal soft theorem up order $q^{d_i-1}$ in the soft momentum ($q$) expansion. That is because otherwise $F_n(z)$ would introduce poles whose residues would be unknown.
\begin{table}[t]
\centering
\begin{tabular}{c|ccccccccccc}
$s^n$ $\backslash$ \# of points & 4 ~ & 5 ~ & 6 ~ & 7 ~ & 8 ~ & ~ $\cdots$ ~
\\ \hline
2 & $\times$ & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark
\\
3 & $\times$ & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark
\\
4 & $\times$ & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark
\\
5 & \checkmark & $\times$ & \checkmark & \checkmark & \checkmark & \checkmark
\\
6 & \checkmark & \checkmark & $\times$ & \checkmark & \checkmark & \checkmark
\\
7 & \checkmark & \checkmark & \checkmark & $\times$ & \checkmark & \checkmark
\\
8 & \checkmark & \checkmark & \checkmark & \checkmark & $\times$ & \checkmark
\\
\vdots & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$
\end{tabular}
\caption{The table is to show that the knowledge of the $k$-point amplitude at order $s^k$ with $k \leq n$ allows one to determine all the amplitudes up to the order $s^n$. The $\times$ is to indicate the amplitudes that have to compute by other means, then all other amplitudes marked with \checkmark are completely determined by the soft theorems as well as the soft-BCFW. One should note that the soft-BCFW recursion relation can only apply to amplitudes in a $D$-dimensional theory with at least $D{+}2$ external legs.
\label{tab:summary}}
\end{table}
Let us first consider the dilaton effective action. Since the dilaton soft theorem is universal up to $\mathcal{O}(q^1)$, this implies that all the $d_i$ can be mostly taken to be $2$, and thus $F_n(z)$ behaves as $z^{2n}$ in the large $z$ limit. Simple power counting shows that for an amplitude at order $s^k$, the recursion formula is valid for $A_n$ if $n>k$. Thus the pure dilaton sector is completely determined if the $n$-point amplitude at order $s^n$ is known, as the higher-point amplitudes are uniquely determined via recursion, while lower-point amplitudes can simply be obtained through leading soft theorems by taking a soft particle away. Therefore if the $k$-point amplitude at order $s^k$ with all $k \leq n$ are given, one can determine all the amplitudes up to the order $s^n$. One should also take into account that the soft BCFW recursion relation is only applicable in $D$-dimensions for at least $D{+}2$ external legs. For instance, in $D=4$, at order $s^4$, knowing the four-point amplitude is not enough to completely fix all higher-point amplitudes. Instead the five-point amplitude is required to fully determine all amplitudes at this order. This general discussion is summarized in table \ref{tab:summary}. For general superconformal theories, one can consider mixed amplitudes with $n_1$ dilatons and $n_2$ R-symmetry Goldstone bosons. Since the R-symmetry soft theorem is only leading, the requisite bound for valid recursion is $n_1+{1\over 2} n_2 > k$ for order-$s^k$ amplitudes.
In some special cases, all the terms marked with ``$\times$" in table \ref{tab:summary}, except the one of order $s^2$, may simply vanish. For instance, this is indeed the case if each scalar insertion carries at most one derivative, namely analogous to the ``constant field-strength approximation''. Thus at $2n$ or $(2n{+}1)$ points, amplitudes go as $s^n$ at most. Let us normalize the four-point amplitude at order $s^2$ as
\begin{eqnarray}
A^{(2)}_4 = c^{(2)}_4 \left( s^2 + t^2 + u^2 \right) \, ,
\end{eqnarray}
then all the amplitudes are completely determined by the soft theorems in terms of the factor $c^{(2)}_4$, which must be non-zero and positive for a non-trivial interacting theory~\cite{Adams:2006sv}. In other words, the theory in the ``constant field-strength approximation'' is uniquely fixed by soft theorems of the (broken) conformal symmetry, which turns out to be the conformal DBI, {\it i.e.} the DBI action in the AdS background, with an appropriate choice of the overall coefficient $c^{(2)}_4$. This conclusion is in the analog of the analysis of~\cite{Cheung:2014dqa}, where the usual flat-space DBI (namely DBI action in the flat-space background) are uniquely determined by the so-called enhanced soft limits. We remark that the flat-space DBI is a special limit of the conformal DBI. The scattering amplitudes in flat-space DBI are in a subset of those in conformal DBI, in particular the highest-derivative amplitudes with even number of external legs.
\section{Constraints from supersymmetry}\label{Sec:SUSY}
Supersymmetry imposes further constraint by relating coefficients of higher dimension operators with different dimensions. More precisely, one considers matrix elements of a susy invariant local operator of a given dimension and multiplicity. If one can conclude that such an operator does not exist, then the contribution of local operators must be proportional to the one produced by factorization channels. This leads to non-renormalization conditions. Indeed recently a whole set of new non-renormalization theorems have been obtained for the effective action of supersymmetric gauge and gravity theories in diverse dimensions~\cite{Wang:2015jna, Lin:2015ixa, Cordova:2015vwa, Chen:2015hpa}. Here we will consider the consequences of combining constraints from maximal SUSY and soft theorems.
\subsection{4D $\mathcal{N}=4$ supersymmetry} \label{section:N=4SUSY}
\subsubsection{Pure dilaton sectors}
Already for $\mathcal{N}=4$ {S}YM, it was shown that operators of the form $F^2_-F^{2\ell}_+$ is $\ell$-loop exact~\cite{Chen:2015hpa}, with the coefficient recursively determined by that of the four-point $F^2_-F^{2}_+$ operator, which was known to be one-loop exact~\cite{Dine:1997nq}. As already shown at one-loop order, generally SUSY invariant local operators for four and five points take the form:\footnote{Note four- and five-point amplitudes admit no factorization channels.}
\begin{eqnarray} \label{eq:45pts}
\mathcal{A}_4 &=& \delta^{8}(Q) {[12]^2 \over \langle 34\rangle^2} \sum_k P^{(k)}_4(s_{ij}) \, , \\
\mathcal{A}_5 &=&v\, \delta^{8}(Q) {m^{(1)}_{1,2,3} m^{(2)}_{1,2,3} +
m^{(3)}_{1,2,3} m^{(4)}_{1,2,3} \over \langle 45\rangle^2} \sum_k P^{(k)}_5(s_{ij})\,,
\end{eqnarray}
where $P^{(k)}_n(s_{ij})$ represents $n$-point degree-$k$ symmetric polynomials of $s_{ij}$. As well-known, for high enough $k$, the polynomials may have diverse structures. In particular $P^{(k)}_4(s_{ij})$ starts to have two independent structures at $k=6$, and $P^{(k)}_5(s_{ij})$ has two structures at $k=4$. Maximal SUSY relates purely gluonic operators $\partial^kF^n$ to operators with scalars $\partial^{k+n} \phi^n$.
As one can see, at least for four and five points, due to maximal SUSY, a degree $k$ operator in $s$ is determined by a polynomial of degree $k{-}2$ which generally has fewer degrees of freedom. This is crucial for mixed operators with both pions and dilaton, which would otherwise not even have the full permutation symmetry. As we will see, this simple fact leads to further non-renormalization theorems: the dilaton effective action up to 10 derivatives is completely determined by two unknown coefficients of the four-point operator at $s^4$ and $s^5$.
For the four-point amplitudes of order $s^2$, the four-point result is one-loop exact, namely it does not receive higher-loop and non-perturbative corrections~\cite{Dine:1997nq}:
\begin{eqnarray} \label{eq:4pts2term}
P^{(0)}_4(s_{ij}) = c^{(0)}_4(g, N) \times 1\,,
\end{eqnarray}
where the one-loop exact coefficient $c^{(0)}_4(g, N) = {g^4 N\over 32\pi^2 m^4}.$
By knowing the four-point amplitude of order $s^2$, the dilaton soft theorems allow us to determine all higher-point amplitudes at this order. Since there are no factorization contributions in the recursion, all coefficients are determined by the four-point amplitude and hence one-loop exact. One can easily see that these higher-point amplitudes are identical to that derived from DBI action in AdS background, that we dubbed conformal DBI earlier on.
For amplitudes at $s^3$, the four and five-point matrix element is simply zero due to the fact that $P^{(1)}_5(s_{ij})=0$ from momentum conservation. Thus the first non-trivial amplitude starts at six-point which is constructible via soft-dilaton recursion. The six-point amplitude receives contributions from local operator ${\partial}^6\phi^6$ as well as from factorization, which can be parametrized as
\begin{eqnarray}
A^{(3)}_6 &=&
a_1 (s_{12}^3 + \mathcal{P}_6) + a_2 (s_{123}^3 + \mathcal{P}_6)
\cr
&+&
\left({g^4 N\over 8\pi^2 m^4} \right)^2 \left(
(s_{12}^2+s_{13}^2 + s_{23}^2) {1 \over s_{123}} (s_{45}^2+s_{46}^2 + s_{56}^2) + \mathcal{P}_6 \right) \,,
\end{eqnarray}
where we have used the result of (\ref{eq:4pts2term}). The soft theorems then fix
\begin{eqnarray}
a_1 =0 \, , \quad a_2 = - \left({g^4 N\over 8\pi^2 m^4} \right)^2 \,.
\end{eqnarray}
We see that the soft theorems fix the coefficient of the local 6-point operator to be the square of that of the 4-point operator of order $s^2$. Since the latter is one-loop exact, the six-derivative operator ${\partial}^6\phi^6$ as well as the amplitude $A^{(3)}_6$ are two-loop exact. The same analysis applies to amplitudes beyond six points, and the recursion implies that order $s^3$ amplitudes for arbitrary multiplicity are two-loop exact. In terms of higher-dimensional operators, the soft theorems fix all the four and six-derivative operators (${\partial}^4 \phi^n$ and ${\partial}^6 \phi^n$) completely, and they are in fact identical to the conformal DBI.
At order $s^4$, an $n$-point amplitude receives contributions from factorization diagrams at order $s^3$ and $s^2$, as well as the contribution from the local operator ${\partial}^8 \phi^n$. As the factorization contributions are identical to those of conformal DBI, it is convenient to separate the contribution from the local operator ${\partial}^8 \phi^n$ into two parts: DBI and non-DBI.\footnote{Such separation was also used for the operator $F^8 \sim F^2_-F^6_+ + F^2_+F^6_- + F^4_-F^4_+$ in~\cite{Buchbinder:2001ui,Chen:2015hpa}, where the first two ``MHV" and ``anti-MHV" operators are three-loop exact, and coincide with DBI.} In this way, the local DBI part combining with factorization channels reproduces the amplitudes generated from conformal DBI, which is three-loop exact at order $s^4$. This separates the amplitude into two independent solutions to the soft equations. The remaining non-DBI part consists of degree-$4$ symmetric polynomials in $s_{ij}$. Again since at this order the amplitude is recursively constructible beyond four points, the non-DBI contribution is completely determined by the coefficient of the four-point operator, which is unique at this order,\footnote{$c^{(2)}_{4}(g,N)$ as well as $c^{(3)}_4(g,N)$ that will appear later at order $s^5$ have been computed at one and two-loop orders in~\cite{Bianchi:2015cta}.}
\begin{eqnarray}
P^{(2)}_4(s_{ij}) = c^{(2)}_{4}(g,N) \left( s_{12}^2 + \mathcal{P}_4 \right) \, .
\end{eqnarray}
Note that non-DBI contributions will be identical to that of the one-loop effective action (since there is no factorization at one loop) up to an overall normalization, namely $c^{(2)}_{4}(g,N)$ at one-loop order, thus we denote this part as $\mathcal{L}^{\ell{=}1}_{{\partial}^8\phi^n}$.
In summary, up to order $s^4$, the dilaton effective action is constrained by $\mathcal{N}=4$ supersymmetry as well as the soft theorems to take the form
\begin{eqnarray} \label{eq:1-loopexact}
\sum_{k \leq 8} \mathcal{L}_{{\partial}^k\phi^n} = \delta_{k, 8}\, c^{(2)}_{4}(g,N)\mathcal{L}^{\ell{=}1}_{{\partial}^8\phi^n}
+ \sum_{k \leq 8} \mathcal{L}^{\rm DBI}_{{\partial}^k\phi^n} \, ,
\end{eqnarray}
namely when $k<8$, the on-shell action is identical to conformal DBI, and at order $k=8$ the all-loop and non-perturbative action is fully determined by a single coefficient $c^{(2)}_{4}(g,N)$ of the four-point amplitude at this order.
For the amplitudes at order $s^5$ there is again a single polynomial both at four and five points, namely
\begin{eqnarray}
P^{(3)}_4(s_{ij}) = c^{(3)}_4(g,N) \times (s_{12}^3 + \mathcal{P}_4 ) \,, \quad
P^{(3)}_5(s_{ij}) = c^{(3)}_5(g,N) \times ( s_{12}^3 + \mathcal{P}_5)\, .
\end{eqnarray}
First of all, the soft theorems requires $c^{(3)}_5(g,N)=-2c^{(3)}_4(g,N)$. From soft-BCFW recursion relations, at order $s^5$ knowing the five-point amplitude allows us to fix the amplitudes of arbitrary multiplicity. At this order, the factorization contributions come from amplitudes of order $s^2$, $s^3$ as well as $s^4$. As we have argued the amplitudes of order $s^2$, $s^3$ are one and two-loop exact and coincide with conformal DBI, while order-$s^4$ amplitudes we separate into DBI and non-DBI parts. So it is again convenient to separate a DBI part from the ten-derivative operator $\partial^{10} \phi^n$, such that it combines with factorization diagrams from the amplitudes of order $s^2$, $s^3$ as well as DBI part of the order-$s^4$ amplitudes, and generates the corresponding amplitude of conformal DBI at this order, which is four-loop exact.
Let us now consider the remaining contributions, which contain the factorization terms of the non-DBI part of the order-$s^4$ amplitudes with the amplitudes at order $s^2$, as well as non-DBI part of the local operator $\partial^{10} \phi^n$. Due to the fact that amplitudes of order $s^2$ and of non-DBI part of the order-$s^4$ are both in the one-loop form, (\ref{eq:1-loopexact}), and clearly they produce the factorization parts that are in the same form as those of two-loop amplitudes. Thus these factorizations can be conveniently combined with a piece from the non-DBI part of $\partial^{10} \phi^n$ to produce the amplitudes as two-loop ones (again up to an overall factor), which we denote as $\mathcal{L}^{\ell{=}2}_{{\partial}^{10}\phi^n}$. The above analysis leads to the following compact representation for the complete $\partial^{10}$ effective action,
\begin{eqnarray}
\mathcal{L}_{{\partial}^{10}\phi^n} = c^{(3)}_{4}(g,N)\mathcal{L}^{\ell{=}1}_{{\partial}^{10}\phi^n} + c^{(0)}_{4}(g,N) \times c^{(2)}_{4}(g,N)\mathcal{L}^{\ell{=}2}_{{\partial}^{10}\phi^n}
+\mathcal{L}^{\rm DBI}_{{\partial}^{10}\phi^n} \, ,
\end{eqnarray}
again the kinematics dependences of $\mathcal{L}^{\ell{=}1}_{{\partial}^{10}\phi^n}$ and $\mathcal{L}^{\ell{=}2}_{{\partial}^{10}\phi^n}$ are completely fixed, and identical to those of the effective action at one and two loops, respectively.
Beyond order $s^5$, the symmetric polynomials $P_4(s_{ij})$ and $P_5(s_{ij})$ can in general be expressed in terms of several independent structures, which may differ at different loop orders and instanton levels. Furthermore, in order to apply the soft-BCFW recursion relations one eventually requires the knowledge of amplitudes beyond four points. For instance at order $s^6$, $P_4(s_{ij})$ and $P_5(s_{ij})$ are of order $s^4$. At this order, $P_4(s_{ij})$ still has only a unique structure, while $P_5(s_{ij})$ have two independent structures, thus there are three independent parameters which can be reduced to two using the soft theorems of going from five points to four points. The six-point amplitude can be generally expressed in terms of a local polynomial term and terms containing factorization poles which are determined by lower-point and lower-dimensional amplitudes~\footnote{At this order only four-point amplitudes of order $s^2$ and $s^5$ contributes since order-$s^3$ four-point amplitude vanishes in $\mathcal{N}=4$ SYM.}. We find that the polynomial term has $13$ independent structures, and soft theorems can fix $10$ of them in terms of those of five-point amplitudes. However, clearly six-point amplitudes should be further constrained by supersymmetry, such as the SUSY Ward identity presented in the Appendix~\ref{Appendix:Ward}. We will discuss this more in the following section of computing mixed amplitudes with both dialtons and pions.
\subsubsection{Dilaton and Pion mixed sectors}
When R-symmetry pions involved, first of all at four and five points, the amplitudes are completely determined by the pure-dilaton amplitudes via maximal supersymmetry as shown in (\ref{eq:45pts}). As for higher-point amplitudes, due to the fact that the soft-pion theorems are only leading order the constraints are slightly less powerful. As we discussed, for mixed amplitudes with $n_1$ dilatons and $n_2$ R-symmetry pions the requisite bound for valid recursion is $n_1+{1\over 2} n_2 > k$ at order $s^{k}$. Thus at order $s^3$, all the amplitudes are fully determined (and again are two-loop exact), except the six-point amplitudes with pions only. Now, $\mathcal{N}=4$ SUSY imposes further constraints that help to completely fix these amplitudes. Let us study this exceptional case in details in what follows.
As shown in details in Appendix \ref{Appendix:Ward}, we find that six-point SU(4)-violating component amplitudes must take the form,
\begin{eqnarray}
A(\phi_{12},\phi_{12},\phi_{34},\phi_{34},\phi_{34},\phi_{34}) = s_{12}^2\, P_6(s_{ij}) \,,
\end{eqnarray}
where $P_6(s_{ij})$ is symmetric polynomials with six external legs. At order $s^3$, $P_6(s_{ij})$ is of order $s^1$ and vanishes due to the momentum conservation. Thus all the six-point amplitudes (with or without dilatons) can be expressed as linear combinations of SU(4)-preserving amplitudes, such as,
\begin{eqnarray}
A(\phi_{12},\phi_{12},\phi_{12},\phi_{34},\phi_{34},\phi_{34}) \,.
\end{eqnarray}
A way of determining these amplitudes is to make an ansatz, and fix unknown parameters using soft theorems. For this particular case, the ansatz can be expressed as a factorization term with (two) four-point amplitudes $A_4(\phi_{12}, \phi_{12}, \phi_{34}, \phi_{34})$ as the residue (thus this term is two-loop exact), as well as a degree $s^3$ polynomial with $S_3 \times S_3$ symmetry which has $7$ independent structures. We find in fact in this case the soft-dilaton theorems alone are enough to determine the amplitudes, and the soft-pion theorems can serve as a consistent check. Explicitly, we find the amplitude to be given by
\begin{eqnarray}
A(\phi_{12},\phi_{12},\phi_{12},\phi_{34},\phi_{34},\phi_{34})
&=& \left( gN^2 \over 32 \pi^2 m^4 \right)^2 \left[ {s_{12}^2 \, s_{56}^2 \over s_{124}} -
{1 \over 6} \left( s_{12}^3 + s_{45}^3 \right) -
\left( s_{12}^2 s_{13} + s_{45}^2 s_{46} \right) \right. \cr
&-& \left. {1 \over 3} \left( s_{12}s_{13}s_{23} + s_{45}s_{56}s_{46} \right) \right]
+ \mathcal{P}_{\{123;456 \}} \, .
\end{eqnarray}
From this amplitude and similar ones with different R-symmetry indices, we can obtain all mixed amplitudes using the map in (\ref{eq:Rpions}). At order $s^4$, we find the same conclusion that with the help of the SUSY Ward identity one can fix all six-point amplitudes at this order, in terms of the four-point one, namely they are fully determined in terms of a single unknown coefficient $c^{(2)}_{4}(g,N)$. We then can apply soft-BCFW to determine all higher-point amplitudes, except a seven-point amplitude with six pions as well as an eight-point amplitude with eight pions. This obstruction can be understood by a simple large-$z$ counting. As we discussed in the previous section, supersymmetry should of course impose further constraints, and we believe they should eventually completely fix all amplitudes at this order in terms of the lowest-point one, especially given the fact that the pure-dilaton sector is fully determined. Similarly at order $s^5$, as far as for the constraints we have used, unlike the pure-dilaton amplitudes not all the mixed amplitudes can be determined in terms of the four-point one. As we discussed previously for pure-dilaton amplitudes at higher points, it is certainly of interest to explore systematically the SUSY Ward identity constraints, which has been very successfully applied to the ``MHV" higher-dimensional operators $F_-^2F_+^{2\ell}$ as well as SU(4)-breaking ones: $\phi^n F_-^2F_+^{2\ell}$. We will leave this investigation as a future research direction.
\subsection{6D $\mathcal{N}=(2,0)$ supersymmetry}
In $D=6$, the $\mathcal{N}=(2,0)$ theory contains a self-dual two-form and 5 scalars as its bosonic field content. It describes the theory of multiple M5-branes, and since it lacks a perturbative expansion parameter, it is a non-lagrangian theory. Moving on to the Coulomb branch provides such an expansion parameter.
On the Coulomb branch 4 of the 5 scalars are R-symmetry Goldstone bosons of SO(5) $\rightarrow$ SO(4) and the remaining one is the dilaton. The generators of SO(4) $\sim$ SU(2) $\times$ SU(2) are conveniently represented using a pair of Grassmann odd variables $(\eta_a,\tilde{\eta}_{a})$, with $a=1,2$ being a chiral spinor index of the SU(2) subgroup of the little group\footnote{Not to be confused with the SO(4) residual R-symmetry.} SO(4) $\subset$ SO(5,1):
\begin{equation}
\{J^+,J^z,J^-\}=\{\eta{\cdot}\eta,\;\;\eta{\cdot}\partial_{\eta}-1,\;\;\partial_\eta{\cdot}\partial_{\eta}\},\quad \{\tilde{J}^i\} = \{{J}^i (\eta\rightarrow \tilde{\eta})\}
\end{equation}
where the inner products are defined via the contraction of the chiral spinor index, {\it i.e.} $\eta{\cdot}\eta\equiv\eta^a\eta_a = 2 \eta^1\eta^2$. Note that the additive constant $-1$ for $J^z$ is required by the commutator $[J^+,J^-]=J^z$. When the operator $J^z$ acts on the on-shell matrix elements one finds
\begin{equation}
J^z\mathcal{A}_n\equiv\sum_{i}\left(\eta_i{\cdot}\partial_{\eta_i}-1\right)\mathcal{A}_n=0
\end{equation}
and similarly for $\tilde{J}^z$. As a result, the $n$-point amplitude turns out to be a polynomial of degree $(n,n)$ in the Grassmann variables. Following almost {\it verbatim} our discussion of $\mathcal{N}=4$ SYM in $D=4$, let us try to construct SUSY invariant local building blocks at four and five-points that are annihilated by the 16 susy operators $Q^{A+}=\lambda^{A}{\cdot}\eta$, $Q^{A-}=\lambda^{A}{\cdot}\partial_{\eta}$ and $\eta\rightarrow \tilde{\eta}$, with $A=1,\ldots 4$ a spinor index\footnote{In $D=6$ light-like momenta can be written as $P^{[AB]} = P^\mu \Gamma_\mu^{[AB]} = \epsilon^{ab}\lambda_a^A \lambda_b^B$.} of SO$(5,1)$. The susy invariant four- and five- point amplitudes read
\begin{eqnarray}
\mathcal{A}_4 &=& \delta^4(Q^+)\delta^4(\tilde Q^+) \sum_k P^{(k)}_4(s_{ij}) \, , \cr
\mathcal{A}_5 &=& \delta^4(Q^+)\delta^4(\tilde Q^+) \left({\sum_i \eta_i{\cdot}\tilde\eta_i} \right) \sum_k P^{(k)}_5(s_{ij})\, .
\end{eqnarray}
Acting with the derivative susy operators gives zero, since it generates terms that are proportional to the sum of total momentum or super-momentum, which vanish on the support of the delta functions. Thus following a similar analysis as in the D=4 case, the dilaton effective action is again completely fixed up to ten derivatives in terms of the three coefficients of the four-point operator.
\section{Scale vs Conformal symmetry} \label{section:scalevsconformal}
The relation between scale invariance and conformal invariance can also be studied for effective field theories (see e.g.~\cite{Nakayama:2013is} for a recent review). In our language, the question can be framed as follows: ``To what extent does the sub-leading soft theorem, due to broken conformal boost symmetry, follow from the leading behaviour stemming from broken dilation symmetry?'' First of all, we find that any five-point amplitude (which is a polynomial in $s_{ij}$) constrained by the leading soft theorem automatically satisfy the sub-leading soft theorem. This fact has been checked up to the very high $s^{11}$ order. For instance, at this particular order, four-point amplitudes involve two different polynomial structures, while five-point amplitudes depend on eleven parameters associated to as many independent polynomial structures. The leading soft theorem fixes two out of the eleven parameters in terms of the rest and those in the four-point amplitudes, and we find that the sub-leading soft theorem does not impose any further constraints.
More generally at higher points, according to the soft BCFW recursion relation, at order $s^{n}$, knowing the $2n$-point amplitude is enough to completely fix all the amplitudes with the same dimension by using the leading soft theorem alone, one may ask whether these amplitudes satisfy the sub-leading soft theorem automatically. Recall that from soft-BCFW recursion relations we have
\begin{eqnarray}
A_{2n+1} = {1 \over 2 \pi i} \oint_{{\cal C}_0} {dz \over z} {A_{2n+1}(z) \over F^{(1)}_{2n+1}(z)} \, .
\end{eqnarray}
From the form of the $(2n{+}1)$-point in the soft BCFW representation, it is highly non-trivial that the amplitude also satisfies the sub-leading soft theorem. At order $s^2$, all the amplitudes are simply
\begin{eqnarray}
A^{(2)}_n = c^{(2)}_n ( s_{12}^2 + \mathcal{P}_n ) \, ,
\end{eqnarray}
and $c^{(2)}_n$ for $n>4$ are determined in terms of $c^{(2)}_{4}$ via the leading soft theorem. With such $c^{(2)}_n$ satisfying the leading soft theorem, in this relatively simple case one can show that $A^{(2)}_n$ also satisfies the sub-leading soft theorem. Beyond order $s^2$, the story becomes more interesting and non-trivial. We have checked explicitly for many non-trivial examples that this is indeed the case for amplitudes at orders $s^3, s^4$ and $s^5$. Let us take $s^3$ as an example to illustrate the idea. The inputs are the five-point amplitude at order $s^3$,
\begin{eqnarray}
A^{(3)}_5 =c^{(3)}_5 ( s_{12}^3 + \mathcal{P}_5 ) \, ,
\end{eqnarray}
as well as the four-point amplitude at order $s^2$, $A^{(2)}_4$. With these inputs one can construct the six-point amplitude using both leading and sub-leading soft theorems, and find, for instance in 4D,
\begin{eqnarray}
A^{(3)}_6 &=& -c^{(3)}_5 ( s_{12}^3 + \mathcal{P}_6 ) - \left( {c^{(3)}_5 \over 2} + (c^{(2)}_4)^2 \right)( s_{123}^3 + \mathcal{P}_{6} ) \cr
&+&
(c^{(2)}_4)^2 \left( (s_{12}^2 + s_{13}^2 + s_{23}^2){1 \over s_{123}} (s_{45}^2 + s_{46}^2 + s_{56}^2) + \mathcal{P}_{6} \right) \, .
\end{eqnarray}
Now, the leading soft theorem alone allows us to determine $A^{(3)}_7$ in terms of lower-point and lower-derivative amplitudes. Explicitly, we find
\begin{eqnarray} \label{eq:A7s3}
A^{(3)}_7 =c^{(3)}_5 ( s_{12}^3 + \mathcal{P}_7 ) + \left( {c^{(3)}_5 } +3 {(c^{(2)}_4)^2 } \right)( s_{123}^3 + \mathcal{P}_7 )
-{ (c^{(2)}_4)^2 } A^{\rm fac}_7 \, ,
\end{eqnarray}
where $A_{\rm fac}$ is the factorization contribution, defined as
\begin{eqnarray}
A^{\rm fac}_7 &=& (s_{12}^2 + s_{13}^2 + s_{23}^2) {1 \over s_{123}}
\left[ s_{45}^2 + s_{46}^2 + s_{47}^2 +
s_{56}^2 + s_{57}^2 + s_{67}^2+ (s_{47} + s_{57} +s_{67})^2 \right. \cr
&+& \left.
(s_{45} + s_{46} + s_{47})^2 + (s_{45} + s_{56} + s_{57})^2 + (s_{46} + s_{56} + s_{67})^2
\right] + \mathcal{P}_7 \, .
\end{eqnarray}
It is then straightforward to verify that $A^{(3)}_7$ with particular parameters fixed by the leading soft theorem as in (\ref{eq:A7s3}) does satisfy the sub-leading soft theorem automatically. Similar construction or the use of recursion relations can be carried out for amplitudes of higher order, as we mentioned we have explicitly checked the statement up to order-$s^5$ local polynomial terms (namely due to the complication at this order, we set the factorization terms to vanish), which requires constructing the amplitudes until $10$ points using both leading and sub-leading soft theorems, and finally obtain the $11$-point amplitude using the leading soft theorem alone, and we find this amplitude does further satisfy the sub-leading soft theorem.
\section{Conclusions} \label{section:conclusion}
In this paper, we initiate the systematic study of constraints on effective actions due to soft theorems of spontaneously broken symmetries where multiple GB modes are mixed under the broken symmetry. Using the one-loop and one-instanton effective action for $\mathcal{N}=4$ SYM in the Coulomb branch, we demonstrated the validity of the dilaton soft theorems as well as that of the newly derived R-symmetry pion soft theorems, both perturbatively and non-perturbatively. We have shown that with maximal susy, the dilaton effective action is completely determined up to ten derivatives in terms of two unknown coefficients parameterising the four-point amplitude.
For CFTs which are non-Lagrangian, the dilaton effective actions are unique in the sense that the coefficients of the irrelevant operators are not functions of continuous parameter such as the coupling constant. However even with maximal SUSY, we've seen that broken and unbroken symmetries leave behind a large number of unknown coefficients. It is interesting to explore what are the other possible constraint that leads us to the unique action. An obvious possibility would be to explore the full implication of UV unitarity. At four points, this manifests itself as positivity constraint \cite{Adams:2006sv, Bellazzini:2015cra}.\footnote{Recently, using unitarity, analyticity and crossing symmetry, \cite{Bellazzini:2016xrt} shows that amplitudes that are softer than $s^2$ does not admit a non-trivial UV completion.} Needless to say that results beyond four points, while complicated, are desirable as this would be an alternative approach to gathering information on consistent CFTs.
In $D=4$, the maximal supersymmetric theory also enjoys S-duality at finite $N$. Furthermore in the large $N$ limit the UV theory on the Coulomb branch enjoys dual conformal symmetry~\cite{Drummond:2008vq}. It is thus a pressing question to understand to what extent does this input allow us to further fix the effective action. Also we have already discussed, it is important to have a better understanding of utilizing supersymmetry constraints at higher multiplicity, which would certainly reduce the independent parameters of higher-point amplitudes. In $D=3$, the massless degrees of freedom for the maximal theory are all Goldstone bosons. The eight scalars are identified as 7 Goldstone bosons from the breaking of SO(8) $R$-symmetry to SO(7), while the remaining one is the dilaton. Thus it would be interesting to explore the extent of uniqueness for its effective action when all broken and unbroken symmetry are taken into account. One may apply similar analysis to the low-energy expansion of string theories, since the string scattering amplitudes satisfy similar soft theorems, in particular the soft ``dilaton" theorems (for the closed-string dilaton)~\cite{DiVecchia:2015jaq}.
We observed and tested many highly non-trivial examples showing that amplitudes determined by recursion relations only based on the leading soft theorem satisfy the sub-leading soft theorem automatically. This observation leads to the supporting evidence that relativistic quantum field theories (under certain assumptions) with scale symmetry necessarily possess the enhanced conformal symmetry. It would certainly be interesting to study more on the possible equivalence between scale invariance and conformal invariance in the context of soft theorems.
Recently it was shown that the soft limit of Born-Infeld theory~\cite{Cachazo:2016njl}, at order $q^1$ in soft momentum is proportional to a larger theory involving the higher dimensional operators that mixes between the field strengths of Born-Infeld photons and Yang-Mills gluons. Given that so far a majority of universal soft behaviours can be explained via symmetry, it will be interesting to study if there exists a hidden symmetry in the larger theory that would dictate such universal soft limits.
\vspace{0.4cm}
\section{Acknowledgements}
We are grateful to Nima Arkani-Hamed, Clay Cordova, Paolo Di Vecchia, Francesco Fucito, Raffaele Marotta, Francisco Morales, David Poland, Yassen Stanev and Xi Yin for discussions. We would also like to thank Oliver Schlotterer and Carlos Mafra for helping to carry out the analysis on 10D SYM integrands.
Y-t.~H. and C-J. L.~are supported by MOST under the grant No.~103-2112-M-002-025-MY3.
|
\section{Introduction.}\label{intro}
\section{Introduction}
Stochastic simulation takes in input models and generates random outputs for subsequent performance analyses. The accuracy of these input model assumptions is critical to the analyses' credibility. In the conventional premise in studying stochastic simulation, these input models are conferred either through physical implication or expert opinions, or observable via input data. In this paper, we answer a converse question: Given only \emph{output} data from a stochastic system, can one infer about the input model?
The main motivation for asking this question is that, in many situations, a simulation modeler plainly may not have the luxury of direct data or knowledge about the input. The only way to gain such knowledge could be data from other sources that are at the output level. For instance, one of the authors has experienced such complication when building a simulation model for a contract fulfillment center, where service agents work on a variety of processing tasks and, despite the abundant transaction data stored in the center's IT system, there is no record on the start, completion, or service times spent by each agent on each particular task. Similarly, in clinic operations, patients often receive service in multiple phases such as initial checkup, medical tests and doctor's consultation. Patients' check-in and check-out times could be accurately noted, but the ``service" times provided by the medical staff could very well be unrecorded. Clearly, these service time distributions are needed to build a simulation model, if an analyst wants to use the model for sensitivity analysis or system optimization purposes.
The problem of inferring an input model from output data is sometimes known as \emph{model calibration}. In the simulation literature, this is often treated as a refinement process that occurs together with iterative comparisons between simulation reports and real-world output data (a task known as \emph{model validation}; \cite{sargent2005verification,kleijnen1995verification}). If simulation reports differ significantly from output data, the simulation model is re-calibrated (which can involve both the input distributions and system specifications), re-compared, and the process is iterated. Suggested approaches to compare simulation with real-world data include conducting statistical tests such as two-sample mean-difference tests (\cite{balci1982some}) and the Schruben-Turing test (\cite{schruben1980establishing}). Beyond that, inferring input from output seems to be an important problem that has not been widely discussed in the stochastic simulation literature (\cite{nelson2016some}).
The setting we consider can be briefly described as follows. We assume an input model is missing and make no particular assumptions on the form of its probability distribution. We assume, however, that a certain output random variable from a well-specified system is observable with some data. Our task is to nonparametrically infer the input distribution, or other quantities related to this input distribution (e.g., a second output measure driven by the same input distribution). One distinction between our setting and model calibration in other literature (e.g., computer experiments) is the intrinsic probabilistic structure of the system. Namely, the input and the output in stochastic simulation are represented as probability distributions, or in other words, the relation that links the observed and the to-be-calibrated objects is a (simulable) map between the spaces of distributions. Our calibration method will be designed to take such a relation into account.
Specifically, we study an optimization-based framework for model calibration, where the optimization, on a high level, entails an objective function associated with the ``input" and constraints associated with the ``output". The decision variable in this optimization is the unknown input distribution. The constraints comprise a confidence region on the the output distribution that is compiled from the observed output statistics. By expressing the region in terms of the input distributions via the simulable map, the optimization objective, which is set to be some target input quantity, will then give rise to statistically valid confidence bounds on this target. Advantageously, this approach leads to valid bounds even if the input model is \emph{non-identifiable}, i.e., there exist more than one input model that give rise to the same observable output pattern, which may occur since the simulable map is typically highly complicated. The tightness of the bounds in turn depends on the degree of non-identifiability (which also leads to a notion of \emph{identifiability gap} that we will discuss). The idea of utilizing a confidence region as the constraint is inspired by distributionally robust optimization (DRO). However, in the conventional DRO literature, the constraints (often called collectively as the uncertainty set or the ambiguity set) are constructed based on direct observation of data. On the other hand, our constraints here serve as a tool to integrate the input-output relation, in addition to the output-level statistical noise, to effectively calibrate the input model. This leads to several new methodological challenges and solution approaches.
Under this general framework, we propose a concrete optimization formulation that balances statistical validity and the required computational efforts. Specifically, we use a nonparametric statistic, namely the Kolmogorov-Smirnov (KS) statistic, to construct the output-level confidence region. This formulation has the strengths of being statistically consistent (implied by the KS statistic) and expressible as expectation-type constraints that can be effectively solved by our subsequent algorithms. It also has an interesting additional benefit in terms of controlling the dimension of the optimization. Because of computational capacity, the decision variable, which is the unknown input distribution and potentially infinite-dimensional, needs to be suitably discretized by randomly generating a finite number of support points. A consistent statistic typically induces a large number of constraints, and one may need to use a large number of support points to retain the discretization error. However, as will be seen, it turns out that the KS constraints allow us to use a moderate support size without compromising the asymptotic statistical guarantees, thanks to their low complexity as measured by the so-called bracketing number in the empirical process theory. This thus leads us to an optimization problem with both a controllable number of decision variables and statistical validity.
Next, due to the sophisticated input-output map, the optimization programs generally involve non-convex stochastic (i.e., simulation-based) constraints. We propose and analyze a stochastic quadratic penalty method, by adding a growing penalty on the squared constraint violation. This method borrows from the quadratic penalty method used in deterministic nonlinear programming. However, while the deterministic version suggests solving a nonlinear program at each particular value of the penalty coefficient and letting the coefficient grows, the stochastic method we analyze involves a stochastic approximation (SA) that runs updates of the solution, slack variables and the penalty coefficient simultaneously. This is motivated from the typical challenge of finding good stopping times for SA, which are needed for each SA run at each penalty coefficient value if one were to mimic the deterministic procedure. Simultaneous updates of all the quantities, however, only need one SA run. We analyze the convergence guarantee of this algorithm and provide guidance on the step sizes of all the constituent updates. Our SA update uses a mirror descent stochastic approximation (MDSA) (\cite{nemirovski2009robust}), in particular the entropic descent (\cite{beck2003mirror}).
The remainder of the paper is organized as follows. Section \ref{sec:lit} reviews the related literature. Section \ref{sec:formulation} introduces the problem setting and presents our general optimization-based framework. Section \ref{sec:KS} refines our framework with the KS-based formulations and demonstrates the statistical guarantees. Section \ref{sec:procedure} presents and analyzes our optimization algorithm. Section \ref{sec:numerics} reports numerical results. Section \ref{sec:discussion} concludes. The Appendix contains all the proofs.
\section{Related literature}\label{sec:lit}
We organize the literature review in two aspects, one related to the model calibration problem, and one related to our optimization approach.
\subsection{Literature Related to Our Problem Setting}
Input modeling and uncertainty quantification in the stochastic simulation focus mostly on the input level. \cite{barton2012tutorial} and \cite{song2014input}, e.g., review some major methods in quantifying the statistical errors from finite input data. These methods include the delta or two-point method (\cite{cheng1998two,cheng2004calculation}), Bayesian methodology and model averaging (\cite{chick2001input,zouaoui2004accounting}) and resampling methods (\cite{barton2001resampling,barton2013quantifying}). Our problem is more related to model calibration. In the simulation literature, this is often considered together with model validation (\cite{sargent2005verification,kleijnen1995verification}). Conventional approaches compare simulation data with real-world historical output data according to statistical or Turing tests (\cite{balci1982some,schruben1980establishing}), conduct re-calibration, and repeat the process until the data are successfully validated (\cite{banks2000dm,kelton2000simulation}).
The model calibration problem is also known as the \emph{inverse problem} (\cite{tarantola2005inverse}) in the literature of other fields. It generally refers to the identification of parameters or functions that can only be inferred from transformed outputs. In the context where the parameters are probability distributions, \cite{kraan2005probabilistic} demonstrates theoretically the characterization of a distribution that leads to the smallest relative entropy with a reference measure, and proposes an entropy maximization to calibrate the distribution from output data. Our work relates to \cite{kraan2005probabilistic} as we also utilize a probabilistic input-output map, but we focus on maps that are evaluable only by simulation, and aim to compute confidence bounds on the true distribution instead of attempting to recover the maximum entropy distribution.
The inverse problem also appeared in many other contexts. In signal processing, the linear inverse problem (e.g., \cite{csiszar1991least,donoho1992maximum}) reconstructs signals from measurements of linear transformations. Common approaches consist of least-square minimization and the use of penalty such as the entropy. In computer experiments (\cite{santner2013design}), surrogate models built on complex physical laws require the calibration of physical parameters. Such models have wide scientific applications such as weather prediction, oceanography, nuclear physics, and acoustics (e.g., \cite{wunsch1996ocean,shirangi2014history}). Bayesian and Gaussian process methodologies are commonly used (e.g., \cite{kennedy2001bayesian,currin1991bayesian}). We point out that Bayesian methods could be a potential alternative to the approach considered in this paper, but because of the nature of discrete-event systems, one might need to resort to sophisticated techniques such as approximate Bayesian computation (\cite{marjoram2003markov}). Other related literature include experimental design to optimize inference for input parameters (e.g., \cite{chick2002simulation}) and calibrating financial option prices (e.g., \cite{avellaneda2001weighted,glasserman2005large}).
Also related to our work is the body of research on inference problems in the context of queueing systems. The first stream, similar to our paper, aims at inferring the constituent probability distributions of a queueing model based on its output data, e.g., queue length or waiting time data, collected either continuously or at discrete time points. This stream of papers focuses on systems whose structures allow closed-form analyses or are amenable to analytic approximations via, for instance, the diffusion limit. The majority of them assume that the inferred distribution(s) comes from a parametric family and use maximum likelihood estimators (\cite{basawa1996maximum,pickands1997estimation,basawa2008parameter,fearnhead2004filtering,wang2006maximum,ross2007estimation,heckmuller2009reconstructing,whitt2012fitting}). Others work on nonparametric inference by exploiting specific queueing system structures (\cite{bingham1999non,hall2004nonparametric,moulines2007,feng2014estimating}). A related stream of literature studies point process approximation (see Section 4.7 of \cite{cooper1972introduction}, \cite{whitt1981approximating,whitt1982approximating}, and the references therein), based on a parametric approach and is motivated from traffic pattern modeling in communication networks. Finally, there are also a number of studies inspired by the ``queue inference engine" by \cite{larson1990queue}. But, instead of inferring the input models, many of these studies use transaction data to estimate the performance of a queueing system directly and hence do not take on the form of an inverse problem (see \cite{mandelbaum1998estimating} for a good survey of the earlier literature and \cite{frey2010queue} and its references for more recent progress). Several papers estimate both the queueing operational performance and the constituent input models (e.g., \cite{daley1998moment,kim2008new,park2011analysis}), and can be considered to belong to both this stream and the aforementioned first stream of literature.
\subsection{Literature Related to Our Methodology}
Our formulation uses ideas from robust optimization (e.g., \cite{bertsimas2011theory,ben2009robust}), which studies optimization under uncertain parameters and suggests to obtain decisions that optimize the worst-case scenarios, subject to a set of constraints on the belief/uncertainty that is often known as the ambiguity set or the uncertainty set. Of particular relevance to us is the setting of distributionally robust optimization (DRO), where the uncertainty is on the probability distribution in a stochastic optimization problem. This approach has been applied in many disciplines such as stochastic control (e.g., \cite{pjd00,doi:10.1287/moor.1120.0540,iyengar2005robust}), economics (\cite{hansen2008robustness}), finance (\cite{gx12a}), queueing control (\cite{jls10}) and dynamic pricing (\cite{ls07}). Its connection to machine learning and statistics has also been recently investigated (\cite{blanchet2016robust,shafieezadeh2015distributionally}). In the DRO literature, common choices of the uncertainty set are based on moments (\cite{delage2010distributionally,goh2010distributionally,wiesemann2014distributionally,bertsimas2005optimal,smith95,bertsimas2007semidefinite}), distances from nominal distributions (\cite{ben2013robust,bayraksan2015data,blanchet2016quantifying,esfahani2015data,gao2016distributionally}), and shape conditions (\cite{popescu2005semidefinite,lam2017tail,li2016ambiguous,hanasusanto2017ambiguous}). The literature of data-driven DRO further addresses the question of calibrating these sets, using for instance confidence regions or hypothesis testing (\cite{bertsimas2014robust}), empirical likelihood (\cite{lam2017empirical,duchi2016statistics,lam2016recovering}) and the related Wasserstein profile function (\cite{blanchet2016robust}), and Bayesian perspectives (\cite{gupta2015near}).
For DRO in the simulation context, \cite{hu2012robust} studies the computation of robust bounds under Gaussian model assumptions, \cite{gx12b,glasserman2016bounding} study distance-based constraints to address model risks, \cite{lam2013robust,lam2016serial} study asymptotic approximations of related formulations, and \cite{ghosh2015robust} studies formulations and solution techniques for DRO in quantifying simulation input uncertainty. \cite{fan2013robust,ryzhov2012ranking} study the use of robust optimization in simulation-based decision-making. Our framework in particular follows the concept in using confidence region such that the uncertainty set covers the true distribution with high probability. However, it also involves the simulation map between input and output that serves as the key in our model calibration goal.
Our optimization procedure builds on the quadratic penalty method (\cite{bertsekas1999nonlinear}), which is a deterministic nonlinear programming technique that reformulates the constraints as squared penalty and sequentially tunes the penalty coefficient to approach optimality. Different from the deterministic technique, our procedure in solving the stochastic quadratic penalty formulation sequentially update the penalty parameter simultaneously together with the solution and slack variables. This involves a specialized version of MDSA proposed by \cite{nemirovski2009robust}. \cite{nemirovski2009robust} analyzed convergence guarantees on convex programs with stochastic objectives. \cite{lanzhou2017,NIPS2017_6741} investigated convex stochastic constraints, and \cite{ghadimi2013stochastic,ghadimi2015accelerated,dang2015stochastic,ghadimi2016mini} studied related schemes for nonconvex and nonsmooth objectives. \cite{wang2008stochastic} introduced a quadratic penalty method for stochastic objectives with deterministic constraints. The particular scheme of MDSA we consider uses entropic penalty, and is known as the entropic descent algorithm (\cite{beck2003mirror}).
\section{Proposed Framework}\label{sec:formulation}
Consider a generic input variate $X$ with an input probability distribution $P_X$. We let $\mathbf X=(X_1,\ldots,X_{T})$, where $X_t\in\mathcal X$, be an i.i.d. sequence of input variates each distributed under $P_X$ over a time horizon $T$. We denote the function $h(\cdot)\in\mathbb R$ as the system logic from the input sequence $\mathbf X$ to the output $h(\mathbf X)$. We assume that $h$ is completely specified and is computable, even though it may not be writable in closed-form, i.e. we can evaluate the output given $\mathbf X$. For example, $\mathbf X$ can denote the sequence of interarrival or service times for the customers in a queue, and $h(\mathbf X)$ is an average queue length seen by the $T$ customers. Note that we can work in a more general framework where $h$ depends on both $\mathbf X$ and other independent input sequences, denoted collectively as $\mathbf W$, that possess known or observable distributions. In other words, we can have $h(\mathbf X,\mathbf W)$ as the output. Our developments can readily handle this case, but for expositional convenience we will assume the absence of these auxiliary input sequences most of the time, and will indicate the modifications of our developments in handling them at various suitable places.
Consider the situation that only $h(\mathbf X)$ can be observed via data. Let $D=\{y_1,\ldots,y_n\}$ be $n$ observations of $h(\mathbf X)$. Our task is to calibrate some quantities related to $P_X$, which we call $\psi(P_X)$. Two types of target quantities we will consider are:
\begin{enumerate}
\item Restricting $X$ to real value, we consider the distribution function of $P_X$, denoted $F_X(x)$, where $x$ can take a range of values. Note that, obviously, $F_X(x)=E_{P_X}[I(X\leq x)]$ where $E_{P_X}[\cdot]$ denotes the expectation with respect to $P_X$ and $I(\cdot)$ denotes the indicator function.\label{case1}
\item We consider a performance measure $E_{P_X}[g(\mathbf X)]$ where $E_{P_X}[\cdot]$ here denotes the expectation with respect to the product measure induced by the i.i.d. sequence $\mathbf X=(X_1,\ldots,X_{S})$ over a time horizon $S$. The function $g(\mathbf X)$ can denote another output of interest different from $h(\mathbf X)$ that is unobservable, and requires information about $\mathbf X$. This case includes the first target quantity above (by choosing $g(\mathbf X)=I(X_1\leq x)$ when $X$ is real-valued), as well as other statistics of $X$ such as power moments (by choosing $g(\mathbf X)=X_1^k$ for some $k$).\label{case2}
\end{enumerate}
To describe our framework, we denote $P_Y=P_{h(\mathbf X)}$ as the probability distribution of the output $Y=h(\mathbf X)$. Since $P_Y$ is completely identified by $P_X$, we can view $P_Y$ as a transformation of $P_X$, i.e., $P_Y=\gamma(P_X)$ for some map $\gamma$ between probability distributions. We denote $\mathcal P_X$ and $\mathcal P_Y$ as the spaces of all possible input and output distributions respectively.
On an abstract level, we use the optimization formulations
\begin{equation}
\begin{array}{ll}
\max&\psi(P_X)\\
\text{subject to}&P_Y\in\mathcal U
\end{array}\label{RO max}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\min&\psi(P_X)\\
\text{subject to}&P_Y\in\mathcal U
\end{array}\label{RO min}
\end{equation}
where the decision variable is the unknown $P_X\in\mathcal P_X$, and $\mathcal U\subset\mathcal P_Y$ is an ``uncertainty set" that covers a set of possibilities for $P_Y$. The objective function $\psi(P_X)$ refers to either $F_X(x)$ in case \ref{case1} or $E_{P_X}[g(\mathbf X)]$ in case \ref{case2} above.
An important element in formulations \eqref{RO max} and \eqref{RO min} is that the constraints represented by $\mathcal U$ are cast on the output level. Since we have available output data, $\mathcal U$ can be constructed using these observations in a statistically valid manner (e.g., by using the confidence region on the output statistic). By expressing $P_Y=\gamma(P_X)$, the region $\mathcal U$ can be viewed as a region on $P_X$, given by $\{P_X\in\mathcal P_X:\gamma(P_X)\in\mathcal U\}$. The following result summarizes the confidence guarantee for the optimal values of \eqref{RO max} and \eqref{RO min} in bounding $\psi(P_X)$ when $\mathcal U$ is chosen suitably:
\begin{proposition}
Let $P_X^0\in\mathcal P_X$ and $P_Y^0\in\mathcal P_Y$ be the true input and output distributions. Suppose $\mathcal U$ is a $(1-\alpha)$-level confidence region for $P_Y^0$, i.e.,
\begin{equation}
\mathbb P_D(P_Y^0\in\mathcal U)\geq1-\alpha\label{confidence setup}
\end{equation}
where $\mathbb P_D(\cdot)$ denotes the probability with respect to the data $D$. Let $\overline Z$ and $\underline Z$ be the optimal values of \eqref{RO max} and \eqref{RO min} respectively. Then we have
$$\mathbb P_D(\underline Z\leq\psi(P_X^0)\leq\overline Z)\geq1-\alpha$$
Similar statements hold if the confidence is approximate, i.e., if
$$\liminf_{n\to\infty}\mathbb P_D(P_Y^0\in\mathcal U)\geq1-\alpha$$
then
$$\liminf_{n\to\infty}\mathbb P_D(\underline Z\leq\psi(P_X^0)\leq\overline Z)\geq1-\alpha$$
\label{basic guarantee}
\end{proposition}
It is worth pointing out that the same guarantee holds, without any statistical adjustment, if one solves \eqref{RO max} and \eqref{RO min} simultaneously for different $\psi(\cdot)$, say $\psi_l(\cdot),l=1,\ldots,L$, i.e., supposing that \eqref{confidence setup} holds, then the confidence statement
$$\mathbb P_D(\underline Z_l\leq\psi_l(P_X^0)\leq\overline Z_l,\ l=1,\ldots,L)\geq1-\alpha$$
holds, so does a similar statement for the limiting counterpart. We provide this extended version of Proposition \ref{basic guarantee} in the appendix (Proposition \ref{extended guarantee}). This allows us to obtain bounds for multiple quantities about the input model at the same time. Note that, in conventional statistical methods, simultaneous estimation like this sort often requires Bonferroni correction or more advanced techniques, but these are not needed in our approach.
We mention an important feature of our framework related to the issue of \emph{non-identifiability} (e.g., \cite{tarantola2005inverse}). When there are more than one input model $P_X$ that leads to the same output distribution, it is statistically impossible to recover exactly the true $P_X$, and methods that attempt to do so may result in ill-posed problems. Our framework, however, gets around this issue by focusing on computing bounds instead of full model recovery. Even though $P_X$ can be non-identifiable, our optimization always produces valid bounds for it. One special case of interest is when we use $\mathcal U=\{P_Y^0\}$, i.e., the true output distribution is exactly known. In this case, \eqref{RO max} and \eqref{RO min} will provide the best bounds for $\psi(P_X)$ given the output. If $\underline Z<\overline Z$, then $P_X$ cannot be exactly identified, implying an issue of non-identifiabilty, but our outputs would still be valid. In fact, the difference $\overline Z-\underline Z$ can be viewed as an \emph{identifiability gap} with respect to $\psi$.
\section{Kolmogorov-Smirnov-based Constraints}\label{sec:KS}
We will now choose a specific $\mathcal U$ that is statistically consistent on the output level, i.e., $\mathcal U$ shrinks to $\{P_Y^0\}$ as $n\to\infty$ (in a suitable sense). In particular, we use $\mathcal U$ implied by the Kolmogorov-Smirnov (KS) statistic, and discuss how this choice enjoys benefits balancing statistical consistency and computation.
\subsection{Statistical Confidence Guarantee}\label{sec:KS confidence}
It is known that the empirical distribution for continuous i.i.d. data $D$, denoted $\hat F_Y(y)$, satisfies $\sqrt n\|\hat F_Y-F_Y^0\|_\infty\Rightarrow\sup_{u\in[0,1]}BB(u)$ where $F_Y^0$ is the true distribution function of $Y$, $\|\cdot\|_\infty$ denotes the sup norm over $\mathbb R$, $BB(\cdot)$ is a standard Brownian bridge, and $\Rightarrow$ denotes weak convergence. This implies that the KS-statistic $\sqrt n\|\hat F_Y-F_Y^0\|_\infty$ satisfies
$$\lim_{n\to\infty}P\left(\|\hat F_Y-F_Y^0\|_\infty\leq\frac{q_{1-\alpha}}{\sqrt n}\right)=1-\alpha$$
where $q_{1-\alpha}$ is the $(1-\alpha)$-quantile of $\sup_{u\in[0,1]}BB(u)$. Therefore, setting
\begin{equation}
\mathcal U=\left\{P_Y\in\mathcal P_Y:\|F_Y-\hat F_Y\|_\infty\leq\frac{q_{1-\alpha}}{\sqrt n}\right\}\label{uncertainty set}
\end{equation}
ensures that \eqref{confidence setup} holds and subsequently the conclusion in Proposition \ref{basic guarantee}. As $n$ increases, the size of \eqref{uncertainty set} shrinks to zero.
The following result states precisely the implication of this construction, and moreover, describes how this leads to a more tractable optimization formulation:
\begin{theorem}
Let $\overline Z$ and $\underline Z$ be the optimal values of the optimization programs
\begin{equation}
\begin{array}{ll}
\max&\psi(P_X)\\
\text{subject to}&\|F_Y-\hat F_Y\|_\infty\leq\frac{q_{1-\alpha}}{\sqrt n}\\
&P_X\in\mathcal P_X
\end{array}\label{KS RO max direct}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\min&\psi(P_X)\\
\text{subject to}&\|F_Y-\hat F_Y\|_\infty\leq\frac{q_{1-\alpha}}{\sqrt n}\\
&P_X\in\mathcal P_X
\end{array}\label{KS RO min direct}
\end{equation}
where $q_{1-\alpha}$ is the $(1-\alpha)$-quantile of $\sup_{u\in[0,1]}BB(u)$, and $\hat F_Y$ is the empirical distribution of i.i.d. output data. Supposing the true output distribution is continuous, we have
\begin{equation}
\liminf_{n\to\infty}\mathbb P_D(\underline Z\leq\psi(P_X^0)\leq\overline Z)\geq1-\alpha\label{KS confidence guarantee}
\end{equation}
where $P_X^0$ is the true distribution of the input variate $X$. Moreover, \eqref{KS RO max direct} and \eqref{KS RO min direct} are equivalent to
\begin{equation}
\begin{array}{ll}
\max&\psi(P_X)\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq E_{P_X}[I(h(\mathbf X)\leq y_j)]\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&P_X\in\mathcal P_X
\end{array}\label{KS RO max}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\min&\psi(P_X)\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq E_{P_X}[I(h(\mathbf X)\leq y_j)]\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&P_X\in\mathcal P_X
\end{array}\label{KS RO min}
\end{equation}
respectively, where $\hat F_Y(y_j+)$ and $\hat F_Y(y_j-)$ refer to the right- and left-limits of the empirical distributions $\hat F_Y$ at $y_j$, and $E_{P_X}[\cdot]$ denotes the expectation taken with respect to the $T$-fold product measure of $P_X$. \label{KS guarantee}
\end{theorem}
A merit of using the depicted KS-based uncertainty set, seen by Theorem \ref{KS guarantee}, is that it can be reformulated into linear constraints in terms of the expectations $E_{P_X}[\cdot]$ of certain ``moments" of $h(\mathbf X)$. These constraints constitute precisely $n$ interval-type conditions, and the moment functions are the indicator functions of $h(\mathbf X)$ falling under the thresholds $y_j$'s. The derivation leading to the reformulation result in Theorem \ref{KS guarantee} has been used conventionally in computing the KS-statistic. Similar reformulations have also appeared in recent work in approximating stochastic optimization via robust optimization (\cite{bertsimas2014robust}).
The asymptotic of the KS-statistic is more complicated if the output distribution is discrete (this happens if the outputs we look at are for instance the queue length). In such cases, the critical values are generally smaller than those for the continuous distribution (\cite{lehmann2006testing}). Consequently, using $q_{1-\alpha}/\sqrt n$ to calibrate the size of the uncertainty set as in \eqref{uncertainty set} is still valid, but could be conservative, i.e., we still have $P\left(\|\hat F_Y-F_Y^0\|_\infty\leq\frac{q_{1-\alpha}}{\sqrt n}\right)$ asymptotically at least $1-\alpha$, but possibly strictly higher. As a remedy, one can use bootstrapping to calibrate the size of a tighter set. Moreover, the constraint of the form $\|\hat F_Y-F_Y^0\|_\infty\leq q$ will now be written as
\begin{equation}
\hat F_Y(w_j)-q\leq E_{P_X}[I(h(\mathbf X)\leq w_j)]\leq\hat F_Y(w_j)+q,j=1,\ldots,K\label{discrete KS}
\end{equation}
where $w_j,j=1,\ldots$ are the ordered support points of $Y$, with $K=\min\{j:\hat F_Y(w_j)=1\}$. These are the points where jumps occur (and the constraints put on the first $K$ of them automatically ensure the rest). If the support size is small, an alternative is to impose constraints on each probability mass, i.e.,
\begin{equation}
\hat P(Y=w_j)-q\leq E_{P_X}[I(h(\mathbf X)=w_j)]\leq\hat P(Y=w_j)+q,j=1,\ldots,K\label{discrete moment}
\end{equation}
where $\hat P(Y=w_j)$ is the observed proportions of $Y$ being $w_j$, and $q$ can be calibrated by a standard binomial quantile and the Bonferroni correction.
The KS-statistic has several advantages over other types of uncertainty sets in our considered settings. Alternatives like $\chi^2$ goodness-of-fit tests could be used, but the resulting formulations would not come as handy when expressed in terms of $P_X$ or $h(\mathbf X)$, which would affect the efficiency of the gradient estimator that we will discuss in Section \ref{sec:gradient}. Another advantage of using KS-statistic relates to the statistical property of a discretization that is needed to feed into an implementable optimization procedure, which we shall discuss next.
\subsection{Randomizing the Decision Space}
\label{sec:discretization}
Note that optimization programs \eqref{KS RO max} and \eqref{KS RO min} involve decision variable $P_X$ that is potentially infinite-dimensional, e.g., when $X$ is a continuous variable. This can cause algorithmic and storage issues. One could appropriately discretize the decision variable by randomly sampling a finite set of support points on $\mathcal X$. Once these support points are realized, the optimization is imposed on the probability weights on these points, or in other words on a discrete input distribution.
Our next result shows that as the support points are generated from a suitably chosen distribution, and the number of these points grows at an appropriate rate relative to the output data size, the discretized KS-implied optimization will retain the confidence guarantee as the original formulation:
\begin{theorem}
Suppose we sample $\{z_i\}_{i=1,\ldots,m}$ in the space $\mathcal X$ from a distribution $Q$. Suppose that $P_X^0$, the true distribution of $X$, is absolutely continuous with respect to $Q$ and $\|dP_X^0/dQ\|_\infty\leq C$ for some $C>0$, where $dP_X^0/dQ$ is the likelihood ratio calculated from the Radon-Nikodym derivative of $P_X^0$ with respect to $Q$, and $\|\cdot\|_\infty$ denotes the essential supremum. Using the notations as in Theorem \ref{KS guarantee}, we solve
\begin{equation}
\begin{array}{ll}
\max&\psi(P_X)\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq E_{P_X}[I(h(\mathbf X)\leq y_j)]\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&P_X\in\hat{\mathcal P}_X
\end{array}\label{KS RO max discretized}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\min&\psi(P_X)\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq E_{P_X}[I(h(\mathbf X)\leq y_j)]\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&P_X\in\hat{\mathcal P}_X
\end{array}\label{KS RO min discretized}
\end{equation}
where $\hat{\mathcal P}_X$ denotes the set of distributions with support points $\{z_i\}_{i=1,\ldots,m}$. Let $\hat{\overline Z}$ and $\hat{\underline Z}$ be the optimal values of \eqref{KS RO max discretized} and \eqref{KS RO min discretized}.
Denote $\mathbb P$ as the probability taken with respect to both the output data and the support generation for $X$. Suppose that $\psi(P_X)$ takes the form $E_{P_X}[g(\mathbf X)]$ (which subsumes both types of target measures discussed in Section \ref{sec:formulation}) where $E_{P_X^0}[g(X_{i_1},\ldots,X_{i_T})^2]<\infty$ for any $1\leq i_1,\ldots,i_T\leq T$. Also suppose that the true output distribution is continuous and that $\mathbb P(\text{for any\ }P_X\in\hat{\mathcal P}_X, \text{supp}(\gamma(P_X))\cap\{y_j\}_{j=1,\ldots,n}\neq\emptyset)=0$ where $\text{supp}(\gamma(P_X))$ denotes the support of the distribution $\gamma(P_X)$. Then, we have
$$\liminf_{n\to\infty,m/n\to\infty}\mathbb P\left(\hat{\underline Z}+O_p\left(\frac{1}{\sqrt m}\right)\leq\psi(P_X^0)\leq\hat{\overline Z}+O_p\left(\frac{1}{\sqrt m}\right)\right)\geq1-\alpha$$
\label{main guarantee}
\end{theorem}
The error term $O_p(1/\sqrt m)$ represents a random variable of stochastic order $1/\sqrt m$, i.e., $a_m=O_p(1/\sqrt m)$ if for any $\epsilon>0$, there exists $M,N>0$ such that $P(|\sqrt m a_m|\leq N)>1-\epsilon$ for $m>M$.
Theorem \ref{main guarantee} guarantees that by solving the finite-dimensional optimization problems \eqref{KS RO max discretized} and \eqref{KS RO min discretized}, we obtain confidence bounds for the true quantity of interest $\psi(P_X^0)$, up to an error of order $O_p(1/\sqrt m)$. Note that the conclusion holds with the numbers of constraints in \eqref{KS RO max discretized} and \eqref{KS RO min discretized} growing in the data size $n$. One significance of the result is that, despite this growth, as long as one generates the supports of $X$ from a distribution with a heavier tail than the true distribution, and with a size $m$ of order higher than $n$, the confidence guarantee is approximately retained. A key element in explaining this behavior lies in the low complexity of the function class $I(h(\cdot)\leq y)$ (parametrized by $y$) appearing in the constraints and interplayed with the likelihood ratio $dP_X^0/dQ$, as measured by the bracketing number. This number captures the richness of the involved function class with the counts of neighborhoods, each formed by an upper and a lower bounding function that is known as a bracket, in covering the whole class (see the discussion in Appendix \ref{sec:complexity}). A slowly growing (e.g., polynomial in our case) bracketing number turns out to allow the statistic on the output performance measure to be approximated uniformly well with a discretized input distribution, by invoking the empirical process theory for so-called $U$-statistics (\cite{arcones1993limit}). On the other hand, using other moment functions (implied by other test statistics) may not preserve this behavior. This connection to function complexity, which informs the usefulness of sampling-based procedures when integrating with output data, is the first of such kind in the model calibration literature as far as we know.
We have focused on a continuous output distribution in Theorem \ref{main guarantee}. The assumption $\mathbb P(\text{for any\ }P_X\in\hat{\mathcal P}_X, \text{supp}(\gamma(P_X))\cap\{y_j\}_{j=1,\ldots,n}\neq\emptyset)=0$ is a technical condition that ensures the distribution of $h(\mathbf X)$ under $P_X\in\hat{\mathcal P}_X$ does not have overlapping support points as $y_j$'s, which allows us to reduce the KS-implied constraint into the $n$ interval constraints depicted in the theorem. This assumption holds in almost every discrete-event performance measure provided that the considered $P_X$ and $P_Y$ are continuous. On the other hand, if $P_Y$ is discrete, then the theorem holds with the first constraints in \eqref{KS RO max discretized} and \eqref{KS RO min discretized} replaced by \eqref{discrete moment} (with $q$ suitably calibrated as discussed there), without needing the assumption $\mathbb P(\text{for any\ }P_X\in\hat{\mathcal P}_X, \text{supp}(\gamma(P_X))\cap\{y_j\}_{j=1,\ldots,n}\neq\emptyset)=0$.
We mention that \cite{ghosh2015robust} provides a similar guarantee for robust optimization problems designed for quantifying input uncertainty. In particular, their analysis allows to give confidence bounds on output performance measures. However, they do not consider the asymptotic confidence guarantee in relation to the data size and the randomized support size. As a consequence, they do not need considering the complexity of the constraints. Moreover, since they handle input uncertainty, the uncertainty sets are more elementary, in contrast to ours which serve as a tool to invert the input-output relation.
We note that, like Proposition \ref{basic guarantee}, all the results in this section can be similarly extended to a simultaneous guarantee when solving $L$ optimization problems, where each problem has a different objective function $\psi_l(P_X)$. For instance, under the same assumptions as Theorem \ref{main guarantee} with $L$ different objectives in \eqref{KS RO max discretized} and \eqref{KS RO min discretized}, and using the same generated set of support points across all optimization problems, we would obtain that
$$\liminf_{n\to\infty,m/n\to\infty}\mathbb P\left(\hat{\underline Z}_l+O_p\left(\frac{1}{\sqrt m}\right)\leq\psi_l(P_X^0)\leq\hat{\overline Z}_l+O_p\left(\frac{1}{\sqrt m}\right),l=1,\ldots,L\right)\geq1-\alpha$$
where $\hat{\underline Z}_l,\hat{\overline Z}_l$ are the minimum and maximum values of the discretized optimization with objective $\psi_l(P_X)$, and each $O_p(1/\sqrt m)$ is the error term corresponding to each optimization program.
Lastly, we point out that all the results in Sections \ref{sec:formulation} and \ref{sec:KS} hold when we consider $h(\mathbf X,\mathbf W)$ and $g(\mathbf X,\mathbf W)$, where $\mathbf W$ consist of other input variate sequences independent from $\mathbf X$ with known probability distributions. This is as long as we treat all the expectations $E_{P_X}[\cdot]$ as taken jointly under both the product measure of $P_X$ and the known distribution of $\mathbf W$. We provide further remarks in the appendix.
\section{Optimization Procedure}\label{sec:procedure}
This section presents our optimization strategy for (locally) solving \eqref{KS RO max discretized} and \eqref{KS RO min discretized}. Without loss of generality, we only focus on the minimization problem \eqref{KS RO min discretized} since maximization can be converted to minimization by negating the objective. Section \ref{sec:transformation} first discusses the transformation of the stochastic constrained program into a sequence of programs with deterministic convex constraints, using the quadratic penalty method in nonlinear programming. Section \ref{sec:MDSA} then investigates how this transformation can be utilized effectively in a fully iterative stochastic algorithm using MDSA. Section \ref{convergence guarantee} provides a convergence theorem. In the appendix, we also provide an alternate approach that has a similar convergence guarantee but differs in the implementation details.
\subsection{A Stochastic Quadratic Penalty Method}\label{sec:transformation}
When restricted to distributions with support points $\{z_i\}_{i=1,\ldots,m}$, the candidate input distribution $P_X$ can be identified by an $m$-dimensional vector $\mathbf p=(p_1,\ldots,p_m)$ on the probability simplex $\mathcal P:=\{\mathbf p:\sum_{i=1}^mp_i=1,p_i\geq 0\text{ for each }i\}$, where the subscript $X$ is suppressed with no ambiguity. By the vector $\mathbf p$, we mean the distribution that assigns probability $p_i$ to the point $z_i$. The optimization program \eqref{KS RO min discretized} can thus be rewritten as
\begin{equation}
\begin{array}{ll}
\min&\psi(\mathbf p)\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&\mathbf p\in\mathcal P.
\end{array}\label{KS RO min discretized p}
\end{equation}
Note that the constraints in \eqref{KS RO min discretized p} are in general non-convex because the i.i.d.~input sequence means that the expectation $E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]$ is a high-dimensional polynomial in $\mathbf p$. Moreover, this polynomial can involve a huge number of terms and hence its evaluation requires simulation approximation. As far as we know, the literature on dealing with stochastic non-convex constraints is very limited. To overcome this difficulty, we first introduce the quadratic penalty method (\cite{bertsekas1999nonlinear}) to transform program \eqref{KS RO min discretized p} into a sequence of penalized programs with deterministic convex constraints
\begin{equation}
\begin{array}{ll}
\min&\lambda\psi(\mathbf p)+\sum_{j=1}^n (E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]-s_j)^2\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq s_j\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&\mathbf p\in\mathcal P
\end{array}\label{KS RO min discretized penalty}
\end{equation}
where $\mathbf s=(s_1,\ldots,s_n)$ are slack variables and $\lambda>0$ is an inverse measure of the cost/penalty of infeasibility. A related scheme is also used by \cite{wang2008stochastic} in the context of nonconvex stochastic objectives (with deterministic constraints). As $\lambda\to 0$, there is an increasing cost of violating the stochastic constraints, therefore the optimal solution of \eqref{KS RO min discretized penalty} converges to that of \eqref{KS RO min discretized p}, as stated in the following proposition.
\begin{proposition}\label{prop:quadratic penalty}
Suppose \eqref{KS RO min discretized p} has at least one feasible solution. Let $(\mathbf p^*(\lambda),\mathbf s^*(\lambda))$ be an optimal solution of \eqref{KS RO min discretized penalty} indexed at $\lambda$. As $\lambda$ decreases to $0$, every limit point of the sequence $\{\mathbf p^*(\lambda)\}$ is an optimal solution of \eqref{KS RO min discretized p}.
\end{proposition}
As suggested in the proof of Proposition \ref{prop:quadratic penalty}, a mathematically equivalent reformulation of \eqref{KS RO min discretized penalty} with the slack variables optimized is
\begin{equation}
\begin{array}{ll}
\min&\lambda\psi(\mathbf p)+\sum_{j=1}^n (E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]-\Pi_j(E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]))^2\\
\text{subject to}&\mathbf p\in\mathcal P
\end{array}\label{KS RO min discretized penalty2}
\end{equation}
where each $\Pi_j$ is the projection onto the interval $[F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n},F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n}]$ defined as
\begin{equation}\label{pie_j}
\Pi_j(x)=
\begin{cases}
\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}&\text{if }x<\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\\
\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n}&\text{if }x>\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n}\\
x&\text{otherwise}.
\end{cases}
\end{equation}
\subsection{Constrained Stochastic Approximation}\label{sec:MDSA}
Although the formulations \eqref{KS RO min discretized penalty}, \eqref{KS RO min discretized penalty2} are still non-convex, their constraints are convex and deterministic, which can be handled more easily using SA than in the original formulation \eqref{KS RO min discretized p}. This section investigates the design and analysis of an MDSA algorithm for finding local optima of \eqref{KS RO min discretized p} by solving \eqref{KS RO min discretized penalty} with decreasing values of $\lambda$. The appendix would illustrate another algorithm that uses formulation \eqref{KS RO min discretized penalty2} instead of \eqref{KS RO min discretized penalty}.
To describe the algorithm, MD finds the next iterate via optimizing the objective function linearized at the current iterate, together with a penalty on the distance of movement of the iterate. When the objective function is only accessible via simulation, the linearized objective function, or the gradient, at each iteration can only be estimated with noise, in which case the procedure becomes MDSA (\cite{nemirovski2009robust}). More precisely, when applied to the formulation \eqref{KS RO min discretized penalty} with slack variables, MDSA solves the following optimization given a current iterate $(\mathbf p^k,\mathbf s^k)$
\begin{equation}
\begin{array}{ll}
\min&\gamma^k(\lambda\hat{\bm\Psi}^k+\hat{\bm\phi}_{\mathbf p}^k)'(\mathbf p-\mathbf p^k)+\beta^k\hat{\bm\phi}_{\mathbf s}^{k\prime}(\mathbf s-\mathbf s^k)+V(\mathbf p^k,\mathbf p)+\frac{1}{2}\Vert\mathbf s-\mathbf s^k\Vert_2^2\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq s_j\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&\mathbf p\in\mathcal P
\end{array}\label{step optimization2}
\end{equation}
where $\hat{\bm\Psi}^k$ carries the gradient information of the target performance measure $\psi$ at $\mathbf p^k$, while $\hat{\bm\phi}_{\mathbf p}^k$ and $\hat{\bm\phi}_{\mathbf s}^k$ contain the gradient information of the penalty function in \eqref{KS RO min discretized penalty} with respect to $\mathbf p,\mathbf s$ respectively. The sum $V(\mathbf p^k,\mathbf p)+\frac{1}{2}\Vert\mathbf s-\mathbf s^k\Vert_2^2$ serves as the penalty on the movement of the iterate, where $\Vert\cdot\Vert_2$ denotes the standard Euclidean distance, and $V(\cdot,\cdot)$ defined as
\begin{equation}
V(\mathbf p,\mathbf q)=\sum_{i=1}^nq_i\log\frac{q_i}{p_i}\label{KL}
\end{equation}
is the KL divergence between two probability measures. This particular choice of $V$ has been shown (\cite{nemirovski2009robust}) to have superior performance to other choices like the Euclidean distance when the decision space is the probability simplex. Different from traditional SA, the step sizes $\gamma^k$ and $\beta^k$, used for updating $\mathbf p$ and $\mathbf s$ in \eqref{step optimization2}, are different, the rationale for which shall be discussed in Section \ref{convergence guarantee}.
However, iterations in the form of \eqref{step optimization2} can only find optima of \eqref{KS RO min discretized penalty} for a particular penalty coefficient $\lambda$ while retrieving the optimal solution of the original problem \eqref{KS RO min discretized p} through \eqref{KS RO min discretized penalty} hinges on sending $\lambda$ to $0$. Literature on deterministic optimization suggests solving the penalized optimization repeatedly for a set of decreasing values of $\lambda$, but it could be difficult to tell when to stop decreasing the $\lambda$ in our stochastic case. In order to output the optimal solution in one single run, we decrease $\lambda$ together with the step size from one iteration to the next, hence arrive at the following sequential joint solution-and-penalty-updating routine
\begin{equation}
\begin{array}{ll}
\min&\gamma^k(\lambda^k\hat{\bm\Psi}^k+\hat{\bm\phi}_{\mathbf p}^k)'(\mathbf p-\mathbf p^k)+\beta^k\hat{\bm\phi}_{\mathbf s}^{k\prime}(\mathbf s-\mathbf s^k)+V(\mathbf p^k,\mathbf p)+\frac{1}{2}\Vert\mathbf s-\mathbf s^k\Vert_2^2\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq s_j\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&\mathbf p\in\mathcal P
\end{array}\label{step optimization2 lambda}
\end{equation}
where $\lambda^k$ is appropriately chosen in conjunction with $\gamma^k,\beta^k$, and decreases to $0$. To implement the fully sequential scheme, we need to investigate: 1) how to obtain $\hat{\bm\Psi}^k,\hat{\bm\phi}_{\mathbf p}^{k}$ and $\hat{\bm\phi}_{\mathbf s}^{k}$, 2) efficient solution method for program \eqref{step optimization2 lambda}, and 3) how to select the parameters $\gamma^k,\beta^k$ and $\lambda^k$. The next two subsections present the first two investigations respectively, while Section \ref{convergence guarantee} will analyze the convergence of the algorithm in relation to the parameter choices.
\subsubsection{Gradient Estimation and Restricted Programs. }\label{sec:gradient}
Denote by $W(\mathbf p)$ the penalty function in \eqref{KS RO min discretized penalty2}, and by $W_s(\mathbf p,\mathbf s)$ the quadratic penalty in \eqref{KS RO min discretized penalty} where the subscript $s$ refers to ``slack variable''. These are functions of variables on the probability simplex, for which naive differentiation may not lead to simulable object since an arbitrary perturbation may shoot out of the simplex. \cite{gl15_1} and \cite{ghosh2015mirror} have used the idea of Gateaux derivative (in the sense described in Chapter 6 of \cite{serfling2009approximation}) to obtain simulable representations of gradients of expectation-type performance measures. We generalize their result to sums of functions of expectations:
\begin{proposition}
We have:
\begin{enumerate}
\item Suppose $\psi,W,W_s(\cdot,\mathbf s)$ are differentiable in the probability simplex $\mathcal P$, then
\begin{align}
\nabla \psi(\mathbf p)'(\mathbf q-\mathbf p)&=\bm\Psi(\mathbf p)'(\mathbf q-\mathbf p)\label{derivative1}\\
\nabla W(\mathbf p)'(\mathbf q-\mathbf p)&=\bm\phi(\mathbf p)'(\mathbf q-\mathbf p)\label{derivative2}\\
\nabla_{\mathbf p} W_s(\mathbf p,\mathbf s)'(\mathbf q-\mathbf p)&=\bm\phi_{\mathbf p}(\mathbf p,\mathbf s)'(\mathbf q-\mathbf p)\label{derivative3}
\end{align}
for any $\mathbf p,\mathbf q\in\mathcal P$, where the Gateaux derivatives $\bm\Psi(\mathbf p)=(\Psi_1(\mathbf p),\ldots,\Psi_m(\mathbf p))'$, $\bm\phi(\mathbf p)=(\phi_1(\mathbf p),\ldots,\phi_m(\mathbf p))'$, $\bm\phi_{\mathbf p}(\mathbf p,\mathbf s)=(\phi_{\mathbf p,1}(\mathbf p,\mathbf s),\ldots,\phi_{\mathbf p,m}(\mathbf p,\mathbf s))'$, and
\begin{align}
\Psi_i(\mathbf p)&=\frac{d}{d\epsilon}\psi((1-\epsilon)\mathbf p+\epsilon\mathbf 1_i)\Big|_{\epsilon=0^+}\label{Gateaux preservation1}\\
\phi_i(\mathbf p)&=\frac{d}{d\epsilon}W((1-\epsilon)\mathbf p+\epsilon\mathbf 1_i)\Big|_{\epsilon=0^+}\label{Gateaux preservation2}\\
\phi_{\mathbf p,i}(\mathbf p,\mathbf s)&=\frac{d}{d\epsilon}W_s((1-\epsilon)\mathbf p+\epsilon\mathbf 1_i,\mathbf s)\Big|_{\epsilon=0^+}\label{Gateaux preservation3}
\end{align}
\item Assume $\mathbf p=(p_1,\ldots,p_m)$ where each $p_i>0$. Then the Gateaux derivatives \eqref{Gateaux preservation1}\eqref{Gateaux preservation2}\eqref{Gateaux preservation3} are finite and can be expressed as
\begin{align}
\Psi_i(\mathbf p)&=E_{\mathbf p}[g(\mathbf X)S_i(\mathbf X;\mathbf p)]\label{score function1}\\
\phi_i(\mathbf p)&=2\sum_{j=1}^n(E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]-\Pi_j(E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]))E_{\mathbf p}[I(h(\mathbf X)\leq y_j)S_i(\mathbf X;\mathbf p)]\label{score function2}\\
\phi_{\mathbf p,i}(\mathbf p,\mathbf s)&=2\sum_{j=1}^n(E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]-s_j)E_{\mathbf p}[I(h(\mathbf X)\leq y_j)S_i(\mathbf X;\mathbf p)]\label{score function3}
\end{align}
where
$$S_i(\mathbf x;\mathbf p)=\sum_{t=1}^{S}\frac{I_i(x_t)}{p_i}-S\text{ for \eqref{score function1}},\ \text{and }\sum_{t=1}^{T}\frac{I_i(x_t)}{p_i}-T\text{ for \eqref{score function2}\eqref{score function3}}.$$
Here $I_i(x)=1$ if $x=z_i$ and 0 otherwise, and $\mathbf X$ is the i.i.d. input process generated under $\mathbf p$.
\label{derivative form}
\end{enumerate}
\label{lemma:derivative1}
\end{proposition}
The representations \eqref{score function1} and \eqref{score function3} suggest the following unbiased estimators for the gradient of $\psi$, $\bm\Psi(\mathbf p)=(\Psi_i(\mathbf p))_{i=1}^m$, and the gradient of the penalty function, $\bm\phi_{\mathbf p}(\mathbf p,\mathbf s)=(\phi_{\mathbf p,i}(\mathbf p,\mathbf s))_{i=1}^m$
\begin{align}
&\hat{\Psi}_i(\mathbf p)=\frac{1}{M_3}\sum_{r=1}^{M_3}g(\mathbf X^{(r)})S_i(\mathbf X^{(r)};\mathbf p)\label{gradient estimator 1}\\
&\hat{\phi}_{\mathbf p,i}(\mathbf p,\mathbf s)=2\sum_{j=1}^n\frac{1}{M_1}\sum_{r=1}^{M_1}(I(h(\mathbf X^{(r)})\leq y_j)-s_j)\frac{1}{M_2}\sum_{r=1}^{M_2}I(h(\tilde{\mathbf X}^{(r)})\leq y_j)S_i(\tilde{\mathbf X}^{(r)};\mathbf p)\label{gradient estimator 3}
\end{align}
where $\mathbf X^{(r)}$ and $\tilde{\mathbf X}^{(r)}$ are independent copies of the i.i.d. input process generated under $\mathbf p$ and are used simultaneously for all $i,j$. By direct differentiation, a straightforward unbiased estimator for $\bm\phi_{\mathbf s}(\mathbf p,\mathbf s)=(\phi_{\mathbf s,j}(\mathbf p,\mathbf s))_{j=1}^n$, the gradient of the penalty function with respect to $\mathbf s$, is
\begin{equation}
\hat{\phi}_{\mathbf s, j}(\mathbf p,\mathbf s)=\frac{-2}{M_1}\sum_{r=1}^{M_1}(I(h(\mathbf X^{(r)})\leq y_j)-s_j).\label{gradient estimator 4}
\end{equation}
Our main procedure (shown in Algorithm \ref{algo2} momentarily) uses the above gradient estimators, while an alternate MDSA depicted in Algorithm \ref{algo1} in the appendix solves \eqref{KS RO min discretized penalty2} using a biased estimator of $\bm\phi(\mathbf p)$ conferred by \eqref{score function2}.
Note that the above gradient estimators are available thanks to the KS-implied constraints we introduced. By the reformulation in Theorem \ref{KS guarantee}, the constraints in \eqref{KS RO max} and \eqref{KS RO min} become ($T$-fold) expectation-type constraints. Thus, when differentiating the squared expectation in the quadratic penalty, the gradient becomes the product of two $T$-fold expectations, one with the extra factor $S_i(\cdot;\mathbf p)$ which can be interpreted as a score function. This then allows unbiased estimation of the gradient by generating two independent batches of simulation runs each for one of the expectations. Using other statistics to induce the constraints may not lead to such a convenient form.
Note that the $S_i(\cdot;\mathbf p)$ in the gradient estimators \eqref{gradient estimator 1} and \eqref{gradient estimator 3} contains $p_i$ at the denominator, so a small $p_i$ can blow up the variances of the estimators and in turn adversely affect the convergence of MDSA. To ensure convergence, we make an adjustment to our procedure and solve the following restricted version of \eqref{KS RO min discretized p}
\begin{equation}
\begin{array}{ll}
\min&\psi(\mathbf p)\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&\mathbf p\in\mathcal P(\epsilon)
\end{array}\label{KS RO min discretized p epsilon}
\end{equation}
where the restricted probability simplex $\mathcal P(\epsilon):=\{\mathbf p\in \mathcal P:p_i\geq \epsilon\text{ for each }i\}$. Accordingly, the full simplex $\mathcal P$ in the penalized program \eqref{KS RO min discretized penalty} and stepwise subproblem \eqref{step optimization2 lambda} has to be replaced by $\mathcal P(\epsilon)$.
To maintain the statistical guarantee provided by Theorem \ref{main guarantee} when solving the restricted programs, the shrinking size $\epsilon$ has to be appropriately chosen. Theorem \ref{main guarantee epsilon} below indicates that it suffices to choose $\epsilon$ smaller than $1/(m\sqrt n)$ in case of bounded $g(\mathbf X)$.
\begin{theorem}\label{main guarantee epsilon}
Denote by $\hat{\overline Z}_{\epsilon}$ and $\hat{\underline Z}_{\epsilon}$ the maximum and minimum of $\psi(\mathbf p)$ in the feasible set of \eqref{KS RO min discretized p epsilon}. In addition to the conditions of Theorem \ref{main guarantee}, further assume that $g(\mathbf X)$ is bounded. If $\epsilon$ is chosen such that $\epsilon=o\big(\frac{1}{m\sqrt n}\big)$ then we have
$$\liminf_{n\to\infty,m/n\to\infty}\mathbb P\left(\hat{\underline Z}_{\epsilon}+O_p\left(m\epsilon+\frac{1}{\sqrt m}\right)\leq\psi(P_X^0)\leq\hat{\overline Z}_{\epsilon}+O_p\left(m\epsilon+\frac{1}{\sqrt m}\right)\right)\geq 1-\alpha.$$
\end{theorem}
In particular, the first type of target quantities we consider has a bounded $g(\mathbf X)$. Note that the original optimization itself already poses an error of size $O_p(1/\sqrt m)$ in the confidence bounds (Theorem \ref{main guarantee}), so to keep the error at the same level one can use an $\epsilon=O(1/m^{\frac{3}{2}})$ (recall that $m/n\to \infty$). Since the variances of our gradient estimators \eqref{gradient estimator 1}\eqref{gradient estimator 3} can be shown inversely proportional to the components $p_i$ (\cite{ghosh2015robust}), such an $\epsilon$ gives rise to variances of order $O(m^{\frac{3}{2}})$. We point out that this is only slightly worse than the best attainable order $O(m)$, which results from the fact that the average size of $\mathbf p$ in the $m$-dimensional probability simplex is $1/m$.
\subsubsection{Solving Stepwise Subproblem in MDSA. }\label{sec:stepwise}
Since we are now solving the restricted version of subproblem \eqref{step optimization2 lambda}, consider the following generic form
\begin{equation}\label{stepwise slack epsilon}
\begin{array}{ll}
\min&\bm\xi'(\mathbf q-\mathbf p)+\bm\eta'(\mathbf t-\mathbf s)+V(\mathbf p,\mathbf q)+\frac{1}{2}\Vert\mathbf t-\mathbf s\Vert_2^2\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq t_j\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n\\
&\mathbf q\in\mathcal P(\epsilon).
\end{array}
\end{equation}
Because the objective and the feasible set are both separable in $\mathbf q$ and $\mathbf t$, the above program can be decomposed into two independent programs. One is
\begin{equation}
\begin{array}{ll}
\min&\bm\xi'(\mathbf q-\mathbf p)+V(\mathbf p,\mathbf q)\\
\text{subject to}&\mathbf q\in\mathcal P(\epsilon)
\end{array}\label{generic_epsilon}
\end{equation}
and the other is
\begin{equation}\label{generic_slack}
\begin{array}{ll}
\min&\bm\eta'(\mathbf t-\mathbf s)+\frac{1}{2}\Vert\mathbf t-\mathbf s\Vert_2^2\\
\text{subject to}&\hat F_Y(y_j+)-\frac{q_{1-\alpha}}{\sqrt n}\leq t_j\leq\hat F_Y(y_j-)+\frac{q_{1-\alpha}}{\sqrt n},j=1,\ldots,n.
\end{array}
\end{equation}
Program \eqref{generic_slack} is exactly the step-wise routine that appears in the standard gradient descent whose solution takes the form
\begin{equation*}
t^*_j=\Pi_j(s_j-\eta_j)
\end{equation*}
where $\Pi_j$ is the projection defined in \eqref{pie_j}.
The solution of program \eqref{generic_epsilon} has a semi-explicit expression as shown in the following proposition.
\begin{proposition}\label{sol:constrained stepwise}
The optimal solution of the stepwise subproblem \eqref{generic_epsilon} with $0\leq \epsilon<1/m$ is
\begin{equation}\label{sol2}
q^*_i=\frac{\max\{\eta^*,p_ie^{-\xi_i}\}}{\sum_{i=1}^m\max\{\eta^*,p_ie^{-\xi_i}\}}
\end{equation}
where $\eta^*\in [0,\max_ip_ie^{-\xi_i})$ solves the equation
\begin{equation}\label{threshold:eta}
\epsilon=\mu(\eta^*):=\frac{\eta^*}{\sum_{i=1}^m\max\{\eta^*,p_ie^{-\xi_i}\}}.
\end{equation}
\end{proposition}
Proposition \ref{sol:constrained stepwise} suggests a procedure for solving \eqref{generic_epsilon} that involves a root-finding problem \eqref{threshold:eta}. To design an efficient root-finding routine, note that the function $\mu(\eta)$ is strictly increasing in $\eta$. More importantly, it consists of at most $m$ smooth pieces, and on the $i$-th piece it takes the form
\begin{equation*}
\mu(\eta)=\frac{\eta}{i\eta+\sum_{i'=i+1}^mp_{(i')}e^{-\xi_{(i')}}},\text{ if }p_{(i)}e^{-\xi_{(i)}}\leq\eta\leq p_{(i+1)}e^{-\xi_{(i+1)}}
\end{equation*}
where $(p_{(1)}e^{-\xi_{(1)}},\ldots,p_{(m)}e^{-\xi_{(m)}})$ is obtained by sorting $(p_{1}e^{-\xi_{1}},\ldots,p_{m}e^{-\xi_{m}})$ in ascending order. Thus one can first locate which piece the root $\eta^*$ lies on by comparing the values of $\mu$ with $\epsilon$ at the points $p_{(i)}e^{-\xi_{(i)}}$ and then compute $\eta^*$ in closed form from the above expression on that piece. This efficient sort-and-search procedure is described in Algorithm \ref{sort_search} whose proof follows from straightforward algebraic verification and hence is omitted.
\begin{algorithm}[h]
\caption{Sort-and-search for solving \eqref{generic_epsilon} with $0\leq \epsilon<1/m$}
\begin{algorithmic}
\State \textbf{1.} Sort $(p_1e^{-\xi_1},\ldots,p_me^{-\xi_m})$ into ascending order $(p_{(1)}e^{-\xi_{(1)}},\ldots,p_{(m)}e^{-\xi_{(m)}})$, and let $p_{(0)}e^{-\xi_{(0)}}=0$
\State \textbf{2.} Search for the $i^*$ from $0$ to $m-1$ such that
\begin{equation*}
\frac{p_{(i^*)}e^{-\xi_{(i^*)}}}{i^*p_{(i^*)}e^{-\xi_{(i^*)}}+\sum_{i=i^*+1}^mp_{(i)}e^{-\xi_{(i)}}}\leq \epsilon< \frac{p_{(i^*+1)}e^{-\xi_{(i^*+1)}}}{(i^*+1)p_{(i^*+1)}e^{-\xi_{(i^*+1)}}+\sum_{i=i^*+2}^mp_{(i)}e^{-\xi_{(i)}}}
\end{equation*}
\State \textbf{3.} Output $q^*_i$ according to \eqref{sol2} with
\begin{equation*}
\eta^*=\frac{\epsilon\sum_{i=i^*+1}^mp_{(i)}e^{-\xi_{(i)}}}{1-\epsilon i^*}
\end{equation*}
\end{algorithmic}\label{sort_search}
\end{algorithm}
\subsection{Convergence Analysis}\label{convergence guarantee}
We depict our MDSA procedure in Algorithm \ref{algo2}. Steps 1, 2 and 3 of the procedure estimate the gradients using the estimators proposed in Section \ref{sec:gradient}, and Step 4 updates the decision variable with step size $\gamma^k$ and the slack variables with step size $\beta^k$. Steps 1-4 combined are in effect solving the stepwise subproblem \eqref{step optimization2 lambda} with $\mathcal P$ replaced by $\mathcal P(\epsilon)$. Therefore by iterating with decreasing penalty coefficient $\lambda^k$, Algorithm \ref{algo2} searches for the optimum of the restricted formulation \eqref{KS RO min discretized p epsilon}.
\begin{algorithm}[h]
\caption{MDSA for solving \eqref{KS RO min discretized penalty}}
\textbf{Input: }A small parameter $\epsilon>0$, initial solution $\mathbf p^1\in\mathcal P(\epsilon)=\{\mathbf p:\sum_{i=1}^mp_i=1,p_i\geq\epsilon\text{\ for\ }i=1,\ldots,m\}$ and $\mathbf s^1\in [\hat F_Y(y_1+)-\frac{q_{1-\alpha}}{\sqrt n},\hat F_Y(y_1-)+\frac{q_{1-\alpha}}{\sqrt n}]\times \cdots\times [\hat F_Y(y_n+)-\frac{q_{1-\alpha}}{\sqrt n},\hat F_Y(y_n+)-\frac{q_{1-\alpha}}{\sqrt n}]$, a step size sequence $\gamma^k$ for $\mathbf p$, a penalty sequence $\lambda^k$, a step size sequence $\beta^k$ for $\mathbf s$, and sample sizes $M_1,M_2,M_3$.
\textbf{Iteration: }For $k=1,2,\ldots$, do the following: Given $\mathbf p^k,\mathbf s^k$,
\begin{algorithmic}
\State \textbf{1.} Estimate $\hat{\bm\phi}_{\mathbf p}^k=(\hat\phi_{\mathbf p,1}^k,\ldots,\hat\phi_{\mathbf p,m}^k)$, the gradient of the penalty term with respect to $\mathbf p$, with
$$\hat\phi_{\mathbf p,i}^k=2\sum_{j=1}^n\frac{1}{M_1}\sum_{r=1}^{M_1}(I(h(\mathbf X^{(r)})\leq y_j)-s^k_j)\frac{1}{M_2}\sum_{r=1}^{M_2}I(h(\tilde{\mathbf X}^{(r)})\leq y_j)S_i(\tilde{\mathbf X}^{(r)};\mathbf p^k)$$
where $\mathbf X^{(r)},\tilde{\mathbf X}^{(r)}$ are $M_1$ and $M_2$ independent copies of the input process generated under $\mathbf p^k$.
\State \textbf{2.} Estimate $\hat{\bm\Psi}^k=(\hat\Psi_1^k,\ldots,\hat\Psi_m^k)$, the gradient of $E_{\mathbf p}[g(\mathbf X)]$, with
$$\hat\Psi_i^k=\frac{1}{M_3}\sum_{r=1}^{M_3}g(\tilde{\tilde{\mathbf X}}^{(r)})S_i(\tilde{\tilde{\mathbf X}}^{(r)};\mathbf p^k)$$
where $\tilde{\tilde{\mathbf X}}^{(r)}$ are another $M_3$ independent copies of the input process generated under $\mathbf p^k$.
\State \textbf{3.} Estimate $\hat{\bm\phi}_{\mathbf s}^k=(\hat\phi_{\mathbf s,1}^k,\ldots,\hat\phi_{\mathbf s,n}^k)$, the gradient of the penalty term with respect to $\mathbf s$, with
$$\hat\phi_{\mathbf s,j}^k=-\frac{2}{M_1+M_2}\big(\sum_{r=1}^{M_1}(I(h(\mathbf X^{(r)})\leq y_j)-s^k_j)+\sum_{r=1}^{M_2}(I(h(\tilde{\mathbf X}^{(r)})\leq y_j)-s^k_j)\big)$$
where ${\mathbf X}^{(r)},\tilde{\mathbf X}^{(r)}$ are the same replications used in Step 1.
\State \textbf{4.} Compute
$\mathbf p^{k+1}=(p_1^{k+1},\ldots,p_m^{k+1})$ by running Algorithm \ref{sort_search} with $p_i=p_i^{k}$ and $\xi_i=\gamma^k(\lambda^k\hat\Psi_i^k+\hat\phi_{\mathbf p,i}^k)$,
and compute $\mathbf s^{k+1}=(s_1^{k+1},\ldots,s_n^{k+1})$ by
\begin{equation*}
s_j^{k+1}=\Pi_j(s_j^k-\beta^k\hat\phi_{\mathbf s,j}^k).
\end{equation*}
\end{algorithmic}\label{algo2}
\end{algorithm}
To provide convergence guarantee for Algorithm \ref{algo2}, we assume the following:
\begin{assumption}\label{cond:constrained}
The restricted program \eqref{KS RO min discretized p epsilon} has a unique optimal solution $\mathbf p^*_{\epsilon}\in \mathcal P(\epsilon)$ such that for any feasible $\mathbf p\in \mathcal P(\epsilon)$ and $\mathbf p\neq \mathbf p^*_{\epsilon}$ it holds $\bm\Psi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon})>0$, and for any infeasible $\mathbf p\in \mathcal P(\epsilon)$ it holds $\bm\phi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon})>0$, where $\bm\Psi,\bm\phi$ are respectively the Gateaux derivatives of the target quantity $\psi$ and the quadratic penalty function $\sum_{j=1}^n (E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]-\Pi_j(E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]))^2$ in \eqref{KS RO min discretized penalty2}.
\end{assumption}
\begin{assumption}\label{cond:penalized}
There is some threshold $\lambda_{\epsilon}>0$ such that
\begin{description}
\item[1.]for any $\lambda\in (0,\lambda_{\epsilon}]$ the optimization problem \eqref{KS RO min discretized penalty2} with $\mathcal P$ replaced by $\mathcal P(\epsilon)$ has a unique optimal solution $\mathbf p^*_{\epsilon}(\lambda)\in \mathcal P(\epsilon)$ such that for any $\mathbf p\in \mathcal P(\epsilon)$ it holds $(\lambda\bm\Psi(\mathbf p)+\bm\phi(\mathbf p))'(\mathbf p-\mathbf p^*_{\epsilon}(\lambda))\geq 0$
\item[2.]$\mathbf p^*_{\epsilon}(\lambda)$ as a function of $\lambda\in (0,\lambda_{\epsilon}]$ has finite total variation, meaning that there exists a constant $M_{\epsilon}>0$ such that $\sum_{i=0}^{K-1}\norm{\mathbf p^*_{\epsilon}(\lambda_i)-\mathbf p^*_{\epsilon}(\lambda_{i+1})}\leq M_{\epsilon}$
for any $0<\lambda_K<\cdots<\lambda_1<\lambda_0\leq \lambda_{\epsilon}$ and $K$.
\end{description}
\end{assumption}
\begin{assumption}\label{cond:lambda}
$\norm{\mathbf p^*_{\epsilon}(\lambda)-\mathbf p^*_{\epsilon}}=O(\lambda)$ as $\lambda\to 0$.
\end{assumption}
The condition $(\lambda\bm\Psi(\mathbf p)+\bm\phi(\mathbf p))'(\mathbf p-\mathbf p^*_{\epsilon}(\lambda))\geq 0$ in Assumption \ref{cond:penalized} is a weakened version of the general convexity criterion that has appeared in online learning (e.g., \cite{bottou1998online}) and SA (e.g., \cite{benveniste2012adaptive,broadie2011general}) literature. For a minimization problem with objective $f(x)$ and minimizer $x^*$, this criterion refers to the condition that $\nabla f(x)'(x-x^*)>0$ for any $x\neq x^*$. A geometric interpretation of it is that the opposite of the gradient direction always points to the optimum. Part 1 of Assumption \ref{cond:penalized} stipulates that the criterion holds weakly for the penalized program \eqref{KS RO min discretized penalty2} when the penalty coefficient $\lambda$ lies in a small neighborhood of zero. Assumption \ref{cond:constrained} can be viewed as the same criterion for the limit case $\lambda=0$. To explain, at a feasible solution $\mathbf p$ of \eqref{KS RO min discretized p epsilon} the derivative $\bm\phi(\mathbf p)$ vanishes hence the criterion in Assumption \ref{cond:penalized} reduces to $\bm\Psi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon}(\lambda))\geq 0$ when $\lambda>0$, which in the limit $\lambda\to 0$ forces $\bm\Psi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon})\geq 0$ since $\mathbf p^*_{\epsilon}(\lambda)\to \mathbf p^*_{\epsilon}$. Whereas for an infeasible solution $\mathbf p$ the derivative $\bm\phi(\mathbf p)$ is non-zero, thus the criterion becomes $\bm\phi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon})\geq 0$ as $\lambda\to 0$ because $\lambda\bm\Psi(\mathbf p)\to \mathbf 0$ and $\mathbf p^*_{\epsilon}(\lambda)\to \mathbf p^*_{\epsilon}$. Note that Assumption \ref{cond:constrained} further requires the two inequalities to hold strictly.
Part 2 of Assumption \ref{cond:penalized} and Assumption \ref{cond:lambda} impose mild regularity conditions on the solution path of \eqref{KS RO min discretized penalty2} parametrized by $\lambda$. In fact, the solution path is expected to be continuously differentiable in $\lambda$, a stronger property than the assumptions. The reason is that the optimal solution $\mathbf p^*_{\epsilon}(\lambda)$ has to satisfy the set of KKT conditions which is smooth in the decision variable $\mathbf p$ and the penalty coefficient $\lambda$, hence an application of the implicit function theorem reveals the continuous differentiability of $\mathbf p^*_{\epsilon}(\lambda)$ in $\lambda$.
When the target quantity $\psi(\mathbf p)=\mathbf c'\mathbf p$ for some $\mathbf c\in \mathbb{R}^m$, which includes the first type of target quantities we consider in Section \ref{sec:formulation}, the condition $\bm\Psi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon})>0$ in Assumption \ref{cond:constrained} is guaranteed to hold. To explain, note that the feasible set of program \eqref{KS RO min discretized p epsilon} is supported by the hyperplane $\{\mathbf p:\mathbf c'\mathbf p=\mathbf c'\mathbf p^*_{\epsilon}\}$ at the optimum $\mathbf p^*_{\epsilon}$ even if the feasible set is non-convex, and any non-optimal solution $\mathbf p$ will lie in the strict half-space $\{\mathbf p:\mathbf c'\mathbf p> \mathbf c'\mathbf p^*_{\epsilon}\}$ which is exactly the condition in Assumption \ref{cond:constrained}. However, the second condition $\bm\phi(\mathbf p)'(\mathbf p-\mathbf p^*_{\epsilon})>0$ could still be hard to verify because of the nonlinearity of the constraint functions $E_{\mathbf p}[I(h(\mathbf X)\leq y_j)]$. In our numerical experiments, we investigate the use of multi-start and show that our procedure appears to perform well empirically.
Our convergence guarantee of Algorithm \ref{algo2} is stated in Theorem \ref{thm:algo2}, whose proof follows the framework in \cite{blum1954multidimensional} that considers SA on unconstrained problems.
\begin{theorem}
Under Assumptions \ref{cond:constrained}, \ref{cond:penalized} and \ref{cond:lambda}, if the step size sequences $\{\gamma^k\},\{\beta^k\}$ and the penalty sequence $\{\lambda^k\}$ of Algorithm \ref{algo2} are chosen as
\begin{equation}\label{stepsize}
\begin{aligned}
&\gamma^k=\frac{a}{k^{\alpha_1}},\ \frac{3}{4}<\alpha_1\leq 1\\
&\beta^k=\frac{b}{k^{\alpha_2}},\ 2-2\alpha_1<\alpha_2<2\alpha_1-1\\
&\lambda^k=
\begin{cases}
\frac{c}{k^{\alpha_3}},\ 0<\alpha_3\leq 1-\alpha_1&\text{if }\frac{3}{4}<\alpha_1<1\\
\frac{c}{\log k}&\text{if }\alpha_1=1
\end{cases}
\end{aligned}
\end{equation}
then $\mathbf p^k$ generated in Algorithm \ref{algo2} converges to $\mathbf p^*_{\epsilon}$ a.s..\label{thm:algo2}
\end{theorem}
Here $\gamma^k$ and $\beta^k$ are chosen in such a way that the slack variables $\mathbf s^k$ is guaranteed to stay close to the projections $\Pi_j(E_{\mathbf p^k}[I(h(\mathbf X)\leq y_j)])$ and hence the MDSA is effectively solving \eqref{KS RO min discretized penalty2}. Note that the choice of penalty coefficient $\lambda^k$ only depends on the step size $\gamma^k$. The rule of thumb is that $\gamma^k\lambda^k$ should sum up to $\infty$, as indicated by the relation between $\alpha_1$ and $\alpha_3$ in \eqref{stepsize}. This ensures sufficient exploration of the feasible region of \eqref{KS RO min discretized p epsilon}, the rationale of which will be further elaborated in Appendix \ref{sec:proofs algorithm}.
Finally, we mention that in the presence of a collection of auxiliary input sequences $\mathbf W$ with known distribution that is independent of $\mathbf X$, namely that we now have $h(\mathbf X,\mathbf W)$ instead of $h(\mathbf X)$ and $g(\mathbf X,\mathbf W)$ instead of $g(\mathbf X)$, all the results in this section hold by viewing $E_{\mathbf p}[\cdot]$ as taken jointly with respect to the product measure of $\mathbf p$ and the true distribution of $\mathbf W$. In Algorithm \ref{algo2} (and also the other algorithms in the appendix), one only needs to simulate the independent $\mathbf W$ in conjunction with $\mathbf X$ in each replication, e.g., $h(\mathbf X^{(r)},\mathbf W^{(r)})$ instead of $h(\mathbf X^{(r)})$. Appendix \ref{sec:procedure proofs} provides further discussion.
\section{Numerical Results}\label{sec:numerics}
This section provides numerical illustration of our methodology. We focus on a stylized M/G/1 queue, where we assume known i.i.d. unit rate exponential interarrival times. Our goal is to calibrate the unknown i.i.d. service time distribution $P_X$ given the output data. Here, we assume the collection of data for the averaged wait time of the first 20 customers, starting from the empty state. Say these observations are i.i.d. (e.g., among different days or work cycles), denoted $y_1,\ldots,y_n$. The data size $n$ varies from $30$ to $100$ in our experiments.
We consider two target quantities of interest $\psi(P_X)$: 1) the expected averaged queue length seen by the first 20 customers. This performance measure, though related to the waiting time data, is not directly observable and depends on the known service time distribution; 2) the distribution function of the service time. We also consider two different ``true" service time distributions, first one is exponential with rate $1.2$, and second one is a mixture of beta distributions that has a bimodal shape. We set the confidence level to be $95\%$, i.e., $\alpha=5\%$.
Since the input distribution of interest and the output distribution are both continuous, we use optimization programs \eqref{KS RO max discretized} and \eqref{KS RO min discretized} to infer the confidence bounds on $\psi(P_X^0)$. From Theorem \ref{main guarantee}, we first randomly sample $m$ support points from some ``safe" input distribution (i.e., distribution believed to have heavier tail than the truth), where $m$ varies from $100$ to $500$ in our experiments. Then we implement Algorithm \ref{algo2}. In our implementation we choose $M_1 = M_2 = M_3 = 100$ , $\gamma^k = a/k^{0.8}$, $\beta^k = b/k^{0.5}$, $\lambda^k = c/k^{0.2}$, in which the constants $a, b, c$ will be determined slightly different in different cases. The iteration stops when $\|\mathbf p^{k+1} - \mathbf p^{k}\|_\infty \leq 0.0005$.
\subsection{Inferring the Average Queue Length}
We first consider inferring the average queue length $E_{P_X}[g(\mathbf X)]$, and consider a small output data size $n = 30$ for the average waiting time. In this setting, the true service time distribution is set as exponential with rate $1.2$. We generate the input support points with a lognormal distribution with parameter $\mu = 0$ and standard deviation $\sigma = 1$. In light of Theorem \ref{main guarantee}, we choose $m = 100$ to make $m$ bigger than $n$. Figure \ref{trend_queue_100_30} shows the trend of the objective value $E_{\mathbf p}[g(\textbf{X})]$ when we apply Algorithm \ref{algo2} to the max and the min problems. The algorithm appears to converge fairly quickly (within about 10 iterations). The jitter of the trend is due to the evaluation of the objective value, for each of whom we use $100,000$ simulation runs. The minimization stops at $0.622$ and the maximization stops at $0.688$ according to our stopping criterion described above. This gives us an interval $[0.622,0.688]$. The true value in this case is $E_{\mathbf p}[g(\textbf{X})] = 0.636$ (from running 1 million simulation using the true service time distribution), thus demonstrating that the confidence interval we obtained covers the truth. Moreover, the interval we obtained is encouragingly tight.
\begin{figure}[htbp]
\centering
\includegraphics{lowerqueue100_30.jpg}
\includegraphics{upperqueue100_30.jpg}
\caption{Objective value of the minimization (left) and maximization (right) for the expected queue length using Algorithm \ref{algo2} against the iteration number; $n=30,m=100$; true service time distribution is exponential}\label{trend_queue_100_30}
\end{figure}
We also investigate the shape of the input distribution when the algorithm stops. This is shown in Figure \ref{dist_queue_100_30}. We observe that both the obtained maximal and minimal distributions place more masses on the lower value than the upper, roughly following the true exponential distribution. We should mention, however, that the shapes of the obtained optimal distributions are not indicative of the performance of our method, as the latter intends to compute valid bounds for a target quantity, namely the average queue length in this example, instead of direct recovery of the input distribution. The shapes in Figure \ref{dist_queue_100_30} should be interpreted as the worst-case distributions that give rise to the lower and upper bounds for the queue length. The resemblance of these distributions to the true one leads us to conjecture that the service time distribution could be close to identifiable with the waiting time data.
\begin{figure}[htbp]
\centering
\includegraphics{dist_lowerqueue100_30.jpg}
\includegraphics{dist_upperqueue100_30.jpg}
\caption{Minimal (left) and maximal (right) distribution of the service time for bounding the expected queue length; $n=30,m=100$; true service time distribution is exponential}\label{dist_queue_100_30}
\end{figure}
Next we increase our support size $m$ to $200$, keeping the output data size $n$ fixed at $30$. Like the previous case, we show the trend of the objective value as the algorithm progresses, in Figure \ref{trend_queue_200_30}. Compared to the case $m=100$, the algorithm appears to stabilize faster, at around 5 iteration, and exhibit a more monotonic trend (which could be due to our initialization). The minimization stops at $0.622$ and the maximization stops at $0.647$. This gives us an interval $[0.622,0.647]$ which again covers the true value $0.636$, and is shorter than the one obtained when $m=100$. Finally, The obtained maximal and minimal distributions, shown in Figure \ref{dist_queue_200_30}, show a pattern even closer to the exponential distribution.
\begin{figure}[htbp]
\centering
\includegraphics{lowerqueue200_30.jpg}
\includegraphics{upperqueue200_30.jpg}
\caption{Objective value of the minimization (left) and maximization (right) for the expected queue length using Algorithm \ref{algo2} against the iteration number; $n=30,m=200$; true service time distribution is exponential}\label{trend_queue_200_30}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics{dist_lowerqueue200_30.jpg}
\includegraphics{dist_upperqueue200_30.jpg}
\caption{Minimal (left) and maximal (right) distribution of the service time for bounding the expected queue length; $n=30,m=200$; true service time distribution is exponential}\label{dist_queue_200_30}
\end{figure}
We increase the support size $m$ further to $300$ or the data size $n$ to $100$. Table \ref{table1} shows the obtained optimal values. These runs provide valid lower and upper bounds for the true value $0.636$, except when $m=300$ and $n=30$ that misses marginally. The interval lengths do not seem to vary much; all are around $0.03-0.06$.
\begin{table}[htbp]
\footnotesize
\centering
\begin{tabular}{cccccccccc}
\toprule
$m$ & $n$ & min value & max value\\
\midrule
100 & 30 & 0.622 & 0.688 \\
200 & 30 & 0.622 & 0.647 \\
300 & 30 & 0.593 & 0.629 \\
100 & 100 & 0.627 & 0.652 \\
\bottomrule
\end{tabular}%
\caption{Optimal values for bounding the expected queue length under different combinations of $n$ and $m$; true service time distribution is exponential}
\label{table1}%
\end{table}%
The selection of $a, b, c$ in $\gamma^k$, $\beta^k$, $\lambda^k$ depends on $m$ and $n$. We have selected $a = 0.2$ when $m = 100$ and $n = 30$, $a = 0.1$ and $0.075$ when $m = 200$ and $300$ while $n=30$, and $a = 0.1$ when $m = 100$ and $n = 100$. We always choose $b = 0.2$ and $c = 1$. These choices appear to work well.
Regarding running times, when $m = 100$ and $n = 30$, each iteration takes about 40 seconds. The running time seems to increase linearly as $m$ and $n$ increase.
Next we check how the initialization of the probability weights in the algorithm affects the obtained optimal values. This is especially important since our algorithm is only guaranteed local convergence. We randomly generate 34 initial distributions of $\mathbf p$ from a Dirichlet distribution to run the algorithm. Figure \ref{different initial boxplot} shows the boxplot of the obtained optimal values under different initial distributions. The minimum value varies from $0.621$ to $0.635$, whereas the maximum value varies from $0.648$ to $0.665$. The differences among the initial distributions seem to be quite small compared to the gap between the minimum and maximum values, and the true value $0.636$ is always covered. This shows that the algorithm tends to converge to the same optimal solution or solutions that have similar objective values.
\begin{figure}[htbp]
\centering
\includegraphics[width = .6\textwidth]{different_initial_boxplot.jpg}
\caption{Minimum and maximum values for the expected queue length under different initializations; $n=30,m=100$; true service time distribution is exponential}\label{different initial boxplot}
\end{figure}
We then test the coverage of our obtained bounds. For this, we repeatedly sample new output data set of size $n=30$ for $100$ times. For each data set, we generate new support points of size $m=100$. Then we run Algorithm \ref{algo2}. Out of $100$ intervals we obtained, five of them cover the true expected queue length. This gives us a $95\%$ confidence interval for the coverage probability $[0.91,0.99]$, which is consistent with the theoretical guarantee provided by Theorem \ref{main guarantee}.
We have also tested the use of randomized stochastic projected gradient (RSPG), proposed by \cite{ghadimi2016mini}, that has been shown to perform well theoretically and empirically for problems with non-convex stochastic objectives. Specifically, we adapt the algorithm in Section 4.1 and 4.2 of \cite{ghadimi2016mini} heuristically for the current problem we face that has stochastic non-convex constraints. Algorithm \ref{RSPG} in the appendix shows the adaptation of a single run procedure, and Algorithm \ref{2-RSPG-V} shows the adaptation of a post-optimization step to boost the final performance. In our algorithmic specification, we choose $N=30$, $S=5$, $M=500$, $M'=500$, $\bar\gamma=0.03$, and we fix $\lambda$ at $0.03$. We run Algorithm \ref{2-RSPG-V} for two realizations of data and support generation when the true service time distribution is exponential, with $n=30$ and $m=100$. For each realization, we also run Algorithm \ref{algo2} for comparison. For the first realization, we obtained $[0.622, 0.640]$ using RSPG, compared with $[0.626,0.658]$ using Algorithm \ref{algo2}. For the second realization, we obtained $[0.616,0.644]$ using RSPG, compared with $[0.621,0.660]$ using Algorithm \ref{algo2}. The RSPG thus appears to perform very similarly as our procedure, at least for this particular setup (which shows that RSPG could be an alternative for future investigation).
We test the sensitivity of the algorithm with respect to the bounds in the constraints provided by the KS statistic. More concretely, in Algorithm \ref{algo2}, we increase the number $q_{1-\alpha}/\sqrt n$ in the constraint interval by a small $\delta$. Table \ref{change_of_delta} shows that the obtained bounds are quite stable and do not show significant changes.
\begin{table}[htbp]
\footnotesize
\centering
\begin{tabular}{cccccccccc}
\toprule
perturbation size & min value & max value\\
\midrule
0.01 & 0.625 & 0.649 \\
0.02 & 0.628 & 0.649 \\
0.03 & 0.624 & 0.643 \\
0.05 & 0.621 & 0.646 \\
\bottomrule
\end{tabular}%
\caption{Effect on optimal values for bounding the expected queue length when perturbing the interval in the optimization constraint; $n=30,m=100$; true service time distribution is exponential}
\label{change_of_delta}%
\end{table}%
Finally, we test with a more ``challenging" service time distribution that is an equally weighted mixture of two beta distributions with parameters $\alpha = 9,\beta = 3$ and $\alpha = 3,\beta = 9$. This bimodal distribution has highest masses around $0.2$ and $0.8$, with a shape shown in Figure \ref{density of bimodal}.
\begin{figure}[htbp]
\centering
\includegraphics[width = .6\textwidth]{density_of_bimodal.jpg}
\caption{Density of a mixture of two beta distributions }\label{density of bimodal}
\end{figure}
We consider the setting with $n=50$ output observations. We randomly select $m=100$ input support points from uniform distribution in $[0,1]$, and run Algorithm \ref{algo2}, using the same specifications as in the previous setup. The minimization stops at the value $0.242$ and the maximization stops at $0.284$. These cover the true value $0.274$ (from running 1 million simulation using the true service time distribution). Thus our method appears to continue working in this case.
Figure \ref{dist_bimodal} shows the minimal and maximal distributions from Algorithm \ref{algo2}. The distributions are quite spread out throughout the support, though the minimal distribution appears to have a noisy bimodal pattern. As we have discussed before, the shapes of these distributions should be interpreted as the worst-case distributions giving rise to the bounds, but are not indicative of the performance of our approach.
\begin{figure}[htbp]
\centering
\includegraphics{dist_lower_bimodal.jpg}
\includegraphics{dist_upper_bimodal.jpg}
\caption{Minimal (left) and maximal (right) distribution of the service time for bounding the expected queue length; $n=50,m=100$; true service time distribution is mixture of betas}\label{dist_bimodal}
\end{figure}
\subsection{Inferring the Input Distribution Function}
We now consider inferring the distribution function of the service time, i.e., $P_X(X\leq a)$ for a range of values $a$. We first use a true service time distribution that is exponential with rate $1.2$. We consider a collection of $n=50$ observations from the average waiting time. We randomly generate $m =100$ support points from a lognormal distribution with $\mu=0$ and $\sigma^2=1$. We use Algorithm \ref{algo2} with parameters $\gamma^k = 0.1/k^{0.8}$, $\beta^k = 0.1/k^{0.5}$, $\lambda^k = 1/k^{0.2}$.
Table \ref{expo_P(X<=a)} shows the obtained maximum and minimum values compared with the true distribution function evaluated at values $a$ ranging from $0.3$ to $1.2$. Figure \ref{expo_ggplot} further plots the trends of these values. The dashed lines represent the maximum and minimum values, and the solid line represents the true values. Note that Proposition \ref{extended guarantee}, and the analogous extension of Theorem \ref{main guarantee} to multiple objective functions discussed at the end of Section \ref{sec:discretization}, allow us to compute the bounds for different $a$ values simultaneously with little sacrifice of statistical accuracy. In Table \ref{expo_P(X<=a)} and Figure \ref{expo_ggplot}, the obtained optimal values cover the truth at all points except the leftmost $a=0.3$. This could be due to the challenge in inferring the tail (either left or right), stemming from perhaps the observed output we use (i.e., the waiting time) or the statistic we use to form our uncertainty set (i.e., the KS-statistic, which is known to not capture well the tail region of a distribution).
\begin{table}[htbp]
\footnotesize
\centering
\begin{tabular}{cccccccccc}
\toprule
$a$ & min value & max value & true value\\
\midrule
0.3 & 0.118 & 0.250 & 0.302 \\
0.4 & 0.302 & 0.441 & 0.381 \\
0.5 & 0.398 & 0.464 & 0.451 \\
0.6 & 0.435 & 0.565 & 0.513 \\
0.7 & 0.506 & 0.579 & 0.568 \\
0.8 & 0.601 & 0.673 & 0.617 \\
0.9 & 0.636 & 0.735 & 0.660 \\
1 & 0.699 & 0.741 & 0.699 \\
1.1 & 0.723 & 0.756 & 0.733 \\
1.2 & 0.756 & 0.798 & 0.763 \\
\bottomrule
\end{tabular}%
\caption{Minimum, maximum and true values of the distribution function $P_X(X\leq a)$ of the service time across $a$; $n=50,m=100$; true service time distribution that is exponential}
\label{expo_P(X<=a)}%
\end{table}%
\begin{figure}[htbp]
\centering
\includegraphics[width = .6\textwidth]{expo_ggplot.jpg}
\caption{Bounds and true distribution function values for the service time, when the true service time distribution is exponential; $n=50,m=100$}\label{expo_ggplot}
\end{figure}
Figure \ref{dist_exp_a = 0.5} shows the minimal and maximal distributions for bounding $P_X(X\leq0.5)$ when the algorithm terminates. We see that the shapes of both distributions resemble exponential, hinting that the service time distribution is close to identifiable in this case.
\begin{figure}[htbp]
\centering
\includegraphics[width = .4\textwidth]{dist_lower_exp_a5.jpg}\includegraphics[width = .4\textwidth]{dist_upper_exp_a5.jpg}
\caption{Minimal (left) and maximal (right) distribution of the service time for bounding $P_X(X\leq0.5)$, when the true service time distribution is exponential; $n=50,m=100$}\label{dist_exp_a = 0.5}
\end{figure}
Next, we investigate the case when the true service time distribution is a mixture of two beta distributions with parameters $\alpha = 9,\beta = 3$ and $\alpha = 3,\beta = 9$. We consider a collection of $n=50$ observations from the average waiting time. We randomly generate $m =100$ support points from a uniform distribution on $[0,1]$.
Like in the previous case, Table \ref{P(x<=a)} shows the maximum and minimum values from Algorithm \ref{algo2}, against the true values of $P_X(X\leq a)$ at different $a$ values. Figure \ref{ggplot} further plots the trends of these values. Here, the obtained optimal values all cover the truth except at $a = 0.35$. The latter could be attributed to the statistical noise when running the many optimization procedures. The point $a=0.35$ is also one that could be ``difficult" to infer intuitively, as it is in between the two modes. Nonetheless, our procedure appears to be reliable in general in bounding the distribution function across the domain of the service time.
\begin{table}[htbp]
\footnotesize
\centering
\begin{tabular}{cccccccccc}
\toprule
$a$ & min value & max value & true value\\
\midrule
0.2 & 0.129 & 0.231 & 0.188 \\
0.25 & 0.208 & 0.266 & 0.267 \\
0.3 & 0.262 & 0.358 & 0.337 \\
0.35 & 0.296 & 0.395 & 0.393 \\
0.4 & 0.362 & 0.413 & 0.435 \\
0.45 & 0.389 & 0.464 & 0.466 \\
0.5 & 0.416 & 0.503 & 0.491 \\
0.55 & 0.504 & 0.577 & 0.516 \\
0.6 & 0.509 & 0.594 & 0.548 \\
0.65 & 0.573 & 0.611 & 0.591 \\
0.7 & 0.628 & 0.679 & 0.649 \\
0.75 & 0.678 & 0.736 & 0.722 \\
0.8 & 0.724 & 0.834 & 0.805 \\
\bottomrule
\end{tabular}%
\caption{Minimum, maximum and true values of the distribution function $P_X(X\leq a)$ of the service time across $a$, under a true service time distribution that is mixture of betas; $n=50,m=100$}
\label{P(x<=a)}%
\end{table}%
\begin{figure}[htbp]
\centering
\includegraphics[width = .6\textwidth]{ggplot.jpg}
\caption{Bounds and true distribution function values for the service time, when the true service time distribution is mixture of betas; $n=50,m=100$}\label{ggplot}
\end{figure}
Figure \ref{dist_bimodal_a = 0.5} shows the minimal and maximal distributions for bounding $P_X(X\leq0.5)$ when the algorithm terminates. The shapes of these distributions are now considerably noisier than the exponential case in Figure \ref{dist_exp_a = 0.5}. Nonetheless, there is a rough bimodal pattern (around $0.2$ and $0.7$).
\begin{figure}[htbp]
\centering
\includegraphics[width = .4\textwidth]{bimodal_lower_a5.jpg}\includegraphics[width = .4\textwidth]{bimodal_upper_a5.jpg}
\caption{Minimal (left) and maximal (right) distribution of the service time for bounding $P_X(X\leq0.5)$, when the true service time distribution is mixture of betas; $n=50,m=100$}\label{dist_bimodal_a = 0.5}
\end{figure}
\section{Conclusion}\label{sec:discussion}
We have studied an optimization-based framework to calibrate input quantities in stochastic simulation with only the availability of output data. Our approach uses an output-level uncertainty set, inspired by the DRO literature, to represent the statistical noise of the output data. By expressing the output distribution in terms of a simulable map of the input distribution, we can set up optimization programs cast over the input distribution that infers valid confidence bounds on the input quantities of interest.
We propose in particular an output-level uncertainty set based on the KS statistic, which exhibits advantages in computation (thanks to reformulation) and statistical accuracy (thanks to a controllable discretization scale needed to retain the confidence guarantee). We have shown these advantages via looking at the complexity of the resulting constraints and invoking the empirical process theory for $U$-statistics. We also study a stochastic quadratic penalty method to solve the resulting optimization problems, including a convergence analysis that informs the suitable tuning of the parameters. Our numerical results demonstrate how our method could provide valid bounds for input quantities such as the input distribution function and other performance measures that rely on the input.
\ACKNOWLEDGMENT{A preliminary conference version of this work has appeared in \cite{goeva2014reconstructing}. We gratefully acknowledge support from the National Science Foundation under grants CMMI-1542020, CMMI-1523453 and CAREER CMMI-1653339. We also thank Peter Haas for suggesting the use of quantile-based moments, and Russell Barton, Shane Henderson and Barry Nelson for other helpful suggestions.}
\bibliographystyle{informs2014}
|
\section{Introduction}\label{S: introduction}
Copulas are a versatile and useful tool for
modeling multivariate distributions.
See, for example, \cite{fan2014copulas}, \cite{patton2009copula},
\cite{durante2015principles} and \cite{trivedi2007copula}.
Modeling non-continuous marginal
random variables is a challenging task due to computational
problems, interpretation difficulties and various other pitfalls and
paradoxes; see \cite{Smith2012}, for example. \textcolor{blue}{The main source of the
computational issues arises from the difficulty of directly evaluating the likelihood. For example, when modeling a vector of $m$ discrete random variables, evaluating the likelihood at one point requires computing $2^m$ terms.} The literature on modeling non-continuous random marginal problems
has mostly focused on cases where all the marginals are discrete, and less
extensively, on cases where some marginals are discrete and some are continuous.
See, for example,
\cite{genest2007primer}, \cite{Smith2012}, \cite{de2013analysis},
and \cite{panagiotelis2012pair}.
Furthermore, a lot of the literature has focused on approaches restricted to certain classes to copulas. For example, this is the case for Gaussian copulas (See for instance \cite{shen2006copula}, \cite{hoff2007extending}, \cite{song2009joint}, \cite{de2011copula}, \cite{he2012gaussian} and
\cite{jiryaie2016gaussian}) or pair-copula constructions (see \cite{stober2015comorbidity}). Relatively little attention has been paid to the case where some variables are a mixture of discrete and continuous components.
In contrast, our approach, presents methodology for an arbitrary copula and can be applied quite generally as long as it is possible to compute certain marginal and conditional copulas either in closed-form or numerically.
Our article extends the Bayesian methodology used
for estimating continuous marginals to the case where each marginal can be
a mixture of an absolutely continuous random variable and a discrete random variable.
In particular, we are interested in applying the new methodology to copulas that are mixtures of Gaussian and Archimedean copulas.
To illustrate the methodology and sampling algorithm we apply them to estimate a multivariate income dynamics model.
In this application, we use the copula framework to
model the dependence structure of random variables that are mixtures of discrete and continuous components, and apply the model
to empirical economic data.
We note that there are many other
real world economic applications that involve such mixtures of random variables as marginals, and these are briefly discussed in Section~\ref{S: conclusion}.
Our proposed methodology extends that introduced in \cite{Pitt2006} and \cite{Smith2012}. \cite{Smith2012} allow the joint modeling of
distributions of random variables such that each component can be either discrete or continuous. However, neither paper covers the case where some random variables can be
a mixture of an absolutely continuous random variable and a discrete random variable. In a financial econometrics application, \cite{brechmann2014}
consider the case where the marginal distributions are mixtures of continuous and \textcolor{blue}{points of probability mass at zero}.
In contrast, our paper derives the likelihood equations in a much more general setting that allows for the
margins to be arbitrarily classified into three groups: absolutely continuous, discrete and mixtures of absolutely continuous and discrete random variables. Furthermore, there is
no restriction on the number or location of the point masses present in each margin. This can occur in many economic data, for instance in cases where earnings are top-coded and have individuals with zero earnings. Equally, our setting covers the case of dependent interval-censored data.
The paper is organized as follows. Section~\ref{S: likelihood definition} outlines the copula model and defines the likelihood as a density with respect to a mixed measure.
Section~\ref{S: estimation and algorithms} presents the simulation algorithms used for inference.
Section~\ref{S: application to individual income dynamics} applies the methods and algorithms
to model multivariate income dynamics. This section describes the data and presents the estimation results.
Section~\ref{S: conclusion} concludes. \textcolor{blue}{The paper has two appendices}.
Appendix~\ref{app_diference_operator} defines the \textcolor{blue}{difference notation which is a handy tool useful when writing
formulas for the likelihood of our model in closed-form.} Appendix \ref{app_integration} presents and proves the results required to define the likelihood
as a density with respect to a mixed measure. The paper also has an online supplement whose sections are denoted as Sections~S1, etc.
Section~\ref{app_sampling_schemes} describes the Gaussian and Archimedean copulas used in the article, as well as
the Markov chain Monte Carlo (MCMC) sampling scheme. \textcolor{blue}{Section~\ref{S: trivariate example}
introduces a new three dimensional example to further illustrate the methods in the paper.
Section~\ref{proof: proof of lemma 3} gives a proof of Lemma~\ref{L: elementary} which is discussed in
Appendix~\ref{app_integration}}. Section~\ref{app: extra empirical results} presents some additional empirical results.
\section{Defining the Likelihood of a general copula\label{S: likelihood definition}}
This section discusses the proposed model and shows how to write the likelihood of an i.i.d. sample from it.
Each random vector is modeled using a marginal
distribution-copula decomposition and each marginal is allowed to be a mixture
of an absolutely continuous component and \textcolor{blue}{ a discrete component}. The MCMC sampling scheme
in the next section is based on this definition of the likelihood.
Let $\ensuremath{\boldsymbol{X}}= (X_1, \ldots, X_m)$ be \textcolor{blue}{an $\mathbbm{R}^m$-valued random vector.
If, for example, $X_j$ is categorical, then
its support would be a finite subset of $\mathbbm{R}$ and thus without loss of generality, we can work with $\mathbbm{R}^m$.} Let
$\mathcal{M}= \{ 1, \ldots, m \}$ be the index set, and $2^{\mathcal{M}}$ its
power-set (or the set of all of its subsets). Let the random variable $X_j$
have cumulative distribution function $F_j$ for $j = 1, \ldots, m$. By the
Lebesgue decomposition theorem \textcolor{blue}{\citep[][Chapter 7, Theorem 1.1]{shorack2000probability},
and assuming there are no continuous singularities \citep[see][for a detailed discussion]{durante2015principles}},
the distribution of each $X_j$ can be written as a mixture
of an absolutely continuous random variable and a discrete random variable.
This means that $F_j$ is allowed to have jumps at a countable number of
points. In order to exploit this result, we would like to be able to
decide at each point of \textcolor{blue}{$\mathbbm{R}^m$}, which indices have jumps in
their corresponding CDFs.
We need a mapping $\mathcal{C}: \mathbbm{R}^m \rightarrow 2^{\mathcal{M}}$
that, for each $\ensuremath{\boldsymbol{x}} \in \mathbbm{R}^m$, picks out the
subset of the indices of $\ensuremath{\boldsymbol{x}}$ where $F_j$ is
continuous at $x_j$ for each $j \in \mathcal{C}
(\ensuremath{\boldsymbol{x}})$.
\begin{align*}
\mathcal{C} : \mathbbm{R}^m & \longrightarrow 2^{\mathcal{M}} \quad \text{with} \quad
\ensuremath{\boldsymbol{x}} \longrightarrow \mathcal{C}
(\ensuremath{\boldsymbol{x}}).
\end{align*}
Similarly, we define the set $\mathcal{D} (\ensuremath{\boldsymbol{x}}) = \mathcal{M} -
\mathcal{C} (\ensuremath{\boldsymbol{x}})$ (the complement of
$\mathcal{C} (\ensuremath{\boldsymbol{x}})$ in $\mathcal{M}$, that is the
set of indices $j$ for which $F_j$ presents jumps at $x_j$). This means
that for all $ \ensuremath{\boldsymbol{x}} \in \mathbbm{R}^m$, $\{
\mathcal{C} (\ensuremath{\boldsymbol{x}}), \mathcal{D}
(\ensuremath{\boldsymbol{x}}) \}$ partitions the index set so that
$\mathcal{C} (\ensuremath{\boldsymbol{x}}) \cap \mathcal{D}
(\ensuremath{\boldsymbol{x}}) = \varnothing$ and $\mathcal{C}
(\ensuremath{\boldsymbol{x}}) \cup \mathcal{D} (\ensuremath{\boldsymbol{x}})
=\mathcal{M}$.
As a first example, consider $\ensuremath{\boldsymbol{X}}= (X_1, X_2)$, where
$X_1 \sim \mathcal{N} (0, 1)$ and $X_2$ is a mixture of an exponential distribution with
parameter $\lambda$ and a point mass at $0$ with probability $p$, i.e., $X_2
\sim p \delta_0 + (1 - p) \mathcal{E} (\lambda)$). Then, $\mathcal{C} (x_1, 0) = \{ 1 \}$ for all
$ x_1 \in \mathbbm{R}$
and $\mathcal{C} (x_1, x_2) = \{ 1, 2 \}$ for all $x_1 \in \mathbbm{R}, x_2 >
0$. Similarly $\mathcal{D} (x_1, 0) = \{ 2 \}$ for all $x_1 \in \mathbbm{R}$
and $\mathcal{D} (x_1, x_2) = \varnothing$.
As a second example, let $\ensuremath{\boldsymbol{X}}= (X_1, X_2)$, where $X_1$
is Bernoulli and $X_2 \sim \mathcal{N} (0, 1)$. Then $\mathcal{C}
(\ensuremath{\boldsymbol{x}}) = \{ 2 \}$ for all $ \ensuremath{\boldsymbol{x}}
\in \{ 0, 1 \} \times \mathbbm{R}$. Similarly $\mathcal{D}
(\ensuremath{\boldsymbol{x}}) = \{ 1 \}$ for all $
\ensuremath{\boldsymbol{x}}$.
Let $\ensuremath{\boldsymbol{U}}= (U_1, \ldots, U_m)$ be a vector of
uniform random variables whose distribution is given by some copula $C$.
We assume that $F_j^{- 1}$ is the quantile function corresponding to $F_j$
(since $F_j$ is not invertible when \textcolor{blue}{ $X_j$ is not absolutely continuous}, this
corresponds to picking one possible generalized inverse function).
The variables $\ensuremath{\boldsymbol{U}}$ are selected to satisfy
the following criteria. If, at coordinate $x_j$, $j \in \mathcal{C}
(\ensuremath{\boldsymbol{x}})$, then $u_j = F_j (x_j)$, resulting in
a deterministic one-to-one relationship when conditioning on either $U_j$ or
$X_j$. Otherwise, $j \in \mathcal{D} (\ensuremath{\boldsymbol{x}})$, and
we require $x_j = F_j^{- 1} (u_j)$, resulting in an infinity of
$U_j$ corresponding to one $X_j$ and spanning the interval $(F_j (X_j^-), F_j
(X_j))$. This interval corresponds to gaps in the range of $F_j$. If
$\mathcal{C} (\ensuremath{\boldsymbol{x}}) =\mathcal{M}$ for every
$\ensuremath{\boldsymbol{x}}$, then $C$ will be the copula of
$\ensuremath{\boldsymbol{X}}$. Otherwise, the copula structure will still
create dependence between \textcolor{blue}{the non-continuous marginal variables but will not be unique in
general}.
Mathematically, the above description leads to the joint density
\begin{equation}
\label{JointDensity} f (\ensuremath{\boldsymbol{x}},
\ensuremath{\boldsymbol{u}}) := c (\ensuremath{\boldsymbol{u}}) \prod_{j \in
\mathcal{C} (\ensuremath{\boldsymbol{x}})} \mathcal{I} (u_j = F_j (x_j))
\prod_{j' \in \mathcal{D} (\ensuremath{\boldsymbol{x}})} \mathcal{I} (F_{j'}
(x_{j'}^-) \leqslant u_{j'} < F_{j'} (x_{j'})),
\end{equation}
where $c$
is the copula density corresponding to $C$ and $\mathcal{I}$ is an indicator
variable. See Lemma~\ref{L: mixed density in x and u}, part~(i), of Appendix~~\ref{app_integration} for a derivation of
\eqref{JointDensity} and the corresponding measure.
Notice that in ~\eqref{JointDensity}, products over the indices $j$ and $ j'$ correspond to different
partitions for each $\ensuremath{\boldsymbol{x}}$.
With a small abuse of notation, we call $\ensuremath{\boldsymbol{U}}$ the vector of latent variables, even though $U_j$ is a deterministic function of
$X_j$ if $F_j$ is invertible.
To derive the likelihood function, that is the marginal density of
$\ensuremath{\boldsymbol{X}}$, from the joint density $f (\ensuremath{\boldsymbol{x}},
\ensuremath{\boldsymbol{u}})$, we introduce some notation. Let
$\ensuremath{\boldsymbol{a}}, \ensuremath{\boldsymbol{b}}$ be two vectors in
$\mathbbm{R}^k$ such that $\ensuremath{\boldsymbol{a}} \leqslant
\ensuremath{\boldsymbol{b}}$ componentwise and let $g$ be an arbitrary
function from $\mathbbm{R}^k$ into $\mathbbm{R}$. \textcolor{blue}{We denote by
$\bigtriangleup_{\ensuremath{\boldsymbol{a}}}^{\ensuremath{\boldsymbol{b}}} g
(.)$ the sum} of $2^k$ terms that are obtained by repeatedly subtracting $g
(., a_j, .)$ from $g (., b_j, .)$ for each $j = 1, \ldots, k$. Appendix
\ref{app_diference_operator} contains more details on using this notation.
For each $\ensuremath{\boldsymbol{x}} \in \mathbbm{R}^m$, denote by
$\ensuremath{\boldsymbol{b}}= (F_1 (x_1), \ldots, F_m (x_m))$ the vector of upper bounds and similarly denote by
$\ensuremath{\boldsymbol{a}}= (F_1 (x_1^-), \ldots, F_m (x_m^-))$ the vector of lower bounds. For each $j
\in \mathcal{C} (\ensuremath{\boldsymbol{x}})$, $\ensuremath{\boldsymbol{b}}
(j) =\ensuremath{\boldsymbol{a}} (j)$, otherwise we have the strict inequality
$\ensuremath{\boldsymbol{b}} (j) >\ensuremath{\boldsymbol{a}} (j)$. Denote the
partitions of $\ensuremath{\boldsymbol{a}}$ and $\ensuremath{\boldsymbol{b}}$
by $\ensuremath{\boldsymbol{a}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}$,
$\ensuremath{\boldsymbol{a}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}$,
$\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}$ and
$\ensuremath{\boldsymbol{b}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}$. For
some sets $A, B \subset \mathcal{M}$, denote by $c_A$ and $c_{A|B}$, \ the
marginal copula density over the indices of $A$, the conditional copula
density where the variables in $A$ are conditioned on the variables
with index set $B$. It is possible to do the same for \textcolor{blue}{ $C_A$ and $C_{A|B}$, the copula
distribution functions.}
If $(\ensuremath{\boldsymbol{X}}, \ensuremath{\boldsymbol{U}})$ has the joint
density given by \eqref{JointDensity}, then the marginal density of
$\ensuremath{\boldsymbol{X}}$ is
\begin{equation}
\label{likelihood} f (\ensuremath{\boldsymbol{x}}) = c_{\mathcal{C}
(\ensuremath{\boldsymbol{x}})} (\ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}})}) \prod_{j \in \mathcal{C}
(\ensuremath{\boldsymbol{x}})} f _j(x_j)
\bigtriangleup_{\ensuremath{\boldsymbol{a}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}}^{\ensuremath{\boldsymbol{b}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}} C_{\mathcal{D} (\ensuremath{\boldsymbol{x}})
|\mathcal{C} (\ensuremath{\boldsymbol{x}})}
(\cdot|\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}),
\end{equation}
which corresponds to writing the formula for the density of
$\ensuremath{\boldsymbol{X}}$ as the product of the (marginal) density of
continuous components at $\ensuremath{\boldsymbol{x}}$
\[ f (\ensuremath{\boldsymbol{x}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})
= c_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}
(\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})
\prod_{j \in \mathcal{C} (\ensuremath{\boldsymbol{x}})} f _j(x_j), \]
and the (conditional) density of the non-continuous components conditional on
the continuous ones
\[ f (\ensuremath{\boldsymbol{x}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}
|\ensuremath{\boldsymbol{x}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}) =
\bigtriangleup_{\ensuremath{\boldsymbol{a}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}}^{\ensuremath{\boldsymbol{b}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}} C_{\mathcal{D}
(\ensuremath{\boldsymbol{x}}) |\mathcal{C} (\ensuremath{\boldsymbol{x}})}
(\cdot|\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}).
\]
See Lemma~\ref{L: mixed density in x and u}, part~(ii), of Appendix~~\ref{app_integration} for a derivation of
\eqref{likelihood} and the corresponding measure.
\textcolor{blue}{We now give a bivariate example to illustrate how the formulas can be used. This example is continued in later
sections. See also Section~\ref{S: trivariate example} for a trivariate illustrative example. }
\begin{example}[running illustrative example]\label{ex: example 1}
Let $X_1$ have a density that is a mixture of point of probability mass at
zero and a normal distribution $f_1 (x_1) \sim \pi \delta_{x_1} (0) + (1 -
\pi) \phi (x_1)$ where $\phi (.)$ is the density of a standard normal.
This implies that the cumulative distribution function of $X_1$ is
\[ F_1 (x_1) = (1 - \pi) \Phi (x_1) + \pi \mathcal{I} (x_1 \geqslant 0), \]
and thus there a discontinuity in $F_1$ at the point 0.
Let $X_2$ be a binary random variable with $\Pr \{ X_2 = 0 \} = \gamma$.
Let $C (\cdot)$ and $c (\cdot)$ be respectively the Clayton copula and
Clayton copula density with parameter $\theta = 1$, so that
\[ C (u_1, u_2) = \left( \frac{1}{u_1} + \frac{1}{u_2} - 1 \right)^{- 1} , \quad
c (u_1, u_2) = \frac{2}{u_1^2 u_2^2} \left( \frac{1}{u_1} + \frac{1}{u_2} -
1 \right)^{- 3} \]
and the conditional copula is given by
\[ C_{2|1} (u_2 |u_1) = \frac{1}{u_1^2} \left( \frac{1}{u_1} + \frac{1}{u_2} -
1 \right)^{- 2}, \]
which has the conditional quantile function $C^{- 1} (\tau |u_1) =
\frac{\sqrt{\tau} u_1}{1 + \sqrt{\tau} (u_1 - 1)}$ and the conditional density
$c_{2|1} (u_2 |u_1) = c (u_1, u_2)$ (because the marginal of $u_1$ is
uniform).
The following details are necessary construct the example.
$\mathcal{C} (\tmmathbf{x}) = \{ 2 \}$ for $x_1 \neq 0$, for all $ x_2$ and
$\mathcal{C} (\tmmathbf{x}) = \{ 1, 2 \}$ for $x_1 = 0$, for all $ x_2$
{\underline{Joint of $\tmmathbf{x}$ and $\tmmathbf{u}$}} ( Eq.~\eqref{JointDensity} )
There are two cases. Case 1: $x_1 \neq 0$
$f (x_1, x_2, u_1, u_2) = c (u_1, u_2) \mathcal{I} (u_1 = F_1 (x_1))
\mathcal{I} (F_2 (x_2 -) \leqslant u_2 < F_2 (x_2))$
Case 2: $x_1 = 0$
$f (x_1, x_2, u_1, u_2) = c (u_1, u_2) \mathcal{I} (F_1 (0 -) \leqslant u_1 <
F_1 (0)) \mathcal{I} (F_2 (x_2 -) \leqslant u_2 < F_2 (x_2))$
{\underline{Likelihood at one point}} (Eq.~\ref{likelihood} )
If $x_1 \neq 0$, then
\begin{eqnarray*}
f (x_1, x_2) & = & f (x_1) \bigtriangleup_{F_2 (x_2 -)}^{F_2 (x_2)} C_{2|1}
(\cdot |F (x_1))\\
& = & f_1 (x_1) \{ C_{2|1} (F_2 (x_2) |F_1 (x_1)) - C_{2|1} (F_2 (x_2 -)
|F_1 (x_1)) \}
\end{eqnarray*}
because $c (u_1) = 1$ as one-dimensional margins of a copula
are all uniform.
If $x_1 = 0$, then
\begin{eqnarray*}
f (0, x_2) & = & \bigtriangleup_{F_1 (0 -)}^{F_1 (0)} \bigtriangleup_{F_2
(x_2 -)}^{F_2 (x_2)} C (\cdot)\\
& = & \bigtriangleup_{F_1 (0 -)}^{F_1 (0)} \{ C (\cdot, F_2 (x_2)) - C
(\cdot, F_2 (x_2 -)) \}\\
& = & C (F_1 (0), F_2 (x_2)) - C (F_1 (0), F_2 (x_2 -)) - C (F_1 (0 -), F_2
(x_2)) + C (F_1 (0 -), F_2 (x_2 -)).
\end{eqnarray*}
\end{example}
The difficult part of implementing a simulation algorithm based on equations \eqref{JointDensity} and
\eqref{likelihood} is that the size of the vectors
$\ensuremath{\boldsymbol{x}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}$ and
$\ensuremath{\boldsymbol{x}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}$
changes with $\ensuremath{\boldsymbol{x}}$. A secondary difficulty is that the
second term is a sum of $2^{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}$ terms
for each $\ensuremath{\boldsymbol{x}}$, where $| \mathcal{D}
(\ensuremath{\boldsymbol{x}}) |$ is the cardinality of the set $\mathcal{D}
(\ensuremath{\boldsymbol{x}})$.
\section{Estimation and Algorithms\label{S: estimation and algorithms}}
\subsection{Conditional distribution of the latent variables\label{SS: conditional distn of latent variables}}
In any simulation scheme (such as MCMC or simulated EM) where the latent
variables $\ensuremath{\boldsymbol{U}}$ are used to carry out inference, it
is necessary to know the distribution of
$\ensuremath{\boldsymbol{U}}|\ensuremath{\boldsymbol{X}}$. This distribution
is singular due to the deterministic relationship over
$\mathcal{C} (\ensuremath{\boldsymbol{x}})$ for each
$\ensuremath{\boldsymbol{x}} \in \mathbbm{R}^m$. For this reason, it is useful
to work only with $\ensuremath{\boldsymbol{U}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})} |\ensuremath{\boldsymbol{X}}$. A second issue is
the need to work with different sizes of vectors
$\ensuremath{\boldsymbol{U}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}$ for
each $\ensuremath{\boldsymbol{x}}$ in our sample (say
$\ensuremath{\boldsymbol{x}}_1, \ldots, \ensuremath{\boldsymbol{x}}_n$), so we
will be working with $n$ distributions over different spaces. Recursively
using Bayes formula and similar integration arguments to the ones described
during the derivation of the $\ensuremath{\boldsymbol{X}}$ density, we obtain
the density for $\ensuremath{\boldsymbol{U}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})} |\ensuremath{\boldsymbol{X}}$ as
\begin{equation}
\label{LatentConditionalDistribution} f
(\ensuremath{\boldsymbol{u}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}
|\ensuremath{\boldsymbol{x}}) = \frac{c_{\mathcal{D}
(\ensuremath{\boldsymbol{x}}) |\mathcal{C} (\ensuremath{\boldsymbol{x}})}
(\ensuremath{\boldsymbol{u}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}
|\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})
\prod_{j \in \mathcal{D} (\ensuremath{\boldsymbol{x}})} \mathcal{I} (a_j
\leqslant u_j <
b_j)}{\bigtriangleup_{\ensuremath{\boldsymbol{a}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}}^{\ensuremath{\boldsymbol{b}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}} C_{\mathcal{D} (\ensuremath{\boldsymbol{x}})
|\mathcal{C} (\ensuremath{\boldsymbol{x}})}
(\cdot|\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})},
\end{equation}
where the denominator is a constant of integration. As seen from the
above conditional density, one of the complexities arising is that the
distribution $\ensuremath{\boldsymbol{U}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})}
|\ensuremath{\boldsymbol{X}}=\ensuremath{\boldsymbol{x}}$ depends on the whole
vector $\ensuremath{\boldsymbol{x}}$ and not just on
$\ensuremath{\boldsymbol{x}}_{\mathcal{D} (\ensuremath{\boldsymbol{x}})}$.
See Lemma~\ref{L: mixed density in x and u}, part~(iii), of Appendix~~\ref{app_integration} for a derivation of
\eqref{LatentConditionalDistribution} and the corresponding measure.
We can now proceed in two ways. We can either draw each $U_j$ in $\ensuremath{\boldsymbol{U}}_{\mathcal{D}(\ensuremath{\boldsymbol{x}})}$ separately conditionally on everything else. This is reminiscent of a single move Gibbs sampler. Alternatively, it turns out that in spite of the difficulties, the above distribution can also
be sampled recursively without having to compute any of the above normalizing
constants. By writing $\mathcal{D} (\ensuremath{\boldsymbol{x}})$ as
$\{ j_1, \ldots, j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |} \}$,
we can use the following scheme
\begin{itemize}
\item $U_{j_1} |\ensuremath{\boldsymbol{X}}$
\item $U_{j_2} |U_{j_1}, \ensuremath{\boldsymbol{X}}$
\item $\vdots$
\item $U_{j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}} |U_{j_1},
\ldots, U_{j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) | - 1}},
\ensuremath{\boldsymbol{X}}$
\end{itemize}
\textcolor{blue}{We now note that the order of the indices $j_1,...,j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}$ is irrelevant for the sampling scheme. Although it might appear that the sampling procedure depends on the ordering of those indices, the acceptance or rejection of such samples also depends on the ordering and the next subsection shows
that such a procedure will always result in a correct MCMC draw from the conditional distribution $\ensuremath{\boldsymbol{U}}_{\mathcal{D}
(\ensuremath{\boldsymbol{x}})} |\ensuremath{\boldsymbol{X}}$.}
The above sampling scheme requires knowing the marginal distribution of
$\ensuremath{\boldsymbol{U}}_{\mathcal{J}} |\ensuremath{\boldsymbol{X}}$ for
$\mathcal{J} \subset \mathcal{D} (\ensuremath{\boldsymbol{x}})$ and the
conditional decomposition $U_j |\ensuremath{\boldsymbol{U}}_{\mathcal{K}},
\ensuremath{\boldsymbol{X}}$ where $ (\{ j \}, \mathcal{K})$ is a partition of
$\mathcal{J}$ (meaning $\{ j \} =\mathcal{J} \backslash \mathcal{K}$, the
complement of $\mathcal{K}$ in $\mathcal{J}$). This distribution can be
derived as
\begin{eqnarray*}
f (\ensuremath{\boldsymbol{u}}_{\mathcal{J}} |\ensuremath{\boldsymbol{x}}) &
= & \frac{c (\ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}})}) \prod_{j \in \mathcal{C}
(\ensuremath{\boldsymbol{x}})} f (x_j)}{f (\ensuremath{\boldsymbol{x}})} c
(\ensuremath{\boldsymbol{u}}_{\mathcal{J}}
|\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})\\
& \times & \left[
\bigtriangleup_{\ensuremath{\boldsymbol{a}}_{\mathcal{J}^c}}^{\ensuremath{\boldsymbol{b}}_{\mathcal{J}^c}}
C_{\ensuremath{\boldsymbol{U}}_{\mathcal{J}^c}
|\ensuremath{\boldsymbol{U}}_{\mathcal{J}},
\ensuremath{\boldsymbol{U}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})}}
(\cdot |\ensuremath{\boldsymbol{u}}_{\mathcal{J}},
\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})
\right] \prod_{j \in \mathcal{J}} \mathcal{I} (a_j \leqslant u_j < b_j)
\end{eqnarray*}
with $\mathcal{J}^c =\mathcal{D} (\ensuremath{\boldsymbol{x}}) \backslash
\mathcal{J}$ and
\begin{eqnarray*}
f (u_j |\ensuremath{\boldsymbol{u}}_{\mathcal{K}},
\ensuremath{\boldsymbol{x}}) & = & c (u_j
|\ensuremath{\boldsymbol{u}}_{\mathcal{K}},
\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})
\mathcal{I} (a_j \leqslant u_j < b_j)\\
& \times &
\frac{\bigtriangleup_{\ensuremath{\boldsymbol{a}}_{\mathcal{J}^c}}^{\ensuremath{\boldsymbol{b}}_{\mathcal{J}^c}}
C_{\ensuremath{\boldsymbol{U}}_{\mathcal{J}^c}
|\ensuremath{\boldsymbol{U}}_{\mathcal{J}},
\ensuremath{\boldsymbol{U}}_{\mathcal{C}}} (\cdot
|\ensuremath{\boldsymbol{u}}_{\mathcal{J}},
\ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}})})}{\bigtriangleup_{\ensuremath{\boldsymbol{a}}_{\mathcal{K}^c}}^{\ensuremath{\boldsymbol{b}}_{\mathcal{K}^c}}
C_{\ensuremath{\boldsymbol{U}}_{\mathcal{K}^c}
|\ensuremath{\boldsymbol{U}}_{\mathcal{K}},
\ensuremath{\boldsymbol{U}}_{\mathcal{C}}} (\cdot
|\ensuremath{\boldsymbol{u}}_{\mathcal{K}},
\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})},
\end{eqnarray*}
where $\mathcal{K}^c =\mathcal{J}^c \cup \{ j \}$.
\textcolor{blue}{We continue to illustrate how to apply the latent variables conditional formulas by considering Example~\ref{ex: example 1}.}
\setcounter{example}{0}
\begin{example}[continued]
If $x_1 \neq 0$, then
\[ f (u_2 |\tmmathbf{x}) = \frac{c_{2|1} (u_2 |F_1 (x_1)) \mathcal{I} (F_2
(x_2 -) \leqslant u_2 < F_2 (x_2))}{C_{2|1} (F_2 (x_2) |F_1 (x_1)) -
C_{2|1} (F_2 (x_2 -) |F_1 (x_1))} \]
($u_1$ is deterministically equal to $F_1 (x_1)$, so we only need to sample
$u_2$).
If $x_1 = 0$
\[ f (u_1, u_2 |\tmmathbf{x}) = \frac{c (u_1, u_2) \mathcal{I} (F_1 (0 -)
\leqslant u_1 < F_1 (0)) \mathcal{I} (F_2 (x_2 -) \leqslant u_2 < F_2
(x_2))}{C (F_1 (0), F_2 (x_2)) - C (F_1 (0), F_2 (x_2 -)) - C (F_1 (0 -),
F_2 (x_2)) + C (F_1 (0 -), F_2 (x_2 -))} \]
\end{example}
\subsection{Metropolis-Hastings sampling}
It is clear from the formulas for $f (u_j
|\ensuremath{\boldsymbol{u}}_{\mathcal{K}}, \ensuremath{\boldsymbol{x}})$
that they are quite intricate. They correspond to a product of a simple term
$c (u_j |\ensuremath{\boldsymbol{u}}_{\mathcal{K}},
\ensuremath{\boldsymbol{b}}_{\mathcal{C} (\ensuremath{\boldsymbol{x}})})
\mathcal{I} (a_j \leqslant u_j < b_j)$ (a truncated conditional copula density)
and a complicated term that depends on ratios of normalizing constants for
$f (\ensuremath{\boldsymbol{u}}_{\mathcal{J}} |\ensuremath{\boldsymbol{x}})$
and $f (\ensuremath{\boldsymbol{u}}_{\mathcal{K}}
|\ensuremath{\boldsymbol{x}})$. One of the most useful aspects of the
Metropolis-Hastings (MH) algorithm is that it does not require knowledge of normalizing
constants. The trick here is that those normalizing constants are obtained
recursively. Assume that we sample
\begin{itemize}
\item $U_{j_1}$ from $c (u_{j_1}) \mathcal{I} (a_{j_1} \leqslant u_{j_1} <
b_{j_1})$
\item $U_{j_2}$ from $c (u_{j_2} |u_{j_1}) \mathcal{I} (a_{j_2} \leqslant
u_{j_2} < b_{j_2})$
\item $\vdots$
\item $U_{j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}}$ from $c
(u_{j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}} |u_{j_1}, \ldots,
u_{j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) | - 1}}) \mathcal{I}
(a_{j_{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}} \leqslant u_{j_{|
\mathcal{D} (\ensuremath{\boldsymbol{x}}) |}} < b_{j_{| \mathcal{D}
(\ensuremath{\boldsymbol{x}}) |}})$
\end{itemize}
that is, if we use as proposal a truncated form of the copula marginal density
over $\mathcal{D} (\ensuremath{\boldsymbol{x}})$, then computing the MH
accept/reject ratio results in the computationally simple formula
\[ \alpha (\ensuremath{\boldsymbol{x}}_i) = \prod_{k = 1}^{| \mathcal{D}
(\ensuremath{\boldsymbol{x}}) |} \frac{C (F_{j_k} (x_{i, j_k}) |u^N_{i,
j_1}, \ldots, u^N_{i, j_{k - 1}}, \ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}}_i), i}) - C (F_{j_k} (x_{i, j_k}^-) |u^N_{i,
j_1}, \ldots, u^N_{i, j_{k - 1}}, \ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}}_i), i})}{C (F_{j_k} (x_{i, j_k}) |u^O_{i,
j_1}, \ldots, u^O_{i, j_{k - 1}}, \ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}}_i), i}) - C (F_{j_k} (x_{i, j_k}^-) |u^O_{i,
j_1}, \ldots, u^O_{i, j_{k - 1}}, \ensuremath{\boldsymbol{b}}_{\mathcal{C}
(\ensuremath{\boldsymbol{x}}_i), i})} \]
where $i$ represents the observation index. The complexity of this formula is
much smaller than $2^{| \mathcal{D} (\ensuremath{\boldsymbol{x}}) |}$.
\textcolor{blue}{We now illustrate the Metropolis-Hastings acceptance probabilities by again considering Example~\ref{ex: example 1}.}
\setcounter{example}{0}
\begin{example}[continued]
If $x_1 \neq 0$, then the ratio is
$ \alpha (x_2) = 1 $
and if $x_1 = 0$ (first draw $u_1^N$ from a uniform on $(F_1 (0^-), F_1 (0))$
and compare to the previous draw $u_1^O$)
\[ \alpha (0, x_2) = \frac{C_{2|1} (F_2 (x_2) |u_1^N) - C_{2|1} (F_2 (x_2^-)
|u_1^N)}{C_{2|1} (F_2 (x_2) |u_1^O) - C_{2|1} (F_2 (x_2^-) |u_1^O)} \]
Note that here the ordering does not matter, as we could have computed the
other ratio (if we draw instead first $u_2^N$ from a uniform on $(F_2 (x_2^-),
F_2 (x_2))$
\[ \alpha (0, x_2) = \frac{C_{1|2} (F_1 (0) |u_2^N) - C_{1|2} (F_1 (0^-)
|u_2^N)}{C_{1|2} (F_1 (0) |u_2^O) - C_{1|2} (F_1 (0^-) |u_2^O)} \]
Even though the ratio are different, both procedures will result in a draw
from $f (u_1, u_2 |\tmmathbf{x})$.
\end{example}
\subsection{Mixtures of Archimedean and Gaussian copulas\label{SS: mixtures of archimedean and gaussian copulas}}
This section applies the previous results to the family of mixtures of Archimedean and Gaussian copulas. Working with mixtures of copulas
provides a simple and yet rich and flexible modeling framework because mixtures of copulas are copulas themselves,
We are particularly interested in having a mixture of three components, two Archimedean copulas, the Clayton copula
$\left(C_{Cl}\right)$ and the Gumbel copula $\left(c_{Gu}\right)$ and a Gaussian copula $\left(c_{G}\right)$
component. We will later apply this mixture to model the dependence between individual income distributions
over 13 years. The copula density of this 3-component mixture is
\begin{equation}
c_{mix}\left(\boldsymbol{u};\Gamma,\theta_{Cl},\theta_{Gu},w_{1},w_{2}\right)=w_{1}c_{G}\left(\boldsymbol{u};\Gamma\right)+w_{2}c_{Cl}\left(\boldsymbol{u};\theta_{Cl}\right)
+w_{3}c_{Gu}\left(\boldsymbol{u};\theta_{Gu}\right), \label{eq:mixture copula model}
\end{equation}
where $w_{1}$, $w_{2}$, and $w_{3}=1-w_{1}-w_{2}$ are the mixture
weights, and $\Gamma$, $\theta_{Cl}$, and $\theta_{Gu}$ are respectively the
dependence parameters of the Gaussian, Clayton, and Gumbel copulas.
Such a mixture of copula models has the additional flexibility
of being to capture lower and upper tail dependence. We will use a Bayesian approach
to estimate the copula parameters and, for simplicity and without loss of generality, we follow \cite{joe2014dependence} and use empirical CDF's
to model the marginal distributions.
Let the parameter $w_{k}$ denote the probability that the $i$-th
observation comes from the $k$-th component in the mixture. Let $\boldsymbol{d}_{i}=\left(d_{i1},d_{i2},d_{i3}\right)^{'}$
be indicator (latent) variables such that $d_{ik}=1$ when the $i$-th
observation comes from the $k$-th component in the mixture. These
indicator variables identify the component of the copula model defined
in equation \eqref{eq:mixture copula model} to which the observation
$\boldsymbol{y}_{i}$ belongs. Then,
\begin{equation}
p\left(d_{ik}=1|\boldsymbol{w}\right)=w{}_{k},
\end{equation}
with $w_{k}>0$ and $\sum_{k=1}^{3}w_{k}=1$.
Given the information on the $n$ independent sample observations
$\boldsymbol{y}=\left(\boldsymbol{y}_{1},...,\boldsymbol{y}_{n}\right)^{'}$
and $\boldsymbol{y}_{i}=\left(y_{i1},...,y_{iT}\right)$, and by using
Bayes rule, the joint posterior density is obtained as
\begin{equation}
p\left(\boldsymbol{w},\boldsymbol{d},\Gamma,\theta_{Cl},\theta_{Gu}|\boldsymbol{y}\right)\propto p\left(\boldsymbol{y}|\boldsymbol{w},\boldsymbol{d},\Gamma,\theta_{Cl},\theta_{Gu}\right)
p\left(\boldsymbol{d}|\boldsymbol{w},\Gamma,\theta_{Cl},\theta_{Gu}\right)p\left(\boldsymbol{w}\right)p\left(\Gamma\right)p\left(\theta_{Cl}\right)
p\left(\theta_{Gu}\right)\label{eq:posterior mixture copula}
\end{equation}
with
\[
p\left(\boldsymbol{y}|\boldsymbol{w},\boldsymbol{d},\Gamma,\theta_{Cl},\theta_{Gu}\right)=\prod_{i=1}^{n}\left[c_{G}\left(\boldsymbol{u};\Gamma\right)\right]^{d_{i1}}
\left[c_{Cl}\left(\boldsymbol{u};\theta_{Cl}\right)\right]^{d_{i2}}\left[c_{Gu}\left(\boldsymbol{u};\theta_{Gu}\right)\right]^{d_{i3}},
\]
and
\begin{equation}
p\left(\boldsymbol{d}|\boldsymbol{w},\Gamma,\theta_{Cl},\theta_{Gu}\right)=p\left(\mathbf{d}|\boldsymbol{w}\right)=
\prod_{i=1}^{n}\prod_{k=1}^{K}w_{k}^{d_{ik}}=\prod_{k=1}^{K}w_{k}^{n_{k}},
\end{equation}
where $n_{k}=\sum_{i=1}^{n}I\left(d_{ik}=1\right)$ and $I\left(d_{ik}=1\right)$
is an indicator variable which is equal 1 if observation
$i$ belongs to the $k$-th component of the copula mixture model,
and is $0$ otherwise. We use a Dirichlet prior for $\boldsymbol{w}$,
$p\left(\boldsymbol{w}\right)=Dirichlet\left(\boldsymbol{\phi}\right)$, which is
defined as
\begin{equation}
p\left(\boldsymbol{w}\right)\propto w_{1}^{\phi_{1}-1}...w_{3}^{\phi_{3}-1}. \label{eq:weight prior}
\end{equation}
The Dirichlet distribution is the common choice in Bayesian mixture
modeling since it is a conjugate of the multinomial distribution
\citep{Diebold1994} . We use the gamma density $G(\alpha,\beta)$ as the prior distribution for $\theta_{Cl}$ and $\theta_{Gu}$.
The hyperparameters
in the prior PDFs are chosen so that the priors are uninformative.
We use a Metropolis within Gibbs sampling algorithm to draw observations
from the joint posterior PDF defined in equation~\eqref{eq:posterior mixture copula}
and use the resulting MCMC draws to estimate the quantities required for inference.
The relevant conditional posterior PDFs are now specified.
The conditional posterior probability that the $i$th observation
comes from the $k$th component in the copula mixture model is
\begin{equation}
p\left(d_{ik}|\boldsymbol{w},\Gamma,\theta_{Cl},\theta_{Gu},\boldsymbol{y}\right)=\frac{p_{ik}}{p_{i1}+...+p_{i3}}, \label{eq:eq 17}
\end{equation}
where $p_{i1}=w_{1}c_{G}\left(\boldsymbol{u};\Gamma\right)$, $p_{i2}=w_{2}c_{Cl}\left(\boldsymbol{u};\theta_{Cl}\right)$,
and $p_{i3}=w_{3}c_{Gu}\left(\boldsymbol{u};\theta_{Gu}\right)$ for
$k=1,2,3$.
The conditional posterior PDF for the mixture weights $\boldsymbol{w}$
is the Dirichlet PDF
\begin{equation}
p(\boldsymbol{w}|\mathbf{d},\Gamma,\theta_{Cl},\theta_{Gu},\boldsymbol{y})=D(\boldsymbol{\phi}+\boldsymbol{n}), \label{eq:eq 28}
\end{equation}
where $\mathbf{n}=(n_{1},...,n_{k})'$ and $\boldsymbol{\phi}=(\phi_{1},...,\phi_{K})'$.
The conditional posterior PDF for the Gaussian copula parameter matrix $\Gamma$
is
\begin{equation}
p\left(\Gamma|\boldsymbol{y},\mathbf{d},\theta_{Cl},\theta_{Gu},\boldsymbol{w}\right)=\prod_{i\in d_{i1}=1}c_{G}\left(\boldsymbol{u};\Gamma\right)p\left(\Gamma\right). \label{eq:gaussian conditiona}
\end{equation}
The conditional posterior PDF for the Clayton copula parameter $\theta_{Cl}$
is
\begin{equation}
p\left(\theta_{Cl}|\boldsymbol{y},\mathbf{d},\Gamma,\theta_{Gu},\boldsymbol{w}\right)=\prod_{i\in d_{i1}=2}c_{Cl}\left(\boldsymbol{u};\theta_{Cl}\right)p\left(\theta_{Cl}\right). \label{eq:clayton conditional}
\end{equation}
The conditional posterior PDF for the Gumbel copula parameter $\theta_{Gu}$
is
\begin{equation}
p\left(\theta_{Gu}|\boldsymbol{y},\mathbf{d},\Gamma,\theta_{Cl},\boldsymbol{w}\right)=\prod_{i\in d_{i1}=3}c_{Gu}\left(\boldsymbol{u};\theta_{Gu}\right)p\left(\theta_{Gu}\right). \label{eq:gumbel conditional}
\end{equation}
Generating the conditional posterior density for $\theta_{Cl}$ and
$\theta_{Gu}$ is not straightforward since the conditional posterior
densities for both $\theta_{Cl}$ and $\theta_{Gu}$ are not in a recognizable
form. We use a random walk Metropolis algorithm to draw
from the conditional posterior densities of both $\theta_{Cl}$
and $\theta_{Gu}$. The generation of the Gaussian copula matrix parameter $\Gamma$
is more complicated and is explained in the next section.
The full MCMC sampling scheme is
\begin{enumerate}
\item Set the starting values for $\boldsymbol{w}^{\left(0\right)}$, $\Gamma^{\left(0\right)}$,
$\theta_{Cl}^{\left(0\right)}$, and $\theta_{Gu}^{\left(0\right)}$
\item Generate $(\boldsymbol{w}^{\left(t+1\right)}|\mathbf{d}^{\left(t\right)},\Gamma^{\left(t\right)},\theta_{Cl}^{\left(t\right)},\theta_{Gu}^{\left(t\right)},\boldsymbol{y})$
from equation \eqref{eq:eq 28}
\item Generate $\left(\Gamma^{\left(t+1\right)}|\boldsymbol{y},\mathbf{d}^{\left(t+1\right)},\theta_{Cl}^{\left(t\right)},\theta_{Gu}^{\left(t\right)},\boldsymbol{w}^{\left(t+1\right)}\right)$
from equation \eqref{eq:gaussian conditiona}
\item Generate $\left(\theta_{Cl}^{\left(t+1\right)}|\boldsymbol{y},\mathbf{d}^{\left(t+1\right)},\Gamma^{\left(t+1\right)},\theta_{Gu}^{\left(t\right)},\boldsymbol{w}^{\left(t+1\right)}\right)$
from equation \eqref{eq:clayton conditional}
\item Generate $\left(\theta_{Gu}^{\left(t+1\right)}|\boldsymbol{y},\mathbf{d}^{\left(t+1\right)},\Gamma^{\left(t+1\right)},\theta_{Cl}^{\left(t+1\right)},\boldsymbol{w}^{\left(t+1\right)}\right)$
from equation \eqref{eq:gumbel conditional}
\item Set $t=t+1$ and return to step 2.
\end{enumerate}
Appendix~\ref{app_sampling_schemes} gives further details on the particulars of the sampling scheme. In particular, it describes how to write the distributions and densities of the Gaussian, Gumbel and Clayton copulas respectively and how to sample from them. It also details how to sample the correlation parameters of the Gaussian copula and summarizes how the one-margin at a time latent variable simulation works.
\section{Application to Individual Income Dynamics\label{S: application to individual income dynamics}}
Longitudinal or panel datasets, such as the Panel Study of Income Dynamics
(PSID), the British Household Panel Survey (BHPS), and the Household Income and
Labour Dynamics Survey in Australia (HILDA) are increasingly used for
assessing income inequality, mobility, and poverty over time.
The income data from these surveys for different years are correlated
due to the nature of panel studies. For such correlated samples, the
standard approach of fitting univariate models to income distributions
for different years may give rise to misleading results. The univariate
approach treats the income distribution over different years as independent
and ignores the dependence structure between incomes for different years. It does not
take into account that those who earned a high income in one year are more likely to earn
a high income in subsequent years and vice versa. A
common way to address this problem is to use a multivariate income
distribution model that takes into account the dependence between
incomes for different years.
The presence of dependence in a sample of incomes from panel datasets
has rarely been addressed in the past. Only recently, \cite{Vinh2010}
proposed using bivariate copulas to model income distributions for
two different years, using maximum likelihood estimation. However, in their applications, they do not take
into account the point mass occurring at zero income. Our methodology
is more general than \cite{Vinh2010}. We estimate a panel of incomes
from the HILDA survey from 2001 to 2013 using a finite mixture of Gaussian, Clayton,
and Gumbel copulas while taking into account the point mass occurring
at zero incomes. Once the parameters for the multivariate income distribution
have been estimated, values for various measures of inequality, mobility,
and poverty can be obtained. Our methodology is Bayesian which enables us to estimate
the posterior densities of the
parameters of the copula models and the inequality, mobility,
and poverty measures. In this example, we consider the \cite{Shorrocks1978a}
and \cite{Foster2009} indices for illustration purposes. Other inequality,
mobility, and poverty indices can be estimated similarly.
For other recent studies on income mobility dynamics, see also \cite{Bonhomme2009}.
Although a number of income related variables are available, we use
the imputed income series \_WSCEI in this example. This variable
contains the average individual weekly wage and salary incomes from
all paid employment over the period considered. It is reported before
taxation and governmental transfers. The income data were also adjusted
to account for the effects of inflation using the Consumer Price Index
data obtained from the Australian Bureau of Statistics, which is based
in 2010 dollars. From these data, a dependence sample was constructed
by establishing whether a particular individual had recorded an income
in all the years. Individuals who only recorded incomes in some
of the years being considered were removed. In addition, we also focus
our attention on individuals who are in the labor force (both employed
and unemployed). We found that 1745 individuals recorded an income
for all 13 years. Table~\ref{tab:Descriptive-statistics-for}
summarizes the distributions of real individual disposable income in Australia for the years 2001
- 2013 and shows that all income distributions
exhibit positive skewness and fat long right tails typical of
income distributions. If the ordering of the
distributions is judged on the basis of the means or the medians, the population
becomes better off as it moves from 2001 to 2013, except between the period
2006 and 2007. These effects are also confirmed by Figures~\ref{fig:Histogram-of-real} to
\ref{fig:Histogram-of-real-2} in appendix~\ref{app: extra empirical results}
\begin{sidewaystable}[]
\centering{}\caption{Descriptive statistics for real individual wages for Australia for
the years 2001 - 2013\label{tab:Descriptive-statistics-for}}
\begin{tabular}{cccccccccccccc}
\hline
& 2001 & 2002 & 2003 & 2004 & 2005 & 2006 & 2007 & 2008 & 2009 & 2010 & 2011 & 2012 & 2013\tabularnewline
\hline
\hline
Mean & 684 & 734 & 766 & 819 & 874 & 923 & 783 & 1067 & 1105 & 1128 & 1188 & 1215 & 1245\tabularnewline
Median & 616 & 673 & 712 & 753 & 803 & 852 & 702& 969 & 1003 & 1048 & 1051 & 1101 & 1100 \tabularnewline
Std. dev. & 551 & 591 & 568 & 645 & 674 & 668 & 694 & 788 & 825 & 869 & 990 & 916 & 950 \tabularnewline
skewness & 2.1 & 3.0 & 2.1 & 3.7 & 2.9 & 1.5 & 2.0& 2.0 & 1.8 & 2.0 & 3.7 & 1.9 & 1.5 \tabularnewline
kurtosis & 15.7 & 26.6 & 16.0 & 40.2 & 27.1 & 8.7 & 11.1& 12.7 & 11.7 & 15.7 & 37.5 & 13.0 & 7.4 \tabularnewline
\hline
\end{tabular}
\end{sidewaystable}
\subsection{Foster's (2009) Chronic Poverty Measures}
The measurement of chronic income poverty is important because it focuses
on those whose lack of income stops them from obtaining the ``minimum necessities
of life'' for much of their life course. Let $z\in \mathbb{R^{+}}$ be the
poverty line. It is the level of income/wages which is just sufficient for someone
to be able to afford the minimum necessities of life. For every $i=1,...,n$
and $t=1,...,T$, the row vector $\mathbf{y}_{i}=\left(y_{i1},...,y_{iT}\right)$
contains individual $i$'s incomes across time and the column vector
$\mathbf{y}_{*t}=\left(y_{1t},...,y_{nt}\right)^{'}$ contains the
income distribution at period $t$.
The measurement of chronic poverty is split into two steps:
an ``identification'' step and an aggregation step. The identification
function $\rho\left(\mathbf{y}_{i};z\right)$ indicates that individual
$i$ is in chronic poverty when $\rho\left(\mathbf{y}_{i};z\right)=1$,
while $\rho\left(\mathbf{y}_{i};z\right)=0$ otherwise. \cite{Foster2009} proposed an identification
method that counts the number of periods of poverty experienced by a particular
individual, $y_{it}<z$, and then expressed it as a fraction $d_{i}$
of the $T$ periods. The identification function $\rho_{\tau}\left(\mathbf{y}_{i};z\right)=1$
if $d_{i}\geq\tau$ and $\rho_{\tau}\left(\mathbf{y}_{i};z\right)=0$
if $d_{i}<\tau$.
The aggregation step combines the information on the chronically poor
people to obtain an overall level of chronic poverty in a given society.
We use the extension of univariate Foster, Greer and Thorbecke (FGT) poverty indices
of \citet{Foster:1984}.
These are given by
\[
FGT^{\alpha}\left(z\right)=\frac{1}{n}\sum_{i=1}^{n}g_{i}^{\alpha}\ ,
\]
where $g_{i}^{\alpha}=0$ if $y_{i}>z$ and $g_{i}^{\alpha}\left(z\right)=\left(\frac{z-y_{i}}{z}\right)^{\alpha}$
if $y_{i}\leq z$, and $\alpha$ measures inequality aversion. The
FGT measure when $\alpha=0$ is called the headcount ratio,
when $\alpha=1$ it is called the poverty gap index and when
$\alpha=2$ it is called the poverty severity index. \cite{Foster2009} proposed duration
adjusted FGT poverty indices: duration adjusted headcount ratio and duration
adjusted poverty gap. Following \cite{Foster2009}, we define the normalized
gap matrix as $G^{\alpha}\left(z\right):=\left[g_{it}^{\alpha}\left(z\right)\right]$,
where $g_{it}^{\alpha}\left(z\right)=0$ if $y_{it}>z$ and $g_{it}^{\alpha}\left(z\right)=\left(\frac{z-y_{it}}{z}\right)^{\alpha}$
if $y_{it}\leq z$. Then, identification is incorporated into the
censored matrix $G^{\alpha}\left(z,\tau\right)=\left[g_{it}^{\alpha}\left(z,\tau\right)\right]$,
where $g_{it}^{\alpha}\left(z,\tau\right)=g_{it}^{\alpha}\left(z\right)\rho_{\tau}\left(\mathbf{y}_{i};z\right)$.
The entries for the non-chronically poor are censored to zero, while the
entries for the chronically poor are left unchanged. When $\alpha=0$,
the measure becomes the duration adjusted headcount ratio and is
the mean of $G^{0}\left(z,\tau\right)$, and when $\alpha=1$,
the measure becomes the duration adjusted poverty gap, and is given
by the mean of $G^{1}\left(z,\tau\right)$.
\subsection{Shorrocks (1978a) Income Mobility Measures\label{SS: income mobility measures}}
The measurement of income mobility focuses on how individuals' income
changes over time. Many mobility measures have been developed and applied
to empirical data to describe income dynamics; see \cite{Shorrocks1978a},
\cite{Shorrocks1978b}, \cite{Formby2004}, \cite{Dardanoni1993},
\cite{Fields1996}, \cite{Maasoumi1986}, and references therein.
However, statistical inference on income mobility has been largely
neglected in the literature. Only recently, some researchers have
developed statistical inference procedures for the measurement of
income mobility \citep{Biewen2002, Maasoumi2001, Formby2004}.
Here, we show that our approach can be used to obtain the posterior densities
of mobility measures which can then be used for making inference on income mobility.
\cite{Shorrocks1978a} proposed a measure of income mobility that
is based on transition matrices. Following \cite{Formby2004}, we
consider the joint distribution of two income variables $y_{1}$
and $y_{2}$ with a continuous CDF $F\left(y_{1},y_{2}\right)$. This
distribution function captures all the transitions between $y_{1}$ and $y_{2}$.
In this application, we consider the mobility between two points in
time. The movement between $y_{1}$ and $y_{2}$ is described by a
transition matrix. To form the the transition matrix from $F\left(y_{1},y_{2}\right)$,
we need to determine the number of, and boundaries between, income classes.
Suppose there are $m$ classes in each of the income variables
and the boundaries of these classes are $0<\tau_{1}^{y_{1}}<...<\tau_{m-1}^{y_{1}}<\infty$
and $0<\tau_{1}^{y_{2}}<...<\tau_{m-1}^{y_{2}}<\infty$. The resulting
transition matrix is denoted $P=\left[p_{ij}\right]$. Each element
$p_{ij}$ is a conditional probability that an individual moves to
class $j$ of income $y_{2}$ given that they are initially in class
$i$ with income $y_{1}$. It is defined as
\[
p_{ij}:=\frac{\Pr\left(\tau_{i-1}^{y_{1}}\leq y_{1}<\tau_{i}^{y_{1}}\;and\;\tau_{j-1}^{y_{2}}\leq y_{2}<\tau_{j}^{y_{2}}\right)}{\Pr\left(\tau_{i-1}^{y_{1}}\leq y_{1}<\tau_{i}^{y_{1}}\right)}\ ,
\]
where $\Pr\left(\tau_{i-1}^{y_{1}}\leq y_{1}<\tau_{i}^{y_{1}}\right)$
is the probability that an individual falls into income class $i$
of $y_{1}$.
A Mobility measure $M\left(P\right)$ can be defined as a function
of the transition matrix $P$. We say that a society with transition matrix
$P_{1}$ is more mobile than one with transition matrix $P_{2}$, according
to mobility measure $M\left(P\right)$, if and only if $M\left(P_{1}\right)>M\left(P_{2}\right)$.
We consider a mobility measure developed by \cite{Shorrocks1978a} and defined as
\[
M_{1}\left(P\right):=\frac{m-\sum_{i=1}^{m}p_{ii}}{m-1};
\]
$M_{1}$ measures the average probability across all classes that
an individual will leave his initial class in the next period.
\subsection{Empirical Analysis\label{SS: empirical analysis}}
This section discusses the results from the analysis
of the real individual wages data after estimating the proposed multivariate
income distribution model using a Bayesian approach.
The univariate income distribution is usually modeled using Dagum or Singh-Maddala distributions \citep{Kleiber1996}.
In this example, the marginal income distribution is modeled using empirical distribution function,
for simplicity. It is straightforward to extend the MCMC sampling scheme in Section~\ref{S: estimation and algorithms}
to estimate both marginal
and joint parameters as in \cite{Pitt2006} and \cite{Smith2012}.
First, we present
the model selection results and the estimated parameters of the copula models.
To select the best copula model, we use the $DIC_{3}$ criterion of \cite{Celeux:2006}
and the cross-validated log predictive score (LPDS) \citep{Good:1952, Geisser1980}. The $DIC_{3}$ criterion is defined as
\[
DIC_{3}:=-4\E_{\theta}\left(\log p\left(\boldsymbol{y}|\theta\right)|\boldsymbol{y}\right)+2\log\widehat{p}\left(\boldsymbol{y}\right),
\]
where $\widehat{p}\left(\boldsymbol{y}\right)=\prod_{i=1}^{n}\widehat{p}\left(y_{i}\right)$.
We next define the $B$-fold cross-validated LPDS.
Suppose that the dataset ${\cal D}$ is split into roughly $B$ equal
parts ${\cal D}_{1},...,{\cal D}_{B}$. Then, the $B-$fold cross validated LPDS is
defined as
\[
LPDS\left(\widehat{p}\right):=\sum_{j=1}^{B}\sum_{y_{j}\in{\cal D}_{j}}\log\widehat{p}\left(\mathbf{y}_{j}|{\cal D}\setminus{\cal D}_{j}\right)
\]
In our work we take $B = 5 $.
Table \ref{tab:Model-Selection-} shows that the best model, according to both criteria, is the
mixture of Gaussian, Clayton, and Gumbel copulas.
We estimate the best model with an initial burnin period of 10000
sweeps and a Monte Carlo sample of 10000 iterates. Next, we use the iterates from the best
model to estimate transition probabilities from 0 to positive
wages and from positive wages to zero, Spearman's correlation coefficient, and the mobility
and poverty measures, by averaging over the posterior distribution of the parameters.
Table~\ref{tab:Some-Estimated-Parameters copula} shows some of the estimated
parameters and corresponding 95\% credible intervals for the chosen
copula mixture model. The parameters and their 95\% credible intervals
are quite tight, indicating that the parameters are well estimated.
It is clear that there are significant differences in the estimated parameters
by taking into account the point mass at zero wages compared to the parameters
estimated by not taking into account this point mass. The estimated
mixture weight parameters show that the Gaussian copula has the
highest weight, followed by the Clayton and Gumbel copulas. As the weight
of the Clayton copula is higher than of the Gumbel copula, it implies that
there are more people with lower tail dependence than upper tail dependence.
This may coincide with a relatively higher degree of income
mobility amongst high income earners.
\begin{table}[H]
\caption{Model Selection of the copula to model 13 years of income distribution with point mass at zero incomes \label{tab:Model-Selection-} }
\centering{}%
\begin{tabular}{ccc}
\hline
Model & $DIC_{3}$ & LPDS-CV\tabularnewline
\hline
Clayton & $-1.21\times10^{4}$ & $6.03\times10^{3}$\tabularnewline
Gumbel & $-1.75\times10^{4}$ & $4.95\times10^{3}$\tabularnewline
Gaussian & $-2.13\times10^{4}$ & $4.29\times10^{4}$\tabularnewline
Mixture (Gaussian, Clayton) & $-2.86\times10^{4}$ & $5.63\times10^{4}$\tabularnewline
Mixture (Gaussian, Gumbel) & $-2.83\times10^{4}$ & $5.54\times10^{4}$\tabularnewline
Mixture (Clayton, Gumbel) & $-1.68\times10^{4}$ & $3.31\times10^{4}$\tabularnewline
Mixture (Gaussian, Clayton, Gumbel) & $-2.89\times10^{4\star}$ & $5.68\times10^{4\star}$\tabularnewline
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\caption{Some of the estimated parameters of the mixture of the Gaussian, Gumbel and Clayton copulas to model 13 years of income distributions. The 95\% credible intervals are in brackets
\label{tab:Some-Estimated-Parameters copula}}
\centering{}%
\begin{tabular}{ccc}
\hline
Parameters & \multicolumn{1}{c}{Copula (Point Mass)} & \multicolumn{1}{c}{Copula (No Point Mass)}\tabularnewline
\hline
\hline
$\theta_{Cl}$ & $\underset{\left(0.12,0.18\right)}{0.15}$ & $\underset{\left(0.29,0.37\right)}{0.33}$\tabularnewline
$\theta_{Gu}$ & $\underset{\left(1.84,2.06\right)}{1.94}$ & $\underset{\left(2.23,2.45\right)}{2.33}$\tabularnewline
$w_{1}$ & $\underset{\left(0.64,0.69\right)}{0.66}$ & $\underset{\left(0.60,0.65\right)}{0.62}$\tabularnewline
$w_{2}$ & $\underset{\left(0.19,0.24\right)}{0.21}$ & $\underset{\left(0.21,0.26\right)}{0.23}$\tabularnewline
\hline
\end{tabular}
\end{table}
Tables~\ref{tab:Transition-Probability aug-2} and
\ref{tab:Transition-Probability aug-1-1} in Appendix~\ref{app: extra empirical results}
present the estimates
of the transition probabilities from 0 to positive wages and from positive
to 0 wages. The estimates of the transition probabilities seem to be
close to their sample (non-parametric) counterparts. The estimates of
transition probabilities from 0 to positive wages are similar
(0.39-0.49) in the period from 2001-2006. Similarly, the
estimates are similar in the period 2008-2013 (0.34-0.38).
However, there are higher estimates for the period 2006-2007 and 2007-2008 (0.83 and 0.87,
respectively). Similar
results are observed for the transition probabilities from positive to zero wages.
The estimates of the transition probabilities are
roughly the same between the periods 2001-2006 and 2008-2013. There
are higher estimates for the period 2006-2007 and 2007-2008. This
phenomenon may indicate that there is very high income mobility between
2006-2007 and 2007-2008. Note that the model that does not take into account the point masses at zero cannot give us the estimate of transition probabilities.
Tables \ref{tab:Spearman-Rho-Dependence} and \ref{tab:Shorrocks-(1978a)-Mobility-1}
show the estimate of Spearman's rho dependence and \cite{Shorrocks1978a}
mobility measure. We can see from these two measures that there are
very high values of the mobility measure and very low values of Spearman's
rho dependence measure between 2006-2007 and 2007-2008. This confirms
our previous analysis that in the period 2006-2008 there is very
high mobility between income earners. Table \ref{tab:Foster's-Chronic-Poverty}
shows the estimates of Foster's chronic poverty measures: duration
adjusted headcount ratio and duration adjusted poverty gap. The two
measures indicate that the chronic poverty is significantly lower
in the 2007-2013 period compared to the 2001-2006 period. The standard of living
in Australia is higher in the period 2007-2013 compared to the period
2001-2006. Furthermore, we can see that the estimates of Spearman's
rho dependence, mobility, and chronic measures are different between
the estimates that take into account the point masses and the estimates
that do not take into account the point masses at zero wages. Figure
\ref{fig:Headcount-Posterior-Density} shows the posterior densities
of duration adjusted headcount ratio for the years 2007-2013 for the
two estimates. The figure shows that the posterior densities almost do
not overlap, indicating that the two estimates are significantly different.
Therefore, whenever the point masses are present, it is strongly recommended
to incorporate them into the model to guard against biased estimates.
\begin{table}[H]
\caption{Estimates of the Spearman rho dependence measure of the mixture of the Gaussian, Gumbel and Clayton copulas and 95\% credible intervals (in brackets) \label{tab:Spearman-Rho-Dependence}}
\centering{}%
\begin{tabular}{ccc}
\hline
Period & \multicolumn{1}{c}{Copula (Point Mass)} & \multicolumn{1}{c}{Copula (No Point Mass)}\tabularnewline
\hline
\hline
2001-2002 & $\underset{\left(0.684,0.722\right)}{0.703}$ & $\underset{\left(0.723,0.757\right)}{0.740}
$\tabularnewline
2002-2003 & $\underset{\left(0.700,0.737\right)}{0.719}$ & $\underset{\left(0.726,0.759\right)}{0.743
}$\tabularnewline
2003-2004 & $\underset{\left(0.702,0.739\right)}{0.721}$ & $\underset{\left(0.727,0.759\right)}{0.743}
$\tabularnewline
2004-2005 & $\underset{\left(0.7040,0.741\right)}{0.723}$ & $\underset{\left(0.730,0.763\right)}{0.747}
$\tabularnewline
2005-2006 & $\underset{\left(0.708,0.745\right)}{0.727}$ & $\underset{\left(0.733,0.766\right)}{0.750}
$\tabularnewline
2006-2007 & $\underset{\left(-0.028,0.068\right)}{0.020}$ & $\underset{\left(-0.020,0.086\right)}{0.030}$\tabularnewline
2007-2008 & $\underset{\left(-0.023,0.073\right)}{0.025}$ & $\underset{\left(-0.013,0.093\right)}{0.037}$\tabularnewline
2008-2009 & $\underset{\left(0.706,0.744\right)}{0.725}$ & $\underset{\left(0.733,0.766\right)}{0.7500}$\tabularnewline
2009-2010 & $\underset{\left(0.716,0.753\right)}{0.735}$ & $\underset{\left(0.741,0.775\right)}{0.758}
$\tabularnewline
2010-2011 & $\underset{\left(0.720,0.758\right)}{0.740}$ & $\underset{\left(0.747,0.781\right)}{0.764}
$\tabularnewline
2011-2012 & $\underset{\left(0.718,0.755\right)}{0.737}$ & $\underset{\left(0.745,0.778\right)}{0.762}
$\tabularnewline
2012-2013 & $\underset{\left(0.714,0.752\right)}{0.733}$ & $\underset{\left(0.742,0.776\right)}{0.759}
$\tabularnewline
\hline
\end{tabular}
\end{table}
\begin{table}[H]
\caption{Estimates of Shorrocks (1978a) Mobility Measure ($m=5$) of the mixture of the Gaussian, Gumbel and Clayton copulas \label{tab:Shorrocks-(1978a)-Mobility-1}}
\centering{}%
\begin{tabular}{cccc}
\hline
Period & Non-Parametric & \multicolumn{1}{c}{Copula (Point Mass)} & \multicolumn{1}{c}{Copula (No Point Mass)}\tabularnewline
\hline
\hline
2001-2002 & $\underset{\left(0.367,0.466\right)}{0.414}$ & $\underset{\left(0.549,0.588\right)}{0.569}$ & $\underset{\left(0.501,0.534\right)}{0.518}$\tabularnewline
2002-2003 & $\underset{\left(0.361,0.461\right)}{0.411}$ & $\underset{\left(0.508,0.543\right)}{0.526}$ & $\underset{\left(0.484,0.516\right)}{0.499}$\tabularnewline
2003-2004 & $\underset{\left(0.324,0.409\right)}{0.366}$ & $\underset{\left(0.483,0.516\right)}{0.500}$ & $\underset{\left(0.463,0.495\right)}{0.479}$\tabularnewline
2004-2005 & $\underset{\left(0.341,0.418\right)}{0.380}$ & $\underset{\left(0.473,0.506\right)}{0.489}$ & $\underset{\left(0.450,0.480\right)}{0.465}$\tabularnewline
2005-2006 & $\underset{\left(0.352,0.427\right)}{0.392}$ & $\underset{\left(0.468,0.5000\right)}{0.484}$ & $\underset{\left(0.444,0.475\right)}{0.459}$\tabularnewline
2006-2007 & $\underset{\left(0.974,1.019\right)}{0.996}$ & $\underset{\left(0.957,0.980\right)}{0.969}$ & $\underset{\left(0.878,0.938\right)}{0.918}$\tabularnewline
2007-2008 & $\underset{\left(0.959,1.015\right)}{0.987}$ & $\underset{\left(0.921,0.945\right)}{0.933}$ & $\underset{\left(0.843,0.906\right)}{0.885}$\tabularnewline
2008-2009 & $\underset{\left(0.384,0.441\right)}{0.411}$ & $\underset{\left(0.493,0.526\right)}{0.510}$ & $\underset{\left(0.465,0.495\right)}{0.480}$\tabularnewline
2009-2010 & $\underset{\left(0.350,0.409\right)}{0.380}$ & $\underset{\left(0.482,0.516\right)}{0.500}$ & $\underset{\left(0.449,0.481\right)}{0.465}$\tabularnewline
2010-2011 & $\underset{\left(0.351,0.411\right)}{0.381}$ & $\underset{\left(0.463,0.500\right)}{0.481}$ & $\underset{\left(0.424,0.456\right)}{0.440}$\tabularnewline
2011-2012 & $\underset{\left(0.353,0.405\right)}{0.380}$ & $\underset{\left(0.475,0.510\right)}{0.492}$ & $\underset{\left(0.437,0.469\right)}{0.453}$\tabularnewline
2012-2013 & $\underset{\left(0.339,0.395\right)}{0.365}$ & $\underset{\left(0.499,0.536\right)}{0.517}$ & $\underset{\left(0.458,0.493\right)}{0.475}$\tabularnewline
\hline
\end{tabular}
\end{table}
\begin{sidewaystable}
\caption{Estimates of Foster's chronic poverty measure of the mixture of the Gaussian, Gumbel and Clayton copulas with 95\% credible intervals (in brackets)\label{tab:Foster's-Chronic-Poverty}}
\centering{}%
\begin{tabular}{ccccc}
\hline
Measure & Period & Non-Parametric & \multicolumn{1}{c}{Copula (Point Mass)} & \multicolumn{1}{c}{Copula (No Point Mass)}\tabularnewline
\hline
\hline
adj. headcount & 2001-2006 & $\underset{\left(0.193,0.229\right)}{0.211}$ & $\underset{\left(0.187,0.197\right)}{0.192}$ & $\underset{\left(0.197,0.205\right)}{0.201}$\tabularnewline
adj. headcount & 2007-2013 & $\underset{\left(0.120,0.149\right)}{0.135}$ & $\underset{\left(0.123,0.130\right)}{0.126}$ & $\underset{\left(0.131,0.138\right)}{0.135}$\tabularnewline
adj. poverty gap & 2001-2006 & $\underset{\left(0.123,0.150\right)}{0.137}$ & $\underset{\left(0.130,0.137\right)}{0.134}$ & $\underset{\left(0.137,0.144\right)}{0.141}$\tabularnewline
adj. poverty gap & 2007-2013 & $\underset{\left(0.095,0.119\right)}{0.108}$ & $\underset{\left(0.101,0.107\right)}{0.104}$ & $\underset{\left(0.108,0.114\right)}{0.111}$\tabularnewline
\hline
\end{tabular}
\end{sidewaystable}
\begin{figure}[H]
\caption{Estimated headcount posterior densities based on including (left density-blue line) and not including point masses (right density-orange line) at 0 (2007-2013)\label{fig:Headcount-Posterior-Density}}
\begin{centering}
\includegraphics[width=15cm,height=10cm]{headcount.png}
\par\end{centering}
\end{figure}
\section{Conclusion and discussion \label{S: conclusion}}
\textcolor{blue}{The paper shows how to define and derive the density of the observations of a general copula model when
some of the marginals are discrete, some are continuous {\em and} some of the marginals are a mixture of discrete and continuous components.
This is done by carefully defining the likelihood as the
density of the observations with respect to a mixed measure
and allows us to define the likelihood for general copula models and hence carry out likelihood based inference.
Our work extends in a very general way the current literature on likelihood based inference which focuses
on copulas where each marginal is either discrete or continuous. The inference in the paper is Bayesian and we show how to construct an
efficient MCMC scheme to estimate functionals of the posterior distribution. Although our discussion and examples focus on Gaussian and Archimedean copulas, our treatment is quite general and can be applied as long as it is possible to compute certain marginal and conditional copulas either in closed-form or numerically.}
\textcolor{blue}{Our article can be extended in the following directions.
First, using our definition of the likelihood also enables maximum likelihood type inference using, for example, simulated EM or simulated maximum likelihood.
Second, copulas based on pair-copula constructions \citep[e.g.][]{aas2009pair} or vine copulas \citep[e.g.][]{bedford2002vines} lend themselves well to our approach because
the methods in this paper apply to arbitrary copulas with the only requirement that it is possible
to write down several conditional marginal copulas and copula densities and being able to compute those either analytically or numerically.
Third, by using pseudo marginal methods \citep[e.g.][]{Andrieu:2010}, our methodology can also be extended to the case where the case where the likelihood of the copula can only be estimated unbiasedly, rather than evaluated. We leave all such extensions to future work.}
Our article illustrates the methodology and algorithms by applying them to
estimate a multivariate income dynamics model.
Examples of further possible applications arise from any setup where one or more of the
following variables are present: wages (where there are points of probability mass at the
minimum wage) individual sales figures, where there is a \textcolor{blue}{point of probability mass} at 0 (many
individuals deciding not to purchase) and a smooth distribution above that
point (corresponding to a continuum of price figures). Another
interesting potential application is to extend the general truncated/censored
variable models in econometrics to a copula framework, e.g., for multivariate tobit and
sample selection models.
\section*{Acknowledgement}
We would like to thank two anonymous referees and the associate editor for suggestions that helped improve the clarity of the paper.
The research of David Gunawan and Robert Kohn was partially supported by an Australian Research Council Discovery Grant
DP150104630.
\bibliographystyle{asa}
|
\section*{Acknowledgements}
This work was supported by an AFOSR grant FA9550-09-0174 to JNK, and by a NIH training grant 5T90DA03243602, as well as NSF grant DMS-1122106 and a Simons Fellowship in Mathematics to ESB.
SLB acknowledges support from the UW department of Mechanical Engineering and the eScience institute.
We thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and support.
\vspace{-.25in}
\bibliographystyle{Science}
|
\section{Introduction}
The investigation of reset systems was started more than fifty years ago with the seminal work of Clegg \cite{clegg}, and \textcolor{black}{carried on with} a series of works by Horowitz and coworkers (\cite{KriHor, HorRos}). The main motivation for the study of reset systems arises from reset control (\cite{BanBarBook}), since reset compensation may achieve fast and robust control solutions for problems under linear limitations, which as it is well-known are particularly severe in the case of control systems with time-delays (\cite{astrom}). A large number of works have shown the advantages of reset control over linear control (\cite{beker2001,davo,moreno, BanBarBook}).
\textcolor{black}{The term {\em reset system} was first introduced by Hollot, Chait, and coworkers (\cite{BekHoll}) to denote a ``linear and time invariant system with (state-dependent) mechanisms and laws to reset their states to zero". Two distinctive characteristics of reset systems are that the resetting law is state-dependent, and that (some) states are reset to zero. Therefore, reset systems can be considered as a special type of impulsive/hybrid systems, in which the system state (or a part of it) is instantaneously zeroed out at those instants in which the system solution intersects some reset set. On the other hand, time-delay systems (see the monographs \cite{hale,gu,F2014}, or for example the works \cite{fridman,Cao2004,Cao2006,GP2006,SG2013}) are a natural target for reset control. As a result, time-delay reset systems have a clear interest in control practice and have been an active research topic in the last decade (\cite{BanBar,BarBan, BanBarBook}). \\
In addition, it should be noted that impulsive systems (without time-delays) have been a very active research topic in different areas of mathematics and systems theory (\cite{BS1989,LBS1989,HCN2006,MHD2008}), where the research effort has been concentrated in systems with impulses at fixed/variable instants. However, as it has been discussed in (\cite{BanBarBook, r_BM2012}) reset systems (with a time invariant base system) are a special case of autonomous impulsive systems, a much less developed research topic in the literature. On the other hand, time-delay impulsive systems, or more specifically impulsive functional differential equations, have been investigated in a number of works. Again, most of the research effort has been concentrated in systems with impulses at fixed instants (\cite{BL2000,delasen,impulsivo2004,impulsivo2007-2,impulsivo2008,SMZ2005}), and to a lesser extent in systems with impulses at variable times (\cite{impulsivo2007-3}). Note that the vast majority of stability results for impulsive functional differential equations uses Lyapunov-Razumikhin techniques (see \cite{S2009} and references therein).}
This work is focused on the internal stability analysis of time-delay reset systems by using a Lyapunov-Krasovskii approach, with the goal of obtaining conditions formulated by using Lyapunov-Krasovskii functionals. \textcolor{black}{To the knowledge of authors, all the previous published results from the impulsive functional differential equations literature are restricted to systems with impulses at fixed times, and thus they are not applicable to our case, or are based on Lyapunov-Razumikhin techniques. In spite of that,} the Lyapunov-Krasovskii approach has been already investigated in the area of reset systems;
in particular, a delay-independent condition is obtained in \cite{BanBar} for reset control systems, and, in addition, extension to delay-dependent conditions is given in \cite{BarBan}, and more recently in \cite{Pri}, considering a more general resetting law (anticipative reset). Also, quadratic stability of time-delay reset control systems with uncertainty in the resetting law has been analyzed in \cite{Guo}. In general, a previous work \cite{DavBan} suggests that it is necessary a deeper analysis of Lyapunov-Krasovskii functionals, focusing into the necessity of obtaining less restrictive conditions. A first attempt to obtain less restrictive conditions is \cite{DavBan}, where it has been proposed criteria based on bounded increments of the functional after the reset instants. In addition, input-output stability has been investigated in \cite{MerDav}, based on the previous Lyapunov-Krasovskii results, and in \cite{MerCar}, based on passivity properties of reset systems \cite{CarBanArj}, and the IQC framework.
To the authors knowledge, all the previous published results are about reset systems with a linear and time-invariant (LTI) base system, and most of them are based on the existence of a Lyapunov-Krasovskii theorem; and therefore, the proof of that theorem has been only sketched, making the generalization to nonlinear and time-varying systems challenging.
\textcolor{black}{On the other hand, some preliminary works \cite{teel,teel2, r_BPTZ2014} have extended the hybrid inclusion model developed in \cite{c13} to investigate hybrid systems with time-delays, based on a generalized concept of solutions.
In addition, the results in \cite{teel2} provide sufficient conditions for the stability analysis of hybrid systems with time-delays, using Lyapunov-Razumikhin functions, but application to reset control system is still unexplored. On the other hand, although \cite{r_BPTZ2014} approaches the case of reset systems, it deals with a restrictive class of reset systems in which the time-delay only affects a part of the state and thus its applicability is very limited in practice.}
The main result of this work is the development of a Lyapunov-Krasovskii theorem for reset systems, \textcolor{black}{in which the
resetting law is generalized in the sense that the state may be reset to a non-zero value after a reset action, and in addition the base system is nonlinear and time-varying.} In this way, this work is devoted to provide a formal and complete proof, and then to analyze reset systems with a LTI base system as a particular case.
The paper is structured as follows. After formally stating the problem in Section \ref{section:preliminares}, the main stability result is given in Section \ref{section:mainresult}. In Section \ref{section:application}, two application cases of the stability result are shown; firstly, a general reset system with a LTI base system and a single time-delay; and secondly, a particular reset system with a nonlinear and time-varying base system. The work concluded in Section \ref{section:conclusion} with some final remarks.\\
{\em Notation:} $\mathds{R}$ is the set of real numbers, $\mathds{R}^+$ is the set of non-negative real numbers, $\mathds{R}^n$ is the n-dimensional euclidean space, where $\|\mathbf{x}\|$ is the euclidean norm for $\mathbf{x} \in \mathds{R}^n$, and $({\bf x}, {\bf y })$, with column vectors ${\bf x} \in \mathds{R}^n$ and ${\bf y} \in \mathds{R}^m$, denotes the column vector $ \left( \ {\bf x}^\top \ \ {\bf y}^\top \ \right)^\top$. ${\cal PC}([a,b],F)$ is the set of piecewise continuous functions from $[a,b]$ to $F$, that is the set of functions that are continuous on $[a,b]$ except in a finite number of points $(t_k)_{k=1}^N$, and with a norm $\|\mathbf{\psi}\| = \sup_{\theta \in [a,b]} \| \psi(\theta)\|$. $\mathcal{R}(A)$ and $\mathcal{N}(A)$, for a matrix $A\in \mathds{R}^{n\times m}$, stand for the column space and the null space of A, respectively. $diag(A,B)$ is a block diagonal matrix composed by the matrices $A\in \mathds{R}^{n\times n}$ and $B\in \mathds{R}^{m\times m}$. For a symmetric matrix $A \in \mathds{R}^{n\times n}$, $\lambda_m(A)$ and $\lambda_M(A)$ stand for the minimum and maximum eigenvalue, respectively.
\section{Preliminaries}
\label{section:preliminares}
Consider a state-dependent time-delay reset system given by the impulsive
differential equation
\begin{equation}
\label{eq:resetsystem}
\left\{ \begin{array}{llll}
\dot{\mathbf{x}}(t) = \mathbf{f}(t,\mathbf{x}_t), & \mathbf{x}(t) \notin {\cal M}, \\
{\mathbf{x}}(t^+) = \mathbf{I}(t,\mathbf{x}_t), & \mathbf{x}(t) \in {\cal M},\\
\mathbf{x}(t) = \phi(t-t_0), & t \in [t_0-h,t_0],
\end{array}\right.
\end{equation}
\noindent where $t_0\in \mathds{R}^+$ is the initial instant, $\mathbf{x}(t) \in \mathds{R}^n$ is the system state at the instant $t \in [t_0-h,\infty)$, $\mathbf{x}_t \in {\cal PC}([-h,0],\mathds{R}^n)$ is the distributed state at the instant $t \in [t_0,\infty)$, that is $\mathbf{x}_t(\theta) = \mathbf{x}(t+\theta)$ for $\theta \in [-h,0]$, and the initial condition is a function $\phi \in {\cal PC} ([-h,0],\mathds{R}^n)$. Henceforth, it will be denoted ${\cal PC} ={\cal PC} ([-h,0],\mathds{R}^n)$ for notational simplicity.
It is considered that the reset is applied at instants $t_k \in [t_0,\infty)$ if $\mathbf{x}(t_k) \in {\cal M}$, \textcolor{black}{where ${\cal M} \subset \mathds{R}^n$ is the reset set.
It is assumed that reset instants are {\em well-posed}, that is} for any initial condition $\phi \in {\cal PC}$ there exists a finite or infinite sequence of well defined reset instants $\mathds{T}_\phi = (t_1, t_2, \cdots)$, such that they are distinct and satisfy $t_0<t_1 < t_2 < \cdots$; \textcolor{black}{and also that reset instants are Zeno-free,} that is if the reset instants sequence is infinite then $t_N \rightarrow \infty$ as $N \rightarrow \infty$. Otherwise, as well as in the case of free-delay impulsive dynamical systems (\cite{HCN2006}), pathological behaviors like beating and existence of Zeno solutions \textcolor{black}{may be present. A simple manner to guaranty that reset instants are well-posed and Zeno-free is to use time regularization} (see for example \cite{BanBarBook}), which means that reset instants satisfy $t_{k+1}-t_k> \Delta$, $k=0,1,2,\ldots$, \textcolor{black}{for some $\Delta >0$} (the initial instant $t_0$ is not a reset instant).
Thus, the existence and uniqueness of solutions of the reset system (\ref{eq:resetsystem}) follows from the existence and uniqueness of the following initial-value problem
\begin{equation}\label{eq:initialvalueproblem}
\left\{ \begin{array}{ll}
\dot{{\bf z}}(t) = \mathbf{f}(t,{\bf z}_t),\\
{\bf z}(t) = {\bf x}_{t_k}(t-t_k), & t\in[t_k-h, t_k), \\
{\bf z}(t_k) = \mathbf{I}(t_k,\mathbf{x}_{t_k}),
\end{array} \right.
\end{equation}
where $t\geq t_k$ and ${\bf x}_{t_k} \in \mathcal{PC}$ is the initial condition. It will be assumed that the initial value problem (\ref{eq:initialvalueproblem}) has a unique solution for $t \geq t_k$ (see Corollary 3.1 in \cite{BL2000}). For example, that there exist constants $M$, $N \in \mathds{R}^+$ such that $\|\mathbf{f}(t,\psi)\| \leq M+N\|\psi\|$ for all $(t,\psi) \in \mathds{R}^+ \times \mathcal{PC}$; \textcolor{black}{and that $\mathbf{f}$ is {\em locally Lipschitz}, that is for each compact set $\Omega \subset \mathds{R}^n$ there exists some constant $K \in \mathds{R}^+$ such that $\|\mathbf{f}(t,\psi_1) - \mathbf{f}(t,\psi_2) \| \leq K \|\psi_1 - \psi_2\|$ for all $t\in \mathds{R}^+$ and all $\psi_1, \psi_2 \in \mathcal{PC} ([-h,0],\Omega)$.}
Thus, the initial-value problem is well-posed and for every $(t_k,{\bf x}_{t_k}) \in \mathds{R}^+ \times \mathcal{PC}$, there exists a continuous and unique solution ${\bf z}(t_k,{\bf x}_{t_k})(t)$ for all $t \in [t_k,\infty)$.\\
Hence, the solution ${\bf x}(t_0,\phi)$ of (\ref{eq:resetsystem
) is made up of an initial condition $\phi$ and a sequence of continuous solution segments ${\bf z}(t_k,{\bf x}_{t_k})$, that is
\begin{equation}
{\bf x}(t_0,\phi)(t)=\left\{ \begin{array}{ll}
{\bf z}(t_0,\phi)(t), & t\in [t_0,t_1], \\
{\bf z}(t_1,{\bf x}_{t_1})(t), & t\in (t_1,t_2], \\
\multicolumn{2}{c}{\vdots} \\
{\bf z}(t_k,{\bf x}_{t_k})(t), & t\in (t_{k},t_{k+1}]. \\
\end{array} \right.
\end{equation}
If the sequence of reset instants is finite then ${\bf x}(t_0,\phi)(t)={\bf z}(t_N,{\bf x}_{t_N})(t)$, $t\in (t_{N},\infty)$. In addition, note that the solution is left-continuous with right limits, and there exists jump discontinuities at the reset instants $t_k$, $k = 1, 2, \cdots$,
that is the limits ${\bf x}(t_k^+)$ and ${\bf x}(t_k^-)$ exist, and ${\bf x}(t_k)={\bf x}(t_k^-)$.\\
Suppose $\mathbf{f}(t,0)=0$ and $\mathbf{I}(t,0)=0$ for all $t\in \mathds{R}^+$. The trivial solution ${\bf x}_t=0$ of system (\ref{eq:resetsystem
) (henceforth named {\em zero solution}) is said to be {\em stable} if for any $t_0\in \mathds{R}^+$ and $\epsilon>0$, there exists $\delta = \delta(t_0, \epsilon)$
such that $\| \phi \| < \delta$ implies $ \| {\bf x}(t)\| < \epsilon$ for $t \geq t_0$. In addition, the solution ${\bf x}_t=0$ is {\em uniformly stable} if $\delta=\delta(\epsilon)$. On the other hand, the zero solution is said to be {\em asymptotically stable} if it is stable and there exists $\delta_0=\delta_0(t_0) > 0$ such that $\lim_{t \rightarrow \infty} \| x(t)\| = 0$ whenever $\| \phi \|< \delta_0$. The solution is {\em uniformly asymptotically stable} if it is uniformly stable and there exists $\delta_0>0$ such that, for every $\eta>0$ there exists a $T=T(\eta)$ such that $\| \phi \| < \delta_0$ implies $\| {\bf x}(t) \| < \eta$ for $t \geq t_0 + T$ and for every $t_0\in \mathds{R}^+$; moreover, if $\delta_0$ can be an arbitrarily large finite number, then ${\bf x}_t=0$ is said to be {\em globally uniformly asymptotically stable}.\\
A function $f:\mathds{R}^+ \rightarrow \mathds{R}^+$ is said to be nondecreasing if $f(b)\geq f(a)$ for all $b>a$, where $a$, $b\in \mathds{R}^+$, if $f(b)> f(a)$ then it is said to be strictly increasing. In addition, $f$ is of class $\mathcal{K}$ if it is continuous, strictly increasing, and $f(0)=0$.\\
Let $V: \mathds{R}^+ \times {\cal PC} \rightarrow \mathds{R^+}$ be continuously differentiable with respect to all of its arguments, and let $\mathbf{x}(t_0,\phi)(t)$ be the solution of the system (\ref{eq:resetsystem
). Thus, $V(t,{\bf x}_t)$ only has (jump) discontinuities at $t\in \mathds{T}_\phi$. In addition, the upper right-hand derivative of $V$ along the solution $\mathbf{x}(t_0,\phi)$ is defined by
\begin{equation}
\dot{V}(t,{\bf x}_t)=\limsup_{\substack{
\epsilon \rightarrow 0,\:
\epsilon >0
}} \frac{V(t+\epsilon,{ \bf x}_{t+\epsilon}) - V(t,{ \bf x}_t)}{\epsilon}
\label{eq:}
\end{equation}
for all $t\in[t_0, \infty) \setminus \mathds{T}_\phi$. In addition, the increment of $V$ along the solution $\mathbf{x}(t_0,\phi)$ is defined by
\begin{equation}
\Delta V(t,\mathbf{x}_t) = V(t,\mathbf{I}(t,\mathbf{x}_t)) - V(t,\mathbf{x}_t)
\end{equation}
for any $t \in \mathds{T}_\phi$
\section{Main Result}
\label{section:mainresult}
In this section, sufficient conditions for stability of the reset system (1) are proposed as a Lyapunov-Krasovskii theorem, generalizing the basic result for retarded functional differential equations (without reset actions) (see for example \cite{hale,gu}). \textcolor{black}{The result takes inspiration from \cite{hale}}; in fact, since in general, for a given initial condition $\phi$, the system (\ref{eq:resetsystem}) may not have reset actions and thus $\mathds{T}_\phi =\varnothing$, the proof is identical in that case.
\\
{\bf Proposition 3.1}: Assume that \textcolor{black}{$\mathbf{f}$ is locally Lipschitz}, $u, v, w:\mathds{R}^+ \rightarrow \mathds{R}^+$ are continuous nondecreasing functions and in addition $u,v \in \mathcal{K}$. If there exists a (Lyapunov-Krasovskii) functional $V:\mathds{R}^+ \times {\cal PC} \rightarrow \mathds{R^+}$ such that
\begin{equation}\label{eq:cond1}
u(\|\psi(0)\|) \leq V(t,\psi) \leq v(\|\psi\|)
\end{equation}
\noindent for any $\psi \in \Omega= \{ \psi\in {\cal PC}: \|\psi\| < \gamma\}$ for some $\gamma>0$ and all $t \in \mathds{R^+}$, and that for every solution $\mathbf{x}(t_0,\phi)$ of the system (\ref{eq:resetsystem}), $V(t,\mathbf{x}_t)$ is continuous for all $t\geq t_0$ and $t_0 \in \mathds{R}^+$ except on the set $\mathds{T}_\phi$, and in addition
\begin{equation}\label{eq:cond2}
\dot{V}(t,{\bf x}_t) \leq -w(\|{\bf x}_t(0)\|), \hspace{1cm} \mathbf{{\bf x}}_t(0) \notin \cal{M},
\end{equation}
\begin{equation}\label{eq:cond3}
\Delta V(t,{\bf x}_t) \leq 0, \hspace{1cm} \mathbf{{\bf x}}_t(0) \in \cal{M},
\end{equation}
\noindent where $\dot{V}$ and $\Delta V$ are evaluated along the trajectories of (\ref{eq:resetsystem}) with ${\bf x}_t \in \Omega$, then the zero solution of (\ref{eq:resetsystem}) is uniformly stable. If $w(s) > 0$ for $s>0$ then the solution is uniformly asymptotically stable. In addition, if $\lim_{s \rightarrow \infty} u(s) = \infty$ and $\Omega={\cal PC}$, then it is globally uniformly asymptotically stable.\\
{\bf Proof}: \textcolor{black}{({\em Uniform stability)}} For a given $\epsilon > 0$, let set $\epsilon_1<\min(\epsilon, \gamma)$, then it can be found some $\delta=\delta(\epsilon_1)$, $0<\delta < \epsilon_1$, such that $v(\delta) < u(\epsilon_1)$. Suppose that ${\bf x}(t_0,\phi)(t)$ is the solution of (\ref{eq:resetsystem}) for $(t_0,\phi)\in \mathds{R}^+ \times {\cal PC}
. Therefore, ${\bf x}(t_0,\phi)(t)$ is continuous on $[t_0-h,\infty)\setminus \mathds{T}_\phi$, where $\mathds{T}_\phi = \{t_1, t_2, \cdots\}$ is the set of reset instants corresponding to the initial condition $\phi$, and $t_0 < t_1$. Now, we will prove that $\|{\bf x}(t_0,\phi)(t)\| < \epsilon_1 < \epsilon$ for \textcolor{black}{any initial condition $\phi$, with $\|\phi \| < \delta$, and $t\geq t_0$. By contradiction, if it is false then $\|{\bf x}(t_0,\phi^\star)(t^\star)\| \geq \epsilon_1$ at some instant $t ^\star\geq t_0$, and for some initial condition $\phi^\star$, with $\|\phi^\star \| < \delta$
. Let $T\geq t_0$ be given by $T = \min\{t^\star \in \mathds{R}^+: \lim_{ s \rightarrow t^\star , s>t^\star} \|{\bf x}(t_0,\phi^\star)(t^\star)\| \geq \epsilon_1\}$.
Note that if
$T\in \mathds{T}_{\phi^\star}$ then $\| \mathbf{I}(T,{\bf x}_T) \| = \epsilon_1$, otherwise $\| {\bf x}(T) \| = \epsilon_1$; thus, as a result, we have $\| {\bf x}(t_0,\phi^\star)(t) \| < \epsilon_1$ for $t\in [t_0,T)$. In addition, let $\mathds{T}_{\phi^\star} = \{t_1^\star, t_2^\star, \cdots\}$ be the sequence of reset instants corresponding to the initial condition $\phi^\star$; since reset instants are well-posed and Zeno-free then there exists $N>0$, defined as the largest integer for which $t_N^\star \leq T$.
Now, since $\delta<\epsilon_1$, conditions (\ref{eq:cond2}) and (\ref{eq:cond3}) imply
\begin{equation}
V(t,\mathbf{x}_t) \leq V(t_k^\star, I(t_k^\star ,\mathbf{x}_{t_k^\star})) \leq V(t_k^\star,\mathbf{x}_{t_k^\star})
\end{equation}
\noindent for $t \in (t_k^\star, t_{k+1}^\star]$, $k= 1, 2, \cdots, N-1$, and
\begin{equation}
V(t,\mathbf{x}_t) \leq V(t_0,\phi^\star)
\end{equation}
for $t \in [t_0, t_1^\star]$. Since, in addition, $V(T,\mathbf{x}_T) \leq V(t_N^\star,\mathbf{x}_{t_N\star})$, and
$\|\phi^\star \| < \delta < \epsilon_1 \leq \gamma$, combining (9)-(10) and (\ref{eq:cond1}) it follows}
\begin{equation}
V(T,\mathbf{x}_T) \leq V(t_0,\phi^\star) \leq v(\|\phi^\star\|) < v(\delta) < u(\epsilon_1),
\end{equation}
But, from (\ref{eq:cond1}) and (\ref{eq:cond3}), it follows
\begin{equation}
u(\epsilon_1)=u(\| {\bf x}(T) \| ) \leq V(T,{\bf x}_T) \ \ \text{if} \ \ T\notin \mathds{T}_{\phi^\star}
\end{equation}
and
\begin{equation}
u(\epsilon_1) = u(\| \mathbf{I}(T,{\bf x}_T) \| ) \leq V(T,\mathbf{I}(T,{\bf x}_T)) \leq V(T,{\bf x}_T)\ \ \text{if} \ \ T \in \mathds{T}_{\phi^\star},
\end{equation}
which is a contradiction in both cases.
\textcolor{black}{({\em Uniform asymptotic stability)}} In this case, the proof is a bit more involved. For $\epsilon > 0$ choose $\delta_a > 0$ such as \textcolor{black}{$v(\delta_a) < u(\min\{\epsilon,\gamma\})$, }thus it is true that $\|\phi\| < \delta_a$ implies $\|\mathbf{x}(t_0,\phi)(t)\| < \min(\epsilon,\gamma)$ for $t \geq t_0$. Now, it will be shown that for any $\eta > 0$ there exists some $T(\delta_a,\eta)$ such that $\|\mathbf{x}(t_0,\phi)(t)\| < \eta$ \textcolor{black}{for any $\phi$, with $\|\phi\| < \delta_a$, and} $t \geq t_0 + T$. This is equivalent to prove that $\|\mathbf{x}_{t_0+T}\| < \delta_b$, where \textcolor{black}{$v(\delta_b) =u(\min\{\eta,\gamma\})$.}
By contradiction, suppose that there not exists such $T$, \textcolor{black}{that is there exist some $\eta>0$ and a solution $\mathbf{x}(t_0,\phi^\star)(t)$, with $\| \phi^\star \| < \delta_a$,} such as $\|\mathbf{x}_{t}\| \geq \delta_b$ for all $t \geq t_0$. Thus, there exists a sequence $(\tau_k)$, $k = 1, 2, \cdots$ such that
\begin{equation}
t_0 + (2k-1)h \leq \tau_k \leq t_0 + 2kh,
\end{equation}
\noindent where $\tau_k \notin \mathds{T}_{\phi^\star} = \{t_1^\star, t_2^\star, \cdots\}$ and $\|{\bf x}(\tau_k)\|\geq\delta_b$. \textcolor{black}{Since it is assumed that the system (\ref{eq:resetsystem}) has well-posed and Zeno-free reset instants, and $t_0$ is not a reset instant, then $t_0 < t_{1}^\star < t_2^\star < \cdots $, and if $\mathds{T}_{\phi^\star}$ is infinite then $t_N^\star \rightarrow \infty$ as $N\rightarrow \infty$. In addition, since ${\bf f}$ is locally Lipschitz, and by uniform stability $\|\mathbf{x}(t_0,\phi^\star)(t)\| < \min(\epsilon,\gamma)$ for $t \geq t_0$, then there exists a constant $L>0$ such that $\|\dot{\mathbf{x}}(t)\| < L$ for all $t \in [t_0,\infty)\setminus \mathds{T}_\phi$.} Therefore, it is possible to build a set of intervals $I_k = [\tau_k - \frac{\delta_b}{2L}\alpha^1_k , \tau_k + \frac{\delta_b}{2L}\alpha^2_k] $ with $\alpha^1_k$, $\alpha^2_k \in \{0,1\}$, $\alpha^1_k+\alpha^2_k\geq1$, that do not contain reset instants and do not overlap (by using a number $L > 0$ large enough and proper values of $\alpha^1_k$, $\alpha^2_k$), that is \textcolor{black}{$I_k \cap \mathds{T}_{\phi^\star} = \varnothing$}, $k = 1, 2, \cdots$, and then by using the mean-value theorem on the intervals $[\tau_k,t] \subset I_k$ and $[t,\tau_k] \subset I_k$
\textcolor{black}{
\begin{equation}
\|\mathbf{x}(t_0,\phi^\star)(t)\| = \| \mathbf{x}(\tau_k) + \dot{\mathbf{x}}(\tau_k + \theta_1 (t-\tau_k))(t-\tau_k)\|,
\end{equation}
\begin{equation}
\|\mathbf{x}(t_0,\phi^\star)(t)\| = \| \mathbf{x}(\tau_k) - \dot{\mathbf{x}}(\tau_k + \theta_2 (\tau_k-t))(\tau_k-t)\|
\end{equation}
}
for some $\theta_1,\theta_2 \in (0,1)$, and then
\textcolor{black}{
\begin{equation}
\|\mathbf{x}(t_0,\phi^\star)(t)\| \geq \| \mathbf{x}(\tau_k)\| - \|\dot{\mathbf{x}}(\tau_k + \theta (t-\tau_k))\| |t-\tau_k| \geq \frac{\delta_b}{2}
\end{equation}
}
for any $t \in I_k$ and some $\theta \in (-1,1)$. In addition, from (\ref{eq:cond2}) it is true that $\dot{V}(t,\mathbf{x}_t) \leq -w(\|\mathbf{x}(t)\|) \leq - w(\frac{\delta_b}{2})< 0$, for any $t \in I_k$, this means that $V(t,\mathbf{x}_t)$ is decreasing with at least a ratio $-w(\frac{\delta_b}{2})$ in each interval $I_k$, $k= 1,2, \cdots$. On the other hand, the reset instants in the sequence \textcolor{black}{$\mathds{T}_{\phi^\star} = \{t_1^\star,t_2^\star, \cdots \}$ may be renamed as $\mathds{T}_{\phi^\star} = \{t_{k,l}^\star\}$, where the reset instant $t_{k,l}^\star$ corresponds to the $l^{th}$-instant prior to $\tau_k \in I_k$, that is
\begin{equation}
\begin{array}{l}
t_0 \leq t_{1,1}^\star < t_{1,2}^\star < \cdots t_{1,N_1}^\star < \check{\tau}_{1} = \tau_{1} - \frac{\delta_b}{2L}\alpha^1_k \\ < \hat{\tau}_{1} = \tau_{1} + \frac{\delta_b}{2L}\alpha^2_k < t_{2,1}^\star < \cdots t_{2,N_2}^\star < \check{\tau}_{2} < \cdots
\end{array}
\end{equation}
\noindent for some integers $N_1, N_2, \cdots \geq 0$. Note that $N_1, N_2, \cdots$ do exist since reset instants are well-posed and Zeno-free.} Therefore, by integrating $\dot{V}(t,\mathbf{x}_t)$ over the interval $[\check{\tau}_1, \hat{\tau}_1]$ and using also (\ref{eq:cond3}), it is obtained
\textcolor{black}{
\begin{equation}
\begin{array}{l}
V(\check{\tau}_{1} ,\mathbf{x}_{\check{\tau}_{1}} ) \leq V(t_{1,N_1}^\star,\mathbf{I}(t_{1,N_1}^\star,\mathbf{x}_{t_{1,N_1}^\star})) \leq V(t_{1,N_1}^\star,\mathbf{x}_{t_{1,N_1}^\star}) \leq \cdots \leq V(t_0,\phi^\star)
\end{array}
\end{equation}
}
and
\begin{equation}
\begin{array}{l}
V(\tau_2,\mathbf{x}_{\tau_2}) \leq V(\hat{\tau}_{1} ,\mathbf{x}_{\hat{\tau}_{1} }) \leq V(\check{\tau}_{1} ,\mathbf{x}_{\check{\tau}_{1} }) - w(\frac{\delta_b}{2})\frac{\delta_b}{2L}(\alpha^1_k+\alpha^2_k) \\ \textcolor{black}{\leq \cdots \leq V(t_0,\phi^\star)- w(\frac{\delta_b}{2})\frac{\delta_b}{2L} (\alpha^1_k+\alpha^2_k).}
\end{array}
\end{equation}
\noindent Finally, since $\alpha^1_k+\alpha^2_k\geq 1$ for $k=1,2,\ldots$, then repeating the reasoning for any $\tau_k$, it follows
\textcolor{black}{
\begin{equation}
V(\tau_k,\mathbf{x}_{\tau_k}) \leq V(t_0,\phi^\star) - w(\frac{\delta_b}{2})(k-1)\frac{\delta_b}{2L},
\end{equation}
}
\noindent and then for a large enough $k$ it results that $V(\tau_k,\mathbf{x}_{\tau_k}) < 0$, which is a contradiction.
Finally, if $\lim_{s \rightarrow \infty} u(s) = \infty$ and $\Omega=\cal{PC}$, then $\delta_a$ above may be chosen arbitrarily large, and $\epsilon$ can be set after $\delta_a$ to satisfy $v(\delta_a) < u(\epsilon)$. Therefore, global uniform asymptotic stability can be concluded. $\Box$ \\
\section{Application cases}
\label{section:application}
\textcolor{black}{In the following cases, it will be assumed that reset instants are time-regularized, and thus they are well-posed and Zeno-free. In addition, the rest of assumptions in Prop. 3.1 can be easily checked to be satisfied, and it will not be explicitly shown. }
\subsection{Reset systems with a LTI base system and single time-delay }
In this section, we establish delay-independent stability conditions for time-delay reset systems with LTI base system as in \cite{BanBar}, here the provided proof is linked to the main stability result of Section \ref{section:mainresult}. Consider a special case of (\ref{eq:resetsystem}), in which the base system is a LTI base system with a single time-delay, given by
\begin{equation}\label{eq:ltiresetsystem}
\left\{ \begin{array}{ll}
\dot{\mathbf{x}}(t) = A {\bf x}(t)+ A_d {\bf x}(t-h), & \mathbf{x}(t) \notin {\cal M}, \\
{\mathbf{x}}(t^+) = A_R {\bf x}(t), & \mathbf{x}(t) \in {\cal M},\\
\mathbf{x}(t) = \phi(t), & t \in [-h,0],
\end{array}\right.
\end{equation}
for arbitrary values of $A$ and $A_d$, and where $t_0=0$ is the initial instant, and the reset matrix $A_R$ takes the form $A_R=diag(I_{\bar{n}_\rho},0_{n_\rho})$, $\bar{n}_\rho = n - n_\rho $. The reset action is applied on the last $n_\rho$ states of the vector ${\bf x}\in \mathds{R}^n$ at those instants in which the state reaches the reset set defined as \textcolor{black}{$\mathcal{M} = \{{\bf x} \in \mathds{R}^n: C\mathbf{x}=0 \}$ for some row vector $C \in \mathds{R}^{1 \times n}$}.
The asymptotic stability of this system can be analyzed by the following proposition.\\
{\bf Proposition 4.1}: If there exist (symmetric) matrices $P$, $Q>0$ such that
\textcolor{black}{
\begin{equation}\label{eq:LTIcond1}
M= \left(
\begin{array}{cc}
A^\top P + PA + Q & PA_d \\
A_d^\top P & -Q
\end{array}
\right)<0
\end{equation}
}
and
\begin{equation}\label{eq:LTIcond2}
\Theta^\top ( A_R^\top P A_R - P ) \Theta \leq 0
\end{equation}
for some $\Theta$ with $\mathcal{R} (\Theta)=\mathcal{N}(C)$, then the zero solution of system (\ref{eq:ltiresetsystem}) is globally asymptotically stable.\\
{\bf Proof}: Consider the Lyapunov-Krasowskii functional $V:{\cal PC} \rightarrow \mathds{R^+}$ given by
\begin{equation}\label{eq:functional}
V(\psi) = \psi(0)^\top P \psi(0) \ + \int_{-h}^{0} \psi^\top(\theta) Q \psi(\theta) \ d\theta,
\end{equation}
with $P$, $Q$ the matrices of the proposition. For this functional, since $P$, $Q > 0$ it is true that
\begin{equation}
V(\psi) \leq \lambda_{M}(P) \|\psi(0)\|^2 + h\lambda_{M}(Q)\|\psi\|^2 \leq (\lambda_{M}(P)+ h\lambda_{M}(Q))\|\psi\|^2 = v(\|\psi\|)
\end{equation}
\noindent and
\begin{equation}
V(\psi) \geq \lambda_{m}(P) \|\psi(0)\|^2 + h\lambda_{m}(Q)\|\psi\|^2 \geq \lambda_{m}(P)\|\psi(0)\|^2 = u(\|\psi(0)\|),
\end{equation}
\noindent where $u,v: \mathds{R}^+ \rightarrow \mathds{R}^+$ are continuous nondecreasing functions and $u,v \in K$. On the other hand, the derivative of $V$ along the solutions of (\ref{eq:ltiresetsystem}), after some manipulation, is given by
\begin{equation}
\dot{V}(\psi) =
\left(
\begin{array}{cc}
{\bf x}^\top(t)& {\bf x}^\top(t-h)
\end{array}
\right)
M\left(
\begin{array}{c}
{\bf x}(t) \\ {\bf x}(t-h)
\end{array}
\right).
\end{equation}
Therefore, condition (\ref{eq:LTIcond1}) is obtained, and it implies
\begin{equation}
- \dot{V}({\bf x}_t) > \lambda_{m}(-M)(\|{\bf x}(t)\|^2 + \|{\bf x}(t-h)\|^2) \geq \lambda_{m}(-M)(\|{\bf x}(t)\|^2) = w(\|{\bf x}(t)\|)
\end{equation}
\noindent for all ${\bf x}(t) \notin {\cal M}$, where $w: \mathds{R}^+ \rightarrow \mathds{R}^+$ is a continuous nondecreasing function and $w \in K$. Finally, condition (\ref{eq:LTIcond2}) is obtained by setting $\Delta V({\bf x}_t)= {\bf x}^\top(t) (A_R^\top P A_R -P) {\bf x}(t)\leq0$ and considering that ${\bf x} \in \cal{M}$ implies that there exist $y$ such that $x=\Theta y$. As a result, all the conditions of Proposition 3.1 are satisfied and thus the zero solution of the system (\ref{eq:ltiresetsystem}) is globally asymptotically stable. $\Box$ \\
\textcolor{black}{
{\bf Example 4.1:} Consider the time-delay reset system (\ref{eq:ltiresetsystem}) with matrices
\begin{equation}
\begin{array}{ccccc}
A=\left( \begin{array}{cc} -2 & 0 \\ 0 & -0.9
\end{array}\right), & &
A_d=\left( \begin{array}{cc} -1 & 1 \\ -1 & -0.5
\end{array}\right), & &
A_R=\left( \begin{array}{cc} a_1 & 0 \\ a_2 & a_3
\end{array}\right).
\end{array}
\end{equation}
The base system is asymptotically stable independently of the time-delay since there exist matrices $P$ and $Q$ such that condition (\ref{eq:LTIcond1}) is satisfied (see for example \cite{gu}). Now suppose that $C= (\ 1 \ \ 0 \ )$, then condition (\ref{eq:LTIcond2}) is satisfied whenever $| a_3 | \leq 1$, and thus, the reset system is globally asymptotically stable. Moreover, if $a_2=0$, $| a_1 | \leq 1$ and $| a_3 | \leq 1$ then the asymptotic stability of the reset control system is guaranteed for any row vector $C$. The trajectory of the reset system with $C= [\ -2 \ \ 1 \ ]$, $h=1$, $a_1=1$, $a_2=0$ and $a_3=0$ is shown in Fig. \ref{fig:LTI_trajectory}. In addition, Fig. \ref{fig:LTI_functional} shows the value of the Lyapunov-Krasovskii functional along this trajectory. Note that the Lyapunov-Krasovskii functional obeys the two conditions in Prop. 4.1, decreasing both during the continuous dynamic and the jumps.
\begin{figure}
\begin{minipage}[b]{.5\textwidth}
\vspace*{\fill}
\centering
\includegraphics[scale=1]{trajectorypdf.pdf}
\subcaption{\textcolor{black}{Trajectory: $x_1$ (dashed) and $x_2$ (solid).} }
\label{fig:LTI_trajectory}
\end{minipage}
\begin{minipage}[b]{.5\textwidth}
\vspace*{\fill}
\centering
\includegraphics[scale=1]{functionalpdf.pdf}
\subcaption{\textcolor{black}{Lyapunov-Krasovskii functional.}}
\label{fig:LTI_functional}
\end{minipage}
\caption{\textcolor{black}{Trajectory and value of the Lyapunov-Krasovskii functional in the Example 4.1.}}
\label{fig:LTIexample}
\end{figure}
}
\subsection{Reset system with a nonlinear and time varying base system }
In general, for a reset system with nonlinear and time-varying base system without a particular structure, there is not systematic procedure for generating Lyapunov-Krasovskii functionals candidates. Therefore, in this section the main result is applied to a reset system with a particular structure.\\
{\bf Example 4.2:} Consider a second-order reset system given by
\begin{equation}\label{eq:nonlinearresetsystem}
\left\lbrace \begin{array}{ll}
\left\lbrace \begin{array}{l}
\dot{x}_1(t) = a_1(t) x_1^3 (t) + b_1(t) x_1^3 (t-h) + c_1(t) x_2(t) \\
\dot{x}_2(t) = a_2(t) x_2 (t) + b_2(t) x_2 (t-h) + c_2(t) x_1^3(t)
\end{array}
\right\rbrace,& \textcolor{black}{{\bf g}(x_1(t),x_2(t))\neq 0, }
\\
\\
\left\lbrace \begin{array}{l}
x_1(t^+) = x_1(t) \\
x_2(t^+) = 0
\end{array}
\right\rbrace,& \textcolor{black}{{\bf g}(x_1(t),x_2(t))= 0. }
\end{array} \right.
\end{equation}
where $a_1(t)$, $b_1(t)$, $c_1(t)$, $a_2(t)$, $b_2(t)$, and $c_2(t)$ are arbitrary continuous bounded functions with $a_1(t), a_2(t) \leq -\delta$, $|b_1(t)|,|b_2(t)| \leq \frac{\delta}{2}$, and $|c_1(t)|,|c_2(t)| \leq \frac{\delta}{4}$ for some given $\delta>0$. \textcolor{black}{Here, the reset set $\mathcal{M}$ is given as $\mathcal{M} = \{(x_1,x_2) \in \mathds{R}^2: \mathbf{g}(x_1,x_2)= 0, \text{for some } \mathbf{g}: \mathds{R}^2 \rightarrow \mathds{R} \}$. }
Consider the candidate functional $V$ be defined by
\textcolor{black}{
\begin{equation}
\label{eq:non_functional}
V({\bf x}_t) = \frac{x_1^4(t)}{4}+ \frac{x_2^2(t)}{2} + \frac{\delta}{2}\int^0_{-h} \left( x_1^6(t+\alpha) + x_2^2(t+\alpha) \right) d\alpha,
\end{equation}
}
where ${\bf x}(t)=(x_1(t),x_2(t))\in\mathds{R}^2$ and ${\bf x}_t=x(t+\alpha)$, $\alpha \in [-h,0]$. Let define the continuous nondecreasing functions $u(s)=\frac{1}{4}s^4$ and $v(s)=2 s^2$, then the above functional satisfies condition (\ref{eq:cond1}),
\begin{equation}
u(\|{\bf x}_t(0)\|)=\frac{1}{4}(x_1^2(t)+x_2^2(t))^2 \leq \frac{1}{4}(x_1^4(t)+x_2^2(t)) \leq V({\bf x}_t)
\end{equation}
and
\begin{equation}
V({\bf x}_t)\leq x_1^2(t)+ x_2^2(t) + \frac{\delta}{2}\int^0_{-h} \left( x_1^2(t+\alpha)+ x_2^2(t+\alpha)\right) d\alpha \leq 2\|{\bf x}_t\|^2.
\end{equation}
It is easy to check that the above conditions are satisfied for any $\|{\bf x}_t\| \leq \gamma=\sqrt{\frac{1}{5} }$.
On the other hand, the derivative of the functional along the solutions of system (\ref{eq:nonlinearresetsystem}) is given by
\begin{equation}
\dot{V}({\bf x}_t) = x_1^3(t)\dot{x}_1(t)+ x_2(t)\dot{x}_2(t) + \frac{\delta}{2}(x_1^6(t) - x_1^6(t-h)) + \frac{\delta}{2}(x_2^2(t) - x_2^2(t-h)).
\end{equation}
After some manipulations the derivative of the functional is bounded by
\begin{equation}
\dot{V}({\bf x}_t)\leq \xi^\top(t) M \xi(t),
\end{equation}
where $\xi(t)=\left( |x_1(t)|^3, |x_1(t-h)|^3, |x_2(t)|, |x_2(t-h)| \right)$ and
\begin{equation}
M=\left( \begin{array}{rrrr}
-\frac{\delta}{2} & \frac{\delta}{4} & \frac{\delta}{4} & 0\\
\frac{\delta}{4} & -\frac{\delta}{2} & 0 & 0\\
\frac{\delta}{4} & 0 & -\frac{\delta}{2} & \frac{\delta}{4}\\
0& 0 & \frac{\delta}{4} & -\frac{\delta}{2}\\
\end{array}\right).
\end{equation}
Since $M<0$ for any $\delta>0$, then defining $w(s)=\lambda_{m}(-M)s^6$ it follows
\begin{equation}
-\dot{V}({\bf x}_t)\geq \lambda_{m}(-M) \left( |x_1(t)|^6 + |x_1(t-h)|^6+|x_2(t)|^2 \right. \left. +|x_2(t-h)|^2 \right)\geq w(\| {\bf x}(t)\|)
\end{equation}
for any $\| {\bf x}_t\| \leq \sqrt{\frac{1}{5}}$, and thus condition (\ref{eq:cond2}) is satisfied. Finally, $\Delta V({\bf x}_t)=-\frac{x^2_2(t)}{2}$, which is negative for any reset instant, \textcolor{black}{ regardless of the function ${\bf g}$.} Hence, the solution ${\bf x}_t=0$ of the system (\ref{eq:nonlinearresetsystem}) is uniformly asymptotically stable.
\textcolor{black}{Now, consider the reset system with functions $a_1(t)$, $b_1(t)$, $c_1(t)$, $a_2(t)$, $b_2(t)$, and $c_2(t)$ given by
\begin{equation}
\begin{array}{lllll}
a_1(t)=50e^{-t}-100\delta, & & b_1(t) =\frac{\delta}{2}\sin(t), & &c_1(t)=\frac{\delta}{4},\\
a_2(t)=-e^{-t}-\delta, & &b_2(t) =-\frac{\delta}{2}, & &c_2(t) =\frac{\delta}{4}\sin(t).
\end{array}
\end{equation}
It can be easily seen that the above functions satisfy the required bounds for any $\delta>0.51$. In addition, consider the function ${\bf g}(x_1,x_2)=-5x_1+x_2$ and the initial condition $\phi(t)=(\frac{1}{2\sqrt{5}}, -\frac{1}{2\sqrt{5}})$, $t \in [t_0-h,0]$ such that $\| \phi \| = \frac{1}{\sqrt{10} }< \frac{1}{\sqrt{5}}$. The evolution of the system with $h=1$ and $t_0=0$ is plotted in Fig. \ref{fig:non_trajectory}. In addition, Fig. \ref{fig:non_functional} shows the value of the functional (\ref{eq:non_functional}) along the trajectory. Once again, it can be observed how the functional always decreases.}
\begin{figure}
\begin{minipage}[b]{.5\textwidth}
\vspace*{\fill}
\centering
\includegraphics[scale=1]{trajectory_non_pdf.pdf}
\subcaption{\textcolor{black}{Trajectory: $x_1$ (dashed) and $x_2$ (solid). }}
\label{fig:non_trajectory}
\end{minipage}
\begin{minipage}[b]{.5\textwidth}
\vspace*{\fill}
\centering
\includegraphics[scale=1]{functional_non_pdf.pdf}
\subcaption{\textcolor{black}{Lyapunov-Krasovskii functional.}}
\label{fig:non_functional}
\end{minipage}
\caption{\textcolor{black}{Trajectory and value of the Lyapunov-Krasovskii functional in the Example 4.2.}}
\label{fig:nonexample}
\end{figure}
\section{ Conclusions}
\label{section:conclusion}
This work provides Lyapunov-Krasovskii based conditions that guarantee the stability of time-delay reset systems. In comparison with previous works, the main contribution has been to consider reset systems, whose base system is nonlinear and time-varying. As an application, the proposed result have been applied to obtain delay-independent condition in terms of LMI for time-delay reset control systems. In addition, the stability of a particular reset system with nonlinear and time-varying base system is analyzed.
\section*{Acknowledgements}
This work has been supported by Ministerio de Econom{\'i}a e Innovaci{\'o}n of Spain under project DPI2013-47100-C2-1-P (including FEDER co-funding).
The authors thank helpful comments of Sophie Tarbouriech and Luca Zaccarian, that have motivated the problem approached in this work.
\section*{References}
|
\section{Introduction}
\begin{flushleft}
•
\end{flushleft}
Given two graphs $G$ and $H$, the Graph Isomorphism problem (GI) asks whether there exists a bijection from the vertices of $G$ to the vertices of $H$ that preserves adjacency. The graph isomorphism problem has a long history in the fields of mathematics, chemistry, and computing science. The problem is known to be in NP, but is not known to be in \textbf{P} or \textbf{NP-complete}. The best current theoretical algorithm is due to Babai and Luks (1983)\textbf{[2]}. The algorithm relies on the classification of finite simple groups. In 2015, Laszlo Babai claimed that the Graph Isomorphism problem can be solved in quasipolynomial time.
\subsection{ Notations and Definitions}
Let $G$ and $H$ be two graphs. Each graph has $n$ vertices. The \textit{cardinality} of a set $B$ is the number of elements in it, denoted by $|B|$. For example, $|G|=|H|=n$.
The automorphism group of graph $G_1$ will be denoted by $Aut(G_1)$
The \textit{neighbourhood} of a vertex $v$ in a graph $G$ is the induced subgraph of $ G$ consisting of all vertices adjacent to $v$.
The neighbourhood is denoted $ N_G(v)$ or (when the graph is unambiguous) $N(v)$. If the neighbourhood does not include $v$ itself then it is open neighbourhood of $v$; it is also possible to define a neighbourhood in which v itself is included, called the closed neighbourhood and denoted by $N_G[v]$.
If all vertices in $G$ have neighbourhoods that are isomorphic to the same graph $G_1$, then $G$ is said to be \textit{locally} $G_1$, and if all vertices in $G$ have neighbourhoods that belong to some graph family $\mathcal{F}, G$ is said to be locally $\mathcal{F} $ (Hell 1978, Sedlacek 1983).
A permutation of a vertex set $G$ is a bijection from $G$ to itself. For example, if $\pi=\left(\begin{matrix} 1 & 2 & 3\dots & n \\5& 2 & 7 \dots& 11 \end{matrix} \right)$, then the first vertex of $G$ moves to fifth position in the $G^\pi$ . $Sym(G$) or $S_n$ denotes the set of all permutations of $G$.
If $G$ is Isomorphic to $H$,then $\exists P \in S_n$(Symmetric Group of $n $ vertices) such that $H^P=G$(notation by Wielandt). We write $G\simeq H$ when $G$ is Isomorphic to $H$.
A \textit{tuple} is a finite ordered list of vertices.
A \textit{search tree} is an undirected graph in which any two vertices are connected by exactly one path.A \textit{rooted search tree} is a tree in which one vertex has been designated the root.The tree elements are called \textit{nodes}. Each of the nodes that is one graph-edge further away from a given node is called a \textit{child}, i.e the vertices adjacent to the root vertex are called its children.A rooted tree naturally imparts a notion of \textit{Levels} (distance from the root), thus for every node a notion of children may be defined as the nodes connected to it a level below. Nodes without children are called \textit{leaf nodes, end-Nodes, leaves}.
A \textit{walk} on a graph is an alternating series of vertices and edges beginning and ending with a vertex in which each edge is incident with the vertex immediately preceding it and the vertex immediately following it.A \textit{trail} is a walk in which all edges are distinct. A \textit{path} is a trail in which all vertices are distinct. A path has a \textit{Sequences of Vertices}. In a search tree, a \textit{Discrete Partition} or \textit{Individualization-Refinement Path} is a path which starts at the root and ends in a leaf \textbf{[1]}.
A subset $B$ of a group is called generating set, if the smallest subgroup containing the subset is the group $S_n$ itself. We write, $\langle B \rangle =S_n$. A generating set is called minimal generating, if the set does not properly contain any generating set.
\section{Overview}
We define graphs $G,H$ as \textit{regular, connected, } and not \textit{locally triangle free}. If $G \simeq H$ then $H$ must have the same structure like $G$. We will rearrange $G$ according to \textbf{2.1}. This rearrangement will split $G$ in to vertex set $G_1, G_2.... G_x$ . In this paper, Graphs are not locally triangle-free. We will construct a search tree $\mathcal{T}_k$ for $H_k$. The search tree $\mathcal{T}_k$ will provide a set of permutations which will create the set $\beta_k$. We will construct generating set of automorphsim from $\beta_k$. The algorithm uses the same technique to find automorphism as [6].
\subsection {\textit{Rearrangement} of a Graph:}
Consider any vertex of $G$, say $v$. We label $v$ with the integer $n$ (put $n$ on it, as label). We use subscript $n$ to denote the $n$ labeled vertex and write $v_n$. Now, construct neighbourhood induced sub-graph $N_G[v_n]$ (closed neighbourhood). We rewrite $ N_G[v_n]$ as $ G_1$ i.e. $ G_1=N_G[v_n]$. Now, select any unlabeled vertex from $G_1$ and label it with $(n-1)$. The vertex with label $(n-1)$ would be denoted as $v_{(n-1)}$.
Now, construct induced sub-graph $ N_{G_{1}}(v_{(n-1)})$ and label any vertex of it as $v_{(n-2)}$ .
We repeat the above procedure again. Label an unlabeled vertex (i.e. an unlabeled vertex of $G \setminus G_1$) as $v_{(n-3)}$ and obtain $G_2$ based on adjacency (neighbourhood) of veretex $v_{(n-3)}$. Using the same procedure, we would be able to obtain $G_3, G_4, ... G_{(n/2)}$ subgraphs and to label all vertices of $G$. Thus $G$ could be splitted into $(n/3)$ subgraphs.
There is a bijection from the vertex set of $G$ to a set $L= \{ 1,2..n\}$ of labels.
Now, let us define, a $n$-tuple (a sequence or ordered list of $n$ vertices) $w_G$ where vertices are ordered according to their labels (e.g. vertex labelled with label $1$ is in $1^{st}$ position in $w_G$).
$w_G$ is our desired arrangement (ordered) of graph $G$. For some $k$, $G_k$ has vertices from $i_k$ to $j_k$ of $w_G$ where
$(i_k+2=j_k)$. Here, $i_k$ is the starting position of $k^{th}$ subgraph $G_k$ in $w_G$ and $j_k$ is the ending position of $k^{th}$ subgraph $G_k$ in $w_G$. Subscript $k$ is used to distinguish $i, j$ for the $k^{th}$ subgraph $G_k$ from other subgraphs.
For example, if $k=1$, then $G_1$ has vertices from $(n-2)$ to $n$ in $w_G$, here $i_1=(n-2)$ and $j_1=n$.
\subsection {Construction of \textit{Search Tree} for Generating Set :}
We have rearranged $G$ (one of the two given graphs $G$ and $H$). Now, we will construct permutations for $H$ with respect to $G$. Conversely, permutations can be generated for $G$ with respect to $H$ using procedure described in this subsection.
We label each vertex of $H$ uniquely with elements from the same set $L= \{ 1,2,...n\}$. This labeling procedure is random. It must make sure that there is a bijection from the vertex set of $H$ to the label set $L$.
$w_H$ is a $n$-tuple (a sequence of $n$ vertices) where vertices are ordered according to their labels . The definition of $w_H$ is similar to $w_G$, except it is defined for graph $H$.
In $w_G$, position starts from left to right, so the subgraph $G_k$ starts at $i_k^{th}$ position and ends at $j_k^{th}$ where $n> j_k > i_k$.
We define the subgraph $H_k $ of $ H$, which has consecutive vertices from $i_k^{th}$ position to $j_k^{th}$ position in $w_H$. Here, $i_k$ and $j_k$ have the same value as they had in $G_k$. So, $H_k$ has $3$ vertices too.
If $G_k \simeq H_k$ then $\exists \pi_k$ such that $H_k^{\pi_k}=G_k$. Permutation $\pi_k$ moves vertices of $H$ to the interval between $i_k^{th}$ position and $j_k^{th}$ position, so that $H_k^{\pi_k}=G_k$. So, we construct a search tree $\mathcal{T}_k $ for constructing all possible permutations which could be $\pi_k$. We will follow the construction method of $G_k$ described in \textbf{2.1}, when we construct all possible permutations that could be $\pi_k$.
Let us define a rooted tree $\mathcal{T}_k$ (for $H_k$) , its nodes are labelled vertices of $H$.
The children of root will be all possible candidates (vertices) for $j_k^{th}$ position of $H$ which could be $v_{j_k}$ where $v_{j_k}$ is the ${j^{th}_k}$ vertex of $w_G$. Since $\pi_k$ moves (or fixes) a vertex of $H$ to the $j_k^{th}$ position of $H$ such that $H_k^{\pi_k}=G_k$, consider all possible vertices that could be the vertex $v_{j_k}$ of $G_k \subset G$. All $n$ vertices of $H$ could be the $j_k^{th}$ vertex of $G$. So, at level $1$ , the children of root would be all $n$ vertices of $H$. It means, each node of $1^{st}$ level, has a unique label in $\mathcal{T}_k $. Thus, we have a bijection form $H$ to the nodes of $1^{st}$ level of $\mathcal{T}_k $. Let, $t_{k_1}$ is a node of level $1$ of $\mathcal{T}_k $.
Each node of $t_{k_1} \in \mathcal{T}_k $ is related to a vertex, say $u_l$ in $H$ ($1 \leq l \leq n $). All vertices that are adjacent to $u_l$ make a subgraph, say $\mathbf{A_H}$.
The children nodes of $t_{k_1} \in \mathcal{T}_k $ will be all vertices of subgraph $\mathbf{A_H}$. Repeat previous procedure $\forall t_{k_1} \in \mathcal{T}_k $. Thus we obtain all nodes of the $2^{nd}$ level of $\mathcal{T}_k $.
We repeat the procedure for all $H_k$ graphs until we find all possible leaf node of $\mathcal{T}_k $. Thus, we construct the search tree $\mathcal{T}_k $. Note that, each level represents a position in $w_H$ of $H$, for example, in $\mathcal{T}_k $ of $H_k$, $1^{st}$ level represents the $j_k^{th}$ position. We construct $x$ such trees.
The height of the $\mathcal{T}_k $ is $2$ .
Note that an individualization-refinement path or discrete partition of $\mathcal{T}_k $ is a permutation of $H_k \subset H$ (concept of \textbf{ [1]}). All such paths, i.e. permutations create the set $\beta_k$. There will be total $x$ number of $\beta_k$.
For example, if a \textit{path} is $5,8,9$ (where $5,8,9$ represent the labeled vertices of $H$) then,
$\pi_k=\left(\begin{matrix} j_k & (j_k-1) & (j_k-2) &=i_k \\5 & 8 & 9 \end{matrix} \right)$.
It means, $\pi_k$ is a permutation that moves $5^{th}$ vertex of $w_H$ to $j_k^{th}$ vertex,$8^{th}$ vertex of $w_H$ to $(j_k-1)^{th}$ vertex, and $9^{th}$ vertex of $w_H$ to $(j_k-2)^{th}$ vertex.
\section{Propositions}
If $G \simeq H$, then $\exists \pi_k \in \beta_k$ such that $H_k^{\pi_k}=G_k$. For each subgraph $H_k$, we have found a set of permutations $\beta_k$(from \textbf{2.2}). We would be able to construct the direct product $P$ such that $H^{p}=G$ \textit{if and only if} $H \simeq G$. If we fail to construct such $P$, it implies that $H \not \simeq G$.
\textbf{Proposition 3.1 :} $|\beta_k| < n^{3}$.
\textbf{Proposition 3.2 :} Given two graphs $G, H$ with $n$ vertices each, deciding whether they are isomorphic is \textit{polynomial time equivalent} to determining generating sets of automorphism group of graphs $G, H$ .
\textbf{Proof:} See $[3]$.$\blacksquare$
So,to decide graph isomorphism of $G, H$, it is sufficient to construct generating sets of automorphism group of graphs $G, H$.
\textbf{Proposition 3.3 :} Let $S_n$ be the finite group of order $n!$, There is a subset of elements of $S_n$ of size at most
$\log_2(n!)$ which generates $S_n$.
\textbf{Proof:} The proof is similar to the Lemma 1 of $[4]$ on page 3.$\blacksquare$
From now on, $G, H$ are adjacency matrices of graphs $G, H$ respectively. $H_k, G_k$ are blocks or sub-matrices of matrix $H, G$ respectively. The adjacency matrix of graph $H_k \cup H_e$ is $M_{(k,e)}$ where $M_{(k,e)} =\left( \begin{array}{ccc} H_e & R_{k,e} \\ R_{k,e}^{T} & H_k\\ \end{array} \right) $, where, $R_{k,e}$ is the non symmetric sub-matrix of adjacency matrix $H$. Here, $R_{k,e}$ represents edges between $H_k, H_e$. Similarly, $S_{k,e}$ represents edges between $G_k, G_e$.
$$H = \begin{bmatrix}
H_{(x)} & R_{(x, x-1)} & R_{(x,x-2)} & \dots & \dots & R_{(x,1)} \\
R_{(x,x-1)} & H_{(x-1)} & R_{(x-1,x-2)} & \dots & \dots & R_{(x-1,1)} \\ \vdots & \vdots & \vdots & \ddots & \ddots & \vdots \\
R_{(x,1)} & R_{(x-1,1)} & R_{(x-2,1)} & \dots & \dots &H_{1} \end{bmatrix}$$
\textbf{Proposition 3.4 :} Generating set of automorphism group of graph $H$ can be constructed in polynomial time if $H$ is not locally triangle-free as defined above .
\textbf{Proof:} An algorithmic proof is presented here.
At $1^{st}$ iteration -
Step 1. Construct all possible direct product $(\pi_1 \times \pi_2)$ where $\pi_1 \in \beta_1$ and $ \pi_2 \in \beta_2$.
There are $| \beta_1 | \times | \beta_2| < n^{9}$ direct products (permutations). All these permutations (direct products) form set $\gamma_1$. Each element of $\gamma_1$ is a permutation that acts on graph $H_1 \cup H_2$.
Step 2. Construct/find -
$\alpha_1 =\{ \pi \in \gamma_1 | (M_{(1,2)}^{\pi}= M_{(1,2)}) \land ( R_{1,2}^{\pi} = S_{1,2}) \land (H_1^{\pi} = G_1) \land (H_2^{\pi} = G_2) \}$
$\alpha_1$ is the set of automorphisms of matrix $M_{(1,2)}$. $|\alpha_1| < n^{9}$.There are two possible cases-
Case 1: If $|\alpha_1| =1$, then for each $\pi_1 \in \beta_1$, there is only one permutation
$\pi_2 \in \beta_2$. So, there could be maximum $n^{2}$ permutations in $\gamma_1$ but only one permutation could be included in $\alpha_1$.
Case 2: If $|\alpha_1| >1$, we would be able to construct a generating set $\mathcal{S}_1$ of an automorphism group of $Aut(M_{(1,2)})$ Note, that if $\exists \pi_a \in Aut(H)$ such that it acts on vertices of $H_1 \cup H_2$, then $ \pi_a \in \langle \mathcal{S}_1 \rangle =Aut(M_{(1,2)})$. So, when we construct direct product of $\mathcal{S}_1$ and another set, $\pi_a$ can be found in the resulting generating set. This concept is similar to \textit{extending an automorphism} described in [6]. Also, see Theorem 7, on page 31 of [5].The theorem showed how to obtain the automorphism group of an arbitrary graph from the intersection of a specific permutation group with a direct product of symmetric groups.
Step 3. Now, we construct the generating set $\mathcal{S}_1$ from $\alpha_1$. This construction of generating set can be done in polynomial time (see [5], page 40, theorem 9). From proposition 3.3, we find that $|\mathcal{S}_1| \leq log(n!)$ .
$\mathcal{S}_1$ is the generating set of automorphism of $M_{(1,2)}$ .
Step 4. We start $2^{nd}$ iteration, for $\beta_3, \mathcal{S}_1$ (instead of $\beta_2$), $ M_{(2,3)}$ where $M_{(2,3)} =\left( \begin{array}{ccc} H_3 & R_{2,3} \\ R_{2,3}^{T} & H_2 \\ \end{array} \right) $. We find $\gamma_2, \alpha_2$ repeating steps $1,2$ and construct $\mathcal{S}_2$ (repeating step $3$) which is the generating set of automorphism of $M_{(2,3)}$, i.e. graph $H_1 \cup H_2 \cup H_3$. Note that, $|\mathcal{S}_2| \leq log(n!)$ .
Step 5. We keep repeating above four processes, until we find the set $\mathcal{S}_{(x-1)} $ which is the generating set of automorphism of graph $H_1 \cup H_2 \cup H_3 \dots \cup H_x=H $. Note that, $|\mathcal{S}_{(x-1)}|\leq log(n!)$, since $ \langle \mathcal{S}_{(x-1)} \rangle= Aut(H) \leq S_n$. $\blacksquare$
\section{ Conclusion}
We repeat the process of construction of $\mathcal{S}_{(x-1)}$ for graph $G$ and obtain set $\mathcal{R}_{(x-1)}$. Once we generate generating sets of $G, H$, we can decide isomorphism betwen them (3.2). The algorithm does not solve graph isomorphism problem in polynomial time if graphs are \textit{locally triangle-free}. If the sub-matrix of edges is not a zero matrix then the problem reduces down to \textit{Bipartite Graph Isomorphism Problem}. We can use the same approach there as above. This should lead to a practical solution which is the core idea of practical graph isomorphism \textbf{[1]}.
|
\chapter*{Statement}
I certify that all material in this thesis that is not my own work has been identified and that no material has previously been submitted and approved for the award of a degree by the University of Glasgow or any other institution.
\clearpage
\chapter*{Abstract}
The topic of this thesis is the application of distributive laws between comonads to the theory of cyclic homology. The work herein is based on the three papers~\cite{2, 1, woohoo}, to which the current author has contributed. Explicitly, our main aims are:
\begin{itemize}
\item To study
how the cyclic homology of associative algebras and
of Hopf algebras in the original sense of Connes and
Moscovici arises from a distributive law, and to clarify
the r\^ole of different notions of bimonad in this
generalisation.
\item To extend the procedure of twisting the cyclic
homology of a unital associative algebra to any duplicial object
defined by a distributive law.
\item To study the universality of B{\"o}hm and {\c S}tefan's approach
to constructing duplicial objects, which we do in terms of a 2-categorical
generalisation of Hochschild (co)homology.
\item To characterise those categories whose nerve admits a duplicial structure.
\end{itemize}
\chapter*{Acknowledgements}
Firstly, I would like to thank my family and friends for the immense amount of love and support offered to me whilst undertaking this venture. I would not have been able to produce this document if not for your patience and understanding in helping to see things clearly when maybe I could not.
Secondly, I must thank my second family at the University of Glasgow for providing the greatest learning environment to me for the last ten years. I will not forget the dedication and passion of my lecturers in their commitment to my learning in my undergraduate years, nor will I forget the insane level of support from my PhD supervisors, Ulrich Kr\"ahmer and Tara Brendle.
Thirdly, thank you to my friends and colleagues in Australia for the incredible hospitality. I learned a great deal of category theory in my two visits to Macquarie University, as well as the fact that it is possible to get sick of hot weather (but seriously, thank you for allowing me to experience two summers this year).
Finally, thank you to those in the musical part of my life. Art is the most crucial aspect of our humanity, and I have been lucky enough to be able to practise two of them.
I gratefully acknowledge the support of the Engineering and Physical Sciences Research Council (grants K503058/1 and P505534/1) in providing the funding necessary to complete this PhD. Thank you also to the Centre of Australian Category Theory and the University of Glasgow College of Science and Engineering for providing support for my research project in Australia.
\tableofcontents
\chapter{Introduction}\label{INTRO}
We begin by giving some context and background for the thesis in terms of the aims presented in the abstract, followed by the conventions that we use throughout.
\section{Background and aims}
The Dold-Kan correspondence generalises chain complexes
in abelian categories to general simplicial objects,
and thus homological algebra to homotopical algebra.
The classical homology theories
defined by an augmented algebra (such as group, Lie
algebra, Hochschild, de Rham and Poisson homology)
become expressed as the homology of suitable comonads
$T$, defined via simplicial objects
$\rCC_T(N,M)$ obtained from the bar
construction (see e.g.~\cite[\S6.5]{MR1269324}). Here $M,N$ are suitable functors
providing homology coefficients.
Distributive laws between monads were originally defined by Beck in~\cite{MR0241502}
and correspond to monad structures on the composite of the two underlying endofunctors. They
have found many applications in mathematics as well as computer science; see
e.g.~\cite{MR2520969, MR2504663, MR1692751, MR2220892, MR2784770}.
The study of monads and comonads arose from homological algebra, which prompts the question: can we go back and apply distributive laws to homological algebra? The answer is yes. Connes' cyclic homology created a new paradigm
of homology theories defined in terms of mixed
complexes \cite{MR883882,MR826872}. The homotopical
counterparts are cyclic
\cite{MR777584} or more generally duplicial
objects \cite{MR826872,MR885102}. B\"ohm and \c Stefan \cite{MR2415479} showed how
$\rCC_T(N,M)$ becomes duplicial in the
presence of a second comonad $S$ together with a distributive law between $T$ and $S$, which is compatible in
a suitable sense with $N$ and $M$.
The paradigmatic example of such a cyclic homology theory is the cyclic homology $\operatorname{HC}(A)$ of
a unital associative algebra $A$~\cite{MR823176, MR695483}. This leads us to our first aim:
\begin{itemize}
\item To study
how the cyclic homology of associative algebras and
of Hopf algebras in the original sense of Connes and
Moscovici \cite{MR1657389}
fits into the monadic formalism of B{\" o}hm and {\c S}tefan, extending
the construction from \cite{MR2803876}, and to clarify
the r\^ole of different notions of bimonad in this
generalisation (Chapters~\ref{DISTRIBUTIVE},~\ref{CYCLIC},~\ref{EXAMPLES}).
\end{itemize}
It was observed by Kustermans, Murphy,
and Tuset~\cite{MR1943179} that the functor $\operatorname{HC}$ can be twisted by automorphisms of $A$. In fact, this twisted cyclic homology
occurs as an instance of B{\" o}hm and {\c S}tefan's construction. Thus, our second aim:
\begin{itemize}
\item To extend the procedure of twisting the cyclic
homology of a unital associative algebra to any duplicial object
defined by a distributive law (Chapters~\ref{DISTRIBUTIVE},~\ref{CYCLIC},~\ref{EXAMPLES}).
\end{itemize}
The construction of simplicial objects via the bar resolution is universal in the sense that comonads on a category $\mathcal{B}$ correspond to strict monoidal functors $\Delta_+^* \to [\mathcal{B}, \mathcal{B}]$ where
$\Delta_+$ denotes the augmented simplicial category (cf.~Definition~\ref{simpldef}), and $*$ denotes the opposite category. Using the symmetric monoidal closed structure of $\mathbf{Cat}$, we obtain the bar resolution of the corresponding comonad as a functor $\mathcal{B} \to [\Delta_+^*, \mathcal{B}]$. What has been missing in the literature so far is a similar universal description of the situation for duplicial objects. Our third main aim is then:
\begin{itemize}
\item To study the universality of B{\"o}hm and {\c S}tefan's approach
to constructing duplicial objects, which we do in terms of a 2-categorical
generalisation of Hochschild (co)homology (Chapter~\ref{AUSTRALIA}).
\end{itemize}
The nerve functor $N \colon \mathbf{Cat} \to [\Delta^*, \mathbf{Set}]$ is full and faithful, embedding categories into simplicial sets. Thus, it makes sense to discuss
a simplicial or duplicial structure on a category. This leads to our final main aim:
\begin{itemize}
\item To characterise those categories whose nerve admits a duplicial structure (Chapter~\ref{AUSTRALIA}).
\end{itemize}
In addition to our main goals, we have some subsidiary aims:
\begin{itemize}
\item To develop some 2-category theory to assist with the above aims (Chapter~\ref{MONAD}).
\item To give a wide variety of examples (Chapters~\ref{DISTRIBUTIVE},~\ref{EXAMPLES}).
\item To pose some questions that the author was not able to answer (Chapter~\ref{GRANDFINALE}).
\end{itemize}
\section{Conventions}
We assume that the reader has familiarity with ordinary category theory (as, for example, in~\cite{MR1712872, MR1291599}), as well as the very basics of the theory of (co)modules over (co)algebras (see e.g.~\cite{MR2455920, MR1269324, MR2012570}). When natural transformations (and similar notions) appear, by abuse of notation we often write the same symbol for a natural transformation and its components, e.g.\ $\alpha \colon F \Rightarrow G$ and $\alpha \colon FX \to GX$. When a commutative diagram is given with unlabelled variables, we rather mean the collection of commutative diagrams where the variables are objects in the relevant category of interest. We completely ignore issues of size throughout (for more information on size considerations, see~\cite[p.~viii]{MR2178101} or~\cite[Ch.~I]{MR1712872} for a more serious discussion).
Each chapter contains relevant background material as well as original work of the author.
Chapters~2,~3,~4 and 6 contain joint work of the author with Ulrich Kr\"ahmer and Niels Kowalzig, although here some of the results of the relevant papers~\cite{1,2} are developed further and more background detail is given.
Chapter~5 contains joint work carried out by the author, Richard Garner and Steve Lack in~\cite{woohoo}.
At the beginning of each chapter it is made precise which parts of the material consist of original work.
\chapter{Monads and comonads}\label{MONAD}
In this chapter, we review all the 2-category theory needed for the thesis. After giving basic definitions we study monads internal to 2-categories and related concepts. Sections~\ref{2-categories} and~\ref{monadsandcomonads} contain the required definitions and preliminary results. Section~\ref{liftingthroughadjunctions}, concerning how one obtains distributive laws from certain types of squares, contains original work which is an expansion of~\cite[\S2.1--2.16]{1}. In Section~\ref{emem} we explicitly give the constructions of the previous sections in the 2-category $\mathbf{Cat}$.
\section{Preliminaries}\label{2-categories}
We begin by recalling some fundamental notions of 2-category theory.
\subsection{2-categories}
Let $\mathbbm{1}$ denote the terminal category, containing one object and one (identity) morphism.
\begin{defn}
A 2-\emph{category} $\mathscr{C}$ consists of
\begin{itemize}
\item a class $|\mathscr{C}|$ whose elements we call 0-\emph{cells}
\item for any $\mathcal{A}, \mathcal{B} \in |\mathscr{C}|$, a category $\mathscr{C}(\mathcal{A}, \mathcal{B})$ whose objects we call 1-\emph{cells}, whose morphisms
we call 2-\emph{cells}, and whose composition law we call \emph{vertical composition}
\item for each 0-cell $\mathcal{A}$, a functor $u_\mathcal{A} \colon \mathbbm{1} \to \mathscr{C}(\mathcal{A}, \mathcal{A})$, called the \emph{unit}
\item for any $\mathcal{A}, \mathcal{B}, \mathcal{C}$ in $|\mathscr{C}|$, a functor $\circ_{\mathcal{A}, \mathcal{B}, \mathcal{C}} \colon \mathscr{C}(\mathcal{B}, \mathcal{C}) \times \mathscr{C}(\mathcal{A}, \mathcal{B}) \to \mathscr{C}(\mathcal{A}, \mathcal{C})$, called
\emph{horizontal composition}
\end{itemize}
satisfying associativity and unitality conditions, that is, commutativity of the two diagrams
$$
\xymatrix@C=3em{
\mathscr{C}(\mathcal{C}, \mathcal{D}) \times \mathscr{C}(\mathcal{B}, \mathcal{C}) \times \mathscr{C}(\mathcal{A}, \mathcal{B})\ar[d]_-{\circ_{\mathcal{B}, \mathcal{C}, \mathcal{D}} \times 1} \ar[rr]^-{1 \times \circ_{\mathcal{A}, \mathcal{B}, \mathcal{C}}} && \mathscr{C}(\mathcal{C}, \mathcal{D}) \times \mathscr{C}(\mathcal{A}, \mathcal{C}) \ar[d]^-{\circ_{\mathcal{A}, \mathcal{C}, \mathcal{D}}} \\
\mathscr{C}(\mathcal{B}, \mathcal{D}) \times \mathscr{C}(\mathcal{A}, \mathcal{B}) \ar[rr]_-{\circ_{\mathcal{A}, \mathcal{B}, \mathcal{D}} } && \mathscr{C}(\mathcal{A}, \mathcal{D})
}
$$
$$
\xymatrix@C=3em{
\mathscr{C}(\mathcal{A}, \mathcal{B}) \ar@{=}[drr] \ar[rr]^-{u_\mathcal{A} \times 1} \ar[d]_-{1 \times u_\mathcal{B}} && \mathscr{C}(\mathcal{B}, \mathcal{B}) \times \mathscr{C}(\mathcal{A}, \mathcal{B}) \ar[d]^-{\circ_{\mathcal{A}, \mathcal{B}, \mathcal{B}}} \\
\mathscr{C}(\mathcal{A}, \mathcal{B}) \times \mathscr{C}(\mathcal{A}, \mathcal{A}) \ar[rr]_-{\circ_{\mathcal{A}, \mathcal{A}, \mathcal{B}} } & & \mathscr{C}(\mathcal{A}, \mathcal{B})
}
$$
for all $\mathcal{A}, \mathcal{B}, \mathcal{C}, \mathcal{D} \in |\mathscr{C}|$.
\end{defn}
From this point onward, we omit the subscripts on the functors $\circ$ and $u$, much as we do for natural transformations.
We denote a 1-cell $F$ in $\mathscr{C}(\mathcal{A}, \mathcal{B})$ by $F \colon \mathcal{A} \to \mathcal{B}$, and a 2-cell $\alpha$ between $F,F'$ is denoted by $\alpha \colon F \Rightarrow F'$. We denote both horizontal and vertical composition by concatenation, or occasionally by the symbol $\circ$. It is always clear from the context to which type of composition we refer.
A 2-cell inside a diagram of 1-cells denotes a 2-cell between their horizontal composites, e.g.\
$$
\xymatrix{
\mathcal{A} \ar[r]^-F \ar[d]_-H & \mathcal{B} \ar[d]^-G \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-\alpha "a";"b"\\
\mathcal{C} \ar[r]_-K \ar[r]_-K & \mathcal{D}
}
$$
means that $\alpha \colon GF \Rightarrow KH$ is a 2-cell. In diagrams consisting only of 2-cells, we usually abandon the double arrows $\Rightarrow$ in favour of regular arrows $\to$ for the sake of readability.
We use the symbol $1$ to denote both the identity 1-cell and 2-cell, i.e.\ the images of the unique morphism and object respectively, under the unit functor $u$. However, when we write a horizontal composite involving an identity 2-cell, we rather write the \emph{corresponding 1-cell}. We do this because it makes it easy to reference individual cells, while also making diagrams easier to interpret; for example, given a diagram
$$
\xymatrix{
\mathcal{A} \rtwocell^F_{F'}{\ \alpha} & \mathcal{B} \ar[r]^-G & \mathcal{C}
}
$$
we denote the horizontal composite with the identity by
$$
\xymatrix{
\mathcal{A} \rrtwocell^{GF}_{GF'}{\ \ \ G \alpha} && \mathcal{C}
}
$$
There are various duals one obtains by reversing some of the structure in a 2-category. We denote by $\mathscr{C}^*$ the
2-category obtained by reversing all 1-cells in
$\mathscr{C}$, and by $\mathscr{C}_*$ we denote the
2-category obtained by reversing all 2-cells in
$\mathscr{C}$. Of course, $(\mathscr{C}_*)^* = (\mathscr{C}^*)_*$ so there is no harm in writing ${\mathscr{C}}^*_*$ to denote either of these duals.
We may view an ordinary category $\mathcal{C}$ as a 2-category, where for two objects $A,B$ in $\mathcal{C}$, the category $\mathcal{C}(A,B)$ is the discrete category on the morphisms $A \to B$. Using the above notation, $\mathcal{C}^*$ can thus be viewed as the dual (or opposite) category to $\mathcal{C}$ in the ordinary sense.
\begin{exa}
The paradigmatic example of a 2-category is $\mathbf{Cat}$. This is the 2-category whose 0-, 1- and 2-cells are categories, functors and natural transformations respectively.
Most examples of 2-categories in this thesis appear as constructions based on $\mathbf{Cat}$.
\end{exa}
\subsection{2-functors and 2-natural transformations}
\begin{defn}
Let $\mathscr{C}, \mathscr{D}$ be 2-categories. A 2-\emph{functor} $\Phi \colon \mathscr{C} \to \mathscr{D}$ consists of
\begin{itemize}
\item a function $\Phi \colon |\mathscr{C}| \to |\mathscr{D}|$
\item for each pair $\mathcal{A}, \mathcal{B} \in \mathscr{C}$, a functor $\Phi_{\mathcal{A}, \mathcal{B}} \colon \mathscr{C}(\mathcal{A}, \mathcal{B}) \to \mathscr{D}(\Phi\mathcal{A}, \Phi\mathcal{B})$
\end{itemize}
that are compatible with both horizontal and vertical composition, that is, the diagrams
$$
\xymatrix{
\mathscr{C}(\mathcal{B}, \mathcal{C}) \times \mathscr{C}(\mathcal{A}, \mathcal{B}) \ar[r]^-{\circ} \ar[d]_-{\Phi_{\mathcal{B}, \mathcal{C}} \times \Phi_{\mathcal{A}, \mathcal{B}}} & \mathscr{C}(\mathcal{A}, \mathcal{C}) \ar[d]^-{\Phi_{\mathcal{A}, \mathcal{C}}} \\
\mathscr{D}(\Phi\mathcal{B}, \Phi\mathcal{C}) \times \mathscr{D}(\Phi\mathcal{A}, \Phi\mathcal{B}) \ar[r]_-\circ & \mathscr{D}(\Phi\mathcal{A}, \Phi\mathcal{C})
}
\qquad
\xymatrix{
\mathbbm{1} \ar[r]^-u \ar[dr]_-u & \mathscr{C}(\mathcal{A},\mathcal{A}) \ar[d]^-{\Phi_{\mathcal{A}, \mathcal{A}}} \\
& \mathscr{D}(\Phi\mathcal{A}, \Phi\mathcal{A})
}
$$
commute for all $\mathcal{A}, \mathcal{B}, \mathcal{C} \in |\mathscr{C}|$.
\end{defn}
Thus a 2-functor is simply a mapping of cells between 2-categories that preserves commutative diagrams of 1-cells and 2-cells, as well as identities.
We now omit the subscripts on a 2-functor $\Phi$, similar again to the case for the composition and unit of a 2-category, and natural transformations.
\begin{exa}\label{hom2functors}
For each 0-cell $\mathcal{A}$ in $\mathscr{C}$, there are \emph{hom 2-functors}
\begin{align*}
\mathscr{C}(\mathcal{A}, -) \colon \mathscr{C} \to \mathbf{Cat} \\
\mathscr{C}(-, \mathcal{A}) \colon \mathscr{C}^* \to \mathbf{Cat}
\end{align*}
defined in the obvious way.
\end{exa}
There is also a notion of morphism between 2-functors:
\begin{defn}
Let $\Phi, \Phi' \colon \mathscr{C} \to \mathscr{D}$ be 2-functors. A 2-\emph{natural transformation}, which we denote by $\nu \colon \Phi \Rightarrow \Phi'$, is given by, for each $\mathcal{A} \in |\mathscr{C}|$, a 1-cell $\nu_{\mathcal{A}} \colon \Phi\mathcal{A} \to \Phi'\mathcal{A}$ such that the diagram
$$
\xymatrix{
\mathscr{C}(\mathcal{A}, \mathcal{B}) \ar[rr]^-\Phi \ar[d]_-{\Phi '} && \mathscr{D}(\Phi\mathcal{A}, \Phi\mathcal{B})\ar[d]^-{\mathscr{D}(\Phi\mathcal{A}, \nu_{\mathcal{B}})} \\
\mathscr{D}(\Phi'\mathcal{A}, \Phi'\mathcal{B}) \ar[rr]_-{\mathscr{D}(\nu_{\mathcal{A}}, \Phi'\mathcal{B})} && \mathscr{D}(\Phi\mathcal{A}, \Phi'\mathcal{B})
}
$$
commutes for all $\mathcal{A}, \mathcal{B} \in |\mathscr{C}|$.\end{defn}
Again, we usually omit the indices on the 1-cell components of 2-natural transformations.
\begin{exa}
There is a 2-category $\operatorname{2-\mathbf{Cat}}$ whose 0-cells are 2-categories, whose 1-cells are 2-functors and whose 2-cells are 2-natural transformations.
\end{exa}
\subsection{Adjunctions}
One advantage of studying 2-categories as opposed to just ordinary categories is that it allows us to study certain phenomena as being \emph{internal}. For example, instead of thinking of an adjunction \emph{between} categories, we can think of an adjunction \emph{inside} a 2-category:
\begin{defn}
Let $F \colon \mathcal{A} \to \mathcal{B}$ and $U \colon \mathcal{B} \to \mathcal{A}$ be 1-cells in a 2-category $\mathscr{C}$. We say that $F$ is \emph{left adjoint} to $U$ if there are 2-cells $\eta \colon 1 \Rightarrow UF$, $\epsilon \colon FU \Rightarrow 1$ called the \emph{unit} and \emph{counit} respectively, such that the two \emph{triangle identities} hold, i.e.\ the diagrams
$$
\xymatrix{
F \ar[r]^-{F \eta} \ar@{=}[dr] & FUF\ar[d]^-{\epsilon F} \\
& F
}\qquad
\xymatrix{
U \ar[r]^-{\eta U} \ar@{=}[dr] & UFU \ar[d]^-{U \epsilon} \\
& U
}
$$
commute. In this situation, we also say that $U$ is \emph{right adjoint to} $F$, and that $F\dashv U$ is an \emph{adjunction}.
\end{defn}
\begin{exa}
An adjunction in $\mathbf{Cat}$ is just an ordinary adjunction between categories.
\end{exa}
\begin{exa}
An adjunction in $\operatorname{2-\mathbf{Cat}}$ is called a \emph{2-adjunction}. A 2-adjunction
$$
\xymatrix{
\mathscr{C} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-\Phi & & \ar@/^0.5pc/[ll]^-\Theta \mathscr{D}
}
$$
with unit $\nu \colon 1 \Rightarrow \Theta\Phi$ and counit $\xi \colon \Phi \Theta \Rightarrow 1$ can be described equivalently as having 2-natural isomorphisms
$$
\mathscr{C}(\mathcal{A}, \Theta \mathcal{E}) \cong \mathscr{D}(\Phi\mathcal{A}, \mathcal{E}).
$$
explicitly given by
\begin{align*}
\xymatrix{
\mathcal{A} \rrtwocell^X_{X'}{\alpha} & & \Theta\mathcal{E} & \longmapsto & \Phi\mathcal{A}\rrtwocell^{\Phi X}_{\Phi X'}{\ \ \Phi \alpha} & & \Phi \Theta \mathcal{E} \ar[r]^-{\xi \mathcal{E}} & \mathcal{E}
} \\
\xymatrix{
\Phi \mathcal{A} \rrtwocell^Z_{Z'}{\beta} & & \mathcal{E} & \longmapsto & \mathcal{A} \ar[r]^-{\nu \mathcal{A}} & \Theta\Phi\mathcal{A} \rrtwocell^{\Theta Z}_{\Theta Z'}{\ \ \Theta \beta} & & \Theta \mathcal{E}
}
\end{align*}
\end{exa}
Often 2-functors preserve these internal properties. Indeed:
\begin{prop}
All 2-functors preserve adjunctions.
\end{prop}
\begin{proof}
If $F \dashv U$ is an adjunction with unit $\eta$ and counit $\epsilon$, and $\Phi$ is a 2-functor, then the triangle identities for $\Phi\eta$ and $\Phi \epsilon$ are satisfied since 2-functors preserve identities and all forms of composition. Hence $\Phi F \dashv \Phi U$.
\end{proof}
\section{Monads and comonads in 2-categories}\label{monadsandcomonads}
Throughout the remainder of this chapter, let $\mathscr{C}$ be a 2-category. In subsequent chapters, the 2-categorical constructions presented here
are considered only in the case $\mathscr{C} = \mathbf{Cat}$, and are explicitly described in that way in Section~\ref{emem}, so if the reader is not entirely comfortable with the language of 2-categories, they may replace appearances of 0-, 1- and 2-cells in $\mathscr{C}$ with the words `category', `functor' and `natural transformation' respectively.
We restate, in our terminology, some of the definitions and results involving monads in \cite{MR0299653}. We also emphasise results for comonads since they and their interplay with monads are an important topic in later chapters.
\subsection{The 2-category of monads}
\begin{defn}
Let $\mathcal{A}$ be a 0-cell in $\mathscr{C}$. A \emph{monad on $\mathcal{A}$ in $\mathscr{C}$} is a triple $(B, \mu, \eta)$ consisting of a 1-cell $B \colon \mathcal{A} \to \mathcal{A}$ together with 2-cells $\eta \colon 1 \Rightarrow B$, $\mu \colon BB \Rightarrow B$, called the \emph{unit} and \emph{multiplication} respectively, such that the following two diagrams commute:
$$
\xymatrix{
BBB \ar[r]^-{B \mu} \ar[d]_-{\mu B} & BB \ar[d]^-\mu \\
BB \ar[r]_-\mu & B
}\qquad
\xymatrix{
B \ar@{=}[dr] \ar[r]^-{B \eta} \ar[d]_-{\eta B} & BB \ar[d]^-{\mu} \\
BB \ar[r]_-{\mu} & B
}
$$
\end{defn}
By abuse of notation, we refer to a monad by its underlying 1-cell, and we always use the symbols $\eta$ and $\mu$ to refer to the unit and multiplication of an arbitrary monad.
Monads in $\mathscr{C}$ constitute the 0-cells of a 2-category $\operatorname{Mnd}(\mathscr{C})$, defined as follows. The 1-cells $(\mathcal{A},B) \to (\mathcal{D},A )$ consist of pairs $(\Sigma, \sigma)$ where $\Sigma \colon \mathcal{A} \to \mathcal{D}$ is a 1-cell in $\mathscr{C}$ and $\sigma \colon A\Sigma \Rightarrow \Sigma B$ is a 2-cell in $\mathscr{C}$, subject to the commutativity conditions
$$
\xymatrix{
AA\Sigma \ar[d]_-{\mu \Sigma} \ar[r]^-{A \sigma} & A \Sigma B \ar[r]^-{\sigma B} & \Sigma BB \ar[d]^-{\Sigma \mu} \\
A \Sigma \ar[rr]_-{\sigma} & & \Sigma B
}\qquad
\xymatrix{
\Sigma \ar[dr]_-{\Sigma \eta} \ar[r]^-{\eta \Sigma} & A \Sigma \ar[d]^-\sigma \\
& \Sigma B
}
$$
We call these 1-cells \emph{morphisms of monads}.
A 2-cell $\alpha \colon (\Sigma, \sigma) \Rightarrow (\Sigma', \sigma')$ is a 2-cell $\alpha \colon \Sigma \Rightarrow \Sigma'$ in $\mathscr{C}$ such that the diagram
$$
\xymatrix{
A \Sigma \ar[d]_-{A \alpha} \ar[r]^-\sigma & \Sigma B \ar[d]^-{\alpha B} \\
A \Sigma ' \ar[r]_-{\sigma '} & \Sigma' B
}
$$
commutes.
Dually, we define a \emph{comonad} $(T, \delta, \epsilon)$ in $\mathscr{C}$ to be a monad in $\mathscr{C}_*$. We define the 2-category of comonads in $\mathscr{C}$ as $\operatorname{Cmd}(\mathscr{C}):= \operatorname{Mnd}(\mathscr{C}_*)_*$. The 1-cells herein are called \emph{morphisms of comonads}. We say \emph{opmorphism of (co)monads} to mean a morphism of (co)monads in $\mathscr{C}^*$.
\begin{exa}\label{monadj}
Suppose that we have an adjunction
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{B}
}
$$
in $\mathscr{C}$ with unit $\eta$ and counit $\epsilon$.
The 1-cell $UF$ becomes a monad on $\mathcal{A}$, and dually $FU$ becomes a comonad on $\mathcal{B}$, and we say that these are \emph{generated by the adjunction}. The
(co)units and (co)multiplications are given by
\begin{align*}
\xymatrix{
1 \ar[r]^-{\eta} & UF} & & & \xymatrix{FU \ar[r]^-\epsilon & 1} \\
\xymatrix{UFUF \ar[r]^-{U \epsilon F} & UF}& & &\xymatrix{FU \ar[r]^-{F \eta U} & FUFU}
\end{align*}
\end{exa}
\begin{rem}
Since the taking of duals is confusing, let us make explicit that if
$$\xymatrix{
(\mathcal{A}, B) \ar[rr]^-{(\Sigma, \sigma)} && (\mathcal{D}, A),
}\qquad
\xymatrix{
(\mathcal{B}, T) \ar[rr]^-{(\Gamma, \gamma)} && (\mathcal{E}, G)
}
$$
are morphisms of monads and comonads respectively, then the underlying 2-cells are given by $\sigma \colon A\Sigma \Rightarrow \Sigma B$ and $\gamma \colon \Gamma T \Rightarrow G \Gamma$. Note the differing positions of $\Gamma, \Sigma$ in each case, telling us that these morphisms are not of the same \emph{variance}. If we take opmorphisms instead, these 2-cells would reverse direction.
\end{rem}
\begin{exa}\label{trivialdist}
Consider a monad $(\mathcal{A}, B)$ in $\mathscr{C}$. The 2-cell
$$
\xymatrix{
\tau \colon BB \ar[rr]^-{BB\eta} && BBB \ar[r]^-{\mu B} & BB
}
$$
defines a morphism of monads
$$
\xymatrix{
(\mathcal{A}, B) \ar[rr]^-{(B, \tau)} && (\mathcal{A}, B).
}
$$
Furthermore, any morphism of monads $(\Sigma, \sigma) \colon (\mathcal{A}, B) \to (\mathcal{D}, A)$ induces a monad 2-cell
$$
\xymatrix{
(\mathcal{A}, B) \ar[rr]^-{(\Sigma, \sigma)} \ar[dd]_-{(B, \tau)} && (\mathcal{D},A) \ar[dd]^-{(A, \tau)} \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\sigma} "a";"b" \\
\\
(\mathcal{A}, B) \ar[rr]_-{(\Sigma, \sigma)} && (\mathcal{D}, A)
}$$
\end{exa}
As in the case of adjunctions, 2-functors preserve the internal property of being a monad. Thus, a 2-functor
$\Phi \colon \mathscr{C} \to \mathscr{D}$ restricts to a 2-functor
$$\xymatrix{ \operatorname{Mnd}(\mathscr{C}) \ar[rr]^-{\operatorname{Mnd}(\Phi)} && \operatorname{Mnd}(\mathscr{D}),}$$ defined diagramatically by
$$
\xymatrix{
(\mathcal{A}, B ) \ddtwocell<9>_{(\Sigma, \sigma)\ \ \ \ \ }^{\ \ \ \ \ \ \ (\Sigma', \sigma')}{^\alpha} & & & & (\Phi\mathcal{A}, \Phi B) \ddtwocell<9>_{(\Phi\Sigma, \Phi\sigma)\ \ \ \ \ \ \ \ \ }^{\ \ \ \ \ \ \ \ \ \ (\Phi\Sigma', \Phi\sigma')}{^\Phi\alpha} \\
& & \longmapsto \\
(\mathcal{D}, A) & & & & (\Phi\mathcal{D}, \Phi A)
}
$$
A 2-natural transformation $\nu \colon \Phi \Rightarrow \Phi'$ clearly induces another 2-natural transformation $\operatorname{Mnd}(\nu)$. We thus obtain the following:
\begin{prop}
The above assignment defines a 2-functor $\operatorname{Mnd} \colon \operatorname{2-\mathbf{Cat}} \to \operatorname{2-\mathbf{Cat}}$.
\end{prop}
Dually, there is a 2-functor $\operatorname{Cmd} \colon \operatorname{2-\mathbf{Cat}} \to \operatorname{2-\mathbf{Cat}}$.
\subsection{Distributive laws}\label{finallyoversection}
\begin{defn}\label{finallyover}
A \emph{distributive law of comonads} is a comonad in $\operatorname{Cmd}(\mathscr{C})$. Explicitly, a distributive law between comonads $T,S$ on the same 1-cell $\mathcal{B}$ is a 2-cell $\chi \colon TS \Rightarrow ST$ such that the four diagrams
\begin{align*}
\xymatrix{
TS \ar[d]_-{\delta S} \ar[rr]^-\chi & & ST \ar[d]^-{S \delta} \\
TTS \ar[r]_-{T \chi} & TST \ar[r]_-{\chi T} & STT
}\qquad
\xymatrix{
TS \ar[dr]_-{\epsilon S} \ar[r]^-\chi & ST \ar[d]^-{S \epsilon} \\
& S
} \\
\xymatrix{
TS \ar[d]_-{T\delta} \ar[rr]^-\chi & & ST \ar[d]^-{\delta T} \\
TSS \ar[r]_-{\chi S} & STS \ar[r]_-{ S\chi } & SST
}
\qquad
\xymatrix{
TS \ar[dr]_-{T \epsilon} \ar[r]^-\chi & ST \ar[d]^-{\epsilon T} \\
& T
}
\end{align*}
commute.
\end{defn}
We denote by $\cDist(\mathscr{C})$ the 2-category $\operatorname{Cmd}(\operatorname{Cmd}(\mathscr{C})^*)^*$. Thus, explicitly,
\begin{itemize}
\item $0$-cells are quadruples $(\mathcal{B}, \chi, T, S)$,
where $\chi \colon TS \Rightarrow ST$ is a
comonad distributive law on $\mathcal{B}$,
\item $1$-cells $$(\mathcal{B}, \chi, T, S) \rightarrow (\mathcal{D},
\tau, G, C)$$ are triples $(\Sigma,
\sigma, \gamma)$, where $(\Sigma, \sigma) \colon (\mathcal{B}, T) \rightarrow (\mathcal{D}, G)$ is
an opmorphism of comonads, and $(\Sigma, \gamma)
\colon (\mathcal{B} , S) \rightarrow (\mathcal{D}, C)$ is a
morphism of comonads satisfying the Yang-Baxter
equation, i.e.\
$$
\xymatrix@R=0.5em{ & \Sigma TS \ar[r]^-{\Sigma
\chi} & \Sigma ST \ar[dr]^-{\gamma T} & \\
G\Sigma S \ar[dr]_-{G \gamma} \ar[ur]^-{\sigma
S} & & & C \Sigma T \\ & GC\Sigma
\ar[r]_-{\tau \Sigma} & CG \Sigma \ar[ur]_-{C
\sigma} & }
$$
commutes, and
\item $2$-cells $(\Sigma,
\sigma, \gamma) \Rightarrow (\Sigma', \sigma', \gamma')$
are 2-cells $\alpha \colon \!\Sigma \Rightarrow
\Sigma'$ in $\mathscr{C}$ for which the diagrams
$$
\xymatrix{
G\Sigma \ar[d]_-\sigma \ar[r]^-{G \alpha}
& G \Sigma '
\ar[d]^-{\sigma '} \\ \Sigma T \ar[r]_-{\alpha T } &
\Sigma' T }
\quad \quad \quad
\xymatrix{ \Sigma S
\ar[r]^-{\alpha S} \ar[d]_-\gamma & \Sigma' S
\ar[d]^-{\gamma '} \\ C \Sigma \ar[r]_-{C \alpha}& C
\Sigma' }
$$
commute.
\end{itemize}
Similarly, we define the $2$-category of \emph{mixed distributive laws} in $\mathscr{C}$ as $$\cMix(\mathscr{C}):=\operatorname{Mnd}(\operatorname{Cmd}(\mathscr{C})).$$
Let us unpack this definition.
Consider an arbitrary 1-cell in the 2-category $\cMix(\mathscr{C})$:
$$
\xymatrix{
((\mathcal{A}, C), (B, \theta) ) \ar[rr]^-{((\Sigma, \gamma), \sigma)} && ((\mathcal{D}, D), (A, \tau) ).
}
$$
We have that:
\begin{itemize}
\item $(\mathcal{A}, C)$ and $(\mathcal{D}, D)$ are comonads in $\mathscr{C}$;
\item $(B, \theta)$ and $(A, \tau)$ are monads in $\operatorname{Cmd}(\mathscr{C})$, meaning that $$\theta \colon BC \Rightarrow CB, \qquad \tau \colon AD \Rightarrow DA$$ are 2-cells in $\mathscr{C}$, compatible with the appropriate monad and comonad structures;
\item $(\Sigma, \gamma)$ is a morphism of comonads, so in particular $\Sigma \colon \mathcal{A} \to \mathcal{D}$ is a 1-cell in $\mathscr{C}$, and $\gamma \colon \Sigma C \Rightarrow D \Sigma$ is a 2-cell in $\mathscr{C}$ compatible with the comonad structures of $C$ and $D$;
\item $\sigma \colon (A, \tau) \circ (\Sigma, \gamma) \Rightarrow (\Sigma, \gamma) \circ (B, \theta)$ is a 2-cell in $\operatorname{Cmd}(\mathscr{C})$ which corresponds to a Yang-Baxter-esque commutative hexagon;
\item $\sigma \colon A\Sigma \Rightarrow \Sigma B$ is a 2-cell compatible with the monad structures of $A$ and $B$.
\end{itemize}
The above data exactly defines a $1$-cell
$$
\xymatrix{
((\mathcal{A}, B), (C, \theta) ) \ar[rr]^-{((\Sigma, \sigma), \gamma)} & & ((\mathcal{D}, A), (D, \tau))
}
$$
in the 2-category $\operatorname{Cmd}(\operatorname{Mnd}(\mathscr{C}))$. Comparing the 2-cells in a similar way shows that:
\begin{lem}\label{mndcmdiso}
There is a 2-isomorphism $$\operatorname{Mnd}(\operatorname{Cmd}(\mathscr{C})) \cong \operatorname{Cmd}(\operatorname{Mnd}(\mathscr{C})).$$
\end{lem}
We represent the 1-cells of $\cMix(\mathscr{C})$ hereafter as
$$
\xymatrix{
(\mathcal{A}, \theta, B, C ) \ar[rr]^-{(\Sigma, \sigma, \gamma)} & & (\mathcal{D}, \tau, A, D ),
}
$$
so our notation aligns with that of the category $\cDist(\mathscr{C})$.
\subsection{Eilenberg-Moore constructions}\label{EM2cat}
Let $\mathcal{A}$ be a 0-cell in $\mathscr{C}$. The identity 1-cell $1 \colon \mathcal{A} \to \mathcal{A}$ is part of a monad with multiplication and unit both given by $1$. This defines an inclusion 2-functor $I \colon \mathscr{C} \to \operatorname{Mnd}(\mathscr{C})$.
\begin{defn}
We say that $\mathscr{C}$ admits \emph{Eilenberg-Moore constructions for monads} if the inclusion 2-functor $I$ has a right 2-adjoint $J \colon \operatorname{Mnd}(\mathscr{C}) \to \mathscr{C}$.
\end{defn}
If this 2-functor exists, its action on a morphism of monads $(\Sigma, \sigma) \colon (\mathcal{A}, B) \to (\mathcal{D}, A)$ is denoted by $\Sigma^\sigma \colon \mathcal{A}^B \to \mathcal{D}^A$, and its action on a 2-cell $\alpha \colon (\Sigma, \sigma) \Rightarrow (\Sigma', \sigma')$ is denoted by $\tilde \alpha \colon \Sigma^\sigma \Rightarrow \Sigma'^{\sigma'}$. The 2-adjunction part of the definition explicitly means that there are 2-natural isomorphisms
$$
\operatorname{Mnd}(\mathscr{C}) ( (\mathcal{A}, 1), (\mathcal{D}, A) ) \cong \mathscr{C}\left(\mathcal{A}, \mathcal{D}^A\right)
$$
of hom-categories.
Let $\nu \colon 1 \Rightarrow JI$ and $\xi \colon IJ \Rightarrow 1$ denote the unit and counit of $I\dashv J$, respectively. For each monad $(\mathcal{A}, B)$ the counit has a component $(\mathcal{A}^B, 1) \to (\mathcal{A}, B)$ . This is a morphism of monads, and we denote it by $(U^B, \kappa)$, where $\kappa \colon BU^B \Rightarrow U^B$ is a 2-cell in $\mathscr{C}$ such that the diagrams
$$
\xymatrix{
U^B \ar[r]^-{\eta U^B}\ar@{=}[dr] & BU^B \ar[d]^-\kappa \\
& U^B
}\qquad
\xymatrix{
BBU^B \ar[r]^-{B\kappa} \ar[d]_-{\mu U^B} & BU^B \ar[d]^-\kappa \\
BU^B \ar[r]_-\kappa & B
}
$$
commute. The monad axioms tell us that $(B, \mu) \colon (\mathcal{A}, 1) \to (\mathcal{A}, B)$ is a morphism of monads, and this corresponds under the 2-adjunction $I\dashv J$ to a 1-cell $F^B\colon \mathcal{A} \to \mathcal{A}^B$ in $\mathscr{C}$, unique such that
\begin{equation}\label{EMmonad}
\begin{array}{c}
\xymatrix{
(\mathcal{A}, 1) \ar[r]^-{(F^B, 1)} \ar[dr]_-{(B, \mu)} & (\mathcal{A}^B, 1) \ar[d]^-{(U^B,\kappa)} \\
& (\mathcal{A}, B)
}
\end{array}
\end{equation}
commutes in $\operatorname{Mnd}(\mathscr{C})$. In particular this tells us that $U^BF^B = B$ and $\kappa F^B = \mu$. We have that
$$
\xymatrix{
(\mathcal{A}^B, 1) \rrtwocell^{<1.5>(U^BFU^B,~\mu U^B)}_{<1.5>(U^B,~\kappa)}{\kappa}&& (\mathcal{A}, B)
}
$$
is a 2-cell in $\operatorname{Mnd}(\mathscr{C})$ corresponding under the 2-adjunction to a 2-cell
$$
\xymatrix{
\mathcal{A}^B \rrtwocell^{<1.5>F^BU^B}_{1}{\epsilon} && \mathcal{A}^B
}
$$
in $\mathscr{C}$, unique such that $U^B \epsilon = \kappa$. It turns out that $\eta$ is the unit and $\epsilon$ is the counit of an adjunction $F^B \dashv U^B$; see~\cite[p.~152]{MR0299653} for the full details of this construction.
As explained in Example~\ref{monadj}, the adjunction $F^B \dashv U^B$ generates the monad $B$ as well a comonad $F^BU^B \colon \mathcal{A}^B \to \mathcal{A}^B$ with counit $\epsilon$ and comultiplication $F^B \eta U^B$. We write $\tilde B$ to denote this comonad.
\begin{prop}
Let $B = 1$ be the identity monad on $\mathcal{A}$. Then $F^B \colon \mathcal{A} \to \mathcal{A}^B$ is the unit of $I \dashv J$ evaluated at $\mathcal{A}$ and is an isomorphism. Furthermore, $\tilde B = 1$.
\end{prop}
\begin{proof}
By diagram~\ref{EMmonad}, we have that $U^BF^B = 1$ and $\kappa = 1$. It follows that $\epsilon = 1$ since it is the unique 2-cell such that $U^B \epsilon = 1$. Hence $\tilde B = F^BU^B = 1$ and $F^B,U^B$ are isomorphisms. By the triangle identities for $I\dashv J$, Diagram~\ref{EMmonad} commutes for the choice $F^B = \nu$, so this must be the only choice by uniqueness.
\end{proof}
We henceforth identify $JI$ with 1.
If $(\Sigma, \sigma) \colon (\mathcal{A}, 1)\to (\mathcal{D}, A)$ is an arbitrary morphism of monads, then it corresponds to
$$
\xymatrix{
\mathcal{A} \ar[r]^-\nu_-\cong & \mathcal{A}^1 \ar[r]^-{\Sigma^\sigma} & J(\mathcal{D}, A)
}
$$
under the 2-adjunction, which we write as $\Sigma^\sigma \colon \mathcal{A} \to \mathcal{D}^A$. In a similar fashion, a morphism of monads $(\Sigma, \sigma) \colon (\mathcal{A}, B) \to (\mathcal{D}, 1)$ corresponds under the 2-adjunction to $\Sigma^\sigma \colon \mathcal{A}^B \to \mathcal{D}$.
\begin{prop}\label{cmdEM}
If $\mathscr{C}$ admits Eilenberg-Moore constructions for monads, then so too does $\operatorname{Cmd}(\mathscr{C})$.
\end{prop}
\begin{proof}
Since 2-functors preserve adjunctions, $\operatorname{Cmd}(J) \colon \operatorname{Cmd}(\operatorname{Mnd}(\mathscr{C})) \to \operatorname{Cmd}(\mathscr{C})$ is a right 2-adjoint to $\operatorname{Cmd}(I)$. After composing with the isomorphism $$\operatorname{Mnd}(\operatorname{Cmd}(\mathscr{C})) \cong \operatorname{Cmd}(\operatorname{Mnd}(\mathscr{C}))$$ of Lemma~\ref{mndcmdiso}, the 2-functor $\operatorname{Cmd}(I)$ becomes the inclusion 2-functor, and thus we have constructed its right 2-adjoint, as required.
\end{proof}
Since $\cMix(\mathscr{C}) = \operatorname{Mnd}(\operatorname{Cmd}(\mathscr{C}))$, we have given a 2-functor $\cMix(\mathscr{C}) \to \operatorname{Cmd}(\mathscr{C})$. Diagramatically it is represented as
$$
\xymatrix{
(\mathcal{A}, \theta, B, C) \ddtwocell<10>_{(\Sigma, \sigma, \gamma)\ \ \ \ \ \ \ }^{\ \ \ \ \ \ \ \ \ \ (\Sigma', \sigma', \gamma')}{^\alpha} & & & & (\mathcal{A}^B, C^\theta) \ddtwocell<10>_{(\Sigma^\sigma, \tilde\gamma)\ \ \ \ \ \ \ }^{\ \ \ \ \ \ \ \ \ \ (\Sigma'^{\sigma '}, \tilde\gamma')}{^\tilde\alpha} \\
& & \longmapsto \\
(\mathcal{D}, \psi, A, D) & & & & (\mathcal{D}^A, D^\psi)
}
$$
where the notation $\tilde\gamma$ is explained as follows: by the Yang-Baxter equation, $\gamma$ is a 2-cell of monads
$$
\xymatrix{
(\mathcal{A}, B) \ar[rr]^-{(C, \theta)} \ar[d]_-{(\Sigma, \sigma)} & & \ar@{}[dll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-\gamma "a";"b" (\mathcal{A}, B) \ar[d]^-{(\Sigma, \sigma)} \\
(\mathcal{D}, A) \ar[rr]_-{(D, \psi)} & & (\mathcal{D}, A)
}
$$
and applying $J \colon \operatorname{Mnd}(\mathscr{C}) \to \mathscr{C}$ to this square gives a 2-cell
$$
\xymatrix{
\mathcal{A}^B \ar[rr]^-{C^\theta} \ar[d]_-{\Sigma^\sigma} && \ar@{}[dll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\tilde\gamma} "a";"b" \mathcal{A}^B \ar[d]^-{\Sigma^\sigma} \\
\mathcal{D}^A \ar[rr]_-{D^\psi} & & \mathcal{D}^A
}
$$
Since $J \colon \operatorname{Mnd}(\mathscr{C}) \to \mathscr{C}$ is a 2-functor, it sends comonads to comonads. In this situation, a mixed distributive law $\theta \colon BC \Rightarrow CB$ is sent to a comonad $C^\theta$.
Using Proposition~\ref{cmdEM} we may freely replace $\mathscr{C}$ with $\operatorname{Cmd}(\mathscr{C})$ in any statement made about a 2-category $\mathscr{C}$ which admits Eilenberg-Moore constructions for monads.
\section{Lifting through adjunctions}\label{liftingthroughadjunctions}
Here we discuss distributive
laws that are compatible in a specific way with adjunctions that generate one of the
involved comonads. We go on to explain, given a 2-category $\mathscr{C}$ which admits Eilenberg-Moore constructions for monads, how the comparison 1-cell of~\cite{MR0299653} lifts to become a 1-cell between comonad distributive laws and that there is a canonical 2-functor $ \cMix(\mathscr{C}) \to \cDist(\mathscr{C})$.
\subsection{The lifting theorem}
Let $\mathscr{C}$ be a 2-category. Consider squares in $\mathscr{C}$ of the form
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{D} \ar[r]_V & \mathcal{C}}
$$
We obtain two 2-categories $\mathrm{Sq}^h(\mathscr{C})$ and $\mathrm{Sq}^v(\mathscr{C})$ by defining such squares to be 1-cells and defining their composition by pasting horizontally and vertically respectively (really these are the horizontal and vertical components of a double category $\mathrm{Sq}(\mathscr{C})$, see~\cite[Observation~76]{MR2399898}). The 2-cells are pairs of 2-cells in $\mathscr{C}$ that satisfy the obvious compatibility condition.
\begin{prop}\label{monadsq}
To give a monad in $\mathrm{Sq}^v(\mathscr{C})$ is the same as to give a morphism of monads in $\mathscr{C}$.
\end{prop}
\begin{proof}
A square
$$
\xymatrix{\mathcal{A} \ar[r]^ \Sigma \ar[d]_ B & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\sigma} "a";"b" \mathcal{D} \ar[d]^ A\\
\mathcal{A} \ar[r]_\Sigma & \mathcal{D}}
$$
is a monad in $\mathrm{Sq}^v(\mathscr{C})$ if and only if $B,A $ are monads and $$\xymatrix{(\mathcal{A}, B) \ar[rr]^-{(\Sigma, \sigma)} && (\mathcal{D}, A)}$$ is a morphism of monads.
\end{proof}
Dually, to give a comonad in $\mathrm{Sq}^v(\mathscr{C})$ is the same as to give an opmorphism of comonads in $\mathscr{C}$.
Suppose that we have a square
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{D} \ar[r]_V & \mathcal{C}}
$$
in $\mathscr{C}$ where $\Omega$ is an isomorphism, and $U,V$ are both right adjoints
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{B}
},\qquad
\xymatrix{
\mathcal{C} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-G & & \ar@/^0.5pc/[ll]^-V \mathcal{D}
}
$$
where $\eta, \epsilon$ denote the units and counits respectively of both adjunctions.
\begin{defn}\label{lift}
In the above situation, we say that $ C$ is an \emph{extension of} $ S$ and $ S$ is a \emph{lift
of $ C$ through the adjunctions} $F \dashv U$ and $G \dashv V$.
\end{defn}
\begin{lem}\label{sq}
The square
$$\xymatrix{
\mathcal{B} \ar[rr]^-S \ar[dd]_-U && \mathcal{D} \ar[dd]^-V \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega^{-1}} "a";"b" \\
\\
\mathcal{A} \ar[rr]_-C && \mathcal{C}
}$$
is a right adjoint in $\mathrm{Sq}^v(\mathscr{C})$.
\end{lem}
\begin{proof}
The left adjoint is constructed as
$$\xymatrix{
\mathcal{A} \ar[r]^-C \ar[d]_-F & \mathcal{C} \ar[d]^-G \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Lambda} "a";"b"\\
\mathcal{B} \ar[r]_-S & \mathcal{D}
}$$
where $\Lambda$ is the uniquely determined \emph{mate}~\cite{MR0357542} of $\Omega$, that is, the composite
$$
\xymatrix{
GC \ar[r]^-{GC \eta} & GCUF \ar[rr]^-{G \Omega F} && GVSF \ar[r]^-{\epsilon SF} & SF
}
$$
The (co)unit for this adunction is given by taking the pair of (co)units for the adjunctions $F\dashv U$ and $G \dashv V$.
\end{proof}
The following theorem,
which closely follows \cite[Lemmata~6.1.1
and~6.1.4]{MR2094071}, constructs a canonical pair of
distributive laws from $\Lambda$.
\begin{thm}\label{arisem}
The 2-cells
$$
\xymatrix{\theta \colon
VGC \ar[rr]^-{V \Lambda}
&& VSF \ar[rr]^-{\Omega^{-1} F} &&
CUF}
$$
and
$$
\xymatrix{\chi \colon
GVS\ar[rr]^-{G \Omega^{-1}} & &GCU
\ar[rr]^-{\Lambda U} && SFU}
$$
define a morphism of monads $(C, \theta) \colon (\mathcal{A}, UF) \to (\mathcal{C}, VG)$ and an opmorphism of comonads
$(S, \chi) \colon (\mathcal{B}, FU) \to (\mathcal{D}, GV)$ respectively.
\end{thm}
\begin{proof}
The adjunctions in Lemma~\ref{sq} generate a monad and a comonad
$$
\xymatrix{
\mathcal{A} \ar[r]^-C \ar[d]_-{UF} & \mathcal{C} \ar[d]^-{VG} \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\theta} "a";"b"\\
\mathcal{A} \ar[r]_-C & \mathcal{C}
}\qquad
\xymatrix{
\mathcal{B} \ar[r]^-S \ar[d]_-{FU} & \mathcal{D} \ar[d]^-{GV} \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\chi} "a";"b"\\
\mathcal{B} \ar[r]_-S & \mathcal{D}
}
$$
respectively in $\mathrm{Sq}^v(\mathscr{C})$. By Proposition~\ref{monadsq} this is equivalent to the theorem. \end{proof}
In fact, $\theta, \chi$ satisfy a universal property:
\begin{prop}\label{univprop}
In the setting of Theorem~\ref{arisem},
\begin{enumerate}
\item The 2-cell $\theta \colon VGC \Rightarrow CUF$ is unique such that the diagram
\begin{equation}\label{univmonad}
\begin{array}{c}
\xymatrix{
VGCU \ar[r]^-{\theta U } \ar[d]_-{VG\Omega} & CUFU \ar[r]^-{CU \epsilon} & CU \ar[d]^-{\Omega} \\
VGVS \ar[rr]_-{V\epsilon S} & & VS
}
\end{array}
\end{equation}
commutes.
\item The 2-cell $\chi \colon GVS \Rightarrow SFU$ is unique such that the diagram
\begin{equation}\label{univcomonad}
\xymatrix{
CU \ar[rr]^-{C\eta U} \ar[d]_-{\Omega} & & CUFU\ar[d]^-{\Omega FU} \\
VS \ar[r]_-{V \eta S} & VGVS \ar[r]_-{V\chi} & VSFU
}
\end{equation}
commutes.
\end{enumerate}
\end{prop}
\begin{proof} We prove part 1, and remark that the proof for part 2 is similar. Consider the diagram
$$
\xymatrix@=3.5em{
VGCU \ar@{=}[d] \ar[r]^-{VGC\eta U} & VGCUFU \ar[d]_-{VGCU\epsilon} \ar[r]^-{VG\Omega FU} & VGUSFU \ar[d]_-{VGUS\epsilon}\ar[r]^-{V\epsilon S FU} & VSFU \ar[d]_-{VS\epsilon}\ar[r]^-{\Omega^{-1} FU} & CUFU \ar[d]^-{CU\epsilon} \\
VGCU\ar@{=}[r] & VGCU \ar[r]_-{VG\Omega} & VGVS \ar[r]_-{V\epsilon S} & VS \ar[r]_-{\Omega^{-1}} & CU
}
$$
The first square commutes by a triangle identities for the adjunction $G \dashv V$ and the other squares commute by compatibility of horizontal and vertical composition in $\mathscr{C}$. Thus, the outer diagram commutes, which means that diagram~\ref{univmonad} commutes also.
Let $\theta \colon VGC \Rightarrow CUF$ be a 2-cell satisfying the hypothesis. Consider the diagram
$$
\xymatrix@=3.5em{
VGC \ar[r]^-{\theta'} \ar[d]_-{VGC\eta} & CUF \ar@{=}[r] \ar[d]^-{CUF\eta} & CUF \ar@{=}[d] \\
VGCUF \ar[r]^-{\theta' UF} \ar[d]_-{VG\Omega F}& CUFUF \ar[r]^-{CU\epsilon F} & CUF \ar[d]^-{\Omega F} \\
VGVSF \ar[rr]_-{V\epsilon SF} & & VSF
}
$$
The leftmost square commutes by compatibility of composition in $\mathscr{C}$, the rightmost square commutes by a triangle identity for $F \dashv U$ and the lower rectangle commutes by assumption. Thus, the outer rectangle commutes, which means exactly that $\theta' = \theta$.
\end{proof}
\begin{rem}\label{univrem}
The commutativity of diagrams~\ref{univmonad} and~\ref{univcomonad} is equivalent to the statement that
$$
\xymatrix{
(\mathcal{B}, 1) \ar[rr]^-{(U, U\epsilon)} \ar[dd]_-{(S,1)} && (\mathcal{A}, UF)\ar[dd]^-{(C, \theta)} \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b"\\
\\
(\mathcal{D}, 1) \ar[rr]_-{(V, V\epsilon)} && (\mathcal{C}, VG)
}
\qquad
\xymatrix{
(\mathcal{B}, FU) \ar[rr]^-{(U, \eta U)} \ar[dd]_-{(S,\chi)} && (\mathcal{A}, 1)\ar[dd]^-{(C, 1)} \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b"\\
\\
(\mathcal{D}, GV) \ar[rr]_-{(V, \eta V)} && (\mathcal{C}, 1)
}
$$
are 2-cells in $\operatorname{Mnd}(\mathscr{C})$ and $\operatorname{Cmd}(\mathscr{C}^*)^*$ respectively.
\end{rem}
We now state an important corollary of Theorem~\ref{arisem}. Since we use this extensively later, we explicitly state the necessary terminology.
\begin{cor}\label{arisec}
Suppose that we have a square
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
in a 2-category $\mathscr{C}$ where $\Omega$ is an isomorphism, and $U$ is a right adjoint
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{B}.
}
$$
Let $\eta$ be the unit, and let $\epsilon$ be the counit of this adjunction, and let $\Lambda$ denote the 2-cell
$$
\xymatrix{
FC \ar[r]^-{FC \eta} & FCUF \ar[rr]^-{F \Omega F} && FUSF \ar[r]^-{\epsilon SF} & SF.
}
$$
Then:
\begin{enumerate}
\item The 2-cells
$$
\xymatrix{\theta \colon
UFC \ar[rr]^-{U \Lambda}
&& USF \ar[rr]^-{\Omega^{-1} F} &&
CUF}
$$
and
$$
\xymatrix{\chi \colon
FUS\ar[rr]^-{F \Omega^{-1}} & &FCU
\ar[rr]^-{\Lambda U} && SFU}
$$
define a morphism of monads $(C, \theta) \colon (\mathcal{A}, UF) \to (\mathcal{A}, UF)$ and an opmorphism of comonads
$(S, \chi) \colon (\mathcal{B}, FU) \to (\mathcal{B}, FU)$ respectively.
\item
If $C$ and $S$ are themselves comonads, and
$(U, \Omega) \colon (\mathcal{B}, S) \to (\mathcal{A}, C)$ is an opmorphism of comonads, then $\theta$ is a mixed distributive law and $\chi$ is a comonad distributive law.
\end{enumerate}
\end{cor}
\begin{proof}
Part 1 follows immediately from Theorem~\ref{arisem}, in the special case where the two adjunctions are equal.
If $C,S$ are comonads and $(U, \Omega)$ is an opmorphism of comonads, then there is an adjunction $(F,\Lambda) \dashv (U, \Omega^{-1})$ in $\operatorname{Cmd}(\mathscr{C})$. This adjunction generates a monad and comonad, yielding the distributive laws $\theta,\chi$.
\end{proof}
\begin{defn}\label{arisedeffo}
A comonad distributive law $ \chi $, or a mixed distributive law $\theta$, as in
Corollary~\ref{arisec} is said to \emph{arise from the
adjunction $F \dashv U$}.
\end{defn}
\subsection{The extremal case}\label{extremalcase}
In this section, we examine what can be said about lifts when we start with a morphism of monads, instead of the other way around.
Suppose that $\mathscr{C}$ admits Eilenberg-Moore constructions for monads, with 2-adjunction $I\dashv J$.
One extremal situation in which specifying a morphism of monads uniquely determines a lift of a 1-cell is the following: consider an arbitrary morphism of monads
$$
\xymatrix{
(\mathcal{A}, B) \ar[rr]^-{(C, \theta)} && (\mathcal{C}, E).
}
$$
By the remarks in Section~\ref{EM2cat}, there are adjunctions
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-{F^B} & & \ar@/^0.5pc/[ll]^-{U^B} \mathcal{A}^B
},\qquad
\xymatrix{
\mathcal{C} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-{F^E} & & \ar@/^0.5pc/[ll]^-{U^E} \mathcal{C}^E
}
$$
with unit, counit denoted by $\eta, \epsilon$ respectively where
$$
\xymatrix{
(\mathcal{A}^B, 1) \ar[rr]^-{(U^B, U^B \epsilon)} && (\mathcal{A}, B)
}, \qquad
\xymatrix{
(\mathcal{C}^E, 1) \ar[rr]^-{(U^E, U^E \epsilon)} && (\mathcal{C}, E)
}
$$
are morphisms of monads given by evaluating the counit $\xi \colon IJ \Rightarrow 1$ at $(\mathcal{A}, B)$ and $ (\mathcal{C}, E)$ respectively.
\begin{thm}\label{unique1cell}
The 1-cell $C^\theta \colon \mathcal{A}^B \to \mathcal{C}^E$ is unique such that the diagram
\begin{equation}\label{liftdiag}
\begin{array}{c}
\xymatrix{
(\mathcal{A}^B, 1) \ar[d]_-{(C^\theta, 1)} \ar[rr]^-{(U^B, U^B\epsilon)} && (\mathcal{A}, B) \ar[d]^-{(C, \theta)} \\
(\mathcal{C}^E, 1) \ar[rr]_-{(U^E, U^E\epsilon)} && (\mathcal{C}, E)
}
\end{array}
\end{equation}
commutes in $\operatorname{Mnd}(\mathscr{C})$.
\end{thm}
\begin{proof}
By 2-naturality of $\xi$, the diagram
$$
\xymatrix{IJ(\mathcal{A}, B) \ar[d]_-{IJ(\mathcal{C}, \theta)} \ar[r]^-\xi & (\mathcal{A}, B) \ar[d]^-{(C, \theta)} \\
IJ(\mathcal{C}, E) \ar[r]_-{\xi} & (\mathcal{C}, E)
}
$$
commutes, which is identical to diagram~\ref{liftdiag}. If $H \colon \mathcal{A}^B \to \mathcal{C}^E$ is another such 1-cell then the diagram
$$
\xymatrix{IJ(\mathcal{A}, B) \ar[d]_-{IH} \ar[r]^-\xi & (\mathcal{A}, B) \ar[d]^-{(C, \theta)} \\
IJ(\mathcal{C}, E) \ar[r]_-{\xi} & (\mathcal{C}, E)
}
$$
commutes. The morphism obtained from going along the top of the diagram corresponds to $H$ under the 2-adjunction $I\dashv J$, and the morphism along the bottom corresponds to $J(C, \theta) = C^\theta$. Hence $H = C^\theta$.
\end{proof}
Note that diagram~\ref{liftdiag} says exactly that the identity is a monad 2-cell. Therefore, by Proposition~\ref{univprop} and Remark~\ref{univrem} (taking $E = VG$ and $B = UF$), we recover $(C,\theta)$ as the canonical morphism given by Theorem~\ref{arisem}. We write $\tilde\theta$ to denote the 2-cell
$$
\chi \colon \tilde E C^\theta \Rightarrow C^\theta \tilde B
$$
from Theorem~\ref{arisem},
\begin{exa}\label{trivv2}
We have a commutative diagram
$$\xymatrix{
\mathcal{A}^B \ar[r]^-{U^B} \ar[d]_-{\tilde B} & \mathcal{A} \ar[d]^-B \\
\mathcal{A}^B \ar[r]_-{U^B} & \mathcal{A}
}
$$
and thus, by Theorem~\ref{arisem}, a morphism of monads $(B, \theta) \colon (\mathcal{A}, B) \to (\mathcal{A}, B)$ defined by
$$
\xymatrix{
\theta \colon BB = UFUF \ar[rr]^-{UFUF\eta} && UFUFUF \ar[rr]^-{U\epsilon F U F} && UFUF = BB
}
$$
However, $U \epsilon F = \mu$ by diagram~\ref{EMmonad}, so in fact $\theta = \tau$ from Example~\ref{trivialdist}. Since both $\tilde B$, $B^\tau$ fit into diagram~\ref{trivialdist}, Theorem~\ref{unique1cell} tells us that $\tilde B = B^\tau$ as 1-cells.
\end{exa}
Now let us specialise to the situation that $\mathcal{C} = \mathcal{A}$, $E = B$, and $C$ is a comonad on $\mathcal{A}$. Suppose also that $\theta$ is a distributive law of comonads.
\begin{prop}\label{everydistlawarises}
The identity 2-cell
$$
\xymatrix{\mathcal{A}^B \ar[r]^-{U^B} \ar[d]_{C^\theta} & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{1} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{A}^B \ar[r]_{U^B} & \mathcal{A}
}
$$
defines an opmorphism of comonads
$$\xymatrix{ (\mathcal{A}^B, C^\theta) \ar[rr]^-{(U^B, 1)} && (\mathcal{A}, C).}$$
\end{prop}
\begin{proof}
There is a commutative diagram
$$
\xymatrix{
(\mathcal{A}^B, 1) \ar[rr]^-{(U^B, U^B\epsilon)} \ar[drr]_-{(U^B, 1)} && (\mathcal{A}, B) \ar[d]^-{(1, \eta)} \\
&& (\mathcal{A}, 1)
}
$$
in $\operatorname{Mnd}(\mathscr{C})$. The top 1-cell is the counit of $I \dashv J$, so applying $J$ maps it to the identity. Therefore $1^\eta = U^B \colon \mathcal{A}^B \to \mathcal{A}$. The result now follows by applying the 2-functor $\cMix(\mathscr{C}) \to \operatorname{Cmd}(\mathscr{C})$ of Proposition~\ref{cmdEM} to the 1-cell
$$
\xymatrix{
(\mathcal{A}, \theta, B, C) \ar[rr]^-{(1, \eta, 1)} & & (\mathcal{A}, 1, 1, C)
}
$$
in $\cMix(\mathscr{C})$.
\end{proof}
Now we are in the situation of Corollary~\ref{arisec} (taking $B = U^BF^B = UF$), and from that we recover the fact that $\theta, \chi$ are distributive laws.
\subsection{The comparison 1-cell}\label{comparisonsection}
Let $\mathscr{C}$ be a 2-category which admits Eilenberg-Moore constructions for monads, with 2-adjunction $I\dashv J$, and let
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{B}
}
$$
be an adjunction therein, which generates a monad $B = UF$ on $\mathcal{A}$ as well as a comonad $T = FU$ on $\mathcal{B}$. Let $\epsilon$ denote the counit $T \to 1$. There is a morphism of monads
$$
\xymatrix{
(\mathcal{B}, 1) \ar[rr]^-{(U, U\epsilon)} && (\mathcal{A}, B)
}
$$
which is mapped by the right adjoint $J \colon \operatorname{Mnd}(\mathscr{C}) \to \mathscr{C}$
to a 1-cell
$$
\xymatrix{
\mathcal{B}^1 = \mathcal{B} \ar[r]^-{U^{U\epsilon}} & \mathcal{A}^B.
}
$$
Equivalently, we can view this as the 1-cell corresponding to the morphism of monads $(U, U\epsilon)$ under the 2-adjunction.
\begin{defn}
We call $U^{U\epsilon}$ the \emph{comparison 1-cell} associated to the adjunction $F\dashv U$.
\end{defn}
\begin{lem}\label{compcomonad}
We have that
$$
\xymatrix{
(\mathcal{B}, T) \ar[rr]^-{(U^{U\epsilon}, 1)} && (\mathcal{A}^B, \tilde B)
}
$$
is an opmorphism of comonads.
\end{lem}
\begin{proof}
There are two commutative squares
$$
\xymatrix{
(\mathcal{B}, 1) \ar[rr]^-{(U, 1)} \ar[d]_-{(U, U\epsilon)} && (\mathcal{A},1) \ar@{=}[d] \\
(\mathcal{A}, B) \ar[rr]_-{(1, \eta)} && (\mathcal{A}, 1)
}\qquad
\xymatrix{
(\mathcal{B}, 1) \ar[rr]^-{(U, U\epsilon)} \ar[d]_-{(T, 1)}&& (\mathcal{A}, B) \ar[d]^-{(B, \tau)} \\
(\mathcal{B}, 1) \ar[rr]_-{(U, U\epsilon)} && (\mathcal{A}, B)
}
$$
in $\operatorname{Mnd}(\mathscr{C})$, mapped to the two diagrams
$$
\xymatrix{
\mathcal{B} \ar[rr]^-U \ar[d]_-{U^{U\epsilon}} && \mathcal{A} \ar@{=}[d] \\
\mathcal{A}^B \ar[rr]_-{U^B} && \mathcal{A}
}\qquad
\xymatrix{
\mathcal{B} \ar[rr]^-{U^{U\epsilon}} \ar[d]_-{T} && \mathcal{A}^B \ar[d]^-{\tilde B} \\
\mathcal{B} \ar[rr]_-{U^{U\epsilon}} && \mathcal{A}^B
}
$$
in $\mathscr{C}$. Applying Theorem~\ref{arisem} to the right-hand square yields an opmorphism of comonads $(U^{U\epsilon}, \chi) \colon (\mathcal{B}, T) \to (\mathcal{A}^B, \tilde B)$. Using the left-hand square, we see that
$$
\xymatrix{
U \ar[rrr]^-{\eta U} \ar@{=}[d] & & & UT \ar@{=}[d] \\
U^B U^{U\epsilon} \ar[rr]_-{\eta U^B U^{U\epsilon}} && \tilde B U^B U^{U\epsilon} \ar@{=}[r] & UT
}
$$
commutes, so by part 2 of Proposition~\ref{univprop}, $\chi = 1$, proving the Lemma.
\end{proof}
Now, consider a square
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
where $(U, \Omega) \colon (\mathcal{B}, S) \to (\mathcal{A}, C)$ is an iso-opmorphism of comonads, so that we are in the context of Definition~\ref{lift}. By Remark~\ref{univrem},
$$
\xymatrix{
(\mathcal{B}, 1) \ar[rr]^-{(U, U\epsilon)} \ar[dd]_-{(S,1)} && (\mathcal{A}, B)\ar[dd]^-{(C, \theta)} \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b"\\
\\
(\mathcal{B}, 1) \ar[rr]_-{(U, U\epsilon)} && (\mathcal{A}, B)
}
$$
is a monad 2-cell, giving rise to a 2-cell
$$
\xymatrix{
\mathcal{B} \ar[r]^-{U^{U\epsilon}} \ar[d]_-{S} & \mathcal{A}^B \ar[d]^-{C^\theta} \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\tilde\Omega} "a";"b" \\
\mathcal{B} \ar[r]_-{U^{U\epsilon}} & \mathcal{A}^B
}
$$
and thus giving an opmorphism of comonads
$$
\xymatrix{
(\mathcal{B}, S) \ar[rr]^-{(U^{U\epsilon}, \tilde\Omega)} && (\mathcal{A}^B, C^\theta)
}
$$
or equivalently, a morphism of comonads
$$
\xymatrix{
(\mathcal{B}, S) \ar[rr]^-{(U^{U\epsilon}, \tilde\Omega^{-1})} && (\mathcal{A}^B, C^\theta).
}
$$
\begin{rem}
The morphism of comonads $(U^{U\epsilon}, \tilde\Omega^{-1})$ is the comparison 1-cell associated to the adjunction
$$
\xymatrix{
(\mathcal{A}, C) \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-{(F, \Lambda)} & & \ar@/^0.5pc/[ll]^-{(U, \Omega^{-1})} (\mathcal{B}, S)
}
$$
in $\operatorname{Cmd}(\mathscr{C})$, where $\Lambda$ denotes the mate of $\Omega$.
\end{rem}
\begin{thm}\label{distcomp1cell}
Let $\theta, \chi$ be the distributive laws given by Corollary~\ref{arisec}. Then
$$
\xymatrix{
(\mathcal{B}, \chi, T, S) \ar[rrr]^-{(U^{U\epsilon}, 1, \tilde\Omega^{-1})} & & & (\mathcal{A}^B, \tilde\theta, \tilde B, C^\theta)
}
$$
is a 1-cell in $\cDist(\mathscr{C})$.
\end{thm}
\begin{proof}
By the above remarks, $(U^{U\epsilon}, \tilde\Omega^{-1})$ is a morphism of comonads, and $(U^{U\epsilon}, 1)$ is an opmorphism of comonads by Lemma~\ref{compcomonad}. Therefore, we need only check that the Yang-Baxter equation holds to prove the Theorem. By definition of $\theta, \chi$ there is a commutative diagram
$$
\xymatrix{
BUS \ar@{=}[d] \ar[r]^-{B \Omega^{-1}} & BCU \ar[r]^-{\theta U} & CBU \ar@{=}[d] \\
UTS \ar[r]_-{U \chi} & UST \ar[r]_-{\Omega^{-1} T} & CUT
}
$$ in $\mathscr{C}$. Therefore, in $\operatorname{Mnd}(\mathscr{C})$, the compositions of all the monad 2-cells in the diagrams
$$
\xymatrix{
(\mathcal{B}, 1) \ar[dd]_-{(T, 1)} \ar[rr]^-{(S, 1)} & & \ar[dd]^-{(T, 1)} (\mathcal{B}, 1) \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\chi} "a";"b"\ar[rr]^-{(U, U\epsilon)} & & (\mathcal{A}, B) \ar[dd]^-{(B, \tau)} \ar@{}[ddll]^(.25){}="c"^(.75){}="d" \ar@{=>}_-{1} "c";"d" \\
& & & &\\
(\mathcal{B}, 1) \ar[rr]^-{(S, 1)} \ar@/_1.5pc/[ddrr]_-{(U U\epsilon)} & & (\mathcal{B}, 1) \ar@{}[dd]^(.25){}="e"^(.75){}="f" \ar@{=>}_-{\Omega^{-1}} "e";"f"\ar[rr]^-{(U, U\epsilon)}& & (\mathcal{A}, B) \\
& & & &\\
& & (\mathcal{A}, B) \ar@/_1.5pc/[uurr]_-{(C, \theta)} & &
}
$$
$$\xymatrix{
(\mathcal{B}, 1) \ar@{}[ddrr]^(.25){}="c"^(.75){}="d" \ar@{=>}^-{\Omega^{-1}} "c";"d"\ar[rr]^-{(U, U\epsilon)} & & (\mathcal{A}, B) \ar@{}[rrdd]^(.25){}="a"^(.75){}="b" \ar@{=>}^-{\theta} "a";"b" \ar[rr]^-{(B, \tau)} & & (\mathcal{A}, B) \\
& & & &\\
(\mathcal{B}, 1) \ar[uu]^-{(S, 1)}\ar[rr]^-{(U, U\epsilon)} \ar@/_1.5pc/[ddrr]_-{(T, 1)} & & (\mathcal{A}, B) \ar[uu]_-{(C, \theta)}\ar@{}[dd]^(.25){}="e"^(.75){}="f" \ar@{=>}_-{1} "e";"f"\ar[rr]^-{(B, \tau)}& & (\mathcal{A}, B) \ar[uu]_-{(C, \theta)} \\
& & & &\\
& & (\mathcal{B}, 1) \ar@/_1.5pc/[uurr]_-{(U, U\epsilon)} & &
}
$$
are the same. Applying $J$ to each composition gives the diagram
$$
\xymatrix{
\tilde B U^{U\epsilon} S \ar@{=}[d] \ar[rr]^-{\tilde B \tilde \Omega^{-1}} & &\tilde B C^\theta U^{U\epsilon} \ar[rr]^-{\tilde \theta U^{U\epsilon}} & & C^\theta \tilde B U^{U\epsilon} \ar@{=}[d] \\
U^{U\epsilon} TS \ar[rr]^-{U^{U\epsilon} \chi} & & U^{U\epsilon} ST \ar[rr]^-{\tilde\Omega^{-1} T} & &C^\theta U^{U\epsilon} T
}
$$
which is the Yang-Baxter equation.
\end{proof}
\subsection{From mixed to comonad distributive laws}
Finally, we explain how one can functorially assign a comonad distributive law to a mixed one. Let $\mathscr{C}$ be a 2-category which admits Eilenberg-Moore constructions for monads. We begin by making a small observation. Let
$$
\xymatrix{ (\mathcal{A}, B) \ar[rr]^-{(\Sigma, \sigma)}&&(\mathcal{D}, A)}
$$
be a morphism of monads. Then
$$
\xymatrix{
(\mathcal{A}, B) \ar[rr]^-{(\Sigma, \sigma)} \ar[dd]_-{(B, \tau)} && (\mathcal{D},A) \ar[dd]^-{(A, \tau)} \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\sigma} "a";"b" \\
\\
(\mathcal{A}, B) \ar[rr]_-{(\Sigma, \sigma)} && (\mathcal{D}, A)
}
$$
is a monad 2-cell, which is mapped by $J \colon \operatorname{Mnd}(\mathscr{C}) \to \mathscr{C}$ to a 2-cell
$$
\xymatrix{
\mathcal{A}^B \ar[r]^-{\Sigma^\sigma} \ar[d]_-{\tilde B} & \mathcal{D}^A \ar[d]^-{\tilde A} \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\tilde\sigma} "a";"b" \\
\mathcal{A}^B \ar[r]_-{\Sigma^\sigma}& \mathcal{D}^A
}
$$
Using this notation, we can now state the following Proposition. We omit the proof, which can be found in~\cite[p.~160]{MR0299653}.
\begin{prop}\label{StreetJ}
There is a 2-functor $\overline J \colon \operatorname{Mnd}(\mathscr{C}) \to \operatorname{Cmd}(\mathscr{C}^*)^*$ defined by
$$
\xymatrix{
(\mathcal{A}, B ) \ddtwocell<9>_{(\Sigma, \sigma)\ \ \ \ \ }^{\ \ \ \ \ \ \ (\Sigma', \sigma')}{^\alpha} & & & & (\mathcal{A}^B, \tilde B) \ddtwocell<9>_{(\Sigma^{\sigma}, \tilde\sigma)\ \ \ \ \ \ \ \ }^{\ \ \ \ \ \ \ \ \ \ (\Sigma'^{\sigma'}, \tilde\sigma')}{^\tilde\alpha} \\
& & \longmapsto \\
(\mathcal{D}, A) & & & & (\mathcal{D}^A, \tilde A)
}
$$
\end{prop}
By applying Proposition~\ref{StreetJ} to the 2-category $\operatorname{Cmd}(\mathscr{C})$, we obtain the following:
\begin{cor}\label{2funcmixdist}
The assignment
$$
\xymatrix{
(\mathcal{A}, \theta, B, C ) \ddtwocell<10>_{(\Sigma, \sigma, \gamma)\ \ \ \ \ \ \ }^{\ \ \ \ \ \ \ \ \ \ (\Sigma', \sigma', \gamma')}{^\alpha} & & & & (\mathcal{A}^B, \tilde\theta, \tilde B, C^\theta) \ddtwocell<10>_{(\Sigma^{\sigma}, \tilde\sigma, \tilde\gamma)\ \ \ \ \ \ \ \ \ }^{\ \ \ \ \ \ \ \ \ \ \ \ (\Sigma'^{\sigma'}, \tilde\sigma', \tilde\gamma')}{^\tilde\alpha} \\
& & \longmapsto \\
(\mathcal{D}, \psi, A, D) & & & & (\mathcal{D}^A, \tilde\psi, \tilde A, \tilde D)
}
$$
defines a 2-functor $\cMix(\mathscr{C}) \to \cDist(\mathscr{C})$.
\end{cor}
Thus the image under this 2-functor of a general
1-cell $\theta \to \psi$ of mixed distributive laws
can be composed with the 1-cell given by the
comparison functor of Proposition~\ref{distcomp1cell} to give a new 1-cell $\chi \to
\tilde\psi$.
\section{Eilenberg-Moore constructions in \texorpdfstring{$\mathbf{Cat}$}{Cat}}\label{emem}
We begin by explicitly describing the general 2-categorical constructions involving monads from the preceding sections, and discussing the dual versions of several of these.
\subsection{Eilenberg-Moore categories}\label{emcatsect}
Let $B$ be a monad on a category $\mathcal{A}$.
\begin{defn}\label{algebra}
The \emph{Eilenberg-Moore category} of the monad $B$, denoted $\mathcal{A}^B$, is the category whose objects are pairs $(X, \beta)$, called $B$-\emph{algebras}, where $X$ is an object of $\mathcal{A}$ and $\beta \colon BX \to X$ is a morphism satisfying associativity and unitality axioms, that is, the diagrams
$$
\xymatrix{
BBX \ar[r]^-{{\mu_X}} \ar[d]_-{B \beta} & BX \ar[d]^-\beta \\
BX \ar[r]_-{\beta} & X
}
\qquad
\xymatrix{
X \ar@{=}[dr]\ar[r]^-{\eta_X} & BX \ar[d]^-{\beta} \\
& X
}
$$
commute. The morphisms $f \colon (X, \beta) \to (X', \beta')$, called \emph{$B$-algebra morphisms}, are morphisms $f \colon X \to Y$ in $\mathcal{A}$ which are compatible with the algebra structures, that is, the diagram
$$
\xymatrix{
BX \ar[r]^-{Bf} \ar[d]_-\beta & BX' \ar[d]^-{\beta '} \\
X \ar[r]_-{f} & X'
}
$$
commutes.
\end{defn}
Let $(\Sigma, \sigma) \colon (\mathcal{A}, B) \to (\mathcal{D}, A)$ be a morphism of monads. The functor $\Sigma \colon \mathcal{A} \to \mathcal{D}$ lifts to a functor $\Sigma^\sigma \colon \mathcal{A}^B \to \mathcal{D}^A$, defined on objects by
$$
\Sigma^\sigma (X, \beta) = (\Sigma X, \xymatrix{A\Sigma X \ar[r]^-{\sigma_X} & \Sigma B X \ar[r]^-{\Sigma \beta} & \Sigma X})
$$
and defined on morphisms by $\Sigma^\sigma f = \Sigma f$.
Also, any monad 2-cell $\alpha \colon (\Sigma, \sigma) \Rightarrow (\Sigma', \sigma')$ lifts in an obvious way to give a natural transformation $\tilde\alpha \colon \Sigma^\sigma \Rightarrow \Sigma'^{\sigma '}$. It is straightforward to check that these assignments define a 2-functor $J \colon \operatorname{Mnd}(\mathbf{Cat}) \to \mathbf{Cat}$, exhibiting $\mathbf{Cat}$ as a 2-category which admits Eilenberg-Moore constructions for monads (cf.\ Section~\ref{EM2cat}).
The canonical adjunction
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-{F^B} & & \ar@/^0.5pc/[ll]^-{U^B} \mathcal{A}^B
}
$$
is defined as follows. $U^B$ is the obvious forgetful functor, and $F^B$ sends an object $X$ to the \emph{free $B$-algebra} $(BX, \mu_X)$, while acting as $B$ on morphisms. Thus the comonad $\tilde B$ generated by this adjunction sends a $B$-algebra $(X, \beta)$ to the free $B$-algebra on $X$.
In fact, $\mathbf{Cat}_*$ also admits Eilenberg-Moore constructions for monads; but of course, a monad in $\mathbf{Cat}_*$ is nothing more than a comonad in $\mathbf{Cat}$. The construction is obtained by dualising the monad case above. In particular:
\begin{defn}
Let $T$ be a comonad on a category $\mathcal{B}$. A \emph{$T$-coalgebra} is a pair $(M, \nabla)$ where $M$ is an object of $\mathcal{B}$, and $\nabla \colon M \Rightarrow TM$ is a natural transformation which satisfies coassociativity and counitality axioms, that is the diagrams
$$
\xymatrix{
X \ar[r]^-\nabla \ar[d]_-\nabla & TX \ar[d]^-{\delta_X } \\
TX \ar[r]_-\nabla & TTX
}\qquad
\xymatrix{
X \ar@{=}[dr] \ar[r]^-{\nabla} & \ar[d]^-{\epsilon_X} TX \\
& X
}
$$commute.
\end{defn}
These, along with the obvious notion of coalgebra morphism, constitute the \emph{Eilenberg-Moore} category of the comonad $T$, denoted $\mathcal{B}^T$.
\subsection{Algebra structures on functors}\label{algebrafunc}
For a fixed category $\mathcal{A}$, there are 2-functors
\begin{align*}
[\mathcal{A}, -] \colon &\mathbf{Cat} \to \mathbf{Cat} \\
[-, \mathcal{A}] \colon &\mathbf{Cat}^* \to \mathbf{Cat}
\end{align*}
where $[\mathcal{Y}, \mathcal{A}] := \mathbf{Cat}(\mathcal{Y}, \mathcal{A})$ is the category of functors $\mathcal{Y} \to \mathcal{A}$ (cf.\ Example~\ref{hom2functors}). In particular, these both map monads to monads and comonads to comonads.
Let $B$ be a monad on $\mathcal{A}$, and let $X \colon \mathcal{Y} \to \mathcal{A}$ be a functor.
\begin{defn}\label{algebrafuncdef}
We say that a natural transformation $\beta \colon BX \Rightarrow X$ is a \emph{$B$-algebra structure on $X$} if $(X, \beta)$ is a $[\mathcal{Y}, \mathcal{A}]$-algebra, i.e.\ an object of the Eilenberg-Moore category
$[\mathcal{Y}, \mathcal{A}]^{[\mathcal{Y}, B]} $. We also say that $(X, \mathcal{Y}, \beta)$ is a \emph{$B$-algebra}.
\end{defn}
So, this definition is basically the same as Definition~\ref{algebra} except the object $X$ becomes a functor, and the structure morphism $\beta$ becomes a natural transformation. Both definitions are actually equivalent: we recover~\ref{algebra} from~\ref{algebrafuncdef} by choosing $\mathcal{Y}$ to be the terminal category $\mathbbm{1}$. Thinking of algebras as functors is advantageous, however, since it allows us to dualise in the right way:
\begin{defn}
Let $Y \colon \mathcal{A} \to \mathcal{Z}$ be a functor. We say that a natural transformation $\omega \colon YB \Rightarrow Y$ is a \emph{$B$-opalgebra structure on $Y$} if $(Y, \omega)$ is a $[B, \mathcal{Z}]$-algebra, i.e.\ an object of the Eilenberg-Moore category $[\mathcal{A}, \mathcal{Z}]^{[B, \mathcal{Z}]}$. We also say that $(Y, \mathcal{Z}, \omega)$ is a \emph{$B$-opalgebra}. Explicitly, there are commutative diagrams
$$
\xymatrix{
YBB\ar[r]^-{Y \mu} \ar[d]_-{\omega B} & YB \ar[d]^-\omega \\
YB \ar[r]_-\omega & B
}
\qquad
\xymatrix{
Y \ar[r]^-{Y \eta} \ar@{=}[dr] & YB \ar[d]^-\omega \\
& Y
}
$$
\end{defn}
Of course, we can dualise in a different way to obtain the notion of \emph{coalgebra} and \emph{opcoalgebra} structures on functors.
\subsection{Kleisli categories}
The 2-category $\mathbf{Cat}^*$ also admits Eilenberg-Moore constructions for monads. We now describe the 0-cell part of the 2-functor $\operatorname{Mnd}(\mathbf{Cat}^*) \to \mathbf{Cat}^*$. Let $B$ be a monad on a category $\mathcal{A}$.
\begin{defn}
The \emph{Kleisli category} of the monad $B$, denoted $\mathcal{A}_B$, is the category whose objects are precisely those of $\mathcal{A}$, and whose morphisms $X \to Y$ are morphisms $X \to BY$ in $\mathcal{A}$. The composite of
$$
\xymatrix{
X \ar[r]^-f & Y
},
\qquad
\xymatrix{
Y\ar[r]^-g & Z
}
$$
in $\mathcal{A}_B$ is given by the composite
$$
\xymatrix{
X \ar[r]^-f & TY \ar[r]^-{Tg} & TTZ \ar[r]^-{\mu_Z} & TZ
}
$$
in $\mathcal{A}$. The identity morphism $X \to X$ in $\mathcal{A}_B$ is given by $\eta_X \colon X \to BX$ in $\mathcal{A}$.
\end{defn}
We also have that $\mathbf{Cat}_*^*$ admits Eilenberg-Moore constructions for monads, giving rise to the dual notion of the Kleisli category $\mathcal{B}_T$ for a comonad $T$ on a category $\mathcal{B}$, defined in a similar way.
\subsection{The comparison functor}
Let
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{B}
}
$$
be an adjunction between categories $\mathcal{A}, \mathcal{B}$. From Section~\ref{emcatsect}, the comparison functor $U^{U\epsilon} \colon \mathcal{B} \to \mathcal{A}^B$ is defined on objects by
$$
U^{U\epsilon} M = (UM, \xymatrix{BUM = UFUM \ar[r]^-{U\epsilon_M} & UM}) .
$$
\chapter{Distributive laws and their coalgebras}\label{DISTRIBUTIVE}
Now we concentrate on the theory of distributive laws in the 2-category $\mathscr{C} = \mathbf{Cat}$.
Inspired by~\cite{MR2787298, MR3020336}, in Section~\ref{galoismapsect} we focus on a special case of Corollary~\ref{arisec}, and how different instances thereof are related by a \emph{Galois map} (Definition~\ref{galoisdeff}). In Section~\ref{coalgsect} we study $\chi$-coalgebras~\cite{MR2411421, MR2415479} for a distributive law $\chi$ of comonads, and view these as coefficient modules of the cyclic homology theories which appear later. We conclude this chapter in Section~\ref{examplesect} by studying some simple examples in the category of sets.
Hereafter we are concerned with distributive laws only in $\mathbf{Cat}$, and so we write $\mathbf{Dist}$ in place of $\cDist(\mathbf{Cat})$, and $\mathbf{Mix}$ in place of $\cMix(\mathbf{Cat})$.
The work in this chapter is original. Sections~\ref{galoismapsect} (inspired by~\cite{MR2787298, MR3020336}) and~\ref{coalgsect} are based on~\cite{2} and~\cite[\S2--3]{1}.
\section{The Galois map}\label{galoismapsect}
Corollary~\ref{arisec} yields comonad distributive laws
from lifts through an adjunction, and different lifts
produce different distributive laws. Here we describe
how these are related in terms of suitable
generalisations of the Galois map from the theory of
Hopf algebras (see Section~\ref{whygalois} below for the
example motivating the terminology).
Suppose that
$$
\xymatrix{\mathcal{A} \ar@/^{0.5pc}/[rr]^-F
\ar@{}[rr]|-{\perp}&& \ar@/^0.5pc/[ll]^-U \mathcal{B} }
$$
is an adjunction between categories $\mathcal{A}, \mathcal{B}$ which generates a monad $B = UF$ on $\mathcal{A}$
and a comonad $T = FU$ on $\mathcal{B}$.
\subsection{Main application of the lifting theorem}\label{mainapp}
Suppose furthermore that $S$ is a lift of an endofunctor $C$ through a single adjunction as in Definition~\ref{lift}, i.e.\ we have a square
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
where $\Omega$ is a natural isomorphism. By Corollary \ref{arisec}, we have two natural transformations
$$
\xymatrix@=2.9em{
\theta \colon BC = UFC \ar[r]^-{UFC\eta} & UFCUF \ar[r]^-{UF\Omega F} & UFUSF \ar[r]^-{U\epsilon SF} & USF \ar[r]^-{\Omega^{-1} F} & CUF = CB}
$$
and
$$
\xymatrix@=2.9em{\chi \colon
TS = FUS \ar[r]^-{F \Omega^{-1}} & FCU \ar[r]^-{FC\eta U} & FCUFU \ar[r]^-{F\Omega FU} & FUSFU \ar[r]^-{\epsilon SFU} & SFU = ST
}
$$
such that $(C, \theta)$ is an morphism of monads, and $(S, \chi)$ is an opmorphism of comonads.
\begin{exa}\label{trivv}
A trivial example which nevertheless plays a r\^ole below is the case where $C = B$, $S = T$, and $\Omega = 1$ (cf.~Example~\ref{trivv2}). In this case, $\theta$ and $\chi$ are given by
\begin{align*}
&\xymatrix{BB = UFUF\ar[r]^-{U\epsilon F} & UF \ar[r]^-{UF\eta} & UFUF = BB} \\
&\xymatrix{TT = FUFU \ar[r]^-{\epsilon FU} & FU \ar[r]^-{F \eta U} & FUFU = TT}
\end{align*}
respectively.
\end{exa}
Functors do not
necessarily lift nor extend through an
adjunction (for example, the functor on $\cSet$
which assigns the empty set to each set does not lift
to the category of vector spaces over a given field), and if they do, they may not do so
uniquely. Theorem~\ref{arisem} says only that once a
lift or extension is chosen, there
is a unique compatible pair of natural transformations $
\theta $ and $\chi$.
Recall from Corollary~\ref{arisec} that if $(U, \Omega)$ is an opmorphism of comonads, then $\theta$ and $\chi$ are distributive laws, in which case we say that they arise from the adjunction $F \dashv U$ cf.~Definition~\ref{arisedeffo}.
\begin{prop}
Every mixed distributive law, and every comonad distributive law, arises from an adjunction.
\end{prop}
\begin{proof}
The statement for mixed distributive laws follows from Proposition~\ref{everydistlawarises} and the following comments. The statement for comonad distributive laws is proved dually: take $\mathcal{A}$ to be the Kleisli category $\mathcal{B}_T$ in which case a comonad distributive law $\chi \colon TS \Rightarrow ST$ yields an extension $C$ of a functor $S$.
\end{proof}
Recall that there is a 2-functor $\mathbf{Mix} \to \mathbf{Dist}$ as in Corollary~\ref{2funcmixdist}. It is those distributive laws in the image of this 2-functor that are the main ones of interest in later chapters.
Analogously, we obtain a $2$-functor $\mathbf{Dist} \rightarrow
\mathbf{Mix}$ by taking extensions to Kleisli
categories.
\subsection{Generalising the Galois map}\label{galoismapsct}
Suppose that in addition to $S$, we have another lift $W \colon \mathcal{B} \to \mathcal{B}$ of the endofunctor $C$ through the adjunction $F\dashv U$. This means that we have two squares
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}} \qquad
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ W & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Phi} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
where $\Omega \colon CU \Rightarrow US$ and $\Phi \colon CU \Rightarrow UW$ are natural isomorphisms. For any object $X$ in $\mathcal{A}$ and any object $Y$ in $\mathcal{B}$, consider the diagram
$$
\xymatrix{
\mathcal{B}(FX, SY) \ar[d]_-\cong \ar@{.>}[rr] & & \mathcal{B}(FX, WY) \\
\mathcal{A}(X, USY) \ar[rr] & & \mathcal{A}(X, UWY) \ar[u]_-\cong
}
$$
where the vertical maps are induced by the adjunction $F \dashv U$, and the lower map is induced by the composition
$$
\xymatrix{
USY \ar[r]^-{\Omega_Y^{-1}} & CUY \ar[r]^-{\Phi_Y} & UWY.
}
$$
The dashed arrow defines one component of a natural
isomorphism
$$
\Gamma^{S, W} \colon \mathcal{B}(F-, S-) \Rightarrow
\mathcal{B}(F-, W-)
$$
of functors $\mathcal{A}^* \times \mathcal{B} \rightarrow \cSet$.
\begin{defn}\label{galoisdeff}
We call the natural isomorphism $\Gamma^{S,W}$ the \emph{Galois map}.
\end{defn}
The following properties
are easy consequences of the definition:
\begin{prop}\label{sunshines}
Let $S$ and $W$ be two lifts
of an endofunctor $C$ through an adjunction $F \dashv
U$. Then:
\begin{enumerate}
\item The inverse of
$\Gamma^{S,W}$ is given by $\Gamma^{W,S}$.
\item The natural transformation $\Gamma^{S,W}$ maps a morphism $f
\colon F X \rightarrow S Y$ to
$$
\xymatrix{ F X
\ar[r]^-{F \eta_X} & FUF X \ar[r]^-{FU f} &
FUS Y \ar[rr]^{F(\Phi_Y \circ \Omega^{-1}_Y)} &
& FUW Y \ar[r]^-{\epsilon_{W Y}} & W Y. }
$$
\item If $\chi^{S}$ and $ \chi^{W}$ denote the natural transformations
determined by the two lifts, then the composite
$$
\xymatrix@C=3em{
\mathcal{B}( FUS, SFU ) \ar[r]^-{\Gamma^{S,W}}_-\cong & \mathcal{B} (FUS, WFU) \ar[r]_-\cong & \mathcal{B}( FUW, WFU)
}
$$
maps $\chi^S$ to $\chi^W$, where the right-hand isomorphism is induced by the composite $\Omega \circ \Phi^{-1} \colon UW \Rightarrow US$.
\end{enumerate}
\end{prop}
So, in the applications of Corollary~\ref{arisec}, all
distributive laws obtained from different lifts
$S,W$ of a
given comonad through an adjunction are obtained from
each other by application of $\Gamma^{S,W}$.
In particular, consider the situation of Example~\ref{trivv}, where we have a lift of $B$ itself through an adjunction, that is, a commutative square:
$$
\xymatrix{
\mathcal{B} \ar[r]^-{U} \ar[d]_-{T} & \mathcal{A} \ar[d]^-B \\
\mathcal{B} \ar[r]_-{U} & \mathcal{A}
}
$$
Let
$W$ be any other lift of $B$ through the
adjunction. By taking $X$ to be $U Y$ for an object
$Y$ of $\mathcal{B}$, one obtains a natural transformation
$ \Gamma^{T,W}
\colon \mathcal{B} (T-,T-) \rightarrow \mathcal{B}(T-,W-) $ that we can
evaluate on $ 1 \colon T Y \rightarrow T Y$, which
produces a natural transformation
$\Gamma^{\mathrm T,\mathrm V}(1) \colon T \Rightarrow W$.
Adapting \cite[Definition~1.3]{MR2651345}, we define:
\begin{defn} We say that $F$ is \emph{$W$-Galois}
if $$ \xymatrix{ \Gamma^{T,W}(1) \colon T=FU
\ar[r]^-{F \eta U} & FUFU = FUT
\ar[r]^-{F \Phi} & FUW \ar[r]^-{\epsilon W}
& W } $$ is an isomorphism.
\end{defn}
The following proposition provides the connection to
Hopf algebra theory:
\begin{prop}\label{wisga}
If $F$ is $W$-Galois and
$ \theta \colon BB \Rightarrow BB$ is the natural transformation arising from the lift $W$ of $B$, then the
natural transformation
$$
\xymatrix{ \beta \colon BB
\ar[r]^-{B \eta B} & BBB \ar[r]^-{\theta B}
& BBB \ar[r]^-{B \mu} & BB }
$$
is an
isomorphism.
\end{prop}
\begin{proof}
If $F$ is
$W$-Galois, then $ U\Gamma^{T,W}(1)F$ is an
isomorphism
$$
\xymatrix@C=2.455em{ UTF=UFUF
\ar[rr]^-{UF \eta UF} & & UFUFUF =
UFUT F \ar[r]^-{UF \Phi F} &
UFUW F \ar[r]^-{U\epsilon WF} &
UWF. }
$$
Let now $ \chi \colon TW \Rightarrow
WT$ be the natural transformation corresponding to $ \theta $
as in Theorem~\ref{arisem} and Corollary~\ref{arisec}. Inserting $\varepsilon
W=(W \varepsilon) \circ \chi$ and $U \chi \circ
UF \Phi=\Phi FU \circ \theta U$ and
$B=UF$, the isomorphism becomes $$ \xymatrix@C=2.5em{
UTF=BB \ar[r]^-{B \eta B} & BBB
\ar[r]^-{\theta B} & BBB=BUFUF
\ar[r]^-{\Phi FUF} & UWFUF
\ar[r]^-{UW\epsilon F} & UWF } $$ Finally,
we have by construction $U \varepsilon F=\mu $, and
using the naturality of $\Phi$ this gives
$UW\varepsilon F \circ \Phi FUF= \Phi F
\circ BU\varepsilon F$. Composing the
above isomorphism with $\Phi^{-1} F$ gives $ \beta $.
\end{proof}
It is this associated map $ \beta $ that is used to
distinguish Hopf algebras amongst bialgebras, see
Section~\ref{whygalois} below.
\section{Coalgebras for distributive laws}\label{coalgsect}
We now discuss $ \chi
$-coalgebras, which serve as
coefficient modules in the constructions of Chapter~\ref{CYCLIC}. Let $\chi \colon TS \Rightarrow ST$ be a distributive law of comonads on a category $\mathcal{B}$.
\subsection{Coalgebras for distributive laws}\label{chicoalgs}
\begin{defn}\label{coalgdef}
A \emph{$\chi$-coalgebra} $(M,\rho)$ consists of an object $M$ in $\mathcal{B}$, together with a morphism $\rho \colon TM \to SM$
in $\mathcal{B}$ such that the following two diagrams commute:
$$
\xymatrix{ TM \ar[r]^-{\Delta M} \ar[d]_-{\rho} &
TTM \ar[r]^{T\rho} &TSM \ar[d]^-{\chi M} \\
SM \ar[r]_-{\Delta M} & SSM & STM
\ar[l]^-{S\rho} } \quad\quad\quad
\xymatrix{ & TM
\ar[dl]_-{\epsilon M} \ar[d]^-{\rho}\\ M & SM
\ar[l]^-{\epsilon M} } $$
\end{defn}
\begin{rem}\label{extrarem}
The distributive law $\chi$ is a comonad in the 2-category $\operatorname{Cmd}(\mathbf{Cat})$, but a $\chi$-coalgebra is \emph{not} a coalgebra for this comonad. Rather, the composite $TS$ becomes a comonad with comultiplication and counit given by
$$
\xymatrix{
TS \ar[r]^-{\delta\delta} & TTSS \ar[rr]^-{T\chi S} && TSTS,
}
\qquad
\xymatrix{
TS \ar[r]^-{\epsilon\epsilon} & 1
}
$$
respectively~\cite{MR0241502}. There is a morphism of comonads
$$
\xymatrix{
(\mathcal{B}, TS) \ar[rr]^-{(1,T\epsilon)} & & (\mathcal{B}, T)
}
$$
which induces a forgetful functor $\mathcal{B}^{TS} \to \mathcal{B}^T$. The $\chi$-coalgebra structures $\rho$ on an object $M$ are equivalent to those $TS$-coalgebra structures on $TM$ whose image under this forgetful functor is cofree, see e.g.~\cite[Prop.~1.9]{MR2415479}.\end{rem}
For a fixed category $\mathcal{Y}$, the 2-functor $\operatorname{Cmd} \colon \operatorname{2-\mathbf{Cat}} \to \operatorname{2-\mathbf{Cat}}$ maps the 2-functor $[\mathcal{Y}, -] \colon \mathbf{Cat} \to \mathbf{Cat}$ to the 2-functor $\operatorname{Cmd}([\mathcal{Y}, -]) \colon \operatorname{Cmd}(\mathbf{Cat}) \to \operatorname{Cmd}(\mathbf{Cat})$. This sends comonads to comonads, i.e.\ it sends distributive laws to distributive laws. Therefore,
$$
[\mathcal{Y}, \chi] \colon [\mathcal{Y}, T] \circ [\mathcal{Y}, S] \Rightarrow [\mathcal{Y}, S] \circ [\mathcal{Y}, T]
$$
is a distributive law of comonads. Similarly, for a fixed category $\mathcal{Z}$,
$$
[\chi, \mathcal{Z}] \colon [T, \mathcal{Z}] \circ [S, \mathcal{Z}] \Rightarrow [S, \mathcal{Z}] \circ [T, \mathcal{Z}]
$$
is a distributive law of comonads. With this in mind:
\begin{defn}
A \emph{$\chi$-coalgebra structure} on a functor $M \colon \mathcal{Y} \to \mathcal{B}$ is a natural transformation $\rho \colon TM \Rightarrow SM$ such that $(M, \rho)$ is a $[\mathcal{Y}, \chi]$-coalgebra in $[\mathcal{Y}, \mathcal{B}]$. \end{defn}
So, in a similar way to Section~\ref{algebrafunc}, it does not really matter whether we talk about $\chi$-coalgebras as objects or as functors since we may choose $\mathcal{Y} = \mathbbm{1}$ to recover Definition~\ref{coalgdef}.
Dually, we have:
\begin{defn}
A \emph{$\chi$-opcoalgebra structure} on a functor $N \colon \mathcal{B} \to \mathcal{Z}$ is a natural transformation $\lambda \colon NS \Rightarrow NT$ such that $(N, \lambda)$ is a $[\chi, \mathcal{Z}]$-coalgebra in $[\mathcal{B}, \mathcal{Z}]$. Explicitly, the two diagrams
$$
\xymatrix{
NS \ar[r]^-{N\delta} \ar[d]_-\lambda & NSS \ar[r]^-{\lambda S} & NTS \ar[d]^-{N \chi} \\
NT \ar[r]_-{N \delta} & NTT & NST \ar[l]^-{\lambda T}
}
\qquad
\xymatrix{
& NS \ar[d]^-{\lambda} \ar[dl]_-{N \epsilon} \\ N & NT \ar[l]^-{N\epsilon}
}
$$
commute.
\end{defn}
\begin{rem}
Dualising Remark~\ref{extrarem} appropriately, the $\chi$-opcoalgebra structures $\lambda$ on $N$ correspond to $TS$-opcoalgebra structures on $NS$ whose underlying $S$-opcoalgebra is opcofree.
\end{rem}
Now suppose we are in the comonad setting of Section~\ref{mainapp}, so we have an adjunction $F\dashv U$ and a square
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
where $(U,\Omega)$ is an iso-opmorphism of comonads. Via Corollary~\ref{arisec}, the square gives rise to a distributive law $\chi \colon TS \Rightarrow ST$. The following characterises $ \chi $-coalgebras
in this case.
\begin{prop}\label{chicoalgprop}
Let $M \colon \mathcal{Y} \rightarrow \mathcal{B}$
be a functor.
\begin{enumerate}
\item
Coalgebra structures for $\chi$ on $M$ correspond to
$C$-coalgebra structures $\nabla$ on the
functor $UM \colon \mathcal{Y} \rightarrow \mathcal{A}$.
\item Let
$S$ and $W$ be two lifts of the functor $C$ through the
adjunction, and let $\chi^S$ and $\chi^W$ denote the
comonad distributive laws determined by the lifts
$S$ and $W$ respectively. Then the composite
$\Gamma^{S,W}$ maps $\chi^S$-coalgebra
structures $\rho^S$ on $M$ bijectively to
$\chi^W$-coalgebra structures $\rho^W$ on $M$.
\end{enumerate}
\end{prop}
\begin{proof} For part (1),
$\chi$-coalgebra structures $\rho \colon FUM
\Rightarrow SM$ are mapped under the adjunction to $\nabla
\colon UM \Rightarrow USM \cong CUM$. Part (2)
follows immediately since $\Gamma^{S,W}$ is
the composition of the adjunction isomorphisms and
$\Phi \circ \Omega^{-1}$.
\end{proof}
Dually, given an adjunction
$V \dashv G$ for the comonad $S$ and an extension
$Q $ of the comonad $T$ through the
adjunction,
$ \chi $-opcoalgebra structures on
$N \colon \mathcal{B} \rightarrow \mathcal{Z}$ correspond in complete analogy
to $Q$-opcoalgebra structures on $NV$.
\subsection{Twisting by 1-cells}\label{twistcoeff}
Here we show how factorisations of distributive laws as
considered in \cite{2} can be used to obtain new
$ \chi $-coalgebras from old ones.
Generalising the notion of a module over a monoidal category, we make the following definition:
\begin{defn}\label{leftCmodule}
Let $\mathscr{C}$ be a 2-category. A \emph{left $\mathscr{C}$-module} $\mathscr{M}$ consists of a category $\mathscr{M}(\mathcal{A})$ for each 0-cell $\mathcal{A}$ in $\mathscr{C}$ and functors $\rhd_{\mathcal{A}, \mathcal{B}} \colon \mathscr{C}(\mathcal{A}, \mathcal{B}) \times \mathscr{M}(\mathcal{A}) \to \mathscr{M}(\mathcal{B})$, called \emph{left actions}, such that the diagrams
$$
\xymatrix{
\mathscr{C}(\mathcal{B}, \mathcal{C}) \times \mathscr{C}(\mathcal{A}, \mathcal{B}) \times \mathscr{M}(\mathcal{A}) \ar[rr]^-{1 \times \rhd_{\mathcal{A}, \mathcal{B}}} \ar[d]_-{\circ \times 1} && \mathscr{C}(\mathcal{B}, \mathcal{C}) \times \mathscr{M}(\mathcal{B}) \ar[d]^-{\rhd_{\mathcal{B}, \mathcal{C}}} \\
\mathscr{C}(\mathcal{A}, \mathcal{C}) \times \mathscr{M}(\mathcal{A}) \ar[rr]_-{\rhd_{\mathcal{A}, \mathcal{C}}} && \mathscr{M}(\mathcal{C})
}
$$
and
$$
\xymatrix{
\mathscr{M}(\mathcal{A}) \ar[rr]^-{u \times 1} \ar@{=}[drr] && \mathscr{C}(\mathcal{A}, \mathcal{A}) \times \mathscr{M}(\mathcal{A}) \ar[d]^-{\rhd_{\mathcal{A}, \mathcal{A}}} \\
& &\mathscr{M}(\mathcal{A})
}
$$
commute in $\mathbf{Cat}$, for all 0-cells $\mathcal{A}, \mathcal{B}, \mathcal{C}$ in $\mathscr{C}$.
\end{defn}
Hereafter, we omit the subscripts from the action functors.
\begin{rem}
The functors
$$\rhd \colon \mathscr{C}(\mathcal{A}, \mathcal{B}) \times \mathscr{M}(\mathcal{A}) \to \mathscr{M}(\mathcal{B})$$
correspond under the closed symmetric monoidal structure of $\mathbf{Cat}$ to functors
$$
\mathscr{C}(\mathcal{A}, \mathcal{B}) \to [\mathscr{M}(\mathcal{A}), \mathscr{M}(\mathcal{B})]
$$
and thus, a left $\mathscr{C}$-module is nothing more than a 2-functor $\mathscr{M} \colon \mathscr{C} \to \mathbf{Cat}$.
\end{rem}
Dually, one defines a \emph{right $\mathscr{C}$-module} $\mathscr{N}$, with right actions $$\lhd \colon \mathscr{N}(\mathcal{B}) \times \mathscr{C}(\mathcal{A}, \mathcal{B}) \to \mathscr{N}(\mathcal{A})$$
which can alternatively be viewed as 2-functor $\mathscr{N} \colon \mathscr{C}^* \to \mathbf{Cat}$.
For each distributive law $\chi \colon TS \Rightarrow ST$ of comonads in a category $\mathcal{B}$, we define a category $\mathscr{R}(\chi)$
as follows. The objects are $\chi$-coalgebras $(M, \mathcal{Y}, \rho)$, i.e.\ $\chi$-coalgebra structures on functors $\mathcal{Y} \to \mathcal{B}$ where $\mathcal{Y}$ is allowed to vary. The morphisms
$$\xymatrix{(M, \mathcal{Y}, \rho) \ar[rr]^-{(F, \phi)} && (M', \mathcal{Y}', \rho')}$$
consist of a functor $F \colon \mathcal{Y} \to \mathcal{Y}'$ and a natural transformation $\phi \colon M \Rightarrow M'F$ such that the diagram
$$
\xymatrix{
TM \ar[r]^-{T\phi} \ar[d]_-\rho & TM'F \ar[d]^-{\rho' F} \\
SM \ar[r]_-{S \phi} & SM'F
}
$$
commutes.
Recall that $\mathbf{Dist}$ denotes the 2-category $\operatorname{Cmd}(\operatorname{Cmd}(\mathbf{Cat})^*)^*$ of comonad distributive laws in $\mathbf{Cat}$, cf.~Section~\ref{finallyover}
\begin{thm}\label{twist}
The categories $\mathscr{R}(\chi)$ define a left $\mathbf{Dist}$-module.
\end{thm}
\begin{proof}
We define the action $\rhd \colon \mathbf{Dist}(\chi, \tau) \times \mathscr{R}(\chi) \to \mathscr{R}(\tau)$ as follows. For a 1-cell
$$\xymatrix{
(\mathcal{B}, \chi, T, S) \ar[rr]^-{(\Sigma, \sigma, \gamma)} && (\mathcal{D}, \tau, G, C)
}
$$
in $\mathbf{Dist}$, we define
$$
(\Sigma, \sigma, \gamma) \rhd (M, \mathcal{Y}, \rho) = (\Sigma M, \mathcal{Y}, \gamma M \circ \Sigma\rho \circ \sigma M)
$$
and on morphisms we define $\alpha \rhd ({\varphi},F)$ to be the pair $(\alpha {\varphi}, F)$. First we check that this assignment is a well-defined functor. Consider the diagram
$$
\xymatrix@C=3.5em{
G \Sigma M \ar[d]_-{\sigma M} \ar[r]^{\delta \Sigma M} & GG\Sigma M \ar[r]^-{G \sigma M} & G\Sigma TM \ar[d]^-{\sigma{TM}} \ar[r]^-{G\Sigma\rho} & G \Sigma SM \ar[d]^-{\sigma SM} \ar[r]^-{G \gamma M} & GC\Sigma M \ar[d]^-{\tau \Sigma M} \\
\Sigma TM \ar[dd]_-{\Sigma \rho} \ar[rr]_-{\Sigma \delta M} & & \Sigma TTM \ar[r]_-{\Sigma T\rho} & \Sigma TSM \ar[d]^-{\Sigma \chi M} & CG\Sigma M \ar[d]^-{C \sigma M } \\
& & & \Sigma STM \ar[r]^-{\gamma{TM}} \ar[d]_-{\Sigma S \rho}& C\Sigma TM \ar[d]^-{C \Sigma \rho} \\
\Sigma SM \ar[d]_-{\gamma{M}}\ar[rrr]_-{\Sigma \delta M}& & & \Sigma SSM \ar[r]_-{\gamma {SM}}& C \Sigma SM \ar[d]^-{C \gamma M} \\
C\Sigma M \ar[rrrr]_-{\delta {\Sigma M}}& & & & CC \Sigma M
}
$$
The top-left and bottom rectangles commute because $\sigma, \gamma$ are compatible with comultiplication, the middle-left rectangle commutes because $M$ is a $\chi$-coalgebra, the top-right diagram commutes by the Yang-Baxter condition, and the remaining squares commute by naturality of $\sigma,\gamma$. Therefore the outer rectangle commutes.
Consider the triangle
$$
\xymatrix@=3em{
G\Sigma M \ar[rrd]_-{\epsilon {\Sigma M}} \ar[r]^-{\sigma M} & \Sigma TM \ar[dr]^-{\Sigma \epsilon M} \ar[rr]^-{\Sigma \rho} & & \Sigma SM \ar[dl]_-{\Sigma \epsilon M} \ar[r]^-{\gamma M} & C \Sigma M \ar[dll]^-{\epsilon {\Sigma M}}\\
& & \Sigma M & &
}
$$
The middle triangle commutes because $M$ is a $\chi$-coalgebra, and the other two inner triangles commute by the compatibility of $\sigma, \gamma$ with the counit. Therefore the outer triangle commutes. This shows that $\rhd$ is well-defined on objects.
Let $({\varphi},F) \colon (M,\mathcal{Y}, \rho) \to (M',\mathcal{Y}', \rho')$ be a morphism of $\chi$-coalgebras, and let $\alpha$ be a 2-cell $ (\Sigma, \sigma, \gamma) \Rightarrow (\Sigma', \sigma', \gamma')$ of distributive laws. Consider the diagram
$$
\xymatrix@=3em{
G \Sigma M \ar[d]_-{\sigma M} \ar[r]^-{G \alpha M} & G \Sigma' M \ar[d]^-{\sigma' M} \ar[r]^-{G \Sigma' {\varphi}} & G \Sigma' M'F \ar[d]^-{\sigma' {M' F}} \\
\Sigma TM \ar[d]_-{\Sigma \rho} \ar[r]^-{\alpha {TM}} & \Sigma' TM \ar[d]^-{\Sigma' \rho} \ar[r]^-{\Sigma' T {\varphi}} & \Sigma' TM'F \ar[d]^-{\Sigma' \rho' F} \\
\Sigma SM \ar[r]_-{\alpha {SM}} & \Sigma' SM \ar[r]_-{\Sigma' S{\varphi}} & \Sigma' S M'F
}
$$
The top-left square commutes since $\alpha$ is a 2-cell, the top-right square commutes by naturality of $\sigma$, the bottom-left square commutes by naturality of $\alpha$, and the bottom-right square commutes since $({\varphi},F)$ is a $\chi$-coalgebra morphism. Thus the outer triangle commutes, which shows that $\alpha \rhd ({\varphi},F)$ is a $\chi$-coalgebra morphism.
It is clear that $\rhd$ respects identities and composition of morphisms (because the vertical and horizontal compositions of natural transformations are compatible with each other),
so $\rhd$ is well-defined on morphisms. It is also routine to check that $\rhd$ satisfies the required axioms of Definition~\ref{leftCmodule}, thus proving the Theorem.
\end{proof}
The axioms of Definition~\ref{leftCmodule} tell us immediately:
\begin{cor}
For a fixed distributive law $\chi$, the category $\mathscr{R}(\chi)$ of $\chi$-coalgebras is a strict left module category for the strict monoidal category $\mathbf{Dist}(\chi, \chi)$.
\end{cor}
This is the main result of~\cite{2}.
Dually, one can construct the category $\mathscr{L}(\chi)$ of $\chi$-opcoalgebras, and show that $\mathscr{L}$ is a right $\mathbf{Dist}$-module, and that $\mathscr{L}(\chi)$ is a strict right module category for $\mathbf{Dist}(\chi, \chi)$.
In summary, we have shown that for a 1-cell of distributive laws
$$
\xymatrix{(\mathcal{B}, \chi, T, S) \ar[rr]^-{(\Sigma, \sigma, \gamma)} & & (\mathcal{D}, \tau, G, C)}
$$
we can canonically twist any $\chi$-coalgebra and any $\tau$-opcoalgebra
$$
(\xymatrix{\mathcal{Y} \ar[r]^-{M} & \mathcal{B}
}, \xymatrix{ TM \ar@{=>}[r]^-{\rho} & SM }),\qquad
(\xymatrix{\mathcal{D} \ar[r]^-{N} & \mathcal{Z}
}, \xymatrix{ NC \ar@{=>}[r]^-{\lambda} & NG })
$$
by $(\Sigma, \sigma, \gamma)$ to yield a $\tau$-coalgebra
$$
(
\xymatrix{\mathcal{Y} \ar[r]^-{M} & \mathcal{B} \ar[r]^-{\Sigma} & \mathcal{D}},
\xymatrix{
G\Sigma M \ar[r]^-{\sigma M} & \Sigma TM \ar[r]^-{\Sigma \rho} & \Sigma SM \ar[r]^-{\gamma M} & C\Sigma M
}
)
$$
and a $\chi$-opcoalgebra
$$
(
\xymatrix{
\mathcal{B} \ar[r]^-{\Sigma} & \mathcal{D} \ar[r]^-N & \mathcal{Z}
},
\xymatrix{
N\Sigma S \ar[r]^-{N\gamma} & NC\Sigma \ar[r]^-{\lambda \Sigma} & NG\Sigma \ar[r]^-{N\sigma} & N\Sigma T
}
).
$$
This will be applied in Section~\ref{twistsec} below in
the context of duplicial functors.
\subsection{From coalgebras for comonads to those for
distributive laws}
In the remainder of this section, we discuss a class of
coefficient objects that
lead
to contractible simplicial objects; see
Proposition~\ref{trivcontract} below. In the Hopf
algebroid setting, these are the Hopf (or entwined)
modules as studied in \cite{MR3020336,MR1604340}.
Note first that
$T$-coalgebras can be equivalently viewed as
$1$-cells \emph{from} the trivial distributive law, and dually, $T$-opcoalgebras correspond to $1$-cells \emph{to} the trivial distributive law:
\begin{prop}\label{triv1cell}
Let $\chi \colon TS\Rightarrow ST$ be a comonad distributive law. Then:
\begin{enumerate}
\item Any $S$-coalgebra $(M, \mathcal{Y}, \nabla^S)$ defines a 1-cell
$$
\xymatrix{
(\mathcal{Y}, 1, 1, 1) \ar[rr]^-{(M, \epsilon M, \nabla^S)} && (\mathcal{B}, \chi, T, S)
}
$$
in $\mathbf{Dist}$, and all 1-cells $1 \to \chi$ are of this form.
\item Any $T$-opcoalgebra $(N, \mathcal{Z}, \nabla^T)$
defines a $1$-cell
$$
\xymatrix{
(\mathcal{B}, \chi, T, S) \ar[rr]^-{(N, \nabla^T, N\epsilon)} && (\mathcal{Z}, 1, 1, 1).
}
$$
in $\mathbf{Dist}$, and all $1$-cells $\chi \to 1$ are of this form.
\end{enumerate}
\end{prop}
Furthermore, these $1$-cells can also be viewed as $ \chi
$-coalgebras:
\begin{prop}\label{triv}
Let $\chi \colon TS \Rightarrow ST$ be a comonad
distributive law. Then:
\begin{enumerate}
\item Any $S$-coalgebra $(M,
\mathcal{Y}, \nabla^S)$ defines a
$\chi$-coalgebra $(M, \mathcal{Y}, \epsilon
\nabla^S)$. \item Any $T$-opcoalgebra $(N,
\mathcal{Z}, \nabla^T)$ defines a
$\chi$-opcoalgebra $(N, \mathcal{Z}, \nabla^T
\epsilon^S)$.
\end{enumerate}
\end{prop}
Note, however, that there is no known way to associate
a $1$-cell in $\mathbf{Dist}$ to an arbitrary $\chi$-(op)coalgebra.
\subsection{Entwined algebras}
Finally, we describe how
$ \chi $-coalgebras as in Proposition~\ref{triv}
are in some sense lifts of
entwined (also called mixed) algebras; throughout,
$\theta \colon BC \Rightarrow CB$ is a mixed
distributive law between a monad $B$ and a
comonad $C$ on a category $\mathcal{A}$.
\begin{defn}
Let $M$ be an object in $\mathcal{A}$ equipped with both a
$B$-algebra structure $\beta \colon B M \rightarrow M$ and
a $C$-coalgebra structure $\nabla \colon M \rightarrow
CM$. We say that the $(M, \beta,
\nabla)$ is an \emph{entwined algebra with respect to
$\theta$}, or a $\theta$-\emph{entwined algebra}, if the diagram
\begin{equation}\label{entwinedcondition}
\begin{array}{c}
\xymatrix{ B M
\ar[d]_-{B \nabla} \ar[r]^-{\beta} & M
\ar[r]^-{\nabla} & CM\\ BC \mathrm M \ar[rr]_{\theta
M} & &
CB M \ar[u]_-{C \beta} }
\end{array}
\end{equation}
commutes.
\end{defn}
As before, we have a version for functors:
\begin{defn}
A functor $M \colon \mathcal{Y} \to \mathcal{A}$ has the structure of a \emph{$\theta$-entwined algebra} if $M$ is a $[\mathcal{Y}, \theta]$-entwined algebra.
\end{defn}
Once again by choosing $\mathcal{Y} = \mathbbm{1}$ we see that these two definitions are equivalent. Without loss of generality,
we consider entwined
algebras as objects in $\mathcal A$.
With the obvious notion of morphism (given by natural
transformations compatible with $\nabla$ and $ \beta
$), entwined algebras form a category; this is
evidently isomorphic to the category
$(\mathcal{A}^B)^{C^\theta}$ of
$C^\theta$-coalgebras in $\mathcal{A}^B$.
Dually we define an entwined opalgebra structure
on a
functor $N \colon \mathcal{A} \rightarrow \mathcal{Z}$
for a distributive law $CB \Rightarrow BC$.
Suppose again that we are in the comonad setting of Section~\ref{mainapp}, so we have an adjunction $F\dashv U$ and a square
$$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
where $(U,\Omega)$ is an iso-opmorphism of comonads, giving rise to distributive laws
$$\chi \colon TS \Rightarrow ST, \qquad \theta \colon BC \Rightarrow CB$$ via Corollary~\ref{arisec}.
The following proposition explains the relation
between $\theta$-entwined algebras and
$\chi$-coalgebras:
\begin{prop}
Let $M \colon \mathcal{Y} \rightarrow \mathcal{B}$
be a functor and let $\nabla \colon M \Rightarrow
SM$ be a natural transformation.
\begin{enumerate}
\item If $\nabla$ is an
$S$-coalgebra structure,
then the structure morphisms
$$
\xymatrix{ BUM = UFUM
\ar[r]^-{U \epsilon M} & UM}, \qquad \xymatrix{
UM \ar[r]^-{U \nabla} & USM
\ar[r]^-{\Omega^{-1}} &CUM}
$$
turn $UM$ into an entwined
algebra with respect to $\theta$.
\item If $\mathcal{B} =
\mathcal{A}^B$, then the converse of (1) holds.
\end{enumerate}
\end{prop}
\begin{proof}
By the remarks in Section~\ref{comparisonsection}, the comparison functor $U^{U\epsilon} \colon \mathcal{B} \to \mathcal{A}^B$ is part of a morphism of comonads
$$
\xymatrix{
(\mathcal{B}, S) \ar[rr]^-{(U^{U\epsilon}, \tilde\Omega^{-1})} &&(\mathcal{A}^B, C^\theta)
}
$$ thus inducing a lifting
$$
\xymatrix{
\mathcal{B}^S \ar[d] \ar[rr]^-{(U^{U\epsilon})^{\tilde \Omega^{-1}}} && (\mathcal{A}^B )^{C^\theta} \ar[d]\\
\mathcal{B} \ar[rr]_-{U^{U\epsilon}} && \mathcal{A}^B
}
$$
where the undecorated arrows denote the appropriate forgetful functors.
The object map of the functor in the top row of this
diagram is the construction in part (1). If $\mathcal{B} =
\mathcal{A}^B$, then $U^{U \epsilon} = 1$, $\Omega =
1$ and $S = C^\theta$, so the top functor in the diagram is just the identity, implying part (2).
\end{proof}
Dually, entwined opalgebra structures on a $B$-opalgebra $(N, \mathcal{Z} , \omega)$ are related to
$\chi$-opcoalgebras if
the codomain $\mathcal{Z}$ of $N$ is a
category with coequalisers. First, we define a functor
$N_B \colon \mathcal{A}^B \rightarrow \mathcal{Z}$
that takes a $B$-algebra morphism $f
\colon (X, \alpha) \rightarrow (Y, \beta)$ to
$N_B(f)$ defined using coequalisers:
$$
\xymatrix@C=5em{NB X \ar[d]_-{N B f}
\ar@<+.5ex>[r]^-{\omega_X} \ar@<-.5ex>[r]_-{N \alpha} &
\ar[d]_-{N f} N X \ar@{->>}[r]^-{q_{(X,\alpha)}} &
N_B (X, \alpha) \ar@{.>}[d]^-{N_B(f)} \\ NB
Y \ar@<+.5ex>[r]^-{\omega_Y}
\ar@<-.5ex>[r]_-{N \beta} & N Y \ar@{->>}[r]_-{q_{(Y,
\beta)}} & NB(Y, \beta) }
$$
Thus $ N_B$ generalises the functor
${-}\otimes_B N$ defined by a left module $N$
over a ring $B$ on the category of right $B$-modules.
Suppose that
$\theta$ is invertible, and that $N$ admits the
structure of an entwined $\theta^{-1}$-opalgebra, with
coalgebra structure $\nabla \colon N \Rightarrow
CN$. There
are two commutative diagrams:
$$
\xymatrix{ NB
X\ar[d]_-{\nabla_{ B X}} \ar[rr]^-{\omega_X} &&
N X
\ar[dd]^-{\nabla_X} \\ N C B X
\ar[d]_-{ N \theta^{-1}_X} & & \\ N B C X
\ar[rr]_-{\omega_{ C X}} & & N C X
}\quad\quad\quad
\xymatrix{ N B X
\ar[d]_-{\nabla_{ B X}}
\ar[rr]^-{ N\alpha} & &
N X \ar[dd]^-{\nabla_X} \\
N C B X
\ar[d]_-{ N \theta^{-1}_X} & & \\
N B C X
\ar[r]_-{ N \theta_X} &
N C B X
\ar[r]_-{ N C \alpha} &
N C X }
$$
Hence, using coequalisers,
$\nabla$ extends to a natural transformation $\tilde
\nabla \colon N_B \Rightarrow N_B
C^\theta$, and in fact it gives $ N_B$ the structure
of a $C^\theta$-opcoalgebra. Since
$\tilde{\theta}^{-1} \colon C^\theta
\tilde{B} \Rightarrow \tilde{B}C^\theta$
is a comonad distributive law on
$\mathcal{A}^B$, Proposition~\ref{triv} gives examples of
homologically trivial
$ \tilde \theta^{-1}$-opcoalgebras
of the form
$(N_B,
\mathcal{Z}, \tilde{\nabla} \epsilon)$.
\section{Examples in \texorpdfstring{$\mathbf{Set}$}{Set}}\label{examplesect}
Our main examples of distributive laws and their coalgebras are homological in flavour, and come later in Chapter~\ref{EXAMPLES}. Here, however, we give some simple concrete examples in $\mathbf{Set}$ to illustrate the concepts of the current chapter.
\subsection{Monads and comonads}
\begin{exa}\label{finordset}
Let $C$ be a set. This has a unique coalgebra structure with respect to $\mathbf{Set}$ with the monoidal product given by the Cartesian product $\times$. Thus, the functor $C \times{-}$ becomes a comonad on $\mathbf{Set}$ with counit and comultiplication given by
\begin{align*}
&\xymatrix@R=1em{C \times X \ar[r]^-\epsilon & X \\ (c,x) \ar@{|->}[r] & x} & &\xymatrix@R=1em{C \times X \ar[r]^-{\delta} & C \times C \times X \\ (c, x) \ar@{|->}[r] & (c, c, x)}
\end{align*}
\end{exa}
\begin{exa}
Assigning to a set $X$ its powerset $P(X)$ forms a monad $P$ on $\mathbf{Set}$, where the unit and multiplication are given by
\begin{align*}
&\xymatrix@R=1em{X \ar[r]^-\eta & P(X) \\ x \ar@{|->}[r] & \{x\}} & &\xymatrix@R=1em{P(P(X)) \ar[r]^-{\mu} & P(X) \\
A \ar@{|->}[r] & \bigcup A
}
\end{align*}
\end{exa}
\begin{exa}\label{listmonad}
The functor $L \colon \mathbf{Set} \to \mathbf{Set}$, defined on objects by the disjoint union
$$
L(X) := \bigsqcup_{n \ge 0} X^n
$$
is part of a monad. We view $L(X)$ as the set of lists (including the empty list) of elements in $X$, and we denote these lists with square brackets as is common in computer science, for example $[x_1, \ldots, x_n]$. The monad structure is given by
\begin{align*}
&\xymatrix@R=1em{X \ar[r]^-\eta & L(X) \\ x \ar@{|->}[r] & [x]} & &\xymatrix@R=1em{L(L(X)) \ar[r]^-{\mu} & L(X) \\ \ar@{|->}[r] [[x_{1,1}, \ldots, x_{1, n_1}], \ldots] &[x_{1,1}, \ldots, x_{1, n_1}, \ldots] }
\end{align*}
Here $\mu$ is concatenation, i.e.\ removal of the inner brackets.
\end{exa}
\begin{exa}\label{nonemptylistmonad}
The non-empty list functor $L^+ \colon \mathbf{Set} \to \mathbf{Set}$, given by
$$
L^+(X) := \bigsqcup_{n > 0} X^n
$$
becomes a monad with the same unit and multiplication of the list monad of Example~\ref{listmonad}. However, it also a comonad, with structure given by
\begin{align*}
&\xymatrix@R=1em{L^+(X) \ar[r]^-\epsilon & X \\ [x_1, \ldots, x_n] \ar@{|->}[r] & x_1} & &\xymatrix@R=1em{L^+(X) \ar[r]^-{\delta} & L^+(L^+(X)) \\ [x_1, \ldots, x_n] \ar@{|->}[r] & [ [x_1, \ldots, x_n], [x_2, \ldots, x_n], \ldots, [x_n]]}
\end{align*}
Explicitly, $\delta$ takes a list and removes the first element iteratively until the list is empty, while at each stage storing the result in the output list. This procedure is known as giving the \emph{tails} of the list.
\end{exa}
\begin{exa}\label{monoidcomonad}
Let $M$ be a monoid (with identity element $e$). The functor $S \colon \mathbf{Set} \to \mathbf{Set}$ given by $S(X) = X^M$ on objects (this is the set of functions from $M$ to $X$) becomes a comonad with structure given by
\begin{align*}
&\xymatrix@R=1em{X^M \ar[r]^-\epsilon & X \\ f \ar@{|->}[r] & f(e)} & &\xymatrix@R=1em{X^M \ar[r]^-{\delta} & (X^M)^M \cong X^{M \times M}\\ f \ar@{|->}[r] & \left((m,n) \mapsto f(mn)\right)}
\end{align*}
\end{exa}
\begin{exa}\label{filter}
Recall that a \emph{filter} $\mathscr{F}$ on a set $X$ is a non-empty subset of $P(X)$ such that:
\begin{enumerate}
\item The set $X$ is in $\mathscr{F}$, and $A,B \in \mathscr{F} \implies A \cap B \in \mathscr{F}$.
\item The empty set $\emptyset$ is not in $\mathscr{F}$.
\item If $B\subseteq X$ and there exists $A \in \mathscr{F}$ with $A \subseteq B$, then $B \in \mathscr{F}.$
\end{enumerate}
An \emph{ultrafilter} $\mathscr{U}$ is a filter such that $A \cup B \in \mathscr{U} \implies A \in \mathscr{U}$ or $B \in \mathscr{U}$.
\end{exa}
There is a functor $F \colon \mathbf{Set} \to \mathbf{Set}$ which assigns to $X$ the set $FX$ of filters on $X$. A function $f \colon X \to Y$ is mapped to the function $Ff \colon FX \to FY$, defined by
$$
F(f)(\mathscr{F}) = \{ A \subseteq Y \mid f^{-1}(A) \in \mathscr{F}\}.
$$
Given $\mathbb G \in FFX$, we define the \emph{Kowalski sum} to be the filter
$$
\mu(\mathbb G) = \{ A \subseteq X \mid \{\mathscr{F} \in FX \mid A \in \mathscr{F} \} \in \mathbb G\},
$$
and given $x \in X$, we define the \emph{principal ultrafilter} on $x$ to be the (ultra)filter
$$
\eta(x) = \{ A \subseteq X \mid x \in A\}.
$$
These define natural transformations $\mu \colon FF \Rightarrow F$ and $\eta \colon 1 \Rightarrow F$ turning $F$ into a monad. By restricting to ultrafilters, we get an additional monad $U$ on $\mathbf{Set}$. See~\cite[Chapter~II]{MR3307673} for more information on these monads.
\begin{exa}
Given a set $X$, let $DX$ be the set of (countable) probability distributions on $X$, that is,
$$
DX = \left\{ p \in [0,1]^X \ \Big|\ \left|p^{-1}(0, 1]\right| \le \aleph_0,\ \sum_{x \in X} p(x) = 1 \right\}.
$$
Given a function $f \colon X \to Y$, we get a function $Df \colon DX \to DY$ which maps a distribution $p \colon X \to [0,1]$ to the distribution $Y \to [0,1]$ defined by
$$
y \mapsto \displaystyle\sum_{x \in f^{-1}(y)} p(x)
$$
This defines a functor $D \colon \mathbf{Set} \to \mathbf{Set}$. Let $\eta \colon X \to DX$ be the function which maps $x$ to the characteristic function $\chi_{\{x\}} \colon X \to [0,1]$, and let $\mu \colon DDX \to DX$ be the function that sends a distribution $P \colon DX \to [0,1]$ to the distribution $\mu(P) \colon X \to [0,1]$ defined by
$$
\mu(P) (x) = \sum_{p \in DX} P(p) \cdot p(x).
$$
These define natural transformations $\eta, \mu$ that turn $D$ into a monad.
The $D$-algebras are \emph{superconvex spaces}, i.e.\ sets where one can take convex combinations of countably many elements, see~\cite{MR659884, MR1421176} for more information.
\end{exa}
\subsection{Distributive laws}
Let $C = \{c_0, \ldots, c_n\}$ be a non-empty, finite, totally-ordered set with $c_0 < \cdots < c_n$, and recall the comonad $C \times -$ of Example~\ref{finordset}.
\begin{exa}\label{suppreserving}
For any set $X$, define $\theta \colon P(C \times X) \to C \times PX$ by
$$
\theta (A \subseteq C \times X) = \begin{cases} (\sup \pi_1(A), \pi_2(A) ) & \mbox{ if } A \neq \emptyset \\
(c_0, \emptyset) & \mbox{ otherwise }
\end{cases}
$$
This defines a mixed distributive law $\theta \colon P(C \times{-}) \Rightarrow C \times P{-}$.
\end{exa}
\begin{exa}\label{supsagain}
There is a mixed distributive law $\theta \colon L(C \times {-}) \Rightarrow C \times L{-}$ defined by
$$
\theta [ (c_1, x_1), \ldots, (c_m, x_m) ] = \begin{cases}
(\sup_i c_i, [x_1, \ldots, x_m]) & \mbox{ if } m >0 \\
(c_0, [~]) & \mbox{ if } m = 0
\end{cases}
$$
where $[~]$ denotes the empty list.
\end{exa}
\begin{exa}
Let $\theta \colon D(C \times X) \to C \times DX$ be the map into the product, defined by the two maps
\begin{align*}
&\xymatrix@R=1em{D(C \times X) \ar[r]^-{\theta_1} & C \\ p \ar@{|->}[r] &\displaystyle\sup_{\sum_{x \in X} p(c,x) \neq 0} c} & &\xymatrix@R=1em{D(C\times X) \ar[r]^-{\theta_2} & DX\\ p \ar@{|->}[r] & \left( x \mapsto \displaystyle\sum_{c \in C} p(c,x)\right)}
\end{align*}
This defines a distributive law $\theta \colon D(C \times{-}) \Rightarrow C \times D{-}$.
\end{exa}
In Examples~\ref{11}--\ref{15}, let $M$ be a monoid, giving rise to the comonad $S$ of Example~\ref{monoidcomonad} which maps a set $X$ to $X^M$.
\begin{exa}\label{11}
The maps $\theta \colon L(X^M) \to (LX)^M$ given by
$$
\theta[f_1, \ldots, f_n](m) = [f_1(m), \ldots, f_n(m)]
$$
define a mixed distributive law $\theta \colon LS \Rightarrow SL$.
\end{exa}
\begin{exa}\label{13}
We define maps $\theta \colon P(X^M) \to (PX)^M$ as follows. For $A \in P(X^M)$ we let
$$
\theta(A ) (m) = \{ f(m) \mid f \in A\}.
$$
This defines a mixed distributive law $\theta \colon PS \Rightarrow SP$.
\end{exa}
\begin{exa}\label{14} Recall the filter monad $F$ of Example~\ref{filter}.
There is a mixed distributive law $\theta \colon FS \Rightarrow SF$, defined as follows. For each $m \in M$, let $\mathrm{ev}^m \colon X^M \to X$ denote the natural map $f \mapsto f(m)$. Then $\theta \colon F(X^M) \to (FX)^M$ is defined by
$$
\theta(\mathscr{F})(m) = F(\mathrm{ev}^m)(\mathscr{F}).
$$
In a similar way, we get a mixed distributive law $US \Rightarrow SU$ by restricting to ultrafilters.
\end{exa}
\begin{exa}\label{15}
Define maps $\theta \colon D(X^M) \to (DX)^M$ as follows. For a distribution $p \colon X^M \to [0,1]$, $\theta(p)$ should be a function $M \to DX$, and so $\theta(p)(m)$ should be a distribution $X \to [0,1]$. We define this by
$$
\theta(p)(m)(x) = \sum_{\substack{f \in X^M \\ x = f(m)}} p(f).
$$
This defines a mixed distributive law $\theta \colon DS \Rightarrow SD$.
\end{exa}
\subsection{$\tilde\theta$-coalgebras and entwined algebras }
Recall that the mixed distributive laws $\theta$ in the preceding section lift to comonad distributive laws $\tilde\theta$ on the Eilenberg-Moore category of the relevant monad, cf.\ Section~\ref{extremalcase}.
\begin{exa}
Consider the powerset monad $P$. A $P$-algebra $(X, \beta)$ is the same as a partially ordered set such that every subset has a supremum. Indeed, the partial order is defined by
$$
x \le y \iff \beta\{x, y\} = y
$$
and $\beta \colon PX \to X$ points out the supremum of a given subset. Now consider the distributive law $\theta \colon P(C \times{-}) \Rightarrow C \times P{-}$ of Example~\ref{suppreserving}. By Proposition~\ref{chicoalgprop}, $\tilde\theta$-coalgebras in $\mathbf{Set}^P$ correspond to $P$-algebras $(X, \beta)$ whose underlying set $X$ admits the structure of a $(C\times{-})$-coalgebra, which means that $X$ comes equipped with a partition of subsets indexed by the `colours' in $C$. We denote this partition by a colour function $\kappa \colon X \to C$. The $\theta$-entwined algebras are $P$-algebras of this form, such that $\kappa$ preserves suprema, i.e.\
$$
\kappa(\sup A) = \sup_{a \in A} \kappa(a)
$$
for all $A \subseteq X$.
\end{exa}
\begin{exa}\label{7}
Consider the list monad $L$. An $L$-algebra $(X, \beta)$ is the same thing as a monoid structure on $X$. The identity element $e$ is given by $\beta([~])$, and the multiplication is defined by $xy := \beta[x,y]$.
Let $\theta \colon L(C\times{-}) \Rightarrow C \times L{-}$ be the distributive law of Example~\ref{supsagain}.
By Proposition~\ref{chicoalgprop}, $\tilde\theta$-coalgebras in $\mathbf{Set}^L$ correspond to monoids $X$ whose underlying set admits a partition $\kappa \colon X \to C$. The
$\theta$-entwined algebras are those monoids of this form such that the colour of a product in $X$ is the supremum of the colours in the product, that is
$$
\kappa(x_1 \cdots x_n ) = \sup_i \kappa (x_i).
$$
Note that the empty list must be included, so in particular the colour of the identity is minimal, that is
$$
\kappa(e) = c_0.
$$
The set $C$ becomes a monoid with unit $c_0$ and multipication given by $\sup$, and so $\kappa$ is a monoid map.
\end{exa}
\begin{exa}
Let $\theta \colon LS \Rightarrow SL $ be the distributive law of Example~\ref{11}. Let $X$ be a $L$-algebra (a monoid). By Proposition~\ref{chicoalgprop}, a $\tilde\theta$-coalgebra structure on $X$ is equivalent to a map $\nabla \colon X \to X^M$, or equivalently a map $\nabla_m \colon X \to X$ for each $m \in M$, such that
$\nabla_e = 1 $ and $\nabla_{mn} = \nabla_n \nabla_m$
for all $m,n \in M$. In other words, $\nabla$ defines a monoid map $M \to \mathbf{Set}(X,X)^*$. The $\theta$-entwined algebras are those monoids $X$ equipped with such a map $\nabla$ whereby the image of $\nabla$ lies in the monoid endomorphisms of $X$.
\end{exa}
\begin{exa}
Let $\theta \colon PS \Rightarrow SP$ be the distributive law of Example~\ref{13}. By Proposition~\ref{chicoalgprop}, a $\tilde\theta$-coalgebra structure on a $P$-algebra $(X, \beta)$ is equivalent to a map $\nabla \colon X \to X^M$ as in Example~\ref{7}. The set $X^M$ can be given a partial order, defined by $f \le g$ if and only if $f(m) \le g(m)$ for all $m \in M$. This poset is itself a $P$-algebra, since given a subset $H \subseteq X^M$ we can define a function $\sup H \colon M \to X$ given by
$$
\sup H (m) = \sup \{ f(m) \mid f \in H \}
$$
which genuinely is the supremum of $H$. Thus the $\theta$-entwined algebra structures on $P$-algebras $(X, \beta)$ correspond to coassociative coactions $\nabla \colon X \to X^M$ which are sup-preserving.
\end{exa}
\chapter{Duplicial objects and cyclic homology}\label{CYCLIC}
We begin this chapter by recalling the definitions of simplicial and duplicial objects as well as cyclic homology in Section~\ref{simpmeth}. Following this,
in Section~\ref{cychomol} we recount the method of constructing duplicial objects from~\cite{MR2415479}. We emphasize
the self-duality of the situation by defining in fact two duplicial objects $C_T(N,M)$ and $C^*_S(N,M)$, arising from bar resolutions using comonads $T,S$, as well as suitable functors $M,N$. There is a canonical
pair of morphisms of duplicial objects between these which are mutual inverses if and
only if the two objects are cyclic (Theorem~\ref{cyc}).
In Section~\ref{duplobjsec}, we further develop the process of twisting a pair of coefficient objects
$M, N$ detailed previously in Section~\ref{twistcoeff}. We refrain from giving concrete examples in this chapter, and instead postpone this until Chapter~\ref{EXAMPLES}.
Apart from the basic definitions and cited material, the work in Sections~\ref{cychomol} and~\ref{duplobjsec} is original. The latter section is based on the work of~\cite[\S4.5--4.11]{1} but goes into a bit more detail.
\section{Simplicial methods}\label{simpmeth}
Here we go over the very basics of simplicial homotopy theory, see e.g.~\cite{MR2840650,MR1950475} for a thorough account. We also discuss the duplicial objects of Dwyer and Kan~\cite{MR826872}.
\subsection{The simplicial and augmented simplicial categories}
\begin{defn}\label{simpldef}
The \emph{simplicial category}, or the \emph{topologists' simplicial category}, is the category $\Delta$ whose objects are finite non-zero ordinals
$$
\mathbf n = \{0, 1, \dots, n\}
$$
and whose morphisms are the (weakly) order-preserving maps between these. The \emph{augmented simplicial category}, or the \emph{algebraists' simplicial category} is the category $\Delta_+$ whose objects are the same as $\Delta$ plus an additional object
$$
\mathbf{-1} = \emptyset
$$
and whose morphisms are again the (weakly) order-preserving maps.
\end{defn}
It is straightforward to show that the morphisms of $\Delta, \Delta_+$ are generated by functions
$$
\xymatrix{
\mathbf{n-1} \ar[r]^-{\delta_i} & \mathbf n
},\qquad
\xymatrix{
\mathbf{n + 1} \ar[r]^-{\sigma_i} & \mathbf n
}
$$
for all suitable $n$ and for all $0\le i \le n$, called \emph{face maps} and \emph{degeneracy maps} respectively, such that the \emph{simplicial identities} are satisfied:
$$
\begin{array}{cc}
\delta_j \delta_i
=
\delta_i\delta_{j-1}
&
\mbox{ if } i < j
\\
\sigma_j \sigma_i
=
\sigma_i \sigma_{j+1}
&
\mbox{ if } i \le j
\end{array}
\qquad\qquad
\sigma_j \delta_i = \begin{cases}
\delta_i \sigma_{j-1} & \mbox{ if } i < j \\
1 &\mbox{ if } i = j, j+1 \\
\delta_{i-1} \sigma_j & \mbox{ if } i > j + 1
\end{cases}
$$
\subsection{The duplicial and cyclic categories}
\begin{defn}
The \emph{duplicial} \emph{category} $\Lambda_\infty$ is constructed as follows: we take the category $\Delta$ and adjoin in each degree an abstract morphism $\tau \colon \mathbf n \to \mathbf n$ subject to the relations:$$
\tau \delta_i =
\begin{cases}
\delta_{i-1} \tau & \mbox{ if } 1 \le i \le n \\
\delta_n & \mbox{ if } i = 0
\end{cases}
\qquad
\tau \sigma_j =
\begin{cases}
\sigma_{j-1} \tau & \mbox{ if } 1 \le j \le n \\
\sigma_n \tau^2 & \mbox{ if } j = 0
\end{cases}
$$
We obtain the \emph{cyclic category} $\Lambda$ by adding the extra requirement that in each degree, the morphism $\tau \colon \mathbf n \to \mathbf n$ satisfies the relation $\tau^{n+1} = 1$.
Similarly, we have augmented versions of these categories by considering $\Delta_+$ rather than $\Delta$ in the above definition.
\subsection{Simplicial, duplicial and cyclic objects}
Let $\mathcal{Z}$ be a category.
\begin{defn}
A \emph{simplicial object} in $\mathcal{Z}$ is an object of the functor category $[\Delta^*, \mathcal{Z}]$, so a contravariant functor from $\Delta$ to $\mathcal{Z}$.
\end{defn}
Thus to give a simplicial object $X$ in $\mathcal{Z}$ is to give a sequence $X_0, X_1, \ldots$ of objects in $\mathcal{Z}$
and in each degree $n \ge 0$ morphisms
$$
\xymatrix{
X_n \ar[r]^-{d_i} & X_{n-1}
},\qquad
\xymatrix{
X_n \ar[r]^-{s_i} & X_{n+1}
}
$$
for all $0\le i \le n$ subject to the conditions:$$
\begin{array}{cc}
d_i d_j
=
d_{j-1}d_i
&
\mbox{ if } i < j
\\
s_i s_j
=
s_{j+1} s_i
&
\mbox{ if } i \le j
\end{array}
\qquad\qquad
d_i s_j = \begin{cases}
s_{j-1} d_i & \mbox{ if } i < j \\
1 &\mbox{ if } i = j, j+1 \\
s_j d_{i-1} & \mbox{ if } i > j + 1
\end{cases}
$$
Similarly, we define \emph{augmented simplicial objects} as functors $\Delta_+^* \to \mathcal{Z}$.
\begin{rem}\label{opsimpobj}
Given a simplicial object $X$, by reversing the order of both the face maps and degeneracy maps we obtain a new simplicial object. We call this the \emph{opsimplicial simplicial object associated to $X$}.
\end{rem}
\begin{defn}[see~\cite{MR826872}]
A \emph{duplicial object} in $\mathcal{Z}$ is an object of the functor category $[\Lambda_\infty^*, \mathcal{Z}]$, so a contravariant functor from $\Lambda_\infty$ to $\mathcal{Z}$.
\end{defn}
Thus, explicitly, a duplicial object $X$ in $\mathcal{Z}$ is a simplicial object $X$ together with a morphism, or \emph{duplicial operator},
$$
\xymatrix{
X_n \ar[r]^-{t} & X_n
}
$$
in each degree subject to the conditions:
$$
d_i t=
\begin{cases}
t d_{i-1} & \mbox{ if } 1 \le i \le n \\
d_n & \mbox{ if } i = 0
\end{cases}
\qquad
s_j t =
\begin{cases}
t s_{j-1} & \mbox{ if } 1 \le j \le n \\
t^2 s_n & \mbox{ if } j = 0
\end{cases}
$$
Similarly again, we have the notion of \emph{cyclic objects}~\cite{MR777584}, which are functors $\Lambda^* \to \mathcal{Z}$, as well as augmented versions.
\begin{rem}
Some authors use term \emph{paracyclic} to mean duplicial in our sense. Beware of this terminology, since it has also been used to describe those duplicial objects where the operator $t$ or $T$ (see Section~\ref{tdef}) is an isomorphism. Furthermore,
the original sense of a duplicial structure~\cite{MR826872} on a simplicial object $X$ is precisely a duplicial structure, in our sense, on the opsimplicial simplicial object associated to $X$, cf.~Remark~\ref{opsimpobj} and~Section~\ref{duplobjsec}.
\end{rem}
\subsection{The Dold-Kan and Dwyer-Kan correspondences}
Now suppose that $\mathcal{Z}$ is an abelian category. Let $\operatorname{C}(\mathcal{Z})$ denote the category of chain complexes in $\mathcal{Z}$, and let
$\operatorname{C}^+(\mathcal{Z})$ denote the full subcategory of $\operatorname{C}(\mathcal{Z})$ consisting of non-negatively graded chain complexes.
\begin{defn}[see~\cite{MR826872,MR883882}]
A \emph{duchain complex} in $\mathcal{Z}$ is a triple $(X, b, B)$ where $(X, b)$ is a chain complex and $(X, B)$ is a cochain complex:
$$
\xymatrix{
\cdots \ar@<1ex>[r]^-b & \ar@<1ex>[l]^-B X_n \ar@<1ex>[r]^-b & \ar@<1ex>[l]^-B X_{n-1} \ar@<1ex>[r]^-b & \ar@<1ex>[l]^-B \cdots
}
$$
A \emph{mixed complex} is a duchain complex $(X, b, B)$ such that $bB + Bb = 0$.
\end{defn}
\begin{rem}
Since both chain and cochain complexes are involved in the definition of a duchain complex, we point out that whenever we use the terminology \emph{quasi-isomorphism}, we mean a morphism which induces an isomorphism on homology, but not necessarily on cohomology.
\end{rem}
Let $\operatorname{D}(\mathcal{Z})$ denote the category of duchain complexes in $\mathcal{Z}$, where a morphism is defined to be one which is simultaneously a morphism of chain complexes and of cochain complexes. Let $\operatorname{M}(\mathcal{Z})$ denote the full subcategory of $\operatorname{D}(\mathcal{Z})$ given by mixed complexes. Let $\operatorname{M}^+(\mathcal{Z}) \subseteq \operatorname{D}^+(\mathcal{Z})$ denote the non-negatively graded versions.
The following theorem is known as the \emph{Dold-Kan} correspondence. We sketch the proof, but for a full proof see~\cite[Thm.~8.4.1]{MR1269324} and its following remarks, or see~\cite[Chapter~III.2]{MR2840650}.
\begin{thm}\label{doldkan}
There is an equivalence of categories
$$
[\Delta^*, \mathcal{Z}] \simeq \operatorname{C}^+(\mathcal{Z}).
$$
\end{thm}
\begin{proof}[Sketch proof]
The equivalence $ N \colon [\Delta^*, \mathcal{Z}] \to \operatorname{C}^+(\mathcal{Z})$ is defined on objects as follows. A simplicial object $X$ is mapped to the chain complex $NX$ given in degree $n$ by
$$
NX_n = \quotient{X_n}{ \sum_{i=0}^{n-1} \operatorname{im} s_i }
$$
whose differential $NX_{n} \to NX_{n-1}$ is defined by the diagram
$$
\xymatrix{
X_n \ar@{->>}[d] \ar[rr]^-{\sum_{i=0}^{n} (-1)^i d_i} && X_{n-1} \ar@{->>}[d] \\
NX_n \ar@{.>}[rr] && NX_{n-1}
}
$$
\end{proof}
The following theorem is known as the \emph{Dwyer-Kan correspondence}, proved in~\cite{MR826872}. We do not make explicit use of it in later sections, but nevertheless we include it here as we feel it highlights the importance of duplicial objects. We give only a sketch proof.
\begin{thm}\label{dwyerkan}
There is an equivalence of categories
$$
[\Lambda_\infty^*, \mathcal{Z}] \simeq \operatorname{D}^+(\mathcal{Z}).
$$
\end{thm}
\begin{proof}[Sketch proof]
The equivalence $ [\Lambda_\infty^*, \mathcal{Z}] \to \operatorname{D}^+(\mathcal{Z}) $ is defined on objects as follows. Given a duplicial object $X$, we define a codifferential on the complex $NX$, where $N$ is the equivalence of Theorem~\ref{doldkan}. The codifferential is defined by the diagram
$$
\xymatrix{
X_n \ar@{->>}[d] \ar[rr]^-{(-1)^n t s_n} && X_{n+1} \ar@{->>}[d] \\
NX_n \ar@{.>}[rr] && NX_{n+1}
}
$$
in degree $n$.
\end{proof}
\section{Cyclic homology}\label{cychomol}
In this section, we define the cyclic homology of various objects and show that the notions coincide in appropriate cases. Throughout, $\mathcal{Z}$ denotes an abelian category.
\subsection{Cyclic homology of a mixed complex}
Given a mixed complex $(X,b,B)$ in $\operatorname{M}^+(\mathcal{Z})$, we construct a bicomplex in the first quadrant:
$$
\xymatrix{
\vdots \ar[d]_-b & \vdots \ar[d]_-b & \vdots \ar[d]_-b \\
X_2\ar[d]_-b&\ar[l]^-B X_1\ar[d]_-b &\ar[l]^-B X_0 \\
X_1 \ar[d]_-b& \ar[l]^-B X_0 \\
X_0
}
$$
This just means that each square anticommutes, so there is an associated total complex $T(X)$ given in degree $n$ by
$$
T(X)_n = \bigoplus_{i \le 0} X_{n + 2i}
$$
(so the first few degrees are $X_0, X_1, X_0 \oplus X_2, X_1 \oplus X_3, X_0 \oplus X_2 \oplus X_4, \ldots$ and one can spot the pattern) with differential $b + B$.
The notion of cyclic homology is due to Connes~\cite{MR823176,MR777584} but the following definition is from~\cite{MR883882}.
\begin{defn}\label{cyclichomology}
The \emph{cyclic homology} of the mixed complex $X$, denoted $\operatorname{HC}(X)$, is the homology of the total complex $T(X)$, i.e.\
$$
\operatorname{HC}(X) := \operatorname{H}(T(X))
$$
\end{defn}
Thus we have defined a functor
$$
\operatorname{HC} \colon \operatorname{M}^+(\mathcal{Z}) \longrightarrow \mathcal{Z}.
$$
Let $U \colon \operatorname{M}^+(\mathcal{Z}) \longrightarrow \operatorname{C}^+(\mathcal{Z})$ denote the obvious forgetful functor to the category of non-negatively graded chain complexes.
The following proposition is a useful one, and we refer the reader to~\cite[Proposition~2.5.15]{MR1600246} for a proof:
\begin{prop}\label{quasi}
Let $f$ be a morphism in $\operatorname{M}^+(\mathcal{Z})$. Then $Uf$ is a quasi-isomorphism if and only if $\operatorname{HC}(f)$ is an isomorphism.
\end{prop}
\subsection{Cyclic homology of a cyclic object}\label{cychomcycobj}
Let $X \colon \Lambda^{*} \to \mathcal Y$ be a cyclic object. Consider the following morphisms, defined in degree $n$:
$$
\tilde t := (-1)^n t \quad \quad \quad s_{-1} = ts_n \quad \quad \quad \mathcal N_n = \sum_{i=0}^n \tilde t^i \quad \quad \quad b_n = \sum_{i=0}^n (-1)^i d_i
$$
By Theorem~\ref{doldkan}, $(X,b)$ is a chain complex. In fact, the morphism $\hat B = s_{-1} \mathcal N$ has the property that $b \hat B + \hat B b = 0$, but unfortunately, we do not have that $\hat B^2 = 0$ and so it is not a codifferential. However, the morphism $B := (1-\tilde t)s_{-1}\mathcal N$ does square to zero, and also anticommutes with $b$. Thus, $(X,b,B)$ is a mixed complex. We have defined a functor
$$
\xymatrix{
[\Lambda^*, \mathcal{Z}] \ar[r]^-K & \operatorname{M}^+(\mathcal{Z})
}
$$
and we define the cyclic homology of a cyclic object to be the composite of this functor with $\operatorname{HC}$ of Definition~\ref{cyclichomology}. By abuse of notation, we denote this also by $\operatorname{HC}$, and so we have a commutative diagram
$$
\xymatrix{
[\Lambda^*, \mathcal{Z}] \ar[r]^-K \ar[dr]_-{\operatorname{HC}} & \operatorname{M}^+(\mathcal{Z}) \ar[d]^-{\operatorname{HC}} \\
& \mathcal{Z}
}
$$
Alternatively, one could do the following: Let $NX$ be the complex associated to the underlying simplicial object of the cyclic object $X$, i.e.\
$$
( N X)_n = \quotient{X_n}{ \sum_{i=0}^{n-1} \operatorname{im} s_i}
$$
as in the proof of Theorem~\ref{doldkan}. Both the morphisms $b$ and $B$ descends to this quotient, defining a functor
$$
\xymatrix{
[\Lambda^*, \mathcal{Z}] \ar[r]^-{K'} & \operatorname{M}^+(\mathcal{Z}).
}
$$ It is well-known that the quotient morphism $q \colon (X,b) \to (NX, b)$ is a quasi-isomorphism (see e.g.~\cite[Thm.~8.3.8]{MR1269324}), and it descends to the quotient too. Therefore, $q$ is actually a morphism $KX \to K'X$ of mixed complexes, defining a natural transformation
$$q \colon K \Rightarrow K'.$$ By Proposition~\ref{quasi}, this gives an isomorphism after composing with $\operatorname{HC}$. This proves the following:
\begin{prop}
We have a square
$$
\xymatrix{
[\Lambda^*, \mathcal{Z}] \ar[rr]^-K \ar[dd]_-{K'} && \operatorname{M}^+(\mathcal{Z}) \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\operatorname{HC} q}^-\cong "a";"b" \ar[dd]^-{\operatorname{HC}} \\
\\
\operatorname{M}^+(\mathcal{Z}) \ar[rr]_-{\operatorname{HC}} && \mathcal{Z}
}
$$
in $\mathbf{Cat}$.
\end{prop}
In summary, we have two equivalent ways of computing the cyclic homology of a cyclic object: either define the structure of a mixed complex on the actual cyclic object, or do so on the normalised complex of the underlying simplicial object.
\subsection{Cyclic homology of a duchain complex}\label{tdef}
Let $(X,b,B)$ be a duchain complex in $\operatorname{D}^+(\mathcal{Z})$. Define an endomorphism
$$
T = 1 - bB -Bb
$$
in each degree. One easily checks that $b T = Tb$ and $BT = TB$, so clearly both $b$ and $B$ descend to the quotient $B/\operatorname{im}(1-T)$, and in that case $bB + Bb = 0$. This defines a functor $F \colon \operatorname{D}^+(\mathcal{Z}) \longrightarrow \operatorname{M}^+(\mathcal{Z})$, which in fact is a left adjoint to the forgetful functor $U$.
$$
\xymatrix{
\operatorname{D}^+(\mathcal{Z}) \ar@/^1pc/[r]^-F \ar@{}[r]|\perp& \operatorname{M}^+(\mathcal{Z}) \ar@/^1pc/[l]^-{U}
}
$$
We define the cyclic homology of a duchain complex to be the composite of $\operatorname{HC}$ with this functor $F$. Note that $FU \cong 1$. Again we abuse notation to give a commutative diagram
$$
\xymatrix{
\operatorname{D}^+(\mathcal{Z}) \ar[dr]_-{\operatorname{HC}} \ar[r]^-F & \operatorname{M}^+(\mathcal{Z}) \ar[d]^-{\operatorname{HC}} \\
& \mathcal{Z}
}
$$
\subsection{Cyclic homology of a duplicial object}
Let $X \colon \Lambda_\infty^* \to \mathcal{Z}$ be a duplicial object. In light of the above story for cyclic objects, there are two natural ways we could proceed to define cyclic homology of $X$: either construct a natural cyclic object from $X$ and take the cyclic homology of that, or instead, construct a duchain complex and take the cyclic homology of that. Here we show these two definitions are equivalent.
Let $\pi \colon \Lambda_\infty^* \longrightarrow \Lambda^*$ denote the quotient map. This induces a forgetful functor $\pi^* \colon [\Lambda^*, \mathcal{Z}]\longrightarrow [\Lambda_\infty^*, \mathcal{Z}]$. By the theory of Kan extensions, this functor has a left adjoint $\pi_!$.
$$
\xymatrix{
[\Lambda_\infty^*, \mathcal{Z}] \ar@/^1pc/[r]^-{\pi_!} \ar@{}[r]|\perp& [\Lambda^*, \mathcal{Z}] \ar@/^1pc/[l]^-{\pi^*}
}
$$
This left adjoint is explicitly defined by $(\pi_! X)_n = X_n / \operatorname{im} (1 - t^{n+1})$. The face maps and degeneracy maps descend to this quotient. Note that $\pi_! \pi^* \cong 1$. We define the cyclic homology of $X$ to be the cyclic homology of $\pi_! X$, giving a commutative diagram
$$
\xymatrix{
[\Lambda_\infty^*,\mathcal{Z}] \ar[r]^-{\pi_!} \ar[dr]_-{\operatorname{HC}} & [\Lambda^*, \mathcal{Z}] \ar[d]^-{\operatorname{HC}} \\
& \mathcal{Z}
}
$$
On the other hand, we can define morphisms $b,\tilde t, s_{-1}, \mathcal N, B, \hat B$ as before. Again, the morphism $\hat B$ does not square to zero, but neither does it satisfy the relation $b \hat B + \hat B b = 0$. In the duplicial case, the morphism $B$ does not satisfy these relations either. On the normalised complex $(NX, b)$ the morphism $B$ descends to the quotient (the proof in the cyclic case only uses the duplicial structure), is equal to $\hat B$, and satisfies $B^2 = 0$. This only gives us a duchain complex $(NX,b,B)$ since $bB + Bb$ is still not zero in the quotient. This defines a functor
$$
\xymatrix{
[\Lambda_\infty^*, \mathcal{Z}] \ar[r]^-{P} & \operatorname{D}^+(\mathcal{Z})
}
$$
and it is important to note that this is \emph{not} the same as the Dwyer-Kan normalisation functor of Theorem~\ref{dwyerkan}.
In a similar way to Section~\ref{cychomcycobj}, the quotient morphism defines a map of mixed complexes $q \colon K \pi_! X \to FPX $ defining a natural transformation
$$
\xymatrix{
[\Lambda_\infty^*, \mathcal{Z}] \ar[rr]^-{\pi_!} \ar[dd]_-{P} && [\Lambda^*, \mathcal{Z}] \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{q}"a";"b" \ar[dd]^-{K} \\
\\
\operatorname{D}^+(\mathcal{Z}) \ar[rr]_-{F} && \operatorname{M}^+(\mathcal{Z})
}
$$
which becomes an isomorphism after composing with $\operatorname{HC} \colon \operatorname{M}^+(\mathcal{Z}) \to \mathcal{Z}$, thus proving:
\begin{prop}
We have a square
$$
\xymatrix{
[\Lambda_\infty^*, \mathcal{Z}] \ar[rr]^-{\pi_!} \ar[dd]_-{P} && [\Lambda^*, \mathcal{Z}] \ar@{}[ddll]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\operatorname{HC} q}^-\cong "a";"b" \ar[dd]^-{\operatorname{HC}} \\
\\
\operatorname{D}^+(\mathcal{Z}) \ar[rr]_-{\operatorname{HC}} && \mathcal{Z}
}
$$
in $\mathbf{Cat}$.
\end{prop}
\end{defn}
\section{From distributive laws to duplicial objects}\label{duplobjsec}
Here we develop B\"ohm and {\c S}tefan's work on duplicial objects~\cite{MR2956318,MR2415479}, by showing that some of the involved morphisms extend to simplicial morphisms. We use these morphisms to characterise the cyclicity of the relevant duplicial functor.
\subsection{The bar and opbar resolutions}
Let $\mathcal{B}$ and $\mathcal{Y}$ be two categories.
\begin{defn}
A \emph{simplicial functor} $\mathcal{Y} \to \mathcal{B}$ is a simplicial object in the category $[\mathcal{Y}, \mathcal{B}]$.
\end{defn}
Similarly, we have notions of \emph{dupicial} and \emph{cyclic functors}, as well as augmented versions of these.
Let $(T, \delta, \epsilon)$ be a comonad on $\mathcal{B}$ and let $M \colon \mathcal{Y} \rightarrow \mathcal{B}$ be a
functor.
\begin{defn}\label{bardeffo}
The \emph{bar resolution of }$M$
is the simplicial functor
$ \rBB (T, M) \colon \mathcal{Y} \rightarrow
\mathcal{B}$ defined by
$$
\rBB (T, M)_n =
T^{n+1}M, \qquad
d_i =T^i \epsilon
T^{n-i} M, \qquad
s_j =T^j \delta T^{n-j}
M,
$$
where the face and degeneracy maps
above are given in degree $n$. The \emph{opbar
resolution of }$M$,
denoted
$\rBB^*(T, M)$, is the simplicial
functor obtained by taking the op\-sim\-pli\-cial
sim\-pli\-cial
functor (cf.~Remark~\ref{opsimpobj}) associated to $\rBB (T, M)$. Explicitly:
$$
\rBB^*(T, M)_n =
T^{n+1}M, \qquad d_i =T^{n-i} \epsilon T^{i} M,\qquad
s_j =T^{n-j} \delta T^{j} M.
$$
\end{defn}
Given any functor $\mathrm N \colon \mathcal{B} \rightarrow
\mathcal{Z}$, we compose it with the above simplicial
functors to obtain new simplicial functors that we
denote by
$$
\rCC_T(N,M):=N\rBB (T,M), \qquad
\rCC^*_T(N,M):=N\rBB ^*(T, M).
$$
\subsection{The B\"ohm-\c Stefan construction}
\label{evidenziatore1}
Let
$$\xymatrix{(\mathcal{B}, T) \ar[rr]^-{(\Sigma, \sigma)} && (\mathcal{D}, G)}$$ be an opmorphism of comonads, so in particular $\sigma $ is a natural transformation $G\Sigma \Rightarrow \Sigma T$. By abuse of notation, we let $\sigma^n$
denote the natural transformation
$G^n \Sigma \Rightarrow \Sigma T^n$ obtained by repeated application of
$\sigma$ (up to horizontal composition of identities),
where $\sigma^0 = 1$. In other words, we define $\sigma^n$ recursively: for all $n \ge 0$, $\sigma^{n+1} \colon G^{n+1}\Sigma \Rightarrow \Sigma T^{n+1}$ is given by the composite
$$
\xymatrix{
G^{n+1}\Sigma \ar[rr]^-{G\sigma^n} && G\Sigma T^n \ar[rr]^-{\sigma T^n} && \Sigma T^{n+1}
}.
$$
\begin{lem}\label{superchiepsilon}
For all $n \ge 0$ and for $0 \le i \le n$, the diagram
$$
\xymatrix{
G^{n+1}\Sigma \ar[rr]^-{\sigma^{n+1}} \ar[d]_-{G^{n-i}\epsilon T^i \Sigma} && \Sigma T^{n+1} \ar[d]^-{\Sigma T^{n-i}\epsilon T^i } \\
G^n\Sigma \ar[rr]_-{\sigma^n} && \Sigma T^n
}
$$
commutes.
\end{lem}
\begin{proof}
We proceed by induction. For $n = 0$ and $i = 0$, the result holds since it becomes the compatibility condition of $\sigma$ with $\epsilon \colon T \Rightarrow 1$. Now, suppose the statement holds for $n - 1$, and let $i$ be such that $0 \le i < n$. Consider the diagram
$$
\xymatrix{
G^{n+1} \Sigma \ar[d]_-{G^{n-i} \epsilon G^i \Sigma } \ar[rr]^-{G\sigma^n} && G\Sigma T ^n\ar[d]_-{G\Sigma T^{n-i-1} \epsilon T^i} \ar[rr]^-{\sigma T^n } && \Sigma T^{n+1} \ar[d]^-{\Sigma T^{n-i} \epsilon T^i } \\
G^n \Sigma \ar[rr]_-{G\sigma^{n-1}} && G\Sigma T^{n-1} \ar[rr]_-{\sigma T^{n-1}} && \Sigma T^n
}
$$
The right-hand square commutes by naturality of $\sigma$, and the left-hand square commutes by the inductive hypothesis. For $i = n$, consider the diagram
$$
\xymatrix{
G^{n+1} \Sigma \ar[d]_-{\epsilon G^n \Sigma } \ar[rr]^-{G\sigma^n} && G\Sigma T^n \ar[drr]_-{\epsilon \Sigma T^n \ \ } \ar[rr]^-{\sigma T^n } && \Sigma T^{n+1} \ar[d]^-{\Sigma\epsilon T^n} \\
G^n \Sigma \ar[rrrr]_-{\sigma^n } && &&\Sigma T^n
}
$$
The right-hand triangle commutes by compatibility of $\sigma$ with $\epsilon$ and the left-hand quadrilateral commutes by naturality of $\epsilon$. In each case, the outer diagram commutes, as required.
\end{proof}
There is a similar statement for the coproducts $\delta$. Thus, we have proved:
\begin{prop}\label{supersigma}
The morphisms $\sigma^n$ define a morphism of simplicial functors
$$
\xymatrix{
\rCC_G(1_{\mathcal D}, \Sigma) \ar[r]^-{\sigma} & \rCC_T(\Sigma, 1_\mathcal B ).
}
$$
\end{prop}
Dually, given a morphism of comonads
$$
\xymatrix{
(\mathcal{B}, S) \ar[rr]^-{(\Sigma, \gamma)} & & (\mathcal{D}, C),
}
$$
the natural transformation $\gamma \colon \Sigma S \Rightarrow C\Sigma$ extends to a morphism of simplicial functors
$$
\xymatrix{
\rCC_S^*(\Sigma,1_\mathcal{B}) \ar[r]^-\gamma & \rCC_C^*(1_\mathcal{D}, \Sigma).
}
$$
Now consider a distributive law $\chi \colon TS \Rightarrow ST$ of comonads on $\mathcal{B}$. This is equivalent to saying that
$$
\xymatrix{
(\mathcal{B}, T) \ar[rr]^-{(S,~\chi)} & & (\mathcal{B}, T),
}$$
is an opmorphism of comonads, and
$$
\xymatrix{
(\mathcal{B}, S) \ar[rr]^-{(T,~\chi)} & & (\mathcal{B}, S)
}
$$
is a morphism of comonads. Therefore, by Lemma~\ref{superchiepsilon} and its dual version, $\chi$ extends in two ways to simplicial functors
$$
\xymatrix{
\rCC_T(1_\mathcal{B}, S) \ar[r]^-\chi & \rCC_T(S, 1_\mathcal{B}),} \qquad
\xymatrix{\rCC_S^*(T, 1_\mathcal{B}) \ar[r]^-\chi & \rCC_S^*(1_\mathcal{B}, T).}
$$
Some abuse of notation occurs here, as we implicitly use the symbol $\chi^n$ to denote natural transformations $T^nS \Rightarrow ST^n$ and $TS^n \Rightarrow S^n T$.
\begin{rem}\label{superchimorphism}
A proof similar to that for Lemma~\ref{superchiepsilon} shows that for all $n \ge 0$,
$$
\xymatrix{
(\mathcal{B}, T) \ar[rr]^-{(S^n,~\chi^n)} & & (\mathcal{B}, T),
}$$
is an opmorphism of comonads, and
$$
\xymatrix{
(\mathcal{B}, S) \ar[rr]^-{(T^n,~\chi^n)} & & (\mathcal{B}, S)
}
$$ is a morphism of comonads.
\end{rem}
Now, let $(M, \mathcal{Y}, \rho)$ be a $\chi$-coalgebra, and let $(N, \mathcal{Z}, \lambda)$
be a $\chi$-opcoalgebra. We define
natural transformations
$$
t_T \colon
\rCC_T(N,M)_n \Rightarrow \rCC_T(N,M)_n,
\quad t_S \colon
\rCC_S^*(N,M)_n \Rightarrow \rCC_S^*(N,M)_n
$$
by the diagrams
$$
\xymatrix@C=3.5em{ NT^{n} SM
\ar[r]^-{N \chi^n M} &
N S T^n M
\ar[d]^-{\lambda T^n M}
\\
N T^{n+1} M
\ar[u]^-{ N T^n \rho}
\ar@{.>}[r]_-{t_T} & N T^{n+1} M }
\quad\quad\quad
\xymatrix@C=3.5em{ N T S^n M
\ar[r]^-{ N \chi^n M} & N S^n T M
\ar[d]^-{ N S^n \rho}
\\ N S^{n+1} M \ar[u]^-{\lambda S^n M}
\ar@{.>}[r]_-{t_S} & N S^{n+1} M }
$$
\begin{thm}[B\"ohm and {\c S}tefan]\label{dup}
The simplicial functors
$
\rCC_T( N, M)
$
and
$
\rCC^*_S( N, M)
$
become duplicial functors with duplicial
operators given by $t_T$ and $t_S$ respectively.
\end{thm}
\begin{proof}
The first operator being duplicial is exactly the case considered
in \cite{MR2415479}, and the second follows from a slight
modification of their proof. \end{proof}
At the end of Section~\ref{decalage} in the next chapter, we give a new way of deriving the duplicial functor $\rCC_T(N,M)$ of Theorem~\ref{dup}, using ideas from Hochschild (co)homology. Following this, we also show that the theorem can in fact be deduced directly from the twisting procedure given in Section~\ref{twistcoeff}.
\subsection{Cyclicity}
\label{evidenziatore2}
For each $n \ge 0$, we define a morphism
$\rho_n \colon T^{n+1} M \Rightarrow S^{n+1} M$
in the following way. For
each $0 \le i \le n$, let $\rho_{i,n}$ denote the morphism
$$
\xymatrix{ S^i T^{n-i + 1} M \ar[rr]^-{ S^i
T^{n-i} \rho} && S^i T^{n-i} S M \ar[rr]^-{ S^i
\chi^{n-i} M} & & S^{i+1} T^{n-i} M.}
$$
Then set $\rho_n$ to be the vertical composite
$$
\rho_n := \rho_{n,n} \rho_{n-1,n} \cdots \rho_{1,n} \rho_{0,n}.
$$
As above, the morphisms $\rho_{n+1} \colon T^{n+2}M \Rightarrow S^{n+2}M$ (for $n \ge 1$) can be defined recursively by either of the composites
\begin{align*}
\xymatrix{
T^{n+2}M \ar[rr]^-{T^{n+1} \rho} && T^{n+1} SM \ar[rr]^-{\chi^{n+1}M} && ST^{n+1}M \ar[rr]^-{S \rho_n} && S^{n+2}M
} \\
\xymatrix{
T^{n+2}M \ar[rr]^-{T\rho_n} && T S^{n+1} M \ar[rr]^-{\chi^{n+1}M} && S^{n+1}TM \ar[rr]^-{S^{n+1} \rho} && S^{n+2}M
}
\end{align*}
Completely symmetrically, we have morphisms $\lambda_{i,n} \colon NT^{n-i} S^{i+1} \Rightarrow NT^{n-i + 1} S^i$ given by
$$
\xymatrix{
N T^{n-i} S^{i+1} \ar[rr]^-{N \chi^{n-i} S^i} && NST^{n-i}S^i \ar[rr]^-{\lambda T^{n-i} S^i} && NT^{n- i + 1}
}
$$
which define a morphism $\lambda_n \colon NS^{n+1} \Rightarrow NT^{n+1}$ given by the composite
$$
\lambda_n := \lambda_{0,n} \lambda_{1,n} \cdots \lambda_{n-1, n} \lambda_{n,n}.
$$
\begin{lem}\label{superrhoepsilon}
For all $n \ge 0$ and for $0 \le i \le n + 1$, the diagram
\begin{equation}\label{superrho}
\begin{array}{cc}
\xymatrix{
T^{n+2}M \ar[rr]^-{\rho_{n+1}} \ar[d]_-{T^{n + 1 - i}\epsilon T^{i} M} && S^{n+2} M \ar[d]^-{S^{i}\epsilon S^{n+1-i }M} \\
T^{n+1}M \ar[rr]_-{\rho_n} && S^{n+1} M
}
\end{array}
\end{equation}
commutes.
\begin{proof}
Again, we proceed by induction. Suppose that $n = 0$ and $i = 1$. Naturality of $\epsilon$ and compatibility of $\rho$ and $\epsilon$ tells us that each inner square of the diagram
$$
\xymatrix{
TTM \ar[d]_-{\epsilon TM} \ar[rr]^-{T\rho} && TSM\ar[d]^-{\epsilon SM} \ar[rr]^-{\chi M} && STM \ar[d]^-{S\epsilon M} \ar[rr]^-{S \rho M} && SSM \ar[d]^-{S \epsilon M} \\
TM \ar[rr]_-{\rho} && SM \ar@{=}[rr] && SM \ar@{=}[rr] && SM
}
$$
commutes, and thus so does the outer diagram. Similarly, when $i = 0$, we have a commutative diagram
$$
\xymatrix{
TTM \ar[d]_-{T \epsilon M} \ar[rr]^-{T\rho} && TSM\ar[d]^-{T\epsilon M} \ar[rr]^-{\chi M} && STM \ar[d]^-{ \epsilon TM} \ar[rr]^-{S \rho M} && SSM \ar[d]^-{ \epsilon SM} \\
TM \ar@{=}[rr] && TM \ar@{=}[rr] && TM \ar[rr]_-\rho && SM
}
$$
proving that diagram~\ref{superrho} commutes in the base case.
Now suppose that the statement holds for $n-1$. Let $i$ be such that $0 < i \le n+1$, and write $j = i - 1$, so that $0 \le j \le n$. Consider the diagram
$$
\xymatrix{
T^{n+2}M \ar[rr]^-{T^{n+1} \rho} \ar[d]_-{T^{n-j} \epsilon T^{j+1} M} && T^{n+1} SM \ar[d]_-{T^{n-j} \epsilon T^j SM} \ar[rr]^-{\chi^{n+1} M} && ST^{n+1}M \ar[d]^-{ST^{n-j} \epsilon T^j M} \ar[rr]^-{S \rho_n} && S^{n+2}M \ar[d]^-{S S^j \epsilon S^{n-j} M} \\
T^{n+1}M \ar[rr]_-{T^n \rho} && T^nSM \ar[rr]_-{\chi^n M} && ST^nM \ar[rr]_-{S \rho_{n-1}} && S^{n+1}M
}
$$
The right-hand square commutes by inductive hypothesis, the middle square commutes by Lemma~\ref{superchiepsilon}, and the left-hand square commutes by naturality of $\epsilon$. The case for $i = 0$ is similarly proved, using compatibility of $\chi^n$ with $\epsilon \colon S \Rightarrow 1$ from Remark~\ref{superchimorphism}.
\end{proof}
\end{lem}
There is a similar statement regarding the coproducts $\delta$. Thus, we have proved the following:
\begin{prop}\label{simplicialrho}
The morphisms $\rho_n$ define a morphism of simplicial functors
$$
\xymatrix{
\rBB(T,M) \ar[r]^-{\rho} & \rBB^*(S,M).
}$$
\end{prop}
\begin{lem}\label{cycliclem}
For all $i,j,k$ with $i \ge 1$, $k \ge i$ and $j \ge 0$, the diagram
\begin{equation}\label{cycliclemma}
\begin{array}{c}
\xymatrix{
NT^jS^i T^{k-i+1} M \ar[dd]_-{\lambda_{i-1, i+j - 1} T^{k-i+1}M} \ar[rrr]^-{NT^j \rho_{i, k} } &&& NT^jS^{i+1}T^{k-i} M \ar[dd]^-{\lambda_{i, i+j} T^{k-i}M} \\
\\
NT^{j+1} S^{i-1} T^{k-i+1} M\ar[rrr]_-{NT^{j+1} \rho_{i-1, k-1}} & && NT^{j+1} S^i T^{k-i} M
}
\end{array}
\end{equation}
commutes.
\end{lem}
\begin{proof}
Consider the diagram
$$
\xymatrix@C=2.31em{
NT^jS^iT^{k-i + 1} M \ar[dd]^-{N \chi^jS^{i-1} T^{k-i+1}} \ar[rr]^-{NT^j S^i T^{k-i}\rho} & & NT^j S^i T^{k-i} SM \ar[dd]^-{N\chi^j S^{i-1}T^{k-i}M} \ar[rr]^-{NT^jS^i \chi^{k-i} M} && NT^j S^{i+1}T^{k-i}M \ar[dd]_-{N\chi^j S^i T^{k-i}M}\\
\\
NST^j S^{i-1} T^{k-i + 1} M \ar[dd]^-{\lambda T^j S^{i-1} T^{k-i+1} M} \ar[rr]_-{NST^j S^{i-1}T^{k-i} \rho} && NST^j S^{i-1}T^{k-i} SM \ar[rr]_-{NST^jS^{i-1}\chi^{k-i} M} \ar[dd]_-{\lambda T^{j}S^{i-1} T^{k-i}SM} && NST^j S^i T^{k-i} M \ar[dd]_-{\lambda T^jS^iT^{k-i}M} \\
\\
NT^{j+1}S^{i-1}T^{k-i+1} M \ar[rr]_-{NT^{j+1}S^{i-1}T^{k-i}\rho} && NT^{j+1}S^{i-1}T^{k-i}SM \ar[rr]_-{NT^{j+1}S^{i-1}\chi^{k-i} M} && NT^{j+1}S^{i} T^{k-i} M
}
$$
The upper squares commute by naturality of $\chi^j$, and the lower squares commute by naturality of $\lambda$. Therefore, the outer square, which is the same as the square~\ref{cycliclemma}, commutes.
\end{proof}
Let $R_n$ denote the morphism $N \rho_n \colon NT^{n+1}M \Rightarrow N S^{n+1}M$, and let $L_n$ denote the morphism $\lambda_n M \colon N S^{n+1} M \Rightarrow
N T^{n+1} M$.
\begin{thm}\label{cyc}
The above construction defines
two morphisms
\begin{align*}
\xymatrix{ \rCC_T(N,M) \ar[r]^-{R } &
\rCC_S^*(N,M),} \qquad
\xymatrix{\rCC_S^*(N,M)
\ar[r]^-{L } & \rCC_T(N,M)}
\end{align*}
of duplicial functors. Furthermore, $L
R = 1$ if and only if $\rCC_T
(N,M)$ is cyclic, and $R L = 1$
if and only if $ \rCC_S^*(N,M)$ is cyclic.
\end{thm}
\begin{proof}
We show that $R$ is a morphism of duplicial functors, and remark that the proof for $L$ is similar. It is clear that $R$ is a morphism of simplicial objects, so it remains to show that it commutes with the duplicial operators. Let $n \ge 0$ and consider the diagram
$$
\xymatrix@C=0.799em{
& & & NS^{n+1} M \ar[dr]^-{\ \lambda S^n M} \ar@/^6.5ex/[dddrrr]^-{t_S }& & & \\
& & NST^n M \ar[ddr]_-{\lambda T^n M}\ar[ur]^-{NS\rho_{n-1}}& & NTS^n M \ar[dr]^-{N \chi^n M} & & \\
& NT^n SM \ar[ur]^-{N \chi^n M} & & & & NS^n TM \ar[dr]_-{N S^n \rho} & \\
\ar@/^6.5ex/[uuurrr]^-{R_n }NT^{n+1}M \ar[ur]_-{NT^n \rho} \ar[rrr]_-{t_T} & & & NT^{n+1} M \ar[uur]_-{NT\rho_{n-1}} \ar[rrr]_-{R_n} & & & NS^{n+1}M
}
$$
The left-hand triangle commutes by definition of $t_T$, and the right-hand semicircle commutes by definition of $t_S$. The right-hand triangle and left-hand semicircle commute by the recursive definitions of $\rho_n$. The middle kite commutes by naturality of $\lambda$, and thus the outer diagram commutes, proving that $R$ is a morphism of duplicial functors.
We finish the proof of the theorem by showing that $(LR)_n = t_T^{n+1}$, which is the identity if and only if $\operatorname{C}_T(N,M)$ is cyclic.
Consider the following two triangles:
$$
\xymatrix@=2.9em{
\smxy{NT^{n+1}M} \ar[r]^-{\smxy{N\rho_{0, n}}} \ar[dr]_-{\smxy{t_T}} & \smxy{NST^n M} \ar[d]^-{\smxy{\lambda_{0,0} T^n M}}\ar[r]^-{\smxy{N\rho_{1,n}}} & \smxy{NS^2T^{n-1}M} \ar[d]^-{\smxy{\lambda_{1,1} T^{n-1}M}}\ar@{.}[r]& \smxy{NS^nTM} \ar[r]^-{\smxy{N\rho_{n,n}}} \ar[d]^-{\smxy{\lambda_{n-1,n-1}TM}} & \smxy{NS^{n+1}M} \ar[d]^-{\smxy{\lambda_{n,n}M}} \\
& \smxy{NT^{n+1}M} \ar@{.}[dr
\ar[r]_-{\smxy{NT\rho_{0, n-1}}} & \smxy{NTST^{n-1}M} \ar@{.}[r] \ar@{.}[d] & \smxy{NTS^{n-1} TM} \ar@{.}[d] \ar[r]_-{\smxy{NT\rho_{n-1,n-1}}} & \ar@{.}[d] \smxy{NTS^n M}\\
& & \smxy{NT^{n+1}M} \ar@{.}[r] \ar@{.}[dr]& \smxy{NT^{n-1} STM} \ar[d]_-{\smxy{\lambda_{0, n-1}M}}\ar[r]_-{\smxy{NT^{n-1} \rho_{1,1}}} & \smxy{NT^{n-1}S^2M} \ar[d]^-{\smxy{\lambda_{1,n} M}} & \\
NT^{n+1}M \ar[drr]_-{t_T} \ar[rr]^-{NT^{n-k} \rho_{0,k}}& & {NT^{n-k}ST^k M} \ar[d]^-{\lambda_{0,n-k}T^kM} & \smxy{NT^{n+1}M} \ar[r]^-{\smxy{NT^n \rho_{0,0}}}\ar[dr]_-{\smxy{t_T}}& \smxy{NT^n SM} \ar[d]^-{\smxy{\lambda_{0,n}M}} \\
& & NT^{n+1}M & & \smxy{NT^{n+1}M}
}
$$
Applying Lemma~\ref{cycliclem} a total of $\frac{n}{2}(n+1)$ times shows that each square of the right-hand triangle commutes. Each triangle along the diagonal is of the form of the lower-left triangle, for some $k$ such that $0 \le k \le n$. This lower-left triangle can also be written as the outer triangle of
$$
\xymatrix@=3em{
NT^nM \ar[ddrrr]_-{t_T}\ar[r]^-{NT^n \rho} & NT^n SM \ar[rrd]_-{N\chi^n M\ }\ar[rr]^-{NT^{n-k} \chi^k M} && NT^{n-k} ST^k M \ar[d]^-{N\chi^{n-k}T^kM} \\ && & NST^n M \ar[d]^-{\lambda T^n M} \\
& & & NT^{n+1}M
}
$$
which commutes since the inner shapes do. Therefore the very large outer triangle above commutes, which means exactly that $(LR)_n = t_T^{n+1}$. A similar argument shows that $(RL)_n = t_S^{n+1}$, completing the proof.
\end{proof}
\subsection{Homologically trivial $ \chi $-coalgebras}
Let $\mathcal{Z}$ be an abelian category.
\begin{defn}
Let $X$ be a non-negative chain complex in $\mathcal{Z}$. We say that $X$ is \emph{contractible} if $X$ is homotopy-equivalent to the zero chain complex.
\end{defn}
A complex $X$ is contractible if and only if the identity chain morphism $1\colon X \to X$ is homotopic to $0$. Explicitly this means that there is a contracting homotopy, i.e.\ a morphism $h \colon X \to X$ of degree 1, such that $hb + bh = 1$, where $b$ denotes the differential of $X$. This implies that the homology objects $\operatorname{H}^n(X)$ are trivial for $n > 0$.
As we had announced above, $ \chi $-coalgebras as in
Proposition~\ref{triv}
lead to contractible chain complexes:
\begin{prop}
\label{trivcontract}
Let $\chi \colon
TS \Rightarrow ST$ be a comonad distributive
law on a category $\mathcal{B}$, and let $(M, \mathcal{Y},
\rho)$ and $(N, \mathcal{Z}, \lambda)$ be a $\chi$-coalgebra and $\chi$-opcoalgebra, respectively. If either of
$(N,Y, \lambda),(M, \mathcal{Y}, \rho)$ arises as in
Proposition~\ref{triv},
then the chain complexes associated to both
$\rCC_T (N, M)$ and $\rCC_S^* (N, M)$
under the Dold-Kan correspondence are contractible.
\end{prop}
\begin{proof}
Let $\nabla \colon N \Rightarrow NT$ be a $T$-opcoalgebra
structure on $N$.
The morphisms
$$
\xymatrix@=4em{h_n \colon NS^{n+1} M
\ar[r]^-{\nabla S^{n+1} M}
&NTS^{n+1}M
\ar[r]^-{N\chi^{n+1}M} &NS^{n+1} T M \ar[r]^-{NS^{n+1}
\rho} &NS^{n+2} M }
$$
provide a contracting homotopy
for the complex associated to $\rCC_S^*
(N,M)$. To see this, consider the following diagram, where $n >0$ and $i$ is such that $0 \le i \le n$:
$$
\xymatrix@C=3.5em{
NS^{n+1}M \ar[r]^-{\nabla S^{n+1}M} \ar[d]_-{NS^{n-i}\epsilon S^i M}& NTS^{n+1}M \ar[d]_-{NTS^{n-i}\epsilon S^i M}\ar[r]^-{N\chi^{n+1}M} & NS^{n+1}TM \ar[d]^-{NS^{n-i}\epsilon S^i T M} \ar[r]^-{NS^{n-1}\rho M} & \ar[d]^-{NS^{n-i}\epsilon S^{i+1}M}NS^{n+2}M \\
NS^nM \ar[r]_-{\nabla S^n M} & NTS^n M \ar[r]_-{N\chi^n M} & NS^n TM \ar[r]_-{NS^n \rho M} & NS^{n+1}M
}
$$
The left-hand square commutes by naturality of $\nabla$, the right-hand square commutes by naturality of $\epsilon$, and the middle square commutes by the version of Lemma~\ref{superchiepsilon} for $\chi^n \colon TS^n \Rightarrow S^n T$. Therefore, the outer rectangle commutes, so we have the relation $h_{n-1}d_i = d_{i+1}h_n$, where $d$ denotes the face morphisms of $\rCC_S^*(N,M)$. Similarly, we have $d_0 h_n = 1$. Therefore, we have
\begin{align*}
bh_n &= \sum_{i=0}^{n+1} (-1)^i d_i h_n \\
&= d_0 h_n + \sum_{i=0}^n (-1)^{i+1} d_{i+1} h_n \\
&= 1 + \sum_{i=0}^n (-1)^{i+1} h_{n-1}d_i \\
&= 1 - h_{n-1}b,
\end{align*}
so $hb + bh = 1$, as required.
A similar proof shows that the morphisms $\nabla T^n M \colon NT^{n+1}M \Rightarrow
NT^{n+2}M$ provide a contracting homotopy for the
complex associated to $\rCC_T (N, M)$. The case when $M$ is an $S$-coalgebra is proved similarly.
\end{proof}
\subsection{Twisting by $1$-cells}
\label{twistsec}
Applying the twisting procedure described in
Section~\ref{twistcoeff},
a $1$-cell
$$
\xymatrix{
(\mathcal{B}, \chi, T, S) \ar[rr]^-{(\Sigma, \sigma, \gamma)} & & (\mathcal{D}, \tau, G, C)
}
$$
in the $2$-category $\mathbf{Dist}$, together with a $\chi$-coalgebra $M$ and a $\tau$-opcoalgebra $N$, give rise to morphisms
between duplicial functors of the form considered
above: Theorems~\ref{dup} and~\ref{twist} yield
duplicial structures on the simplicial functors $$
\rCC_T(N\Sigma,M),\quad \rCC_S^*(N\Sigma,M),
\quad \rCC_G(N,\Sigma M),
\quad \rCC_C^*(N,\Sigma M),
$$
and from Proposition~\ref{cyc} we obtain morphisms
\begin{align*}
&\xymatrix{ \rCC_T(N\Sigma,M)
\ar[r]^{R^\chi} &
\rCC_S^*(N\Sigma,M), } &
\xymatrix{ \rCC_S^*(N\Sigma,M)
\ar[r]^{L^\chi} & \rCC_T(N\Sigma,M)} \\
&\xymatrix{ \rCC_G(N,\Sigma M) \ar[r]^-{R^\tau}
& \rCC_C^*(N,\Sigma M),} & \xymatrix{
\rCC_C^*(N,\Sigma M) \ar[r]^-{L^\tau} &
\rCC_G(N,\Sigma M)}
\end{align*}
of duplicial
objects which determine the cyclicity of each functor.
We now prove that a generalised Yang-Baxter equation holds:
\begin{lem}\label{superyb}
For all $n > 0$, the diagram
$$
\xymatrix@C=3em{
\ar[d]_-{\sigma^n S}G^n\Sigma S\ar[r]^-{G^n \gamma} & G^n C\Sigma \ar[r]^-{\tau^n \Sigma} & CG^n \Sigma \ar[d]^-{C\sigma^n} \\
\Sigma T^n S \ar[r]_-{\Sigma \chi^n} & \Sigma ST^n \ar[r]_-{\gamma T^n} & C\Sigma T^n
}
$$
commutes.
\end{lem}
\begin{proof}
When $n = 1$, this is the Yang-Baxter equation. If the diagram commutes for $n$, consider the diagram
$$
\xymatrix@C=3em{
G^{n+1} \ar[d]_-{G\sigma^n S} \Sigma S \ar[r]^-{G^{n+1}\gamma} & G^{n+1}C\Sigma \ar[r]^-{G\tau^n \Sigma} & GCG^n\Sigma \ar[r]^-{\tau G^n\Sigma} \ar[d]_-{GC\sigma^n} & CG^{n+1} \Sigma \ar[d]^-{CG \sigma^n} \\
G\Sigma T^n S \ar[d]_-{\sigma T^n S} \ar[r]^-{G\Sigma \chi^n} & \ar[d]_-{\sigma ST^n} G\Sigma ST^n \ar[r]^-{G\gamma T^n} & GC\Sigma T^n \ar[r]^-{\tau \Sigma T^n}& CG\Sigma T^n \ar[d]^-{C\sigma T^n} \\
\Sigma T^{n+1}S \ar[r]_-{\Sigma T\chi^n}& \Sigma TST^n \ar[r]_-{\Sigma \chi T^n}& \Sigma ST^{n+1} \ar[r]_-{\gamma T^{n+1}}& C\Sigma T^{n+1}
}
$$
The upper-right square commutes by naturality of $\tau$, the lower-left square commutes by naturality of $\sigma$, the lower-right rectangle commutes by the Yang-Baxter equation, and the top left-diagram commutes by inductive hypothesis. Therefore, the outer diagram commutes, proving the lemma.
\end{proof}
By applying the functors $M,N$ to the simplicial morphisms $\sigma,\gamma$ of Proposition~\ref{supersigma} and its dual, we obtain two simplicial morphisms
$$
\xymatrix{ \rCC_G(N,\Sigma M) \ar[rr]^-{N\sigma M} &&
\rCC_T(N\Sigma, M),} \quad \quad \quad
\xymatrix{ \rCC_S^*(N\Sigma, M) \ar[rr]^-{N\gamma M} &&
\rCC_C^*(N,\Sigma M).}
$$
\begin{prop}\label{rhocyclic}
The morphisms $N\sigma M$, $N\gamma M$ are morphisms of duplicial functors.
\end{prop}
\begin{proof}
We prove this only for $N\sigma M$. It suffices to show this commutes with the duplicial operators $t_T, t_G$. Recall, from Section~\ref{twistcoeff}, that the (op)coalgebra structure morphisms on $\Sigma M$ and $N \Sigma$ are respectively given by the composites
$$
\xymatrix@R=1em{
G\Sigma M \ar[r]^-{\sigma M} & \Sigma TM \ar[r]^-{\Sigma\rho} & \Sigma S M \ar[r]^-{\gamma M} & C \Sigma M,
\\
N \Sigma S \ar[r]^-{N \gamma} & N C \Sigma \ar[r]^-{\lambda \Sigma} & NG\Sigma \ar[r]^-{N \sigma} & N \Sigma T.
}
$$
Let $n \ge 0$. In the diagram
$$
\xymatrix@C=2.74em{
\smmxy{NG^n\Sigma TM} \ar[d]^-{N \sigma^n TM} \ar[r]^-{NG^n \Sigma \rho} & \smmxy{NG^n\Sigma SM} \ar[d]^-{N\sigma^n SM }\ar[r]^-{NG^n \gamma M} & \smmxy{NG^nC\Sigma M} \ar[r]^-{N\tau^n \Sigma M} & \smmxy{NCG^n\Sigma M} \ar[d]_-{NC\sigma^n M} \ar[r]^-{\lambda G^n \Sigma M} & \smmxy{NG^{n+1}\Sigma M} \ar[d]_-{NG\sigma^n M}\\
\smmxy{N\Sigma T^{n+1} M} \ar[r]_-{N\Sigma T^n \rho} & \smmxy{N\Sigma T^n SM} \ar[r]_-{N\Sigma \chi^n M} & \smmxy{N\Sigma ST^n M} \ar[r]_-{N\gamma T^n M} & \smmxy{NC\Sigma T^n M} \ar[r]_-{\lambda \Sigma T^n M} & \smmxy{NG\Sigma T^n M}
}
$$
the middle square commutes by Lemma~\ref{superyb}, the right-hand square commutes by naturality of $\lambda$, and the left-hand square commutes by naturality of $\sigma^n$. The outer diagram commutes therefore, and after pre-composing with $NG^n \sigma M$ and post-composing with $N\sigma T^n M$, we have the result.
\end{proof}
\begin{exa}
Let
$$
\xymatrix{
\mathcal{A} \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{B}
}
$$
be an adjunction, generating a monad $B = UF$ on $\mathcal{A}$ and a comonad $T = FU$ on $\mathcal{B}$. Let $$
\xymatrix{\mathcal{B} \ar[r]^ U \ar[d]_ S & \ar@{}[dl]^(.25){}="a"^(.75){}="b" \ar@{=>}_-{\Omega} "a";"b" \mathcal{A} \ar[d]^ C\\
\mathcal{B} \ar[r]_U & \mathcal{A}}
$$
be a square where $(U, \Omega) \colon (\mathcal{B}, S) \to (\mathcal{A}, C)$ is an iso-opmorphism of comonads, giving rise to a mixed distributive law $\theta$ and a comonad distributive law $\chi$ by Corollary~\ref{arisec}. By Theorem~\ref{distcomp1cell}, the comparison functor $\mathcal{B} \to \mathcal{A}^B$ gives rise to a 1-cell
$$\xymatrix{(\mathcal{B}, \chi, T, S) \ar[rrr]^-{(U^{U\epsilon}, 1, \tilde\Omega^{-1})} & & & (\mathcal{A}^B, \tilde\theta, \tilde B, C^\theta)
}
$$
in $\mathbf{Dist}$. Given any $\chi$-coalgebra $M$ and $\tilde\theta$-opcoalgebra $N$, by Proposition~\ref{rhocyclic} we have isomorphisms of duplicial functors
$$
C_{\tilde B}\left(N, U^{U\epsilon} M\right) \cong C_T\left(NU^{U\epsilon}, M\right), \qquad
C_S^*\left(NU^{U\epsilon}, M\right) \cong C_{C^\theta}^* \left(N, U^{U\epsilon} M\right).
$$
Since $\sigma = 1$ in this example, the left-hand isomorphism is actually an equality.
\end{exa}
\subsection{Cyclicity and the reflection equation}
We conclude this chapter by observing that there is an interesting connection between the cyclicity of duplicial functors of the above form, and the reflection equation of physics. Let $\chi \colon TS \Rightarrow ST$ be a distributive law of comonads, and let $M$ and $N$ be a $\chi$-coalgebra and a $\chi$-opcoalgebra respectively. Let us represent the natural transformations diagrammatically as
$$
\chi = \begin{array}{c}
\begin{tikzpicture}
\draw [thick] (-1, 1) node[anchor=south]{$T$} -- (0, -1) node[anchor=north]{$T$};
\draw [thick] (0, 1) node[anchor=south]{$S$} -- (-1,- 1) node[anchor=north]{$S$};
\end{tikzpicture}
\end{array},\qquad \qquad
\rho =
\begin{array}{c}
\begin{tikzpicture}
\draw [thick] (-1, 1) node[anchor=south]{$T$} -- (0, -0);
\draw [thick](0, 1) node[anchor=south]{$M$} -- (0,- 1) node[anchor=north]{$M$};
\draw[thick] (0,0) -- (-1, -1) node[anchor=north]{$S$};
\end{tikzpicture}
\end{array},\qquad \qquad
\lambda =
\begin{array}{c}
\begin{tikzpicture}
\draw [thick](0,1) node[anchor=south]{$N$} -- (0, -1) node[anchor=north]{$N$};
\draw [thick](1,1) node[anchor=south]{$S$} -- (0,0);
\draw [thick](0,0) -- (1,-1) node[anchor=north]{$T$};
\end{tikzpicture}
\end{array}
$$
Using graphical calculus and reading from top to bottom, in degree 1, the two morphisms $t_T^2, t_S^2$ look like
$$
\begin{array}{c}
\begin{tikzpicture}
\draw [thick](0, 5) node[anchor=south]{$N$} -- (0, 0) node[anchor=north]{$N$};
\draw [thick](1,5) node[anchor=south]{$T$} -- (3,3) -- (0,1) --(1,0) node[anchor=north]{$T$};
\draw [thick](2,5) node[anchor=south]{$T$} -- (3,4) --(0,2) -- (2,0) node[anchor=north]{$T$};
\draw [thick](3,5) node[anchor=south]{$M$} -- (3,0) node[anchor=north]{$M$};
\end{tikzpicture}
\end{array},
\qquad
\begin{array}{c}
\begin{tikzpicture}
\draw [thick](0, 5) node[anchor=south]{$N$} -- (0, 0) node[anchor=north]{$N$};
\draw[thick](1,5) node[anchor=south]{$S$} -- (0,4) -- (3,2) --(1,0) node[anchor=north]{$S$};
\draw[thick](2,5) node[anchor=south]{$S$} -- (0,3) -- (3,1) --(2,0) node[anchor=north]{$S$};
\draw[thick](3,5) node[anchor=south]{$M$} -- (3,0) node[anchor=north]{$M$};
\end{tikzpicture}
\end{array}
$$
We can view $N,M$ as walls and the two inner strands may be viewed as the trajectories of two distinct particles bouncing between them. Now, instead of viewing the particles as bouncing off the right-hand walls, let us ignore the wall and see the particle's trajectory as a straight line. In other words, we remove the right-hand wall, straighten out each of the right-hand kinks, but \emph{preserve} the points where the lines cross. Heuristically, as long as a particle $T$ is furthest to the right, it may undergo a `state-change' and turn into the particle $S$ which allows us to use the distributive law $\chi$ to cross the particles. Ignoring the labels, if we redraw the above diagrams with this viewpoint, we have:
$$
\begin{array}{c}
\begin{tikzpicture}
\useasboundingbox (0,4) rectangle (1,0);
\clip (-1, 4) rectangle (1,0);
\draw [thick] (0,4) -- (0,0);
\draw [thick] (2, 4) -- (0,3);
\draw [red] [thick] (1,4) -- (0,2) -- (1,0);
\draw [thick] (0,3) -- (2,2);
\end{tikzpicture}
\end{array}, \qquad
\begin{array}{c}
\begin{tikzpicture}
\useasboundingbox (0,5) rectangle (1,1);
\clip (-1, 5) rectangle (1,1);
\draw [thick] (0,5) -- (0,1);
\draw [thick] (2, 3) -- (0,2);
\draw [red] [thick] (1,5) -- (0,3) -- (1,1);
\draw [thick] (0,2) -- (2,1);
\end{tikzpicture}
\end{array}
$$
Each diagram is one side of the reflection equation, which describes the trajectories of two particles with different velocities bouncing off a wall in different orders. We use the colour red to distinguish between the particles in the physical interpretation.
We may construct similar pairs of diagrams for $n > 1$, as well as reflection equations for more than two particles. For example, the reflection equation for $n = 2$ (i.e.\ for 3 particles) is
$$
\begin{array}{c}
\begin{tikzpicture}
\useasboundingbox (0,6) rectangle (2,0);
\clip (-1,6) rectangle (2,-1);
\draw [thick] (0,6) -- (0,0);
\draw [thick] (3,6) -- (0,5);
\draw [thick] [red] (2,6) -- (0,4);
\draw [thick] [blue] (1,6) -- (0,3) -- (1, 0);
\draw [thick] [red] (0,4) -- (3,1);
\draw[thick] (0,5) -- (3,4);
\end{tikzpicture}
\end{array} \quad = \quad
\begin{array}{c}
\begin{tikzpicture}
\useasboundingbox (0,6) rectangle (2,0);
\clip(-1, 6) rectangle (2, -1);
\draw [thick] (0,6) -- (0,0);
\draw[thick] [blue] (1,6) -- (0,3);
\draw[thick] [red] (3,5) -- (0,2);
\draw [thick] (3, 2) -- (0, 1);
\draw[thick] (0,1) -- (3,0);
\draw[thick] [red] (0, 2) -- (3,-1);
\draw[thick] [blue] (0,3) -- (1,0);
\end{tikzpicture}
\end{array}
$$Therefore, we immediately obtain another characterisation of cyclicity.
\begin{prop}
The duplicial functor $\rCC_T(N,M)$ is cyclic if and only if the appropriate side of the reflection equation in each degree is equal to the identity.
\end{prop}
Of course, there is a similar statement for the duplicial functor $C_S^*(N,M)$.
\chapter{Hochschild viewpoint of dupliciality}\label{AUSTRALIA}
The purpose of this chapter is to explain that Hochschild homology and cohomology of algebras can be imported to the world of 2-categories, in turn giving an insight into the nature of duplicial structure. We begin by recounting the classical case (Section~\ref{hochclassic}). In Sections~\ref{HOCHLAX} and~\ref{duplapp} we construct an upgraded version of Hochschild (co)homology and see that we recover the duplicial object $C_T(N,M)$ of Chapter~\ref{CYCLIC}. We also characterise duplicial structure on the nerve of a category $\mathcal{C}$ in terms of left adjoint functors from groupoids into $\mathcal{C}$.
The work in these latter sections is original and is based on that carried out in~\cite{woohoo}.
\section{The classical case}\label{hochclassic}
Let $k$ be a commutative ring and let $A$ be a (unital, associative) $k$-algebra. We denote the monoidal product $\otimes_k$ of $k\mbox{-}\mathbf{Mod}$ by an unadorned tensor product $\otimes$. Let $A^*$ denote the opposite algebra to $A$, and let ${A^\mathrm{e}}$ denote the enveloping algebra $A \otimes A^*$. If we denote the category of $(A,A)$-bimodules (with symmetric action of $k$) by $A\mbox{-}\mathbf{Mod}\mbox{-}A$, there are isomorphisms
$$
A\mbox{-}\mathbf{Mod}\mbox{-}A \cong {A^\mathrm{e}}\mbox{-}\mathbf{Mod} \cong \mathbf{Mod}\mbox{-}{A^\mathrm{e}}.
$$
and we therefore identify all of these.
Consider the endofunctor $B = {-}\otimes A$ on $A\mbox{-}\mathbf{Mod}$ which takes a left $A$-module $X$ to the module $X \otimes A$ with left $A$-action given by $a\cdot(x \otimes a') := ax \otimes a'$.
The natural maps
\begin{align*}
&\xymatrix@R=1em{X \otimes A \otimes A \ar[r]^-{\mu} & X \otimes A \\ x\otimes a\otimes a' \ar@{|->}[r] & x\otimes aa'} & &\xymatrix@R=1em{ X \ar[r]^-{\eta} & X \otimes A \\ x \ar@{|->}[r] & x\otimes 1}
\end{align*}
turn the functor $B = {-}\otimes A$ into a monad on $A\mbox{-}\mathbf{Mod}$, which lifts to a comonad $\tilde B$ on the Eilenberg-Moore category $A\mbox{-}\mathbf{Mod}^B$. In fact, there is another isomorphism
$$
A\mbox{-}\mathbf{Mod}^B \cong \mathbf{Mod}\mbox{-}{A^\mathrm{e}}
$$
so we view $\tilde B$ as a comonad on the category of bimodules, defined by the morphisms
\begin{align*}
&\xymatrix@R=1em{M \otimes A \ar[r]^-{\delta} & M \otimes A \otimes A \\ m\otimes a \ar@{|->}[r] & m\otimes 1 \otimes a}
& &\xymatrix@R=1em{ M \otimes A \ar[r]^-{\epsilon} & M \\ m \otimes a \ar@{|->}[r] & ma.}
\end{align*}
So, for every bimodule $M$, we have that $\tilde B(M) = M \otimes A$ becomes a bimodule with actions given by
$$
a\cdot(m \otimes a') \cdot a'' := am \otimes a'a''.
$$
\subsection{Hochschild homology and cohomology}\label{hochsub}
Viewing $A$ as a left ${A^\mathrm{e}}$-module, there is a functor $${-} \otimes_{{A^\mathrm{e}}}A \colon \mathbf{Mod}\mbox{-}{A^\mathrm{e}} \to k\mbox{-}\mathbf{Mod}.$$
Also, let $M$ be any $(A,A)$-bimodule, viewed as a right ${A^\mathrm{e}}$-module.
\begin{defn}\label{hochdef}
The \emph{Hochschild homology of $A$ with coefficients in $M$}, denoted by $\mathrm H_\bullet(A,M)$, is the graded $k$-module given by the homology of the chain complex associated to the simplicial $k$-module $\rCC_{\tilde B}({-}\otimes_{{A^\mathrm{e}}} A, M)$.
\end{defn}
The above simplicial object is seen again in Chapter~\ref{EXAMPLES}. However, there is a less complicated simplicial object which may be used to define Hochschild homology. For every left $A$-module $P$, there are $k$-module isomorphisms
$$
\xymatrix@R=1em{
(P \otimes A)\otimes_{{A^\mathrm{e}}} A \ar[r]^-{\cong} & P \\
(p \otimes a) \otimes_{{A^\mathrm{e}}} a' \ar@{|->}[r] & aa'p,}
$$ and so we have isomorphisms for $n \ge 0$
$$
\rCC_{\tilde B}(- \otimes_{{A^\mathrm{e}}} A, M)_n = \left(M \otimes A^{\otimes(n+1)}\right) \otimes_{{A^\mathrm{e}}} A \cong M \otimes A^{\otimes n}.
$$
Thus, we get a simplicial $k$-module $C_\bullet(A, M)$ defined by $C_n(A, M) = M \otimes A^{\otimes n}$, with face maps and degeneracy maps
$$d_i \colon C_n(A,M) \to C_{n-1}(A,M), \qquad s_i \colon C_{n}(A,M) \to C_{n+1}(A,M)$$ given by
\begin{align*}
d_i(m \otimes a_1 \otimes \cdots \otimes a_n) &=
\begin{cases}
a_n m \otimes a_1 \otimes \cdots \otimes a_{n-1} &\mbox{ if } i = 0 \\
m \otimes a_1 \otimes \cdots \otimes a_{n-i}a_{n-i+1} \otimes \cdots \otimes a_n & \mbox{ if } 1 \le i < n \\
ma_1 \otimes a_2 \otimes \cdots \otimes a_n & \mbox{ if } i = n
\end{cases} \\
s_i(m \otimes a_1 \otimes \cdots \otimes a_n) &=
m \otimes \cdots \otimes a_{n-i} \otimes 1 \otimes a_{n-i+1} \otimes \cdots \otimes a_n
\end{align*}
The homology of the chain complex associated to this complex is therefore isomorphic to $\mathrm H_\bullet(A, M)$. The zeroth Hochschild homology is given by
$$
\mathrm H_0(A, M) \cong {M}/\langle am - ma \rangle
$$
where $\langle am - ma \rangle$ denotes the submodule of $M$ generated by all expressions of the form $ma - am$ for $m \in M$, $a \in A$.
Consider the contravariant functor
$$
\operatorname{Hom}_{A^\mathrm{e}}({-}, M) \colon \mathbf{Mod}\mbox{-}{A^\mathrm{e}} \to k\mbox{-}\mathbf{Mod}
$$
which assigns to a bimodule $N$ the $k$-module of $(A,A)$-bimodule maps $N \to M$.
\begin{defn}
The \emph{Hochschild cohomology of $A$ with coefficients in $M$}, which we denote by $\mathrm H^\bullet(A,M)$, is the graded $k$-module given by the cohomology of the cochain complex associated to the cosimplicial $k$-module $\rCC_{\tilde B}(\operatorname{Hom}_{A^\mathrm{e}}({-}, M), A).$
\end{defn}
The zeroth Hochschild cohomology is given by the centre of $M$, explicitly:
$$
\mathrm H^0(A, M) \cong ZX = \{ m \in M \mid ma = am \mbox{ for all } m \in A\}.
$$
Note that both $\mathrm H_0 (A, {-} )$ and $\mathrm H^0(A, {-})$ define functors $A\mbox{-}\mathbf{Mod}\mbox{-}A \to k\mbox{-}\mathbf{Mod}$.
We do not delve into Hochschild cohomology further, as we do not discuss anything beyond this zeroth homology module.
\subsection{Universal properties}\label{hochunivprop}
For any $k$-module $X$, the set
$$
[M,X] := \operatorname{Hom}_k(M,X)
$$
becomes an $(A,A)$-bimodule with actions given by
$$
(a \cdot f \cdot b ) (m) = f(bma).
$$
We now present a universal coefficients theorem:
\begin{thm}\label{hochuniv}
There is an isomorphism of $k$-modules
$$
[\mathrm H_0(A, M) , X] \cong \mathrm H^0(A, [M, X] ),
$$
natural in $M$ and $X$.
\end{thm}
\begin{proof}
By the above remarks, to prove this is equivalent to proving that
$$
\left[{M}/{\langle am - ma \rangle}, X \right] \cong Z([M,X]).
$$
Given $f \in Z([M,X])$, we have that $a f = f a$ for all $a \in A$, which means exactly that $f(ma) = f(am)$ for all $a \in A, m \in M$. Therefore, $f$ descends to a map $\bar f$ on the quotient:
$$
\xymatrix{
M \ar@{->>}[r] \ar[dr]_-{f} & M / \langle am-ma\rangle \ar[d]^-{\bar f} \\
& X
}
$$
The assignment $f \mapsto \bar f$ defines the required isomorphism, whose inverse is given by composing with the quotient map. Naturality follows easily.
\end{proof}
\begin{thm}
We have an adjunction
$$
\xymatrix{
A\mbox{-}\mathbf{Mod}\mbox{-}A \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-{\mathrm H_0(A, {-})} & & \ar@/^0.5pc/[ll]^-{[A, {-}]} k\mbox{-}\mathbf{Mod}.
}
$$
\end{thm}
\begin{proof}
We have that $\mathrm H_0(A,{-}) \cong {-} \otimes_{A^\mathrm{e}} A$ which is left adjoint to $[A,{-}]$ by the tensor-hom adjunction (see e.g.~\cite[Theorem 2.75]{MR2455920}).
\end{proof}
Let us now apply the cohomology functor $\mathrm H^0(A, -)$ to the unit of the adjunction $\mathrm H_0(A, {-} ) \dashv [A, {-}]$, giving a morphism
$$
\xymatrix{
\mathrm H^0(A, M) \ar[rr]^-{\mathrm H^0(A, \eta)} && \mathrm H^0 (A, [A, \mathrm H_0(A, M)]) \ar[r]^-{\cong} & [\mathrm H_0(A,A), \mathrm H_0 (A, M) ]
}
$$
where the isomorphism is that of Theorem~\ref{hochuniv}. Under the closed monoidal structure of $k\mbox{-}\mathbf{Mod}$, this morphism corresponds to one
$$
\xymatrix{
\mathrm H^0(A, M) \otimes \mathrm H_0(A, A) \ar[r]^-\cap &\mathrm H_0 (A, M),
}
$$
which we call the \emph{cap product.}
\begin{rem}
The reason for this terminology is that it is a special case of the cap product in homological algebra (see~\cite{MR1731415} and~\cite{MR3281654} for a more general version for Hopf algebroids).
\end{rem}
\section{The lax categorical case}\label{HOCHLAX}
We now carry out similar constructions to those of Section~\ref{hochclassic} in the context of categories. Prior to this, we remark that what follows can probably be done with reference to an arbitrary monoid in a symmetric monoidal closed bicategory (so in particular Section~\ref{hochclassic} becomes a special case of this section), but it is beyond the scope of this thesis.
Throughout, let $\mathcal{A}$ be a monoidal category (playing the r\^ole of $A$ in the previous section) with tensor product $\otimes$ and unit $I$. For simplicity, we assume that $\mathcal{A}$ is strict monoidal, and any module categories upon which it acts are strict also.
\subsection{The 2-category $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$}
Let $\mathcal{M}$ be a category which is both a left-module category and a right-module category for $\mathcal{A}$, with actions
$$
\xymatrix{
\mathcal{A} \times \mathcal{M} \ar[r]^-{\rhd} & \mathcal{M}, \qquad \mathcal{M} \times \mathcal{A} \ar[r]^-{\lhd} & \mathcal{M}
}
$$
that are \emph{laxly compatible}, in the sense that we have morphisms
$$
\xymatrix{
A \rhd (M \lhd B) \ar[r]^-\chi & (A \rhd M) \lhd B
}
$$
natural in $A,M,B$ satisfying some coherence conditions: namely, the diagrams
\begin{align*}
\xymatrix@C=2.6em{
(A \otimes A') \rhd (M \lhd B) \ar[rr]^-{\chi} \ar@{=}[d] & & ( (A \otimes A') \rhd M ) \lhd B \ar@{=}[d] \\
A \rhd (A' \rhd (M \lhd B)) \ar[r]_-{A \rhd \chi} & A \rhd((A\ \rhd M) \lhd B) \ar[r]_-\chi & (A \rhd(A' \rhd M)) \lhd B
} \\
\xymatrix@C=2.6em{
A \rhd (M \lhd (B \otimes B') ) \ar[rr]^-\chi \ar@{=}[d] & & (A \rhd M) \lhd (B \otimes B') \ar@{=}[d] \\
A \rhd((M \lhd B) \lhd B') \ar[r]_-{\chi} & (A \rhd (M \lhd B)) \lhd B' \ar[r]_-{\chi \rhd B'} & ((A \rhd M) \lhd B) \lhd B'
}
\end{align*}
commute, and such that $\chi = 1$ whenever $A$ or $B$ is the unit $I$. A 1-cell $P \colon \mathcal{M} \to \mathcal N$ of such categories is defined to be a functor $P$ such that $PM \lhd B = P(M \lhd B)$ naturally, together with morphisms
$$
\xymatrix{
A \rhd PM \ar[r]^-{r} & P(A \rhd M)
}
$$
natural in $A,M$, such that the three diagrams
$$
\begin{array}{c}
\xymatrix@C=2.5em{
(A \otimes B) \rhd PM \ar[rr]^-r \ar@{=}[d] &&\ar@{=}[d] P( (A \otimes B) \rhd M) \\
A \rhd(B \rhd PM) \ar[r]_-{A \rhd r} & A \rhd P(B \rhd M) \ar[r]_-r & P(A \rhd (B \rhd M))
}
\\
\xymatrix@C=2.5em{
A\rhd P(M \lhd B) \ar@{=}[d] \ar[rr]^-r && \ar@{=}[d] P(A \rhd (M \lhd B) ) \\
A \rhd(PM \lhd B) \ar[r]_-\chi & ( A \rhd PM) \lhd B \ar[r]_-{r \lhd B} & P(A \rhd M) \lhd B
}
\\
\xymatrix{
I \rhd PM \ar@{=}[dr] \ar[r]^-r & P(I \rhd M) \ar@{=}[d] \\
& PM
}
\end{array}
$$
commute. In other words, $P$ preserves the right action, but only laxly preserves the left action. A 2-cell $\alpha \colon (P, r) \Rightarrow (P', r')$ is a natural transformation $\alpha \colon P \Rightarrow P'$ such that the diagram
$$
\xymatrix{
A \rhd PM \ar[r]^-r \ar[d]_-{A \rhd \alpha} & P(A \rhd M) \ar[d]^-\alpha \\
A \rhd P'M \ar[r]_-{r'} & P'(A \rhd M)
}
$$
commutes. These structures constitute a 2-category, which we denote by $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$.
\subsection{(Co)homology}\label{cohomo}
We now define the analogous structures to the zeroth Hochschild (co)homology modules of Section~\ref{hochclassic}. Let $\mathcal{M}$ be a $0$-cell in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$. We begin with cohomology, as it is slightly simpler.
\begin{defn}
The category $\mathrm H^0(\mathcal{A}, \mathcal{M})$ has as objects pairs $(M, \rho)$ where $M$ is an object of $\mathcal{M}$, and $\rho \colon A \rhd M \Rightarrow M \lhd A$ is a natural morphism such that $\rho \colon I \rhd X \to X \lhd I$ is the identity and the diagram
$$
\xymatrix@C=3em{
(A \otimes B) \ar[d]_-\rho \rhd M \ar@{=}[r] & A \rhd(B \rhd M) \ar[r]^-{A \rhd \rho} & A \rhd(M \lhd B) \ar[d]^-\chi\\
M \rhd (A \otimes B) \ar@{=}[r] & (M \lhd A) \lhd B & \ar[l]^-{\rho \lhd B} (A \rhd M) \lhd B
}
$$
commutes. A morphism $f \colon (M, \rho) \to (M', \rho')$ is a morphism $f \colon M \to M'$ in $\mathcal{M}$ such that the diagram
$$
\xymatrix@C=3em{
A\rhd M \ar[d]_-\rho \ar[r]^-{A \rhd f} & A \rhd M' \ar[d]^-{\rho '} \\
M \lhd A \ar[r]_-{f \lhd A} & M' \rhd A
}
$$
commutes.
\end{defn}
The category $\mathrm H^0(\mathcal{A}, \mathcal{M})$ can also be described as a \emph{lax descent object}, see~\cite{MR1935980, MR903151, woohoo}.
\begin{defn}
The category $\mathrm H_0(\mathcal{A}, \mathcal{M})$ is constructed as follows: it has the same objects as the category $\mathcal{M}$, but the morphisms are given by taking the morphisms of $\mathcal{M}$ and adjoining morphisms
$$
\xymatrix{
M \lhd A \ar[r]^-{\phi} & A \rhd M
}
$$
natural in $A, M$, such that $ \phi \colon M \lhd I \to I \rhd M$ is the identity and the diagram
$$
\xymatrix@C=2.5em{
M \lhd (A \otimes B) \ar@{=}[r] \ar[d]_-\phi& (M \lhd A) \lhd B \ar[r]^-\phi & B \rhd( M \lhd A) \ar[d]^-\chi \\
(A \otimes B) \rhd M \ar@{=}[r] & A \rhd (B \rhd M) & \ar[l]^-{\phi} (B \rhd M) \lhd A
}
$$
commutes.
\end{defn}
Dually to cohomology, we can describe the category $\mathrm H_0(\mathcal{A}, \mathcal{M})$ as a \emph{lax codescent object}.
Both $\mathrm H_0(\mathcal{A}, {-})$ and $\mathrm H^0(\mathcal{A}, {-})$ define 2-functors $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A \to \mathbf{Cat}$.
\subsection{Universal properties}
For any category $\mathcal{Z}$, the functor category $[\mathcal{M}, \mathcal{Z}]$ becomes a 0-cell in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$, with left and right actions on a functor $F \colon \mathcal{M} \to \mathcal{Z}$ given by
$$
(A \rhd F \lhd B) (M) = F(B \rhd M \lhd A)
$$
on objects. This defines a 2-functor
$$
\xymatrix{
\mathbf{Cat} \ar[rr]^-{[\mathcal{A},{-}]} && \mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A.
}$$
Given a category $\mathcal{Y}$, the functor category $[\mathcal{Y}, \mathcal{M}]$ becomes a 0-cell in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$ in an analagous way.
\begin{thm}\label{hochgen}
There is an isomorphism of categories
$$
[\mathrm H_0(\mathcal{A}, \mathcal{M}) , \mathcal{Z}] \cong \mathrm H^0(\mathcal{A}, [\mathcal{M}, \mathcal{Z}] ),
$$
natural in $\mathcal{M}$ and $\mathcal{Z}$.
\end{thm}
\begin{proof}
We only define the isomorphisms on objects as their action on morphisms is clear. The functor $$\xymatrix{ [\mathrm H_0(\mathcal{A}, \mathcal{M}) , \mathcal{Z}] \ar[r]^-\Phi & \mathrm H^0(\mathcal{A}, [\mathcal{M}, \mathcal{Z}] )}$$ is defined as follows. Given a functor $F \colon \mathrm H_0(\mathcal{A}, \mathcal{M}) \to \mathcal{Z}$, we define $\Phi(F)$ to be the pair $( F, \rho)$ where the natural morphisms
$$\rho \colon A\rhd F \Rightarrow F \lhd A$$ are defined by
$$
\xymatrix{
F(M \lhd A) \ar[rr]^-{F\phi} && F(A \rhd M).
}
$$
The inverse functor
$$
\xymatrix{ \mathrm H^0(\mathcal{A}, [\mathcal{M}, \mathcal{Z}] ) \ar[r]^-\Theta & [\mathrm H_0(\mathcal{A}, \mathcal{M}) , \mathcal{Z}] }
$$
maps a pair $(G, \rho)$ to the functor $G^\rho \colon \mathrm H_0(\mathcal{A}, \mathcal{M}) \to \mathcal{Z}$ given by $G$ on the morphisms of $\mathcal{A}$, and on the extra morphisms
$$\xymatrix{
X \lhd A \ar[r]^-{\phi} & A \rhd X }
$$
by
$$
\xymatrix{
G(X \lhd A) = (A \rhd G)(X) \ar[r]^-{\rho} & (G \lhd A)(X) = G(A \rhd X).
}
$$
\end{proof}
\begin{thm}
There is a 2-adjunction
$$
\xymatrix{
\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A \ar@{}[rrr]|-{\perp}\ar@/^0.7pc/[rrr]^-{\mathrm H_0(\mathcal{A}, {-})} & & & \ar@/^0.7pc/[lll]^-{[\mathcal{A}, {-}]} \mathbf{Cat}.
}
$$
\end{thm}
\begin{proof}
Let $\mathcal{M}$ be an object of $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$. For any object $M$ in $\mathcal{M}$, consider the functor $\eta(M) \colon \mathcal{A} \to \mathrm H_0(\mathcal{A}, \mathcal{M})$ which maps an object $A$ to $A\rhd{M }$. We have the functorial equalities
\begin{align*}
(A \rhd \eta(M)) (B) :&= \eta(M)(B \otimes A) \\
&= (B \otimes A) \rhd M \\
&= B \rhd (A \rhd M) \\
&= \eta(A \rhd M) (B)
\end{align*}
thus proving that $(A \rhd \eta(M) ) = \eta( A \rhd M)$. This means that $\eta$ itself is a 1-cell in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$, so we have defined a 2-natural transformation $$\eta \colon 1 \Rightarrow [\mathcal{A}, \mathrm H_0(\mathcal{A}, {-})].$$
For any category $\mathcal{Z}$, we define a functor
$$
\xymatrix{
\mathrm H_0(\mathcal{A}, [\mathcal{A}, \mathcal{Z}]) \ar[r]^-\epsilon & \mathcal{Z}
}
$$
on objects by $F \mapsto F(I)$. On a morphism of $[\mathcal{A}, \mathcal{Z}]$, i.e.\ a natural transformation, $\epsilon$ maps this to the component at the unit object $I$. The extra morphisms
$$\xymatrix{
F \lhd A \ar[r]^-\phi & A \rhd F
}$$
are mapped to the identity morphism
$$
(F \lhd A) (I) = F(A \otimes I) = F(I \otimes A) = (A\rhd F)(I).
$$
We have thus defined a 2-natural transformation
$$
\epsilon \colon \mathrm H_0 (\mathcal{A}, [\mathcal{A}, {-} ] ) \Rightarrow 1
$$
and it is routine to check that $\eta, \epsilon$ satisfy the triangle identities.
\end{proof}
For an object $\mathcal{M}$ in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$, we take the unit $\eta \colon \mathcal{M} \to [\mathcal{A}, \mathrm H_0(\mathcal{A}, \mathcal{M})]$ and apply the functor $\mathrm H^0(\mathcal{A},{-})$ as well as the isomorphism of Theorem~\ref{hochgen} to construct a morphism
$$
\xymatrix{
\mathrm H^0(\mathcal{A}, \mathcal{M}) \ar[rr]^-{\mathrm H^0(\mathcal{A}, \eta)} & & \mathrm H^0(\mathcal{A}, [\mathcal{A}, \mathrm H_0(\mathcal{A}, \mathcal{M})]) \ar[r]^-{\cong} & [\mathrm H_0(\mathcal{A}, \mathcal{A}), \mathrm H_0(\mathcal{A}, \mathcal{M})].
}
$$
which we call the \emph{cap product} (cf.\ Section~\ref{hochunivprop}) which we denote by $\cap$.
\section{Application to duplicial objects}\label{duplapp}
We now focus on a special case of the constructions in the previous section. Consider the augmented simplicial category $\Delta_+$ (cf.~Definition~\ref{simpldef}). This becomes a monoidal category with tensor product $\oplus$ defined on objects by
$$
\mathbf{m \oplus n = m + n+1},
$$
and the tensor product of two morphisms $f \colon \mathbf{m} \to \mathbf{m'}$, $g \colon \mathbf{n} \to \mathbf{n'}$ is given by
$$
(f \oplus g )(i) = \begin{cases} f(i) & \mbox{ if } 0 \le i \le m \\ g(i - m - 1) &\mbox{ if } m < i \le m+n+1\end{cases}
$$ The unit is given by $\mathbf{-1} = \emptyset$.
\subsection{Actions}
We now set $\mathcal{A} = \Delta_+^*$ for the remainder of this chapter, and apply the results of Section~\ref{HOCHLAX}.
Let $\mathcal{M}$ be a category.
\begin{prop}\label{leftaction}
Strict left actions
$$\xymatrix{
\mathcal{A} \times \mathcal{M} \ar[r] & \mathcal{M}}
$$
correspond to comonads on $\mathcal{M}$.
\end{prop}
\begin{proof}
An action $\mathcal{A} \times \mathcal{M} \to \mathcal{M}$ corresponds to a strict monoidal functor $\mathcal{A} \to [\mathcal{M}, \mathcal{M}]$, which in turn corresponds to a comonad in the monoidal category $[\mathcal{M}, \mathcal{M}]$ which finally corresponds to a comonad in $\mathcal{M}$.
\end{proof}
Explicitly, given a left action $\rhd \colon \mathcal{A} \times \mathcal{M} \to \mathcal{M}$, the corresponding comonad $T$ on $\mathcal{M}$
is defined as $T = \mathbf 0 \rhd{-}$, with counit $\epsilon$ given on components $X$ by
$$
\xymatrix{
T(X) = \mathbf 0 \rhd X \ar[rr]^-{d_0 \rhd X} && \mathbf{-1} \rhd X = X
}
$$
where $d_0 \colon \mathbf{0 \to -1}$ is the unique face map in $\mathcal{A}$. The coproduct $\delta$ is defined on components by
$$
\xymatrix{
T X = \mathbf{0} \rhd X \ar[rr]^-{s_0 \rhd X} & & \mathbf{1} \rhd X = \mathbf 0 \rhd (\mathbf 0 \rhd X) = T T X
}
$$
The augmented simplicial category $\Delta_+$ becomes a monoidal category with the reverse tensor product, given explicitly on objects by
$$
\mathbf{m \oplus^{\operatorname{rev}} n = n \oplus m}.
$$
We denote this by $\Delta_+^{\operatorname{rev}}$. There is an isomorphism $\Delta_+^{\operatorname{rev}} \cong \Delta_+$ of monoidal categories~\cite{woohoo}, which of course carries over to the dual categories $\mathcal{A}^{\operatorname{rev}} \cong \mathcal{A}$. Right actions of $\mathcal{A}$ correspond by to right actions of $\mathcal{A}^{\operatorname{rev}}$ by the monoidal isomorphism, and these in turn correspond to left actions of $\mathcal{A}$, which correspond to comonads by Proposition~\ref{leftaction}.
Thus we have:
\begin{prop}
Strict right actions
$$\xymatrix{
\mathcal{M} \times \mathcal{A} \ar[r] & \mathcal{M}
}$$
correspond to comonads on $\mathcal{M}$.
\end{prop}
Given a right action $\lhd$, the corresponding comonad $S$ is explicitly defined in an analagous way to comonads corresponding to left actions.
\begin{lem}\label{whenwhen} Distributive laws of comonads correspond to 0-cells of $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$.
\end{lem}
\begin{proof}
Suppose that $\mathcal{M}$ is an object in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$, so that we have natural morphisms
$$
\xymatrix{
\mathbf{m} \rhd (M \lhd \mathbf n) \ar[r]^-{\chi} & (\mathbf{m} \rhd M) \lhd \mathbf n.
}
$$
By choosing $\mathbf{m} = \mathbf{n} = \mathbf 0$, we get morphisms
$$
\xymatrix{
T S(M) \ar[r]^-\chi & ST(M)
}
$$
natural in $M$. The naturality of $\chi$ in $\mathbf{m},\mathbf n$, combined with compatibility with the unit object $\mathbf{-1}$, tells us that $\chi$ is compatible with the comultiplication and counits of both comonads, realising it as a distributive law.
Conversely, let $\chi$ be a distributive law with components
$$
\xymatrix{
T S(M) \ar[r]^-\chi & ST(M).
}
$$
These morphisms can easily be upgraded by iterating $\chi$ up to horizontal composition of identities, giving morphisms $T^nS^m (M) \to S^mT^n(M)$ for any $m,n \ge 0$ \`a la Section~\ref{evidenziatore1}. By definition, the coherence conditions required for $\chi$ to turn $\mathcal{M}$ into an object of $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$ are satisfied.
\end{proof}
\subsection{(Co)homology}
For this subsection, fix a distributive law $\chi \colon TS\Rightarrow ST$. Let $\mathcal{M}$ denote the $0$-cell of $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$ corresponding to $\chi$ under Lemma~\ref{whenwhen}.
\begin{lem}\label{chicoalghoch}
An object of $\mathrm H^0(\mathcal{A}, \mathcal{M})$ corresponds to a $\chi$-coalgebra in $\mathcal{M}$.
\end{lem}
\begin{proof}
Let $\rho \colon TM \Rightarrow SM$ be a $\chi$-coalgebra structure on an object $M$ in $\mathcal{M}$. By Proposition~\ref{simplicialrho}, $\rho$ extends to a morphism of simplicial objects
$$
\xymatrix{
\rBB(T,M) \ar[r]^-{\rho} & \rBB^*(S,M),
}$$
and in fact extends to a morphism of augmented simplicial objects by setting $\rho$ to be $1 \colon M \to M$ in degree $-1$. Then, by construction, the pair $(M,\rho)$ is an object of $\mathrm H^0(\mathcal{A}, \mathcal{M})$.
Conversely, given an object $(M, \rho)$ in $\mathrm H^0(\mathcal{A}, \mathcal{M})$, the morphism $\rho$ in degree 0 gives a $\chi$-coalgebra structure on $M$.
\end{proof}
Similarly, we have the following:
\begin{lem}
Functors $\mathrm H_0(\mathcal{A}, \mathcal{M}) \to \mathcal{Z}$ correspond to $\chi$-opcoalgebras $\mathcal{M} \to \mathcal{Z}$. \end{lem}
\subsection{The d\'ecalage comonads}\label{decalage}
Let $X$ be an augmented simplicial object in a category $\mathcal{Z}$.
\begin{defn}
The \emph{right d\'ecalage} (French for \emph{shift}) of $X$, denoted $\operatorname{Dec}^{\operatorname{r}}(X)$, is the simplicial object given in degree $n$ by
$$\operatorname{Dec}^{\operatorname{r}}(X)_n = X_{n+1},$$ whose faces are given by discarding the last face of $X$, and similarly for the degeneracies.
\end{defn}
See~\cite{MR2399898,MR3065943} for more on d\'ecalage.
Pictorially (ignoring the degeneracies), the right d\'ecalage of an augmented simplicial object
$$
\xymatrix{
\cdots X_2 \ar@<2.5ex>[r]^-{d_0} \ar[r]^-{d_1} \ar@<-2.5 ex>[r]^-{d_2}& X_1 \ar@<1.25ex>[r]^-{d_0} \ar@<-1.25ex>[r]^-{d_1} & X_0 \ar[r]^-{d_0} & X_{-1}
}
$$
looks like
$$
\xymatrix{
\cdots X_3 \ar@<2.5ex>[r]^-{d_0} \ar[r]^-{d_1} \ar@<-2.5 ex>[r]^-{d_2}& X_2 \ar@<1.25ex>[r]^-{d_0} \ar@<-1.25ex>[r]^-{d_1} & X_1 \ar[r]^-{d_0} & X_{0}
\\ \\}
$$
In fact, $\operatorname{Dec}^{\operatorname{r}}$ is a comonad on $[\mathcal{A}, \mathcal{Z}]$, where the counit $\epsilon \colon \operatorname{Dec}^{\operatorname{r}}(X) \to X$ is given in each degree by the missing face map, and the comultiplication $\delta \colon \operatorname{Dec}^{\operatorname{r}}(X) \to \operatorname{Dec}^{\operatorname{r}} \operatorname{Dec}^{\operatorname{r}}(X)$ is given by the missing degeneracy.
In a similar way, we define the left d\'ecalage of an augmented simplicial set by discarding the zeroth face and degeneracy maps. We denote this by $\operatorname{Dec}^{\operatorname{l}}$. Since $\operatorname{Dec}^{\operatorname{r}}\operatorname{Dec}^{\operatorname{l}}$ is naturally equal to $\operatorname{Dec}^{\operatorname{l}} \operatorname{Dec}^{\operatorname{r}}$, we get the following:
\begin{prop}\label{decalagep}
The identity natural transformation defines a distributive law $$\xymatrix{{\operatorname{Dec}^{\operatorname{r}}} {\operatorname{Dec}^{\operatorname{l}}} \ar@{=>}[r]^-\chi & {\operatorname{Dec}^{\operatorname{l}}} {\operatorname{Dec}^{\operatorname{r}}}.}$$
\end{prop}
Now, let us view this distributive law on $[\mathcal{A}, \mathcal{Z}]$ as an object of $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$.
\begin{thm}\label{ultimate}
The category $\mathrm H^0(\mathcal{A}, [\mathcal{A}, \mathcal{Z}] )$ is isomorphic to the category of augmented duplicial objects in $\mathcal{Z}$.
\end{thm}
\begin{proof}
By Lemma~\ref{chicoalghoch}, to give an object $(X, t)$ of $\mathrm H^0(\mathcal{A}, [\mathcal{A}, \mathcal{Z}] )$ is the same as to give a $\chi$-coalgebra in $[\mathcal{A}, \mathcal{Z}]$. This is a morphism $t \colon \operatorname{Dec}^{\operatorname{r}} X \to \operatorname{Dec}^{\operatorname{l}} X$, so we get an operator
$$
\xymatrix{
X_n \ar[r]^-t & X_n
}
$$ in each degree. That $t$ commutes with the faces and degeneracies is precisely that the equations
\begin{align*}
d_i t = t d_{i-1}, \qquad s_i t = ts_{i-1}
\end{align*}
hold, and that $t$ is compatible with $\epsilon$ and $\delta$ as a $\chi$-coalgebra structure is precisely that the equations
$$
d_i t = d_n, \qquad s_it = t^2s_n
$$
respectively hold; that is, $t$ is a duplicial operator.
\end{proof}
\begin{rem}\label{remrem}
The category $\Delta_+$ becomes both a left-module and right-module over itself with actions given by the monoidal product $\oplus$. These actions are strictly compatible, and moreover they restrict to the simplicial category $\Delta$. This carries over to the dual categories, and so both $\Delta^*$ and $[\Delta^*, \mathcal{Z}]$ become $0$-cells in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$. We can then consider the (co)homology categories
$$
\mathrm H_0(\mathcal{A}, \Delta^*), \qquad \mathrm H_0(\mathcal{A}, [\Delta^*, \mathcal{Z}]),
$$
the latter of which is isomorphic to the category of unaugmented duplicial objects in $\mathcal{Z}$.
\end{rem}
Thus, we have another way of deriving the duplicial functor at the heart of B\"ohm and {\c S}tefan's Theorem~\ref{dup}. Indeed, let $M$ be a $\chi$-coalgebra in $\mathcal{M}$, viewed as an object of $\mathrm H^0(\mathcal{A}, \mathcal{M})$. Let $N \colon \mathcal{M} \to \mathcal{Z}$ be a $\chi$-opcoalgebra, viewed as a functor $\mathrm H_0(\mathcal{A}, \mathcal{M}) \to \mathcal{Z}$. Applying the cap product
$$
\xymatrix{
\mathrm H^0(\mathcal{A}, \mathcal{M}) \ar[r]^-{\cap} & [\mathrm H_0(\mathcal{A}, \mathcal{A}), \mathrm H_0(\mathcal{A}, \mathcal{M})]
}
$$
to $M$ gives us a functor
$$
\xymatrix{ \mathrm H_0(\mathcal{A}, \mathcal{A}) \ar[rr]^-{\cap(M)} && \mathrm H_0(\mathcal{A}, \mathcal{M})}
$$
that we may compose with $N$ to obtain a functor
$$
\xymatrix{ \mathrm H_0(\mathcal{A}, \mathcal{A}) \ar[rr]^-{X} && \mathcal{Z}}.
$$
This is of course an object in $[\mathrm H_0(\mathcal{A}, \mathcal{A}), \mathcal{Z}]$ which is isomorphic to $\mathrm H^0(\mathcal{A}, [\mathcal{A}, \mathcal{Z}])$ by Theorem~\ref{hochgen}. By Theorem~\ref{ultimate}, this is in turn isomorphic to the category of augmented duplicial objects in $\mathcal{Z}$. By replacing $\mathcal{M}$ with $[\mathcal{Y}, \mathcal{M}]$ for a general category $\mathcal{Y}$, we recover the duplicial functor $\rCC_T(N,M)$ of Theorem~\ref{dup}.
\subsection{Duplicial objects from twisting}\label{bonusect}
Here, we show that Theorem~\ref{dup} can be deduced directly from the twisting procedure given in Section~\ref{twistcoeff}. Recall from Proposition~\ref{decalagep} that for an arbitrary category $\mathcal{Z}$, the identity defines a distributive law $1\colon \operatorname{Dec}^{\operatorname{r}}\operatorname{Dec}^{\operatorname{l}} \Rightarrow \operatorname{Dec}^{\operatorname{l}}\operatorname{Dec}^{\operatorname{r}}$.
Suppose we have a distributive law $\chi \colon TS \Rightarrow ST$ of comonads on a category $\mathcal{B}$, together with a $\chi$-opcoalgebra $(N, \lambda) \colon \mathcal{B} \to \mathcal{Z}$.
Consider the functor
$$
\xymatrix{
\mathcal{B} \ar[rr]^-{\mathrm C_T(N,{-})} && [\mathcal{A}, \mathcal{Z}]
}
$$
that assigns to every object $X$, the simplicial object $\mathrm C_T(N,X)$ of Section~\ref{duplobjsec}.
For every $n\ge 0$, we have
$$
C_T(N, SX)_n = NT^{n+1} SX, \qquad \operatorname{Dec}^{\operatorname{l}} \mathrm C_T(N, X)_n = \mathrm C_T(N, X)_{n+1} = NT^{n+2} X
$$
The morphisms
$$
\xymatrix{
NT^{n+1}SX \ar[rr]^-{N \chi^{n+1}_X} & & NST^{n+1}X \ar[rr]^-{\lambda_{T^{n+1} X}} & & NT^{n+2}X
}
$$
then define a natural transformation
$$
\xymatrix{
\mathrm C_T(N,{-}) \circ S \ar@{=>}[r]^-{\gamma} & {\operatorname{Dec}^{\operatorname{l}}} \circ~\mathrm C_T(N, {-}).
}
$$
We also have
$$
{\operatorname{Dec}^{\operatorname{r}}} \mathrm C_T(N, X)_n = \mathrm C_T(N, X)_{n+1} = NT^{n+2}X, \qquad \mathrm C_T(N, TX)_n = NT^{n+2} X
$$
and it turns out that the identity defines a natural transformation
$$
\xymatrix{
{\operatorname{Dec}^{\operatorname{r}}}\circ~\mathrm C_T(N,{-}) \ar@{=>}[r]^-{1} & \mathrm C_T(N, {-}) \circ T.
}
$$
It is routine to check the following:
\begin{prop}\label{ultimate1}
We have that
$$
\xymatrix{
(\mathcal{B}, \chi, T, S) \ar[rrr]^-{(\mathrm C_T(N, -), 1, \gamma)} & & & ([\mathcal{A}, \mathcal{Z}], 1, \operatorname{Dec}^{\operatorname{r}}, \operatorname{Dec}^{\operatorname{l}})
}
$$
is a 1-cell in $\mathbf{Dist}$.
\end{prop}
Now let $M$ be a $\chi$-coalgebra in $\mathcal{B}$. The 1-cell in Proposition~\ref{ultimate1} acts on this to give a $\chi$-coalgebra in $[\mathcal{A}, \mathcal{Z}]$ which is precisely an augmented duplicial object in $\mathcal{Z}$ by Lemma~\ref{chicoalghoch}. Given a $\chi$-coalgebra structure on a functor $M \colon \mathcal{Y} \to \mathcal{B}$, we recover the more general duplicial functor of Theorem~\ref{dup} by carrying out the replacements
$$
\xymatrix@=1em{
\mathcal{B} \ar@{|->}[r] & [\mathcal{Y}, \mathcal{B}], & \mathcal{Z} \ar@{|->}[r] & [\mathcal{Y}, \mathcal{Z}], & N \ar@{|->}[r] & [\mathcal{Y}, N], \\
& T \ar@{|->}[r] & [\mathcal{Y}, T], & S \ar@{|->}[r] & [\mathcal{Y}, S]
}
$$
in the above arguments.
\subsection{Restriction to $\mathbf{Cat}$}\label{nerve}
In this final section of the chapter, we view $\mathbf{Cat}$ as a 0-cell of $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$ via the nerve functor, and study what it means to be an object of the cohomology category $\mathrm H^0(\mathcal{A}, \mathbf{Cat})$. Throughout, $\mathcal{C}$ denotes a (small) category.
\begin{defn}
The \emph{nerve of $\mathcal{C}$}, denoted by $N \mathcal{C}$, is the simplicial set where $(N\mathcal{C})_n$ is the set of composable $n$-tuples of morphisms. In degree $n$, the faces and degeneracies, defined on an $n$-tuple
$$
\xymatrix{
A_0 \ar[r]^-{f_0} & A_1 \ar[r]^-{f_1} & \cdots \ar[r]^-{f_{n-2}} & A_{n-1} \ar[r]^-{f_{n-1}} & A_{n}
}
$$
are given by
\begin{align*}
d_i (f_{n-1}, \ldots, f_{0} ) &= \begin{cases} (f_{n-1}, \ldots, f_1) & \mbox{ if } i=0 \\
(f_{n-1}, \ldots, f_i f_{i-1}, \ldots, f_0) & \mbox{ if } 0 < i < n \\
(f_{n-2}, \ldots, f_0) & \mbox{ if } i = n \end{cases} \\
s_i (f_{n-1}, \ldots, f_{0} ) &= (f_{n-1}, \cdots, f_i, 1, f_{i-1}, \cdots, f_0 ).
\end{align*}
\end{defn}
The nerve actually defines a fully faithful functor $N \colon \mathbf{Cat} \to [\Delta^*, \mathbf{Set}]$ (see e.g.~\cite[II.4.22]{MR1950475}) so we may view $\mathbf{Cat}$ as a full subcategory of the category of simplicial sets. We do not give precise details, but the actions of $\mathcal{A}$ on $[\Delta^*, \mathbf{Set}]$ mentioned in Remark~\ref{remrem} restrict to actions on $\mathbf{Cat}$, and so $\mathbf{Cat}$ becomes a 0-cell in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$ in its own right (for a more detailed explanation, see~\cite{woohoo}). We are then able to consider the cohomology category $\mathrm H^0(\mathcal{A}, \mathbf{Cat})$. Explicitly, an object of this category is a (small) category $\mathcal{C}$ together with a duplicial structure on its nerve $N\mathcal{C}$.
To give a duplicial structure on $N\mathcal{C}$ amounts to giving the following structure on $\mathcal{C}$:
\begin{itemize}
\item for each object $A$, an object $tA$
\item for each morphism $f\colon A \to B$ a morphism $tf \colon tB \to A$
\end{itemize}
such that $t^2(1_A) = 1_{tA}$ for objects $A$ of $\mathcal{C}$, and $t$ maps any commutative triangle of morphisms as
$$
\begin{array}{c}
\xymatrix{
A \ar[r]^-f \ar[dr]_-{h} & B \ar[d]^-g \\
& C
}\end{array}
\mapsto
\begin{array}{c}
\xymatrix{
tC \ar[r]^-{th} \ar[dr]_-{tg} & A \ar[d]^-f \\
& B
}
\end{array}
$$
\begin{defn}\label{corefdef}
We call a morphism $t$ as described above a \emph{coreflector}.
\end{defn}
The following theorem characterises such structures, and Corollary~\ref{motivation} makes clear the motivation for the terminology in Definition~\ref{corefdef}.
\begin{thm}\label{auspara}
A category $\mathcal{C}$ has a duplicial structure on its nerve if and only if there exist a groupoid $\mathcal G$ and an adjunction
$$
\xymatrix{
\mathcal G \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-I & & \ar@/^0.5pc/[ll]^-R \mathcal{C}.
}$$
\end{thm}
\begin{proof}
Suppose that we are given such an adjunction, with unit $\eta$ and counit $\epsilon$. For an object $A$ in $\mathcal{C}$ we define $tA := IRA$, and for a morphism $f\colon A \to B$ in $\mathcal{C}$ we define $tf \colon tB \to A$ to be the composite
$$
\xymatrix{
IRB \ar[rr]^-{I(Rf)^{-1}} & &I RA \ar[r]^-{\epsilon_A} & A.
}
$$
Note that $tf$ is well defined because $Rf$ is invertible, being a morphism in a groupoid.
Consider an identity morphism $1 \colon A\to A$. We have that $t^2(1)$ is equal to
$$
\xymatrix{
IRA \ar[rr]^-{IR(\epsilon_A)^{-1}} && IRIRA \ar[rr]^-{\epsilon_{IRA}} & & IRA.
}
$$
One triangle identity for $I \dashv R$ implies that the diagram
$$
\xymatrix{RA \ar@{=}[dr] \ar[r]^-{\eta_{RA}} & RIRA \ar[d]^-{R\epsilon_A} \\
& RA}
$$
commutes, which
tells us that $R(\epsilon_A)^{-1} = \eta_{RA}$. Therefore, using the other triangle identity, we have that the diagram
$$
\xymatrix@C=3.5em{
IRA \ar@{=} [dr]\ar[r]^-{IR(\epsilon_A)^{-1}} & IRIRA \ar[d]^-{\epsilon_{IRA}} \\
& RA
}
$$
commutes, that is $t^2(1) = 1$.
Now suppose that we have a commutative triangle
$$\xymatrix{
A \ar[r]^-f \ar[dr]_-{h} & B \ar[d]^-g \\
& C
}$$
Since $h = gf$ we have $(Rh)^{-1} = (Rf)^{-1} (Rg)^{-1}$ and so $Rf (Rh)^{-1} = (Rg)^{-1}$. In the diagram
$$
\xymatrix@C=3.5em{
IRC \ar[r]^-{I(Rh)^{-1}} \ar[dr]_-{I(Rg)^{-1}\ } & IRA \ar[d]^-{IRf} \ar[r]^-{\epsilon_A} & A \ar[d]^-f \\
& IRB \ar[r]_-{\epsilon_B} & B
}
$$
the inner triangle commutes by the aforementioned relation, and the right-hand square commutes by naturality of $\epsilon$. The outer diagram commutes, which says that $tg = f(th)$, completing one direction of the proof.
Conversely, suppose that we have a coreflector $t$ giving rise to a duplicial structure on $N\mathcal{C}$. There is an induced functor $G \colon \mathcal{C} \to \mathcal{C}$, defined on objects and morphisms respectively by
$$
G(A) := tA, \qquad G\!\xymatrix{(A \ar[r]^-f & B)} := \xymatrix{tA \ar[r]^-{t^2 f} & tB.}
$$
Let $f\colon A \to B$ be any morphism in $\mathcal{C}$. By applying $t$ to two commutative triangles as follows
$$
\begin{array}{c}
\xymatrix{
A \ar[dr]_-f \ar[r]^-f & B \ar[d]^-1 \\
& B
}
\end{array}
\mapsto
\begin{array}{c}
\xymatrix{
tB \ar[dr]_-{t1} \ar[r]^-{tf} & A \ar[d]^-f \\
& B
}
\end{array}, \qquad
\begin{array}{c}
\xymatrix{
tB \ar[dr]_-{tf} \ar[r]^-{tf} & A \ar[d]^-1 \\
& A
}
\end{array}
\mapsto
\begin{array}{c}
\xymatrix{
tA \ar[dr]_-{t1} \ar[r]^-{t^2f} & B \ar[d]^-{tf} \\
& B
}
\end{array}
$$
we see that the diagram
$$
\xymatrix{
tA\ar[d]_-{t^2 f} \ar[r]^-{t1} & A \ar[d]^-f \\
tB\ar[ur]_-{tf} \ar[r]_-{t1} & B
}
$$
commutes. By defining $\epsilon_A := t\!\xymatrix{(A \ar[r]^-1 & A)}$, we therefore obtain a natural transformation $\epsilon \colon G \Rightarrow 1$. Also, together with $\epsilon$, $G$ becomes a well-copointed endofunctor: that is, for any $A$ we have
$$
G(\epsilon_A) = t^2(\epsilon_A) = t^3(1_A) = t(1_{tA} ) = \epsilon_{GA}.
$$
Let $f\colon A \to B$ be any morphism. By applying $t$ twice:
$$
\begin{array}{c}
\xymatrix{
t A \ar[dr]_-{f\epsilon_A} \ar[r]^-{\epsilon_A} & A \ar[d]^-f \\
& B
}
\end{array}
\mapsto
\begin{array}{c}
\xymatrix{
tB \ar[dr]_-{tf} \ar[r]^-{t(f\epsilon_A)} & tA \ar[d]^-{\epsilon_A} \\
& A
}
\end{array}
\mapsto
\begin{array}{c}
\xymatrix{
tA \ar[dr]_-{t(\epsilon_A)} \ar[r]^-{t^2 f} & tB \ar[d]^-{t(f\epsilon_A)} \\
& t A
}
\end{array}
$$
we have that $$t(f\epsilon_A)\circ t^2f = t(\epsilon_A) = 1.$$ Furthermore, by naturality of $\epsilon$ we have $$f\epsilon_A = \epsilon_B \circ t^2 f$$ and applying $t$ once more:
$$
\begin{array}{c}
\xymatrix{
tA \ar[dr]_-{f\epsilon_A} \ar[r]^-{t^2 f} & tB \ar[d]^-{\epsilon_B} \\
& B
}
\end{array}
\mapsto
\begin{array}{c}
\xymatrix{
tB \ar[dr]_-{t(\epsilon_B)} \ar[r]^-{t(f\epsilon_A)} & tA \ar[d]^-{t^2 f} \\
& tB
}
\end{array}
$$
tells us that $$
t^2f \circ t(f\epsilon_A) = t^2 f \circ t(\epsilon_B \circ t^2 (f) ) = 1.
$$
Thus $Gf = t^2f$ is invertible for any morphism $f$. It follows that $G$ becomes an (idempotent) comonad with counit $\epsilon$ and comultiplication $ \delta := G(\epsilon)^{-1}$. The groupoid $\mathcal G$ that we require is the Eilenberg-Moore category $\mathcal{C}^G$, together with the canonical associated adjunction (cf.\ Section~\ref{EM2cat}), which is indeed a groupoid since $G$ inverts morphisms.
\end{proof}
Later, in Section~\ref{finally}, we see that this example of a duplicial structure can be derived from a special case of one involving enriched functor categories.
\begin{cor}\label{motivation}
A category $\mathcal{C}$ admits a duplicial structure on its nerve if and only if $\mathcal{C}$ has a coreflective subcategory which is a groupoid.
\end{cor}
\begin{proof}
Given two categories $\mathcal G, \mathcal{C}$ and a left-adjoint functor $I \colon \mathcal G \to \mathcal{C}$, the unit $\eta \colon 1 \Rightarrow RI$ is necessarily an isomorphism so $I$ is full and faithful~\cite[Prop.~3.4.1]{MR1291599}. Thus, the corollary follows immediately from Theorem~\ref{auspara}.
\end{proof}
\begin{rem}\label{catrem}
Suppose that we have two groupoids $\mathcal G$ and $\mathcal G'$, as well as two adjunctions
$$
\xymatrix{
\mathcal G \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-I & & \ar@/^0.5pc/[ll]^-R \mathcal{C},
}\qquad
\xymatrix{
\mathcal G' \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-{I'} & & \ar@/^0.5pc/[ll]^-{R'} \mathcal{C}.}
$$
The functor
$$
\xymatrix{
\mathcal G \ar[r]^-{I} & \mathcal{C} \ar[r]^-{R'} & \mathcal G'
}
$$
defines an equivalence of categories $\mathcal G \simeq \mathcal G'$. In fact, in the context of Theorem~\ref{auspara}, we may always choose $\mathcal G$ to be the fundamental groupoid $\Pi_1(\mathcal{C})$ (this is the localisation of $\mathcal{C}$ at all morphisms~\cite{MR0210125}, see also~\cite{woohoo}).
\end{rem}
Thus we have another way to state Theorem~\ref{auspara}:
\begin{cor}
A category $\mathcal{C}$ admits a duplicial structure on its nerve if and only if the universal functor $\pi\colon \mathcal{C} \to \Pi_1(\mathcal{C})$ has a left adjoint.
\end{cor}
\begin{cor}
A category $\mathcal{C}$ admits a cyclic structure on its nerve if and only if $\mathcal{C}$ is a groupoid.
\end{cor}
\begin{proof}
Suppose that a category $\mathcal{C}$ is equipped with a cyclic structure on its nerve, induced by a coreflector $t$. By assumption, $t$ is the identity on objects, and $t^2$ is the identity on morphisms. By the proof of Theorem~\ref{auspara}, for all morphisms $f$ in $\mathcal{C}$, the morphism $t^2 f = f$ is invertible and thus $\mathcal{C}$ is a groupoid.
Conversely, if $\mathcal{C}$ is a groupoid, then we apply Theorem~\ref{auspara} to the the identity functor $1\colon \mathcal{C} \to \mathcal{C}$, which is trivially a left-adjoint functor.
\end{proof}
\begin{rem}
Note that a groupoid $\mathcal{C}$ may have multiple cyclic structures on its nerve. Indeed, given a natural transformation
$$
\xymatrix{
\mathcal{C} \rrtwocell^1_{1}{\ \alpha} & & \mathcal{C}
}
$$
we define a coreflector $t$ on objects and morphisms respectively by
$$
tA := A, \qquad t\!\xymatrix{(A \ar[r]^-f & B)} := \xymatrix{B \ar[r]^-{\alpha_B} & B \ar[r]^-{f^{-1}} & A.}
$$
\end{rem}
\begin{rem}
In~\cite{woohoo} we also characterise duplicial structure on the nerve of a bicategory~\cite{MR920944,MR1897816} (and so in particular, on the nerve of a monoidal category).
\end{rem}
\chapter{Examples}\label{EXAMPLES}
In this chapter we give examples with an algebraic flavour.
In Section~\ref{cychomalg} we explain how the cyclic homology of an associative algebra~\cite{MR823176,MR777584} arises as an instance of Theorem~\ref{dup}. We also show that an algebra map $\sigma$ induces a 1-cell in $\mathbf{Dist}$ which acts on cyclic homology to give $\sigma$-twisted cyclic homology~\cite{MR1943179}. Next, in Section~\ref{brugsec} we show that Hopf-cyclic homology of bialgebroids (as in~\cite{MR2803876}) arises from Theorem~\ref{dup} and analyse this phenomenon in terms of opmonoidal adjunctions and their opmodules~\cite{MR3020336}. In Section~\ref{hopfywopfy} we study the various notions of bimonad and Hopf monad~\cite{MR3020336,MR2793022,MR1942328,MR3175323,MR1887157}, as well as give a new example of a bimonad that is not a Hopf monad (Section~\ref{newbimonad}). We conclude the chapter in Section~\ref{finally} by showing that the category of enriched functors from a Hopf category~\cite{Hop, MR1458415} contains a duplicial object under certain conditions.
The work contained herein is original; Section~\ref{cychomalg} is based on~\cite[\S4]{2} and Sections~\ref{brugsec} and~\ref{hopfywopfy} are based on\cite[\S5--6]{1}.
\section{Cyclic homology of algebras}\label{cychomalg}
\subsection{Flat connections}
Let $B$ be a monad on a category $\mathcal{A}$, and let
$$
\xymatrix{ \mathcal{A}
\ar@/^{0.5pc}/[rr]^-F \ar@{}[rr]|-{\perp}&&
\ar@/^0.5pc/[ll]^U\mathcal{A}^B
}
$$
be the canonical adjunction (cf.\ Sections~\ref{EM2cat}, \ref{emcatsect}). As before, let $\tilde B$ denote the comonad $FU$ on $\mathcal{A}^B$ generated by the adjunction, and let $\Sigma \colon \mathcal{A}^B \to \mathcal{A}^B$ be an endofunctor.
For any $B$-algebra $(X, \beta)$, we have natural isomorphisms
$$
\mathcal{A}^B ( \tilde B \Sigma (X,\beta), \Sigma \tilde B (X, \beta) ) \cong \mathcal \mathcal{A} ( U\Sigma (X, \beta) , U \Sigma \tilde B (X, \beta) )
$$
given by the adjunction, so there is a one-to-one correspondence between natural transformations $\sigma \colon \tilde B \Sigma \Rightarrow \Sigma \tilde B$ and natural transformations $\nabla \colon U \Sigma \Rightarrow U \Sigma \tilde B$. In fact, $(\Sigma, \sigma)$ is an opmorphism of comonads if and only if the diagrams
$$
\xymatrix{
U\Sigma \ar[r]^-{\nabla} \ar[d]_-\nabla & U\Sigma \tilde B \ar[d]^-{\nabla \tilde B} \\
U \Sigma \tilde B \ar[r]_-{U \Sigma \tilde \delta} & U \Sigma \tilde B \tilde B
}
\quad\quad\quad
\xymatrix{
U \Sigma \ar[r]^-\nabla \ar@{=}[dr] & U \Sigma \tilde B \ar[d]^-{U \Sigma \tilde \epsilon} \\
& U \Sigma
}
$$
commute, that is, if $(U\Sigma, \nabla)$ is a $\tilde B$-opcoalgebra.
\begin{defn}
We say that the natural transformation $\sigma$ is a \emph{connection} if
$\epsilon $ is compatible with $\sigma$,
i.e.\ the second diagram above commutes for the
corresponding natural transformation $\nabla$. We say
that a connection $\sigma$ is \emph{flat} if $
\delta $ is compatible with $\sigma$, i.e.\ $(\Sigma,\sigma)$ is an opmorphism of comonads, or equivalently, both diagrams above commute.
\end{defn}
The terminology is motivated by the special case
discussed in detail in Section~\ref{connectexample}.
\subsection{Cyclic and twisted cyclic homology}
Let $A$ be a unital associative algebra over a commutative ring $k$. Consider the simplicial $k$-module $C_\bullet(A,A)$ of Section~\ref{hochsub}, defined degreewise by $$C_n(A,A) = A^{\otimes (n+1)}$$ which gives rise to the Hochschild homology module $\mathrm H_\bullet(A, A)$. This simplicial object is in fact a cyclic object. Indeed, in degree $n$ we define a cyclic operator $t$ on $A^{\otimes (n+1)}$ by
$$
t(a_0 \otimes \cdots \otimes a_n) = a_1 \otimes \cdots \otimes a_n \otimes a_0.
$$
\begin{defn}
The \emph{cyclic homology of $A$}, denoted $\operatorname{HC}(A)$, is the cyclic homology of the cyclic object $C_\bullet(A,A)$.
\end{defn}
Now let $\sigma \colon A \to A$ be an algebra map. We define a simplicial module $C_\bullet (A, A)_\sigma$ as follows:
take the degeneracies and faces of the Hochschild simplicial object $C_\bullet(A,A)$ but in each degree, we change the last face $d_n \colon A^{\otimes(n+1)} \to A^{\otimes n}$ to the map $d^\sigma_n$ defined by
$$
d^\sigma_n (a_0 \otimes \cdots \otimes a_n) = (a_0 \sigma(a_1) , a_2, \cdots, a_n).
$$
In fact, this is a duplicial object with duplicial operator $t^\sigma$ defined by
$$
t^\sigma(a_0 \otimes \cdots \otimes a_n) = \sigma(a_1) \otimes \cdots \otimes a_n \otimes a_0.
$$
This is a cyclic object if and only if $\sigma = 1$.
\begin{defn}
The \emph{cyclic homology of $A$ twisted by $\sigma$}, denoted $\operatorname{HC}_\sigma(A)$, is the cyclic homology of the duplicial object $C_\bullet(A,A)_\sigma$.
\end{defn}
It is the aim of this section to show that the assignment
$$
\operatorname{HC} (A)\mapsto \operatorname{HC}_\sigma(A)
$$
can be realised in the abstract framework of the previous chapters.
\subsection{$(A,A)$-bimodules}\label{connectexample}
Let $\mathcal{A}= A\mbox{-}\mathbf{Mod}$ be the category of left $A$-modules and let $B$ be the comonad $= {-}\otimes A$, so that $\mathcal{A}^B$ is the category of $(A,A)$-bimodules as in Section~\ref{hochclassic}.
The functor $C= A \otimes {-} \colon \mathcal{A} \to \mathcal{A}$, together with the natural morphisms
\begin{align*}
\delta \colon A \otimes X &\longrightarrow A \otimes A \otimes X & \epsilon \colon A \otimes X & \longrightarrow X \\
a \otimes x &\longmapsto a \otimes 1 \otimes x & a \otimes x & \longmapsto ax
\end{align*}
defines a comonad on $\mathcal{A}^B$. There is a mixed distributive law $\theta \colon BC \Rightarrow CB$ given by rebracketing on components
$$
\theta \colon (A \otimes X) \otimes A \to A \otimes (X \otimes A)
$$
so by Corollary~\ref{arisec}, this lifts to a comonad distributive law $\tilde\theta \colon \tilde B C^\theta \Rightarrow C^\theta \tilde B$ on $\mathcal{A}^B$.
Let $N$ be an $(A,A)$-bimodule and $\Sigma \colon
\mathcal{A}^B \to \mathcal{A}^B$ be the endofunctor defined by $\Sigma (M) = M \otimes_A N$. We have that $\Sigma C^\theta = C^\theta \Sigma$ so that
$$
\xymatrix{
(\mathcal{A}^B, C^\theta) \ar[rr]^-{(\Sigma, 1)} && (\mathcal{A}^B, C^\theta)
}
$$
is a morphism of comonads.
The component of a natural transformation $\nabla \colon U \Sigma \Rightarrow U \Sigma \tilde B$ is given by a left $A$-linear map
\begin{align*}
\nabla_M \colon M \otimes_A N & \to (M \otimes A) \otimes N \cong M \otimes N
\end{align*}
The natural transformation $\nabla$ defines a
connection if and only if each $\nabla_M$ splits the
quotient map $M \otimes N \to M \otimes_A N$.
Taking $M=A$ yields an $A$-linear splitting of the action $A
\otimes N \to N$, so $N$ is $k$-relative
projective~\cite[Definition~8.6.5]{MR1269324}. Conversely, given a splitting
$n \mapsto n_{(-1)} \otimes n_{(0)}$ of the action, we
obtain $\nabla_M$ as
$\nabla_M(m \otimes_A n)=mn_{(-1)} \otimes n_{(0)}$.
Thus we have:
\begin{prop}
The functor $\Sigma$ admits a connection $ \sigma $ if and only if $N$ is
$k$-relative projective as a left $A$-module.
\end{prop}
Composing $\nabla_A$ with the noncommutative De Rham
differential
$$
\operatorname d \colon A \longrightarrow
\Omega^1_{A,k},\quad
a \longmapsto 1 \otimes a-a \otimes 1
$$
gives the notion of connection in noncommutative
geometry \cite[III.3.5]{MR1303779}.
If $N$ is not just $k$-relative projective but
$k$-relative free, i.e.\
$N \cong A \otimes V$ as left $A$-modules, for some
$k$-module $V$, then the assignment $\nabla_M ( m \otimes_A (a \otimes v) ) = ma \otimes (1 \otimes v)$ defines a flat connection. Explicitly, the flat connection $\sigma \colon\tilde B \Sigma \Rightarrow \Sigma \tilde B$ is given by
\begin{align*}
\sigma_M \colon (M \otimes_A N) \otimes A &\longrightarrow (M \otimes A) \otimes_A N \cong M \otimes N \\
(m \otimes_A n) \otimes b &\longmapsto \nabla_M (m \otimes_A n)b.
\end{align*}
Thus we have:
\begin{prop}\label{bimoduletwist}
The triple
$(\Sigma, \sigma, 1)$ defines a 1-cell
$$
\xymatrix{
(\mathcal{A}^B, \tilde\theta, \tilde B, C^\theta) \ar[rr]^-{(\Sigma, \sigma, 1)} & & (\mathcal{A}^B, \tilde\theta, \tilde B, C^\theta)
}
$$
in $\mathbf{Dist}$.
\end{prop}
In particular, let $ \sigma \colon A \to A$ be an
algebra map and let $N=A_\sigma$, the $(A,A)$-bimodule
which is $A$ as a left $A$-module with right action of $a
\in A$ given by right multiplication by $ \sigma
(a)$.
Then we have
$\Sigma (M) = M \otimes_A A_\sigma \cong M_\sigma $.
Since $A_\sigma$ is free as a left $A$-module we get a 1-cell
$(\Sigma, \sigma,1)$ by Proposition~\ref{bimoduletwist},
where $\sigma \colon \tilde{B} \Sigma \Rightarrow \Sigma\tilde B$ is the flat connection defined on components by
\begin{align*}
\sigma_M \colon M_\sigma \otimes_k A & \longrightarrow (M \otimes_k A)_\sigma \\
m \otimes a &\longmapsto m \otimes \sigma(a).
\end{align*}
Note that we use $ \sigma $ to denote both the algebr
map and the flat connection it induces.
\subsection{Twisting by 1-cells}
From the general theory developed in
Section~\ref{twistcoeff} we obtain therefore an action of
the group of endomorphisms of $A$ on the categories $\mathscr{R}(\tilde\theta)$ and $\mathscr{L}(\tilde\theta)$ of $\tilde\theta$-coalgebras and $\tilde\theta$-opcoalgebras, respectively.
In particular, we can act
on the standard cyclic object $C_\bullet(A,A)$ associated to $A$, which corresponds to
the following data.
Consider $A$ as an object of $\mathcal{A}^B$. Since $\tilde B A = C^\theta A = A \otimes A$ we have a morphism $\rho = 1 \colon \tilde B A \to C^\theta A$. It follows that $(A, \rho)$ is a $\tilde\theta$-coalgebra in $\mathcal{A}^B$.
Considering $(A,A)$-bimodules as either left or
right
${A^\mathrm{e}} = A \otimes A^*$-modules, we view the zeroth
Hochschild homology as a functor
$H = - \otimes_{{A^\mathrm{e}}} A \colon
\mathcal{A} ^ B \to k\mbox{-}\mathbf{Mod}$. Then, as in Section~\ref{hochclassic}, we have an isomorphism of simplicial objects
$$
C_\bullet(A,A) \cong \mathrm C_{\tilde B}(H, A).
$$
We define a natural transformation
$\lambda \colon H C^\theta \Rightarrow H \tilde B$ by
\begin{align*}
\lambda_M \colon ( A\otimes M) \otimes_{{A^\mathrm{e}}} A &\longrightarrow (M \otimes A )\otimes_{{A^\mathrm{e}}} A \cong M \\
(a \otimes m) \otimes_{{A^\mathrm{e}}} b &\longmapsto mba
\end{align*}
Therefore, by Theorem~\ref{dup}, $C_\bullet(A,A)$ becomes a duplicial $k$-module, and indeed is
the cyclic object defining the cyclic homology
$\operatorname{HC}(A)$.
By acting on the $\tilde\theta$-coalgebra $M$ with the $1$-cell $(\Sigma, \sigma, 1)$ and applying Theorem~\ref{dup}, we obtain a duplicial structure on the simplicial $k$-module
$$
C_\bullet(A,A)_\sigma
$$
whose cyclic homology is the $\sigma$-twisted cyclic homology of $A$.
Thus the action of 1-cells in $\mathbf{Dist}$
generalises this twisting procedure.
\section{Hopf-cyclic homology}\label{brugsec}
In this section we assume that the monoidal categories mentioned and their actions on other categories are strict, with the symbols
$
I, \otimes, \rhd
$
used to denote the unit, tensor product, and left action respectively.
\subsection{Opmonoidal adjunctions}\label{notsure}
Let $\mathcal{E}, \mathcal{H}$ be monoidal categories.
\begin{defn}
An \emph{opmonoidal functor} $\mathcal{E} \to \mathcal{H}$ is a triple $(P, \Xi, \Xi_0)$ where $P \colon \mathcal{E} \to \mathcal{H}$ is a functor,
$$\Xi \colon P({-}_{1} \otimes {-}_{2} ) \Rightarrow P({-}_{1}) \otimes P({-}_{2})$$
is a natural transformation, and $\Xi_0 \colon PI \to I$ is a morphism, such that the three diagrams
$$\xymatrix{
P(X \otimes Y \otimes Z) \ar[rr]^-{\Xi} \ar[d]_-{\Xi} && P(X \otimes Y) \otimes P(Z) \ar[d]^-{\Xi \otimes PZ} \\
PX \otimes P(Y \otimes Z) \ar[rr]_-{PX \otimes {\Xi}} && PX \otimes PY \otimes PZ
}$$
$$
\xymatrix{
P(X \otimes I) \ar[r]^-\Xi \ar@{=}[dr] & PX \otimes PI \ar[d]^-{PX \otimes \Xi_0} \\
& PX
} \qquad
\xymatrix{
P(I \otimes X) \ar[r]^-\Xi \ar@{=}[dr] & PI \otimes PX \ar[d]^-{\Xi_0 \otimes PX} \\
& PX
}
$$
commute.
\end{defn}
Let $(P, \Xi, \Xi_0)$ and $(G, \Phi, \Phi_0)$ be two monoidal functors $\mathcal{E} \to \mathcal{H}$.
\begin{defn}
An \emph{opmonoidal natural transformation} $\alpha \colon (P, \Xi, \Xi_0) \Rightarrow (G, \Phi, \Phi_0)$ is a natural
transformation $\alpha \colon P \Rightarrow G$ such that the two diagrams
$$
\xymatrix{
P(A \otimes B) \ar[r]^-{\Xi} \ar[d]_-\alpha & PA \otimes PB \ar[d]^-{\alpha \otimes \alpha} \\
G(A \otimes B) \ar[r]_-{\Phi} & GA \otimes GB
}\qquad
\xymatrix{
PI \ar[r]^-{\alpha} \ar[dr]_-{\Xi_0} & GI \ar[d]^-{\Phi_0} \\
& I
}
$$
commute.
\end{defn}
This defines a 2-category $\mathbf{OpMonCat}$ of monoidal categories, opmonoidal functors, and opmonoidal natural transformations~\cite[Example~2.4]{MR2664622}. One example of Corollary~\ref{arisec} is provided by an adjunction in this 2-category. Explicitly:
\begin{defn}
An \emph{opmonoidal adjunction}
$$
\xymatrix{ \mathcal{E}
\ar@/^{0.5pc}/[rr]^-{(P, \Xi, \Xi_0)} \ar@{}[rr]|-{\perp}&&
\ar@/^0.5pc/[ll]^-{(Q, \Psi, \Psi_0)} \mathcal{H}
}
$$
consists of an adjunction $P \dashv Q$ between the underlying categories, such that the unit and counit are opmonoidal natural transformations.
\end{defn}
Opmonoidal adjunctions are a special case of doctrinal adjunctions, so it follows that $\Psi$ and $\Psi_0$ as above are in fact isomorphisms~\cite[Theorem~1.4]{MR0360749}. Some authors call opmonoidal adjunctions \emph{comonoidal adjunctions} or
\emph{bimonads}. We refer {e.g.} to~\cite{MR3020336,MR2793022,MR1942328,MR3175323,MR1887157}
for more information.
It follows that
$$
PI \otimes {-}
\qquad QPI
\otimes {-}
$$
form a compatible pair of comonads
as in Corollary~\ref{arisec}
whose comonad structures are induced by the natural
coalgebra (comonoid) structures on $I$.
\subsection{Opmodule adjunctions}\label{opmoduleadj}
The examples we are more interested in are
given by opmodule adjunctions, as defined below. They were introduced under the
name comodule adjunctions in
\cite[Definition~4.1.1]{MR3020336}.
Let $(P, \Xi, \Xi_0) \colon \mathcal{E}
\rightarrow \mathcal{H}$ be an
opmonoidal functor and let $\mathcal{A}$ and
$\mathcal{B}$ be (left)
$\mathcal{E}$-module $\mathcal{H}$-module categories, respectively.
\begin{defn}
A $(P, \Xi, \Xi_0)$-\emph{opmodule} is a functor
$F \colon \mathcal{A} \rightarrow \mathcal{B}$ together with a
natural transformation
$$
\Theta \colon F({-}_1 \rhd {-}_2) \Rightarrow
P ({-}_1) \rhd F({-}_2)
$$ such that the diagrams
$$
\xymatrix{
\mathrm F((X \otimes Y) \rhd Z)\ar@{=}[d]
\ar[rr]^-\Theta & & P(X \otimes Y) \rhd FZ \ar[d]^-{\Xi \otimes FZ} \\
F(X \rhd (Y \rhd Z) ) \ar[d]_-\Theta & & (PX \otimes PY ) \rhd FZ \ar@{=}[d] \\
PX \rhd F(Y \rhd Z) \ar[rr]_-{PX \rhd \Theta} & & PX \rhd (PY \rhd FZ) \\
}
$$
and
$$
\xymatrix{
F(I \rhd Z) \ar[r]^-\Theta
\ar@{=}[d] &
PI \rhd F Z \ar[d]^-{\Xi_0
\rhd FZ}\\
FZ \ar@{=}[r] &
I \rhd F Z}
$$
commute.
\end{defn}
Let $(P, \Xi, \Xi_0)$ and $(G, \Phi, \Phi_0)$ be two opmonoidal functors $\mathcal{H} \to \mathcal{E}$, and let $(F, \Theta)$ and $(K, \kappa)$ be two opmodules $\mathcal{A} \to \mathcal{B}$ over $(P, \Xi, \Xi_0)$ and $(G, \Phi, \Phi_0)$ respectively.
\begin{defn}
An \emph{opmodule morphism} $(\alpha, \beta) \colon (F, \Theta) \Rightarrow (K, \kappa)$ consists of an opmonoidal natural transformation
$$\alpha \colon (P, \Xi, \Xi_0) \Rightarrow (G, \Phi, \Phi_0) $$
and a natural transformation
$$
\beta \colon F \Rightarrow K
$$
such that the diagram
$$
\xymatrix{
F(X \rhd Z) \ar[r]^-{\Theta} \ar[d]_-\beta \ar[r]^-{\Theta} & PX \rhd FZ \ar[d]^-{\alpha \rhd \beta} \\
K(X \rhd Z) \ar[r]_-{\kappa} & GX \rhd KZ
}
$$
commutes.
\end{defn}
This defines a 2-category $\mathbf{OpMod}$ whose 0-cells are actions of a monoidal category on another category, whose 1-cells are opmodules over opmonoidal functors, and whose 2-cells are opmodule morphisms~\cite[Remark~4.3]{MR3020336}. We now give an explicit definition of adjunctions in this 2-category:
\begin{defn}
An \emph{opmodule
adjunction}
$$
\xymatrix{ \mathcal{A}
\ar@/^{0.5pc}/[rr]^-{(F,\Theta)} \ar@{}[rr]|-{\perp}&&
\ar@/^0.5pc/[ll]^-{(U, \Omega)} \mathcal{B}
}
$$
over an opmonoidal adjunction
$$
\xymatrix{ \mathcal{E}
\ar@/^{0.5pc}/[rr]^-{(P, \Xi, \Xi_0)} \ar@{}[rr]|-{\perp}&&
\ar@/^0.5pc/[ll]^-{(Q, \Psi, \Psi_0)} \mathcal{H}
}
$$
is an adjunction $F
\dashv U$ such that
$$
\xymatrix{
X \rhd Z
\ar[d]_-{\eta}
\ar[rr]^-{\eta \rhd \eta}
& &
QP X \rhd U F Z \\
UF (X\rhd Z) \ar[rr]_-{U \Theta} & &
U(P X \rhd F Z) \ar[u]_-{\Omega}}
$$
and
$$
\xymatrix{
P Q L \rhd FU M
\ar[rr]^-{\varepsilon \rhd
\varepsilon} & &
L \rhd M\\
F(Q L \rhd U M) \ar[u]^\Theta & &
F U (L \rhd M) \ar[ll]^-{F \Omega}
\ar[u]_-{\varepsilon}}
$$
commute.
\end{defn}
By the theory of doctrinal adjunctions, it follows that
$\Omega$ is an isomorphism
(see \cite[Proposition~4.1.2]{MR3020336} and again
\cite[Theorem~1.4]{MR0360749}).
Now any coalgebra $D$ in $\mathcal{H}$
defines a compatible pair of comonads
$$
S= D \rhd{-},\quad
C=Q D \rhd{-}$$
on $\mathcal{B}$
and $\mathcal{A}$ respectively. It is such an
instance of Corollary~\ref{arisec} that provides the
monadic generalisation of the setting from
\cite{MR2803876}, see Section~\ref{coeffsforhopf}.
\subsection{Bialgebroids and Hopf algebroids}
Opmonoidal
adjunctions can be seen as categorical
generalisations of bialgebras and more generally
(left) bialgebroids. We briefly recall the definitions
but refer to \cite{MR2553659,MR1458415,MR2803876,MR1984397} for further
details and references.
\begin{defn}
If $E$ is a $k$-algebra,
then an \emph{$E$-ring} is a $k$-algebra map $ \eta : E
\rightarrow H$.
\end{defn}
In particular, when $E={A^\mathrm{e}}:=A \otimes A^*$ is the
enveloping algebra of a $k$-algebra $A$, then
$H$ carries two left actions $\smalltriangleright, \blacktriangleright$ and two right actions $\smalltriangleleft, \blacktriangleleft$ of $A$, given by
$$
a \smalltriangleright h \smalltriangleleft
b:=\eta(a \otimes b)h,\quad
a \blacktriangleright h \blacktriangleleft b:=
h \eta(b \otimes a).
$$
These actions give rise to four $A$-bimodule structures on $H$, and so we use the actions as subscripts to make clear which of these structures is under discussion. For example, ${}_\blacktriangleright H_{\smalltriangleleft}$ denotes the $A$-bimodule $H$, with left $A$-action $a \blacktriangleright h$ and right $A$-action $h \smalltriangleleft b$.
Recall that ${A^\mathrm{e}}\mbox{-}\mathbf{Mod}$ is a monoidal category, with tensor product $\otimes_A$ and unit $A$.
\begin{defn}[see~\cite{MR0506407}]
A \emph{bialgebroid} is an ${A^\mathrm{e}}$-ring
$ \eta : {A^\mathrm{e}} \rightarrow H$
for which ${}_\smalltriangleright H_\smalltriangleleft$
is a coalgebra in ${A^\mathrm{e}}\mbox{-}\mathbf{Mod}$
whose coproduct
$
\Delta \colon {}_\smalltriangleright H_\smalltriangleleft \rightarrow {}_\smalltriangleright H_\smalltriangleleft \otimes_A
{}_\smalltriangleright H_\smalltriangleleft
$
satisfies
$$
a \blacktriangleright \Delta(h) =\Delta (h) \blacktriangleleft a,\quad
\Delta
(gh)=\Delta(g)\Delta(h),
$$
and whose counit
$
\varepsilon \colon {}_\smalltriangleright H_\smalltriangleleft \rightarrow A
$
defines a unital $H$-action on $A$ given by
$h(a):=\varepsilon (a \blacktriangleright h)$.
\end{defn}
Finally, by a Hopf algebroid we mean \emph{left}
rather than \emph{full} Hopf algebroid, so there is
in general no antipode \cite{3}:
\begin{defn}[see~\cite{MR1800718}]\label{hadef}
A \emph{Hopf algebroid} is a bialgebroid with bijective
\emph{Galois map}
$$
\beta \colon {}_\blacktriangleright H_\smalltriangleleft \otimes_{A^*}
{}_\smalltriangleright H_\smalltriangleleft \rightarrow {}_\blacktriangleright H_\smalltriangleleft \otimes_A {}_\smalltriangleright H_\smalltriangleleft,\quad
g \otimes_{A^*} h \mapsto \Delta (g)h. $$ \end{defn}
As usual, we abbreviate
\begin{equation}
\label{hunger}
\Delta (h) =: h_{(1)} \otimes_A
h_{(2)},\qquad
\beta^{-1}(h \otimes_A 1) =: h_+
\otimes_{A^*} h_-.
\end{equation}
This symbolic notation we use here resembles \emph{Sweedler notation} for Hopf algebras (see e.g.~\cite[\S 1.4]{MR1243637}), and is fully explained in
~\cite[\S 2.3]{MR3281654}.
\subsection{The opmonoidal adjunction}\label{bimfromhopf}
Every $E$-ring $H$ defines a forgetful functor
$$
Q \colon H\mbox{-}\mathbf{Mod} \rightarrow E\mbox{-}\mathbf{Mod}
$$
with left
adjoint $P=H \otimes_E -$.
In the next section, we abbreviate
$\mathcal{H}:=H\mbox{-}\mathbf{Mod}$ and $\mathcal{E}:=E\mbox{-}\mathbf{Mod}$.
If $H$ is a bialgebroid (so $E = {A^\mathrm{e}}$)
then $\mathcal{H}$ is monoidal with tensor product
$K \otimes_\mathcal{H} L$ of
two left $H$-modules $K$ and $L$ given by the tensor
product $K \otimes_A L$ of the underlying
$A$-bimodules whose $H$-module structure is given by
$$
h(k \otimes_\mathcal{H} l):=
h_{(1)}(k) \otimes_A h_{(2)}(l).
$$
So by definition, we have $Q(K \otimes _\mathcal{H} L) =
Q K \otimes_A Q L$. The opmonoidal structure $\Xi$
on $P$ is defined by the map \cite{MR2793022,MR3020336}
\begin{align*}
P(X \otimes_A Y)=H \otimes_{A^\mathrm{e}} (X \otimes_A Y)
&\rightarrow P X \otimes_\mathcal{H} P Y= (H \otimes_{A^\mathrm{e}}
X) \otimes_A (H \otimes_{A^\mathrm{e}} Y), \\
h \otimes_{A^\mathrm{e}} (x \otimes_A y) &\mapsto
(h_{(1)} \otimes_{A^\mathrm{e}} x) \otimes_A
(h_{(2)} \otimes_{A^\mathrm{e}} y).
\end{align*}
Schauenburg proved that this establishes a bijective
correspondence between bialgebroid structures on $H$
and monoidal structures on $H\mbox{-}\mathbf{Mod}$
\cite[Theorem~5.1]{MR1629385}:
\begin{thm}
The following data are equivalent for an ${A^\mathrm{e}}$-ring
$ \eta \colon {A^\mathrm{e}} \rightarrow H$:
\begin{enumerate}
\item A bialgebroid structure on $H$.
\item A monoidal structure on ${H\mbox{-}\mathbf{Mod}}$
such that the adjunction
$$
\xymatrix{{A^\mathrm{e}}\mbox{-}\mathbf{Mod} \hspace{-10mm}&
\ar@/^2mm/[r] & \ar@/^2mm/[l] & \hspace{-10mm}
H\mbox{-}\mathbf{Mod}}
$$
induced by $ \eta $ is
opmonoidal.
\end{enumerate}
\end{thm}
Consequently,
we obtain an opmonoidal monad
$$
QP={}_\blacktriangleright H_\blacktriangleleft
\otimes_{A^\mathrm{e}} -
$$
on $\mathcal{E}={A^\mathrm{e}}\mbox{-}\mathbf{Mod}$. This
takes the unit object $I = A$ to the cocentre $H
\otimes_{A^\mathrm{e}} A$ of the $A$-bimodule ${}_\blacktriangleright
H_\blacktriangleleft$, and the comonad $PI
\otimes_A -$ is given by
$$
(H \otimes_{A^\mathrm{e}} A) \otimes_A -,
$$
where the $A$-bimodule structure on the
cocentre is given by the actions
$\smalltriangleright,\smalltriangleleft$ on $H$.
The lift to $\mathcal{H}={H\mbox{-}\mathbf{Mod}}$ takes a left
$H$-module $L$ to
$
(H \otimes_{A^\mathrm{e}} A)
\otimes_A L
$
with action
$$
g \cdot ((h \otimes_{A^\mathrm{e}} 1) \otimes_A l)=
(g_{(1)}h \otimes_{A^\mathrm{e}} 1) \otimes_A g_{(2)}l,
$$
and the
distributive law resulting from Corollary~\ref{arisec} is
given by
$$
\chi \colon g \otimes_{A^\mathrm{e}} ((h \otimes_{A^\mathrm{e}} 1)
\otimes_A l) \mapsto (g_{(1)}h \otimes_{A^\mathrm{e}} 1) \otimes_A
(g_{(2)} \otimes_{A^\mathrm{e}} l).
$$
That is, it is the map
induced by the \emph{Yetter-Drinfel'd braiding}
$$
H_\blacktriangleleft \otimes_A {}_\smalltriangleright H \to H_\smalltriangleleft \otimes_A
{}_\smalltriangleright H ,\quad g \otimes_A h \mapsto g_{(1)}h
\otimes_A g_{(2)}.
$$
For $A=k$, that is, when $H$ is a Hopf algebra, and
also trivially when $H={A^\mathrm{e}}$, the monad and the comonad
on ${A^\mathrm{e}}\mbox{-}\mathbf{Mod}$ coincide and are also a bimonad in the sense of
Mesablishvili and Wisbauer,
cf.\ Section~\ref{wisbimonad}. An example where the two
are different is the Weyl algebra, or more generally,
the universal enveloping algebra of a
Lie-Rinehart algebra \cite{MR1625610}. In these
examples, $A$ is commutative but
not central in $H$ in general.
\subsection{Doi-Koppinen data}
The instance of Corollary~\ref{arisec}
that we are most interested in is an opmodule
adjunction associated to the following structure:
\begin{defn}
Following e.g.~\cite{MR1877862}, a \emph{Doi-Koppinen datum} is a triple
$(H,D,W)$ of an $H$-module coalgebra $D$ and an $H$-comodule
algebra $W$ over a bialgebroid $H$.
\end{defn}
This means that $D$ is a coalgebra in
the monoidal category ${H\mbox{-}\mathbf{Mod}}$.
Similarly, the category $H\mbox{-}\mathbf{Comod}$ of left $H$-comodules is
also monoidal (see {e.g.}~\cite[Section~3.6]{MR2553659}), and this defines the notion of a
comodule algebra. Explicitly, $W$ is an $A$-ring
$
\eta_W \colon A \rightarrow W
$
together with a coassociative coaction
$$
\delta \colon W \rightarrow H_\smalltriangleleft \otimes_A
W,\quad b
\mapsto b_{(-1)} \otimes_A b_{(0)},
$$
which is counital and an algebra map,
$$
\eta _W (\varepsilon
(b_{(-1)}))b_{(0)}=b,\quad
(b d)_{(-1)} \otimes (b d)_{(0)} =
b_{(-1)} d_{(-1)}
\otimes b_{(0)} d_{(0)}.
$$
Here again we use a Sweedler-esque notation to denote the coaction, as in~\cite[\S 2.5]{MR1243637}. Similarly, as in the
definition of a bialgebroid itself, for this condition
to be well-defined one must also require
$$
b_{(-1)}
\otimes_A b_{(0)} \eta_W(a)= a \blacktriangleright b_{(-1)}
\otimes_A b_{(0)}.
$$
The key example that reproduces
\cite{MR2803876} is the following.
\subsection{The opmodule
adjunction}\label{associatedsec}
For any Doi-Koppinen datum $(H,D,W)$, the
$H$-coaction $ \delta $ on $W$ turns the Eilenberg-Moore adjunction
\!\!
$\xymatrix{{A\mbox{-}\mathbf{Mod}} \ar@/^{0.3pc}/[r] &
\ar@/^0.3pc/[l] {W\mbox{-}\mathbf{Mod}}}$
\!\!
for the monad $ B:=W \otimes_A - $ into an opmodule
adjunction for the opmonoidal adjunction
$\xymatrix{\mathcal{E} \ar@/^{0.3pc}/[r]
& \ar@/^0.3pc/[l] \mathcal{H}}$ defined in
Section~\ref{bimfromhopf}. The $\mathcal{H}$-module
category structure of $W\mbox{-}\mathbf{Mod}$ is given by the left
$W$-action
$$
b(l \otimes_A m):=
b_{(-1)}l \otimes_A
b_{(0)}m,
$$
where $b \in W$,
$l \in L$ (an $H$-module), and $m \in
M$ (a $W$-module).
Hence, as explained in Section~\ref{opmoduleadj},
$D$ defines a compatible pair of
comonads $ D \otimes_A - $ on ${W\mbox{-}\mathbf{Mod}}$ and ${A\mbox{-}\mathbf{Mod}}$. The
distributive law resulting from Corollary~\ref{arisec}
generalises the Yetter-Drinfel'd braiding, as it is
given for a $W$-module $M$ by
\begin{eqnarray*}
\chi \colon W \otimes_A (D
\otimes_A M) &\rightarrow& D \otimes_A (W \otimes_A M),
\\
b \otimes_A (c \otimes_A m) &\mapsto& b_{(-1)}c
\otimes _A (b_{(0)} \otimes_A m).
\end{eqnarray*}
\subsection{The main example}\label{coeffsforhopf}
If $H$ is a bialgebroid, then
$D:=H$ is a module coalgebra with
left action given by multiplication and coalgebra
structure given by that of $H$.
If $H$
is a Hopf algebroid, then $W:=H^*$ is
a comodule algebra with unit map $\eta_W(a):=\eta(1
\otimes_k a)$ and coaction
$$
\delta \colon H^* \rightarrow
H_\smalltriangleleft \otimes_A {}_\blacktriangleright H^*, \quad
b \mapsto b_- \otimes_A b_+.
$$
In the sequel we write $B$ as $-
\otimes_{A^*} H$ rather than $H^* \otimes_A -$
to work with $H$ only. Then the distributive law
becomes
\begin{eqnarray*}
\nonumber
\chi \colon (H \otimes_A M) \otimes_{A^*} H
&\rightarrow& H \otimes_A (M \otimes_{A^*} H),
\\
(c \otimes_A m) \otimes_{A^*} b &\mapsto& b_- c \otimes _A (m
\otimes_{A^*} b_+),
\end{eqnarray*}
for $b, c \in H$.
Proposition~\ref{chicoalgprop} completely
characterises the $ \chi $-coalgebras: in this
example, they are given by right $H$-modules and left
$H$-comodules $M$ with $ \chi $-coalgebra
structure
$$
\rho: m \otimes_{A^*} h \mapsto
h_-m_{(-1)}
\otimes_A m_{(0)} h_+.
$$
In general, the characterisation of
$ \chi $-opcoalgebras mentioned after Proposition~\ref{chicoalgprop}
does not provide us with such an explicit description.
Note, however, that
one obtains $ \chi $-opcoalgebras from
(left-left) Yetter-Drinfel'd modules:
\begin{defn}
A \emph{Yetter-Drinfel'd module} over $H$ is a left
$H$-comodule and left $H$-module $N$ such that
for all $ h\in H,n \in N$, one has
$$
(hn)_{(-1)} \otimes_A (hn)_{(0)}=
h_{+(1)} n_{(-1)} h_{-} \otimes_A
h_{+(2)}n_{(0)}.
$$
\end{defn}
Indeed, each such Yetter-Drinfel'd
module defines a $ \chi$-opcoalgebra
$$
- \otimes_H N \colon H^*\mbox{-}\mathbf{Mod} \rightarrow
k\mbox{-}\mathbf{Mod}
$$
whose $ \chi $-opcoalgebra structure is given by
\begin{equation*}
\lambda: (h \otimes_A x) \otimes_H n
\mapsto (xn_{(-1)+} h_+ \otimes_{A^*} h_- n_{(-1)-})
\otimes_H n_{(0)}.
\end{equation*}
The resulting duplicial object
$\rCC_T({-}\otimes_H N,M)$ is the one
studied in \cite{MR2803876, Kow:GABVSOMOO}.
Identifying
$(- \otimes_{A^*} H) \otimes_H N \cong - \otimes_{A^*} N$,
the $ \chi $-opcoalgebra structure
becomes
\begin{equation*}
\lambda: (h \otimes_A x) \otimes_H n
\mapsto
xn_{(-1)+} h_+ \otimes_{A^*} h_- n_{(-1)-} n_{(0)}.
\end{equation*}
Using this identification, we give
explicit expressions of the operators $L$ and $R$
as well as $t_\mathbb{T}$ that appeared in
Sections~\ref{evidenziatore1} and~\ref{evidenziatore2}:
first of all,
observe that the right $H$-module structure on
$
S M := {}_\smalltriangleright H_\smalltriangleleft \otimes_A M
$
is given by
$$
(h \otimes_A m)g :=
g_- h \otimes_A mg_+,
$$
whereas the right $H$-module structure on
$
T M := M \otimes_{A^*} {}_\blacktriangleright H_\smalltriangleleft
$
is given by
$$
(m \otimes_{A^*} h)g :=
m \otimes_{A^*} hg.
$$
The cyclic operator from Section \ref{evidenziatore1}
then results as
\begin{equation*}
\begin{split}
&t_T
(m \otimes_{A^*} h^1 \otimes_{A^*} \cdots \otimes_{A^*} h^n \otimes_{A^*} n)
\\
&=
m_{(0)} h^1_+ \otimes_{A^*} h^2_+ \otimes_{A^*} \cdots \otimes_{A^*} h^n_+ \\
& \qquad
\otimes_{A^*} (n_{(-1)} h^n_- \cdots h^1_- m_{(-1)})_+
\otimes_{A^*} (n_{(-1)} h^n_- \cdots h^1_- m_{(-1)})_-
n_{(0)},
\end{split}
\end{equation*}
and for the operators $L$ and $R$ from Section~\ref{evidenziatore2}
one obtains with the help of the properties
\cite[Prop.~3.7]{MR1800718} of the translation map
(\ref{hunger}):
\begin{equation*}
\begin{split}
L: (h^1 & \otimes_A \cdots \otimes_A h^{n+1}
\otimes_A m) \otimes_H n
\mapsto \\
& (mn_{(-1)+} h^1_+ \otimes_{A^*}
h^1_- h^2_+ \otimes_{A^*} \cdots \otimes_{A^*} h^{n+1}_- n_{(-1)-})
\otimes_H n_{(0)},
\end{split}
\end{equation*}
along with
\begin{equation*}
\begin{split}
R: (m & \otimes_{A^*} h^1
\otimes_{A^*} \cdots \otimes_{A^*} h^n \otimes_{A^*} 1)
\otimes_H n
\mapsto \\
&
(m_{(-n-1)} \otimes_A m_{(-n)}h^1_{(1)} \otimes_A
m_{(-n+1)}h^1_{(2)}h^2_{(1)} \otimes_A \cdots \\
& \quad
\otimes_A m_{(-1)} h^1_{(n)}
h^2_{(n-1)} \cdots h^n_{(1)}
\otimes_A m_{(0)}) \otimes_H h^1_{(n+1)} h^2_{(n)} \cdots h^n_{(2)} n.
\end{split}
\end{equation*}
Compare these maps with those obtained in \cite[Lemma~4.10]{MR2803876}.
Hence, one has:
\begin{equation*}
\begin{split}
(L &\circ R)\big((m \otimes_{A^*} h^1
\otimes_{A^*} \cdots \otimes_{A^*} h^n \otimes_{A^*} 1)
\otimes_H n\big)
= \\
&
m_{(0)} (h^1_{(n+1)}h^2_{(n)} \cdots h^n_{(2)} n)_{(-1)+} m_{(-n-1)+}
\otimes_{A^*} m_{(-n-1)-} m_{(-n)+} h^1_{(1)+}
\\
&
\qquad
\otimes_{A^*}
h^1_{(1)-} m_{(-n)-} m_{(-n+1)+} h^1_{(2)+} h^2_{(1)+} \otimes_{A^*}
\cdots
\\
&
\qquad
\otimes_{A^*} h^n_{(1)-} \cdots h^1_{(n)-} m_{(-1)-} (h^1_{(n+1)} \cdots h^n_{(2)} n)_{(-1)-} (h^1_{(n+1)} \cdots h^n_{(2)} n)_{(0)}
\\
&
=
m_{(0)} \big((h^1_{(2)} \cdots h^n_{(2)} n)_{(-1)} m_{(-1)}\big)_+
\otimes_{A^*} h^1_{(1)+} \otimes_{A^*} \cdots
\\
&
\quad
\otimes_{A^*} h^n_{(1)+}
\otimes_{A^*} h^n_{(1)-} \cdots h^1_{(1)-} \big((h^1_{(2)} \cdots h^n_{(2)} n)_{(-1)} m_{(-1)} \big)_- (h^1_{(2)} \cdots h^n_{(2)} n)_{(0)}.
\end{split}
\end{equation*}
Finally, if $M \otimes_{A^*} N$ is a stable anti
Yetter-Drinfel'd module \cite{MR2415479},
that is, if
$$
m_{(0)}(n_{(-1)}m_{(-1)})_+ \otimes_{A^*} (n_{(-1)}m_{(-1)})_- n_{(0)} = m \otimes_{A^*} n
$$
holds for all $n \in N$, $m \in M$,
we conclude by observing that
\begin{equation*}
\begin{split}
(L \circ R)(m & \otimes_{A^*} h^1 \otimes_{A^*} \cdots
\otimes_{A^*} h^n \otimes_{A^*} n)
\\
&
=
m \otimes_{A^*} h^1_{(1)+} \otimes_{A^*} \cdots \otimes_{A^*} h^n_{(1)+}
\otimes_{A^*} h^n_{(1)-} \cdots h^1_{(1)-} h^1_{(2)} \cdots h^n_{(2)} n
\\
&
=
m \otimes_{A^*} h^1 \otimes_{A^*} \cdots \otimes_{A^*} h^n
\otimes_{A^*} n.
\end{split}
\end{equation*}
Observe that in \cite{Kow:GABVSOMOO} this
cyclicity condition was obtained for a different complex
which, however, computes the same homology.
\subsection{The antipode as a $1$-cell}
\label{viviverde}
If $A=k$, then the four actions
$\smalltriangleright,\smalltriangleleft,\blacktriangleright,\blacktriangleleft$ coincide and
$H$ is a Hopf algebra with antipode
$S \colon H \rightarrow H$ given by
$S(h)=\varepsilon (h_+)h_-$. The
aim of this brief section is to remark that
this defines a 1-cell that connects the two instances
of Corollary~\ref{arisec} provided by the opmonoidal
adjunction and the opmodule adjunction considered
above.
Indeed, in this case we have ${A^\mathrm{e}}\mbox{-}\mathbf{Mod}\cong A\mbox{-}\mathbf{Mod}=k\mbox{-}\mathbf{Mod}$,
but (unless $H$ is commutative) $H^*\mbox{-}\mathbf{Mod} \neq H\mbox{-}\mathbf{Mod}$. However, $S$ defines
a morphism of monads
$$\xymatrix{
(k\mbox{-}\mathbf{Mod}, H \otimes -) \ar[rr]^-{(1, \sigma)} && (k\mbox{-}\mathbf{Mod}, - \otimes H)
}
$$ where $\sigma \colon{-} \otimes H \Rightarrow H \otimes{-}$ is given in components by
$$
\sigma \colon X \otimes H
\rightarrow H \otimes X, \quad
x \otimes h \mapsto S(h) \otimes x.
$$
The fact that this $(1, \sigma)$ is a morphism of monads is
equivalent to the fact that $S$ is an algebra
anti-homomorphism.
Also, the lifted comonads agree and are given by
$H \otimes{-}$ with comonad structure given by the
coalgebra structure of $H$;
clearly, $\gamma = 1
\colon H \otimes {-} \Rightarrow H \otimes {-}$
defines an opmorphism of monads
$$
\xymatrix{
(k\mbox{-}\mathbf{Mod}, H \otimes{-}) \ar[rr]^-{(1, \gamma)} & & (k\mbox{-}\mathbf{Mod}, H \otimes{-})
}
$$ Furthermore, the Yang-Baxter
condition is satisfied, so we have that $(1,
\sigma, \gamma)$ is a $1$-cell in the $2$-category of mixed
distributive laws. If we apply the $2$-functor of Corollary~\ref{2funcmixdist} to
this, we get a $1$-cell $(\Sigma, \tilde \sigma, \tilde
\gamma)$ between a comonad distributive law on the
category of left $H$-modules and one on the category of
right $H$-modules. The identity lifts to the functor
$\Sigma \colon H\mbox{-}\mathbf{Mod} \rightarrow
\mathbf{Mod}\mbox{-}H$ which sends a left $H$-module
$X$ to the right $H$-module with right action given by
$$
x \cdot h := S(h) x.
$$
\section{Hopf monads \`a la Mesablishvili-Wisbauer}\label{hopfywopfy}
\label{wisbimonad}
\subsection{Bimonads}
A \emph{bimonad} in the sense of
\cite[Def.~4.1]{MR2787298} is a sextuple
$(B,\mu,\eta,\delta, \varepsilon,\theta)$,
where $B
\colon \mathcal{A} \rightarrow \mathcal{A}$ is a functor, $(B, \mu,\eta)$ is
a monad, $(B,\delta,\varepsilon)$ is a comonad and
$\theta \colon BB \Rightarrow BB$ is a mixed distributive
law satisfying a list of compatibility conditions. A \emph{Hopf monad} as in \cite[Def.~5.2]{MR2787298} is a bimonad $B$ equipped with
a natural transformation $B \Rightarrow B$, called the \emph{antipode}, satisfying various compatibility conditions mirroring those
for Hopf algebras.
In particular, for a bimonad $B$, the multiplication $ \mu $ and comultiplication $ \delta $ are required to
be compatible in the sense that there is a commutative
diagram \begin{equation}\label{wisbauerdiagram}
\begin{array}{c} \xymatrix{BB \ar[d]_{B
{\delta}}
\ar[r]^\mu & B \ar[r]^{\delta} & BB\\
BBB
\ar[rr]_{\theta B} & & BBB \ar[u]_{B\mu}} \end{array}
\end{equation}
The other defining conditions govern the
compatibility between the unit and the counit with each
other and with $ \mu $ and $ \delta$ respectively, see
\cite{MR2787298} for the details.
It follows immediately that we also obtain an instance
of Corollary~\ref{arisec} in this situation: if we take
$\mathcal{B}=\mathcal{A}^B$ to be the Eilenberg-Moore
category of the monad $B$ as in
Section~\ref{extremalcase}, then the mixed distributive
law $\theta$ defines a lift
$W$ of the comonad
$B$ to $\mathcal{B}$.
Note that in general, neither $\mathcal{A}$ nor $\mathcal{B}$ need to
be monoidal, so $B$ is in general not an opmonoidal
monad. Conversely, recall that for the examples of
Corollary~\ref{arisec} obtained from opmonoidal monads,
$B$ need not equal $C$ as functors.
\subsection{Examples from bialgebras}\label{whygalois}
In the main
example of bimonads in the above sense, we in fact do
have $B=C$ and we are in the situation of
Section~\ref{bimfromhopf} for a bialgebra $H$ over
$A=k$. The commutativity of (\ref{wisbauerdiagram})
amounts to the fact that the coproduct is an algebra
map.
This setting provides an instance of
Proposition~\ref{sunshines} since there are two lifts
of $B=C$ from $\mathcal{A}={k\mbox{-}\mathbf{Mod}}$ to $\mathcal{B}={H\mbox{-}\mathbf{Mod}}$: the
canonical lift $S=T=FU$ which takes a left
$H$-module $L$ to the $H$-module $H \otimes L$ with
$H$-module structure given by multiplication in the
first tensor component, and the lift $W$ which takes
$L$ to $H \otimes L$ with $H$-action given by the
codiagonal action
$
g(h \otimes y)=
g_{(1)}h \otimes
g_{(2)}y,
$
that is, the one defining the monoidal
structure on $\mathcal{B}$.
In this example, the map $ \beta $ from
Proposition~\ref{wisga} is given by
$$
H \otimes L \rightarrow H \otimes
L,\quad
g\otimes y \mapsto g_{(1)} \otimes
g_{(2)}y
$$
which for $L=H$ is the Galois map
from Definition~\ref{hadef}. This is bijective for all
$L$ if and only if it is so for $L=H$, which is also
equivalent to $H$ being a Hopf algebra. However,
this Galois map should not just be viewed as a $k$-linear
map, but as a natural $H$-module morphism between the two
$H$-modules $ TL$ and $WL$, and this is
the natural transformation $ \Gamma ^{T,W}(1)$ from
Section~\ref{galoismapsct}.
As shown in \cite[Theorem~5.8(c)]{MR3320218}, this
characterisation of Hopf algebras in terms of the
bijectivity of the Galois map extends straightforwardly
to Hopf monads.
\subsection{An example not from bialgebras}\label{newbimonad}
Another
example of a bimonad is the \emph{nonempty list monad}
$L^+$ on $\cSet$ (see Example~\ref{nonemptylistmonad}), which assigns to a
set $X$ the set $L^+X$ of all nonempty lists
of elements in $X$, denoted $[x_1, \ldots, x_n]$. The mixed
distributive law
$$
\theta \colon L^+ L^+ \Rightarrow
L^+ L^+
$$
is defined as follows: given a list
$$
[ [ x_{1,1}, \ldots, x_{1,
n_1} ] , \ldots, [x_{m,1}, \ldots, x_{m, n_m}]]
$$
in
$L^+L^+ X$,
its image under $\theta$ is the list with $$
\sum_{i=1}^m n_i (m-i+1) $$ terms, given by
\begin{align*}
\Big[[x_{1,1}, x_{2,1}, x_{3,1} \ldots, x_{m,1}],
\ldots, [x_{1, n_1}, x_{2,1}, x_{3,1}, \ldots,
x_{m,1}]&, \\ [x_{2,1}, x_{3,1} \ldots, x_{m,1}],
\ldots, [x_{2, n_2}, x_{3,1}, \ldots x_{m,1} ]&,\\
\ldots&, \\ [ x_{m,1} ] , [ x_{m,2} ], \ldots,
[x_{m,n_m} ] \Big].&
\end{align*}
One verifies straightforwardly:
\begin{prop} $L^+$ becomes a bimonad on $\mathbf{Set}$ whose (monad) Eilenberg-Moore category is
$\cSet^{L^+} \cong \mathbf{SemiGrp}$, the
category of (nonunital) semigroups. \end{prop}
The second lift $W \colon \mathbf{SemiGrp \to SemiGrp}$ of the comonad $L^+$ that
one obtains from the bimonad structure is as follows. Given a semigroup $X$,
we have $W X = L^+ X$ as sets, but the binary
operation is given by \begin{align*} W X \times W X
&\rightarrow W X \\ [x_1, \ldots, x_m][y_1, \ldots, y_n] &:=
[x_1y_1, \ldots, x_my_1, y_1, \ldots, y_n]. \end{align*}
Following Proposition~\ref{chicoalgprop}, given a
semigroup $X$, the unit turns the underlying set of $X$ into an
$L^+$-coalgebra and hence we get a $
\chi$-coalgebra structure on $X$. Explicitly, $\rho \colon T X \rightarrow W X$ is given by
$$
\rho[ x_1, \ldots, x_n] = [x_1 \cdots x_n, x_2 \cdots
x_n, \ldots, x_n].
$$
The image of $\rho$ is known as
the \emph{left machine expansion} of $X$
\cite{MR745358}.
\begin{prop}
The only $\theta$-entwined algebra is the trivial semigroup $\emptyset$.
\end{prop}
\begin{proof}
An $L^+$-coalgebra structure $\beta \colon X \to L^+ X$ is equivalent to $X$ being a forest of at most countable height (rooted) trees, where each level may have arbitrary cardinality. The structure map $\beta$ sends $x$ to the finite list of predecessors of $x$. A $\theta$-entwined algebra is therefore such a forest, which also has the structure of a semigroup such that for all $x,y \in X$ with $\beta(y) = [y, y_1, \ldots, y_n]$ we have
$$
\beta(xy) = [xy, xy_1, \ldots, xy_n , y, y_1, \ldots, y_n].
$$
Let $X$ be a $\theta$-entwined algebra. If $X$ is non-empty, then there must be a root. We can multiply this root
with itself to generate branches of arbitrary height. Suppose that we have a
branch of height two; that is to say, an element $y \in X$ with $\beta(y) = [y,x]$
(so, in particular, $x \neq y$). Then $\beta(xy) = [xy, y]$, but
$\beta(xx) = [xx, xy, x, y]$. This is impossible since $x$ and $y$ cannot both be the predecessor of $xy$.
\end{proof}
\section{Enriched functor categories}\label{finally}
We conclude the main part of the thesis by showing how to construct a duplicial object that generalises some of the examples we have previously seen.
\subsection{Enriched categories}
Let $\mathcal{V}$ be a monoidal category.
\begin{defn}
A $\mathcal{V}$-\emph{category} $\mathcal{H}$ consists of
\begin{itemize}
\item a class $|\mathcal{H}|$ whose elements we call objects
\item for any $A, B \in |\mathcal{H}|$, an object $\mathcal{H}(A, B) \in |\mathcal{V}|$
\item for each object $A \in |\mathcal{H}|$, a morphism $u_\mathcal{A} \colon I \to \mathcal{H}(A, A)$ in $\mathcal{V}$, called the \emph{unit}
\item for any $A, B, C$ in $|\mathcal{H}|$, a morphism $\circ_{A,B,C} \colon \mathcal{H}(B, C) \otimes \mathcal{H}(A, B) \to \mathcal{H}(A, C)$, called
\emph{composition}
\end{itemize}
satisfying associativity and unitality conditions, that is, commutativity of the two diagrams
\begin{align*}
\xymatrix{
\mathcal{H}({}C, {}D) \otimes \mathcal{H}({}B, {}C) \otimes \mathcal{H}({}A, {}B)\ar[d]_-{\circ_{B,C,D} \otimes 1} \ar[rr]^-{1 \otimes {\circ_{A,B,C}}} && \mathcal{H}({}C, {}D) \otimes \mathcal{H}({}A, {}C) \ar[d]^-{\circ_{A,C,D}} \\
\mathcal{H}({}B, {}D) \otimes \mathcal{H}({}A, {}B) \ar[rr]_-{\circ_{A,B,D}} && \mathcal{H}({}A, {}D)
} \\
\\
\xymatrix{
\mathcal{H}({}A, {}B) \ar@{=}[drr] \ar[rr]^-{u_B \otimes 1} \ar[d]_-{1 \otimes u} && \mathcal{H}({}B, {}B) \otimes \mathcal{H}({}A, {}B) \ar[d]^-{\circ_{A,B,B}} \\
\mathcal{H}({}A, {}B) \otimes \mathcal{H}({}A, {}A) \ar[rr]_-{\circ_{A,A,B}} & & \mathcal{H}({}A, {}B)
}
\end{align*}
for all ${}A, {}B, {}C, {}D \in |\mathcal{H}|$.
\end{defn}
\begin{exa}
A $\mathbf{Set}$-category is an ordinary category.
\end{exa}
\begin{exa}
A $\mathbf{Cat}$-category is a 2-category.
\end{exa}
We now suppose that $\mathcal{V}$ is a complete, cocomplete, closed symmetric monoidal category. Furthermore, we assume for simplicity that the monoidal structure of $\mathcal{V}$ is strict, but we make no assumption that the symmetry is strict. The tensor product is denoted by $\otimes$, the unit by $I$, and the closed structure is given by the functor
$$
[{-},{-}] \colon \mathcal{V}^* \times \mathcal{V} \to \mathcal{V}.
$$
Since left adjoint functors preserve limits, it follows that that the tensor product commutes with coproducts. That is, we have natural isomorphisms
$$
A \otimes \sum_{i \in I} B_i \cong \sum_{i \in I} A \otimes B_i
$$
where $\sum$ denotes the coproduct, and $\{B_i\}_{i \in I}$ is a family of objects in $\mathcal{V}$.
For any $\mathcal{V}$-categories $\mathcal{H}$ and $\mathcal K$, there is an associated (ordinary) $\mathcal{V}$-functor category $[\mathcal{H}, \mathcal K]$. This can also be given the structure of a $\mathcal{V}$-category, but we do not need it here and thus do not give any further details. We are interested particularly in the case that $\mathcal K = \mathcal{V}$. By definition, a $\mathcal{V}$-functor $F \colon \mathcal{H} \to \mathcal{V}$ consists of
\begin{itemize}
\item for each object $X$ in $\mathcal{H}$, an object $FX$ in $\mathcal{V}$
\item for any objects $X,Y$ in $\mathcal{H}$, a morphism $F_{X,Y} \colon \mathcal{H}(X,Y) \to [FX, FY]$ in $\mathcal{V}$
\end{itemize}
satisfying appropriate unitality and associativity axioms. We now omit subscripts and just write $F$ in place of $F_{X,Y}$.
Every morphism
$$
F\colon \mathcal{H}(X, Y) \to [FX, FY]
$$
corresponds to an \emph{action}
$$
\overline{F} \colon \mathcal{H}(X,Y) \otimes FX \to FY
$$
using the closed structure of $\mathcal{V}$,
and thus the axioms defining a $\mathcal{V}$-functor can be rewritten as the two commutative diagrams
$$
\xymatrix@C=3em{
\mathcal{H}(Y,Z) \otimes \mathcal{H}(X,Y) \otimes FX \ar[rr]^-{1 \otimes \overline F} \ar[d]_-{\circ \otimes 1} && \mathcal{H}(Y,Z) \otimes FY \ar[d]^-{\overline F} \\
\mathcal{H}(X,Z) \otimes FX \ar[rr]_-{\overline F}& & FZ
}
$$
$$
\xymatrix@C=3em{
FX \ar[rr]^-{u \otimes 1} \ar@{=}[drr] & & \mathcal{H}(X,X) \otimes FX \ar[d]^-{\overline F} \\
& & FX
}
$$
Similarly, a $\mathcal{V}$-natural transformation $\alpha \colon F \Rightarrow G$ is defined as a collection of morphisms $\alpha \colon FX \to GX$ in $\mathcal{V}$ such that the diagram
$$
\xymatrix@C=3em{
\mathcal{H}(X,Y) \otimes FX \ar[rr]^-{\overline F} \ar[d]_-{1 \otimes \alpha} && FY \ar[d]^-\alpha \\
\mathcal{H}(X,Y) \otimes GX \ar[rr]_-{\overline G} && GY
}
$$
commutes.
\subsection{Hopf categories}
\begin{defn}
A \emph{comonoidal $\mathcal{V}$-category} is a category $\mathcal{H}$ enriched over the category of coalgebras in $\mathcal{V}$. Explicitly, each object $\mathcal{H}(X,Y)$ is a coalgebra in $\mathcal{V}$ in such a way that the composition and unit of $\mathcal{H}$ are coalgebra morphisms.
\end{defn}
\begin{defn}
A \emph{Hopf $\mathcal{V}$-category} is a comonoidal $\mathcal{V}$-category $\mathcal{H}$ equipped with a collection of morphisms
$$
\xymatrix{
\mathcal{H}(X,Y) \ar[r]^-{S} & \mathcal{H}(Y,X)
}
$$
such that the diagrams
$$
\xymatrix{
\mathcal{H}(X,Y) \otimes \mathcal{H}(X,Y) \ar[rr]^-{1\otimes S} && \mathcal{H}(X,Y) \otimes \mathcal{H}(Y,X) \ar[d]^-{\circ} \\
\mathcal{H}(X,Y) \ar[u]^-{\delta} \ar[r]_-{\epsilon} & I \ar[r]_-{u} & \mathcal{H}(X,X)
}
$$
and
$$
\xymatrix{
\mathcal{H}(X,Y) \otimes \mathcal{H}(X,Y) \ar[rr]^-{S \otimes 1} && \mathcal{H}(Y,X) \otimes \mathcal{H}(X,Y) \ar[d]^-{\circ} \\
\mathcal{H}(X,Y) \ar[u]^-{\delta} \ar[r]_-{\epsilon} & I \ar[r]_-{u} & \mathcal{H}(Y,Y)
}
$$
commute.
\end{defn}
Our terminology comes from~\cite{Hop} but these are called Hopf $\mathcal{V}$-algebroids in~\cite{MR1458415}.
\subsection{The comonad $T_\mathcal{H}$}
Let $\mathcal{H}$ be any $\mathcal{V}$-category.
We now construct a pair of comonads on the contravariant enriched functor category $[\mathcal{H}^*, \mathcal{V}]$ and a distributive law between them. We do this however, by first defining a comonad on the covariant enriched functor category and pulling a few tricks. We define an endofunctor $T_\mathcal{H}$ on $[\mathcal{H},\mathcal{V}]$ as follows: given a $\mathcal{V}$-functor $\mathcal{H} \to \mathcal{V}$, let $T_\mathcal{H}(F)$ be defined on objects by
$$
T_\mathcal{H}(F)(X) = \sum_{Y} \mathcal{H}(Y,X) \otimes FY.
$$
The action $\overline{T_\mathcal{H}(F)}$ is defined by lifting the composition
$$
\xymatrix@C=2.5em{
\mathcal{H}(X,Y) \otimes \mathcal{H}(Z,X) \otimes FZ \ar[r]^-{\circ \otimes 1 }
& \mathcal{H}(Z, Y) \otimes FZ
}
$$
to the coproduct, so it is clear that $T_\mathcal{H}(F)$ is a well-defined $\mathcal{V}$-functor. Given a $\mathcal{V}$-natural transformation $\alpha \colon F \Rightarrow G$, we define $T_\mathcal{H}(\alpha) \colon T_{\mathcal{H}}(F) \Rightarrow T_\mathcal{H}(G)$ by lifting the morphisms
$$\xymatrix{
\mathcal{H}(Y,X) \otimes FX \ar[rr]^-{1 \otimes \alpha} & &\mathcal{H}(Y,X) \otimes GX}
$$ to the coproduct, and so clearly $T_\mathcal{H}$ is a well-defined functor.
The morphisms
$$
\xymatrix@C=2.5em{
\mathcal{H}(Y,X) \otimes FY \ar[rr]^-{1 \otimes u \otimes 1} & & \mathcal{H}(Y,X) \otimes \mathcal{H}(Y,Y) \otimes FY,
}
$$
and
$$
\xymatrix{
\mathcal{H}(Y,X) \otimes FY \ar[r]^-{\overline F} & FX
}
$$
lift to the coproduct and define natural transformations $\delta \colon T_\mathcal{H} \Rightarrow T_\mathcal{H} T_\mathcal{H}$ and $\epsilon \colon T_\mathcal{H} \Rightarrow 1$ respectively, which endow $T_\mathcal{H}$ with the structure of a comonad.
Now, suppose that $\mathcal{H}$ is a comonoidal $\mathcal{V}$-category. Then, the category $[\mathcal{H}, \mathcal{V}]$ is monoidal~\cite[p.~143]{MR1458415} with tensor product $\otimes$ and unit $I$ given pointwise by those of $\mathcal{V}$, i.e.\
$$
(F \otimes G)(X) = FX \otimes GX, \qquad I(X) = I
$$
and with respect to this structure, $T_\mathcal{H}$ is in fact an opmonoidal comonad. Therefore, there is a comonad $S_\mathcal{H} = T_\mathcal{H}(I) \otimes {-}$ on $[\mathcal{H}, \mathcal{V}]$ and a distributive law $T_{\mathcal{H}}S_{\mathcal{H}} \Rightarrow S_{\mathcal{H}}T_{\mathcal{H}}$ (cf.~Section~\ref{notsure} and e.g.~\cite{MR2948490}). Explicitly, $S_\mathcal{H}$ is defined on objects $F$ by
$$
S_\mathcal{H} (F) (X) = \sum_Y \mathcal{H}(Y,X) \otimes FX
$$
and the composites $T_{\mathcal{H}}S_{\mathcal{H}}$ and $S_{\mathcal{H}}T_{\mathcal{H}}$ are given by
\begin{align*}
(T_\mathcal{H} S_\mathcal{H}) (F) (X) &= \sum_{Y,Z} \mathcal{H}(Y,X) \otimes \mathcal{H}(Z,Y) \otimes FY, \\
(S_\mathcal{H} T_\mathcal{H}) (F) (X) &= \sum_{Y,Z} \mathcal{H}(Y,X) \otimes \mathcal{H}(Z,X) \otimes FZ .
\end{align*}
The distributive law is defined by the diagram
$$
\xymatrix{
\mathcal{H}(Y,X) \otimes \mathcal{H}(Y,Z) \otimes FY \ar@{.>}[r] \ar[d]_-{\delta \otimes 1 \otimes 1} & \mathcal{H}(Z,X) \otimes \mathcal{H}(Y,X) \otimes FY \\
\mathcal{H}(Y,X) \otimes \mathcal{H}(Y,X) \otimes \mathcal{H}(Z,Y) \otimes FY \ar[r]_-{\cong} &
\mathcal{H}(Y,X) \otimes \mathcal{H}(Z,Y) \otimes \mathcal{H}(Y,X) \otimes FY \ar[u]_-{\circ \otimes 1 \otimes 1}
}
$$
where the bottom isomorphism swaps the two inner tensorands.
Of course, since $\mathcal{H}$ was arbitrary, we can replace it with $\mathcal{H}^*$ to obtain a comonad $T_{\mathcal{H}^*}$.
\subsection{The comonad $R_\mathcal{H}$}
Now suppose that $\mathcal{H}$ is a Hopf $\mathcal{V}$-category. Then $[\mathcal{H}^*, \mathcal{V}]$ becomes a left module category for $[\mathcal{H}, \mathcal{V}]$, with action
$$
\xymatrix{
[\mathcal{H}, \mathcal{V}] \times [\mathcal{H}^*, \mathcal{V}] \ar[rr]^-{\rhd} & & [\mathcal{H}^*, \mathcal{V}]
}
$$ defined on objects by $$
(F \rhd G)(X) = FX \otimes GX.
$$
The action of the $\mathcal{V}$-functor $\overline{F \rhd G}$ is given by
$$
\xymatrix@C=4em{
\mathcal{H}(Y,X) \ar@{.>}[ddd]_-{\overline{F \rhd G}}\otimes FX \otimes GX \ar[r]^-{\delta \otimes 1 \otimes 1} & \mathcal{H}(Y,X) \otimes \mathcal{H}(Y,X) \otimes FX \otimes GX \ar[d]^-{1 \otimes S \otimes 1 \otimes 1}\\
& \mathcal{H}(Y,X) \otimes \mathcal{H}(X,Y) \otimes FX \otimes GX \ar[d]^-{1 \otimes \overline F \otimes 1} \\
& \mathcal{H}(Y,X) \otimes FY \otimes GX \ar[d]^-{\cong} \\
FY \otimes GY & FY \otimes \mathcal{H}(Y,X) \otimes GX \ar[l]^-{1 \otimes \overline G}
}
$$
In particular, if we choose $F$ to be the coalgebra $T_\mathcal{H}(I)$, then we get a comonad
$$
\xymatrix{
[\mathcal{H}^*, \mathcal{V}]\ar[rr]^-{T_\mathcal{H} (I) \rhd {-}} & & [\mathcal{H}^*, \mathcal{V}]
}
$$
which we denote by $R_\mathcal{H}$. Unravelling everything, we have that $R_\mathcal{H}$ is defined on objects $F$ by
$$
R_\mathcal{H} (F)(X) =S_\mathcal{H} (F) (X) = \sum_Y \mathcal{H}(Y,X) \otimes FX
$$
with action $\overline{R_\mathcal{H}(F)}$ given by lifting to the coproduct the morphisms
$$
\xymatrix@C=4em{
\mathcal{H}(Y,X) \ar@{.>}[ddd] \otimes \mathcal{H}(Z,X) \otimes FX \ar[r]^-{\delta \otimes 1 \otimes 1} & \mathcal{H}(Y,X) \otimes \mathcal{H}(Y,X) \otimes \mathcal{H}(Z,X) \otimes FX \ar[d]^-{1 \otimes S \otimes 1 \otimes 1}\\
& \mathcal{H}(Y,X) \otimes \mathcal{H}(X,Y) \otimes \mathcal{H}(Z,X) \otimes FX \ar[d]^-{1 \otimes \circ \otimes 1} \\
& \mathcal{H}(Y,X) \otimes \mathcal{H}(Z, Y) \otimes FX \ar[d]^-{\cong} \\
\mathcal{H}(Z,Y) \otimes FY & \mathcal{H}(Z,Y) \otimes \mathcal{H}(Y,X) \otimes FX \ar[l]^-{1 \otimes \overline F}
}
$$
The comonad structure is induced in the obvious way from the coalgebra structure on each $\mathcal{H}(Y,X)$.
\subsection{The distributive law $\chi$}
Again, we automatically have a distributive law (cf.~Section~\ref{opmoduleadj})
$$
\chi \colon T_{\mathcal{H}^*} R_{\mathcal{H}} \Rightarrow R_{\mathcal{H}} T_{\mathcal{H}^*}.
$$
Explicitly we have
\begin{align*}
(T_{\mathcal{H}^*} R_{\mathcal{H}} ) (F)(X) &= \sum_{Y,Z} \mathcal{H}(X,Y) \otimes \mathcal{H}(Z,Y) \otimes FY, \\
(R_{\mathcal{H}} T_{\mathcal{H}^*} ) (F) (X) &= \sum_{Y,Z} \mathcal{H}(Y,X) \otimes \mathcal{H}(X,Z) \otimes FZ,
\end{align*}
and the distributive law is induced by the composites
$$
\xymatrix@C=4em{
\mathcal{H}(X,Y) \ar@{.>}[dd] \otimes \mathcal{H}(Z,Y) \otimes FY \ar[r]^-{\delta \otimes 1 \otimes 1} & \mathcal{H}(X,Y) \otimes \mathcal{H}(X,Y) \otimes \mathcal{H}(Z,Y) \otimes FY \ar[d]^-{1 \otimes S \otimes 1 \otimes 1}\\
& \mathcal{H}(X,Y) \otimes \mathcal{H}(Y,X) \otimes \mathcal{H}(Z,Y) \otimes FY \ar[d]^-{1 \otimes \circ \otimes 1} \\
\mathcal{H}(Z,X) \otimes \mathcal{H}(X,Y) \otimes FY & \mathcal{H}(X,Y) \otimes \mathcal{H}(Z, X) \otimes FY \ar[l]^-{\cong}
}
$$
\subsection{The $\chi$-(op)coalgebras and duplicial object}
Consider the unit object $I$ in $[\mathcal{H}^*, \mathcal{V}]$. We define a $\mathcal{V}$-natural transformation $$\rho \colon T_{\mathcal{H}^*}(I) \Rightarrow R_{\mathcal{H}}(I)$$ as follows. The components are
defined by
$$
\rho \colon T_{\mathcal{H}^*}(I)(X) = \sum_{Y} \mathcal{H}(X,Y) \longrightarrow \sum_{Y} \mathcal{H}(Y,X) = R_{\mathcal{H}}(I)(X)
$$
to be the coproduct of the antipodes $S \colon \mathcal{H}(X,Y) \to \mathcal{H}(Y,X)$. It follows that $(I, \rho)$ is a $\chi$-coalgebra in $[\mathcal{H}^*, \mathcal{V}]$.
Now, let $N \colon [\mathcal{H}^*, \mathcal{V}] \to \mathcal{V}$ be the functor which maps a $\mathcal{V}$-functor $F$ to the coend
$$
\int^{X} FX
$$
which is, by definition, the coequaliser
$$
\xymatrix@C=3.5em{
\displaystyle\sum_{X,Y} FY \otimes \mathcal{H}(X,Y) \cong \sum_{X,Y} \mathcal{H}(X,Y) \otimes FY \ar@<1ex>[r]^-{\sum \epsilon \otimes 1} \ar@<-1ex>[r]_-{\sum\overline F}& \displaystyle\sum_X FX \ar@{->>}[r] & \displaystyle\int^{X} FX.
}
$$
We define morphisms
$$
\xymatrix{
L \colon \mathcal{H}(Y,X) \otimes FX \ar[r]^-{\overline F} & FY \ar[rr]^-{u \otimes 1} && \mathcal{H}(Y,Y) \otimes FY
}
$$
Since we have
$$
\sum_X R_{\mathcal{H}}(F)(X) = \sum_{X,Y} \mathcal{H}(Y,X) \otimes FX, \qquad \sum_X T_{\mathcal{H}^*}(F)(X) = \sum_{X,Y} \mathcal{H}(X,Y) \otimes FX
$$
the morphisms $L$ define a morphism
$$
\sum_X R_\mathcal{H} (F) (X) \longrightarrow \sum_X T_{\mathcal{H}^*} (F) (X)
$$
which coequalises the morphisms defining the coend $\displaystyle\int^X R_{\mathcal{H}}(F)$. Thus, there is an induced morphism
$$
\xymatrix{
\displaystyle \int^X R_{\mathcal{H}} (F) (X)\ar[rr]^-{\lambda} & & \displaystyle\int^X T_{\mathcal{H}^*} (F)(X)
}
$$
which defines a natural transformation $\lambda \colon N R_\mathcal{H} \Rightarrow N T_{\mathcal{H}^*}$ which turns the triple $(N, \mathcal{V}, \lambda)$ into a $\chi$-opcoalgebra. Therefore, by Theorem~\ref{dup}, we have:
\begin{cor}\label{endcor}
The simplicial object $$ \mathrm C_{T_{\mathcal{H}^*} } (N, I)$$
is a duplicial object in $\mathcal{V}$.
\end{cor}
\begin{exa}
Let $\mathcal{V} = \mathbf{Set}$ with monoidal structure given by the Cartesian product $\times$. A Hopf $\mathbf{Set}$-category $\mathcal G$ is precisely a groupoid~\cite[Prop.~2.4]{Hop}. The terminal object $I = \{ * \}$ is the unit, and the coend functor $N$ is given by the colimit functor
$$
\operatorname{colim} \colon [\mathcal G^*, \mathbf{Set}] \to \mathbf{Set}.
$$
In this case
$$
\mathrm C_{T_{\mathcal G^*} } (\operatorname{colim}, I) \cong N \mathcal G
$$
where $N\mathcal G$ denotes the nerve of the groupoid $\mathcal G$, cf.~Section~\ref{nerve}.
We can say even more: suppose that we have an adjunction
$$
\xymatrix{
\mathcal G \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-F & & \ar@/^0.5pc/[ll]^-U \mathcal{C}
}$$
where $\mathcal{C}$ is any category. The presheaf category $[\mathcal{C}^*, \mathbf{Set}]$ becomes a monoidal category with the pointwise Cartesian product, and the comonad $T_{\mathcal{C}^*}$ is opmonoidal (since every comonad is with respect to this particular monoidal structure) and so there is a distributive law
$$
T_{\mathcal{C}^*} \circ (T_{\mathcal{C}^*} (I) \otimes {-}) \Rightarrow (T_{\mathcal{C}^*}(I) \otimes {-}) \circ T_{\mathcal{C}^*} .
$$ The functor $$[F^*, \mathbf{Set}] \colon [\mathcal{C}^*, \mathbf{Set}] \to [\mathcal G^*, \mathbf{Set}]$$ is the functor part of a 1-cell in $\mathbf{Dist}$ between the aforementioned distributive laws. However, since $F$ is necessarily full and faithful (see the proof of~Corollary~\ref{motivation}), we have
$$
\operatorname{colim} \circ~[F^*, \mathbf{Set}] \cong \operatorname{colim} \colon [\mathcal{C}^*, \mathbf{Set}] \to \mathbf{Set}.
$$Therefore we may act on the above duplicial object with the 1-cell $[F^*, \mathbf{Set}]$ as in Section~\ref{twistsec} to obtain another one
$$
\mathrm C_{T_{\mathcal{C}^*}}(\operatorname{colim}, I) \cong N \mathcal{C}.
$$
The duplicial structure induced on $N\mathcal{C}$ is the same as that given by Theorem~\ref{auspara}.
\end{exa}
\begin{exa}
Let $k$ be a commutative ring, and let $\mathcal{V} = k\mbox{-}\mathbf{Mod}$, with tensor product $\otimes = \otimes_k$ and unit $k$. Any Hopf algebra $H$ over $k$ can be viewed as a one-object Hopf $\mathcal{V}$-category $\mathcal{H}$, and $[\mathcal{H}, \mathcal{V}] \cong H\mbox{-}\mathbf{Mod}$. The coend functor becomes
$$
{-}\otimes_H k \colon H^*\mbox{-}\mathbf{Mod} \to k\mbox{-}\mathbf{Mod}
$$
and the duplicial $k$-module $$
\mathrm C_{T_{\mathcal{H}^*}} ({-}\otimes_H k, k)
$$
is precisely the one given in Section~\ref{coeffsforhopf} in the case that $A = M = N = k$, whose ordinary homology is given by ${\rm Tor}_{H/ k} (k, k)$.
\end{exa}
\chapter{Unanswered questions}\label{GRANDFINALE}
In this brief final chapter, we state some problems that came up during the preparation of this thesis that remain unsolved.
\begin{ques}
In Chapter~\ref{EXAMPLES} we saw related algebraic examples of duplicial objects arising from Theorem~\ref{dup}. What is missing is a brand new cyclic homology theory. We do however, show that the non-empty list functor $L^+$ becomes a bimonad (Section~\ref{newbimonad}). Is there a functor $N \colon \mathbf{SemiGrp} \to \mathcal{Z}$ for some category $\mathcal{Z}$ that admits the structure of an opcoalgebra over a distributive law, which gives rise to an interesting cyclic homology theory of semigroups?
\end{ques}
\begin{ques}
The duplicial object $\mathrm C_{T_{\mathcal{H}^*}}(N, I)$ of Corollary~\ref{endcor} can be used to describe duplicial structures on the nerves of categories (Section~\ref{nerve}) as well as on the simplicial object with homology ${\rm Tor}_{H/k} (k, k)$ for a Hopf algebra $H$ over a commutative ring $k$ (Section~\ref{coeffsforhopf}). However, it says nothing of more general Hopf algebroids over a noncommutative base algebra $A$. Is there a way to upgrade the construction method of $\mathrm C_{T_{\mathcal{H}^*}}(N, I)$ so that it admits a special case of the duplicial object in Section~\ref{coeffsforhopf} as an example?
\end{ques}
\begin{ques}
In Theorem~\ref{auspara} we describe duplicial structures on the nerves of categories $\mathcal{C}$ in terms of
adjunctions
$$
\xymatrix{
\mathcal G \ar@{}[rr]|-{\perp}\ar@/^0.5pc/[rr]^-I & & \ar@/^0.5pc/[ll]^-R \mathcal{C}
}$$
where $\mathcal G$ is a groupoid. If there exists two groupoids $\mathcal G, \mathcal G'$ equipped with left adjoints into $\mathcal{C}$, then we have $\mathcal G \simeq \mathcal G'$, as explained in Remark~\ref{catrem}. How are the two duplicial structures on $N\mathcal{C}$ induced by $\mathcal G$ and $\mathcal G'$ related?
\end{ques}
\begin{ques}
Suppose that we are in the situation of Section~\ref{bimfromhopf}, in the special case that we have a bialgebra $H$ over a commutative ring $k$. There is a distributive law $\chi$ defined by
$$
\xymatrix@R=1em{
H \otimes H \otimes X \ar[r]^-{\chi_X} & H \otimes H \otimes X \\
h \otimes g \otimes x \ar@{|->}[r] & h_{(1)}g \otimes h_{(2)} \otimes x.
}$$
If $H$ is a Hopf algebra, the functor $k \otimes_H{-} \colon H\mbox{-}\mathbf{Mod} \to k\mbox{-}\mathbf{Mod}$ becomes a $\chi$-opcoalgebra, with structure morphism $\lambda$ defined by
$$
\xymatrix@R=1em{
k \otimes_H (H \otimes X) \ar[r]^-{\lambda_X} & k \otimes_H (H \otimes X) \cong X \\
1 \otimes_H (h \otimes x) \ar@{|->}[r] & S(h)x.
}
$$
Does the converse hold; i.e.~does a $\chi$-opcoalgebra structure on $k \otimes_H{-}$ imply that $H$ must be a Hopf algebra?
\end{ques}
\begin{ques}
Is there an interesting application of the work in Section~\ref{HOCHLAX} other than duplicial objects?
\end{ques}
\begin{ques}
Yetter-Drinfel'd modules seem to be of some importance with regard to dupliciality/cyclicity. For example:
\begin{itemize}
\item One requires a Yetter-Drinfel'd module $N$ in Section~\ref{coeffsforhopf} to define a $\chi$-opcoalgebra.
\item In the language of Section~\ref{coeffsforhopf} it is shown that the duplicial object $\rCC_T({-}\otimes_H N,M)$ is cyclic if $M \otimes_{A^*} N$ is a stable anti Yetter-Drinfel'd module.
\item A monoidal category $\mathcal{A}$ becomes a 0-cell in $\mathcal A \mbox{-}\mathbf{Mod}\mbox{-}\mathcal A$ with the regular actions (c.f.~Section~\ref{HOCHLAX}). We get that $\mathrm H^0(\mathcal{A}, \mathcal{A})$ is isomorphic to the lax centre of the monoidal category $\mathcal{A}$ (see~\cite{MR2381533,MR1800718}). When $\mathcal{A}$ is the category of modules over a Hopf algebra, this is the category of Yetter-Drinfel'd modules.
\end{itemize}
Is there a general phenomenon at work which explains these connections?
\end{ques}
|
\section{Related Work}
\label{sec:background}
The IR approach to visual model validation is informed by advances in several
different areas of research. These include the topics of uncertainty,
predictive visualization, and progressive or incremental visualization. Also
relevant are visualization systems that utilize inferential statistics methods
and conceptual models of the visualization pipeline.
\subsection{Visualization of Uncertainty}
The visualization of uncertainty has been an active research area within the
visualization community for many years. Studies have explored the problem
from many perspectives, including taxonomies that have
examined types of uncertainty \cite{skeels_revealing_2010} as well as
visualization methods for conveying uncertainty
\cite{potter_quantification_2012}. In addition, there have been many efforts to
formally study alternative methods for depicting uncertainty
measures~\cite{maceachren_visual_2012,sanyal_user_2009-1,tak_perception_2013}
through user studies that explore the perceptual understanding of
various uncertainty representations. However, these studies focus on the visual
representation rather than methods for determining the degree of uncertainty.
Perhaps most relevant to the IR approach proposed in this paper is work that
has focused on estimating uncertainty via measures of entropy within a dataset
rather than by using carefully constructed statistical models
\cite{potter_visualization_2013}. Like IR, this work adopts
a non-parametric approach which does not require formal modeling nor make
assumptions about specific distributions within the data.
Finally, the distinction between the ``visualization of uncertainty'' and ``the
uncertainty of visualization'' has been highlighted \cite{brodlie_review_2012}.
The latter is a related but separate concept from traditional uncertainty
visualization. This work highlights that the rendered graphics of a
visualization can convey a sense of authority which may not be warranted, even
when the underlying data itself is considered to be beyond reproach. This
challenge is a key motivation for IR, as outlined in the discussion presented in
Section~\ref{sec:prediction}.
\subsection{Predictive Visual Analytics}
Visualization has long been used to support predictive analysis tasks.
However, most often, the ``prediction'' is performed by users reviewing historical
data and making assumptions about what might happen in the future for similar
situations. In fact, the relatively limited history of work on visualizations
that incorporate more formal predictive modeling methods was the topic for a
workshop at the most recent IEEE VIS Conference in 2014 \cite{perer_ieee_2014}.
The work that does exist in this area has often focused on model
development and evaluation rather than supporting end users' predictive
analysis tasks. For example, BaobabView \cite{van_den_elzen_baobabview:_2011}
supported interactive construction and evaluation of decision trees. More
recent work has focused on building and evaluating regression models
\cite{muhlbacher_partition-based_2013}. This method, like ours, adopts a
partition-based approach to avoid making structural assumptions about the data.
However, the focus on building regression models leads to an overall workflow
that is very different from the proposed IR approach.
Others have focused on visualizing the output produced by predictive models.
For example, Gosink et al. have visualized prediction uncertainty based on
formalized ensembles of multiple predictors \cite{gosink_characterizing_2013}.
This approach, however, requires careful modeling to develop the predictors, including
the specification of priors that enable the Bayesian method that
they propose.
Outside the visualization literature, where novel visual or interaction methods
are not a concern, predictive features are typically visualized using
traditional statistical graphics, for example, systems that visually prioritize
and threshold p-values to rank features for prediction (e.g.,
\cite{sipes_predictive_2011}). Such methods are fully compatible with the IR
process proposed in this paper.
\subsection{Progressive/Incremental Visualization}
Overfit models and other sampling challenges are common to ``Big Data''
visualizations that rely on progressive or incremental techniques
(e.g., \cite{fisher_exploratory_2012,stolper_progressive_2014}). Initial
samples are small, grow over time, and can change in distribution as time
proceeds. Some have addressed this challenge by including confidence intervals
along with partial sets of query results \cite{fisher_trust_2012}. However,
relying on the query platform to assess confidence in data subsets does not
easily support interactive zoom and filter operations after the query, because
these changes in visual focus do not necessarily result in new queries that
generate new result sets. Moreover, these papers do not propose methods for
computing confidence intervals, but rather, assume that such data will be
provided by the the database.
\subsection{Inferential Statistics}
Statistical inference is a discipline with a very long and distinguished history.
Most relevant to the IR method described in this paper are challenges related
to statistical significance and null hypotheses, and in particular Type~1 and
Type~2 errors. Type~1 errors refer to improper rejections of the null
hypothesis which lead to conclusions that are not real effects, while Type~2
errors refer to falsely retaining the null hypothesis which can lead to
assumptions that a true effect is false~\cite{sheskin_handbook_2003}.
These types of errors are of critical concern
in high-dimensional exploratory visualization where computational
methods can quickly access vast numbers of dimensions for statistical
significance. Statistical correction methods have been proposed to reduce Type~1
errors \cite{stoline_status_1981}, but arguments have also been made against
this approach. Those arguments suggest that parameterized models or
assumptions of ``default'' null hypotheses don't match real world situations where
distributions are rarely straightforward or independent. Suggesting that these correction
methods are the wrong approach for exploratory work, Rothman argues that
``scientists should not be so reluctant to explore leads that may turn out to be
wrong that they penalize themselves by missing possibly important findings''
\cite{rothman_no_1990}.
This tension is present in many interactive exploratory systems which make it
easy to generate vast numbers of potential hypotheses. As a result, many
methods have been proposed for modeling measures of confidence or significance
\cite{chen_uncertainty-aware_2015,feng_matching_2010,wu_visualizing_2012}.
These efforts, however, typically rely formal statistical methods that make assumptions
about distributions and variable independence.
This approach is problematic for exploratory visualizations which allow users to quickly apply filters or
constraints that can quickly change the underlying assumptions. The IR method we
propose provides provides a way for users to visually assess the
reliability of hypotheses. Similar approaches that rely on user judgement have
been shown to be quite effective \cite{majumder_validation_2013}.
\subsection{Models of the Visualization Pipeline}
The traditional visualization pipeline model describes the process of transforming raw data
to an analytical abstraction, to visualization abstraction, and then finally to a
rendered graphic for interaction \cite{card_readings_1999,chi_taxonomy_2000}.
We add partitioning and aggregation stages to this flow to support the IR
approach. As we will describe, a special case of the IR model (in which we
generate just one partition) is equivalent to the traditional model. By extending the canonical pipeline, our work has similarities with Correa et al.'s paper describing pipeline extensions for an uncertainty framework focused on the data transformation process
\cite{correa_framework_2009}.
\section{Conclusion}
\label{sec:conclusion}
Traditional data visualizations show retrospective views of existing
datasets with little or no focus on prediction or generalizability. However, users often base
decisions about future events on the findings made using retrospective
visualizations. In this way, visualization can be considered to be a visual
predictive model that is subject to the same problems of overfitting as
traditional modeling methods. As a result, visualization users can often make
invalid inferences based on unreliable visual evidence.
This paper described an approach to visual model validation called \emph{Inline
Replication} (IR). Similar to the cross-validation technique used widely
in machine learning, IR provides a nonparametric and broadly applicable
approach to visual model assessment and repeatability. The IR pipeline was
defined, including three key functions: the partition function, the metric
function, and the aggregation function. In addition, methods for visual display
and interaction were discussed. Uses cases were described, including a new
IR-based implementation of the existing DecisionFlow system for exploratory analysis. The
use cases demonstrated the successful compatibility of IR with a variety of
visual metaphors and derived measures.
While the results presented in this paper are promising, they represent only
one step in a growing effort to bring high repeatability and predictive power
to visualization-based analysis systems. There are many areas for future work
including: improved techniques for detecting and conveying issues related to
missing data, techniques for addressing and visually warning users regarding
selection bias, and improved methods for conveying the degree of compatibility
between a given statistical model's assumptions and the actual underlying data.
\section{Introduction}
\label{sec:intro}
Visualizations are most often designed to depict the entirety of a
dataset--subject to a set of filters applied to focus the analysis--as
accurately as possible. In this typical pattern, the goal is to provide a
person using the visualization with an accurate understanding of \emph{all of
the data} in the underlying dataset that matches the active set of filters.
This ethos was captured, perhaps most famously, in Shneiderman's Visual
Information Seeking Mantra: \emph{overview first, zoom and filter, then
details-on-demand} \cite{shneiderman_eyes_1996}. Variations of this basic
approach have since been adopted in most modern visualization systems.
The foundation for these systems are visual mappings that specify a graphical
representation for the underlying data. For small and low-dimensional data
sources, these mappings can be direct (e.g., a scatter plot for a small two
dimensional dataset). As problems grow in data size or dimensionality,
algorithmic data transformation methods can be used to filter, manipulate, and
summarize raw data into a more easily visualized form.
On top of these mappings, interactive controls are often provided to allow
users even more flexibility to filter or zoom to specific subsets of data.
These interactions can be linked to more detailed information about data
objects, for example via levels-of-detail or multiple coordinated views. The
result, when well designed, is an effective visual interface for data
exploration and insight discovery.
For this reason, these steps form the core stages of the canonical
visualization pipeline \cite{card_readings_1999, chi_taxonomy_2000}.
This approach can be enormously informative, and it has led
to revolutions in how people seek to understand information. This approach
can be used, for example, to visualize file systems (showing the
space used by various directories) to help computer users navigate through large
directories; to visualize medical records to help doctors
understand patient histories; and to visualize maps of weather data
to identify regions most impacted by a recent storm.
Critically, however, these visualization use cases are all \emph{retrospective}
in nature. Moreover, they describe visualizations that faithfully report data
as it was observed. Users aim to see an overview of the entirety of a dataset.
If a user applies constraints to focus the visual investigation (e.g., via zoom
and filter), the visualization is expected to show the full set of data that
satisfies the applied constraints.
In many visualization scenarios, however, users are in fact more interested in
conducting \emph{prospective analysis}: using historical data to reason about
future or not-yet-observed data. For example, medical experts examining data
for a cohort of patients might be most interested in what treatments would work
best for a future patient with similar characteristics. Visualizations of
historical sports statistics are often used to inform strategic decisions that
are used in upcoming competitions. Financial visualization tools are often
used to inform future investment decisions. In each of these use cases,
visualizations of historical data are used to inform future decisions.
For such prospective analysis tasks, retrospective visualizations are often
used as naive \emph{visual predictive models} with the assumption that
historical data can be predictive of future observations. In fact, in many
cases retrospective representations are indeed very informative.
However, just like the underlying descriptive statistics that such
visualizations often depict \cite{ostle_statistics_1963}, traditional
retrospective visualizations often provide insufficient evidence for making
predictive inferences.
This critical gap between (a) retrospective visualization designs and (b) the
predictive requirements of many users has been recognized within the
visualization community \cite{perer_ieee_2014}.
Some have attempted to bridge this gap by adding support for
inferential statistics within the visualization. Typically, this approach
combines carefully designed statistical models with visualizations of the
model's results. For example, visualizations can be instrumented to estimate
and display uncertainty, confidence intervals, or statistical significance.
Alternatively, predictive modeling methods can be used to generate additional
data, with the predictions themselves being incorporated into the
visualization. These systems go beyond traditional descriptive reporting, but
they typically require a careful and sometimes onerous focus
on modeling, including estimates for underlying statistical distributions.
This paper presents \emph{Inline Replication} (IR), an alternative approach
to enabling inferential interpretation that is designed to overcome many of
the above challenges. Our method, made practical by the ever-larger datasets now
available in many applications, is motivated by the cross-validation technique
used widely in machine learning. The IR approach is nonparametric, making it
easy to apply and use generically within a visualization system without arduous
modeling. In addition, IR is ideal for use in large-scale visualization
systems where progressive or sample-based approaches are required. Finally, our
method provides users with validation information that is both intuitive and
easy to interpret.
This remainder of this paper is organized as follows. It begins with a review
of related work, then describes the details of the IR methodology. We then
share example results from a variety of proof-of-concept systems that have
adopted the IR technique. These examples range from simple bar charts to more
sophisticated interactive visualizations of large scale event data collections
\cite{gotz_decisionflow:_2014}. The paper concludes with a discussion of
limitations and outlines key areas for future work.
\section{Acknowledgements}
This research was made possible, in part, by funds from a 2015 Data Fellow Award from the National Consortium for Data Science (NCDS).
\newpage
\newpage
\bibliographystyle{abbrv}
\section{Visualization as a Predictive Model}
\label{sec:prediction}
Visualization design is often conceptualized as a mechanism for reporting.
This retrospective approach is so ubiquitous that terms such as
\emph{prediction}, \emph{forecast}, and \emph{inference} cannot be found within
the indices of many leading visualization texts from the past 25 years (e.g.,
\cite{munzner_visualization_2014,tufte_envisioning_1990,ware_information_2004}).
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{figs/workflow_overview.pdf}
\caption{The Inline Replication (IR) visualization pipeline sends each
derived measure's subset of data ($m_i$) through a partition function
to create multiple folds ($f_j$) prior to mapping and visualization.
A metric function is applied to each fold independently, and an
aggregation function recombines the folds to form an aggregate measure
($m_i'$) for subsequent visualization and interaction.}
\label{fig:overall_flow}
\end{figure*}
Many visualization \emph{consumers}, however, use graphical representations of
historical data as the basis for decisions about future performance. This is
done even when the underlying data and transformations do not support such
prospective conclusions. Despite potentially fatal flaws
in terms of generalizability and repeatability, retrospective
visualizations are in essence being used as predictive models.
The tendency to assume predictive power in visualization can be seen, for
example, in modern casinos. Roulette wheels, for instance, commonly include an
electronic display (e.g., \cite{roulette_display_firm}) which shows the table's
recent history. Assuming a fair table, ``red'' and ``black'' numbers should be
equally likely to appear. However, as illustrated in Figure
\ref{fig:roulette}, the history provided to gamblers is not sufficiently long
to learn if the table has any systemic bias.
\begin{figure}[t]
\centering
\includegraphics[width=8.2cm]{figs/roulette_vis.pdf}
\caption{Roulette wheels allow users to bet on ``black'' or ``red'' squares.
Casinos often display a simple visualization of ``recent spins'' to provide gamblers
with a false sense of predictive knowledge. In this example, the display
shows a recent preponderance of black numbers with the implication to
gamblers that this may influence future spins of the wheel.}
\label{fig:roulette}
\vspace{-0.2cm}
\end{figure}
Why then is the gambler presented with a simple visualization of the history?
The data is visualized to provide gamblers with a false sense of knowledge; to
suggest to a hesitant gambler that a bet is an informed decision rather than a
random choice. A gambler may infer that the recent streak of black suggests
more black spins. Alternatively, the gambler may infer an imminent return to
red. To the casino, it does not matter what predictive inference is drawn as
long as it provides a sense of confidence that leads to increased betting.
It is tempting to dismiss this scenario as one in which the gambler should be
more informed about basic statistics. The small sample size and the
independence of each roulette spin should make it clear that the display is not
especially informative. However, relatively sophisticated users performing
visual analysis of data from more complex underlying systems can make similarly
poor predictive assessments on the basis of visual representations that don't
properly convey the underlying limits of their predictive power.
For example, consider a business analyst attempting to learn about why sales
are declining, or a physician using historical patient data to compare treatment
efficacy. In these complex real-world cases, in which it is essentially
impossible to fully understand the underlying statistical processes, it is
natural for analysts to turn to visualization as a predictive model for their
problem. Visualization allows these users to see what has happened and, based
on trends or patterns in the representation, make assumptions about what will
happen in the future.
However, just as the casino gambler draws inference from a not-so-meaningful
visualization, these power users can be led to make poor predictions on the
basis of visualizations that are essentially ``overfit'' models based on poor
representations of the underlying process. This problem has even been
documented even in highly quantitative fields such as epidemiology, where
public health analysts have had trouble discounting statistics from small
sample sizes when visualized \cite{sutcliffe_developing_2014}.
Issues of poor sampling and overfit are especially problematic during
exploratory visualization in which users can interactively apply arbitrary
combinations of filters to produce new ad hoc subsets of data for
visualization. Such systems are at greater risk of generating
non-representative visualizations that occur ``by chance'' rather than due to
real properties of the underlying problem. The same is true for visualization
systems that utilize sampled or progressive queries to address issues of scale.
The potential for this sort of ``visual model overfitting'' is
analogous to the overfitting problem in more traditional modeling tasks.
In the machine learning community, this is addressed in
part by cross-validation, a widely used technique for assessing the
quality and generalizability of a model \cite{kohavi_study_1995}. Rather than
relying on a single model, cross-validation methods create and compare multiple
models, one for each of several partitions of a dataset (often called
``folds''). This allows for an assessment of model repeatability, with models
that work consistently across partitions considered more trustworthy.
If one considers---as we argue here---that a visualization is often used as a
form of predictive model, then validation becomes a critical guard against
problems associated with overfitting. When a visualization is zoomed and
filtered to focus on a specific subset, is the visual representation
repeatable? Are the conclusions drawn from the visualization generalizable?
Can an approach similar to cross-validation be embedded within the
visualization pipeline so that each new view produced during user interaction
is evaluated? The IR method outlined in the next section is designed to
support this form of validation as an integrated part of the visualization
process.
\section{Inline Replication}
\label{sec:theory}
Inline Replication (IR) is an approach to visualization in which the dataset
associated with each visualized measurement is partitioned into multiple subsets (which
we call \emph{folds}), processed independently to calculate derived statistics
or metrics, then aggregated together to be mapped to a set of graphical marks
and rendered. This partitioned approach
embeds an automated and non-parametric workflow for replication within the
visualization pipeline as illustrated in Figure~\ref{fig:overall_flow}. The
result is that visualizations based on IR provide users with important
information about the repeatability of observed visual trends, reducing the
likelihood of certain types of erroneous conclusions.
The IR pipeline begins with the same initial step as a traditional
visualization pipeline. As illustrated in Figure~\ref{fig:overall_flow}, a set
of query or filter constraints is applied to a primary data source $D$ to
produce a focused dataset $d \subset D$. The data in $d$ is then organized
into subsets for which statistical measurements are calculated, creating
measure-specific subsets of data which we note as $m_i$. For example, a
visualization pipeline configured to generate the bar chart in
Figure~\ref{fig:roulette}, showing the distribution between black and red spins
for a roulette wheel, would include the subset $m_{recent}$ containing
data for the most recent spin results (black or red). If the visualization included multiple bar charts (e.g., past 10 spins, past 100 spins, and past 1000 spins) then multiple subsets $m_i$ would be defined because each requires the calculation of a distinct set of measurements.
Traditionally, the data for each subgroup $m_i$ would immediately be processed
to compute the measurements required for visualization (e.g., the fraction of
spins resulting in black, and the fraction of spins resulting in red). Those
measures would then be mapped to visual properties of objects within the
visualization (e.g., the size of each bar in the bar chart).
The IR pipeline, however, behaves differently. Each $m_i$ is first partitioned
into distinct folds $f_i$, each of which is analyzed independently via a metric
function. The results are then aggregated to form a merged dataset $m_i'$. It
is this merged representation of the measures, $m_i'$, that is mapped to the
visual representation and rendered to the screen for interaction using methods
designed to convey the repeatability of the visual model across each of the
folds.
This section provides an overview of the IR pipeline, focusing on the three
functions at the core of the design: the partition function, the metric
function, and the aggregation function. It then describes the IR approach to
visual display and interaction, and concludes with a discussion of useful
variations to the core design.
\subsection{Partition Function}
\label{sec:partition}
Conceptually, the \emph{partition function} is designed to subdivide the data
in a given measure-specific subset $m_i$ into multiple partitions. The goal of
this stage in the IR pipeline is the creation of several independent datasets,
which we call \emph{folds}, to use as the basis for calculating each
measurement. Later in the IR process, derived measures (e.g., proportions, or
statistical significance) will be calculated for each fold.
Formally, we define the $Partition$ function as an operator that subdivides a
measure-specific set of data $m_i$ into $n$ folds such that
each fold $f_j \subset m_i$.
\begin{equation}
Partition(m_i, n) \rightarrow \{f_0, f_1, \cdots, f_j, \cdots, f_n\}
\label{eq:partition}
\end{equation}
This function is applied to the raw data in $m_i$, prior to any other
aggregating transformations (such as the summation in the roulette example).
Following an approach inspired by $k$-fold
cross-validation~\cite{kohavi_study_1995}, the baseline partition function
creates $n$ folds that are disjoint, approximately equal in size, and randomly
partitioned such that:
\begin{equation}
\bigcup_{j=0}^{n}f_j=m_i
\label{eq:comprehensive}
\end{equation}
As discussed previously, multiple folds are created with the goal of supporting
repeated calculations for each measure. Increasing the value of $n$ to produce
more folds increases the replication factor. However, higher $n$ values also
produce smaller $f_j$. If $n$ is too large for given $m_i$, the folds may be
too small to compute useful measures. Therefore, $n$ can be dynamically
determined so as to require a minimum fold size. If $m_i$ represents a
``large enough'' subset of data, it will produce a full set of folds. If,
however, $m_i$ is too small for the minimum fold size, fewer than $n$ folds
will be produced. The threshold for ``large enough'' depends on many factors,
including the specific metrics that will be calculated.
Partitioning with $n=1$ results in the \emph{identity partition function} where
$f_0 = m_i$ regardless of the size of $m_i$. Because no partitioning is
performed, an IR process using the identity partition function produces results
that are identical to a traditional visualization pipeline: a single metric is
calculated and visualized. In this way, the traditional approach to
visualization can be seen as a special case of the IR process in which
replication is not performed because there is only one fold.
Choosing a proper $n$ value is necessarily a compromise between increased
replication and smaller sample size. We can look to the machine learning
community for guidance, however, where empirical studies have shown that there
is no meaningful benefit for values of $n$ over 10~\cite{kohavi_study_1995}.
Moreover, as datasets grow larger in many fields, smaller samples become less
of a concern.
Finally, there are certain conditions (e.g., very small datasets with little data
to partition, or very large datasets where sampled queries are required) where
the basic formulation for the partition function can be problematic.
Variations to the partitioning process, designed to help address these
challenges of scale, are discussed in Section~\ref{sec:variations}.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{figs/test_prototype_fullscreen.pdf}
\vspace{-0.05cm}
\caption{The IR-based prototype shown here was developed to test the proposed
pipeline and to explore the parameter space with two baseline visualization
types: bar charts and linear regression lines. The left panel shows the
query and IR controls, the middle panel shows the visualization space,
and the left panel shows detailed descriptive statistics computed for both the
aggregate representation and the individual folds.}
\label{fig:baseline_prototype}
\vspace{-0.1cm}
\end{figure*}
To illustrate the partitioning process, consider the roulette example from
earlier in this paper. The example bar chart showing the fraction of spins
resulting in black or red is based on a single measure-specific subset of
data $m_{recent}$. The $Partition$ function would be applied to this subset to
create a set of multiple folds, each of which would contain a subset of the
recent spin results. For example, $Partition(m_{recent}, 5)$ would produce a
set of five folds, each containing results from one-fifth of the overall set of
recent spins.
\subsection{Metric Function}
The folds produced during partitioning are sent to a \emph{metric
function} which is applied independently to each fold as illustrated in
Figure~\ref{fig:overall_flow}. The goal of the metric function is
to derive a set $s_j$ of one or more derived statistics for each fold $f_j$.
Because the metric function is applied to all folds, multiple sets of
statistics are created for each $m_i$. These statistics can then be aggregated
and compared during the eventual visualization rendering process.
The specific measures computed by the metric function are necessarily
application specific, but could range from simple descriptive statistics
(e.g., sums, averages) to more complex analyses (e.g., classification,
regression). Generally speaking, the metric function is defined to produce the
same derived values that would normally be computed as part of a more
traditional visualization process. The key difference in IR is
that the metrics are computed multiple times for $m_i$ (once for each
fold), where traditionally such values would be computed just once.
For example, consider the roulette use case described earlier. The metric
function in this example would compute the fraction of spins resulting in black
and red in each fold $f_j$. This fraction is the same measure that the
original bar chart is designed to display. However, with the IR approach, the
metric is calculated for each of the five folds produced by
$Partion(m_{black},5)$.
An actual implementation of IR using a similar ``fraction of the population''
metric function is discussed in Section~\ref{sec:usecase}. However, more
sophisticated systems may adopt more advanced measures. For example,
correlation statistics, p-values, metrics of model ``fit'', and regression lines
are all compatible with the IR approach. Examples of IR using linear
regression, correlation, and statistical significance testing are all described
in Section~\ref{sec:usecase}.
\subsection{Aggregation Function}
The metric function produces a set of statistical measures $s_j$ for each of
the $n$ folds $f_j$ that are produced by the partition function. Prior to
visualization, the multiple $s_j$ metrics must be aggregated to a single
representation $m_i'$ to invert the partition process. As illustrated in
Figure~\ref{fig:overall_flow}, this is accomplished via an \emph{aggregation
function} which we define as follows.
\begin{equation}
Aggregate(\{(f_j, s_j)\}) \rightarrow m_i'
\end{equation}
The \emph{Aggregate} function is designed to produce one aggregate value for
each of the different measures computed by the metric function. For instance,
if a metric function computes two measures for each fold (e.g., count and
correlation), then the aggregation function would produce two corresponding
aggregate measures.
A variety of aggregation algorithms can be employed, with different approaches
appropriate to different types of metrics. For example, for count-based
metrics which capture the frequency of data items in each fold, a summation
across all folds might be the most appropriate because a sum of counts for each
fold provides an accurate total for the overall data subset $m_i$. For a
metric that captures a mean or rate, averaging the values across all folds may be
most appropriate. For categorical metrics, meanwhile, such as those produced
by classification algorithms, a ``majority vote'' aggregation method
\cite{lam_application_1997} can be applied to capture the most frequently
assigned category. The same voting approach can be used when aggregating thresholded measures (e.g., tests of statistical significance) across all folds. This approach is demonstrated in Section~\ref{sec:decisionflow_usecase}.
The summary measures produced by the aggregation function are combined with the
set of $s_j$ statistics computed for the $n$ folds to form a merged data
representation $m_i'$. This merged representation is then used to drive the
mapping and rendering process of the final visualization.
As a concrete example, consider again the roulette scenario. The metric
function described previously computed the fraction of spins resulting in black
and red numbers for each of the five folds created by $Partition(m_{recent},
5)$. Because the partitions are by definition equal in size, aggregate rates
for both colors can be obtained by averaging the five fold-specific rates. The
overall average, along with the individual values for each fold, are combined
to form $m_{recent}'$.
\subsection{Visual Display And Unfolding of Partition Data}
Once aggregation has been performed, the merged data $m_i'$ is mapped to its
corresponding visual marks and displayed as part of the visualization. This is
shown as the final step in Figure~\ref{fig:overall_flow}. The IR approach to
visualizing $m_i'$ has two elements, which correspond to the two distinct types
of information in the merged data structure: (a) the aggregate measures and (b)
the individual fold measures.
First, an initial visualization is created using only the aggregate measures.
The process for this stage is similar to a traditional visualization pipeline.
The aggregate measures are mapped to visual properties of the corresponding
graphical marks, which we call \emph{aggregate marks}. These marks
are then rendered to the screen for display and interaction.
In the roulette example, for instance, the aggregate data for black and red
spin rates (produced by the $Aggregate$ function) can be used to generate a basic bar chart that is identical to what is shown in Figure~\ref{fig:roulette}.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{figs/increasing_confidence_orig_captures.pdf}
\caption{Six charts produced by the IR prototype system. The top three charts (a-c) show the gender distribution for three
different sets of ICU patients. The relatively similar bar charts
suggest that the underlying populations are comparable. However, when
the same populations are visualized with 7 folds
(d-f), a different story appears. The charts now clearly demonstrate
that we know less about the population visualized in the left column
than we do about the population on the right. In this
case, the difference is due largely to the size of the respective populations.}
\label{fig:confidence}
\end{figure*}
Second, an IR visualization allows aggregate marks to be \emph{unfolded}. An
unfolding operation---typically triggered by a user interaction event such as
selection or brushing---augments the aggregate marks with a visualization of
the individual fold statistics that contribute to the aggregate measures. In
the ongoing roulette example, the fold data would show the variation in
proportion of spins that result in black and red numbers across each of the
$n=5$ independent folds. Additional examples from our experimental prototypes
are described in Section~\ref{sec:usecase}.
\subsection{Discussion}
The ability to unfold aggregate measures into repeated measurements
is a central contribution of the IR approach. By
graphically depicting the repeatability of a particular measure across
multiple folds, IR provides users with important and easy-to-interpret cues as
to the variability of a given measure. Traditional visualization methods do
not convey this information, meaning it is often not considered when
predictive conclusions are made by users.
Another benefit of IR comes from the aggregation function. In particular,
embedding within the visualization pipeline an ability to aggregate
categorical values such as statistical significance classification can lead to
more accurate results. Repeated measures combined with voting-based
aggregation can, for instance, reduce the exposure to Type 1 errors when
looking for statistically significant p-values. For example, a statistically
significant ($p<0.05$) run of black spins on the roulette wheel is less likely
to occur ``by chance'' across a majority of $n$ folds than it is across a
single group of spins. This is a major benefit for exploratory visualization
techniques that allow users to visually ``mine'' through large numbers of
variables in search for meaningful correlations.
\subsection{Variations}
\label{sec:variations}
Following the traditional approach to $k$-fold cross-validation, the baseline
$Partition$ function defined in Section~\ref{sec:partition} specifies that the
constructed folds are disjoint, randomly partitioned, and exhaustive (Equation~\ref{eq:comprehensive}). However, relaxing these constraints leads to several valuable variations to the baseline IR procedure.
{\bf Partial Partitioning.} Relaxing the requirement of
Equation~\ref{eq:comprehensive} allows for the creation of partitions that do not
contain all data points within $m_i$. For very large datasets, this can allow
for approximate analyses that use only a subset of the available data. This
approach provides significant performance benefits for metric functions that
have poor scaling properties, and it allows IR to work directly with recently
proposed techniques for progressive visualization (e.g.,
\cite{fisher_exploratory_2012,stolper_progressive_2014}).
{\bf Partitioning With Replacement.} Relaxing the requirement that all folds
are disjoint allows for partitioning with replacement. Similar to statistical bootstrapping, this approach allows the same data point to be included in multiple
folds (or even multiple times within the same fold). When allowing replacement,
the dataset in $m_i$ becomes a sample distribution from which the partitioning
algorithm can generate a larger population.
This is especially useful for small datasets---a frequent occurrence in
exploratory visualization where multiple filters can be quickly
applied---because the larger generated population can allow the IR process to
run with less concern about producing fold sizes that are too small.
{\bf Incremental Partitioning.} A number of progressive or sampled methods have
been proposed in recent years to address the challenges of ``Big Data''
visualization (e.g., \cite{fisher_trust_2012,stolper_progressive_2014}). In
these approaches, the full dataset $m_i$ is often never fully retrieved. To
utilize an IR approach in these cases, an incremental partitioning process is
needed. During this process, data points should be distributed to folds as
they are retrieved such that all $n$ folds are kept roughly equal in size.
This will allow IR to work with continuing improvement in metric quality as
more data arrives. However, we note that IR will not overcome selection
bias that may be introduced as part of the progressive query process. Therefore,
the determination of a progressive sampling order that is both representative
and balanced remains a critical concern.
\section{Use Cases}
\label{sec:usecase}
The IR approach is compatible with a broad range of visual metaphors and
interaction models, from basic charts to more sophisticated exploratory visual analysis
systems. To demonstrate this flexibility and to explore the impact of adopting an IR
pipeline, we developed two prototype IR systems: (a) a reference prototype to
study IR in isolation, and (b) a more sophisticated visual analysis system to
examine IR within a more complex analysis environment.
\subsection{Prototype 1: Reference Prototype}
\label{sec:baseline_usecase}
We developed a reference IR implementation as part of a simplified visual
analysis prototype with the goal of exploring the IR parameter space in
isolation, without concern for the more complex interactions that are part of
a real-world application such as the one described in Section~\ref{sec:decisionflow_usecase}. The prototype supports
two basic visualization metaphors: (a) bar charts and (b) scatter plots with
linear regression lines. The prototype was tested using a dataset of
electronic medical data containing over 40,000 intensive care unit (ICU) stays
\cite{goldberger_physiobank_2000}.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{figs/regression_real_data_corrected_band.pdf}
\vspace{-0.1cm}
\caption{Weight versus height distribution for patients admitted to a
neonatal intensive care unit. Simulating the results from a progressive
visualization system, this figure shows both the raw data and best fit
regression line (shown in blue) for (a) 500 patients, (b) 1,000
patients, and (c) 2,500 patients. In all three cases, the IR pipeline has
computed a regression across five folds, shown in red. The decreasing
spread across the red regression lines conveys the expected---but often
overlooked---change in variation between folds as the sample size
increases. The gray band across all three charts has been added to this figure
to emphasize these differences and reflects the variation across folds in (c)
at the maximum observed weight.}
\label{fig:weight_regression}
\vspace{-0.15cm}
\end{figure*}
The prototype interface, shown in Figure~\ref{fig:baseline_prototype}, includes
three panels. In the center is the visualization canvas itself. A left-side
panel allows users to issue queries and control key parameters to the IR
process. Options include the number of folds ($n$) for the partition function,
the use of sampling with replacement, support for random or ordered incremental
sampling, and controls to unfold the merged statistics to show individual folds
within the visualizations. The right-side panel shows detailed descriptive
statistics for both the individual folds ($s_j$) and the aggregate dataset
($m_i'$).
Figure~\ref{fig:confidence} shows a series of bar charts rendered
using the IR prototype to visualize the gender distribution across three
subpopulations from the ICU stay database. This example is directly analogous
to the roulette wheel bar chart example introduced in
Section~\ref{sec:prediction}, as both summarize the distribution of a binary
variable in a given population.
The top row of charts in Figure~\ref{fig:confidence} shows the aggregate gender distribution
for each of the three populations. The charts show a relatively similar
distribution across all three populations, with a moderate increase in female
representation moving from panel (a) to (b) to (c). The bar chart shows the
gender breakdown in each population quite clearly. However, there is no
indication of the distribution's stability across different groups of patients.
Consumers of the visualization are left to assume that the bar charts provide
an accurate depiction.
Panels (d-f) in Figure~\ref{fig:confidence} show the exact same populations
as panels (a-c), respectively. However, these views incorporate measures
computed for multiple folds ($n=7$) using the IR process of partitioning and
merging. These unfolded views provide a more accurate picture regarding the
repeatability of the gender distributions in the top row of the figure. In
particular, we see from Figure~\ref{fig:confidence}(d) that the population
visualized in the left column of the figure is not very predictable.
Meanwhile, far less variation across folds is visible in
Figure~\ref{fig:confidence}(f). In this case, the major difference is the size
of the respective populations which range from about 100 to roughly 10,000
patients. This is a critical factor to interpretation which is invisible in
the original bar charts and easily overlooked even by expert users (e.g.,
\cite{sutcliffe_developing_2014}).
Figure~\ref{fig:weight_regression}, meanwhile, shows three screenshots of the
linear regression portion of the IR prototype applied to data from the same ICU
repository used for the bar charts. In this case, the examples show data for
populations of neonates on a scatter plot, with the x position determined by
weight and the y position determined by height. A linear regression model was
calculated in all three cases using the IR pipeline with $n=5$. The five
regression lines, one for each fold, are visible (``unfolded'') in the
visualizations as red lines. In addition, an aggregate best-fit linear model
is shown in blue.
To explore how IR helps convey uncertainty during progressive analysis, we
used the incremental sampling feature of the prototype to vary the number of
samples while keeping all IR parameters constant. In
Figure~\ref{fig:weight_regression}(a), only 500 patients are included in the
scatter plot. As the varying slopes between the five red lines captures, there
is relatively large disagreement across folds in the linear models they
produce. This uncertainty would be invisible in a traditional plot rendered without
the folds.
As expected, the spread between the individual fold regression lines decreases
as more patients are retrieved by the incremental query feature. For example,
Figure~\ref{fig:weight_regression}(b) shows the same visualization with the
same $n=5$ folds. However, this version includes data for 1,000 patients. The
larger sample size results in increased stability across the folds. Part (c)
of the same figure shows the same visualization with 2,500 patients. We see
little improvement in agreement across folds compared to 1,000 patients,
suggesting that the rate of further gains in agreement will be slower to develop.
As previously stated, the improvement in agreement as sample size increases is
as expected. However, as evidenced by the ``recent history'' charts at casino
roulette tables and the other examples referenced throughout this paper,
visualizations are often assumed to be accurate. Users often fail to consider
issues of sample size or variation. This use case shows that IR can
effectively convey this variation in the data without careful modeling, and in
a non-parametric way that avoids assumptions about the underlying
distributions.
\subsection{DecisionFlow2}
\label{sec:decisionflow_usecase}
To test IR within a more fully-featured exploratory visual analysis
environment, we developed DecisionFlow2, a new IR-based version of our existing
visual analysis system for high-dimensional temporal event sequence
data \cite{gotz_decisionflow:_2014}. A screen capture of the DecisionFlow2 interface
is shown in Figure~\ref{fig:teaser}.
\subsubsection{Original DecisionFlow Design}
The original version of DecisionFlow made heavy use of
p-values to help users identify event types that had a statistically
significant correlation to a user-specified outcome measure. When visualizing
medical data, for example, this approach allows users to find types of medical events
(such as specific diagnoses, medications, and procedures) that---when appearing in a
particular pattern in a patient's history---are associated with better or worse
medical outcomes.
In the original DecisionFlow design, an interactive timeline at the top of the
screen allows users to segment a cohort of event sequences based on the
presence of so-called ``milestone'' events. For a given subgroup, DecisionFlow
visualizes statistics for the potentially thousands of different types that
occur between milestones with the goal of helping users identify good
candidates for new milestones. DecisionFlow conveys the event type statistics
via an interactive bubble chart similar to the one seen in Figure~\ref{fig:teaser}.
In the bubble chart, each event type is represented by a circle whose x-axis
position is determined by its positive support (the fraction of ``good
outcome'' event sequences that contain the event type). Similarly, each
circle's y-axis position is determined by its negative support (the fraction of
``bad outcome'' sequences with the event type). Circle size and color encode
correlation and odds ratio, respectively. Importantly, circles representing
event types whose presence correlates significantly ($p<0.05$) with outcome are
drawn with a distinct border to make it easier for users to visually
distinguish between expected variation and potentially meaningful associations.
\subsubsection{Design Adaptation for IR}
In the IR-based DecisionFlow2 system developed for this paper, a similar
bubble chart design design is adopted to visualize the event type statistics.
However, rather than showing data for measures computed for the overall
population $m_i$, the circles encode aggregate measures computed by an
aggregation function. For example, Figure~\ref{fig:teaser} shows the system
with a bubble chart focused on subset of data containing 45,278 individual
events with 1,148 distinct event types. The support values (used to position
the circles) and other measures were all computed across 5 folds.
The aggregate view in Figure~\ref{fig:teaser}(a) looks essentially identical to the original
DecisionFlow design. This is as intended, with the goal of making IR
compatible with typical visualization designs. However, while the visual
encoding is similar, the number of statistically significant correlations
scores is reduced. In particular, a number of event types that were labeled as
statistically significant in the original design were no longer found to be
significant once majority-voting across the five folds was used to determine
which event types were significant. This makes the visualization system more
selective in rejecting the null hypothesis. The result is a reduction in the
likelihood of Type~1 errors, which are a common problem in high-dimensional
exploratory analysis. More detailed results and discussion are provided in
Section~\ref{sec:df_results}.
Another important part of the IR-based DecisionFlow2 is the ability to
unfold the aggregate statistics for each event type. Users can unfold an event
type by hovering the mouse pointer over the corresponding circle. For example,
after hovering the mouse pointer for a few seconds over the circle shown in
Figure~\ref{fig:teaser}(a), the unfolded representation shown in
Figure~\ref{fig:teaser}(b) is added to the visualization.
As this example shows, the DecisionFlow2 displays the unfolded
data as a convex region drawn around the original circle and outlined with a dashed
border. This region corresponds to the convex hull determined by the $(x,y)$
locations for each of the $n$ folds that contribute to the aggregate measures
that determine the position of the original circle. In other words, the size
and position of the unfolded region represent the variation across folds in
both the positive and negative support measures. Smaller unfolded regions
indicate that the values have little variation across folds. Larger unfolded
regions, such as the one shown in Figure~\ref{fig:teaser}(b), suggest a high
degree of variation between folds and therefore lower confidence in the
repeatability of the aggregate measure.
The typical behavior observed when utilizing the IR-based implementation of
DecisionFlow2 is shown in Figure~\ref{fig:sample_sizes}. Part (a) of the figure
shows an event type from a very large subset of data that shows very limited
variation across folds. This is represented by the very small unfolded region
located near the center of the red circle just above the mouse pointer. Part
(b) of the figure, meanwhile, shows an event type with much higher variation.
This figure, visualizing data from a smaller sample size, demonstrates
what one might expect: findings based on smaller sample sizes have more
variability and therefore should be given less weight in a decision making
process.
However, this very critical difference is not observable via the original
bubble chart. The size of the dataset is made available elsewhere in the user
interface for users who consciously seek it out, but the implications of the
differences in data size are left to the user's imagination. It is only
through the unfolding process that the visualization itself conveys the
difference in confidence that users should place in one view versus the other.
\begin{figure}[t]
\centering
\includegraphics[width=9cm]{figs/big_vs_small_sample_size.pdf}
\caption{In general, (a) smaller differences between folds are seen
when sample sizes are larger, while (b) higher levels of variation are seen
for smaller sample sizes.}
\label{fig:sample_sizes}
\vspace{-0.14cm}
\end{figure}
Moreover, it is critical to note that the size of the dataset is not the sole
determinant of repeatability for a given measure across folds. Major
differences in measure values can be seen even for similarly sized datasets.
For example, Figure~\ref{fig:variation_same_sample_size} shows three different
event types from the exact same subset of event sequences. While the number of
event sequences was the same, the association between \emph{ACE Inhibitors} and
the user-defined outcome (eventual diagnosis with heart failure) was far more consistent across folds.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{figs/variation_same_sample_size.pdf}
\caption{Even with the same sample size, different measures can have
different levels of repeatability across folds. In this example, both (a) and
(c) show relatively high levels of variability, while the small unfolded
region in (b) suggests that the relationship between outcome and ACE
Inhibitors was fairly consistent across all five folds. All three views were
calculated using identical sample sizes.}
\label{fig:variation_same_sample_size}
\vspace{-0.2cm}
\end{figure*}
\subsubsection{Results and Analysis}
\label{sec:df_results}
The IR-based DecisionFlow2 prototype provides visual feedback regarding the
variation in positive and negative support. As previously described, the
system also uses IR to assess the statistical significance of each event type's
correlation with patient outcome. For a given event type, correlation
coefficients and p-values are computed for each fold, then aggregated via
majority-vote. Event types with more than $n/2$ folds showing statistical
significant correlation are displayed in the visualization with a distinct
visual encoding.
To better understand the impact of IR and the choice of $n$ on the visualized
results, we conducted a quantitative experiment in which we compared performance
for a sample user interaction sequence under various conditions. More
specifically, we experimented repeatedly by performing the exact same
exploratory analysis steps using DecisionFlow2, using the exact same input
data, varying only the number of folds. The experiment was conducted at three
partition settings: $n=\{1,3,5\}$.
In all three cases, the input dataset consisted of event data from the medical
records of 2,899 patients containing 1,074,435 individual medical events.
These timestamped events contained 3,631 distinct medical event types: specific
diagnoses, lab tests, or medication orders that were present in the patients'
records.
Of the 3,631 distinct event types, 381 were deemed prevalent enough by the
DecisionFlow2 system to be the target of correlation analysis within the metric
function. The same threshold was used across all three partition settings,
allowing us to compare analysis results across the exact same control
conditions. The results of our analysis are shown in
Table~\ref{tab:df_results}.
With $n=1$, the DecisionFlow2 system flagged statistically significant results
in the same way as in the original paper \cite{gotz_decisionflow:_2014}. Using
a threshold of $p<0.05$, 144 statistically significant event types were
detected. When $n$ was increased to three, the numbers were reduced
dramatically. Only 50 of the original 144 statistically significant event
types remained after applying a majority-vote aggregation algorithm. Of those
50, only 43 were significant across all three folds. For $n=5$, the number of
significant event types was even smaller. The stricter requirements for
replication resulted in just 24 event types being flagged as significant given
a majority-vote aggregation algorithm, and just 15 event types were significant
across all 5 folds.
\begin{table}[t]
\centering
\vspace{0.3cm}
\begin{tabular}{lccc}
\toprule
Number of Folds & $n=1$ & $n=3$ & $n=5$ \\
\midrule
Unanimously Significant & 144 & 43 & 15 \\
Majority Significant & 144 & 50 & 24 \\
At Least One Significant & 144 & 56 & 29 \\
\midrule
Total Number of Measurements Made & \multicolumn{3}{c}{381} \\
Total Number of Event Types & \multicolumn{3}{c}{3,631} \\
\bottomrule
\end{tabular}
\caption{A comparison of statistically significant findings in
three different IR configurations with DecisionFlow2 applied to the same data.
The number of event types flagged as significantly associated with outcome
was largest for $n=1$. This setting corresponds to a traditional visualization
approach with no partitioning. Larger $n$ values dramatically reduced the
number of significant findings.}
\label{tab:df_results}
\end{table}
As expected---and as intended---the number of statistically significant findings
is reduced as $n$ grows from one to five. There are two primary reasons for this
reduction. First, because each condition is applied to the same set of event
sequences for the same patients, the partition size is smaller as $n$
increases. The smaller number of patients reduces the statistical power for
each partition. The expected impact of this is higher p-values and fewer
statistically significant findings. With the ever-growing size of datasets in
many applications, however, the impact on statistical power due to partitioning
should be minimal in many use cases. At the same time, the majority vote
aggregation function requires that a significant level be repeatedly observed
across multiple partitions ($2$ for $n=3$, or $3$ for $n=5$). This reduces the
likelihood of random variation being misinterpreted.
While statistical significance based on p-value thresholds has known
limitations to medical research and beyond (e.g., \cite{goodman_toward_1999}),
it is a widely used metric in exploratory visualization because it allows for
a rough filtering of data to manage visual complexity and the user's analytic
attention. Follow-up analysis of any discovered insights is required. For
this reason, reducing Type~1 errors becomes critical for modern visual analysis
applications where vast numbers of data points can be tested and prioritized
for user analysis. As the results presented here show, IR applies a higher
bar for statistical significance which has the potential to limit unsupported
conclusions from the data in cases where users make quick predictive assessments directly from a visualization. It can also save significant effort in cases where follow-up analysis is performed by reducing the number of falsely generated hypotheses.
\section{Discussion of Limitations}
\label{sec:discussion}
The IR approach is designed to embed the process of replication directly within
the visualization pipeline, providing a non-parametric approach to calculating
and visualizing the repeatability of derived measures. As the examples in
Section~\ref{sec:usecase} demonstrate, the approach can be effective when
applied to a variety of different measures and visual metaphors.
However, there are limitations to IR that must be acknowledged.
First, the proposed approach does nothing to combat selection bias or other
problems in the creation of the original dataset. Any systemic sampling
biases in the original data will be present across all folds created by the
partitioning algorithm. Therefore, even measures that generalize well across
multiple partitions are not necessarily generalizable to entirely new
datasets.
Second, the IR approach is not truly predictive in nature. While information
about the ability of various measures to replicate across multiple folds can
be useful in vetting potential conclusions, findings uncovered via IR should
be considered hypotheses that require testing using more rigorous
methods when important decisions are to be made.
In particular, hypothesis testing often requires the collection and analysis of
new data to fully understand the conditions under which a given insight holds
true. Our method does not replace this step. Instead, IR helps reduce the
number of Type~1 errors, which can lower the number of conclusions that need
testing. However, IR does not eliminate the necessity of a post-hypothesis
validation process.
|
\section*{Abstract}
The cumulant analysis plays an important role in
non Gaussian distributed data analysis. The shares' prices returns are good
example of such data. The purpose of this research is to
develop the cumulant based algorithm and use it to determine eigenvectors that
represent investment portfolios with low variability. Such algorithm is based
on the Alternating Least Square method and involves the
simultaneous minimisation $2$'nd -- $6$'th
cumulants of the multidimensional random variable (percentage shares' returns
of many companies). Then the algorithm was tested during the recent crash on
the Warsaw Stock Exchange. To determine incoming crash and provide enter and
exit signal for the investment strategy the Hurst exponent was calculated using
the local DFA. It was shown that introduced algorithm is on average better that
benchmark and other portfolio determination methods, but only within
examination window determined by low values of the Hurst exponent. Remark
that the algorithm of is based on cumulant tensors up to the $6$'th order
calculated for a multidimensional random variable, what
is the novel idea. It can be expected that the algorithm would be useful in the
financial data analysis on the world wide scale as well as in the analysis of
other types of non Gaussian distributed data.
\paragraph{Keywords} cumulant tensors, ALS--class algorithm, Hurst exponent,
financial data analysis, stock exchange.
\section{Introduction}
Let us consider the multidimensional frequency distribution of shares' prices'
percentage returns. The optimization (minimization) of higher cumulants of this
distribution is used to determine investment portfolios, to test if they are
better on average than the benchmark, during the crash.
The proposed procedure is based on \cite{morton2009algebraic} and
implies the investigation of cumulants tensors -- the $n$'th cumulant of
the
multidimensional
random variable is represented by the $n$--dimensional tensor
\cite{morton2009algebraic,
kolda2009tensor}. For this purpose, I introduce the generalisation of the
classical
Value at Risk (VaR) procedure \cite{best2000implementing}, where the
left Eigenvector Decomposition (EVD) of the second cumulant
(the covariance) matrix is performed, and the multidimensional normal
distribution of financial data is assumed. In classical EVD approach, the
portfolio with minimal variance corresponds to the last
eigenvector. However, the classical EVD method fails to anticipate the risk of
investment
portfolios
since the
second cumulant fails to represent the extreme events, where drops of shares'
prices values are high and cross--correlated. This happens mainly due to the
break down of the central limit theorem resulting from the time--varying
variance of financial data. The Autoregressive Conditional Heteroskedasticity
(ARCH), that violates both independence and identical distribution assumptions
of the central limit theorem, was recorded for many types of financial data
\cite{akgiray1989conditional, bollerslev1987conditionally,
bollerslev1992arch, engle1995multivariate, schwert1990heteroskedasticity}.
Recall also the impact of long
range auto--correlations of
shares' returns
\cite{mandelbrot1997variation, grech2008local, czarnecki2008comparison,
vasconcelos2004guided, domino2011use, domino2012use}. It is worth to mention
the work
\cite{malevergne2002multi, rubinstein2006multi}, where authors shows that
moments or cumulants (of order $6$ or $8$)
may be necessary to account for the severe price fluctuations, that are usually
observed at short time scales (e.g. portfolio rebalanced at a weekly, time
scales). In my research I would examine portfolio rebalanced at the $20$
trading days (approximately monthly scale) as it is often performed in practice
in assets management. To search for the severe price fluctuations I used the
Hurst exponent indicator.
Following this arguments, high cumulants analysis should anticipate
extreme events, improving the search for portfolios with low variability. There
are some works
implying the use of $2$'nd, $3$'rd and $4$'th cumulant of
multivariate shares' returns \cite{arismendi2014monte, jondeau2015moment}.
In this research I use the $5$'th and the $6$'th cumulant as well, what
is a new approach for multivariate shares' returns. In
general the proposed algorithm is based on the High Order Singular
Value Decomposition (HOSVD) and Alternating Least Square (ALS) procedure
\cite{kolda2009tensor}.
To compare the proposed method with others (such as EVD), the author, for each
method, creates the family of investment portfolios which are supposed to be
safer than a benchmark. Then
portfolios are compared using the result function that is an average percentage
change of portfolios' values -- an average portfolio results.
Other result functions are also discussed:
\begin{enumerate}
\item a mode of percentage change of portfolios' values,
\item a maximal loss / minimal gain -- the result of the ``worst
portfolio'',
\item a minimal loss/ maximal gain -- the result of the ``best portfolio''.
\end{enumerate}
The major motivation for this research is to introduce the automation method of
analysis of data that are not Gaussian distributed. Good example of such data
are financial data, especially during the rupture and crash period.
It is why, I focus in this work, on the financial data analysis. To determine
the rupture and crash period and introduce the enter and exit signal of an
investment strategy, I use the Hurst exponent indicator calculated for the
WIG20 index, using the local DFA. This paper give some additional incentive for
the development of cumulants tensors calculation method at low computational
complexity. Afterwards the
multi--cumulant analysis may be applied for large financial data sets and
tested against many crashes on many markets. Additionally the method may be
used to analyse other (non--financial) data that are not Gaussian distributed.
\section{The classical approach, the covariance matrix EVD}
Let us take the
$M$--dimensional random
variable of size $T$, $\textbf{X} \in \mathbb{R}^{( T \times M)}$, being
the percentage
returns of $M$ shares. Its marginal variables are $X_i$, and values are
$x_{t,i}$:
\begin{equation}\label{eq::rw}
\textbf{X} = [X_1, \ldots, X_i, \ldots, X_M] = \left[ \begin{array}{ccc}
x_{t=1, 1} & \cdots & x_{t=1, M} \\
\vdots & \vdots & \vdots \\
x_{t=T, 1} & \cdots & x_{t=T, M} \\
\end{array} \right].
\end{equation}
An unbiased estimator of variance of the $i$'th marginal random variable
($X_i$) is:
\begin{equation}
\sigma^2_i = \frac{1}{T-1} \sum_{t=1}^{T}(x_{t,i} -
\overline{X_{i}})^2,
\end{equation}
and an unbiased estimator of covariance between ($X_i$) and ($X_j$) is:
\begin{equation}
\textrm{cov}_{i,j} = \frac{1}{T-1} \sum_{t=1}^{T}(x_{t,i} -
\overline{X_{i}})(x_{t,j} - \overline{X_{j}}).
\end{equation}\label{f::cov2}
The variance and the covariance can be represented by the $M \times M$
symmetric
covariance matrix, called also the
second cumulant matrix -- $C_2$ (notice $\sigma_i^2 = \textrm{cov}_{i,i}$):
\begin{equation}\label{f::cov3}
C_2 = \left[ \begin{array}{cccc}
\sigma^{2}_{1} & \textrm{cov}_{1,2} & \cdots & \textrm{cov}_{1,L} \\
\textrm{cov}_{2,1} & \sigma^{2}_{2} & \cdots & \textrm{cov}_{2,L} \\
\vdots & \vdots & \ddots & \vdots \\
\textrm{cov}_{L,1} & \textrm{cov}_{L,2} &\cdots & \sigma^{2}_{L} \\
\end{array} \right].
\end{equation}
\begin{definition}{The Eigenvalue Decomposition -- EVD.}\label{d::EVD}
Consider the covariance (second cumulant) symmetric matrix. The matrix can be
diagonalized in the following way:
\begin{equation}
C_2 = V \Sigma V^{\intercal},
\end{equation}\label{f::cov4}
where $\Sigma = V^{\intercal} C_2 V$ is the diagonal matrix with diagonal
values $\sigma'^2_{i} = (V^{\intercal} C_2 V)_{ii}$ and $V$ is unitary $M
\times M$ factors matrix, such that $\sigma'^2_{i}$ are sorted in descending
order:
\begin{equation}
\Sigma = \left[ \begin{array}{cccc}
\sigma'^2_{1} & 0 & \cdots & 0 \\
0 & \sigma'^2_{2} & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 &\cdots & \sigma'^2_{M} \\
\end{array} \right].
\end{equation}
The $i$'th column of $V$ is the eigenvector that corresponds with the
eigenvalue
$\sigma'^2_{i}$. Rows in the $i$'th column of $V$ are factors that give the
linear combination of marginal random variables with the combination's variance
$\sigma'^2_{i}$.
The last eigenvector
would
give the linear combination of marginal random variables with the smallest
combination's
variance -- $\sigma'^2_{M}$.
\end{definition}
The classical EVD procedure has been often used in the portfolio risk
determination.
However, it requires the multidimensional Gaussian distribution of
shares' returns, where all information about the variability of the frequency
distribution is stored in the covariance matrix. As mentioned before the
financial data
(shares' returns) are not Gaussian
distributed and the classical EVD procedure has often
failed in the investment portfolio's risk
determination \cite{cherubini2004copula}. It is why the author proposes to
extend the
classical EVD procedure by taking into consideration also cumulants of order
higher
than $2$ -- the higher cumulants.
\section{Cumulants}
Let us consider the $M$ dimensional random variable
$\textbf{X} = [X_1, \ldots, X_M]$. The $n$'th cumulant $C_n$ of
such variable is
the $n$--mode tensor \cite{kolda2009tensor}, with elements
$\kappa_{\alpha_1,\ldots,\alpha_n}(\textbf{X})$ \cite{kendall1946advanced,
lukacs1970characteristics}:
\begin{equation}\label{eq:generating-func}
\kappa_{\alpha_1, \ldots, \alpha_n}(\textbf{X}) =
\frac{\partial^n}{\partial \tau_{\alpha_1},\partial \tau_{\alpha_2},\ldots,
\partial \tau_{\alpha_n}} \log\left(E\left(\exp(
\tau\cdot\textbf{X}^{\intercal})\right)
\right) \bigg|_{\tau = 0}.
\end{equation}
where $\tau$ is the argument vector $\tau = \underbrace{[\tau_1, \cdots,
\tau_i, \cdots \tau_M]}_M$, and $E()$ is the expected value operator.
Formulas used to calculate cumulants up to $4$'th order are well known
\cite{kendall1946advanced, lukacs1970characteristics}. The author
has calculated $5$'th and $6$'th cumulants by the direct use of
(\ref{eq:generating-func}). Here analysed data were substituted for the
random variable
$\textbf{X} \in \mathbb{R}^{( T \times M)}$, and
computer
differentiations were performed
at point $\tau =
\underbrace{[\tau_1, \cdots
\tau_M]}_M = 0$, using ForwardDiff and DualNumbers library in Julia
programming \cite{juliafd}.
\subsection{The multi--cumulant decomposition.}
To investigate the financial data the author takes many cumulant tensors $
C_2, \dots, C_n$, where $n = 4$ or $n = 6$. The calculation of cumulants
of order $n > 6$ might require larger data series, but non--stationary of
financial data
\cite{grech2008local} makes the investigation of long time series less adequate
than shorter data series. To achieve the factor matrix $V$, the author proposes
the following ALS--class
algorithm, where the
search for the local maximum of the function $\Phi(V)$
is performed \cite{de2004dimensionality, savas2010quasi}.
Following the maximisation procedure which can not be solved precisely, the
author will
find the local maximum using the iteration procedure \cite{savas2010quasi} and
show that the results are meaningful.
\begin{definition}{The $\Phi(V)$ function.}
Consider the $i$'th core--tensor $T_i$ that is the contraction of
$C_i$ tensor and $i$ factor matrices $V$:
\begin{equation}
(T_{i})_{l_1, \cdots, l_i} = \sum_{j_1, \cdots, j_i} (C_{i})_{j_1, \cdots, j_i}
V_{j_1 l_1} \cdots V_{j_i l_i}.
\end{equation}
The ALS procedure
proposed in
\cite{morton2009algebraic,
savas2010quasi} refers
to the search for the common factor matrix $V$ that maximise $\Phi_4(V)$.
\begin{equation}\label{eq::f4}
\Phi_{4}(V) = \frac{1}{2!} ||V^{\intercal} C_2 V||^2 + \sum_{i =
3}^4\frac{1}{i!}
||T_i||^2.
\end{equation}
The author proposes to extend the analysis up to the $6$'th cumulant which are
more
sensitive to extreme ``tail events''. Hence the author defines $\Phi_6(V)$:
\begin{equation}\label{eq::f5}
\Phi_6(V) = \frac{1}{2!}||V^{\intercal} C_2 V||^2 + \sum_{i = 3}^6\frac{1}{i!}
||T_i||^2.
\end{equation}
\end{definition}
To find the common factor matrix $V$, the ALS--based algorithm is proposed by
author and presented at subsection~(3.2). The idea of the algorithm is
based on the algorithm proposed in
\cite{de2004dimensionality} where the iteration procedure was used for the
search for the local maximum of the following function:
\begin{equation}\label{eq::f3}
\Phi'(V) = ||V^{\intercal} C_2 V||^2 + \alpha_n||T_n
||^2.
\end{equation}
The proposed algorithm works for the general case (any $\Phi_n(V)$), but
computations were
performed for $n = 4$ and $n = 6$. Racall that ALS
algorithms move information into the upper left corner of the
core--tensor and order the information in the
sense of the Frobenius Norm. Take the linear
transformation of analysed data $\textbf{X}$:
\begin{equation}\label{eq::comb}
\textbf{Y} = \textbf{X} V,
\end{equation}
where $\textbf{Y} = [Y_1, \ldots, Y_j, \dots, Y_M]$. Here $Y_j$ represents
percentage returns of the $j$'th portfolio. Elements of $\textbf{Y}$ are:
\begin{equation}\label{eq::combs}
y_{t,j} = \sum_{i= 1}^M x_{t, i}V_{i,j}
\end{equation}
The rear columns of the factor matrix would
give the investment portfolio with little variability.
\subsection{The algorithm.}
The algorithm used to determine the factor matrix $V$ given cumulant symmetric
tensors $C_2 \cdots C_n$, it is a general algorithm and work for each $n \geq
3$. Let ${C_i}_{(1)}$ be the unfold of the tensor $C_i$ in the first mode
\cite{kolda2009tensor}. The first factor matrix anzatz is computed as a matrix
that columns are left eigenvectors of the following matrix:
\begin{equation}
\left[\frac{C_2}{2!} \cdots \frac{{C_i}_{(1)}}{i!} \cdots
\frac{{C_n}_{(1)}}{n!}\right]
\end{equation}
At $k$'th interaction, we have the $(V_{k-1})$ factor matrix. Now the following
procedure is performed. The contraction of the
$C_i$ tensor (matrix) and $i-1$ factor matrices $(V_{k-1})$ is performed:
\begin{equation}
(S_{i})_{j_1, l_2 \cdots, l_i} = \sum_{j_2, \cdots, j_i} (C_{i})_{j_1, j_2
\cdots, j_i} (V_{k-1})_{j_2 l_2} \cdots(V_{k-1})_{j_i l_i}.
\end{equation}
To compute $V_k$ we takes left eigenvectors of the following matrix:
\begin{equation}
\left[\frac{S_2}{2!} \cdots \frac{{S_i}_{(1)}}{i!} \cdots
\frac{{S_n}_{(1)}}{n!}\right]
\end{equation}
The procedure is repeated to satisfaction the stop condition.
\section{The investigation of financial data.}
The cumulant analysis was performed in the optimal portfolios
searching
problem. Let us consider the price of a $i$'th share at time $t$ -- $p_{t,i}$.
Its percentage return is
\begin{equation}
x_{t,i} = \frac{p_{t,i}-p_{(t-1),i}}{p_{(t-1),i}} \cdot 100 \%.
\end{equation}
In our case $t$ numerates trading days (the analyse of daily
returns was performed) and $p_{t,i}$ the closing price of $i$'th share the
given trading day numbered by
$t$. Next the multidimensional random variable $\textbf{X}$ of percentage
returns is
constructed.
To construct investment portfolios we use the factor matrix
$V$. The $j$'th portfolio returns are one dimensional random variable
$Y_j$ with elements $y_{t,j}$.
The naive method of factor matrix determination
uses the Eigenvalue Decomposition (EVD) of the
covariance matrix \cite{best2000implementing}. This procedure is not fully
adequate since shares returns are not
Gaussian distributed, especially the rupture and crisis period \cite{
mandelbrot1997variation, grech2008local, czarnecki2008comparison,
vasconcelos2004guided} -- importantly such period can be predicted by the use
of the Hurst
exponent.
To anticipate higher cumulants of shares returns as well, the author proposes to
determine the factor matrix $V$ by searching for the local maximum of the
$\Phi_4(V)$ function as well as $\Phi_6(V)$ function -- using cumulant tensors
up to the $6$'th order, what is a new approach. The proposed $\Phi_6(V)$ method
is used to chose portfolios with returns that have low absolute values of high
cumulants. Hence the method is supposed to work well where the portfolio's
variability is a disadvantage. It happens during the crash of the financial
market, hence the author tests the method
during the last rupture and crisis on the Warsaw Stock Exchange.
\subsection{The data analysis.}
The author has examined $M = 10$ dimensional random variable Tab.~(1),
being daily percentage returns of the shares of $10$ most liquid companies from
the WIG20 index at the time 12.05.2010 -- 04.08.2016 (the WIG20 index includes
$20$ most liquid companies traded on the Warsaw Stock
Exchange). Recent composition of the WIG20 index is presented in
Fig.~(\ref{f::wh})
\begin{figure}
\includegraphics{WHt}
\caption{WIG20 index and the Hurst exponent.}
\label{f::wh}
\end{figure}
The WIG20 index reached maximum at 14.05.2015 and then has fallen rapidly --
the crash has occurred. To introduce the signal of incoming crash, the Hurst
exponent was calculated for the WIG20 index using the local Detrended
Fluctuation Analysis (DFA) \cite{grech2008local, domino2011use}.
\begin{table}\label{tab:portions}
\centering
\begin{tabular}{|c|c|c|c|}
\hline $i$ & company & contribution & contribution to \\ & & to WIG20
\% &
benchmark
\% $(BP_i)$ \\
\hline 1& PKOBP & 14.64 & 18.31 \\
\hline 2& PZU & 14.04 & 17.55 \\
\hline 3& PEKAO & 11.65 & 14.57 \\
\hline 4& PKNORLEN & 8.45 & 10.57 \\
\hline 5& PGE & 7.52 & 9.40\\
\hline 6& KGHM & 7.14 & 8.93 \\
\hline 7& BZWBK & 5.21 & 6.51 \\
\hline 8& LPP & 4.77 & 5.96 \\
\hline 9& PGNIG & 3.55 & 4.43 \\
\hline 10& MBANK & 3.00 & 3.75 \\
\hline
\end{tabular}
\caption{the 10 most liquid companies of the WIG20 index, their value
contribution to the
WIG20 index (at 20.03.2015) and as their value contribution to proposed
benchmark
portfolio.}
\end{table}
\subsubsection{The Hurst exponent.} To determine the rupture and crisis period
of the stock exchange, where the examined investment strategy was tested the
Hurst exponent was calculated using the local DFA. The parameters for DFA were
the same as in \cite{domino2014use}: $500$ days long observation window was
used to examine past closing value of the WIG20 index. Having the Hurst
exponent, I introduce the
signals of entry and exit for proposed
investment strategy. Recall that in \cite{grech2008local} the Hurst exponent
was calculated
using the local DFA for the index of Polish Stock Exchange, and it was shown
that
before a crash (near a rupture point), the Hurst Exponent has minima $\lesssim
0.4$. Hence the entry threshold value was chosen as $H_{entry} = 0.4$. The exit
threshold was chosen as $H_{exit} = 0.425$ -- data with high negative
auto--correlation was chosen for a test.
Regarding the recent crash the entry signal
occurred at 19.12.2014 and the exit signal at 10.09.2015. To examine the
algorithm, I introduce the $20$ trading days (approx. 1 months) long investment
windows -- as it is performed in practice assets management. First window
starts a day after the
enter signal -- 22.12.2014, and there are $9$ windows within a test period
22.12.2014 -- 27.01.2015, 27.01.2015 --
24.02.2015, 24.02.2015 -- 24.03.2015, 24.03.2015 -- 23.04.2015, 23.04.2015 --
22.05.2015, 22.05.2015 -- 22.06.2015, 22.06.2015 -- 20.07.2015, 20.07.2015 --
17.08.2015, 17.08.2015 -- 14.09.2015 (the last window
ends just after exit point). For each window, cumulants are
calculated using a test series of length $T = 1100$, that ends
just before the examination window. Next
investment returns are analysed for data in given window -- the testing
set. In next subsections the
analysis is discussed in details for the $7$'th window of 22.06.2015 --
20.07.2015.
Then the analysis results are presented for other windows.
\subsubsection{Optimal portfolios determination -- training.}
Let us discuss in details the procedure for the exemplary window of 22.06.2015
-- 20.07.2015.
Given the training set, the factor matrix is determined using different methods,
such as EVD, $\Phi_4(V)$ and $\Phi_6(V)$. Here also the Independent Component
Analysis
(ICA) was used for more general comparison. The $\Phi_4(V)$ method requires
the
calculation of $3$'rd and $4$'th cumulants. For $\Phi_6(V)$ also $5$'th and
$6$'th cumulant tensors are required, which were calculated by the direct use of
Eq.~(\ref{eq:generating-func}).
Given $\Phi_4(V)$ and $\Phi_6(V)$ the algorithm introduced in subsection~(3.2)
was
used for the
factor matrix $V$
determination.
In Fig.~(\ref{f::cu}) some cumulant value of the one dimensional random
variable, that is
the $j$'th investment portfolio $Y_j$ (with elements $y_{t,i} = \sum_{i=1}^{M}
x_{t,i} V_{i,j}$) are
presented for different methods of the factor matrix determination.
Generally large cumulants values were stored in first
portfolios where $j = 1, 2, \cdots$. For further investigation I took the $5$
rear portfolios, where $j \in [5,10]$ as those that have low cumulants'
absolute values.
\begin{figure}
\subfloat[second cumulant]{
\includegraphics{cumulant21false.pdf}}
\subfloat[third cumulant (normalized)]{
\includegraphics{cumulant31true.pdf}} \\
\subfloat[fourth cumulant (normalized)]{
\includegraphics{cumulant41true.pdf}}
\subfloat[sixth cumulant (normalized)]{
\includegraphics{cumulant61true.pdf}}
\caption{Values of cumulants for portfolios, the larger $j$ the portfolio
is supposed to be less ``variable''.}
\label{f::cu}
\end{figure}
\subsubsection{Testing optimal portfolios.}
After the training (the determination of $V$) has been completed, the testing
of
portfolios is performed. The factor matrices ($V$) columns contain both
positive
and negative values, the later corresponds to the negative value
of shares in the portfolio -- the short sale. To diminish the use of the sort
sale,
the test portfolios were compared with the benchmark portfolio. Shares values
contributions in benchmark portfolio -- $BP_i$ are given in Tab.~(1).
In proposed test portfolios the value contribution of the $i$'th
share in the $j$'th portfolio would be:
\begin{equation}{\label{eq::test}}
TV_{i,j} = \frac{\alpha BP_i + V_{i,j}}{\sum_{i =
1}^{10} \left(\alpha BP_i + V_{i,j}\right)},
\end{equation}
the $\alpha
= 7$ was taken, to make cases of the short sale rare.
For testing, shares prices of companies, see Tab.~(1) were
taken. Testing set is represented by:
${p_{t',i}}$, where $t'$ is time in the testing window. The percentage return
of $j$'th portfolio after $L$ trading days is:
\begin{equation}
Pr_j(L) =
\frac{\sum_{i=1}^{10} TV_{i,j}\left(\frac{p_{(t'=L+1,i)}}{p_{(t'=1,i)}}\right)
- \sum_{i= 1}^{10}TV_{i,j}}{\sum_{i=
1}^{10} TV_{i,j}} = \sum_{i=1}^{10}
TV_{i,j}\left(\frac{p_{(t'=L+1,i)}}{p_{(t'=1,i)}} \right).
\end{equation}
In Fig.~(\ref{f::por}), returns after $10$ and $20$ trading days are
presented. Remark, in this research transaction costs were
not taken into account. The benchmark portfolio contributions can be reproduced
by simply substituting $\forall_{i,j} \ V_{i,j} = 0$ to Eq.~(\ref{eq::test}).
\subsection{Discussion.} Analysing Fig.~(\ref{f::por}), one can see that
the $\Phi_6(V)$ method gives in both cases
$3$ portfolios that are better than the benchmark and $2$ that are as good as
benchmark. In the reminding part of the paper, I discuss the statistics of
returns of such $5$ portfolios.
\begin{figure}
\subfloat[After $10$ trading days]{
\includegraphics{portfolio106true.pdf}}
\subfloat[After $20$ trading days]{
\includegraphics{portfolio206true.pdf}}
\caption{Returns of $5$ portfolios, investments window
22.06.2015 -- 20.07.2015.}
\label{f::por}
\end{figure}
One can also conclude, that each method of factor matrix determination
($\Phi_6(V)$, $\Phi_4(V)$, EVD, ICA) produces the worst
portfolio
\begin{equation}
\min_{j \in [6,10]}Pr_j(L),
\end{equation}
which return is minimal and
often smaller than benchmark's return. Those minimum of portfolios' returns are
presented in Fig.~(\ref{eq::min1}). Analysing minimum of
portfolios' returns one can
conclude that out of all methods ($\Phi_6(V)$, $\Phi_4(V)$, EVD, ICA) the
$\Phi_6(V)$ method gives smallest loss -- its worst portfolio is almost as good
as benchmark. The worst results gives the ICA method, this is due to large
variability of returns -- see Fig.~(\ref{f::por}), such method is not desirable
during a crisis.
Similarly best portfolios can found:
\begin{equation}
\max_{j \in [6,10]}Pr_j(L),
\end{equation}
Their results are presented in Fig.~(\ref{eq::max1}). Best
results presents the ICA, however this method is not safe. Next best are
$\Phi_6(V)$ and $\Phi_4(V)$.
It is worth checking now, which portfolio determination method is best on
average. In Fig.~~(\ref{eq::m1}), mean values of portfolios'
returns are
presented:
\begin{equation}
\frac{1}{5}\sum_{j=6}^{10}Pr_j(L).
\end{equation}
Remark that all methods but $\Phi_6(V)$ give an average return similar to or
worse than the benchmark. It is a worthy result, since it is hard to beat
the benchmark on average.
\begin{figure}
\subfloat[minimum of portfolio's returns]{\label{eq::min1}
\includegraphics{test20150622minimum1206}}
\subfloat[maximum of portfolio's returns]{\label{eq::max1}
\includegraphics{test20150622maximum5206}}\\
\subfloat[mean of portfolio's returns]{\label{eq::m1}
\includegraphics{test20150622mean2206}}
\subfloat[mode of portfolio's returns]{\label{eq::mo1}
\includegraphics{test20150622mode4206}}
\caption{Statistics of investment returns, $7$'th window of
22.06.2015 -- 20.07.2015.}
\label{eq::stat3}
\end{figure}
To examine a typical portfolio, the mode of portfolios' returns can
be mentioned as well -- see Fig.~(\ref{eq::mo1}), here
$\Phi_6(V)$ gives results, better than other methods, and slightly better the
the benchmark. Concluding, statistics of the $\phi_6(V)$ method are better than
other methods and the benchmark. Results of other $20$ days
observation windows within the observation period determined by the Hurst
exponent and outside it are discussed in next subsection.
\subsection{All observation windows.}
The analysis was performed for following observation windows: 22.12.2014 --
27.01.2015, 27.01.2015 --
24.02.2015, 24.02.2015 -- 24.03.2015, 24.03.2015 -- 23.04.2015, 23.04.2015 --
22.05.2015, 22.05.2015 -- 22.06.2015, 22.06.2015 -- 20.07.2015, 20.07.2015 --
17.08.2015, 17.08.2015 -- 14.09.2015. The first window starts a trading day
after the enter signal recorded at 19.12.2014. The
WIG20 index increased in first for windows, the maximum appeared in the $5$'th
window where the crisis started, the last ($9$'th window) ends just after the
exit signal recorded at 10.09.2015. Windows $5$ -- $9$ are crisis windows.
Since I an interested in the investment strategy outcome between the enter and
the exit signal, I present
the cumulative results of investment that starts a trading day after the enter
signal and ends just after exit signal. For each window factor matrices are
calculated separately, investment is made at the first point in a window, at
the last point of the window shares are sold and the mean of returns of $5$
portfolios is calculated. Cumulative of such mean returns are presented in
Fig.~(\ref{eq::1w}). In Fig.~(\ref{eq::1w1}) the cumulative results are
presented for crisis portfolios $5$ - $9$, here investment starts at
23.04.2015.
\begin{figure}
\subfloat[whole investment 22.12.2014 -- 14.09.2015]{\label{eq::1w}
\includegraphics{mean_cr91}}
\subfloat[crisis windows 23.04.2015 -- 14.09.2015]{\label{eq::1w1}
\includegraphics{mean_cr95}}
\caption{Cumulative of means of portfolios returns.}
\label{eq::stat5}
\end{figure}
In Fig.~(\ref{eq::stat6}) similar results are presented, but now mode of
returns of $5$ portfolios is calculated in each window and the cumulative
results are presented. Analysing
Fig.~(\ref{eq::stat5}, \ref{eq::stat6}) one can conclude that the $\Phi_6(V)$
method on average gives best results at the exit point and during the crisis.
\begin{figure}
\subfloat[whole investment 22.12.2014 -- 14.09.2015]{\label{eq::1mo}
\includegraphics{mode_cr91}}
\subfloat[crisis windows 23.04.2015 -- 14.09.2015]{\label{eq::1mo1}
\includegraphics{mode_cr95}}
\caption{Cumulative of modes of portfolios returns.}
\label{eq::stat6}
\end{figure}
In Fig.~(\ref{eq::min}) cumulative results are presented, if in each window the
worst portfolio was chosen (unlucky choice) -- there $\Phi_6(V)$ method is
worse than a benchmark, but slightly better than other methods. In
Fig.~(\ref{eq::max}) cumulative results are presented, if in each window the
best portfolio was chosen (lucky choice) -- there $\Phi_6(V)$ method is better
than
all other methods apart from ICA. However the ICA produces also very bad
portfolios (worst minimum), and hence is not adequate for a crisis.
\begin{figure}
\subfloat[Cumulatives of minima of returns]{\label{eq::min}
\includegraphics{min_cr95}}
\subfloat[Cumulatives of maxima of returns]{\label{eq::max}
\includegraphics{max_cr95}}
\caption{Best and worst cumulative result at crisis windows 23.04.2015 --
14.09.2015.}
\label{eq::stat7}
\end{figure}
To test a method a bit more, I introduced observation windows after the exit
signal, and number them as $10$'th to $15$'th, the cumulative results of means
and modes of portfolio returns are presented in
(\ref{eq::stat8}). Investment starts at 14.09.2015 and investment windows are
14.09.2015 -- 12.10.2015, 12.10.2015 -- 09.11.2015,
09.11.2015 -- 08.12.2015, 08.12.2015 -- 12.01.2016 and 12.01.2016 -- 09.02.2016.
It can be concluded that beyond the exit signal the $\Phi_6(V)$ method gives
results similar to other methods and the benchmark. Hence the use of the Hurst
exponent to determine the proper enter and exit signal appears to be crucial.
\begin{figure}
\subfloat[Cumulatives of means of returns]{\label{eq::mean}
\includegraphics{mean_cr1410}}
\subfloat[Cumulatives of modes of returns]{\label{eq::mode}
\includegraphics{mode_cr1410}}
\caption{Statistics for investment after exit signal, 14.09.2015 --
09.02.2016}
\label{eq::stat8}
\end{figure}
\section{Conclusions}
The author has used the multi--cumulant tensor analysis to
analyse financial data and determine optimal investment portfolio with low
absolute values of cumulants of their percentage returns. For this purpose, the
author has analysed daily
returns of shares traded on the Warsaw Stock
Exchange to determine the factor matrix that represents such
portfolios and test them during the recent rupture and crash period on the
Warsaw Stock Exchange.
The main result of this
work is the introduction of the algorithm that uses $2$'nd -- $6$'th cumulant
tensors to analyse multivariate financial data and determine the investment
portfolios that have low variability (low
cumulants' absolute values). The Hurst exponent, calculated by the local DFA
for the WIG20 index, indicates the auto--correlation phase on the stock
market (the rupture period and the early stage of the crisis). At this phase,
the introduced method is on average better than the benchmark and other tested
methods. Importantly the Hurst exponent condition appears to be necessary to
achieve this result. The examination of the
method can be extended in further research, e.g. the algorithm can be tested on
many stock exchanges. The algorithm can also be used to analyse other
(non--financial) data that are
non--Gaussian distributed.
\section*{Acknowledgements}
The research was partially financed by the National Science Centre, Poland -
project number 2014/15/B/ST6/05204
\bibliographystyle{ieeetr}
|
\section{\label{intro}Introduction}
$\beta$-decay spectroscopy is a fundamental tool for the investigation of the nuclear structure of unstable nuclei \cite{Bohr1969,Blank2008,Rubio2009,PhysRevLett.112.222501}. Many neutron-deficient $fp$-shell nuclei lie on the astrophysical $rp$-process reaction pathway. Accordingly the study of the $\beta$ decay of such nuclei is of importance because it provides input to calculations of the $rp$-process and models of X-ray bursters \cite{Schatz1998167,Parikh2013225}. The investigation of the decay of these nuclei is difficult because they lie far away from stability. The odd-odd nuclei are particularly difficult to study because there are often two long-lived states, one of which is the ground state, with similar half-lives. This makes it hard to disentangle the two decays. This is because both states are in general members of the same two-particle multiplet and have therefore very similar structure. The only difference is how the spins of the valence nucleons couple to make the final spin. One strong contribution to the half-life is given by the Fermi transition, which is very fast and has identical strength in the two cases. How different the total half-life will be for the two states will thus depend on the details of the Gamow-Teller transitions. Here we present information on one such case. $^{52}$Co is a $T_z = -1$ odd-odd isotope in the $f_{7/2}$ shell that was first observed in an experiment performed at GANIL \cite{Pougheon1987}. There had been previous indications of the existence of a long-lived $\beta$-decaying excited state but it had not been isolated experimentally \cite{Hagberg1997183}.
The $^{52}$Mn nucleus, the mirror of $^{52}$Co, has a 2$^+$ isomeric state at 378 keV above the 6$^+$ ground state. This $^{52}$Mn isomer, having a half-life of 21.1(2) min \cite{PhysRev.113.602,NDS2015}, decays via two branches, 98.22(5)\% by $\beta^+$ decay to $^{52}$Cr and 1.78(5)\% via an internal transition to the ground state \cite{PhysRevC.16.1581,NDS2015}. Assuming isospin symmetry, a 2$^+$ isomeric state is also expected in $^{52}$Co at a similar energy. This would mean that we have the case of two states with \mbox{$J^{\pi}$ = 2$^+$} and 6$^+$, corresponding to the 2$^+$ and 6$^+$ members of the $(\pi f_{7/2})^{-1}(\nu f_{7/2})^{-3}$ multiplet. The Fermi partial half-life will be of the order of 200 ms, and the total half-life will depend on the distribution and population of the low-lying 1$^+$, 2$^+$, 3$^+$ states in the $^{52}$Fe daughter for the decay of the 2$^+$ isomer, and of the 5$^+$, 6$^+$, 7$^+$ states in $^{52}$Fe for the decay of the 6$^+$ ground state. For instance, in the very similar case of $^{44}$V, with probable structure $(\pi f_{7/2})^{3}(\nu f_{7/2})^{1}$, the two states with \mbox{$J^{\pi}$ = 6$^+$} and 2$^+$ have half-lives of 150(3) ms and 111(7) ms \cite{Hagberg1997183}, respectively.
We have studied the $\beta^+$ decay of $^{52}$Ni to $^{52}$Co in Ref. \cite{PhysRevC.93.044336}. A study of the high-spin states in $^{52}$Co has been carried out recently \cite{Bentley2016}. The $\beta^+$ decay of $^{52}$Co to $^{52}$Fe was studied in Ref. \cite{Hagberg1997183}. The $^{52}$Co ground state, having \mbox{$J^{\pi}$ = 6$^+$} and $T$ = 1, undergoes $\beta$ decay to its Isobaric Analogue State (IAS) in $^{52}$Fe at 5655 keV \cite{Hagberg1997183}. Since the proton separation energy is 7378(7) keV \cite{NDS2015} proton emission is not possible here. A cascade of four $\gamma$-rays (1329, 1942, 1535 and 849 keV) was reported in Ref. \cite{Hagberg1997183} corresponding to the de-excitation of the IAS in $^{52}$Fe through the sequence \mbox{6$^+ (T = 1) \rightarrow 6^+ \rightarrow 4^+ \rightarrow 2^+ \rightarrow 0^+$} [also shown in our decay scheme in Fig. \ref{decay}(b)]. The measured $\gamma$-ray intensities in Ref. \cite{Hagberg1997183} implied the existence of a 31(14)\% $\beta$ feeding to the first excited state in $^{52}$Fe at 849 keV (\mbox{$J^{\pi}$ = 2$^+$}), which is quite unlikely to be due to direct feeding from the 6$^+$ state considering the \mbox{$\Delta L$ = 4} difference between the parent and daughter states. It was therefore suggested in Ref. \cite{Hagberg1997183} that this anomaly could be explained by extra feeding associated with the decay of $^{52m}$Co, although no clear evidence could be found.
In the present paper we report the first observation of the 2$^+$ isomer in $^{52}$Co, which was populated in the $\beta$ decay of $^{52}$Ni. The trick here was not to look at $^{52}$Co as a direct product of the fragmentation reaction, but as a product of the decay of the 0$^+$, $^{52}$Ni ground state (see the partial decay scheme in Fig. \ref{decay}(a)]. This decay process directly populates the 0$^+ (T = 2)$ IAS in $^{52}$Co, which then de-excites via the sequence \mbox{0$^+ \rightarrow 1^+ \rightarrow 2^+$} \cite{PhysRevC.93.044336}. This is a much cleaner way to populate the expected 2$^+$ isomeric state. We have observed the $\gamma$-rays emitted following the $\beta$ decay of the isomer and measured its half-life. The $\beta$-decay Fermi and Gamow-Teller transition strengths, $B$(F) and $B$(GT), have been determined [an upper limit for $B$(F)]. Moreover, selecting the direct production of $^{52}$Co we could obtain data on the $\beta$ decay of the $^{52}$Co, 6$^+$ ground state, which allowed us to add new information on this decay and measure the half-life with improved precision.
\section{\label{exp}The experiment}
\begin{figure}[!b]
\centering
\includegraphics[width=1\columnwidth,trim={0 0.2cm 3.1cm 5.5cm},clip]{ID1.pdf}
\vspace{-5.5 mm}
\caption{$\Delta E$ versus ToF identification plot for the dataset optimized to implant $^{56}$Zn close to the middle of the DSSSD (see Ref. \cite{PhysRevC.93.044336} for details). The positions of the $^{52}$Co and $^{52}$Ni implants are shown.}
\label{ID1plot}
\end{figure}
\begin{figure}[!b]
\centering
\includegraphics[width=1\columnwidth,trim={0 0.2cm 3.0cm 5.6cm},clip]{ID2.pdf}
\vspace{-5.5 mm}
\caption{$\Delta E$ versus ToF identification plot for the dataset optimized for $^{48}$Fe. The positions of $^{52}$Co and $^{52}$Ni are shown.}
\label{ID2plot}
\vspace{-5.0 mm}
\end{figure}
We have studied the $\beta^+$ decay of $^{52}$Ni to $^{52}$Co in an experiment done at GANIL \cite{PhysRevC.93.044336}. $^{52}$Ni was produced by fragmenting a $^{58}$Ni$^{26+}$ primary beam, accelerated to 74.5 MeV/nucleon, on a 200-$\mu$m-thick natural Ni target. $^{52}$Co was also produced directly in the same experiment. After selection of the fragments in the LISE3 separator \cite{Anne1992276}, they were implanted into a 300-$\mu$m-thick Double-Sided Silicon Strip Detector (DSSSD). The DSSSD was used to detect both the implanted heavy ions and subsequent charged-particle ($\beta$ particles and protons) decays. Four EXOGAM Germanium clovers \cite{exogam} surrounding the DSSSD were used to detect the $\beta$-delayed $\gamma$-rays.
The ions were identified by combining the energy loss signal generated in a silicon $\Delta E$ detector located 28 cm upstream from the DSSSD and the Time-of-Flight (ToF), defined as the time difference between the cyclotron radio-frequency and the $\Delta E$ signal (see Figs. \ref{ID1plot} and \ref{ID2plot}). An implantation event was defined by simultaneous signals from both the $\Delta E$ detector and the DSSSD. A decay event was defined by a signal above threshold in the DSSSD and no coincident $\Delta E$ signal.
The experimental setup is described in detail in Ref. \cite{PhysRevC.93.044336}, as well as the data analysis procedures employed.
\section{\label{results}The $^{52}$Co, 2$^+$ isomer}
In order to study the decay of the 2$^+$ isomer in $^{52}$Co, we have selected the events where $^{52}$Ni was implanted (see Figs. \ref{ID1plot} and \ref{ID2plot}). For the following discussion we refer to the partial decay scheme shown in Fig. \ref{decay}(a), which starts from the $\beta^+$ decay of $^{52}$Ni to $^{52}$Co and then to $^{52}$Fe. The $\beta$ decay of $^{52}$Ni \cite{PhysRevC.93.044336} populates the 0$^+ (T = 2)$ IAS in $^{52}$Co at 2926(50) keV with a $\beta$ feeding of 56(10)\%, consistent with the expected Fermi strength \mbox{$B$(F) = $|N-Z|$ = 4}. Thereafter the decay of the IAS proceeds 25(5)\% of the time by proton emission to $^{51}$Fe and 75(23)\% of the time by a $\gamma$-ray cascade. The cascade consists of $\gamma$-rays of 2407 and 141 keV energy, with intensities $I_{\gamma}$ of 42(10)\% and 43(8)\%, respectively, and populating in sequence the levels at 519(50) \mbox{($J^{\pi}$ = 1$^+$)} and 378(50) \mbox{($J^{\pi}$ = 2$^+$)} keV in $^{52}$Co. As explained in detail in Ref. \cite{PhysRevC.93.044336}, we have assumed for the last level an energy of 378(50) keV from the value in the mirror nucleus $^{52}$Mn, 377.749(5) keV \cite{PhysRevC.7.677,NDS2015}, fixing in this way the excitation energies for the $^{52}$Co levels. The error of 50 keV on the 378 keV level energy, which accounts for possible mirror energy differences (MED), was estimated in Ref. \cite{PhysRevC.93.044336} by looking at the energies of the levels up to 400 keV in mirror nuclei with $T_z = +1/2, \text{-}1/2, +1, \text{-}1$. MED data for 2$^+$ states in the $A$ = 42-54 region \cite{Bentley2015} shows that our uncertainty is realistic and conservative. No $\gamma$-ray was observed from the 378 keV level, which is expected to be an isomeric state. This is not surprising since the $^{52}$Co, 2$^+$ level can decay by a Fermi $\beta$ transition, in contrast with its homologous state in the mirror nucleus. In $^{52}$Mn the Fermi transition is not possible and this, together with the smaller $Q_\beta$-value, makes the $\beta$ decay much slower and the slow E4 transition can compete with it.
\begin{figure*}[!ht]
\centering
\includegraphics[height=0.89\textheight,trim={0 0.35cm 0 0.4cm},clip]{decay_scheme_complete.pdf}
\vspace{-3.0 mm}
\caption{(a) Partial decay scheme of $^{52}$Ni, including the decay of the $^{52}$Co, 2$^+$ isomer. Other 1$^+$ states populated in this decay and de-exciting via proton decay (see Ref. \cite{PhysRevC.93.044336}) are not included in the figure. The proton branching of the $^{52}$Co IAS is 25(5)\%. The energy of the isomeric level in $^{52}$Co, 378(50) keV, is assumed from the mirror $^{52}$Mn \cite{PhysRevC.7.677}. Two 2$^+$ levels separated by 10 keV are reported as IAS candidates in $^{52}$Fe \cite{DECOWSKI1978186}. The dashed $\gamma$-ray was reported in the literature \cite{JPSJ43} but not seen in the present work. The quoted $I_\beta$ branchings refer to 100 decays from $^{52m}$Co (2$^+$) estimated using the intensity of the 141 keV $\gamma$ line. (b) Decay scheme of the $^{52}$Co, 6$^+$ ground state deduced from the results of the present experiment. The quoted $I_\beta$ branchings refer to 100 decays from $^{52}$Co (6$^+$).}
\label{decay}
\end{figure*}
\begin{figure}[!ht]
\begin{minipage}{1.0\linewidth}
\centering
\vspace{-5.0 mm}
\includegraphics[width=1\columnwidth]{Gammas_all.pdf}
\end{minipage}
\begin{minipage}{1.0\linewidth}
\includegraphics[width=1\columnwidth]{52Fe_849.pdf}
\end{minipage}
\begin{minipage}{1.0\linewidth}
\includegraphics[width=1\columnwidth]{52Fe_1910.pdf}
\end{minipage}
\begin{minipage}{1.0\linewidth}
\includegraphics[width=1\columnwidth]{52Fe_5185.pdf}
\end{minipage}
\caption{(a) $\gamma$-ray spectrum observed for the decay of $^{52}$Ni. (b) Zoom of the 849 keV $\gamma$ line. (c) Zoom of the 1910 keV $\gamma$ line. (d) Zoom of the 5185 keV $\gamma$ line, detected using the low-amplification electronic chain (see text). The energy given for the peak includes the calibration made with the $\gamma$ lines from the decay of $^{52}$Co ($6^+$).}
\label{gammas}
\end{figure}
\begin{figure}[!htb]
\centering
\vspace{-5.0 mm}
\includegraphics[width=1\columnwidth]{halflife_849_errors.pdf}
\caption{Fit of the correlation-time spectrum gated on the 849 keV $\gamma$ line, giving a $T_{1/2}$ = 102(6) ms for the $^{52}$Co, 2$^+$ isomer.}
\label{halflife}
\end{figure}
To detect the population of the $^{52}$Co, 6$^+$ ground state would require the observation of four $\gamma$-rays in cascade (1329, 1942, 1535 and 849 keV) de-exciting the 6$^+$ IAS in $^{52}$Fe \cite{Hagberg1997183}, see Fig. \ref{decay}(b). In contrast the $\beta$ decay of the expected $^{52}$Co, 2$^+$ isomer should proceed to its IAS with 2 units of Fermi strength, and to a series of 1$^+$, 2$^+$ and 3$^+$ levels via GT transitions. Since there is no known 1$^+$ or 3$^+$ level below 6 MeV excitation energy in $^{52}$Fe, one can expect to observe the feeding to the IAS and some other 2$^+$ states. As discussed in Ref. \cite{Hagberg1997183}, the most intense $\gamma$-ray should be the 849 keV line \mbox{($2^+ \rightarrow 0^+$)}, which is also emitted in the cascade de-exciting the $6^+$ IAS in $^{52}$Fe. The specific signature of $^{52m}$Co (2$^+$) should be the strong population of the IAS, with the observation of its de-exciting $\gamma$-rays. In addition, in Ref. \cite{Hagberg1997183} it is proposed that the observation of a weak $\gamma$-ray at 1910 keV, belonging to a \mbox{$2^+ \rightarrow 2^+$} transition between the 2759 and 849 keV levels, should also be a typical feature of the population of $^{52m}$Co.
Ref. \cite{DECOWSKI1978186} reports two 2$^+$ levels in $^{52}$Fe separated in energy by 10 keV only, at 6034(5) and 6044(5) keV, both candidates to be the IAS of the $^{52}$Co, 2$^+$ isomeric state. If the existence of these 2$^+$ levels could be confirmed, they may provide another example of isospin mixing in the IAS. The mixing would be strong because of the very small energy separation. A similar situation has been observed, e.g., in Ref. \cite{PhysRevLett.112.222501} where the energy separation was of the order of 100 keV.
The $\gamma$-ray spectrum observed for the decay chain of $^{52}$Ni is shown in Fig. \ref{gammas}(a). In addition to the 141 keV $\gamma$-ray mentioned above (for the 2407 keV $\gamma$-ray see below) and the 511 keV $\gamma$ line associated with the annihilation of the positrons emitted in the $\beta$ decay, two other lines are observed at 849 and 1910 keV [they are also shown in Figs. \ref{gammas}(b) and (c), respectively]. The $\gamma$-ray seen at 1910 keV corresponds to a \mbox{$2^+ \rightarrow 2^+$} transition between the 2759 and 849 keV known levels \cite{NDS2015} in $^{52}$Fe [Fig. \ref{decay}(a)] and it is expected to be seen in the decay of the $^{52}$Co, 2$^+$ isomer \cite{Hagberg1997183}. The 1910 keV $\gamma$-ray, indeed, cannot be observed in the decay of the $^{52}$Co, 6$^+$ ground state because it does not populate the 2$^+$ state at 2759 keV. Moreover, the population of the 2759 keV state starting from the $^{52}$Fe, 4$^+$ level at 3584 keV would require the observation of a $\gamma$-ray of 825 keV, which we do not see [Fig. \ref{gammas}(b)].
The $\gamma$-ray spectrum shown in Fig. \ref{gammas} (a, b, c) was obtained using the high-amplification electronic chain \cite{PhysRevC.93.044336} and it allows the study of $\gamma$-rays up to 2 MeV. $\gamma$-rays of higher energy were detected using a low-amplification electronic chain, where a problem was observed during the data analysis consisting in a distortion of the peaks (see Ref. \cite{PhysRevC.93.044336} for details). In the $\gamma$-ray spectrum obtained with the low-amplification chain, in addition to the 2407 keV $\gamma$-ray in $^{52}$Co (good agreement was found in both energy and intensity when compared with values in the literature \cite{PhysRevC.93.044336,Dossat200718}), a $\gamma$-ray was observed at around 5 MeV [Fig. \ref{gammas}(d)]. The energy calibration at high energy was performed using the $\gamma$ lines observed in the decay of $^{52}$Co, $6^+$ (Section \ref{gs}). This procedure gave an energy of 5185(10) keV for the above $\gamma$-ray, which was then attributed to the \mbox{$2^+ \rightarrow 2^+$} transition between the $2^+$ IAS in $^{52}$Fe (at 6034(5) and/or 6044(5) keV \cite{DECOWSKI1978186,NDS2015}, where having one or both states does not change our conclusions) and the 849 keV level [Fig. \ref{decay}(a)]. A further confirmation comes from the fact that the 5185(10) keV $\gamma$-ray was also observed [Fig. \ref{gammasgs}(c) in Section \ref{gs}] when selecting events where $^{52}$Co was implanted, where one expects an admixture of both ground and isomeric states.
Therefore the observed 5185 keV $\gamma$-ray establishes clear evidence of the 2$^+$ isomer in $^{52}$Co, which is supported by the observed 1910 keV $\gamma$ line. Together, they constitute the first experimental evidence of the decay of $^{52m}$Co (2$^+$).
\begin{table}[!t]
\vspace{-5.0 mm}
\caption{$\gamma$-ray energies $E_{\gamma}$, $\gamma$ intensities $I_{\gamma}$ relative to $^{52}$Ni implants, and $\gamma$ intensities normalized to 100 decays from $^{52m}$Co (2$^+$) (using the intensity of the 141 keV $\gamma$-ray).}
\label{table}
\centering
\begin{ruledtabular}
\begin{tabular}{r r r}
$E_{\gamma}$ (keV) & $I_{\gamma}$ (\%) ($^{52}$Ni) & $I_{\gamma}/100$ decays (\%) ($^{52m}$Co) \\ \hline
849(1) & 42(8) & 97(26) \\
1910(1) & 5(1) & 12(3) \\
5185(10) & 17(4) & 39(12) \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}[!t]
\caption{Results on the $\beta^{+}$ decay of $^{52m}$Co (2$^+$) to $^{52}$Fe. Excitation energies $E_X$ in $^{52}$Fe, $\beta$ feedings $I_{\beta}$, Fermi $B$(F) and Gamow-Teller $B$(GT) transition strengths to the $^{52}$Fe levels.}
\label{table2}
\centering
\begin{ruledtabular}
\begin{tabular}{c r r r}
$E_X$ (keV) & $I_\beta$ (\%) & $B$(F) & $B$(GT) \\ \hline
849(1) & 46(28) & & 0.06(4) \\
2759(2) & 12(3) & & 0.05(1) \\
6034(5)$^a$ - 6044(5)$^a$ & 39(12) & 1.6(5)$^b$ & \\
\end{tabular}
\end{ruledtabular}
\raggedright{$^a$ IAS, $E_X$ from Refs. \cite{DECOWSKI1978186,NDS2015}. \\ $^b$ Calculated assuming all the strength is Fermi.}
\end{table}
Besides the $\gamma$-rays described above, a $\gamma$-ray of 2760(1) keV was seen in Ref. \cite{JPSJ43} and attributed to a \mbox{$2^+ \rightarrow 0^+$} transition between the 2759 keV level and the ground state in $^{52}$Fe. Considering the intensity measured in Ref. \cite{JPSJ43} for this $\gamma$-ray and our low $\gamma$-efficiency at that energy, we do not expect to see this $\gamma$ line in our low-amplification spectrum, and indeed we do not observe it.
The half-life associated with a given $\gamma$ line is determined from the fit of the correlation-time spectrum gated on that $\gamma$ line, which was created according to the procedure described in Ref. \cite{PhysRevC.93.044336}. The fit performed for the 849 keV $\gamma$ line is shown in Fig. \ref{halflife}. The fit function includes the decay of the parent nucleus ($^{52}$Ni, with a known half-life of 42.8(3) ms \cite{PhysRevC.93.044336}, which was kept fixed), the growth of the daughter activity ($^{52m}$Co, of unknown half-life) and a constant background. From this fit we obtained a half-life of 102(6) ms for $^{52m}$Co (2$^+$). The much lower statistics prevented us from extracting the half-lives associated with the 1910 and 5185 keV $\gamma$-rays when selecting $^{52}$Ni implants. However, by selecting events where $^{52}$Co was implanted we were able to extract a half-life of 91(15) ms for the 5185 keV $\gamma$-ray, in agreement with the value quoted above.
The results are summarized in the decay scheme shown in Fig. \ref{decay}(a) and in Tables \ref{table} and \ref{table2}. The value \mbox{$Q_{\beta}$ = 11571(54) keV}, given in Fig. \ref{decay}(a) for the decay of $^{52}$Ni, was determined as explained in Section V of Ref. \cite{PhysRevC.93.044336}, where we deduced the ground state mass excesses for $^{52}$Ni and $^{52}$Co. Adding to that information the measured mass excess for $^{52}$Fe, -48332(7) keV \cite{Audi2012}, we can determine a value \mbox{$Q_{\beta}$ = 13845(52) keV} for the decay of the $^{52}$Co ground state, given in Fig. \ref{decay}(b).
Table \ref{table} gives the energies $E_{\gamma}$ and intensities $I_{\gamma}$ of the observed $\gamma$ peaks (both relative to $^{52}$Ni implants and normalized to 100 decays from $^{52m}$Co). The $\gamma$ intensities relative to $^{52}$Ni implants are determined as in Ref. \cite{PhysRevC.93.044336}. The $\gamma$-efficiency calibration, shown in Fig. 8 of Ref. \cite{PhysRevC.93.044336}, has been extended to higher $\gamma$ energies by Monte Carlo simulations. Since the $^{52}$Co $\gamma$-ray at 141 keV only populates $^{52m}$Co (2$^+$) and its intensity is 43(8)\%, the $\gamma$ intensities can be normalized to 100 decays from $^{52m}$Co using the intensity of the 141 keV $\gamma$-ray.
Table \ref{table2} gives the $\beta$ feedings $I_{\beta}$ and the Fermi and Gamow-Teller transition strengths for the $\beta$-decay of the $^{52}$Co, 2$^+$ isomer to $^{52}$Fe. The $\beta$ feedings to the levels populated in $^{52}$Fe are deduced from the $\gamma$ intensities for 100 decays from $^{52m}$Co (2$^+$). As discussed above, there are two possible candidate levels in $^{52}$Fe (at 6034(5) and 6044(5) keV \cite{DECOWSKI1978186,NDS2015}) for the IAS of $^{52m}$Co (2$^+$), which are expected to be strongly mixed. Unfortunately the resolution of our low-amplification $\gamma$ spectrum did not allow the disentanglement of the two contributions based on the 5185 keV peak [Fig. \ref{gammas}(d)]. Moreover, in the population of the IAS, both Fermi and Gamow-Teller contributions are possible. Thus we have calculated an upper limit to $B$(F) assuming all the strength is due to the Fermi transition and taking an average excitation energy of 6039(7) keV. A maximum $B$(F) of 1.6(5) is obtained, in agreement with the expected value $|N-Z|$ = 2. This confirms that the intensity extracted for the 5185 keV $\gamma$-ray is meaningful, even with the peak distortion.
\section{\label{gs}Half-life of the $^{52}$Co, 6$^+$ ground state}
In the same experiment $^{52}$Co fragments were also produced directly. This enabled us to add new information on the $\beta$ decay of the $^{52}$Co 6$^+$ ground state and measure its half-life with improved precision. In order to study the $\beta$ decay of the ground state, we have selected the events where $^{52}$Co was implanted (see Figs. \ref{ID1plot} and \ref{ID2plot}). They should be a mixture of the ground and isomeric states. The high-amplification $\gamma$-ray spectrum obtained for the decay of $^{52}$Co is shown in Fig. \ref{gammasgs}(a). There, we observed known $\gamma$-rays (at 849, 1288, 1329, 1535, 1556 and 1942 keV) \cite{NDS2015}, expected from the decay of the levels populated in $^{52}$Fe, and a further $\gamma$-ray at 782 keV. In the low-amplification spectrum [Fig. \ref{gammasgs}(b)] we saw in addition other expected $\gamma$-rays from $^{52}$Fe (at 2488, 2735 and 2755 keV). In the latter spectrum we also observed the 5185(10) keV $\gamma$-ray from the decay of $^{52m}$Co (2$^+$) [Fig. \ref{gammasgs}(c)].
\begin{figure}[!b]
\begin{minipage}{1.0\linewidth}
\centering
\includegraphics[width=1\columnwidth]{52Co_gammas_a.pdf}
\end{minipage}
\begin{minipage}{1.0\linewidth}
\includegraphics[width=1\columnwidth]{52Co_gammas_b.pdf}
\end{minipage}
\begin{minipage}{1.0\linewidth}
\includegraphics[width=1\columnwidth]{52Co_gammas_c.pdf}
\end{minipage}
\caption{(a) $\gamma$-ray spectrum observed for the decay of $^{52}$Co with the high-amplification electronic chain. (b) and (c) Zoom of the low-amplification spectrum in the regions of interest.}
\label{gammasgs}
\end{figure}
The $\beta$ decay of the $^{52}$Co, 6$^+$ ground state is summarized in Fig. \ref{decay}(b), where $\beta$ feeding is expected to the 5$^+$, 6$^+$ and 7$^+$ levels in $^{52}$Fe. The 6$^+$ IAS at 5656 keV \cite{NDS2015} in $^{52}$Fe de-excites by $\gamma$-ray cascades starting with the 782 and 1329 keV $\gamma$-rays. A possible 516 keV $\gamma$-ray connecting the IAS and the 5140 keV level would be hidden below the 511 keV annihilation peak.
\begin{figure}[!th]
\centering
\includegraphics[width=1\columnwidth]{52Cogs_halflife_errors.pdf}
\caption{Fit of the correlation-time spectrum obtained as a sum of the spectra gated on the 1329, 1535 and 1942 keV $\gamma$-rays from the decay of the $^{52}$Co, 6$^+$ ground state, giving a $T_{1/2}$ = 112(3) ms.}
\label{halflifegs}
\end{figure}
We have observed two $\gamma$-rays at 1556 and 2755 keV, corresponding to the de-excitation of the level at 5140(5) keV. A level at 5134(8) keV was observed in Ref. \cite{DECOWSKI1978186}, where a \mbox{$J^{\pi}$ = 5$^-$} was attributed to it. In Ref. \cite{JPSJ43} a level was seen at 5138(4) keV, de-exciting by two $\gamma$-rays of 2380 and 4286 keV which we do not observe. Finally, in Ref. \cite{PhysRevC.58.3163} a level was observed at 5137 keV, de-exciting by three $\gamma$-rays of 740.6, 1553 and 2753 keV. The last two $\gamma$-ray energies agree marginally with our observed $\gamma$-rays at 1556 and 2755 keV, however we did not observe the 740.6 keV $\gamma$-ray, which is supposed to be stronger than the 1553 keV $\gamma$ line according to Ref. \cite{PhysRevC.58.3163}. We believe we see the same level as in Ref. \cite{PhysRevC.58.3163}, and not the level reported in Ref. \cite{JPSJ43}. We also do not know if the level we observed corresponds to that in Ref. \cite{DECOWSKI1978186}, consequently we have put the 5$^-$ assignment in parenthesis in Fig. \ref{decay}(b).
Table \ref{table3} gives the energies $E_{\gamma}$ of the observed $\gamma$ peaks (first column) and their intensities $I_{\gamma}$ normalized to the 849 keV $\gamma$-ray (second column). We could not extract the intensities for the $\gamma$-rays above 2 MeV, only observed in the low-amplification spectrum affected by the peak distortion. We obtained the intensity of the 2735 keV $\gamma$-ray by summing those of the 1288 and 1556 keV $\gamma$-rays, i.e., we assumed that the 3584 keV level is not directly populated in either the $\beta$ decay of the 6$^+$ ground state or the $\beta$ decay of the 2$^+$ isomeric level.
Looking at the intensities normalized to that of the 849 keV $\gamma$-ray, we get 24(6)\% and 27(7)\% for the 1329 and 1942 keV $\gamma$-rays, respectively. Within the errors, a small amount of $\beta$ feeding to the 4327 keV level is possible. In Ref. \cite{Hagberg1997183} the intensity of the 1942 keV $\gamma$-ray was reported to be 17\% lower than that of the 1329 keV $\gamma$-ray. This was probably because the 1942 keV peak was not resolved from a 1944 keV peak from the decay of $^{50m}$Mn.
The summed intensities of the 1535 and 2735 keV $\gamma$-rays, which both go to the 849 keV level, are 83(31)\%. This means that the 2$^+$ level at 849 keV may have an extra feeding of 17(31)\% that may be attributed to the $\beta$ decay of the $^{52}$Co (2$^+$) isomeric state. With this information one can normalize the $\gamma$ intensities to 100 decays from the $^{52}$Co (6$^+$) ground state; these values are presented in the third column of Table \ref{table3}. They are also used to calculate the $\beta$ feedings shown in the second column of Table \ref{table4}. We expect that the levels at 4872 and 5140 keV get some direct feeding, which we cannot estimate because we miss the intensities of the 2488 and 2755 keV $\gamma$-rays. Thus we attribute the missing $\beta$ feeding (50\%) to these 4872 and 5140 keV levels. Besides the $\beta$ feedings, Table \ref{table4} gives $B$(F) and $B$(GT). Also in this case both Fermi and Gamow-Teller contributions are possible in the population of the IAS, thus we calculated an upper limit to $B$(F). We obtained a maximum $B$(F) of 1.7(3), which agrees with the expected value $|N-Z|$ = 2.
\begin{table}[!t]
\vspace{-5.0 mm}
\caption{Column one shows the $\gamma$-ray energies $E_{\gamma}$. Column two gives the $\gamma$ intensities $I_{\gamma}$ relative to the 849 keV $\gamma$-ray, including both the 6$^+$ and the 2$^+$ $^{52}$Co decays. Column three presents the $\gamma$ intensities normalized to 100 decays from the $^{52}$Co (6$^+$) ground state.}
\label{table3}
\centering
\begin{ruledtabular}
\begin{tabular}{r r r}
$E_{\gamma}$ (keV) & $I_{\gamma}$/$I_{\gamma}$(849) (\%) & $I_{\gamma}/100$ decays (\%) ($^{52}$Co$_{gs}$) \\ \hline
782(1) & 15(4) & 18(5) \\
849(1) & 100(26) & 100(21) \\
1288(1) & 8(2) & 10(3) \\
1329(1) & 24(6) & 29(8) \\
1535(1) & 67(17) & 81(21) \\
1556(1) & 7(2) & 9(2) \\
1942(1) & 27(7) & 32(8) \\
2488(5) & & \\
2735(5) & 16(3) & 19(3) \\
2755(5) & & \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}[!t]
\caption{Results on the $\beta^{+}$ decay of the $^{52}$Co (6$^+$) ground state to $^{52}$Fe. Excitation energies $E_X$ in $^{52}$Fe, $\beta$ feedings $I_{\beta}$, Fermi $B$(F) and Gamow-Teller $B$(GT) transition strengths to the $^{52}$Fe levels.}
\label{table4}
\centering
\begin{ruledtabular}
\begin{tabular}{c r r r}
$E_X$ (keV) & $I_\beta$ (\%) & $B$(F) & $B$(GT) \\ \hline
849(1) & & & \\
2385(1) & & & \\
3584(5) & & & \\
4327 (2) & 3(11) & & 0.03(12) \\
4872(5) & \rdelim\}{2}{8mm}[$\leq$50] & & \\
5140(5) & & & \\
5656(2)$^a$ & 47(9) & 1.7(3)$^b$ & \\
\end{tabular}
\end{ruledtabular}
\raggedright{$^a$ IAS. \\ $^b$ Calculated assuming all the strength is Fermi.}
\end{table}
As mentioned above, in fragmentation experiments both the 6$^+$ ground state and the 2$^+$ isomer will be implanted together and cannot be separated with the available information on the implants. This has to be taken into account in the determination of the half-life of the $^{52}$Co ground state. In Ref. \cite{Hagberg1997183}, indeed, because of the ambiguity of the origin of the 849 keV $\gamma$-rays their apparent half-life [104(11) ms] was not used to determine the half-life of the $^{52}$Co ground state [$T_{1/2}$ = 115(23) ms]. More recently, the $\beta$ decay of $^{52}$Co was revisited in Ref. \cite{PhysRevC.87.024312} and a value of 103(7) ms was extracted for the half-life of the ground state by gating on the $\beta$ particles. Combining this with the previous measurement \cite{Hagberg1997183} gives a weighted average value of 104(7) ms which is the value reported in the most recent compilation for mass $A$ = 52 \cite{NDS2015}. However, in Ref. \cite{PhysRevC.87.024312} the possible implantation of the $^{52}$Co isomer together with the $^{52}$Co ground state was not considered.
To determine the half-life of the $^{52}$Co, 6$^+$ ground state in a isomer-free way, we have constructed a correlation-time spectrum as the sum of the spectra gated on the 1329, 1535 and 1942 keV $\gamma$-rays. The fit to this spectrum, shown in Fig. \ref{halflifegs}, gives $T_{1/2}$ = 112(3) ms. This result agrees with the value from Ref. \cite{Hagberg1997183} but the precision is improved.
\section{\label{concl}Conclusions}
We reported the first experimental observation of the decay of the 2$^+$ isomeric state in $^{52}$Co, which was produced in the $\beta$ decay of $^{52}$Ni. We observed the decay of $^{52m}$Co to $^{52}$Fe, where it populates various 2$^+$ states including the IAS. These 2$^+$ levels then de-excite by $\gamma$-ray emission and we observed three $\gamma$-rays at 849, 1910 and 5185 keV. The observed de-excitation of the IAS (by the 5185 keV $\gamma$-ray) is clear evidence of the population of the 2$^+$ isomer, which is reinforced by the observation of the expected \cite{Hagberg1997183} $\gamma$-ray at 1910 keV. The $\beta$ feedings for the decay of the $^{52}$Co isomer to the 2$^+$ levels in $^{52}$Fe and the Fermi and Gamow-Teller transition strengths have been determined. We have extracted a half-life of 102(6) ms for the $^{52}$Co, 2$^+$ isomer using the 849 keV $\gamma$ line.
We have also studied the $\beta$ decay of the $^{52}$Co, 6$^+$ ground state by gating on the events where $^{52}$Co was implanted, obtaining new information. Many $\gamma$-rays were observed, including a previously unobserved $\gamma$-ray at 782 keV, and their intensities were determined. The $\beta$ feedings for the decay of $^{52}$Co (6$^+$) to the 6$^+$ levels in $^{52}$Fe and the $B$(F) and $B$(GT) were deduced. A half-life of 112(3) ms was measured for the $^{52}$Co (6$^+$) ground state, improving the uncertainty in comparison with the values reported in the literature.
The $^{52}$Co nucleus lies in the $rp$-process pathway, where the proton-absorption and $\beta$-decay processes compete. Hence the existence of a $\beta$-decaying isomer as well as its decay properties are important.
\begin{acknowledgments}
This work was supported by the Spanish MICINN grants FPA2008-06419-C02-01, FPA2011-24553, FPA2014-52823-C2-1-P; Centro de Excelencia Severo Ochoa del IFIC SEV-2014-0398; $Junta~para~la~Ampliaci\acute{o}n~de~Estudios$ Programme (CSIC JAE-Doc contract) co-financed by FSE; ENSAR project 262010; MEXT, Japan 18540270 and 22540310; Japan-Spain coll. program of JSPS and CSIC; UK Science and Technology Facilities Council (STFC) Grant No. ST/F012012/1; Region of Aquitaine. E.G. acknowledges support by TUBITAK 2219 International Post Doctoral Research Fellowship Programme. R.B.C. acknowledges support by the Alexander von Humboldt foundation and the Max-Planck-Partner Group. We acknowledge the EXOGAM collaboration for the use of their clover detectors.
\end{acknowledgments}
|
\section{Introduction}\label{s.introduction}
In queueing resources that are shared by multiple traffic streams, smooth streams potentially experience poor performance
when they are mixed with less regular streams. Indeed, under a first-come-first-serve (FCFS) discipline, users that correspond to a
highly variable input process may negatively affect the quality-of-service of other users. This motivates the attention paid
to more sophisticated queueing disciplines, in which firm (per-user) performance guarantees can be given. One such a policy is the
{\it generalized processor sharing} (GPS) discipline. In GPS all users classes are guaranteed a certain service rate, whereas
the residual capacity is distributed according to a given allocation rule. The earliest (packet-based) implementations of GPS,
usually referred to as {\it weighted fair queueing} (WFQ), date
back to the late 1980s \cite{DEM}.
In many real-life systems, input streams may exhibit rather extreme types of irregularities.
For instance in the domain of communication networks, measurement studies show that
traffic patterns are typically heavy-tailed, in that there is a relatively high likelihood of an extremely large
amount of traffic being generated over a short time interval. Under FCFS all streams
would perceive roughly the same performance, which is essentially determined by the input
class with the heaviest tail. GPS can be considered as a viable way to remedy this complication,
by offering each class a guaranteed service rate.
\vspace{2mm}
In this paper we consider a two-class GPS system, in which the inputs are
L\'evy processes with heavy-tailed
marginals; in our context, `heavy-tailed' refers to the complementary distribution function having a regularly varying tail.
This class of L\'evy processes covers many practically relevant processes.
In the first place, it contains the class of compound Poisson processes, in which independent
and identically distributed (i.i.d.) regularly-varying jobs arrive according to a Poisson process. In the second place, it covers the
class of $\alpha$-stable L\'evy motion; this class is particularly relevant, as it appears as the limiting process
for random walk models with increments that have infinite variance \cite{TaS94,WHI}.
Our main findings are the exact asymptotics of the tail distributions of both queues.
More specially, with $Q_i$ denoting the stationary workload of the $i$-th queue, we
find explicit functions $f_i(\cdot)$ such that ${\mathbb P}(Q_i>u)/f_i(u)
\to 1$ as $u\to\infty$; we write ${\mathbb P}(Q_i>u)\sim f_i(u)$.
As it turns out, depending on the interplay between the heaviness of both inputs'
tail distributions and
the stability of the queues while working in isolation,
one can distinguish four scenarios, each of them leading to qualitatively
different asymptotics. The resulting asymptotics lend themselves to an
intuitive explanation, in that they reveal the most likely way that
the workload under consideration exceeds $u$, for $u$ large. The proofs rely on
combining bounds that were derived earlier for related queueing systems, as well as
a set of newly derived inequalities. Related results for settings that are special cases of ours
can be found in e.g.\ \cite{BBJ,Lelarge}, whereas in \cite{MvU} the focus is on GPS systems
with Gaussian inputs.
The paper is organized as follows. Notation, assumptions and preliminaries are
presented in Section 2. Then Section 3 states the main results, in terms of the
exact asymptotics for all four scenarios. These results are used in Section 4 to
give the corresponding expressions for the compound Poisson and $\alpha$-stable
cases. All proofs are given in Section~5.
\section{Notation and model description}\label{s.model}
In this paper we consider a queueing system that consists of two queues and one server.
Each queue, which has infinite storage capacity, is fed by an own traffic class;
the corresponding input processes are assumed to be mutually independent.
The total service rate of the server is $c>0$. Class $i$ is assigned
a guaranteed service rate $\phi_i c>0$ (or `weight'), where $\phi_1 c + \phi_2 c = c$.
This effectively means that if both classes are backlogged, then class $i$ is served
at rate $\phi_i c$, for $i=1,2$. If class $i$ has no backlog, then the other
class obtains the excess service rate.
Throughout this paper, we intensively use the concept of {\it cumulative input processes.}
We define by $Z_i(s,t)$ the cumulative
input to queue $i$ in interval $(s,t]$, for $i=1,2$ and $s<t$.
We assume that
\[
Z_i(s,t)=Z_i(t)-Z_i(s), \:\:\: i=1,2,
\]
where $\{Z_1(t):t\in {\bb R} \}$ and $\{Z_2(t):t\in {\bb R} \}$ are mutually independent
L\'evy processes.
As pointed out in the introduction, we specially consider the situation in which the L\'evy input processes are heavy-tailed. In
more concrete terms, this means that in the sequel we impose the following assumptions:
\vspace{2mm}
{\bf A1}~~${\bb P}(Z_{i}(1) > u) \sim u^{-\alpha_i}L_i(u)$, with
$\alpha_i>1$ and $L_i(\cdot)$ slowly varying at $\infty$, for $i=1,2$;
{\bf A2}~~${\bb E}[Z_{i}(1)] = \mu_{i}$, with $\mu_1+\mu_2=\mu<c.$
\vspace{2mm}
We let
$\{Q_i(t):t\ge0\}$ denote the stationary buffer content processes
for class $i$, for $i= 1,2$.
Observe that
condition {\bf A2} guarantees stability of the system, implying existence of the
stationary buffer content processes. To shorten the notation we
throughout write
\[Q_i\stackrel{\rm d}{:=} Q_i(0),\:\:\:i =1,2.\]
Notice that the system's stability does not rule out that one of the queues `is in overload'
(if it would operate in isolation, that is):
one could have that $\mu_i>\phi_i c$ for one of the queues.
We denote by $B_i(s,t)$, for $i=1,2$,
the amount of service obtained by the $i${-th}
class in time interval $(s,t]$. Then there is the obvious identity\begin{eqnarray}\label{e2}
Q_i(t) = Q_i(s)+Z_i(s,t)-B_i(s,t), \quad \forall s<t.
\end{eqnarray}
According to Reich's formula \cite{REICH} (see also \cite{DvU} in the context of GPS queues)
we have the following distributional representation for the stationary workloads:
\[Q_i \stackrel{\rm d}{=} \sup_{t\geq 0} \{ Z_i(-t,0) - C_i(-t,0)\},\]
where $C_i(s,t)$ is the amount of the service available to class $i$ in the interval $(s,t]$.
The relation $C_i(s,t) \geq \phi_i c\,(t-s)$ holds for all $s<t$.
Additionally, it is convenient to introduce, for $\lambda_i>\mu_i$ and $\lambda>\mu$,
\[
Q_i^{\lambda_i}(t) :=\sup_{s\geq t}\{ Z_i(-s,t) - \lambda_i (t+s)\},\:\:\:\:
Q^{\lambda}(t):=\sup_{s\geq t} \{Z_1(-s,t)+Z_2(-s,t) - \lambda (t+s)\}.\]
Observe that
$Q_i^{\lambda_i} :=Q_i^{\lambda_i} (0)$ is distributed as the stationary buffer content of queue $i$
working in isolation, if it were served at rate $\lambda_i$ all the time. Likewise,
$Q^\lambda :=Q^{\lambda} (0)$ corresponds to the total stationary buffer content of the system, if it were
served at rate~$\lambda$.
Since the queues interact
symmetrically, we focus on
just
${\bb P}(Q_1>u),$ for $u\to\infty.$
\section{Main results}\label{s.main}
In this section we present the main results of the paper.
We distinguish four scenarios, that differ in terms of (i)~the heaviness of the individual
input processes, and (ii)~the individual queues being underloaded or overloaded.
The proofs of all the results presented in this section are relegated to Section \ref{s.proofs}.
\subsection{Second queue in overload}\label{s.case1}
We first consider the scenario
that the second queue is unstable when working in isolation: $\mu_2>\phi_2 c$.
In this case, if the input process of the second queue
generates traffic at its mean rate (which does not correspond to a rare event), then
it will be using its full guaranteed service rate. This pattern would
leave the first queue as if working in isolation.
Based on this observation, we expect that
\[{\bb P}(Q_{1} > u) \sim {\bb P}(Q_{1}^{\phi_1c}>u).\]
The following theorem formalizes this heuristic. Notice that in this scenario we
necessarily have $\mu_1<\phi_1 c.$
\begin{Theorem}\label{th.secUnstab}
Suppose that $Z_1,Z_2$ satisfy {\bf A1-A2}.
If $\mu_2>\phi_2 c$, then, as $u\to\infty$,
$${\bb P}(Q_{1} > u) \sim \frac{1}{(\phi_1 c - \mu_1)(\alpha_1-1)}u^{1-\alpha_1}L_1(u).$$
\end{Theorem}
\subsection{Second queue in underload, first class is heavier}
\label{s.case2}
In the other three scenarios the second queue is stable while working in isolation, i.e.,
we consider the situation that
$\mu_2<\phi_2 c$.
As it turns out, under this condition the interplay between both input processes
plays a key role. We first concentrate on the case that
the first class is heavier than the second one, i.e., $\alpha_1<\alpha_2.$
Since the second queue is stable while working in isolation
and `is lighter' than the first one, the most likely way to generate a large workload in the
first queue does not involve a large buffer
content in the second queue. The most probable way the first buffer
reaches a large level corresponds to (i)~the second class generating traffic
at its mean level $\mu_2$, and (ii)
the remaining service capacity $c-\mu_2$ being allocated to the first queue.
Hence the so-called \textit{reduced-load}
equivalence holds in this case, cf.\ e.g.\ \cite{BBJ2}:
\[{\bb P}(Q_{1} > u) \sim {\bb P}(Q_{1}^{c-\mu_2}>u).\]
This leads to the following theorem.
\begin{Theorem}\label{th.secStabH}
Suppose that $Z_1,Z_2$ satisfy {\bf A1-A2}.
If $\mu_2<\phi_2 c$ and $\alpha_1<\alpha_2$, then, as $u\to\infty$,
$${\bb P}(Q_{1} > u) \sim \frac{1}{(c-\mu)(\alpha_1-1)}u^{1-\alpha_1}L_1(u).$$
\end{Theorem}
\subsection{Second queue in underload, second class is heavier}
\label{s.case3}
In the remaining two scenarios the second queue is stable while working in isolation, and
the second class is heavier than the first one, i.e., $\alpha_2<\alpha_1.$ Two cases still need to be distinguished:
the first queue being in underload or not.
In this third scenario we suppose that both the first and the second queue are stable while working in isolation, i.e., $\mu_i<\phi_i c$ for $i=1,2$
(and, as mentioned above, the second class is the heavier).
For this scenario it turns out that again the reduced load equivalence holds:
\[{\bb P}(Q_{1} > u) \sim {\bb P}(Q_{1}^{c-\mu_2}>u).\]
Intuitively, this means that the most probable way in which queue 1 grows large
is as follows: the second class generates traffic roughly at its mean rate, and
the first queue builds up as acting in isolation
with service rate $c-\mu_2$ (which can be interpreted as the service rate left by the second queue).
Although the asymptotics coincide with those obtained in Theorem \ref{th.secStabH},
the proof of the upper bound for this case needs an entirely different approach (which motivates why we treat
them as separate cases).
\begin{Theorem}\label{th.secStabL}
Suppose that $Z_1,Z_2$ satisfy {\bf A1-A2}.
If $\mu_1<\phi_1 c$, $\mu_2<\phi_2 c$ and $\alpha_2<\alpha_1$, then, as $u\to\infty$,
$${\bb P}(Q_{1} > u) \sim \frac{1}{(c-\mu)(\alpha_1-1)}u^{1-\alpha_1}L_1(u).$$
\end{Theorem}
\subsection{First queue in overload, second class is heavier}\label{s.case4}
Finally, we consider the scenario that
the first queue is in overload (i.e., unstable when working in isolation: $\mu_1>\phi_1 c$), and
the second class is the heavier (i.e., $\alpha_2<\alpha_1$).
We in addition assume that $Z_2$ be spectrally positive.
In this case
the most probable way in which the first queue reaches a high level is
such that the first class generates traffic roughly at its average rate $\mu_1$ (which does not correspond to a rare event).
Now the crucial issue concerns the fraction of its service rate that is left by the second class to the first class.
As it turns out, the most likely behavior of the second queue
can be linked with
the downstream queue of a fictitious two-node tandem queue
fed by $Z_2$ with service rate $\phi_2c$ at the upstream queue and
service rate $c-\mu_1$
at the downstream queue, in the sense that
\[{\bb P}(Q_{1} > u) \sim {\bb P}(V>u),\]
where
\[V \colonequals \sup_{t \geq 0}\{Z_2(-t,0) - (c-\mu_1)t\} - \sup_{s \geq 0} \{Z_2(-s,0) - \phi_2 cs\}.\]
This relation was also observed in
GPS models with fractional Brownian input in \cite{DMGPS}, whereas \cite{BBJ}
finds a similar relation for the case of heavy tailed on-off input.
Combining this with results from \cite{DMvUTand} on tandem queues with spectrally positive input,
we thus arrive at the following asymptotics.
\begin{Theorem}\label{th.reqc}
Suppose that $Z_1,Z_2$ satisfy {\bf A1-A2},
$\mu_1>\phi_1 c$, $\alpha_2<\alpha_1$ and $Z_2$ is spectrally positive
with $\alpha_2\notin {\bb N}$.
Then, as $u\to\infty$,
\[{\bb P}(Q_{1} > u) \sim \left( \frac{\mu_1-\phi_1c}{\phi_2c-\mu_2} \right)^{\alpha_2-1}\frac{1}{(c-\mu)(\alpha_2-1)}u^{1-\alpha_2}L_2(u).\]
\end{Theorem}
\begin{Rem}{\em
In the proof of
Theorem \ref{th.reqc} it plays a crucial role that $Z_2$ is assumed to be spectrally positive.
We strongly believe that this assumption is of a technical nature, in that the statement of
Theorem
\ref{th.reqc} is valid for general $Z_2$.
We anticipate, however, that a proof for general $Z_2$ would be considerably more complicated,
and would go along entirely different lines; see also Remark \ref{rem}.}
\end{Rem}
Observe that in the first three scenarios the workload of the first queue inherits the tail behavior of its input process: the complementary distribution function ${\mathbb P}(Q_1>u)$ essentially behaves as
$u^{1-\alpha_1}$. We conclude that in these cases the GPS mechanism succeeds in protecting the first stream against the second stream. Only in the last scenario, ${\mathbb P}(Q_1>u)$ becomes heavier, which issue to the relatively large weight allocated to the second stream.
\section{Special cases}\label{s.examples}
In this section we use the general results, as presented in the previous section,
to find the asymptotics for
GPS systems fed by compound Poisson processes with heavy-tailed
jumps (Section \ref{ss.comp}) and by $\alpha$-stable L\'evy input (Section \ref{ss.stable}).
\subsection{Compound Poisson input}\label{ss.comp}
This subsection concentrates on the case of {compound Poisson} inputs.
More concretely, we assume that $Z_i(t)$ is of the form
\[Z_i(t)=\sum_{k=1}^{N_i(t)}B_{k,i},\:\:\:\:
i=1,2.\]
In this definition of $Z_i(t)$, we assume that the processes $N_i(\cdot)$ are independent Poisson processes with rates $\lambda_i>0$.
In addition,
$(B_{k,1})_k$ and $(B_{k,2})_k$ are both sequences of i.i.d.\ non-negative random variables, which are
independent of the processes $N_1(\cdot)$ and $N_2(\cdot)$.
We denote by $B_1, B_2$ the
generic random variables corresponding to the sequences $B_{k,1}$ and $B_{k,2}$, where $F_1(\cdot)$ and $F_2(\cdot)$ denote their respective distribution functions.
The following proposition translate the
findings of the previous section into the setting of the
compound Poisson input model.
\begin{Proposition}
Assume that both $Z_1(\cdot)$ and $Z_2(\cdot)$ are independent compound Poisson processes
with $\mu_i={\bb E}[Z_i(1)]=\lambda_i{\bb E}[B_i]$ and
$1-F_i(x)\sim x^{-\alpha_i}L_i(x)$, as $x\to\infty$, for $i=1,2$ and $L_i(\cdot)$
being slowly varying at $\infty$, with $\alpha_i>1$.
1) If $\mu_2>\phi_2 c$, then, as $u\to\infty$,
$${\bb P}(Q_1>u) \sim \frac{\lambda_1}{\phi_1c-\mu_1}\frac{1}{\alpha_1 -1}u^{1-\alpha_1}L_1(u).$$
\\
2) If $\mu_2<\phi_2 c$ and $\alpha_1<\alpha_2$, then, as $u\to\infty$, $${\bb P}(Q_1>u) \sim \frac{\lambda_1}{c-\mu}\frac{1}{\alpha_1 -1}u^{1-\alpha_1}L_1(u).$$
3) If $\mu_2<\phi_2 c$, $\mu_1<\phi_1 c$ and $\alpha_2<\alpha_1$,
then, as $u\to\infty$, $${\bb P}(Q_1>u) \sim \frac{\lambda_1}{c-\mu}\frac{1}{\alpha_1 -1}u^{1-\alpha_1}L_1(u).$$
4) If $\mu_1>\phi_1 c$, $\alpha_2<\alpha_1$ and $Z_2$ is spectrally positive, then, as $u\to\infty$, $${\bb P}(Q_1>u) \sim \frac{\lambda_2}{c-\mu}\left( \frac{\mu_1-\phi_1c}{\phi_2c-\mu_2}
\right)^{\alpha_2-1}\frac{1}{\alpha_2 -1}u^{1-\alpha_2}L_2(u).$$
\end{Proposition}
\proof
The proof follows straightforwardly from
Theorems \ref{th.secUnstab}, \ref{th.secStabH}, \ref{th.secStabL} and \ref{th.reqc}, respectively, in combination with Theorem 2.1 in \cite{ASM}.
\endproof
\subsection{$\alpha$-stable L\'evy input}\label{ss.stable}
In this second subsection we focus on the special case of $Z_1(\cdot)$ and $Z_2(\cdot)$
being independent $\alpha_j$-stable L\'evy motions. This formally means that its law is given in terms of its characteristic function:
\[
\log{\bb E} e^{i\theta Z_j(1)} =
- |\theta|^{\alpha_j} (1-i
\beta_j{\rm sign}(\theta)\tan(\pi\alpha_j/2))+i \mu_j\theta,
\]
where $\alpha_j\in(1,2]$, $\beta_j\in(-1,1]$,
$\mu_j\in{\mathbb R}$, and ${\rm
sign}(x):=1_{(0,\infty)}(x)-1_{(-\infty,0)}(x).$
We write $Z_j\in \mathbb{S}(\alpha_j,\beta_j,\mu_j),$
see e.g., \cite{TaS94} or \cite{DeM15}.
Using that
\[{\bb P}(Z_j(1)>x) \sim
c_{\alpha_j}(1+\beta_j)x^{-\alpha_j},\:\:\:\:\mbox{}\:\:\:\:c_{\alpha} :=
\frac{1-\alpha}{2\Gamma(2-\alpha)\cos (\pi\alpha/2)},\]
see \cite{TaS94}, in combination with the results presented in Section \ref{s.main},
we arrive at the following proposition.
\begin{Proposition}
\label{alphacol}
Suppose that $Z_i\in \mathbb{S}(\alpha_i,\beta_i,\mu_i),$ with $\alpha_i\in(1,2)$ for $i=1,2$.
1) If $\mu_2>\phi_2 c$, then, as $u\to\infty$, $${\bb P}(Q_1>u) \sim \frac{c_{\alpha_1}(1+\beta_1)}{(\phi_1c-\mu_1)(\alpha_1-1)}u^{1-\alpha_1}.$$
2) If $\mu_2<\phi_2 c$ and $\alpha_1<\alpha_2$, then, as $u\to\infty$, $${\bb P}(Q_1>u) \sim \frac{c_{\alpha_1}(1+\beta_1)}{(c-\mu)(\alpha_1-1)}u^{1-\alpha_1}.$$
3) If $\mu_2<\phi_2 c$, $\mu_1<\phi_1 c$ and $\alpha_2<\alpha_1$, then, as $u\to\infty$, $${\bb P}(Q_1>u) \sim \frac{c_{\alpha_1}(1+\beta_1)}{(c-\mu)(\alpha_1-1)}u^{1-\alpha_1}.$$
4) If $\mu_1>\phi_1 c$, $\alpha_2<\alpha_1$ and $\beta_2 = 1$, then, as $u\to\infty$,
$${\bb P}(Q_1>u) \sim \frac{2c_{\alpha_2}}{(c-\mu)(\alpha_2-1)}
\left( \frac{\mu_1-\phi_1c}{\phi_2c-\mu_2}
\right)^{\alpha_2-1}u^{1-\alpha_2}.$$
\end{Proposition}
\begin{Rem}\label{rem}{\em
Complementary to case 4) of Proposition \ref{alphacol},
for $\beta_2 \in (-1,1]$ (and $\mu_1>\phi_1 c$, $\alpha_2<\alpha_1$),
we can find asymptotic upper and lower bounds on ${\bb P}(Q_1>u)$
that are tight up to a constant.
In particular,
combining the proof of Theorem \ref{th.reqc} with Theorem 5.3 and Lemma 5.4 in \cite{DMvUTand} we obtain,
as $u\to\infty$,
$$\limsup_{u\to\infty}
{\bb P}(Q_1>u)u^{-(1-\alpha_2)} \leq
\left( \frac{c_{\alpha_2}(1+\beta_2)}{(c-\mu)(\alpha_2-1)}+
\frac{c_{\alpha_2}(1+\beta_2)}{\phi_2c}\right)
\left( \frac{\mu_1-\phi_1c}{\phi_2c}
\right)^{\alpha_2-1},$$\\
and
$$\liminf_{u\to\infty}
{\bb P}(Q_1>u)u^{-(1-\alpha_2)}\geq \frac{c_{\alpha_2}(1+\beta_2)}{(c-\mu)(\alpha_2-1)}\left( \frac{\mu_1-\phi_1c}{\phi_2c}
\right)^{\alpha_2-1}.$$}
\end{Rem}
\section{Proofs}
\label{s.proofs}
Before we provide detailed proofs of the
results of Section \ref{s.main}, we present some
useful lemmas.
We begin with the classical result by Port \cite{Por89}, describing the asymptotics of the tail distribution of a single queue that is emptied at rate $c$.
\begin{Lemma}
\label{th.Port}
Suppose that $Z_1$ satisfies {\bf A1}, {\bf A2} with
$c>\mu_1$. Then, as $u\to\infty$,
$${\bb P}(Q_1^c > u) \sim \frac {1}{c-\mu_1}\frac{1}{\alpha_1-1}u^{1-\alpha_1}L_1(u
.$$
\end{Lemma}
The following result is due to Willekens \cite{Wil87}, describing the asymptotic distribution of the
supremum of $Z_1(-t,0) - ct$ over a finite interval.
\begin{Lemma}
\label{th.Willekens}
Suppose that $Z_1$ satisfies {\bf A1}. Then, for each $T>0$, as $u\to\infty$,
$$
{\bb P} \left( \sup_{t \in [0, T]}\{ Z_1(-t,0) - ct\}>u \right)\sim {\bb P} (Z_1(1)>u).
$$
\end{Lemma}
Whereas the previous lemma considers the supremum over a finite interval, in the next lemma
the interval grows with the exceedance level $u$. This result may have appeared in some form in the literature, but we decided to include it here, as it has a natural and insightful proof.
\begin{Lemma}\label{lemma2}
Suppose that $Z_1$ satisfies {\bf A1}
with $c>\mu_1$ and $\lim_{u\to\infty}{T(u)}/{u}=\infty$. Then, as $u\to\infty$,
$${\bb P}(Q_1^c > u)\sim {\bb P}\left(\sup_{t\in [0, T(u)]}\{Z_1(t)-ct\}>u\right).$$
\end{Lemma}
\proof
Observe that the following trivial inequality holds:
\[
{\mathbb P}({\mathscr E}(u))\leq {\bb P}(Q_1^c > u)\leq {\mathbb P}({\mathscr E}(u))+{\mathbb P}({\mathscr F}(u)),\]
with
\[{\mathscr E}(u):=\left\{
\sup_{t\in [0, T(u)]}\{Z_1(t)-ct\}>u\right\},\:\:\:\:
{\mathscr F}(u):=\left\{\sup_{t\geq T(u)}\{Z_1(t)-ct\}>u\right\}.
\]
Let $\widetilde{Q}_1^{c}\stackrel{\rm d}{=} Q_1^{c}$, with $\widetilde{Q}_1^{c}$ being independent of $\{Z_1(t), t\in\mathbb{R}\}$.
Then using the independence and stationarity of the increments of $Z_1$, we have, with $\varepsilon\in(0, c-\mu_1)$,
\begin{eqnarray*}
{\mathbb P}({\mathscr F}(u))
&=&{\bb P}\left(Z_1(T(u))-cT(u)+\sup_{t\geq T(u)}\{Z_1(t)-Z_1(T(u))-c(t-T(u))\}>u\right)\\
&=&{\bb P}\left(Z_1(T(u))+\widetilde{Q}_1^{c}>u+cT(u)\right)\\
&=&{\bb P}\left(Z_1(T(u))-(\mu_1+\varepsilon)T(u)+\widetilde{Q}_1^{c}>u+(c-\mu_1-\varepsilon)T(u)\right).
\end{eqnarray*}
By applying ${\bb P}(X+Y\geq z) \leq {\bb P}(X\geq fz) +{\bb P}(Y\geq (1-f)z)$ for $f\in(0,1)$, this quantity is in turn bounded from above by, with $\Delta:=c-\mu_1-\varepsilon>0$,
\[{\bb P}\left(Z_1(T(u))-(\mu_1+\varepsilon)T(u)\geq \frac{1}{2}\,\Delta T(u)\right)
+{\bb P}\left({Q}_1^{c}>u+\frac{1}{2}\,\Delta T(u)\right).\]
We prove for each of these probabilities that they are $o\left({\bb P}(Q_1^c > u)\right)$
as $u\to\infty$.
The first probability is majorized by
\[{\bb P}\left(\sup_{t\geq 0}\{Z_1(t)-(\mu_1+\varepsilon)t\}\geq \frac{1}{2}\,\Delta T(u)\right) =
{\bb P}\left(Q_1^{\mu_1+\varepsilon} \geq \frac{1}{2}\,\Delta T(u)\right),\]
Now it follows from Lemma \ref{th.Port} that, recalling that $u=o(T(u))$,
$${\bb P}\left(Q_1^{\mu_1+\varepsilon} \geq \frac{1}{2}\,\Delta T(u)\right)=o\left({\bb P}(Q_1^c > u)\right), \ \ {\bb P}\left({Q}_1^{c}>u+\frac{1}{2}\,\Delta T(u)\right)=o\left({\bb P}(Q_1^c > u)\right). $$
\iffalse
This we do by intersecting ${\mathscr E}(u)$ with three appropriately chosen events: with $\Delta:= \frac{1}{2}(c-\mu_1)$,
\begin{eqnarray*}
{\bb P}\left({\mathscr E}(u)\right)& \leq&{\bb P}\left({\mathscr E}(u), Z_1(T(u))-cT(u)<-\Delta T(u)\right)\\
&& +\:{\bb P}\left({\mathscr E}(u), -\Delta T(u)\leq Z_1(T(u))-cT(u)\leq \sqrt{u}\right)\\
&& +\:{\bb P}\left({\mathscr E}(u), Z_1(T(u))-cT(u)>\sqrt{u}\right)\\
&\leq& {\bb P}\left(\widetilde{Q}_1^{c}>u+\Delta T(u)\right)+{\bb P}\left(Z_1(T(u))-cT(u)>\sqrt{u}\right)\\
&& +\:{\bb P}\left(\widetilde{Q}_1^{c}>u-\sqrt{u}\right){\bb P}\left( Z_1(T(u))-cT(u)\geq -\Delta T(u)\right).
\end{eqnarray*}
It follows from Lemma \ref{th.Port} that
$${\bb P}\left(\widetilde{Q}_1^{c}>u+\Delta T(u)\right)=o\left({\bb P}(Q_1^c > u)\right), \ \ {\bb P}\left(\widetilde{Q}_1^{c}>u-\sqrt{u}\right)\sim {\bb P}(Q_1^c > u). $$
Note that ${\bb P}\left( Z_1(T(u))-cT(u)\geq-\Delta T(u)\right)\rightarrow 0$ as $u\rightarrow\infty$ (by the law of large numbers), so that $${\bb P}\left(\widetilde{Q}_1^{c}>u-\sqrt{u}\right){\bb P}\left( Z_1(T(u))-cT(u)\geq -\Delta T(u)\right)=o\left({\bb P}(Q_1^c > u)\right).$$
Moreover, in light of Lemma \ref{th.Port} again, for $0<\varepsilon<c-\mu_1$, we have
\begin{eqnarray*}
{\bb P}\left(Z_1(T(u))-cT(u)>\sqrt{u}\right)&=& {\bb P}\left(Z_1(T(u))-(c-\varepsilon)T(u)>\sqrt{u}+\varepsilon T(u)\right)\\
&\leq& {\bb P}(Q_1^{c-\varepsilon} > \sqrt{u}+\varepsilon T(u))=o\left({\bb P}(Q_1^c > u)\right)
\end{eqnarray*}
\fi
Therefore, we conclude that
${\mathbb P}({\mathscr F}(u))=o\left({\bb P}(Q_1^c > u)\right),$
which completes the proof. \endproof
Define, for $\lambda<\mu_2,$ ${T_\lambda}(u):=u/\bar {T_\lambda}(u)$, where
\[\bar T_{\lambda}(u):={\sqrt{{\bb P} \left(\check Q_2^\lambda(0)>u/2\right)\vee (1/\log u) }},\:\:\:
\check Q_2^{\lambda}(s):=\sup_{t\geq s}\{Z_2(s)-Z_2(t)+\lambda (t-s)\}.\]
\begin{Lemma}\label{Lemma1} Suppose that $Z_2$ satisfies {\bf A1} with $\lambda<\mu_2$. Then,
as $u\to\infty$,
\[\xi(u):={\bb P}\left(\sup_{s\in [0,{T_\lambda}(u)]}\check Q_2^{\lambda}(s)>u\right)\to 0.\]
\end{Lemma}
\proof It is immediate that, with $W(t):=Z_2(t)-\lambda t$,
\begin{eqnarray*}
\sup_{s\in [0,{T_\lambda}(u)]}\check Q_2^{\lambda}(s)&=&\sup_{s\in [0,{T_\lambda}(u)]}\sup_{t\geq s}\{Z_2(s)-Z_2(t)+\lambda (t-s)\}\\
&\leq& \sup_{s\in [0,{T_\lambda}(u)]}\sup_{t\geq 0}\{Z_2(s)-Z_2(t)+\lambda (t-s)\}=\check Q_2^{\lambda}(0)+\sup_{s\in[0,{T_\lambda}(u)]}W(s).
\end{eqnarray*}
Setting $$g(u):=u\left(\bar {T_\lambda}(u) \right)^{1/2}, \ \
h(u):=u\left(\bar {T_\lambda}(u) \right)^{1/3},$$
we have
\begin{eqnarray*}
\zeta(u)&:=&{\bb P}\left(\sup_{t\in [0,g(u)]}{\check Q}_2 ^{\lambda}(t)>u\right) \leq {\bb P}\left({\check Q}_2 ^{\lambda}(0)+\sup_{s\in[0, g(u)]}W(s)>u\right)\\
&\leq& {\bb P}\left({\check Q}_2 ^{\lambda}(0)+\sup_{s\in[0,g(u)]}W(s)>u, \sup_{s\in[0,g(u)]}W(s)\leq h(u)\right)\\
&& +\, {\bb P}\left({\check Q}_2 ^{\lambda}(0)+\sup_{s\in[0,g(u)]}W(s)>u, \sup_{s\in[0,g(u)]}W(s)> h(u)\right)\\
&\leq& {\bb P}\left({\check Q}_2 ^{\lambda}(0)>u-h(u)\right)+{\bb P}\left(\sup_{s\in[0,g(u)]}W(s)> h(u)\right)\\
&\leq& {\bb P}\left({\check Q}_2 ^{\lambda}(0)>u/2\right)+{\bb P}\left(\sup_{s\in[0,g(u)]}\{Z_2(s)-(\mu_2+1)s\}> h(u)-(\mu_2-\lambda+1)g(u)\right)\\
&\leq& {\bb P}\left({\check Q}_2 ^{\lambda}(0)>u/2\right)+{\bb P}\left(Q_2^{\mu_2+1}> h(u)/2\right).
\end{eqnarray*}
Using the stationarity of ${\check Q}_2 ^{\lambda}(t)$, and noting that $m(u):=T_\lambda(u)/g(u)\to\infty$ as $u\to\infty$,
\[\xi(u) \leq \sum_{i=1}^{[ m(u)]+1}{\bb P}\left(\sup_{t\in[(i-1)g(u),\,i\,g(u)]}{\check Q}_2 ^{\lambda}(t)>u\right)=([ m(u)]+1) \,\zeta(u).\]
Now, by applying Lemma \ref{th.Port}, we have (noting that $h(u)\to\infty$ as $u\to \infty$)
that there is a positive constant $\kappa$ such that
\begin{eqnarray*}
\xi(u)&\leq& \frac{2{T_\lambda}(u)}{g(u)}\zeta(u)
\leq\frac{2{T_\lambda}(u)}{g(u)}\left({\bb P}\left({\check Q}_2 ^{\lambda}(0)>u/2\right)+{\bb P}\left(Q_2^{\mu_2+1}> h(u)/2\right)\right)\\
&\leq& 2\left({\bb P}\left({\check Q}_2 ^{\lambda}(0)>u/2\right)\right)^{1/4}+\kappa\,(\log u)^{3/4}\,L_2(h(u))(h(u))^{1-\alpha_2}\to 0.
\end{eqnarray*}
This completes the proof. \endproof
The following lemma plays an important role in the proof of Theorem \ref{th.secStabL}.
\begin{Lemma}\label{BIG}
Suppose that $Z_2$ satisfies {\bf A1} with $\lambda<\mu_2$. Then
$$\frac{Q_2^\lambda(t)}{t}\rightarrow 0, \:\:\:{\rm a.s.},\:\:\: {\rm as}\: t\rightarrow\infty.$$
\end{Lemma}
\proof It suffices to prove that for some $\beta>0$,
\begin{eqnarray}\label{e1}
\sup_{t\in [n^\beta, (n+1)^\beta]}\frac{Q_2^{\lambda}(t)}{n^{\beta}}
\rightarrow 0, \:\:\:{\rm a.s.},\:\:\: {\rm as} \: n\rightarrow\infty.
\end{eqnarray}
Let $\beta>({\alpha_2-1})^{-1}>0$ hereafter.
Then, with $n_+:=n+1$, $I_n:=[n^\beta, n_+^\beta]$,
\begin{eqnarray*}
\sup_{t\in I_n}Q_2^{\lambda}(t)&=&
\sup_{t\in I_n}\sup_{s\leq t}\{Z_2(t)-Z_2(s)-\lambda(t-s)\}\\
&\leq &\sup_{t\in I_n}\sup_{s\leq n_+^\beta}\{Z_2(n_+^\beta)-Z_2(s)-\lambda(n_+^\beta-s)+Z_2(t)-Z_2(n_+^\beta) -\lambda(t-n_+^\beta)\}\\
&=&
Q_2^\lambda(n_+^\beta)+\sup_{t\in I_n}\{W(t)-W(n_+^\beta)\},
\end{eqnarray*}
where, as before, $W(t):=Z_2(t)-\lambda t$.
In light of Lemma \ref{th.Port} we have that for any $\varepsilon>0$ and $\gamma\in(1, \beta(\alpha_2-1))$ there exists $N_0\in{\mathbb N}$ such that (where we recall that $\beta>({\alpha_2-1})^{-1}$)
\[\sum_{n=N_0}^\infty{\bb P}\left(\frac{Q_2^\lambda(n_+^\beta)}{n^{\beta}}>\varepsilon\right)=\sum_{n=N_0}^\infty{\bb P}\left(Q_2^\lambda(0)>\varepsilon n^{\beta}\right)\leq \sum_{n=N_0}^\infty n^{-\gamma}<\infty.\]
Using the Borel-Cantelli lemma, we obtain that $Q_2^\lambda(n_+^\beta)/n^\beta \rightarrow 0$ a.s.\ as $n\rightarrow\infty.$
In order to establish (\ref{e1}) we are now left to prove that $n^{-\beta} \sup_{t\in I_n}\{W(t)-W(n_+^\beta)\} \rightarrow 0$ a.s., as $n\rightarrow\infty.$ This convergence is established as follows.
By the fact that
${Z_2(t)}/t \rightarrow \mu_2,$ a.s., as $t\rightarrow\infty$, we have ${W(t)}/t \rightarrow \mu_2-\lambda,$ a.s., as $t\rightarrow\infty$, and
$$\sup_{s,t\geq n}\left\{\frac{W(s)}{s}-\frac{W(t)}{t}\right\}\rightarrow 0, \:\mbox{a.s.},\:\:\mbox{as $n\rightarrow\infty$}.$$
Therefore,
\begin{eqnarray}
0&\leq&\frac{1}{n^{\beta}}\sup_{t\in I_n}\{W(t)-W(n_+^\beta) \}\leq\frac{n_+^\beta}{n^\beta}\sup_{t\in I_n}\left\{\frac{W(t)}{t}-\frac{W(n_+^\beta)}{t}\right\}
\nonumber\\
&\leq& \frac{n_+^\beta}{n^\beta}\left(\sup_{t\in I_n}\left\{\frac{W(t)}{t}-\frac{W(n_+^\beta)}{n_+^\beta}\right\}
+\left|W(n_+^\beta)\right|\sup_{t\in I_n}
\left\{\left| \frac{1}{n_+^\beta}-\frac{1}{t}\right|\right\}
\right)
\rightarrow 0,\: \:\mbox{a.s.}, \nonumber
\end{eqnarray}
as $n\rightarrow\infty$.
This confirms (\ref{e1}) and thus the proof has been completed. \endproof
\subsection{Proof of Theorem \ref{th.secUnstab}}
{\it Upper bound}:
Observe that
\begin{equation}
\label{upper1}
{\bb P}(Q_1>u) \leq {\bb P}(Q_1^{\phi_1 c} >u) \sim \frac{1}{(\phi_1 c-\mu_1)(\alpha_1-1)}u^{1-\alpha_1}L_1(u),
\end{equation}
by Lemma \ref{th.Port}.
\vspace{2mm}
{\it Lower bound}:
Since
\begin{equation*}
\begin{split}
Q_1
& = \sup_{t \geq 0}\left\{Z_1(-t,0) + Z_2(-t,0) - ct - \sup_{s\in [0,t)}\{Z_2(-s,0) - C_2(-s,0)\} \right\} \\
& \geq \sup_{t \geq 0}\left\{Z_1(-t,0) + Z_2(-t,0) - ct - \sup_{s\in [0,t)}\{Z_2(-s,0) - \phi_2cs\} \right\},
\end{split}
\end{equation*}
for any $\varepsilon\in(0,1)$, application of Lemmas \ref{lemma2}-\ref{Lemma1} yields,
with $\lambda=\phi_2c$,
\begin{eqnarray}\label{star}
\lefteqn{
{\bb P}(Q_1 > u)
\geq
{\bb P} \left( \sup_{t \geq 0} \left\{ Z_1(-t,0) + Z_2(-t,0) - ct - \sup_{s\in [0,t)}\{Z_2(-s,0) - \phi_2cs\} \right\} > u \right)}\nonumber \\
&=&
{\bb P} \left( \sup_{t \geq 0} \left\{ Z_1(t) - \phi_1 c t - \sup_{s \in [0,t)} \{ Z_2(s)-Z_2(t) - \phi_2c(s-t) \} \right\} > u \right) \nonumber\\
&\geq &
{\bb P} \left( \sup_{0\leq t \leq {T_\lambda}(\varepsilon u)} \left\{ Z_1(t) - \phi_1 c t - \sup_{0\leq s \leq {T_\lambda}(\varepsilon u), v> s} \{ Z_2(s)-Z_2(v) - \phi_2c(s-v) \} \right\} > u \right) \nonumber\\
&= &
{\bb P} \left( \sup_{0\leq t \leq {T_\lambda}(\varepsilon u)} \left\{ Z_1(t) - \phi_1 c t \right\} - \sup_{0\leq s \leq {T_\lambda}(\varepsilon u)} \check Q_2^{\phi_2c}(s) > u
\right)
\nonumber\\
&\geq&
{\bb P} \left( \sup_{0\leq t \leq {T_\lambda}(\varepsilon u)} \left\{ Z_1(t) - \phi_1 c t \right\} - \sup_{0\leq s \leq {T_\lambda}(\varepsilon u)} \check Q_2^{\phi_2c}(s)> u ,
\sup_{0\leq s \leq {T_\lambda}(\varepsilon u)} \check Q_2^{\phi_2c}(s)\leq \varepsilon u\right)\nonumber\\
&\geq&
{\bb P} \left( \sup_{0\leq t \leq {T_\lambda}(\varepsilon u)} \left\{ Z_1(t) - \phi_1 c t \right\} > (1+\varepsilon)u \right){\bb P}\left( \sup_{0\leq s \leq {T_\lambda}(\varepsilon u)} \check Q_2^{\phi_2c}(s)\leq \varepsilon u\right)\nonumber\\
&\sim& {\bb P} \left( Q_1^{\phi_1 c} > u \right), \ \ u\to\infty, \,\varepsilon\downarrow 0,\nonumber
\end{eqnarray}
which combined with Lemma \ref{th.Port}
leads to the asymptotic upper bound that matches the lower bound.
This completes the proof.
\hfill $\Box$
\subsection{Proof of Theorem \ref{th.secStabH}}
{\it Upper bound}:
Combining
\begin{equation}
\label{upper2}
{\bb P}(Q_1>u) \leq {\bb P}(Q_1 + Q_2 >u) = {\bb P} \left( \sup_{t \geq 0} \{ Z_1(-t,0) + Z_2(-t,0) - ct \} > u \right)
\end{equation}
with ${\bb P}(Z_1(1)+Z_2(1)>u)\sim {\bb P}(Z_1(1)>u)$ as $u\to\infty$,
together with Lemma \ref{th.Port}, straightforwardly
gives that
\begin{eqnarray}\label{upper22}
{\bb P}(Q_1>u)
\sim
\frac{1}{c-\mu}\int_u^{\infty}{\bb P}(Z_1(1)>x){\rm d}x \sim \frac{1}{(c-\mu)(\alpha_1-1)}u^{1-\alpha_1}L_1(u),
\end{eqnarray}
as $u\to\infty$.
\vspace{2mm}
{\it Lower bound}:
Let $\varepsilon >0$ be given.
Following the same argument as in the lower bound of the proof of
Theorem \ref{th.secUnstab}, we have, with $\lambda=\mu_2^\varepsilon:= \mu_2-\varepsilon$,
\begin{eqnarray*}
&&{\bb P}(Q_1 > u)\\
&& \ \ \geq\:
{\bb P} \left( \sup_{t \geq 0}\left\{ Z_1(-t,0) + Z_2(-t,0) - ct - \sup_{s\in [0,t)} \left\{ Z_2(-s,0) - \phi_2cs \right\} \right\} > u \right) \\
&& \ \ =\:
{\bb P} \left( \sup_{t\geq 0} \left\{ Z_1(t) - (c-\mu_2^\varepsilon)t -
\sup_{s \in [0,t)} \{ Z_2(s)-Z_2(t) - \phi_2cs +\mu_2^\varepsilon t\} \right\} > u \right)\\
&&\ \ \geq\: {\bb P} \left( \sup_{t\in [0,{T_\lambda}(\varepsilon u)]} \left\{ Z_1(t) - (c-\mu_2^\varepsilon)t -
\sup_{s \in [0,{T_\lambda}(\varepsilon u)), s< t} \{ Z_2(s)-Z_2(t) +\mu_2^\varepsilon(t-s)\} \right\} > u \right)\\
&& \ \ \geq \:{\bb P} \left( \sup_{t\in [0,{T_\lambda}(\varepsilon u)]} \left\{ Z_1(t) - (c-\mu_2^\varepsilon)t -
\sup_{s \in [0,{T_\lambda}(\varepsilon u)]}\check Q_2^{\mu_2^\varepsilon} (s)\right\} > u \right)\\
&& \ \ \geq \: {\bb P} \left( \sup_{t\in [0,{T_\lambda}(\varepsilon u)]} \left\{ Z_1(t) - (c-\mu_2^\varepsilon)t
\right\} > (1+\varepsilon)u \right){\bb P}\left(\sup_{s \in [0,{T_\lambda}(\varepsilon u)]}\check Q_2^{\mu_2^\varepsilon}(s) <\varepsilon u\right).
\end{eqnarray*}
By Lemmas \ref{lemma2}--\ref{Lemma1} in combination with Lemma \ref{th.Port}, we obtain
$${\bb P}(Q_1 > u)\geq {\bb P}(Q_1^{c-\mu_2^\varepsilon}>(1+\varepsilon)u)(1+o(1))\sim \frac{(1+\varepsilon)^{1-\alpha_1}}{(c-\mu+\varepsilon)(\alpha_1-1)}u^{1-\alpha_1}L_1(u), \ \ u\rightarrow\infty.$$
Letting $\varepsilon\downarrow 0$, and recalling (\ref{upper22}), completes the proof.
\hfill $\Box$
\subsection{Proof of Theorem \ref{th.secStabL}}
{\it Upper bound}:
The starting point is the following evident equality:
$$Q_1 =\sup_{t \in [0, u^{1-\varepsilon})}U_1(t)\vee\sup_{t \geq u^{1-\varepsilon}}U_1(t),\:\:\:\:\:U_1(t):=
Z_1(-t,0) - C_1(-t,0).
$$
with $\varepsilon$ strictly between ${\alpha_1}/({1+\alpha_1})$ and $1$. Then, for
$$\mathscr{U}_{\varepsilon} \colonequals \{ \forall t\geq u^{1-\varepsilon} \,:\, Z_2(-t,0)+Q_2^{\phi_2c}(-t) \leq (\mu_2+\varepsilon)t \}$$
we have
\begin{equation}
\label{upper3a}
{\bb P} \left( \sup_{t \geq u^{1-\varepsilon}}U_1(t)> u \right )=
{\bb P} \left ( \sup_{t \geq u^{1-\varepsilon}}U_1(t) > u; \, \mathscr{U}_{\varepsilon} \right )
+ {\bb P} \left( \sup_{t \geq u^{1-\varepsilon}}U_1(t) > u; \, \mathscr{U}_{\varepsilon}^c \right).
\end{equation}
It follows from (\ref{e2}) that on the event $\mathscr{U}_{\varepsilon}$, for $t\geq u^{1-\varepsilon}$,
\begin{eqnarray*}
B_2(-t,0)=Z_2(-t,0)+Q_2(-t)-Q_2(0)\leq Z_2(-t,0)+Q_2^{\phi_2c}(-t)\leq (\mu_2+\varepsilon)t,
\end{eqnarray*}
which together with the fact that $C_1(s,t)+B_2(s,t)=c(t-s)$ for all $s\leq t$ yields, for $t\geq u^{1-\varepsilon}$
$$C_1(-t,0)\geq (c-\mu_2-\varepsilon)t.$$
Moreover,
\begin{eqnarray}
\nonumber\label{upper3aa}
{\bb P} \left( \sup_{t \geq u^{1-\varepsilon}}U_1(t) > u; \, \mathscr{U}_{\varepsilon} \right)
&\leq& {\bb P} \left( \sup_{t\geq u^{1-\varepsilon}}\{Z_1(-t,0) -(c-\mu_2-\varepsilon)t \}>u;\, \mathscr{U}_{\varepsilon} \right) \\
&\leq &{\bb P} \left( \sup_{t\geq 0}\{Z_1(-t,0) -(c-\mu_2-\varepsilon)t\}>u \right) {\bb P}( \mathscr{U}_{\varepsilon} );
\label{upper3aa}
\end{eqnarray}
the first term in (\ref{upper3aa}) is roughly of the order $u^{1-\alpha_1}$, whereas ${\bb P}( \mathscr{U}_{\varepsilon}) \to 1$, as $u\to\infty$,
as a consequence of the law of large numbers
and Lemma \ref{BIG}.
In addition,
\begin{eqnarray}\nonumber
{\bb P} \left( \sup_{t \geq u^{1-\varepsilon}}U_1(t)> u; \, \mathscr{U}_{\varepsilon}^c \right)
& \leq& {\bb P} \left( \sup_{t\geq u^{1-\varepsilon}}\{Z_1(-t,0) -\phi_1ct \}>u;\, Y\mathscr{U}_{\varepsilon}^c \right)\\
&\leq &{\bb P} \left( \sup_{t\geq 0}\{Z_1(-t,0) -\phi_1ct\}>u \right) {\bb P}(\mathscr{U}_{\varepsilon}^c ),
\label{upper3ab}
\end{eqnarray}
where the first term in (\ref{upper3ab}) essentially vanishes as $u^{1-\alpha_1}$, but ${\bb P}(\mathscr{U}_{\varepsilon}^c ) \to 0$ due to the law of large numbers
and Lemma \ref{BIG}. We conclude it is negligible relative to (\ref{upper3aa}).
Combining (\ref{upper3aa}) and (\ref{upper3ab}) gives that
\[
{\bb P} \left(\sup_{t \geq u^{1-\varepsilon}}\{ Z_1(-t,0) - C_1(-t,0)\} > u \right) \leq {\bb P}(Q_1^{c-\mu_2-\varepsilon}>u ), \quad \textrm{as} \, u\to \infty.
\]
Now we are left with showing that
\[
{\bb P}(Q_1 > u) \sim {\bb P} \left (\sup_{t \geq u^{1-\varepsilon}}U_1(t)>u \right), \,\:\: \textrm{as} \:u \to \infty.
\]
Since we have that
\begin{equation}
\label{negl}
{\bb P}(Q_1>u) = {\bb P} \left( \sup_{t\geq 0}U_1(t)>u \right) \leq
{\bb P} \left( \sup_{t \geq u^{1-\varepsilon}}U_1(t)>u \right)
+ {\bb P} \left(\sup_{t\in [0, u^{1-\varepsilon}]}U_1(t)>u \right)
\end{equation}
and we have already showed that both
${\bb P}\left( \sup_{t \geq u^{1-\varepsilon}}U_1(t)>u \right)$ and ${\bb P}(Q_1 >u)$ are of order $u^{1-\alpha_1}$,
we see that it suffices to prove that the last term in (\ref{negl}) is negligible.
Let
$$S_n:=\sup_{s\in [n,n+1]}Z_1(s)-Z_1(n)-\phi_1c(s-n), \:\:\:n\in \mathbb{N}.$$
Using that $S_n, n\in \mathbb{N}$ are i.i.d., we get
\begin{eqnarray*}
{\bb P} \left( \sup_{t\in [0, u^{1-\varepsilon}]}U_1(t)>u \right)
&\leq& {\bb P} \left( \sup_{t\in [0, u^{1-\varepsilon}]}\{ Z_1(-t,0)-\phi_1ct)\}>u \right) \\
& = &{\bb P} \left( \sup_{t\in [0,u^{1-\varepsilon}]}\{Z_1(t)-\phi_1ct)\}>u \right)\\
& \leq &{\bb P} \left( \sum_{i=0}^{[u^{1-\varepsilon}]} S_i>u \right)\leq ([u^{1-\varepsilon}]+1){\bb P}\left(S_0>\frac{u}{[u^{1-\varepsilon}]+1}\right).
\end{eqnarray*}
Hence, Lemma \ref{th.Willekens}, in combination with the fact that $\varepsilon$ lies strictly between ${\alpha_1}/({1+\alpha_1})$ and $1$, leads to
\begin{eqnarray*}
{\bb P} \left( \sup_{s\in [0, u^{1-\varepsilon}]}U_1(s)>u \right)\leq \bar \kappa \,L_1(u^{\varepsilon})u^{1-\varepsilon-\varepsilon \alpha_1}=o\left(L_1(u)u^{1-\alpha_1}\right),
\end{eqnarray*}
where $\bar\kappa$ is a positive constant.
This completes the proof of the upper bound.
\vspace{2mm}
{\it Lower bound}:
The proof of the lower bound is the same as in the proof of Theorem \ref{th.secStabH}.
Relying on Lemma \ref{th.Port}, we then obtain the equivalence of the asymptotic upper and lower bound.
\hfill $\Box$
\subsection{Proof of Theorem \ref{th.reqc} }
Since the proof of this scenario needs a case-specific approach that involves the notion of tandem systems,
we begin with some notation and auxiliary results.
For $\varepsilon$ such that $\phi_1 c-\mu_1<\varepsilon<c-\mu$, let
$$V^{\varepsilon} \colonequals \sup_{t \geq 0}\{Z_2(-t,0) - (c-\mu_1 - \varepsilon)t\} - \sup_{s \geq 0} \{Z_2(-s,0) - \phi_2cs\}.$$
Recall that
$$Q_1^d \colonequals \sup_{ t \geq 0}\{Z_1(-t,0) - dt\}, \quad d>\mu_1$$
and introduce
$$ \check Q_1^d \colonequals \sup_{t \geq 0}\{dt - Z_1(-t,0)\}, \quad d<\mu_1.$$
The following lemma states a straightforward counterpart of Lemma 2.1 in \cite{DMGPS}.
\begin{Lemma}
\label{lemtand}
For $\varepsilon >0$ small enough, any $u$ and $x$, and for $\delta \in (0,1)$, we have
$${\bb P}(V^{-\varepsilon} > u+x){\bb P}(\check Q_1^{\mu_1 - \varepsilon}\leq x) \leq {\bb P}(Q_1>u) \leq {\bb P}(V^{\varepsilon}>(1-\delta)u) + {\bb P}(Q_1^{\mu_1 +\varepsilon}>\delta u)$$
\end{Lemma}
A combination of Theorem 4.7 in \cite{LiM} (see also Theorem 12.9 in \cite{DeM15})
with Lemma \ref{th.Port} leads to the following lemma.
\begin{Lemma}
\label{tandAs}
Let $|\varepsilon|<\min(c-\mu,\mu_1- \phi_1 c)$ and $Z_2$ be spectrally positive
with $\alpha_2\notin {\bb N}$. Then, as $u\to\infty$,
\[
{\bb P}(V^\varepsilon>u)\sim
\left( \frac{\mu_1-\phi_1c +\varepsilon}{\phi_2c-\mu_2} \right)^{\alpha_2-1}\frac{1}{(c-\mu-\varepsilon)(\alpha_2-1)}u^{1-\alpha_2}L_2(u) .
\]
\end{Lemma}
\textit{Proof of Theorem \ref{th.reqc}}:
Let $\delta\in (0,1)$ and $\varepsilon>0$ be such that $\varepsilon<\min(c-\mu,\mu_1- \phi_1 c)$.
Then, following Lemma \ref{tandAs}, as $u\to\infty$,
\begin{eqnarray*}
{\bb P}(\check Q_1^{\mu_1-\varepsilon}<\sqrt{u})&\to& 1,\\
{\bb P}(V^{-\varepsilon}>u+\sqrt{u})
&\sim&
\left( \frac{\mu_1-\phi_1c -\varepsilon}{\phi_2c-\mu_2} \right)^{\alpha_2-1}\frac{1}{(c-\mu+\varepsilon)(\alpha_2-1)}u^{1-\alpha_2}L_2(u)\\
{\bb P}(V^{\varepsilon}>(1-\delta)u)
&\sim&
\left( \frac{\mu_1-\phi_1c +\varepsilon}{\phi_2c-\mu_2} \right)^{\alpha_2-1}\frac{1}{(c-\mu-\varepsilon)(\alpha_2-1)}(1-\delta)^{1-\alpha_2}u^{1-\alpha_2}L_2(u).
\end{eqnarray*}
Since $\alpha_1>\alpha_2$, we find by Lemma \ref{th.Port} that for each $\delta \in(0,1)$,
\begin{eqnarray*}
{\bb P}(Q_1^{\mu_1+\varepsilon}>\delta u)=o({\bb P}(V^{\varepsilon}>(1-\delta)u)),
\end{eqnarray*}
as $u\to\infty$.
Thus, by Lemma \ref{lemtand}, passing with $\varepsilon,\delta\downarrow 0$,
we obtain that
\[
{\bb P}(Q_{1} > u) \sim \left( \frac{\mu_1-\phi_1c}{\phi_2c-\mu_2} \right)^{\alpha_2-1}\frac{1}{(c-\mu)(\alpha_2-1)}u^{1-\alpha_2}L_2(u).
\]
This completes the proof.
\hfill $\Box$
\section*{Acknowledgments}
{\small
K. D\c ebicki
was partially supported by NCN Grant No 2015/17/B/ST1/01102 (2016-2019) whereas
P. Liu was partially supported by the Swiss National Science Foundation Grant 200021-166274.
{M.\ Mandjes'} research is partly funded by the NWO Gravitation project N{\sc etworks}, grant number 024.002.003. He is also affiliated to (A)~CWI, Amsterdam, the Netherlands; (B)~E{\sc urandom}, Eindhoven University of Technology, Eindhoven, the Netherlands;
and (C)~Amsterdam Business School, Faculty of Economics and Business, University of Amsterdam,
Amsterdam, the Netherlands.}
|
\section{An introduction to the project}
This is the first in a series of three papers by the authors concerned with the structure theory of \emph{cubespaces}, the others being \cite{GMV2} and \cite{GMV3}. Informally, a cubespace is a compact metric space $X$, together with some notion of when a collection of $2^k$ points of $X$ form a ``$k$-cube'', subject to certain further axioms.
The study of cubespaces as axiomatic objects is becoming established as a major theme in the nascent area of higher order Fourier analysis. This programme has its origins in work of Host and Kra \cite{HK08}, where these objects appeared under the name of ``parallelepiped structures''. The study of these objects was furthered by Antol\'\i n Camarena and Szegedy \cite{CS12}, who in the same work formulated a strong structure theorem for cubespaces, subject to certain further hypotheses. The gist of the structure theorem is that, subject to these extra assumptions, all cubespaces arise in some sense from \emph{nilmanifolds} $X = G/\Gamma$; i.e.~they come from compact homogeneous spaces of nilpotent Lie groups.
The papers of Candela \cites{Can1,Can2} expand on \cite{CS12}, providing more detailed proofs. He also includes
several additional results implicit in \cite{CS12}, particularly about continuous systems of measures.
This structure theory has applications in two broad areas: additive combinatorics, and in particular the inverse theorem for the Gowers norms; and topological dynamics and ergodic theory. In this paper, we will approach this project from the point of view of someone interested primarily in understanding the former, and in particular Szegedy's proof \cite{S12} of the inverse theorem for the Gowers norms, which relies heavily on this structural result. The third paper in the series \cite{GMV3} will contain an introduction to the project focussed instead on applications to topological dynamics. Although the results proved in each paper are strongly relevant to the other, a reader more interested in the dynamical perspective might prefer to start there and refer to this paper only later.
The main purpose of our entire project is to provide a self-contained proof of the main structural result of \cite{CS12}. In many places our approach will follow that of \cite{CS12}, or of previous work \cites{HK05,HM07,HK08,HKM10,GT10}, very closely. At other times, we give notably different arguments: sometimes because they are arguably simpler; sometimes to avoid certain technical difficulties (although perhaps at the expense of introducing different ones); and sometimes to obtain sharper conclusions. Furthermore, we do obtain new results in particular in the dynamical setting, and some of our proofs are optimized so as to prove these concurrently.
Our primary goal, however, is to obtain a fuller understanding of cubespaces and related structures. In our view, this understanding will continue to find new relevance as the field of higher order Fourier analysis matures.
\subsection{Obstructions to Gowers uniformity}
It is not immediately clear why the structure theory of cubespaces should be relevant to proving an inverse theorem for the Gowers norms: indeed, the deduction of the latter from the former is by no means straightforward \cite{S12}. To provide some motivation, we will instead sketch a proof of a kind of converse. That is, we will argue that cubespaces (subject to certain additional hypotheses) are obstructions to Gowers uniformity, and that therefore the inverse theorem itself implies that cubespaces are somehow related to nilmanifolds.
We will not recall in full all the relevant definitions (of the uniformity norms, nilmanifolds, polynomial maps, nilsequences etc.), referring the reader to \cites{GT10,ben-book} but will informally sketch the set-up to motivate our discussion. For notational simplicity we will focus only on the $U^3$ norm, although these remarks apply more generally.
\begin{enumerate}
\item Given a function $f \colon \ZZ/N\ZZ \to \CC$ for $N$ a prime (say), the uniformity norm $\|f\|_{U^3}$ is defined in terms of an average over \emph{cube} or \emph{parallelepiped} configurations in $\ZZ/N\ZZ$, e.g.
\[
\|f\|_{U^3}^8 = \EE_{c \in C^3(\ZZ/N\ZZ)} f(c_{000}) \overline{f(c_{001}) f(c_{010})} f(c_{011}) \overline{f(c_{100})} f(c_{101}) f(c_{110}) \overline{f(c_{111})}
\]
where $C^3(\ZZ/N\ZZ) \subseteq (\ZZ/N\ZZ)^8$ consists of all tuples
\[
(c_\omega)_{\omega \in \{0,1\}^3} = x + \omega_1 h_1 + \omega_2 h_2 + \omega_3 h_3
\]
for $x, h_1, h_2, h_3 \in \ZZ/N\ZZ$. It is informative to think of these tuples as elements written on the vertices of a $3$-dimensional cube, as shown.
\inlinetikz{\threecube{$x$}{$x+h_1$}{$x+h_2$}{$x+h_1+h_2$}{$x+h_3$}{$x+h_1+h_3$}{$x+h_2+h_3$}{$x+h_1+h_2+h_3$}{1.6}}
\item Given a nilmanifold $G/\Gamma$ (with suitable additional structure) there is also a notion of cubes on $G/\Gamma$, given by a construction due to Host and Kra \cites{HK05,HK08}. Specifically, suppose $G$ is a $2$-step nilpotent Lie group and $\Gamma$ a discrete co-compact subgroup; then there is a closed subset $C^3(G/\Gamma) \subseteq (G/\Gamma)^8$ somewhat analogous to the parallelepipeds in an abelian group.
\item There is a plentiful supply of maps $p \colon \ZZ/N\ZZ \to G/\Gamma$ which send cubes to cubes; that is, $p(c) \in C^3(G/\Gamma)$ for any $c \in C^3(\ZZ/N\ZZ)$ (with $p$ applied pointwise).
\item The cubes on $G/\Gamma$ satisfy a \emph{corner constraint}: given $c \in C^3(G/\Gamma)$, if we know $c_{000}, \dots, c_{110}$ then the last vertex $c_{111}$ is uniquely determined by the others.
\item By a \emph{nilsequence} on $\ZZ/N\ZZ$ we mean a function of the form $\phi = F \circ p$ where $p$ is as above and $F \colon G/\Gamma \to \CC$ is Lipschitz.
\end{enumerate}
The inverse theorem states roughly that (for very large $N$) if $f \colon \ZZ/N\ZZ \to \CC$, $|f| \le 1$ has $\|f\|_{U^3} \ge \delta$ then $|\langle f, \phi \rangle| \gtrsim_\delta 1$ for some nilsequence $\phi$ whose ``complexity'' is bounded in terms of $\delta$. One can say that nilsequences are the only ``obstructions to Gowers uniformity'': the only reason for a function to have large Gowers norm is if it correlates with a nilsequence.
All known proofs of this statement are significantly hard. By contrast, the converse statement -- if a function correlates with a nilsequence then it has large Gowers norm -- is quite straightforward. We will sketch a result along these lines, following the argument from \cite{green-tao-u3}*{Proposition 12.6}. Although the details of the proof are not logically required in what follows, it is useful to record them to motivate Observation \ref{obs:obstruct} below.
\begin{claim*}
Suppose $\phi = F \circ p$ is a nilsequence in the sense of (v) with $\|F\|_{\infty} \le 1$. Let $f \colon \ZZ/N\ZZ \to \CC$ be such that $|f| \le 1$ and $|\langle f, \phi\rangle| \ge \delta$. Then $\|f\|_{U^3} = \Omega_{F, \delta}(1)$.
\end{claim*}
The key point is that the lower bound on $\|f\|_{U^3}$ depends only on the choice of $F$ (and so implicitly of $G/\Gamma$) and on $\delta$; \emph{not} on $N$ or $p$.
\begin{proof}[Proof sketch]
By property (iv), there is a closed subset $Y \subseteq (G/\Gamma)^7$ and a function $\tau \colon Y \to (G/\Gamma)$ such that
\[
C^3(G/\Gamma) = \{( \tau(y), y) \colon y \in Y \} \ .
\]
Hence we get a continuous function $F \circ \tau$ on $Y$. By Tietze's extension theorem, we can extend this to a bounded continuous function $H$ on $(G/\Gamma)^7$. Any continuous function on a product space can be approximated (up to a small error in $L^\infty$) by a finite sum of products of functions on the factors: that is, we can decompose
\[
H(x_1,\dots,x_7) = \sum_{i=1}^k R_1^{(i)}(x_1) \dots R_7^{(i)}(x_7) + H_{\text{err}}
\]
for some bounded continuous functions $R_j^{(i)} \colon G/\Gamma \to \CC$, and some bounded continuous $H_{\text{err}} \colon (G/\Gamma)^7 \to \CC$ such that $\|H_{\text{err}}\|_\infty = o_{F;k \to \infty}(1)$.
Now, for any $x, h_1, h_2, h_3$ in $\ZZ/N\ZZ$ we have that $(p(x), p(x+h_1), p(x+h_2), p(x+h_1+h_2), \dots)$ is in $C^3(G/\Gamma)$, and so
\begin{align*}
\phi(x) &= F(p(x)) \\
&= F(\tau(p(x+h_1), p(x+h_2), p(x+h_1+h_2), \dots, p(x+h_1+h_2+h_3))) \\
&= \sum_{i=1}^k R_1^{(i)}(p(x+h_1)) R_2^{(i)}(p(x+h_2)) \dots R_7^{(i)}(p(x+h_1+h_2+h_3)) + o_{k\to\infty}(1) \ .
\end{align*}
Since $|\langle f, \phi \rangle|$ is bounded away from zero, we deduce that
\[
\Bigl| \EE_{x,h_1,h_2,h_3} f(x) \overline{R_1^{(i)}(p(x+h_1)) R_2^{(i)}(p(x+h_2)) \dots R_7^{(i)}(p(x+h_1+h_2+h_3))} \Bigr |
\]
is bounded away from zero for some $i \in \{1,\dots,k\}$ (after choosing $k$ appropriately in terms on $F$ and $\delta$). But this expression is a ``Gowers inner product'' of eight functions, and by the Gowers--Cauchy--Schwarz inequality (essentially multiple applications of Cauchy--Schwarz), this quantity is bounded above by
\[
\|f\|_{U^3} \left\|R_1^{(i)} \circ p\right\|_{U^3} \dots \left\|R_7^{(i)} \circ p\right\|_{U^3}
\]
and noting that $\left\|R_j^{(i)} \circ p \right\|_{U^3} \le \left\|R_j^{(i)}\right\|_\infty$ which is bounded, we get a lower bound on $\|f\|_{U^3}$ as required.
\end{proof}
The key point is that the only properties of nilmanifolds, nilsequences etc.~that we have used are those described in (ii), (iii) and (iv) above. So we have in fact shown:
\begin{observation}
\label{obs:obstruct}
If $X$ is \emph{any} compact metric space equipped with some suitable notion of ``cubes'' as in (ii), having an abundance of cube-preserving maps $\ZZ/N\ZZ \to X$ as in (iii), and satisfying a corner constraint as in (iv), then functions of the form $F \circ p$ as in (v) obstruct Gowers uniformity on $\ZZ/N\ZZ$ in the sense of the above claim.
\end{observation}
Let us refer to such a space informally for now as a ``nil-object'' (the formal notion of a \emph{nilspace} will be introduced later). Then the above observation can be summarized as follows.
\begin{slogan}
Any ``nil-object'' is an obstruction to Gowers uniformity.
\end{slogan}
But now, the inverse theorem for the $U^3$ norm tells us that any function $F \circ p$ coming from this construction must have something to do with a genuine nilsequence. Given some technical hypotheses, one can deduce that any such nil-object $X$ must be very closely related to an actual nilmanifold $G/\Gamma$.
The conclusion of work of Szegedy \cite{S12} is that it is possible to go in the other direction. He argues that all functions $f$ with $\|f\|_{U^3}$ somewhat large correlate with something of the form $F \circ p$ where $p \colon \ZZ/N\ZZ \to X$ and $F \colon X \to \CC$ is continuous, for \emph{some} space $X$ equipped with a notion of cubes, and some cube-preserving $p$, obeying some fairly reasonable additional axioms. Moreover, this proof is essentially purely analytic in nature, making no attempt to say anything about the structure of $X$.
Assuming this, we conclude:
\begin{slogan}
The class of all ``nil-objects'' corresponds \emph{precisely} to the obstructions to Gowers uniformity.
\end{slogan}
Hence, the inverse theorem for the Gowers norms is essentially \emph{equivalent} to classifying nil-objects $X$, showing in effect that they are all -- essentially -- nilmanifolds. This structural result is the goal of \cite{CS12}.
\subsection{An outline of this paper}
The formal notion capturing properties (ii)-(iv) above and replacing the informal concept of a ``nil-object'', is what we term a \emph{nilspace}. A nilspace is a compact topological space satisfying certain ``nilspace axioms'', which are both very abstract (e.g.~they do not explicitly mention any group structure) but simultaneously strong enough for a strong structural result to hold.
The remaining tasks of this paper are as follows:
\begin{enumerate}
\item to explain the nilspace axioms formally;
\item to state a precise version of the structure theorem;
\item to outline the stages in the proof of this theorem; and
\item to prove some weaker structural results that constitute the first stage of the proof.
\end{enumerate}
The remaining parts of the proof of the structure theorem appear in the companion papers \cites{GMV2, GMV3}, and we will provide pointers to the relevant sections of these works in the outline.
Most of our discussion will be of an expository or heuristic nature. The only parts that constitute steps in the rigorous proof of the structure theorem are Section \ref{sec:elementary-v2}, and in a way the definitions in Section \ref{sec:cubespace-defns}. A reader already familiar with this whole approach and seeking only a complete proof could therefore read only these, together with the companion papers \cites{GMV2, GMV3}; needless to say we do not recommend this strategy.
Most results proved in this paper are due to Antol\'\i n Camarena and Szegedy \cite{CS12}, and the majority of the ideas we will discuss originate in \cite{HK08} or \cite{CS12}.
The primary exception is the ``relative'' version of this ``weak structure theory'' treated in Section \ref{sec:elementary-v2}.
This generalization is fairly mechanical when stated in the language of \emph{fibrations}, which are a class of maps between cubespaces satisfying certain additional hypotheses. This notion is almost equivalent to that of a \emph{fiber-surjective morphism} appearing in \cite{CS12}; indeed, the latter is only defined for maps between two nilspaces, and in that case the definitions are equivalent. Our reasons for introducing the former are twofold: the definition of a fibration makes sense in greater generality, which we do actually use; and in our view, the alternative definition makes the analogy between relative and non-relative versions more transparent.
Before embarking on (i), we recall a small amount of background about Host--Kra cube groups, which puts the abstract definitions in some context. This is done in Section \ref{sec:hk}.
The definitions themselves are then expounded in Section \ref{sec:cubespace-defns}.
With these in place, we are in a position to state the full structural result in Section \ref{sec:structural-thm}. Then, we give a very heuristic outline of the high-level stages of the proof.
This outline is expanded upon in Section \ref{sec:remaining-summary-v2}, which gives a detailed overview of each stage of the proof.
The final two sections, Section \ref{sec:weak-overview} and Section \ref{sec:elementary-v2}, are concerned with the first of these stages, which we have been calling the ``weak structure theory''. Section \ref{sec:weak-overview} deals with the ``standard'' theory as it appears in \cite{CS12}. In Section \ref{sec:elementary-v2}, we consider \emph{relative} versions of essentially all the elementary theory, which generalize the earlier results and which will be needed elsewhere in this project.
In the Appendix, we resume the discussion of Host--Kra cubes from Section \ref{sec:hk}, and make a detailed study of cubes on nilmanifolds in some example cases.
\subsection{Acknowledgments}
First and foremost we owe gratitude to Bernard Host who introduced us to the subject and to Omar Antol\'\i n Camarena and Bal\'azs Szegedy
whose groundbreaking work \cite{CS12} was a constant inspiration
for us.
We would like to thank Emmanuel Breuillard, J\'er\^ome Buzzi,
Yves de Cornulier, Sylvain Crovisier, Eli Glasner, Ben Green, Bernard Host, Micha\pol{} Rams, Bal\'azs Szegedy, Anatoly Vershik and Benjamin
Weiss for helpful discussions. We are grateful to Pablo Candela and Bryna Kra for a careful reading of a preliminary version.
We are grateful to the referee for her or his careful reading of our paper and for her or his many
helpful comments, which greatly improved the presentation of the paper.
\section{Nilmanifolds and their Host--Kra cubes}
\label{sec:hk}
The key motivating examples of nilspaces (being in some sense the only examples) are nilmanifolds $G/\Gamma$ equipped with their Host--Kra cube structures. Certainly anything we define or prove about general nilspaces should hold true for these spaces, and they provide a good source of intuition to guide the definitions and arguments in the abstract setting.
With this in mind, we will now briefly recall the relevant constructions and a few properties. A much more substantial exposition of this theory, including a number of examples explored in depth, is given in Appendix \ref{app:hk}; this also includes omitted proofs from this section. The reader unfamiliar with this area might wish to read this exposition before continuing with the bulk of the paper.
The notions and ideas presented in this section originate from \cites{HK05,HK08, GT10}.
We first recall the notion of a filtration.
\begin{definition}
Let $G$ be a topological group. By a \emph{filtration of degree $s$}\index{filtration} on $G$, we mean a sequence $G = G_0 \supseteq G_1 \supseteq \dots \supseteq G_{s+1} = \{\id\}$ of closed subgroups, with the property that $[G_i, G_j] \subseteq G_{i+j}$ for all $i,j \ge 0$. (Here $[\cdot,\cdot]$ denotes commutation, and we use the convention $G_i = \{\id\}$ for all $i \ge s+1$.)
The filtration is called {\em proper}\index{filtration!proper} if $G_0=G_1$.
\end{definition}
For most of the theory we will only consider proper filtrations, but we find it useful to permit $G_0\neq G_1$ in
the definition.
Note that if $G$ admits a proper filtration of degree $s$, then $G$ is necessarily nilpotent, with nilpotency class at most $s$.
We write $G_\bullet$ in place of $G$ if we wish to emphasize that a group is equipped with a particular filtration.
The standard example of a filtration is the \emph{lower central series}\index{lower central series} filtration on a group, given by $G_0 = G_1 = G$ and $G_{i+1} = [G, G_i]$ for each $i \ge 1$ (for a proof that this is indeed a filtration see \cite{MKS66}*{Theorem 5.3}). This is the minimal proper filtration on $G$, in the sense that every proper filtration contains it termwise.
The fundamental construction concerning filtered groups is that of the \emph{Host--Kra cube groups}. These are designed to be the appropriate analogues of parallelepipeds in abelian groups, in the setting of general filtered groups $G_\bullet$. To describe them, we will first set up some notation.
For a set $X$, we use the notation $X^{\{0,1\}^k}$ to mean the space of all functions $\{0,1\}^k \to X$. Concretely, this is just $X^{2^k}$ but the reader should always imagine the elements written at the vertices of the discrete cube $\{0,1\}^k$, as in:
\inlinetikz{\threecube{$x(000)$}{$x(001)$}{$x(010)$}{$x(011)$}{$x(100)$}{$x(101)$}{$x(110)$}{$x(111)$}{1}}
We denote by $[k]$\index[nota]{$[k]$} the set $\{1,\ldots, k\}$.
We find it convenient to identify subsets of $[k]$
with vertices of the discrete cube $\{0,1\}^k$, mapping sets to their indicator functions.
In particular, we write $\omega_1\subseteq\omega_2$ for two vertices
if $\omega_1(j)\le\omega_2(j)$ for all $j\in [k]$.
By a \emph{face}\index{face} of the discrete cube $\{0,1\}^k$ we mean a sub-cube obtained by fixing some subset of the coordinates. An \emph{upper face}\index{face!upper} is one obtained by fixing some subset of the coordinates to equal $1$; we can write this as $\{ \omega \in \{0,1\}^k \colon \omega \supseteq S \}$ for some $S \subseteq [k]$. For instance, in
\inlinetikz{
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=3]
\draw[fill=black!40] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[thin] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\draw[thin] (1,1,1) -- (0,1,1) -- (0,0,1) -- (1,0,1) -- cycle;
\draw[thin] (0,0,0) -- (1,0,0) -- (1,1,0) -- (0,1,0) -- cycle;
\draw[thin] (1,1,1) -- (1,1,0) -- (1,0,0) -- (1,0,1) -- cycle;
\draw[thin] (1,1,1) -- (0,1,1) -- (0,1,0) -- (1,1,0) -- cycle;
\draw[thin] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[line width=0.3cm,black!70] (1,1,0) -- (1,1,1);
\draw[fill=black!70,black!70] (1,1,1) circle[radius=0.05cm];
\draw[fill=black!70,black!70] (1,1,0) circle[radius=0.05cm];
\end{scope}
}
an upper face of codimension two and a face of codimension one are indicated.
\begin{definition}
\label{def:cube-group}
Let $G_\bullet$ be a filtered topological group, and let $k \ge 0$ be an integer. The $k$-th Host--Kra cube group\index{Host--Kra cube group}, denoted $\HK^k(G_\bullet)$\index[nota]{$\HK^k(G_\bullet)$}, is a sub-group of $G^{\{0,1\}^k}$ defined as follows.
For a subset $S \subseteq [k]$, let $F_S \subseteq \{0,1\}^k$\index[nota]{$F_S$} denote the upper face corresponding to $S$, i.e.~$F_S = \{ \omega \in \{0,1\}^k \colon \omega \supseteq S \}$. For any $x \in G$ and any face $F$ of $\{0,1\}^k$, let $[x]_F$\index[nota]{$[x]_F$} donote the configuration $\{0,1\}^k \to G$ given by
\[
[x]_F(\omega) = \begin{cases} x &\colon \omega \in F , \\ \id &\colon \text{otherwise} \ . \end{cases}
\]
Then $\HK^k(G_\bullet)$ is the group generated by elements of the form $[x]_{F_S}$ where $S \subseteq \{0,1\}^n$ and $x \in G_{|S|}=G_{\codim(F_S)}$.
\end{definition}
We will illustrate this in the case $k=3$. First, taking $S = \emptyset$ we are free to take any constant configuration:
\inlinetikz{\threecube{$x$}{$x$}{$x$}{$x$}{$x$}{$x$}{$x$}{$x$}{1}}
for any $x \in G$.
Taking $S$ of size $1$ allows elements that are the identity on a lower face and equal to $x$ on the corresponding upper face for some fixed $x \in G_1$, e.g.:
\inlinetikz{\threecube{$\id$}{$\id$}{$\id$}{$\id$}{$x$}{$x$}{$x$}{$x$}{1}}
Similarly, for $|S| = 2$ we get a configuration equal to $x$ on some upper face of codimension $2$, and the identity elsewhere, where now $x \in G_2$ is fixed, e.g.:
\inlinetikz{\threecube{$\id$}{$\id$}{$\id$}{$\id$}{$\id$}{$\id$}{$x$}{$x$}{1}}
Finally, we can put any element of $G_3$ on the topmost vertex.
\inlinetikz{\threecube{$\id$}{$\id$}{$\id$}{$\id$}{$\id$}{$\id$}{$\id$}{$x$}{1}}
\begin{remark}
At first glance, these definitions are inherently asymmetric in the sense of treating upper faces preferentially to other faces. However, it is easy to see that $[x]_F$ is in $\HK^k(G_\bullet)$ for any face $F$ of codimension $r$, provided $x \in G_r$.
For instance, if $S = \{i\}$ for some $i \in [k]$, and let $F$ denote the lower face $\{\omega_i = 0\}$. Then
\[
[x]_F = [x^{-1}]_{F_S} [x]_{F_{\emptyset}}
\]
and this is clearly in $\HK^k(G_\bullet)$. In general one can argue by induction on the co-dimension.
\end{remark}
Ultimately, though, we are not interested in nilpotent groups but in nilmanifolds $G/\Gamma$, their compact homogeneous spaces. The notion of Host--Kra cubes over $G$ goes over directly to a notion on $G/\Gamma$.
\begin{definition}
\label{defn:nilmanifold}
Let $G_\bullet$ be a degree $s$ filtered Lie group. Suppose for simplicity that $G$ is connected.\footnote{In \cites{GMV2,GMV3} there are good reasons to relax this assumption, but they do not apply for now.} Also, let $\Gamma$ be a discrete and co-compact subgroup of $G$ (the latter meaning the quotient $G/\Gamma$ is compact). Under these hypotheses, the quotient space $G/\Gamma$ (which need not be a group) is termed a \emph{nilmanifold}.\index{nilmanifold}
Suppose furthermore that $\Gamma \cap G_i$ is discrete and co-compact in $G_i$ for each $i \ge 0$. If this property holds, we say that $\Gamma$ is {\em compatible}\index{compatible (with filtration)} with the filtration. Then for each $k \ge 0$, we define the Host--Kra cubes\index{Host--Kra cubes} $\HK^k(G_\bullet)/\Gamma$ to be the subset of $(G/\Gamma)^{\{0,1\}^k}$ given by the image of $\HK^k(G_\bullet)$ under the quotient map
\begin{align*}
\pi \colon G^{\{0,1\}^k} &\to (G/\Gamma)^{\{0,1\}^k} \\
(g_\omega)_{\omega \in \{0,1\}^k} &\mapsto (g_\omega \Gamma)_{\omega \in \{0,1\}^k} \ .
\end{align*}
Note that we abuse notation to let $\HK^k(G_\bullet)/\Gamma$ denote the \emph{pointwise} quotient by $\Gamma$ rather than a conventional quotient of groups: this notation is not meant to identify $\Gamma$ with a subgroup of $\HK^k(G_\bullet)$. Equivalently, this quotient may be identified with $\HK^k(G_\bullet) / \left(\Gamma^{\{0,1\}^k}\cap\HK^k(G_\bullet) \right)$.
\end{definition}
The topological conditions are chosen to allow the following conclusion.
\begin{proposition}
For $G_\bullet$ and $\Gamma$ as in the above definition, the space $\HK^k(G_\bullet)/\Gamma$ is a compact subset of $(G/\Gamma)^{\{0,1\}^k}$ for all $k \ge 0$.
\end{proposition}
See \cite{GT10}*{Lemma E.10} for a proof.
Finally, we summarize some properties of Host--Kra cubes on nilmanifolds that will be of significance in the next section. The proofs of these properties appear in Appendix \ref{app:hk}, or follow easily from those results.
\begin{proposition}
\label{prop:nilmanifold-props}
Let $G_\bullet$ and $\Gamma \subseteq G$ be as in Definition \ref{defn:nilmanifold}, and suppose in particular $G_\bullet$ has degree $s$. Then the following hold.
\begin{enumerate}[label=(\roman*)]
\item (Symmetries) The space $\HK^k(G_\bullet)/\Gamma$ is invariant under
a permutation of the $k$ coordinate axes, reflecting in a coordinate axis, or any combination of these.
\item (Compatibility) If $c \in \HK^k(G_\bullet)/\Gamma$ then any face of $c$ of dimension $\ell$ (or more generally any ``subcube'', allowing some diagonal slicing) is an element of $\HK^\ell(G_\bullet)/\Gamma$.
\item (Corner constraint) Suppose $c, c' \in \HK^{s+1}(G_\bullet)/\Gamma$ and $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{1}$. Then $c = c'$.
\item (Corner completion) Suppose $\lambda \colon \{0,1\}^k \setminus \{\vec{1}\} \to G/\Gamma$ is a configuration such that for every lower face of $\{0,1\}^k$ of codimension $1$, the restriction of $\lambda$ to that face is in $\HK^{k-1}(G_\bullet)/\Gamma$. Then $\lambda$ can be extended to an element of $\HK^k(G_\bullet)/\Gamma$.
\end{enumerate}
\end{proposition}
Here, and throughout, $\vec{1}$\index[nota]{$\vec{1}$} denotes the element $(1,1,\dots,1) \in \{0,1\}^k$.
\begin{proof}
Parts (i) and (ii) follow directly from the definition. For (iii) and (iv), see Proposition \ref{prop:nilmanifold-uniqueness} and Proposition \ref{prop:nilmanifold-completion}.
\end{proof}
\section{Cubespaces and nilspaces}
\label{sec:cubespace-defns}
\subsection{Definitions}
We now give the formal definition and axioms of \emph{nilspaces}, and related notions. In fact, the notion of a nilspace captures several distinct hypotheses of differing strength. We will outline these individually, starting with the weakest, before amalgamating them into a final definition. We follow \cites{HK08, CS12} closely, although our terminology differs.
The very weakest structure we will wish to consider is termed a \emph{cubespace}. Informally, this is just a topological space equipped with some notion of when $2^k$ points form a cube,\footnote{Following \cite{CS12}, we use the term ``cube'' throughout to refer to these distinguished collections of $2^k$ points. These objects in fact seldom resemble geometric cubes, and the term ``parallelepiped'' used by Host and Kra \cite{HK08} is more accurate. However, ``cube'' has a significant advantage in brevity.} and satisfying several fairly basic conditions.
To define these conditions, we need to define some nomenclature for certain maps on the discrete cube $\{0,1\}^k$.
\begin{definition}
A map $\rho \colon \{0,1\}^k \to \{0,1\}^\ell$ is called a \emph{morphism of discrete cubes}\index{morphism of discrete cubes} if it is the restriction to $\{0,1\}^k$ of an affine-linear map $\ZZ^k \to \ZZ^\ell$.
Equivalently, this holds if and only if $\rho$ has the form
\[
\rho(x_1, \dots, x_k) = (\sigma_1(x_1, \dots, x_k), \dots, \sigma_\ell(x_1, \dots, x_k))
\]
where each function $\sigma_i$ is one of:
\begin{itemize}
\item identically $0$;
\item identically $1$;
\item equal to $x_j$ for some $j$;
\item equal to $1 - x_j$ for some $j$.
\end{itemize}
\end{definition}
It is straightforward to see that these definitions are indeed equivalent.
Informally, these morphisms of the discrete cube correspond to the fairly natural operations:
\begin{itemize}
\item permute the coordinates of $\{0,1\}^k$;
\item reflect in any coordinate axis;
\item embed $\{0,1\}^k$ as a ``slice'' in $\{0,1\}^\ell$ for $\ell > k$;
\item ``project'' $\{0,1\}^k$ onto $\{0,1\}^\ell$ for $\ell < k$ by deleting a coordinate;
\end{itemize}
and functions obtained from these by composition.
\begin{definition}
A \emph{cubespace}\index{cubespace} is a metric space $X$, together with closed sets $C^k \subseteq X^{\{0,1\}^k}$ of \emph{$k$-cubes} for each $k \ge 0$, satisfying the following condition. Suppose $\rho \colon \{0,1\}^k \to \{0,1\}^\ell$ is a morphism of discrete cubes and $c \colon \{0,1\}^\ell \to X$ is in $C^\ell$. Then $c \circ \rho \colon \{0,1\}^k \to X$ is in $C^k$.
Furthermore, it is always assumed that $C^0 = X$.
\end{definition}
We will typically abuse notation to allow $X$ to refer either to the underlying topological space, or to the full cubespace structure $(X, C^k)$. We write $C^k(X)$\index[nota]{$C^k(X)$} in place of $C^k$ whenever there is ambiguity about which cubespace we are referring to.
Let us unpack what this means in terms of the elementary operations above.
\begin{itemize}
\item If $c \in X^{\{0,1\}^k}$ is a $k$-cube, then permuting the $k$ coordinates gives another $k$-cube; e.g.:
\inlinetikz{
\begin{scope}[shift={(-2,0)}]
\singlesquare{$x_{00}$}{$x_{01}$}{$x_{10}$}{$x_{11}$}
\draw[dashed] (-0.1, -0.1) -- (1.1, 1.1);
\end{scope}
\draw[->] (0,0.5) -- (1,0.5);
\begin{scope}[shift={(2,0)}]
\singlesquare{$x_{00}$}{$x_{10}$}{$x_{01}$}{$x_{11}$}
\end{scope}
}
\item Similarly, reflecting in any coordinate axis gives another $k$-cube; e.g.:
\inlinetikz{
\begin{scope}[shift={(-2,0)}]
\singlesquare{$x_{00}$}{$x_{01}$}{$x_{10}$}{$x_{11}$}
\draw[dashed] (0.5, 1.2) -- (0.5, -0.2);
\end{scope}
\draw[->] (0,0.5) -- (1,0.5);
\begin{scope}[shift={(2,0)}]
\singlesquare{$x_{01}$}{$x_{00}$}{$x_{11}$}{$x_{10}$}
\end{scope}
}
\item If $c \in X^{\{0,1\}^k}$ is a $k$-cube, then restricting to an $\ell$-dimensional ``slice'' gives an $\ell$-cube:
\inlinetikz{
\begin{scope}[shift={(-5,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=3]
\draw (0,0,0) node[below left] {$x_{000}$} -- (0,0,1) node[below right] {$x_{010}$} -- (0,1,1) node[above] {$x_{110}$} -- (0,1,0) node[above left] {$x_{100}$} -- cycle;
\draw (0,0,0) -- (1,0,0) node[below right] {$x_{001}$} -- (1,0,1) node[right] {$x_{011}$} -- (0,0,1) -- cycle;
\draw (1,1,1) node[above right] {$x_{111}$} -- (0,1,1) -- (0,0,1) -- (1,0,1) -- cycle;
\draw (1,1,1) -- (1,1,0) -- (1,0,0) -- (1,0,1) -- cycle;
\draw (1,1,1) -- (0,1,1) -- (0,1,0) -- (1,1,0) -- cycle;
\draw[fill=black!50, opacity=.8] (0,1,0) -- (0,1,1) -- (1,0,1) -- (1,0,0) -- cycle;
\draw (0,0,0) -- (1,0,0) -- (1,1,0) node [above left] {$x_{101}$} -- (0,1,0) -- cycle;
\end{scope}
\end{scope}
\draw[->] (0,2) -- (1,2);
\begin{scope}[shift={(2,1.5)}]
\singlesquare{$x_{100}$}{$x_{001}$}{$x_{110}$}{$x_{011}$}
\end{scope}
}
\item If $c \in X^{\{0,1\}^k}$ is a $k$-cube, the configuration obtained by placing two copies of $c$ adjacent to each other is a $(k+1)$-cube, and so on:
\inlinetikz{
\begin{scope}[shift={(-2,0.7)}]
\singlesquare{$x_{00}$}{$x_{01}$}{$x_{10}$}{$x_{11}$}
\end{scope}
\draw[->] (0,1.2) -- (1,1.2);
\begin{scope}[shift={(2,0)}]
\threecube{$x_{00}$}{$x_{01}$}{$x_{10}$}{$x_{11}$}{$x_{00}$}{$x_{01}$}{$x_{10}$}{$x_{11}$}{1}
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[ultra thick] (0,0,0) -- (0,1,0);
\draw[ultra thick] (1,0,0) -- (1,1,0);
\draw[ultra thick] (0,0,1) -- (0,1,1);
\draw[ultra thick] (1,0,1) -- (1,1,1);
\end{scope}
\end{scope}
}
\end{itemize}
\begin{remark}
\label{rem:constant-cube}
Note that applying this last point repeatedly, we deduce that any constant configuration (i.e.~$c(\omega) = x$ for all $\omega$) is automatically a cube.
\end{remark}
It is intuitively fairly reasonable that any well-behaved notion of ``cube'' analogous to parallelepipeds or the Host--Kra construction on nilmanifolds should at the very least obey these properties.
We now consider the ``corner constraint'' discussed previously. Recall we said informally that, for some specific dimension of cube, all but one of the vertices should determine the last one.
\begin{definition}
We say a cubespace $X$ has \emph{$k$-uniqueness}\index{$k$-uniqueness} if the following holds: whenever $c, c' \in C^k(X)$ and $c(\omega) = c'(\omega)$ for all $\omega \in \{0,1\}^k \setminus \{\vec{1}\}$ then $c = c'$.
\end{definition}
The ``dual'' of this property is equally important but harder to motivate: if all but one vertex of a $k$-cube is specified in a consistent way, then there is \emph{at least one} way to complete the missing vertex to give a cube.
\begin{definition}
\label{defn:k-completion}
We say a cubespace $X$ has \emph{$k$-completion}\index{$k$-completion} if the following holds. Suppose $\lambda \colon \{0,1\}^k \setminus \{\vec{1}\} \to X$ has the property that every ``lower face'' is a $(k-1)$-cube, i.e.~for each $1 \le i \le k$ the map
\begin{align*}
\{0,1\}^{k-1} &\to X \\
(\omega_1, \dots, \omega_{k-1}) &\mapsto \lambda(\omega_1, \dots, \omega_{i-1}, 0, \omega_i, \dots, \omega_{k-1})
\end{align*}
is in $C^{k-1}(X)$. Then there exists some $x \in X$ such that
\begin{align*}
c \colon \{0,1\}^k &\to X \\
\omega &\mapsto \begin{cases} \lambda(\omega) &\colon \omega \ne \vec{1} \\ x &\colon \omega = \vec{1} \end{cases}
\end{align*}
is in $C^k(X)$.
\end{definition}
\inlinetikz{
\begin{scope}{scale=0.67}
\begin{scope}[shift={(-5,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=3]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,1,0) -- (0,1,0) -- cycle;
\end{scope}
\end{scope}
\draw[->] (0, 2) -- (1, 2);
\begin{scope}[shift={(2,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=3]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,1,0) -- (0,1,0) -- cycle;
\draw[dashed] (1,0,1) -- (1,1,1) -- (0,1,1);
\draw[dashed] (1,1,1) -- (1,1,0);
\draw[fill=black] (1,1,1) circle[radius=0.02cm];
\end{scope}
\end{scope}
\end{scope}
}
We call a configuration of the form $\lambda$ (with the same properties) a \emph{$k$-corner}.\index{$k$-corner}
While uniqueness will typically be specified for one particular value of $k$, a well-behaved space will have $k$-completion for \emph{all} $k \ge 0$.
\begin{definition}
We say a cubespace $X$ is a \emph{nilspace of degree $s$} if $s \ge 0$ is the smallest nonnegative integer such that $X$ has $(s+1)$-uniqueness, and if $X$ has $k$-completion for all $k$. We say it is simply a \emph{nilspace}\index{nilspace} if it is a nilspace of degree $s$ for some $s$.
\end{definition}
\begin{remark}
One way of motivating the $k$-completion hypothesis is that it guarantees that $X$ is not missing any points, in the following sense.
Suppose $X$ is a compact nilspace of degree $s$, and $S \subseteq X$ an arbitrary closed subset. We can make $S$ into a cubespace by taking $C^k(S) = C^k(X) \cap S^{\{0,1\}^k}$, i.e.~just taking those cubes of $X$ whose vertices lie in $S$.
We call this cubespace the subcubespace of $X$ {\em induced}\index{induced subcubespace} by $S$.
With this cubespace structure, it is easy to verify that $(S, C^k(S))$ is still a compact cubespace with $(s+1)$-uniqueness; i.e., it satisfies all the hypotheses of being a (compact) nilspace of degree $s$ except for the $k$-completion one. However, $(S, C^k(S))$ will typically \emph{not} have $k$-completion for any $k > 1$, unless $S$ takes a very particular form.
So, if we were to omit the $k$-completion axiom from the definition of a nilspace, it would become fairly hopeless to ask for a rigid classification theorem for nilspaces, since there are a huge variety of examples of this type. By contrast, if we do impose $k$-completion, we will see that this rules out such examples in general and imposes significantly more rigidity on the class of nilspaces.
\end{remark}
We will be interested not only in cubespaces and nilspaces in isolation, but also in maps between them. Informally, a map $X \to Y$ preserves the relevant structures if it takes any cube of $X$ to a cube of $Y$.
\begin{definition}
Suppose $X$, $Y$ are cubespaces with underlying spaces $X$, $Y$ respectively. Let $\phi \colon X \to Y$ be a continuous map. We say $\phi$ is a \emph{cubespace morphism}\index{cubespace!morphism} or just \emph{morphism} if $\phi(C^k(X)) \subseteq C^k(Y)$ for all $k \ge 0$.
\end{definition}
We will need one further technical definition.
\begin{definition}
A cubespace $X$ is called \emph{ergodic}\index{cubespace!ergodic} if $C^1(X) = X^2$, i.e.~if every pair $(x,x')$ is a $1$-cube. Furthermore we say it is $k$-ergodic if $C^k(X) = X^{\{0,1\}^k}$, i.e.~if every configuration is a $k$-cube.
\end{definition}
Note that our terminology differs slightly from that in \cite{CS12}, in that we do not insist that a nilspace be ergodic: this is built into the definition of a nilspace in that work. As a consequence, the phrase ``ergodic nilspace'' will occur frequently in our statements.
It can be shown (although we will not need to do so) that a suitably nice non-ergodic space -- in particular, a non-ergodic nilspace -- decomposes as a disjoint union of ergodic components, whose cubespace structures essentially do not interact at all. Hence one loses almost no generality by working in the ergodic setting.
Also, we remark that \cite{CS12} has the separate notions of ``abstract'' nilspaces, which have no topology, and ``compact'' ones, which have a topology but are always assumed to be compact. By contrast, we assume all nilspaces have a topology (though it could be the discrete topology), but do not insist they be compact. This discrepancy has no particular significance, except that it allows us to consider e.g.~nilpotent Lie groups as nilspaces with their natural topologies. Again, the phrase ``compact, (ergodic) nilspace'' will occur often in what follows.
\subsection{Host--Kra nilmanifolds are nilspaces}
Having motivated these nilspace axioms (and other definitions) in terms of Host--Kra cubespaces and nilmanifolds, it is reasonable to want to check that the latter are actually instances of the former.
We state some facts of this nature now. Most aspects of the proofs are straightforward; some, surprisingly, are much less so, and these are deferred to Appendix \ref{app:hk} which expounds the theory of Host--Kra cubespaces more fully than we have done so far.
\begin{proposition}
\label{prop:hk-cubespace-properties}
Let $G/\Gamma$ be a (filtered) nilmanifold of degree $s$, in the sense of the previous section, equipped with its Host--Kra cubes. Then it is a nilspace of degree $s$.
If the filtration $G_\bullet$ is \emph{proper} (i.e.~$G_0 = G_1$) then $G/\Gamma$ is ergodic.
If $G'/\Gamma'$ is another nilmanifold, and $\psi \colon G \to G'$ a group homomorphism such that $\psi(G_i) \subseteq G'_i$ for all $i$ and $\psi(\Gamma) \subseteq \Gamma'$, then the induced map $\phi \colon G/\Gamma \to G'/\Gamma'$ is a cubespace morphism.
Similarly, for any $x \in G_0$, $g \in G_1$, the map $n \mapsto g^n x\, \Gamma$ is a cubespace morphism $\ZZ \to G/\Gamma$.
\end{proposition}
Here, as throughout these papers, $\ZZ$ is considered as a filtered group with filtration $\ZZ_0 = \ZZ_1 \supseteq \{0\}$, and as a cubespace by setting $C^k(\ZZ) = \HK^k(\ZZ_\bullet)$.
\begin{proof}
The fact that $G/\Gamma$ is a nilspace is essentially a restatement of Proposition \ref{prop:nilmanifold-props}.
The ergodicity statement is immediate from the definition of $\HK^1(G_\bullet)$.
The fact that a group homomorphism $\psi \colon G \to G'$ taking $G_i$ into $G_i'$ maps $\HK^k(G_\bullet)$ into $\HK^k(G_\bullet')$ is straightforward by considering the images of the generators. Recovering the corresponding statement for the quotients $G/\Gamma$, $G'/\Gamma'$ is routine.
Finally, we note that for any $x \in G_0$, $g \in G_1$, $a \in \ZZ$ and $h \in \ZZ^k$, the configuration
\begin{align*}
\{0,1\}^k &\to G \\
\omega &\mapsto g^{a + \omega \cdot h} x
\end{align*}
is in $\HK^k(G_\bullet)$, since it is a product of $(k+1)$ generators:
\[
[g^{h_1}]_{F_{\{1\}}}\cdots [g^{h_k}]_{F_{\{k\}}}\cdot[g^ax]_{F_\varnothing},
\]
and the last part follows.
\end{proof}
We have been light on detail in this proof: partly because this result is not logically necessary in what follows. However, working through the details is an excellent exercise for the reader unfamiliar with either these definitions or the machinery of Host--Kra cubes (or both).
\subsection{High-dimensional cubes}
One slightly unsatisfactory feature of our definitions is that the data of a cubespace involves spaces $C^k(X)$ for infinitely many $k \ge 0$. We are used to considering cubes or parallelepipeds of only bounded dimension in any given problem (e.g.~of dimension $3$ when considering the $U^3$ norm).
In fact, under good hypotheses we can see that for sufficiently large $k$ the data of $C^k(X)$ contains no new information.
As one further reassuring sanity-check, and to see these definitions in action, we state this now.
\begin{proposition}
\label{prop:high-cubes-boring}
Let $s \ge 1$ be an integer, $X$ a nilspace of degree $s$, and $k \ge s+1$. Then a configuration $c \colon \{0,1\}^k \to X$ is in $C^k(X)$, if and only if every face of $c$ of dimension $(s+1)$ is in $C^{s+1}(X)$.
\end{proposition}
\begin{proof}
The ``only if'' direction is direct from the cubespace axioms. For the ``if'' direction, we argue by induction on $k$, the base case $k=s+1$ being trivial.
Given such a $c \colon \{0,1\}^k \to X$, consider the restriction $c|_{\{0,1\}^k \setminus \{\vec{1}\}}$. By inductive hypothesis it is a $k$-corner. Hence, there is a cube $c' \in C^k(X)$ such that $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{1}$. But now, considering any upper face $c'$ of dimension $(s+1)$ and the corresponding one of $c$, and invoking $(s+1)$-uniqueness, we see that $c(\vec{1}) = c'(\vec{1})$ and hence $c$ is a $k$-cube.
\end{proof}
\section{The structure theorem}
\label{sec:structural-thm}
\subsection{Statement}
We are now in a position to state the main structural result.
This result is essentially due to Antol\'\i n Camarena and Szegedy \cite{CS12}*{Theorems 4 and 7};
however, our formulation is somewhat stronger.
We point out the differences below.
\begin{theorem}
\label{main-structural-v1}
Let $X = (X, C^k)$ be a compact ergodic nilspace of degree $s$. Further suppose that the spaces $C^k(X)$ are connected for each $k \ge 0$; so in particular $X = C^0(X)$ is connected.
Then $X$ is an ``inverse limit of nilmanifolds'' in the following sense. There exists:
\begin{itemize}
\item a sequence $G^{(n)}$ of connected Lie groups equipped with filtrations $G^{(n)}_\bullet$ of degree at most $s$ (with $G^{(n)}_i$ also connected for each $i$);
\item discrete co-compact subgroups $\Gamma^{(n)}$ of $G^{(n)}$ such that $\Gamma^{(n)} \cap G^{(n)}_i$ is discrete and co-compact in $G^{(n)}_i$ for each $i$; and
\item surjective group homomorphisms $\phi_{n,m} \colon G^{(n)} \to G^{(m)}$ for each $n \ge m$, such that $\phi_{n,m}\left(G^{(n)}_i\right) = G^{(m)}_i$ for each $i \ge 0$ and also $\phi_{n,m}\left(\Gamma^{(n)}\right) \subseteq \Gamma^{(m)}$, so that we get a map of nilmanifolds $G^{(n)}/\Gamma^{(n)} \to G^{(m)}/\Gamma^{(m)}$ that sends cubes (surjectively) to cubes;
\end{itemize}
such that
\begin{itemize}
\item $X = \varprojlim \left(G^{(n)}/\Gamma^{(n)}\right)$ as a topological space, and
\item the cubes $C^k(X)$ coincide with the inverse limit of Host--Kra cubes; i.e., writing $\pi_n \colon X \to \left(G^{(n)}/\Gamma^{(n)}\right)$ for the projection map arising from the above inverse limit, we have
\[
C^k(X) = \bigcap_n \pi_n^{-1} \left(C^k\left(G^{(n)} / \Gamma^{(n)}\right)\right) \ .
\]
\end{itemize}
\end{theorem}
\begin{definition}
We call a cubespace satisfying the hypothesis that $C^k(X)$ is connected for all $k$ \emph{strongly connected}\index{cubespace!strongly connected}.
\end{definition}
\begin{remark}
\label{rem:inverse-limit}
The last two conditions, together with the fact that $\phi_{n,m}$ are cubespace morphisms between nilmanifolds (see Proposition \ref{prop:hk-cubespace-properties}), say in some sense that $X$ is an ``inverse limit of nilmanifolds, in the category of cubespaces''.
This is roughly the formulation one obtains from \cite{CS12}, combining \cite{CS12}*{Theorems 4 and 7}. (In fact the results from \cite{CS12} give slightly more detail than that, specifically that the maps $\phi_{n,m}$ are ``fiber-surjective'', or in our terminology ``fibrations''; see Definition \ref{def:fibration} and Remark \ref{rem:fiber-surjective}. This same condition comes out of \cite{GMV3}*{Theorem 1.26}, which is equivalent to \cite{CS12}*{Theorem 4}.)
However, the conclusion of Theorem \ref{main-structural-v1} is stronger than this, in that it gives even more information about the maps $\phi_{n,m}$: namely, that they come from group homomorphisms $G^{(n)} \to G^{(m)}$, and moreover that these group homomorphisms are surjective on the filtrations in the obvious sense. Neither of these properties holds in general for a cubespace morphism between nilmanifolds.
\end{remark}
\begin{remark}\label{rem:connectivity}
Without some similar connectivity hypothesis, no complete structure theorem is currently available in general.
Using Theorem \ref{thm:invlim} (see below) or equivalently \cite{CS12}*{Theorem 4}, one can get some partial information by identifying a compact nilspace with an inverse limit of ``finite rank'' nilspaces, which are in particular e.g.~topological manifolds of finite dimension. However, the structure of such spaces remains unclear: they need not necessarily be nilmanifolds. Some structural results in the special case of finite nilspaces appear in \cite{szegedy2010structure}.
The analogous result to Theorem \ref{main-structural-v1} implied by \cite{CS12} has this strong connectivity hypothesis replaced by an assumption on the ``structure groups'' of $X$ (see below). That assumption easily implies strong connectivity (see \cite{GMV2}*{Proposition 2.4}),
but for the converse, we need the full force of our structure theorem.
\end{remark}
For clarity, we summarize the dictionary between statements in these papers and those in \cite{CS12} in Section \ref{sec:comparison} below, once we have introduced the relevant concepts.
\begin{remark}
It is reasonably straightforward to argue that any cubespace $X$ that is an inverse limit of nilmanifolds, subject to some further technical hypotheses on this inverse limit, satisfies the hypotheses of Theorem \ref{main-structural-v1}. Hence in some sense this result is best possible.
One can certainly not remove the inverse limit, as can be seen by considering any compact connected abelian group that is not a Lie group: such an object is a nilspace of degree $1$ but is not a degree $1$ nilmanifold, as the latter are all just tori.
It is also possible to find an example of a nilspace satisfying the hypotheses for $s=2$, which is not a homogeneous space of any group. Such an example is due to Rudolph \cite{R95}. It seems likely that a sharper description of such spaces is possible, but we will not pursue such issues here.
\end{remark}
\subsection{The case $s=1$}\label{sc:1step}
To provide some motivation for this structural result, we briefly comment on the case $s=1$. In this setting, Theorem \ref{main-structural-v1} is much easier.
\begin{proposition}
A compact ergodic nilspace $X$ of degree $1$ naturally has the structure of a compact abelian group.
\end{proposition}
\begin{proof}[Proof sketch]
We give only a partial account of the proof, since these ideas are covered in detail in Section \ref{sec:structure-groups}.
Essentially our task is to use the data we have -- namely, unique corner-completion of $2$-cubes -- to recover a group operation on $X$.
Fix any element $e \in X$, which will be our identity element. Given $x, y \in X$ consider the $2$-corner:
\inlinetikz{
\draw[thin,black] (0,0) node[below left] {$e$} -- (1, 0) node[below right] {$x$};
\draw[thin,black] (0,0) -- (0, 1) node[above left] {$y$};
\draw[dashed] (1,0) -- (1,1) node[above right] {$\ast$};
\draw[dashed] (0,1) -- (1,1);
}
and define $x+y$ to be the element obtained by completing it, i.e.~$x+y=\ast$. We can define inverses by a similar process, and define the identity to be $e$.
It now suffices to check that this defines a topological abelian group operation. Many of the corresponding properties are immediate from the nilspace axioms. The hardest part is checking associativity, which requires the completion property on $3$-dimensional cubes; we omit this part.
\end{proof}
The remaining content of Theorem \ref{main-structural-v1} in this case is that any connected compact abelian group is an inverse limit of tori, which is a classical result.
\subsection{Steps in the proof}
The proof of Theorem \ref{main-structural-v1} splits into three parts.
The first is a weaker, fairly elementary structure theorem. Roughly speaking, this says that a nilspace of degree $s$ has the structure of a tower of extensions $X \to X_{s-1} \to \dots \to X_0 = \{\ast\}$ where each of the fibers is a compact abelian group. We refer to this as the \emph{weak structure theorem}.
The second says that, under some strong technical assumptions, a compact ergodic strongly connected nilspace of degree $s$ actually \emph{is} a nilmanifold $G/\Gamma$ with its usual Host--Kra cubes (and a degree $s$ filtration). These technical assumptions are phrased in terms of the compact abelian groups appearing in the weak structure theorem; specifically, that they are (abelian) Lie groups, i.e.~$(\RR/\ZZ)^d \times K$ for some $d \ge 0$ and $K$ finite. This is the step where a non-abelian group operation is recovered from just the cubespace structure, the key step being to consider the automorphism group of the cubespace $X$, and to show that it is large enough in a certain sense.
The final step requires us to show that a cubespace satisfying the hypotheses of Theorem \ref{main-structural-v1} can be written as an inverse limit of ones which satisfy the strong technical assumptions alluded to in part two. Given the previous discussion, this is very closely related to the statement that a connected compact abelian group is an inverse limit of tori, and this fact is a key ingredient in the proof. However, bolting everything together is one of the more technically difficult aspects of the whole argument.
Only the first part of this plan will be proven fully in this paper, although we will give rather fuller outlines of the latter two below. These are proven fully in two further papers by the authors, the second part in \cite{GMV2} and the third in \cite{GMV3}.
\section{A more detailed summary of the argument}
\label{sec:remaining-summary-v2}
We will now expand on the outline given above of each of the three stages in the proof of Theorem \ref{main-structural-v1}.
\subsection{An outline of the weak structure theory}\label{sc:outline-weak}
The simplest examples of nilspaces are compact abelian groups $A$ together with their Host--Kra cubes (with the degree $1$ filtration $A = A_0 = A_1 \supseteq A_2 = \{0\}$). The $k$-cubes of these spaces are precisely the $k$-dimensional parallelepipeds on $A$ in the usual sense:
\begin{align*}
\{0,1\}^k &\to A \\
\omega &\mapsto x_0 + \sum_{i=1}^k \omega_i x_i
\end{align*}
where $x_0, \dots, x_k \in A$.
The simplest higher-degree examples come from the Host--Kra construction on $A$ (still a compact abelian group) given the degree $s$ filtration
\[
A = A_0 = A_1 = \dots = A_s \supseteq A_{s+1} = \{0\}
\]
where cubes are given by $C^k(A) = \HK^k(A_\bullet)$. We denote this cubespace by $\cD_s(A)$\index[nota]{$\cD_s(A)$}. It is easy to check (or to deduce from general considerations) that $\cD_s(A)$ is a compact, ergodic (and in fact $s$-ergodic) nilspace of degree $s$.
In fact there are several equivalent ways to define $\cD_s(A)$. We briefly note some equivalent characterizations.
\begin{proposition}\label{pr:Ds}
If $c \colon \{0,1\}^{s+1} \to A$ is a configuration, then $c \in C^{s+1}(\cD_s(A))$ if and only if
\[
\sum_{\omega \in \{0,1\}^{s+1}} (-1)^{|\omega|} c(\omega) = 0 \ .
\]
Moreover, for any $k \ge 0$ and $c \colon \{0,1\}^k \to A$, we have $c \in C^k(\cD_s(A))$ if and only if every face of $c$ of dimension $(s+1)$ is a cube; or, if and only if $c \circ \eta \in C^{s+1}(\cD_s(A))$ for every morphism of discrete cubes $\eta \colon \{0,1\}^{s+1} \to \{0,1\}^k$.
\end{proposition}
\begin{proof}
The first claim is proved in \ref{ex:Ds}.
It is clear from the cubespace axioms that if $c \in C^k(\cD_s(A))$ and $\eta \colon \{0,1\}^{s+1} \to \{0,1\}^k$ is a morphism of discrete cubes then $c \circ \eta \in C^{s+1}(\cD_s(A))$, and this includes the case of faces of $c$.
Conversely, by Proposition \ref{prop:nilmanifold-props}, we know that $\cD_s(A)$ is a nilspace of degree $s$. So if $c \colon \{0,1\}^k \to A$ and every face of $c$ of dimension $(s+1)$ is a cube, then $c$ is a cube by Proposition \ref{prop:high-cubes-boring}.
\end{proof}
The weak structure theorem states that, while these are certainly not the only examples of nilspaces, any (compact, ergodic) nilspace $X$ of degree $s$ can be \emph{built up} from spaces of the form $\cD_k(A_k)$ for some compact abelian groups $A_k$ for $1 \le k \le s$.
Indeed, $X$ can be realized as a tower of extensions
\[
X\to \pi_{s-1}(X)\to\pi_{s-2}(X)\to\cdots \to \pi_0(X)=\{*\}.
\]
such that $\pi_{i}(X)$ is a nilspace of degree $i$ and the fibers of the map $\pi_{k}(X)\to \pi_{k-1}(X)$
admit a free transitive action of a compact abelian group $A_{k}$ through which the cubespace structure on the
fiber can be identified by $\cD_k(A_k)$.
Before we can formulate these results more precisely we need to discuss how we can put a cubespace structure on
a quotient of a cubespace.
\begin{definition}\label{df:quotient}
Let $(X, C^k(X))$ be a compact cubespace, and suppose $\sim$ is a closed equivalence relation on $X$. The quotient space $X / \sim$ (with the quotient topology) is considered to be a cubespace\index{cubespace!quotient}, by setting $C^k(X/\sim)$ to be $C^k(X) / \sim$, i.e.~the image of $C^k(X)$ under pointwise application of $\sim$.
In other words this means that a configuration in $X/\sim$ is a cube if and only if there is some representative of it that is a cube in $X$.
\end{definition}
It is straightforward from the definition that this indeed defines a cubespace structure.
In the next definition, we specify the equivalence relation that gives rise to the quotients
$\pi_i(X)$.
This may be compared with \cite{CS12}*{Definition 2.3}.
\begin{definition}\label{df:canonical-rel}
For any $k \ge 0$, we define $\sim_k$\index[nota]{$\sim_k$} on a (compact, ergodic) cubespace $X$ with $n$-completion for all $n$, by $x \sim_k y$ if and only if there exist $c, c' \in C^{k+1}(X)$ such that $c(\vec{1}) = x$, $c'(\vec{1}) = y$ and $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{1}$.
We call $\sim_k$ the $k$-th canonical equivalence relation.\index{canonical equivalence relation}
We denote by $\pi_k$\index[nota]{$\pi_k$} the quotient map $X\to X/\sim_k$ and call it the $k$-th canonical projection\index{canonical projection}.
\end{definition}
If $\sim$ is an equivalence relation such that the factor $X/\sim$ has $(s+1)$-uniqueness, then it must contain
$\sim_s$, which is the motivation for the definition.
However, it requires proof that $\sim_s$ is indeed an equivalence relation and the quotient is indeed a nilspace.
We will discuss these facts and further properties of $\sim_s$ in Section \ref{sec:canonical-factors}.
We are now ready to state the weak structure theorem, which is proved in Sections \ref{sec:weak-overview} and \ref{sec:elementary-v2}.
See also \cite{CS12}*{Theorem 1}, and also \cite{HK08}*{Section 5} for a related discussion.
\begin{theorem}[Weak Structure Theorem]
\label{basic-cs-structure-theorem}
Let $X$ be a compact ergodic nilspace of degree $s$ for some $s \ge 1$.
Then there exists a compact abelian group $A=A_s(X)$, the ``$s$-th structure group'' of $X$,
acting continuously on $X$, such that:
\begin{enumerate}
\item the action of $A$ is free on $X$, and its orbits are precisely the fibers of $X\to\pi_{s-1}(X)$;
\item this induces a pointwise action of $C^k(\cD_s(A))$ on $C^k(X)$: again this action is free, and its orbits are precisely the fibers of the induced map $C^k(X) \to C^k(\pi_{s-1}(X))$.
\end{enumerate}
\end{theorem}
We define the $k$-th structure group\index{structure group} $A_k(X)$\index[nota]{$A_k(X)$} of $X$ as the group $A_k(\pi_{k}(X))$ that
arises when we apply this theorem to the nilspace $\pi_{k}(X)$ (with $s=k$).
We will see in the proof that these groups are defined canonically in terms of the cubespace structure.
\begin{remark}\label{rem:cases}
We note that the weak structure theorem allows the following partial reconstruction of cubes on $X$.
Let $c\in C^k(\pi_{s-1}(X))$ be a cube.
If we know that $\widetilde c\in C^k(X)$ is a cube that projects to $c$ (that is $\pi_{s-1}(\widetilde c)=c$), then
we can find all cubes projecting to $\widetilde c$.
Indeed, let $c':\{0,1\}^k\to X$ be a configuration with $\pi_{s-1}(c')=c$, and write $a:\{0,1\}^k\to A_s(X)$
for the unique configuration such that $a(\omega).\widetilde c(\omega)=c'(\omega)$.
Then $c'$ is a cube if and only if $a\in C^k( \cD_s(A_s))$.
However, the weak structure theorem provides no insight for finding the first lift $\widetilde c$ of a cube in $C^k(\pi_{s-1}(X))$.
\end{remark}
\begin{remark}
We note that no topological input is really used in this theorem, except to show topological conclusions. For instance, if $X$ is a possibly infinite \emph{discrete} nilspace (i.e.~ignoring topology), the same proofs will apply, resulting in a discrete abelian group $A$ acting on the space and so on.
We will not consider the result in such generality, as we have no need for such statements in applications.
\end{remark}
\subsection{Automorphisms and recovering a nilmanifold structure}
Stage II in our plan for proving Theorem \ref{main-structural-v1} was to establish a strong structural result, stating that $X$ actually is a nilmanifold $G/\Gamma$, under some additional technical assumptions. We state this result, which is a variant of \cite{CS12}*{Theorem 7}, now.
\begin{theorem}
\label{toral-structure-theorem}
Suppose $X$ is a compact, ergodic, strongly connected nilspace of degree $s$. Further suppose that the structure groups $A_t$ for $1 \le t \le s$ are all (compact abelian) Lie groups, i.e.~isomorphic to tori $(\RR/\ZZ)^d \times K$ for some $d=d(A_t)\ge 0$ and some finite $K$.
Then $X$ is isomorphic to a nilmanifold $G/\Gamma$, equipped with its Host--Kra cubes. That is, there exists a connected Lie group $G$ equipped with a filtration of degree $s$ (and such that $G_i$ is also connected for each $i$), and a discrete and co-compact subgroup $\Gamma$ of $G$ compatible with the filtration (cf. Definition \ref{defn:nilmanifold}), such that
\begin{itemize}
\item $X$ is homeomorphic to $G/\Gamma$, and
\item under this homeomorphism, $C^k(X)$ is identified with $\HK^k(G_\bullet)/\Gamma$.
\end{itemize}
\end{theorem}
A complete proof of this theorem can be found in another paper of the authors \cite{GMV2}*{Theorem 2.18}.
\begin{remark}
\label{rem:strong-connectivity}
The difference between Theorem \ref{toral-structure-theorem} and \cite{CS12}*{Theorem 7} is that the latter makes the assumption that the structure groups are tori $\RR^d/\ZZ^d$ -- i.e.~that the finite group $K$ is necessarily trivial, or equivalently that $A_k(X)$ is connected -- in place of strong connectivity.
As we mentioned in Remark \ref{rem:connectivity}, these two conditions are equivalent, though exactly one of the implications is difficult. For the purposes of the current sketch, we will assume the stronger hypothesis (that the structure groups are tori), as this makes the exposition somewhat simpler while still addressing the core of the argument. The theorem is proved as stated in \cite{GMV2}*{Theorem 2.18}, and considerations arising from the weaker (strong connectivity) hypothesis are handled there.
\end{remark}
\begin{remark}
It is possible to formulate a version of Theorem \ref{toral-structure-theorem} without any reference to the structure groups. Then one needs to add some hypotheses on the topology of the space $X$, e.g. that it is locally connected and has finite Lebesgue covering dimension.
However, the proof of this statement requires the general structure theorem discussed in Section \ref{sec:structural-thm}.
For details, we refer to the Appendix of the companion paper \cite{GMV3}.
\end{remark}
At the heart of the challenge in proving the above theorem is recovering the non-abelian group operation on $G$. Although we were able to find a very explicit and combinatorial construction of the group operations on the structure groups $A_t$ just by exploiting concatenation and cube completion, there are seemingly fatal obstacles to such a simple approach working for the non-abelian group $G$. Essentially this is because Theorem \ref{toral-structure-theorem} is simply not true, even in spirit, if the strong connectivity assumptions on $X$ are dropped, and hence any proof must have some non-elementary topological aspect.\footnote{An extreme case is to consider finite (so necessarily totally disconnected) nilspaces. An example of one of these which is not of the form $G/\Gamma$ is given in \cite{HK08}*{Example 6}.}
Instead, the group operation is recovered in a completely different way: by considering the group of automorphisms of a nilspace $X$.
\begin{definition}
Write $\Aut(X)$ for the (topological) group of cubespace automorphisms\index{cubespace!automorphism} of $X$; that is, of homeomorphisms $\phi \colon X \to X$ such that both $\phi$ and $\phi^{-1}$ are cubespace morphisms.
\end{definition}
The topology being used here is compact-open. It turns out that $\Aut(X)$ has attached to it a canonical filtration,
\[
\Aut(X) = \Aut_0(X) \supseteq \Aut_1(X) \supseteq \dots \supseteq \Aut_s(X) \supseteq \{\id\}
\]
of degree $s$, assuming $X$ is a nilspace of degree $s$. The definition of the filtration is very explicit and depends only on the cubes of $X$, but we will not go into the details here; for a full description, see \cite{GMV2}*{Definition 2.7}.
It will typically be the case that the first containment $\Aut_0(X) \supseteq \Aut_1(X)$ is strict, and so we have no guarantee that $\Aut(X)$ is nilpotent -- in fact, often it will not be.\footnote{This can be seen even when $X = (\RR/\ZZ)^2$ with the degree $1$ filtration: $\Aut(X)$ contains a copy of $\SL_2(\ZZ)$.} For this reason it is better to work with the subgroup $\Aut_1(X)$, sometimes called the group of \emph{$1$-translations}. This inherits a filtration
\[
\Aut_1(X) = \Aut_1(X) \supseteq \dots \supseteq \Aut_s(X) \supseteq \Aut_{s+1}(X) = \{\id\}\index[nota]{$\Aut_i(X)$}
\]
which guarantees that $\Aut_1(X)$ is nilpotent of nilpotency class at most $s$.
In general, the group $\Aut_1(X)$ need not be connected. It turns out to be vital to work instead with the identity component of $\Aut_1(X)$, which we term $\Aut^\circ_1(X)$, and similarly with the revised filtration $\Aut^\circ_i(X)$, the identity component of $\Aut_i(X)$. Note this need \emph{not} be the same as $\Aut^\circ_1(X) \cap \Aut_i(X)$, and hence one should check that this is actually a filtration, but this is not hard to show.
Given this filtration, we can consider the Host--Kra cube-groups $\HK^k(\Aut_1(X)_\bullet)$ or $\HK^k(\Aut_1^\circ(X)_\bullet)$. Again these have a natural interpretation in terms of the cubes of $X$. One consequence is that if $x_0 \in X$ and $(\phi_\omega)_{\omega \in \{0,1\}^k} \in \HK^k(\Aut_1(X)_\bullet)$ then the configuration $(\phi_\omega(x_0))_{\omega \in \{0,1\}^k}$ is a cube in $C^k(X)$.
In other words given $x_0 \in X$, the map
\begin{align*}
\Aut_1(X) &\to X \\
\phi &\mapsto \phi(x_0)
\end{align*}
is a cubespace morphism if $\Aut_1(X)$ is given its Host--Kra cubes (and hence this restricts to a cubespace morphism on $\Aut^\circ_1(X)$). It will not be an isomorphism, as it is typically not injective: $x_0$ may have non-trivial stabilizer. But we can define another map of cubespaces
\begin{align*}
\Aut^\circ_1(X) / \stab(x_0) &\to X \\
\phi \, \stab(x_0) &\mapsto \phi(x_0)
\end{align*}
which is well-defined and again a cubespace morphism.
The hope is that this is in fact surjective, and furthermore an isomorphism of cubespaces (which is a strictly stronger property). If in addition $\Aut^\circ_1(X)$ were a Lie group and $\stab(x_0)$ a discrete and co-compact subgroup, we would have identified $X$ with a nilmanifold, as required.
However, it is not completely obvious \emph{a priori} that $\Aut(X)$ is not just the trivial group, in which case this map would be far from surjective. Establishing that $X$ has ``enough'' automorphisms is the most difficult step in this strategy.
We will offer a few further clues as to the structure of this part of the proof. Suppose for now that $s = 2$, and so $X$ is a nilspace of degree $2$. One source of non-trivial automorphisms of $X$ is provided by the action of the top structure group $A_2(X)$ on $X$ described above. Since this action fixes fibers of the quotient $X \to \pi_1(X)$ this still falls short of showing surjectivity.
However, we similarly have plenty of automorphisms of $\pi_1(X)$, given by the action of the next structure group $A_1(X)$ on $\pi_1(X)$. In fact (assuming $X$ is ergodic) this action is simply transitive.
So, we know that we can move by elements of $\Aut^\circ_1(X)$ along fibers of $\pi_1$,\footnote{Note we have used here the extra assumption that $A_2(X)$ is connected.} and also that we can move between fibers by automorphisms of $\pi_1(X)$. If we could show that these latter elements of $\Aut_1^\circ(\pi_1(X))$ can be lifted to elements of $\Aut_1^\circ(X)$, this would show that the action of $\Aut_1^\circ(X)$ is transitive, as desired.
With the formal definition of $\Aut_1(X)$ in place (which we deferred to \cite{GMV2}*{Definition 2.7}), it is not hard to show the following fact; this is done in \cite{GMV2}*{Proposition 3.2}.
\begin{lemma}
Elements of $\Aut_1(X)$ commute with $\pi_{s-1}$, and so there is a natural group homomorphism $\Aut_1(X) \to \Aut_1(\pi(X))$. Hence similarly we get a group homomorphism $\Aut^\circ_1(X) \to \Aut^\circ_1(\pi(X))$.
\end{lemma}
By the preceding discussion, the problem of showing surjectivity of the map $\Aut_1^\circ(X) \to X$ reduces to proving the following.
\begin{proposition}\label{prop:toral implies surjective}
Suppose $X$ is an ergodic nilspace of degree $s$, whose structure groups $A_t$ are tori. Then the map $\Aut^\circ_1(X) \to \Aut^\circ_1(\pi(X))$ from the previous proposition is surjective.
\end{proposition}
In the $s=2$ case, this would tell us that $\Aut^\circ_1(X)$ acts transitively on the fibers of $\pi_1$ (i.e.~given any two fibers, we can map some point on one fiber to some point on the other); and since we already know that $A_2$ acts transitively on each fiber, this would show that $\Aut_1^\circ(X)$ acts transitively on $X$ and so the map $\Aut_1^\circ(X) \to X$ is surjective.
Moving from surjectivity to isomorphism (and from the $s=2$ case to the general case) involves some extra work, but it is mainly technical.
To prove Proposition \ref{prop:toral implies surjective}, one begins by showing that if a $1$-translation on $\pi_1(X)$ is a small perturbation of the identity, then it lifts to a $1$-translation on $X$ (if $X$ is a nilspace of degree $2$). The proof of this is roughly to relate failure of lifting to some kind of ``cocycle'', and therefore express obstructions to lifting in terms of some kind of ``cohomology'' (where we use the term in a very loose sense). One then argues that this ``cohomology'' is somehow discrete, and so provided the $1$-translation is small enough, no obstructions arise.
These notions of ``cocycles'' and ``cohomology'' are useful tools, and we will allude to them again below. For the relevant formal definitions of cocycles and coboundaries, see \cite{GMV2}*{Definition 4.8} and the subsequent discussion; for a more in-depth discussion of cohomology more generally, and its relation to extensions of cubespaces, see \cite{CS12}*{Section 2.10} (or \cites{Can1,Can2}).
The proof of Theorem \ref{toral-structure-theorem} is very closely modeled on the arguments of Antol\'\i n Camarena and Szegedy \cite{CS12}. The main difference is the following. In constructing the lift of a translation on the canonical quotient, we first construct (in the paper \cite{GMV2}) a continuous lift, which is not necessarily a translation. Then we use the action of the structure group to ``correct it'' to a translation. On the other hand in \cite{CS12}, a measurable lift is constructed first, which is shown to be continuous a posteriori.
\subsection{The inverse limit statement}
Given this structure theorem for (ergodic, compact, strongly connected) nilspaces whose structure groups are Lie, our remaining task is to deduce something for general (ergodic, compact, strongly connected) nilspaces.
Essentially this task reduces to showing some version of the following (see also \cite{CS12}*{Theorem 4}).
\begin{theorem}
\label{thm:invlim}
Let $X$ be an ergodic compact nilspace of degree $s$.
Then there exists a sequence $X_n$ of ergodic compact nilspaces of degree $s$, such that $A_t(X_n)$ are Lie groups for all $n$ and all $1 \le t \le s$, and such that $X \cong \varprojlim X_n$ (with the inverse limit being in the sense of cubespaces).
\end{theorem}
In fact the result as proven in \cite{GMV3}*{Theorem 1.26} is marginally stronger, but in ways that use terminology we have not yet discussed.
Note that this gives some partial information about general (ergodic, compact) nilspaces without any connectivity assumption.
This is not quite enough to deduce Theorem \ref{main-structural-v1} as we do not get the precise description of the maps appearing in the inverse limit. A complete proof of a suitably stronger result can be found in \cite{GMV3}*{Theorems 1.27 and 1.28}.
We give an outline below, ignoring these finer points.
For simplicity, we again assume that the structure groups $A_k(X)$ are connected for all $k \ge 1$,
rather than just that $X$ itself is strongly connected
The starting point is again the weak structure theorem, together with the result that any compact connected abelian group is an inverse limit of tori.
Again consider first the case $s=2$. Since $A_2(X)$ is a compact connected abelian group, we may write $A_2(X) = \varprojlim K_m$ where $K_m$ are tori (with given surjective maps between them). Furthermore, it is straightforward to quotient $X$ by the action of a subgroup of $A_2$ (the kernel of one of the projections $A \to K_m$) to deduce the following:
\begin{fact*}
Under these hypotheses, $X$ is an inverse limit of nilspaces $X_m$ of degree $2$, with the property that $A_2(X_m)$ is a torus for all $m$.
\end{fact*}
This will allow us to assume -- given a bit of work -- that $A_2(X)$ is itself a torus, as otherwise we can reduce to that case.
Meanwhile, $A_1(X)$ is also a connected compact abelian group so we can write $A_1(X) = \varprojlim M_r$ for some tori $M_r=(\RR/\ZZ)^{d_{r}}$. We can deduce that $\pi_1(X)$ is an inverse limit of tori (in the sense of cubespaces).
The remaining challenge is to prove the following.
\begin{lemma}
Let $X$ be an ergodic compact nilspace of degree $2$, such that $A_2(X)$ is a torus. Suppose further that $A_1(X) = \varprojlim M_r$ for some tori $M_r$. Then for all sufficiently large integers $r$, there exists a canonical (degree $2$, ergodic, compact) nilspace $X_r$ and a quotient map $X \to X_r$ such that $A_2(X_r) \cong A_2(X)$, $A_1(X_r) \cong M_r$ and the diagram
\\[0.5\baselineskip]
\[
\begin{CD}
X @>>> X_r \\
@VV\pi_1V @VV\pi_1V \\
\cD_1(A_1(X)) @>\phi>> \cD_1(M_r)
\end{CD}
\]
\\[0.5\baselineskip]
commutes, where $\phi \colon A_1(X) \to M_r$ is the map from the inverse limit.
\end{lemma}
The problem is essentially one of ``pushing forward'' the degree $2$ cubespace structure on $X$ along the map $\phi$. Unfortunately this operation does not make sense in general.
Again, though, it will turn out that the obstructions to building such a push-forward space are measured by a kind of cohomology; in fact the same one as was used to prove the toral structure theorem. Again, we use a kind of discreteness result for this cohomology to argue that, if $M_r$ is sufficiently close to $\pi_1(X)$ in some sense, then these obstructions do not arise.
One way of measuring this ``closeness'' is the statement that the fibers of the map $\pi_1(X) \to M_r$ should have small diameter with respect to the metric on $\pi_1(X)$. Since $\pi_1(X) = \varprojlim M_r$ holds in the sense of metric spaces, this will be true provided $r$ is sufficiently large.
Again, the general case follows a very similar pattern, with this same argument repeated once for each $1 \le t \le s$.
As before our approach has much in common with that in \cite{CS12}, but differs in some important respects. Notably, the argument in that paper proceeds by establishing a correspondence between extensions of a given nilspace $X$ by a given compact abelian group $A$, up to isomorphism, and classes of measurable cocycles on $X$. An intricate argument is required to recover a topological object (the extended nilspace) from the measurable data of the cocycle. Some cocycle theory is then used to ``push forward'' the cocycle arising from $X \to \pi_1(X)$ onto $\cD_1(M_r)$ to build $X_r$. By contrast, our argument realizes $X_r$ as an explicit quotient of $X$, although many related tools are needed in the process.
\subsection{A dictionary of statements}
\label{sec:comparison}
For clarity and convenience, we summarize which of the main statements in these papers correspond to which in \cite{CS12}.
\begin{itemize}
\item The result concerning nilspaces whose structure groups are Lie, Theorem \ref{toral-structure-theorem}, is analogous to \cite{CS12}*{Theorem 7}; see Remark \ref{rem:strong-connectivity} for a detailed comparison.
\item The inverse limit theorem, Theorem \ref{thm:invlim}, is stated identically to \cite{CS12}*{Theorem 4} (though the proof strategies differ, as remarked above).
\item Our main structure theorem, Theorem \ref{main-structural-v1}, has no explicit counterpart in \cite{CS12}, but should be compared to what one obtains by concatenating \cite{CS12}*{Theorems 4 and 7}; see Remarks \ref{rem:inverse-limit} and \ref{rem:connectivity}.
\end{itemize}
Theorem \ref{main-structural-v1} also depends on \cite{GMV3}*{Theorem 1.27},
to obtain the sharper inverse limit statement (in the sense of the second paragraph of Remark \ref{rem:inverse-limit}). The latter has no counterpart in \cite{CS12}.
\section{The weak structure theory}
\label{sec:weak-overview}
We now turn to detailed statements and proofs of the weak structure theory.
In fact, we will approach this in two stages. This section will present the ``standard'' weak structure theorem \ref{basic-cs-structure-theorem} that appears in \cite{CS12}*{Theorem 1}, and give most of the proofs (although these may differ in some places from those in \cite{CS12}). However, elsewhere in the project we will need a ``relative'' analogue of all of these statements. This relative theory is developed in Section \ref{sec:elementary-v2}; many of the proofs there will be closely analogous to those from this section.
The relative versions are a strict generalization of the non-relative ones, so logically the results in Section \ref{sec:elementary-v2} suffice. Sometimes, therefore, we omit part of a non-relative proof and refer to the corresponding result from Section \ref{sec:elementary-v2}. Usually we do not, but this is for essentially pedagogical reasons.
\subsection{Glueing}
As a preliminary, we introduce one further property of cubespaces.
\begin{definition}
\label{defn:glueing}
We say a cubespace $X$ has the \emph{glueing property}\index{glueing} if ``glueing'' two cubes along a common face yields another cube.
Formally, suppose $c, c' \in C^k(X)$, and $c(\omega 1) = c'(\omega 0)$ for all $\omega \in \{0,1\}^{k-1}$. (Here we use $\omega 0$ to denote $(\omega_1, \dots, \omega_{k-1}, 0)$ and so on.)
Then the configuration
\[
c'' \colon \omega \mapsto \begin{cases} c(\omega) &\colon \omega_k = 0 \\ c'(\omega) &\colon \omega_k = 1 \end{cases}
\]
is in $C^k(X)$.
\end{definition}
The reason we have not defined this previously is that it follows from $k$-completion (for all $k$).
\begin{proposition}
\label{prop:fibrant-glueing}
Suppose a cubespace $(X, C^k(X))$ has $k$-completion for all $k$. Then it satisfies the glueing property.
\end{proposition}
\newcommand*{\glueingpic}[0]{
\inlinetikz{
\begin{scope}[shift={(-3,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\end{scope}
\end{scope}
\draw[->] (0, 1.5) -- (1, 1.5);
\begin{scope}[shift={(2,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\draw[dashed] (1,0,1) -- (1,1,1) -- (0,1,1);
\draw[dashed] (1,0,0) -- (1,1,0) -- (0,1,0);
\draw[dashed] (1,1,1) -- (1,1,0);
\draw[fill=black] (1,1,0) circle[radius=0.02cm];
\draw[fill=black] (1,1,1) circle[radius=0.02cm];
\end{scope}
\end{scope}
\draw[->] (5.5, 1.5) -- (6.5, 1.5);
\begin{scope}[shift={(7.5,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\draw (1,1,1) -- (0,1,1) -- (0,0,1) -- (1,0,1) -- cycle;
\draw (1,1,1) -- (1,1,0) -- (1,0,0) -- (1,0,1) -- cycle;
\draw (1,1,1) -- (0,1,1) -- (0,1,0) -- (1,1,0) -- cycle;
\draw[fill=black!50, opacity=.8] (0,1,0) -- (0,1,1) -- (1,0,1) -- (1,0,0) -- cycle;
\draw (0,0,0) -- (1,0,0) -- (1,1,0) -- (0,1,0) -- cycle;
\end{scope}
\end{scope}
}
}
\begin{proof}[Proof sketch]
We will defer a complete proof until Section \ref{sec:elementary-v2}, by which time we will have developed the machinery for a clean argument. Meanwhile, we will draw some pictures in the case $k=2$.
The idea is to place the two $2$-cubes to be glued as faces of a partial $3$-cube configuration; complete that (in two steps) to a $3$-cube; and then observe that the ``glued'' configuration is a sub-cube of this, and hence a cube by the axioms.
\glueingpic{}
\end{proof}
\subsection{The Heisenberg example}
\label{sec:hberg-example}
As motivation, we will examine an example called the Heisenberg nilmanifold.
This is the nilspace $\HK(\cH_\bullet)/\Gamma$ constructed as in Appendix \ref{app:hk}, where
\[
\cH = \left\{ \heis{x}{y}{z} \colon x,y,z \in \RR \right\}
\]
is the Heisenberg group equipped with the filtration
$\cH = \cH_0 = \cH_1 \supseteq \cH_2 \supseteq \{\id\}$ where $\cH_2$ is the center,
\[
\cH_2 = \left\{ \heis{0}{0}{z} \colon z \in \RR \right\}
\]
and $\Gamma$ is the discrete co-compact subgroup that consists of the elements of $\cH$ with integral entries.
Note that, for any filtered group, $G_j$ is a normal subgroup of $G$ (since $[G_0, G_j] \subseteq G_j$ by the filtration property), and so we may consider the group quotient $G \to G / G_j$. This comes with an induced filtration $(G_i / G_j)_{i \ge 0}$ of degree $j-1$.
For $\cH$, the non-trivial case is $\cH / \cH_2$. This is isomorphic to the abelianization $\RR^2$ with the degree $1$ filtration $\RR^2 = \RR^2 \supseteq \{0\}$.
We also get an induced map on the nilmanifold $\cH / \Gamma$, i.e.~$\pi \colon \cH / \Gamma \to (\cH / \cH_2) / (\Gamma / (\Gamma \cap \cH_2)) = \RR^2 / \ZZ^2$. This map corresponds to
\[
\pi \colon \heis{x}{y}{z} \Gamma \mapsto (x, y) \bmod 1 \in \RR^2 / \ZZ^2
\]
and one can check this is well-defined. Hence, the nilmanifold $\cH / \Gamma$ has a quotient isomorphic to a torus $\RR^2 / \ZZ^2$, sometimes called the ``horizontal torus''. Moreover, it is essentially automatic that this map is well-behaved with respect to the Host--Kra cubes $\HK^k(\cH_\bullet)/\Gamma$: the image of these cubes is precisely $\HK^k(\RR^2/\ZZ^2)$.
The fibers of this quotient map are of the form
\[
\left\{ \heis{x}{y}{z'} \Gamma \colon z' \in \RR / \ZZ \right\}
\]
i.e.~each fiber has an action by the center $\cH_2$, or more precisely a simply transitive action by $\cH_2 / (\cH_2 \cap \Gamma) = \RR / \ZZ$. Again, this action respects cubes in the following sense: given an element
\[
c(\omega) = \heis{x_\omega}{y_\omega}{z_\omega} \Gamma
\]
in $\HK^3(\cH_\bullet)/\Gamma$, and another configuration
\[
c'(\omega) = \heis{x_\omega}{y_\omega}{z'_\omega} \Gamma
\]
so $\pi(c) = \pi(c')$, then one can check that $c'$ is a cube if and only if
\[
z_{000} - z_{001} - \dots + z_{110} - z_{111} =
z'_{000} - z'_{001} - \dots + z'_{110} - z'_{111} \ .
\]
Equivalently, $c'$ is a cube if and only if it is obtained by acting on $c$ pointwise by an element of $\HK^3(\RR/\ZZ)$, where $\RR/\ZZ$ is given the degree $2$ filtration $\RR/\ZZ = \RR/\ZZ = \RR/ \ZZ \supseteq \{0\}$.
In summary, for the Heisenberg nilmanifold we have maps
\[
\cH / \Gamma \xrightarrow{\pi} \RR^2 / \ZZ^2 \to \{\ast\}
\]
where for each map the fibers have a simply transitive action by a compact abelian group ($\RR/\ZZ$ and $\RR^2/\ZZ^2$ respectively); and there is an induced map
\[
\HK^k(\cH_\bullet) / \Gamma \xrightarrow{\pi} \HK^k(\RR^2 / \ZZ^2) \to \{\ast\}
\]
where the fibers are given by $\HK^k$ of the corresponding compact abelian groups, equipped with the ``trivial'' filtrations of degree $2$ and $1$ respectively.
It is not hard to argue that a similar property holds for any nilmanifold of any degree, not just the Heisenberg nilmanifold. For a filtered group of degree $s$, one gets a sequence of maps
\[
G \to G / G_s \to G/G_{s-1} \to \dots \to G/G_2 \to G/G_1
\]
where typically $G = G_1$ and so the last step is the trivial group. The kernel of each of these maps is an abelian group. Again, these maps induce maps of nilmanifolds
\[
G/\Gamma \to (G / G_s) / (\Gamma / (\Gamma \cap G_s)) \to (G / G_{s-1}) / (\Gamma / (\Gamma \cap G_{s-1})) \to \dots
\]
where the fibers of each map are now compact abelian groups. This is a tower of extensions of nilmanifolds. A similar compatibility property of the Host--Kra cubes to that stated above also holds.
In this discussion, we have made heavy use of the group structure of $G$ and the precise description of Host--Kra cubes. It is a slightly surprising but very important fact that this sequence of quotient maps can be recovered in the completely abstract setting of general nilspaces. Indeed, for any nilspace $X$ we can define a tower
\[
X \to X / \sim_{s-1} \to X / \sim_{s-2} \to \dots
\]
where $\sim_{s-1}$ is the equivalence relation from Definition \ref{df:canonical-rel}. When $X = G/\Gamma$ as above, it turns out this is the same tower that we just constructed. For general $X$, many of the properties discussed above -- notably, that the fibers are compact abelian groups, which act compatibly on cubes in the sense described -- can be proven in the abstract setting.
In other words, we do not need to know about the global group operation on $G$ to recover the tower of quotients of the nilmanifold $G/\Gamma$, and indeed the same conclusions hold even for nilspaces that are not of the form $G/\Gamma$ for any group $G$.
\subsection{The canonical factors}
\label{sec:canonical-factors}
As just discussed, for a general compact cubespace $X$ with $k$-completion we are interested in the quotient $X / \sim_{s-1}$ of $X$, which is the general analogue of the quotients $(G / G_s) / (\Gamma / (\Gamma \cap G_s))$ of a filtered nilmanifold $G/\Gamma$. The space $X / \sim_{s-1}$ was defined in Section \ref{sc:outline-weak}: see Definition \ref{df:quotient} (quotient cubespace) and Definition \ref{df:canonical-rel} (canonical equivalence relation).
Our first task is to verify the properties of $X / \sim_s$ that we claimed in Section \ref{sc:outline-weak} immediately after the definition (and a few more). This discussion is very similar to \cite{CS12}*{Section 2.4} and \cite{HK08}*{Section 3.3}.
\begin{proposition}
\label{prop-canonical-factors}
Let $(X, C^k(X))$ be a compact cubespace with the glueing property, and let $s \ge 0$ be an integer. Then the canonical equivalence relation $\sim_s$ is indeed a closed equivalence relation and satisfies the following
{\em universal replacement property}.\index{universal replacement property}
If $k \le s+1$ and $c \in C^{k}(X)$, $c' \colon \{0,1\}^k \to X$, $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{1}$ and $c(\vec{1}) \sim_s c'(\vec{1})$, then $c' \in C^{k}(X)$.
Moreover, $X / \sim_s$ has $(s+1)$-uniqueness, and $\sim_s$ is the smallest equivalence relation with this property. Finally, if $X$ has $k$-completion for all $k$ then so does $X / \sim_s$, and hence $X / \sim_s$ is a compact nilspace of degree $s$.
\end{proposition}
\begin{remark}
Note we are using a convention of stating results using $\vec{1}$ as the ``special'' vertex of the discrete cube $\{0,1\}^k$, whenever one is required. However, using the various morphisms of discrete cubes from the definitions of a cubespace, which act transitively on $\{0,1\}^k$, these properties hold just as well for any fixed vertex in $\{0,1\}^k$.
\end{remark}
\begin{remark}
\label{rem:sim_0}
Consider the case $s=0$. Then $x \sim_0 y$ if and only if $[x, y]$ is a $1$-cube. It is immediate from the glueing property that this is an equivalence relation. If $X$ is ergodic, then $X / \sim_0 = \{\ast\}$ is a one-point space. More generally, this identifies the ``ergodic components'' of $X$.
\end{remark}
The following more explicit characterization of $\sim_s$ will be very helpful.
See also \cite{CS12}*{Lemma 2.3} and \cite{HK08}*{Proposition 3}.
\begin{lemma}
\label{lem:canonical-characterization}
We have $x \sim_s y$ if and only if the configuration $c\colon \{0,1\}^{s+1} \to X$ given by
\[
c(\omega) = \begin{cases} y & \colon \omega = \vec{1} \\ x & \colon \omega \ne \vec{1} \end{cases}
\]
is an $(s+1)$-cube.
\end{lemma}
\begin{proof}[Proof sketch]
The ``if'' direction is straightforward, since the constant configuration $c'(\omega) = x$ for all $\omega$ is in $C^{s+1}(X)$ by the axioms (see Remark \ref{rem:constant-cube}). We illustrate the ``only if'' direction with a picture when $s=1$; the general case is similar but notationally awkward.
We know there exist $c, c'$ with $c(\omega) = c'(\omega)$ for $\omega \ne \vec{1}$, $c(\vec{1}) = x$ and $c'(\vec{1}) = y$. Consider the picture
\inlinetikz{
\begin{scope}[scale=2]
\draw[thin,black] (0,0) grid (2, 2);
\node [below left] at (0, 0) {$x$};
\node [above left] at (0, 2) {$x$};
\node [below right] at (2, 0) {$x$};
\node [above right] at (2, 2) {$y$};
\node [left] at (0, 1) {$c(01)$};
\node [right] at (2, 1) {$c(01)$};
\node [below] at (1, 0) {$c(10)$};
\node [above] at (1, 2) {$c(10)$};
\node [above right] at (1, 1) {$c(00)$};
\node at (0.5, 0.5) {$c$};
\node at (0.5, 1.5) {$c$};
\node at (1.5, 0.5) {$c$};
\node at (1.5, 1.5) {$c'$};
\end{scope}
}
It is clear each of the small cubes is a cube, as they are just rotated and reflected copies of $c$ or $c'$. By two applications of the glueing property, the outer square is also a cube, as required.
\end{proof}
\begin{proof}[Proof sketch of Proposition \ref{prop-canonical-factors}]
First we check that $\sim_k$ is an equivalence relation. Symmetry and reflexivity are immediate from the original definition. For transitivity, we apply the lemma; again we will give a pictorial sketch when $s=1$. Suppose $x \sim_s y$, $y \sim_s z$ and consider
\inlinetikz{
\doublesquare{$x$}{$y$}{$z$}{$y$}{$y$}{$y$}{1}
}
where the left and right squares are cubes by the lemma and a reflection. Then the outer rectangle is a cube by the glueing property. But so is (say)
\inlinetikz{
\singlesquare{$y$}{$y$}{$x$}{$x$}
}
by a duplication operation (i.e.~using the cubespace axioms), and so $x \sim_s z$ from the definition, as required.
Note that it is clear from the hypotheses about the spaces $C^k(X)$ being closed and $X$ being compact that $\sim_s$ is a closed relation.
Next, we prove the universal replacement property. First suppose $k=s+1$. Take $c, c'$ as in the statement, and write $x = c(\vec{1})$, $y = c'(\vec{1})$. Again we give a picture:
\inlinetikz{
\begin{scope}[scale=1.5]
\draw[thin,black] (0,0) grid (2, 2);
\node [below left] at (0, 0) {$c(00)$};
\node [above left] at (0, 2) {$c(01)$};
\node [below right] at (2, 0) {$c(10)$};
\node [above right] at (2, 2) {$y$};
\node [left] at (0, 1) {$c(01)$};
\node [right] at (2, 1) {$x$};
\node [below] at (1, 0) {$c(10)$};
\node [above] at (1, 2) {$x$};
\node [above right] at (1, 1) {$x$};
\end{scope}
}
The bottom left square is just $c$, the top right is a cube by the lemma, and the other two small squares are cubes again by a duplication. Hence the outer square is a cube by glueing, as required.
Now suppose $k < s+1$. Given a cube $c \in C^k(X)$, using appropriate morphisms of discrete cubes and the cubespace axioms, we may
\begin{itemize}
\item duplicate $c$ up to an $(s+1)$ cube $\tilde{c} \in C^{s+1}(X)$;
\item change entries of $\tilde{c}$ repeatedly using the above; and
\item restrict to some appropriate face of $\tilde{c}$ to obtain a cube of dimensions $k$ with the desired properties.
\end{itemize}
Hence we have universal replacement for all $k \le s+1$.
We now argue the definition of $\sim_s$ and the universal replacement property imply $(s+1)$-uniqueness of the quotient. Suppose $\tilde{c}$ and $\tilde{c}'$ are two cubes in $C^{s+1}(X/\sim_s)$ such that $\tilde{c}(\omega) = \tilde{c}'(\omega)$ for all $\omega \ne \vec{1}$. By definition of the quotient cubespace, there are cubes $c, c' \in C^{s+1}(X)$ such that $\pi_s(c) = \tilde{c}$, $\pi_s(c') = \tilde{c}'$. But by repeated application of the universal replacement property, \emph{any} configuration $c \colon \{0,1\}^{s+1} \to X$ such that $\pi_s(c) = \tilde{c}$ is a cube, and similarly for $c'$. Hence we are free to choose $c, c'$ such that $c(\omega) = c'(\omega)$ for each $\omega \ne \vec{1}$. By definition of $\sim_s$, we now see that $c(\vec{1}) \sim_s c'(\vec{1})$, and so $\tilde{c}(\vec{1}) = \tilde{c}'(\vec{1})$ as required.
As remarked after Definition \ref{df:canonical-rel}, the statement that $\sim_s$ is the smallest equivalence relation $\sim$ for which $X / \sim$ has $(s+1)$-uniqueness is clear.
The final statement to prove is the completion property for $X/\sim_s$. In fact this follows from more general statements about ``fibrations'', defined in Section \ref{sec:elementary-v2}, so we defer the proof to there (see Remark \ref{rem:completion for quotient}).
\end{proof}
Note it is clear that $X/\sim_s$ is also ergodic if $X$ is.
\begin{remark}
We observe that this quotient $\pi_s$ as constructed is completely canonical: firstly in that it depends only on the cubespace $(X, C^k(X))$, and secondly that it has the following universal property: if $(Y, C^k(Y))$ is any other nilspace of degree $s$ then any cubespace morphism $X \to Y$ factors through $\pi(X)$.
In particular, $\sim_s$ is trivial if and only if $X$ has $(s+1)$-uniqueness.
\end{remark}
The full strength of these canonical factors comes when they are chained together. That is, as before we can construct a tower of maps
\[
X \xrightarrow{\pi_s} \pi_s(X) \xrightarrow{\pi_{s-1}} \pi_{s-1}(X) \rightarrow \dots \xrightarrow{\pi_0} \pi_0(X)
\]
where $\pi_0(X) = \{\ast\}$, the one-point space, provided $X$ is ergodic. To check this makes sense, we need to verify the following trivial observation.
\begin{proposition}
\label{equiv-nested}
The equivalence relations $\sim_s$ are nested, i.e.~if $t \ge s$ and $x \sim_t y$ then $x \sim_s y$. Moreover, $\pi_s(\pi_t(X)) = \pi_s(X)$, i.e.~the definition of $\sim_s$ is not affected by first quotienting by $\sim_t$.
\end{proposition}
\begin{proof}
If $c, c' \in C^{t+1}(X)$ are any cubes verifying $x \sim_t y$ then any sub-cube of dimension $(s+1)$ containing $\vec{1}$ will verify $x \sim_s y$.
\end{proof}
\subsection{Structure groups and the weak structure theorem}
\label{sec:structure-groups}
Having established the existence of the tower
\[
\dots \xrightarrow{\pi_s} \pi_s(X) \xrightarrow{\pi_{s-1}} \pi_{s-1}(X) \rightarrow \dots \xrightarrow{\pi_0} \pi_0(X)
\]
in the abstract, to further the analogy with the Heisenberg case (or more generally, the case of any nilmanifold) we need to describe the \emph{fibers} of each map $\pi_t$. Recall these are expected to have the structure of a compact abelian group; or more precisely, to have a free and transitive action by a compact abelian group.
As in Section \ref{sc:outline-weak} we let $\cD_s(A)$ denote the Host--Kra cubespace on an Abelian group
equipped with the filtration $A = A_0 = A_1 = \dots = A_s \supseteq \{0\}$, which is a nilspace of degree $s$.
Recall also the statement of the weak structure theorem (Theorem \ref{basic-cs-structure-theorem}).
\begin{remark}
Suppose $X$ is an ergodic nilspace of degree $1$, and consider the weak structure theorem (Theorem \ref{basic-cs-structure-theorem}) in the case $s=1$. Since $\pi_0(X) = \{\ast\}$ (see Remark \ref{rem:sim_0}), the theorem asserts precisely that $X$ is isomorphic to $\cD_1(A)$ for some compact abelian group $A$, i.e.~$X$ bijects with $A$ and the cubes of $X$ are identified with the Host--Kra cubes of $A$ with the degree $1$ filtration.
(This isomorphism is found by fixing $x_0 \in X$ and identifying $A \leftrightarrow X$ by $a \leftrightarrow a(x_0)$. In particular this is canonical only up to the choice of $x_0$.)
\end{remark}
A proof of this case is sketched above in Section \ref{sc:1step}. However, that proof does not generalize entirely cleanly to the case $s > 1$. We will now sketch a slightly different argument that does generalize, but working again in the case $s=1$ for ease of notation.
First we will sketch how the proof would look if we already knew that $X = \cD_1(A)$ for some $A$. We consider all the edges \footnote{An edge is just another name for a $1$-cube.} $[a, b]$ of $X$, and associate to each one the group element $(b - a)$. We define an equivalence relation on edges by $[a,b] \sim [a', b']$ if $b - a = b' - a'$; so $A$ is precisely the set of equivalence classes of edges under this relation.
Given edges $[a, b]$ and $[b, c]$ we can concatenate them to get $[a, c]$. If the associated elements of $A$ are $r = b - a$ and $s = c - b$ then $[a,c]$ is associated to $c - a = r + s$. Thus we can recover the group addition operation on $A$ in a combinatorial fashion by concatenation. The other operations are similarly easy to define.
The task is now to show that these same constructions make sense without any \emph{a priori} assumptions on $X$ other than that it is a nilspace of degree $1$.
\begin{proof}[Sketch proof of Theorem \ref{basic-cs-structure-theorem}, $s=1$]
Since $X$ is ergodic, the edges $C^1(X)$ are in bijection with $X \times X$. For two edges $[x, y]$ and $[x', y']$ we write $[x,y] \sim [x',y']$ whenever
\inlinetikz{\singlesquare{$x$}{$y$}{$x'$}{$y'$}}
is a $2$-cube. This is an equivalence relation by the cubespace axioms and glueing. As a set, we define $A := (X \times X) / \sim$.
\begin{claim}\label{claim:uniqueness}
Let $[x,y]$ be an edge. For any fixed $e \in X$ there is an unique $a \in X$ such that $[x,y] \sim [e,a]$. In other words, the class of $[x,y]$ in $A$ has an unique representative of the form $[e, a]$.
\end{claim}
\begin{proof}[Proof of claim]
This follows trivially from unique completion of the corner
\inlinetikz{\singlesquare{$x$}{$y$}{$e$}{$\ast$}}
giving $a$ in the top right.
\end{proof}
We now define addition on $A$ by concatenation as suggested above. More precisely, given $g, h \in A$, we fix an $e \in X$ and choose the unique representatives $g \sim [x, e]$ and $h \sim [e, y]$; then $g + h$ is defined to be the class of the concatenation $[x, y]$.
\begin{claim}
\label{well-defined-claim-v1}
This definition of $g + h$ is well-defined; i.e.~the outcome does not depend on the choice of $e$.
\end{claim}
\begin{proof}[Proof of claim]
Considering the diagram
\inlinetikz{\doublesquare{$x$}{$e$}{$y$}{$x'$}{$e'$}{$y'$}{1.0}}
where $[x, e] \sim [x', e'] \sim g$ and $[e, y] \sim [e', y'] \sim h$, since the outer square is a $2$-cube by glueing we have $[x, y] \sim [x', y']$ as required.
\end{proof}
The identity $0 \in A$ is given by the class of constant edges $[x, x]$ (by $2$-uniqueness this is indeed a class of $\sim$), and inversion by the operation $[x, y] \mapsto [y, x]$.
\begin{claim}
The set $A$ with these operations forms a (topological) abelian group.
\end{claim}
\begin{proof}[Proof of claim]
That the proposed identity and inverse operation do what they claim is a trivial check. Associativity follows from associativity of concatenation, together with Claim \ref{well-defined-claim-v1}. For commutativity, suppose $[x, e] \sim g$, $[e, y] \sim h$; by completing the corner to get $e'$ in
\inlinetikz{\singlesquare{$e$}{$y$}{$x$}{$e'$}}
we have $[e', y] \sim g$, $[x, e'] \sim h$ and hence $h + g \sim [x, y]\sim g+h$ as required.
Issues of continuity are not difficult to justify. Note that the equivalence relation $\sim$ is closed (as $C^2(X)$ is a closed subspace of $X^4$) and so $A$ is a compact metric space. We now sketch the proof that $+:A\times A\rightarrow A$ is continuous. For a fixed $e \in X$, the map $r_e \colon X \to A$ given by $x \mapsto [e,x]$ is clearly continuous (by the definition of the product and quotient topologies) and a bijection (by Claim \ref{claim:uniqueness}) so is a homeomorphism. The same holds for $\ell_e \colon x \mapsto [x,e]$. Hence, the composite
\begin{align*}
A \times A &\to X \times X \to A \\
(g,h) &\mapsto (\ell_e^{-1}(g),r_e^{-1}(h)) \mapsto [\ell_e^{-1}(g),r_e^{-1}(h)] / \sim
\end{align*}
is also continuous; but this is the definition of $+$, as required.
\end{proof}
%
\vspace{\baselineskip}
%
We now define the group action of $A$ on $X$. Given $g \in A$ and $x \in X$ we take $g(x)$ to be the unique element of $X$ such that $g \sim [x, g(x)]$. It is clear $0 \in A$ acts trivially; looking at the diagram
\inlinetikz{\doublesquare{$x$}{$g(x)$}{$h(g(x))$}{$c$}{$e$}{$d$}{1.0}}
where $g \sim [c, e]$ and $h \sim [e, d]$, we conclude this is indeed a group action by considering the outer square. Moreover, it is trivially simply transitive.
Finally we must investigate the cubes of $X$ in terms of $A$, i.e.\ prove (ii) from Theorem \ref{basic-cs-structure-theorem}. For $0$ and $1$-cubes there is nothing to say. Suppose $c$ is the configuration
\inlinetikz{\singlesquare{$x$}{$y$}{$z$}{$w$}}
in $X^{\{0,1\}^2}$. By definition of $\sim$, this is a cube if and only if $[x, y] \sim [z, w]$. Furthermore it is obtained by acting on the constant cube
\inlinetikz{\singlesquare{$e$}{$e$}{$e$}{$e$}}
by the elements of $A$ represented by $r := [e, x]$, $s := [e, y]$, $t := [e, z]$, $u := [e, w]$ respectively.
By considering the concatenation of $[x,e]$, $[e,y]$ and using the definitions of the group operation on $A$, we find that $[x, y] \sim s - r$, and similarly $[z, w] \sim u - t$. Hence, $c$ is a cube if and only if $[x,y] \sim [z,w]$, if and only if $s - r = u - t$; i.e.~if and only if $c$ is obtained by acting on a constant cube by an element of $C^2(\cD_1(A))$. This is precisely what is required by (ii) in the case $k=2$.
It is straightforward to deduce the cases $k \ge 3$ of Theorem \ref{basic-cs-structure-theorem}(ii) from the $k=2$ case, but we defer this argument to the more general setting of Section \ref{sec:elementary-v2} (see Theorem \ref{relative-structure-thm} and especially the proof of Lemma \ref{lem:relative cubes}).
\end{proof}
Several adjustments are required to make this argument work in the full case of Theorem \ref{basic-cs-structure-theorem}. However, we will address these at the same time as stating and proving a version of Theorem \ref{basic-cs-structure-theorem} in greater generality.
Although the statement of Theorem \ref{basic-cs-structure-theorem} refers only to the map $\pi_{s-1} \colon X \to \pi_{s-1}(X)$, we can exploit the fact that $\pi_{s-1}(X)$ is itself a nilspace of degree $(s-1)$ to apply the theorem repeatedly, and obtain a tower
\[
X \xrightarrow{\pi_{s-1}} \pi_{s-1}(X) \to \dots \xrightarrow{\pi_0} \pi_0(X) = \{\ast\}
\]
as promised. We write $A_t$ (for $1 \le t \le s$, if $X$ is a nilspace of degree $s$) for the compact abelian groups that arise from the application of Theorem \ref{basic-cs-structure-theorem} to $\pi_t(X)$ at each stage; collectively, these are referred to as the \emph{structure groups} of $X$.
\section{The relative weak structure theory}
\label{sec:elementary-v2}
With these overviews concluded, we will now return to a formal account of the weak structure theory. In this second pass, we will both give complete proofs, and also introduce a slightly greater degree of generality which will be useful to us in future.
\subsection{Fibrations}
One key definition missing from our initial treatment of cubespaces and nilspaces is the following notion of a \index{fibration}\emph{fibration} \footnote{This term was chosen by analogy with the notion of a Kan fibration in simplicial homotopy theory. A simplicial set is called a \emph{Kan complex} or \emph{fibrant} if for all $k \ge 1$, any collection of $k$ compatible $(k-1)$-simplices (called a ``horn'') can be completed (or ``filled'') to a $k$-simplex with these as faces, analogously to completion of $k$-corners in a cubespace. The relative analogue of this for a map between two simplicial sets is known as a \emph{Kan fibration}. The survey \cite{friedman} is a very approachable introduction to these ideas. There are other resemblances between these two theories which may not be entirely superficial, but we will not pursue this here.}, which is a particular kind of cubespace morphism. This can be thought of as a relative version of the notion of the corner completion property. An alternative heuristic is that fibrations are morphisms which are ``properly'' surjective in the category of cubespaces.
\begin{definition}
\label{def:fibration}
Let $f \colon X \to Y$ be a morphism between cubespaces $(X, C^k(X))$ and $(Y, C^k(Y))$. We say $f$ is a \emph{fibration} if the following holds for any $k \ge 0$.
Suppose $c \in C^k(Y)$ is a $k$-cube, and $\lambda \colon \{0,1\}^k \setminus \{\vec{1}\} \to X$ is a $k$-corner (see Definition \ref{defn:k-completion}) such that $f(\lambda(\omega)) = c(\omega)$ for all $\omega \in \{0,1\}^k \setminus \{\vec{1}\}$. Then there exists $x \in X$ such that
\begin{itemize}
\item the configuration
\begin{align*}
\tilde{c} \colon \{0,1\}^k &\to X \\
\omega &\mapsto \begin{cases} \lambda(\omega) &\colon \omega \ne \vec{1} \\ x &\colon \omega = \vec{1} \end{cases}
\end{align*}
is in $C^k(X)$; and
\item $f(x) = c(\vec{1})$ (and hence $f(\tilde{c}) = c$).
\end{itemize}
\end{definition}
Informally, this definition can be stated as follows: given any $k$-corner $\lambda$ of $X$ and any completion of $f(\lambda)$ to a cube of $Y$, we can complete $\lambda$ to a cube in a compatible fashion.
\begin{remark}
By taking $k=0$, we find that any fibration is surjective as a map $X \to Y$. However, this is strictly weaker, as can be seen by taking $X = \cD_1(\RR/\ZZ)$, $Y = \cD_2(\RR/\ZZ)$ and $f \colon \RR/\ZZ \to \RR/\ZZ$ the identity map. This is a surjective cubespace morphism, but taking $k=2$ and choosing $c$ to be any configuration in $(\RR/\ZZ)^4$ that is not a cube of $C^2(X)$ (but is in $C^2(Y) = (\RR/\ZZ)^4$) it is seen not to be a fibration.
\end{remark}
\begin{remark}
Let $X$ be a cubespace and consider the unique morphism $f \colon X \to \{\ast\}$, the one-point cubespace. Then $f$ is a fibration if and only if $X$ has $k$-completion for all $k$.
Hence, we refer to spaces that have $k$-completion for all $k$ as \emph{fibrant}.\index{fibrant}
\end{remark}
By the previous remark, whenever we prove a statement about fibrant cubespaces, it is reasonable -- and sometimes useful -- to ask for a relative version that holds for fibrations.
\begin{remark}
\label{rem:fiber-surjective}
The notion of a fibration is intimately related to that of a \emph{fiber-surjective morphism} which appears in \cite{CS12}*{Section 2.8}. There, a cubespace morphism $f \colon X \to Y$ is called fiber-surjective if for each $k \ge 0$, the image of any $\sim_k$ class in $X$ is a $\sim_k$ class of $Y$. It is assumed in this context that both $X$ and $Y$ are nilspaces.
It is not hard to check that if $X$ and $Y$ are nilspaces and $f \colon X \to Y$ is a cubespace morphism, then $f$ is a fibration if and only if it is fiber-surjective. If $X$ and $Y$ are general cubespaces, it is not clear how to extend the definition of fiber-surjectivity in general, so no comparison is possible. If one did extend it to a wider class of cubespaces, we note that the map $X \to \{\ast\}$ will surely be vacuously fiber-surjective for any $X$, whereas it is a fibration if and only if $X$ is fibrant, and hence the notions would be inequivalent whenever $X$ is not assumed to be fibrant.
Our reasons for working with the definition of a fibration given here rather than with fiber-surjective morphisms are twofold.
\begin{itemize}
\item We will have cause later in the project to work with fibrations $f \colon X \to Y$ where $X$ and $Y$ are not nilspaces. Here the distinction matters, and the notion of a fibration is the correct one for what we need.
\item The fact that the definition of a fibration cleanly extends the completion axiom means that many proofs that work for nilspaces tend to transfer without significant modification to relative versions for fibrations. In the authors' view, this makes the theory cleaner in places, and suggests the fibration definition is perhaps the more natural one.
\end{itemize}
However, we stress that the difference is a technical one, and logically unimportant in almost all cases.
\end{remark}
We will now show a couple of lemmas, together with their proofs, to illustrate this concept in action.
\begin{lemma}[Lifting of ``partial cubes'']
\label{extension-property}
Let $f \colon X \to Y$ be a fibration between cubespaces $X$ and $Y$.
Let $S \subseteq \{0,1\}^k$ be a downwards-closed subset, i.e.~if $\omega \in S$ and $\omega' \subseteq \omega$ then $\omega' \in S$. Let $T$ be another such subset with $S \subseteq T$.
Suppose we have configurations $A \colon S \to X$, $B \colon T \to Y$ such that
\begin{enumerate}
\item for each $\omega \in S$, we have $f(A(\omega)) = B(\omega)$;
\item for each $\omega \in S$, the configuration
\begin{align*}
\{ \omega' \colon \omega' \subseteq \omega \} &\to X \\
\omega' &\mapsto S(\omega')
\end{align*}
is a cube of $X$; and
\item similarly for $Y$ and $T$.
\end{enumerate}
Then we can extend $A$ to a function $\tilde{A} \colon T \to X$ such that $f(\tilde{A}) = B$ and $\tilde{A}$ now has property (ii) with respect to $T$.
\end{lemma}
Here we again used the identification of $\{0,1\}^k$ with the set of subsets of $[k]$ as in Section \ref{sec:hk}.
This rather general statement has the following more natural corollaries.
\begin{corollary}
If $f \colon X \to Y$ is a fibration then the induced map $f \colon C^k(X) \to C^k(Y)$ is surjective for all $k \ge 0$.
\end{corollary}
\begin{proof}
Set $S = \emptyset$, $T = \{0,1\}^k$ in the lemma.
\end{proof}
\begin{corollary}
\label{cor:fibrant-extension-property}
If $X$ is fibrant, $S \subseteq \{0,1\}^k$ is downward closed and $A \colon S \to X$ is a configuration satisfying property (ii) from the lemma, then $A$ extends to a cube, i.e.~there is some cube $c \in C^k(X)$ with $c|_S = A$.
\end{corollary}
\begin{proof}
Apply the lemma with $Y = \{\ast\}$, $T = \{0,1\}^k$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{extension-property}]
If $T = S$ there is nothing to do. Suppose $\omega_0$ is a minimal element of $T \setminus S$. If we can extend $A$ to a configuration on $S \cup \{\omega_0\}$ that still satisfies (i) and (ii), we will be done by iterating this process.
An example of this set-up is shown in the following diagram, where $T = \{0,1\}^3 \setminus \{\vec{1}\}$, $S = \{000,001,010,011,100\}$ and $\omega_0 = 101$:
\inlinetikz{
\begin{scope}[shift={(-3,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[line width=0.05cm,black!70] (0,0,0) -- (1,0,0);
\end{scope}
\end{scope}
\draw[->,dashed] (0, 1) -- (1, 1);
\begin{scope}[shift={(2,0)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[line width=0.05cm,black!70] (0,0,0) -- (1,0,0);
\draw[thick,dashed] (0,0,1) -- (1,0,1) -- (1,0,0);
\draw[pattern=dots,draw=none] (0,0,0) -- (0,0,1) -- (1,0,1) -- (1,0,0) -- cycle;
\draw[fill=black] (1,0,1) circle[radius=0.03cm] node[right] {$\omega_0$};
\end{scope}
\end{scope}
\draw[->] (-1,-0.5) -- (0, -1);
\draw[->] (2,-0.5) -- (1, -1);
\begin{scope}[shift={(-0.5,-4)}]
\begin{scope}[x={(1, 0)}, y={(0, 1)}, z={(0.352, 0.317)}, scale=2]
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (0,0,1) -- (0,1,1) -- (0,1,0) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,0,1) -- (0,0,1) -- cycle;
\draw[fill=black!40,fill opacity=.4] (0,0,0) -- (1,0,0) -- (1,1,0) -- (0,1,0) -- cycle;
\end{scope}
\end{scope}
\draw (-5, 1) node {$X\colon$};
\draw (-5, -3) node {$Y\colon$};
}
Let $V = \{ \omega \in \{0,1\}^k \colon \omega \subseteq \omega_0 \}$; clearly this is a discrete sub-cube of dimension $k' = |\omega_0|$ (shown as the dotted square in the diagram, where $k'=2$). Then $A|_{V \setminus \{\omega_0\}}$ is a $k'$-corner, and furthermore a partial lift of the corresponding cube $B|_V$ of $Y$.
By the fibration definition we can choose $x \in X$ such that $f(x) = B(\omega_0)$, and such that setting $A(\omega_0) = x$ we have $A|_V \in C^{k'}(X)$. Hence this extended $A$ satisfies (i) and (ii) as required.
\end{proof}
We can also use this to give a full proof of Proposition \ref{prop:fibrant-glueing}; recall this stated that a fibrant space satisfies the glueing property.
\begin{proof}[Proof of Proposition \ref{prop:fibrant-glueing}]
Essentially we formalize the diagram given above when $k=2$; for convenience we reproduce it now.
\glueingpic
Formally: given $c, c' \in C^k(X)$ with $c(\omega 1) = c'(\omega 0)$ for all $\omega \in \{0,1\}^{k-1}$, let $S = \{0,1\}^{k+1} \setminus \left(\{0,1\}^{k-1} \times \{(1, 1)\}\right)$ and consider the configuration
\begin{align*}
d \colon S &\to X \\
\omega 0 0 & \mapsto c(\omega 1) = c'(\omega 0) \\
\omega 0 1 & \mapsto c'(\omega 1) \\
\omega 1 0 & \mapsto c(\omega 0) \ .
\end{align*}
One can verify this satisfies the hypotheses of Corollary \ref{cor:fibrant-extension-property}, and so there is a cube $\tilde{d} \in C^{k+1}(X)$ whose restriction to $S$ is $d$. But then the ``glued'' configuration $c''$ as in Definition \ref{defn:glueing} is a sub-cube of $\tilde{d}$ as shown, and so is in $C^k(X)$ by the cubespace axioms.
\end{proof}
It is clear from the definition that a composition of fibrations is a fibration.
Another useful fact is the following universal property.
\begin{lemma}[``Universal property'']
\label{lem:universal1}
Let $f_{YX}:X\to Y$ and $f_{ZX}:X\to Z$ be fibrations between compact cubespaces.
Suppose that for every $y\in Y$ there is $z\in Z$ such that $f_{YX}^{-1}(y)\subseteq f_{ZX}^{-1}(z)$.
Then there is a unique fibration $f_{ZY}: Y\to Z$ such that $f_{ZX}=f_{ZY}\circ f_{YX}$.
Equivalently, the following holds.
Let $f:X\to Y$ be a fibration and $g: Y\to Z$ be a map between two compact cubespaces.
If $g\circ f$ is a fibration then so is $g$.
\end{lemma}
\begin{proof}
Since $X$, $Y$ and $Z$ are compact metric spaces, $f_{YX}$ and $f_{ZX}$ are quotient maps.
Hence the map $f_{ZY}$ (which is uniquely defined thanks to the condition imposed on the fibres
of $f_{YX}$ and $f_{ZX}$) is continuous.
It remains to show that it is also a fibration, which is precisely the second statement.
For the second statement, we first show that $g$ is a cubespace morphism.
To this end, fix a cube $c\in C^k(Y)$ and let $\widetilde c\in C^k(X)$ be a cube such that $f(\widetilde c)=c$.
This exists since $f$ is a fibration.
This implies that $g(c)=g\circ f(\widetilde c)\in C^{k}(Z)$, hence $g$ is indeed a cubespace morphism.
Suppose $\lambda$ is a $k$-corner in $Y$ and $c \in C^k(Z)$ a compatible $k$-cube. By Lemma \ref{extension-property} on $f$ with $S = \emptyset$, $T = \{0,1\}^k \setminus \{\vec{1}\}$, we can choose $\tilde{\lambda}$ a $k$-corner of $X$ such that $f(\tilde{\lambda}) = \lambda$ and so in particular $g \circ f(\tilde{\lambda})$ is compatible with $c$.
Since $g \circ f$ is a fibration, we may extend $\tilde{\lambda}$ to a cube $c' \in C^k(X)$ such that $g \circ f(c') = c$. So, $f(c') \in C^k(Y)$ has the required property.
\end{proof}
\begin{remark}
Taking $Z=\{*\}$ in the lemma, we see that if $f:X\to Y$ is a fibration and $X$ is fibrant then so is $Y$.
It also holds that the property of $s$-uniqueness is inherited by the image of a cubespace under a fibration, hence the image of a nilspace under a fibration is a nilspace.
Indeed, let $f:X\to Y$ be a fibration and suppose that $X$ has $s$-uniqueness.
Suppose that $x\sim_s y$ for some $x,y\in Y$.
We aim to show that $x=y$.
We first note that the configuration $c:\{0,1\}^s\to Y$ defined by
$\omega\mapsto x$ for $\omega\neq\vec1$ and $\vec1\mapsto y$ is a cube in $C^s(Y)$.
Let $\widetilde x\in f^{-1}(x)$ be arbitrary and let $\lambda:\{0,1\}^s\backslash\{\vec1\}\to X$ be the constant
corner configuration $\omega\mapsto \widetilde x$.
Since $f$ is a fibration, $\lambda$ can be completed to a cube $\widetilde c$ such that $f(\widetilde c(\vec 1))=y$.
Since $X$ has $s$-uniqueness, we have $\widetilde x=\widetilde c(\vec 1)$, which implies $x=f(\widetilde x)=f(\widetilde c(\vec 1))=y$,
as required.
\end{remark}
We also need to record the correct ``relative'' version of uniqueness.
\begin{definition}
We say a cubespace morphism $f \colon X \to Y$ has \emph{$k$-uniqueness}\index{relative(ly)!$k$-uniqueness} if the following holds: if $c, c' \in C^k(X)$ are two cubes such that $f(c) = f(c')$ and $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{1}$, then in fact $c = c'$.
If $f$ is a fibration and $k$ is the smallest number such that $f$ has $k$-uniqueness, we say that $f$ has \emph{degree $(k-1)$}.
\index{fibration!degree}
\end{definition}
Again it is clear that a space $X$ has $k$-uniqueness if and only if the map $X \to \{\ast\}$ does.
\subsection{The structure theory in terms of fibrations}
We will now modify -- and prove -- statements of the weak structure theory discussed above, in ``relative form'', i.e.~in terms of general fibrations.
First we consider the \emph{canonical factors} $\pi_s$. Everything here is a reasonably straightforward generalization of the corresponding arguments in Section \ref{sec:canonical-factors}.
\begin{definition}
Let $f \colon X \to Y$ be a fibration, and let $k \ge 0$. Define an equivalence relation $\sim_{f, s}$\index[nota]{$\sim_{f, s}$} on $X$ as follows: $x \sim_{f, s} x'$ if there exist two $(s+1)$-cubes $c, c'$ in $X$ such that $f(c)=f(c')$, $c(\vec{1}) = x$, $c'(\vec{1}) = x'$ and $c(\omega) = c'(\omega)$ at all other vertices $\omega \in \{0,1\}^{s+1} \setminus \{\vec{1}\}$.
\end{definition}
\begin{proposition}
\label{relative-canonical-factor}
Suppose $f \colon X \to Y$ is a morphism between compact cubespaces $X$, $Y$ that have the glueing property.
Then the relation $\sim_{f, s}$ is a closed equivalence relation. Moreover a ``universal replacement''\index{relative(ly)!universal replacement property} property holds: if $x \sim_{f, s} x'$, $k \le s+1$ and $c \in C^{k}(X)$ is a cube with $c(\vec{1}) = x$, then the configuration $c'$ given by
\[
c'(\omega) = \begin{cases} x' &\colon \omega = \vec{1} \\ c(\omega) &\colon \omega \ne \vec{1} \end{cases}
\]
is a cube.
Now suppose further that $f$ is a fibration. Writing $\pi = \pi_{f,s} \colon X \to X / \sim_{f,s}$\index[nota]{$\pi_{f,s}$} for the projection map, we have that $\pi$ is a fibration and $f$ factors as $f \colon X \xrightarrow{\pi} X / \sim_{f, s} \xrightarrow{g} Y$ where $g$ is a fibration of degree at most $s$.
\end{proposition}
\begin{remark}\label{rem:completion for quotient}
Again we can consider the case $Y = \{\ast\}$ and $f \colon X \to \{\ast\}$ is the unique map. So, $f$ is a fibration if and only if $X$ is a fibrant cubespace. In this case, Proposition \ref{relative-canonical-factor} exactly restates Proposition \ref{prop-canonical-factors}.
Note in particular that we now have a proof that $X / \sim_s$ is fibrant (since $g \colon X / \sim_s \to \{\ast\}$ is a fibration), which was previously omitted.
For general $X$ and $Y$, since $x \sim_{f,s} y$ precisely if $x \sim_s y$ and $f(x) = f(y)$, the equivalence relation and universal replacement statements in Proposition \ref{relative-canonical-factor} are easily deducible from the non-relative case $Y = \{\ast\}$; i.e.\ these follow from Proposition \ref{prop-canonical-factors}. However, since we gave only a proof sketch of these parts in the non-relative case, we lose nothing by staring over in full generality.
\end{remark}
Notwithstanding the increased generality, all the key ideas required for this result are contained in the sketches in Section \ref{sec:canonical-factors}.
We will first introduce some symbology for manipulating high-dimensional cubes.
\begin{definition}
If $c$ and $c'$ are $k$-cubes, we will denote by $[c, c']$\index[nota]{$[c, c']$} the $(k+1)$-configuration
\[
\omega \mapsto \begin{cases} c(\omega_1 \dots \omega_k) &\colon \omega_{k+1} = 0 \\ c'(\omega_1 \dots \omega_k) &\colon \omega_{k+1} = 1 \ . \end{cases}
\]
Given an element $x \in X$, the notation $\square^k(x)$\index[nota]{$\square^k(x)$} denotes the constant $k$-cube $(\omega \mapsto x)$. Given $x, y \in X$, we denote by $\llcorner^k(x;y)$\index[nota]{$\llcorner^k(x;y)$} the configuration
\begin{align*}
\{0,1\}^k &\to X \\
\omega &\mapsto \begin{cases} x &\colon \omega \ne \vec{1} \\ y &\colon \omega = \vec{1} \ . \end{cases}
\end{align*}
\end{definition}
We may combine these pieces of notation freely with each other and also with pictorial representations of cubes. For instance, the notation $[\llcorner^2(x;y), [\square^1(z), \square^1(w)]]$ is a synonym for
\inlinetikz{
\threecube{$x$}{$x$}{$x$}{$y$}{$z$}{$z$}{$w$}{$w$}{1}
}
and
\inlinetikz{
\begin{scope}[shift={(-3,0.6)}]
\singlesquare{$\square^1(x)$} {$\square^1(y)$} {$\square^1(z)$} {$\llcorner^1(w;w')$}
\end{scope}
\draw (0,1) node {$=$};
\begin{scope}[shift={(1,0)}]
\threecube{$x$}{$y$}{$x$}{$y$}{$z$}{$w$}{$z$}{$w'$}{1}
\end{scope}
}
although we will rarely need to denote anything quite so unpleasant.
We will be generally cavalier about the ordering of the indices $\{1, \dots, k\}$ implied by successive applications of this notation, since it should always be clear from context what is meant, and because (thanks to the axioms) it makes little or no difference most of the time.
Finally, we will need to introduce some notation about \emph{tri-cubes} of the form appearing in the proof of Proposition \ref{prop-canonical-factors}, i.e.~a collection of $2^k$ $k$-cubes that all glue together at the middle. These are very useful gadgets introduced by Antol\'\i n Camarena and Szegedy in \cite{CS12}*{Section 2.3}.
\begin{definition}
We say a collection $t = (t_\nu)_{\nu \in \{0,1\}^k}$\index[nota]{$(t_\nu)_{\nu \in \{0,1\}^k}$} constitutes a \emph{tri-cube}\index{tri-cube} if $t_\nu \in C^k(X)$ for each $\nu$, and if for each $\nu, \nu', \omega \in \{0,1\}^k$ such that for each $i \in \{1,\dots,k\}$ either $\nu_i = \nu'_i$ or $\omega_i = 1$, we have $t_\nu(\omega) = t_{\nu'}(\omega)$.
We say the \emph{outer cube}\index{outer cube} of $t$ is the configuration $\omega \mapsto (t_\omega(\vec{0}))$.
\end{definition}
Note this corresponds to a picture (when $k=2$):
\inlinetikz{
\begin{scope}[scale=2]
\draw[thin,black] (0,0) grid (2, 2);
\node [below left] at (0, 0) {$t_{00}(00)$};
\node [above left] at (0, 2) {$t_{01}(00)$};
\node [below right] at (2, 0) {$t_{10}(00)$};
\node [above right] at (2, 2) {$t_{11}(00)$};
\node [below left] at (0, 1) {$t_{00}(01)$};
\node [above left] at (0, 1) {$t_{01}(01)$};
\node [below right] at (2, 1) {$t_{10}(01)$};
\node [above right] at (2, 1) {$t_{11}(01)$};
\node [below left] at (1, 0) {$t_{00}(10)$};
\node [below right] at (1, 0) {$t_{10}(10)$};
\node [above left] at (1, 2) {$t_{01}(10)$};
\node [above right] at (1, 2) {$t_{11}(10)$};
\node [below left] at (1, 1) {$t_{00}(11)$};
\node [below right] at (1, 1) {$t_{10}(11)$};
\node [above left] at (1, 1) {$t_{01}(11)$};
\node [above right] at (1, 1) {$t_{11}(11)$};
\end{scope}
}
and hence some of the $t_\nu$ are reflected relative to the standard orientation.
\begin{proposition}
If $X$ is a cubespace with the glueing property and $(t_\nu)$ is a tri-cube in $X$, then the outer cube is in $C^k(X)$.
\end{proposition}
\begin{proof}
This follows from repeated application of the glueing property. By assumption we can glue together each pair $t_{\eta 0}$ and $t_{\eta 1}$ for $\eta \in \{0,1\}^{k-1}$, and then repeat for each of the other $(k-1)$ coordinates in turn.
\end{proof}
\begin{proof}[Proof of Proposition \ref{relative-canonical-factor}]
The proof is strongly analogous to that from Section \ref{sec:canonical-factors}. We first prove an analogue of Lemma \ref{lem:canonical-characterization}.
\begin{lemma}
\label{lem:relative-canoncial-characterization}
We have that $x \sim_{f,s} y$ if and only if $f(x) = f(y)$ and also $\llcorner^{s+1}(x; y) \in C^{s+1}(X)$.
\end{lemma}
\begin{proof}
The strategy is to construct a tri-cube whose outer cube is $\llcorner^{s+1}(x; y)$.
By assumption we are given cubes $c, c' \in C^{s+1}(X)$ such that $f(c) = f(c') = d$, $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{0}$ and $c(\vec{0}) = x$, $c'(\vec{0}) = y$. (Note we have switched $\vec{1}$ for $\vec{0}$ for convenience.) From this it is immediate that $f(x) = f(y)$.
Now define the family $(t_\nu)_{\nu \in \{0,1\}^{s+1}}$ by
\[
t_\nu = \begin{cases} c &\colon \nu \ne \vec{1} \\ c' &\colon \nu = \vec{1} \ . \end{cases}
\]
It is clear from the hypotheses that this is a tri-cube with outer cube $\llcorner^{s+1}(x; y)$ as required.
\end{proof}
\begin{claim*}
The relation $\sim_{f,s}$ is an equivalence relation.
\end{claim*}
\begin{proof}[Proof of claim]
Again, symmetry and reflexivity are immediate. For transitivity, assume $x \sim_{f,s} y \sim_{f,s} z$. Then by the lemma, $f(x) = f(y) = f(z)$ and $\llcorner^{s+1}(y; x)$, $\llcorner^{s+1}(y; z)$ are cubes. By glueing along the common face $\square^s(y)$, we find that $\llcorner^s(\square^1(y); [x,z])$ is an $(s+1)$-cube. But so is $\llcorner^s(\square^1(y); [x,x])$, by restricting to a face of $\llcorner^{s+1}(y; x)$ and then duplicating. Since these latter two cubes have the same image under $f$, we have $x \sim_{f,s} z$ by the definition.
\end{proof}
Again, the fact that this relation is closed is straightforward from the definition, the hypothesis that $C^{s+1}(X)$ is closed, continuity of $f$ and compactness.
We turn to the ``universal replacement'' statement.
\begin{claim*}
If $x \sim_{f, s} x'$, and $c \in C^{s+1}(X)$ is a cube with $c(\vec{1}) = x$, then the configuration $c'$ given by
\[
c'(\omega) = \begin{cases} x' &\colon \omega = \vec{1} \\ c(\omega) &\colon \omega \ne \vec{1} \end{cases}
\]
is a cube.
\end{claim*}
As in the proof of Proposition \ref{prop-canonical-factors}, this easily implies the same statement for all $k < s+1$.
\begin{proof}[Proof of claim]
We construct a tri-cube whose outer cube is $c'$. We set
\[
t_\nu(\omega) = \begin{cases} c((\min(\nu_i + \omega_i, 1))_{i=1}^{s+1}) &\colon \nu \ne \vec{1} \\ x &\colon \nu = \vec{1}, \omega \ne \vec{0} \\ x' &\colon \nu = \vec{1}, \omega = \vec{0} \ . \end{cases}
\]
which is a very elaborate way of stating the generalization of the diagram from the proof of Proposition \ref{prop-canonical-factors}. Now, $t_{\vec{0}}$ is just $c$, the other $t_\nu$ except $t_{\vec{1}}$ are obtained from $c$ by duplication of an upper face, and $t_{\vec{1}}$ is a reflected copy of $\llcorner^{s+1}(x; x')$. So, all the $t_\nu$ are cubes, and a tedious check verifies that $(t_\nu)$ is a tri-cube with outer cube $c'$ as required.
\end{proof}
So, we have a projection $\pi_{f,s} \colon X \to X / \sim_{f,s}$, and by construction $f$ factors through $\pi_{f,s}$, i.e.~$f = g \circ \pi_{f,s}$ for some well-defined function $g \colon X / \sim_{f,s} \to Y$.
\begin{claim*}
The map $g$ has $(s+1)$-uniqueness.
\end{claim*}
\begin{proof}[Proof of claim]
Indeed, given two $(s+1)$-cubes $c, c'$ of $X / \sim_{f,s}$ with $c(\omega) = c'(\omega)$ for all $\omega \ne \vec{1}$ and $g(c) = g(c')$, we know there are cubes $\tilde{c}, \tilde{c}'$ of $X$ such that $\pi_{f,s}(\tilde{c}) = c$, $\pi_{f,s}(\tilde{c}') = c'$. By repeated application of universal replacement we may assume $\tilde{c}(\omega) = \tilde{c'}(\omega)$ for each $\omega \ne \vec{1}$. But then $\tilde{c}(\vec{1}) \sim_{f,s} \tilde{c}'(\vec{1})$ by the definition of $\sim_{f,s}$, and so $c = c'$.
\end{proof}
\begin{claim*}
The map $\pi_{f,s}$ is a fibration.
\end{claim*}
\begin{proof}[Proof of claim]
Suppose we are given a $k$-corner $\lambda$ in $X$ and a compatible $c \in C^k(X) / \sim_{f,s}$.
First consider the case $k < s+1$. By definition there is a cube $\tilde{c}$ of $X$ with $\pi_{f,s}(\tilde{c}) = c$, and furthermore by repeated application of universal replacement, \emph{any} configuration $\tilde{c} \colon \{0,1\}^k \to X$ with $\pi_{f,s} \circ \tilde{c} = c$ is a cube. Hence taking $\lambda(\vec{1})$ to be an arbitrary point of $\pi_{f,s}^{-1}(c(\vec{1}))$ must work.
If $k \ge s+1$, we may use the fact that $f$ is a fibration to complete $\lambda$ to a cube $\tilde{c}$ such that $f(\tilde{c}) = g(c)$. But now $\pi_{f,s}(\tilde{c})$ and $c$ are two cubes of $X / \sim_{f,s}$ whose image under $g$ is the same, and which are equal except possibly at the vertex $\vec{1}$. Now we may restrict $\pi_{f,s}(\tilde{c})$ and $c$ to an upper face $F$ of $\{0,1\}^k$ of dimension $(s+1)$ (if necessary) and use $(s+1)$-uniqueness of $g$ to deduce that $\pi_{f,s}(\tilde{c}(\vec1)) = c(\vec1)$, and the claim follows.
\end{proof}
We now note that $g$ is therefore a fibration by the universal property (Lemma \ref{lem:universal1}). This completes the proof of Proposition \ref{relative-canonical-factor}.
\end{proof}
\subsection{The (relative) structure groups}
With this in place, we are in a position to attack the relative analogue of Theorem \ref{basic-cs-structure-theorem}. First, we need one further ``relative'' definition.
\begin{definition}
A morphism $f \colon X \to Y$ is called \emph{(relatively) $k$-ergodic} if for any $1 \le \ell \le k$ and $c \in C^\ell(Y)$, any configuration in $f^{-1}(c)$ is a cube of $X$.
We say it is \emph{relatively ergodic}\index{relative(ly)!ergodic} if it is relatively $1$-ergodic.
\end{definition}
The direct analogue of Theorem \ref{basic-cs-structure-theorem} is as follows.
\begin{theorem}
\label{relative-structure-thm}
Suppose $X$ and $Y$ are compact ergodic cubespaces that obey the glueing condition. Let $f \colon X \to Y$ be a relatively $s$-ergodic fibration of degree at most $s$.
Then there exists a compact Abelian group $A=A(f)$\index[nota]{$A(f)$} acting continuously on $X$ such that:
\begin{enumerate}
\item the action of $A$ on $X$ is free, and its orbits are precisely the fibers of $f$;
\item this induces a (free) pointwise action of $C^k(\cD_s(A))$ on $C^k(X)$, whose orbits are precisely the fibers of the map $f \colon C^k(X) \to C^k(Y)$.
\end{enumerate}
\end{theorem}
We will see that the proof is constructive and $A$ is defined canonically in terms of the cube structures and $f$.
The key point about these hypotheses is that they hold whenever $f$ is the canonical projection map $\pi_{g,s-1} \colon \pi_{g,s}(X) \to \pi_{g,s-1}(X)$ defined with respect to some other fibration $g \colon X \to Y$. In this case, we write $A = A_s(g)$, the ``$s$-th structure group of the fibration''. In this way we can iterate this theorem to obtain:
\begin{corollary}
\label{relative-tower}
Let $f\colon X \to Y$ be a fibration of degree at most $s$ between compact ergodic cubespaces $X$, $Y$ that obey the glueing condition. Then we have a tower
\[
X = \pi_{f,s}(X) \xrightarrow{\pi_{f,s-1}} \pi_{f,s-1}(X) \to \dots \xrightarrow{\pi_{f,0}} \pi_{f,0}(X) = Y
\]
where the fibers of $\pi_{f,i-1}$ are identified with the compact abelian group $A_i(f)$ for each $i$, in the sense of Theorem \ref{relative-structure-thm}.
\end{corollary}
\begin{proof}[Proof of Corollary \ref{relative-tower} from Theorem \ref{relative-structure-thm}]
It suffices to know that each map $\pi_{f,t} \colon \pi_{f, t+1}(X) \to \pi_{f, t}(X)$ is well-defined, a fibration, has $(t+2)$-uniqueness and is $(t+1)$-ergodic.
Well-definedness follows from the nested nature of the relations $\sim_{f,t}$; the proof of this is unchanged from Proposition \ref{equiv-nested}.
The fact that it is a fibration follows from the fact that $\pi_{f,t+1} \colon X \to \pi_{f,t+1}(X)$ and $\pi_{f,t} \colon X \to \pi_{f,t}(X)$ are fibrations, and the universal property (Lemma \ref{lem:universal1}).
The $(t+2)$-uniqueness statement is supplied by Proposition \ref{relative-canonical-factor} applied to $\pi_{f,t+1}(X)$.
Finally, the $(t+1)$-ergodicity statement is precisely what is encoded by the ``universal replacement'' property from that same proposition as long as $t\geq 1$. For $t=0$ it does not follow from Proposition \ref{relative-canonical-factor} (indeed in general it does not hold). Instead for $t=0$ we use that assumption that $X$ is ergodic.
\end{proof}
Before embarking on the proof of Theorem \ref{relative-structure-thm}, we note a relative analogue of Proposition \ref{prop:high-cubes-boring} that will be useful for discussing high-dimensional cubes.
\begin{proposition}
\label{prop:relative-boring-high-dimension}
Let $s \ge 1$ be an integer, $f \colon X \to Y$ a fibration of degree at most $s$ and $k \ge s +1$. Further let $c \colon X \to \{0,1\}^k$ be given. Then $c \in C^k(X)$ if and only if $f(c)$ is a cube of $Y$ and further every face of $c$ of dimension $(s+1)$ is a cube of $X$.
\end{proposition}
\begin{proof}
The ``only if'' part is clear. For the converse, again we argue by induction on $k$, the case $k=s+1$ being trivial. Given $c$, we consider the restriction to $\{0,1\}^k \setminus \{\vec{1}\}$, which by inductive hypothesis is a $k$-corner. Hence we may complete this corner relative to $f(c)$; i.e.~there is a cube $c'$ such that $c'(\omega) = c(\omega)$ for all $\omega \ne \vec{1}$ and $f(c'(\vec{1})) = f(c(\vec{1}))$. Now by restricting to any upper face of $c$ and $c'$ of dimension $(s+1)$ and invoking $(s+1)$-uniqueness, we find $c(\vec{1}) = c'(\vec{1})$ also and hence $c = c'$ is a cube.
\end{proof}
The remainder of this section is spent proving Theorem \ref{relative-structure-thm}.
\begin{proof}[Proof of Theorem \ref{relative-structure-thm}]
Again, this account is very similar to the proof of Theorem \ref{basic-cs-structure-theorem}, and the reader will find it significantly more comprehensible if they are already familiar with the sketch proof of that theorem given above in Section \ref{sec:structure-groups}.
Before, we defined the structure group in terms of equivalence classes of edges in $C^1(X)$, under the equivalence relation $[x,y] \sim [z,w]$ if and only if $[[x,y],[z,w]] \in C^2(X)$.
One way to rationalize this construction is in terms of the \emph{edge cubespace} $\cE(X)$, which we define now.
\begin{definition} \label{defn:edge-cubespace}
Let $X$ be a cubespace. The \emph{edge cubespace}\index{edge cubespace} of $X$, denoted $\cE(X)$\index[nota]{$\cE(X)$}, is the cubespace whose base space consists of the edges $C^1(X) \subseteq X \times X$, and whose $k$-cubes are the configurations $c \in (X \times X)^{\{0,1\}^k} \cong X^{\{0,1\}^{k + 1}}$ that correspond to $(k+1)$-cubes of $X$.
\end{definition}
It is easy to check that this does indeed define a cubespace. Moreover, if $X$ has the glueing property, this is inherited by $\cE(X)$. It follows that we may consider the canonical equivalence relations $\sim_{k}$ on $\cE(X)$; by Proposition \ref{relative-canonical-factor} (in the case $Y = \{\ast\}$; see Remark \ref{rem:completion for quotient}) these are equivalence relations and have the universal replacement property.
It follows that $[x,y] \sim [z,w]$ as above if and only if $[[x,y],[z,w]]$ is an edge of $\cE(X)$, hence if and only if $[x,y] \sim_0 [z,w]$ where $\sim_0$ denotes the canonical equivalence relation on $\cE(X)$.
In order to generalize the proof of Theorem \ref{basic-cs-structure-theorem} to $s > 1$, it is therefore tempting to try replacing $\sim_0$ by $\sim_{s-1}$ in the above and attempting to re-run the argument. A little further thought shows this is not quite the right thing to do, since
\begin{itemize}
\item the space $\cE(X)$ is too large when $s > 1$: we want a pair to represent the element of the structure group that takes one to the other, but this only makes sense for pairs $[x,y]$ that lie in the same fiber of $f$;
\item the condition $[x,y] \sim_{s-1} [x',y']$ turns out to be too restrictive on $Y$, since it implies that $x$ and $x'$ lie in the same fiber of $f$, and hence we cannot relate pairs lying over different fibers.
\end{itemize}
However, a second attempt at this does work, and generalizes Theorem \ref{basic-cs-structure-theorem} fairly cleanly.
We define
\[
\cM= \cM_f(X) = \{ [x,y] \in X\times X \colon f(x) = f(y) \}
\]
i.e.~the space of pairs that lie in the same fiber of $f$, and specify an equivalence relation $\sim$ on $\cM$ by $[x,x'] \sim [y, y']$ if and only if $[x,y] \sim_{s-1} [x',y']$ in $\cE(X)$, where $\sim_{s-1}$ denotes the canonical equivalence relation on that space. Note that this definition involves a ``$90^\circ$ rotation'', e.g.~pairing up $[x,y]$ rather than $[x,x']$; this is not an error and is necessary for the definition to make sense.
Note that, by Lemma \ref{lem:relative-canoncial-characterization}, this is equivalent to saying $[x,x'] \sim [y,y']$ if and only if $[\llcorner^s(x;x'), \llcorner^s(y;y')]$ is a cube of $X$.
We will define the structure group $A = A_f$ to be $\cM / \sim$. As above, addition on $A$ will be given by concatenation of edges, negation by $[x,y] \mapsto [y,x]$ and the action on $X$ by a cube completion.
We now turn to the details.
\begin{claim*}
The relation $\sim$ is a closed equivalence relation on $\cM$.
\end{claim*}
\begin{proof}[Proof of claim]
For reflexivity we are asked to verify that for all $[x, x'] \in \cM$ the configuration $[\llcorner^s(x; x'), \llcorner^s(x; x')]$ is an $(s+1)$-cube of $X$; equivalently by duplication, that $\llcorner^s(x; x')$ is an $s$-cube. Since $f(x) = f(x')$ this is immediate from the assumption of relative $s$-ergodicity.
Symmetry is clear: if $[\llcorner^s(x; x'), \llcorner^s(y; y')]$ is a cube then so is $[\llcorner^s(y; y'), \llcorner^k(x; x')]$.
For transitivity, suppose $[x, x'], [y, y'], [z, z'] \in \cM$ and $[x, x'] \sim [y, y']$, $[y, y'] \sim [z, z']$. Equivalently, $[\llcorner^s(x; x'), \llcorner^s(y; y')]$ and $[\llcorner^s(y; y'), \llcorner^s(z; z')]$ are cubes. It is immediate by glueing that $[\llcorner^s(x;x'), \llcorner^s(z;z')]$ is a cube, and hence $[x,x'] \sim [z,z']$ as required.
Once again, the topological properties are an easy check.
\end{proof}
Hence, as a set, $A = \cM / \sim$ is well-defined. As in the proof of Theorem \ref{basic-cs-structure-theorem}, we have considerable freedom to choose representatives for $a \in A$.
\begin{claim*}
Given $a \in A$ and $x \in X$, there exists an unique $x' \in X$ such that $a \sim [x,x']$.
\end{claim*}
\begin{proof}[Proof of claim]
Suppose $[y,y']$ is any representative of $a$. Necessarily $f(y) = f(y')$. Hence the configuration $[\llcorner^s(y; y'), \llcorner^s(x; -)]$ is an $(s+1)$-corner of $X$ lying above the cube $[\square^s(f(y)), \square^s(f(x))]$ of $Y$ (which is a cube by ergodicity of $Y$ and duplication): the fact that it is a corner is trivial by $s$-ergodicity of $f$. Since $f$ is a fibration of degree at most $s$, there is an unique $x' \in f^{-1}(f(x))$ completing this $(s+1)$-corner to a cube of $X$, as required.
\end{proof}
Hence we may now define addition on $A$ by concatenation of edges as before. Given $a, b \in A$ we may fix $e \in X$ and choose representatives $[x, e] \sim a$, $[e, y] \sim b$ and define $a + b$ to be the class of $[x,y]$. The identity is the class of $[e,e]$ for any $e$, and negation is done by sending $[x,y] \mapsto [y,x]$.
\begin{claim*}
These operations are well-defined, and give $A$ the structure of a (topological) abelian group.
\end{claim*}
\begin{proof}[Proof of claim]
We first check well-definedness: that is, that these operations do not depend on the choice of representatives, or -- in the case of addition -- on the choice of $e$.
It is clear from the definition of $\sim$ (in terms of $\sim_s$ in $\cE(X)$) that $[x,x] \sim [y,y]$ for any $x,y \in X$ and hence the identity class really is a class. Similarly, if $[x,y] \sim [x',y']$ then $[y,x] \sim [y',x']$ and hence negation is well-defined.
Now suppose $[x,y] \sim [x',y']$ and $[y,z] \sim [y',z']$; we wish to show $[x,z] \sim [x',z']$ as this will show addition is well-defined. But this says $[x,x'] \sim_{s-1} [y,y']$ and $[y,y'] \sim_{s-1} [z,z']$ in $\cE(X)$, and hence $[x,x'] \sim_{s-1} [z,z']$ in $\cE(X)$ by transitivity and so $[x,z] \sim [x',z']$ as required.
Associativity of addition is now clear from associativity of concatenation. Similarly, that the identity and inverses behave as they should is straightforward.
We now argue commutativity. We will first need a lemma.
\begin{lemma}
If $x,y,z,w \in X$ all lie in the same fiber of $f$ and $[x,y] \sim [z,w]$ then $[x,z] \sim [y,w]$.
\end{lemma}
\begin{proof}[Proof of lemma]
Note if $s=1$ this is trivial: both are equivalent to
\inlinetikz{\singlesquare{$x$}{$y$}{$z$}{$w$}}
being a cube.
For $s>1$, we argue as follows.
Choose $u$ in the same fiber arbitrarily.
By relative $s$-ergodicity,
\inlinetikz{\singlesquare{$\llcorner^{s-1}(u;x)$}{$\llcorner^{s-1}(u;x)$}{$\llcorner^{s-1}(u;z)$}{$\llcorner^{s-1}(u;z)$}}
is an $(s+1)$-cube in $X$, or equivalently an $s$-cube in $\cE(X)$.
Note that $[x,y]\sim[z,w]$ if and only if $[x,z]\sim_{s-1}[y,w]$ in $\cE(X)$. In the previous diagram, interpreted as an $s$-cube in $\cE(X)$, the edge $[x,z]$ appears twice. If we replace one of these edges by $[y,w]$, giving
\inlinetikz{\singlesquare{$\llcorner^{s-1}(u;x)$}{$\llcorner^{s-1}(u;y)$}{$\llcorner^{s-1}(u;z)$}{$\llcorner^{s-1}(u;w)$}}
then this lies in $C^s(\cE(X))$ if and only if $[x,z] \sim [y,w]$. Indeed, ``only if'' follows from the definition of $\sim_{s-1}$ on $\cE(X)$, and for ``if'' we apply the universal replacement property for $\cE(X)$ (which has the glueing property, as was mentioned after Definition \ref{defn:edge-cubespace}).
However, thinking of this configuration now as an element of $X^{\{0,1\}^{s+1}}$, it is unchanged, up to a morphism of discrete cubes, on exchanging $y$ and $z$. Hence, we see that it is a cube in $C^{s+1}(X)$ if and only if $[x,z]\sim[y,w]$, and also if and only if $[x,y] \sim [z,w]$, as required.
\end{proof}
So, now suppose we have $a, b \in A$, and pick representatives $a \sim [x,e]$, $b \sim [e,y]$ and hence $a+b \sim [x,y]$. Further pick $e'$ such that $[x,e'] \sim b$. Note $x,y,e,e'$ all lie in the same fiber and $[x,e'] \sim [e,y]$, so by the lemma $[x,e] \sim [e',y]$. But then by concatenation,
\[
b + a \sim [x,e'] + [e',y] \sim [x,y] \sim a + b
\]
as required.
Once again, arguing continuity of these operations is straightforward from closedness of the equivalence relations, closedness of $C^k(X)$, continuity of $f$ and compactness.
This concludes the proof of the claim.
\end{proof}
We now define the group action of $A$ on $X$. Given $a \in A$ and $x \in X$ we take $a(x)$ (also notated $a . x$) to be the unique element of $X$ such that $a \sim [x, a(x)]$. It is clear $0 \in A$ acts trivially. To confirm it is an action, we consider that $a \sim [x, a(x)]$ and $a' \sim [a(x), a'(a(x))]$ but then by definition of addition $a+a' \sim [x, a'(a(x))]$.
Moreover, it is clear that this action respects $f$ (i.e.~$x$ and $a(x)$ lie in the same fiber), and that it is simply transitive on fibers of $f$. Again, continuity of the action is straightforward. Hence we have shown part (i) of Theorem \ref{relative-structure-thm}.
We now turn to part (ii), which describes the cubes $C^k(X)$ in terms of $C^k(Y)$ and $A$. Recall that we wanted to show that the action of $A$ on $X$ induces a pointwise action of $C^k(\cD_s(A))$ on $C^k(X)$ whose orbits are precisely the fibers of $f \colon C^k(X) \to C^k(Y)$. This can be rephrased as follows.
\begin{lemma}\label{lem:relative cubes}
Suppose $k\ge 0$, $c \in C^k(X)$ and $c' \colon \{0,1\}^k \to X$ is another configuration such that $f(c) = f(c')$. Define a configuration $a \colon \{0,1\}^k \to A$ by
\[
a(\omega) \sim [c(\omega), c'(\omega)] \ .
\]
Then $c'$ is a cube if and only if $a \in C^{k}(\cD_s(A))$.
\end{lemma}
Recall again that the cubespace $\cD_s(A)$ was defined by the Host--Kra construction applied to the compact abelian group $A$ with the filtration $A_0 = \dots = A_s \supseteq A_{s+1} = \{0\}$. We note the following features:
\begin{itemize}
\item it is $s$-ergodic, i.e.~every element of $A^{\{0,1\}^k}$ is a cube for $0 \le k \le s$; and
\item it is a nilspace of degree $s$.
\end{itemize}
\begin{proof}[Proof of Lemma \ref{lem:relative cubes}]
We will consider three cases.
\setcounter{case}{0}
\begin{case}
$0 \le k \le s$. By relative $s$-ergodicity, we have that $c'$ is always a cube of $X$. But by $s$-ergodicity of $\cD_s(A)$, $a$ is always a $k$-cube, so this is consistent.
\end{case}
\begin{case}
$k = s+1$. For fixed $c$, let $T \subseteq A^{\{0,1\}^{s+1}}$ denote the set of all elements $a \colon \{0,1\}^{s+1} \to A$ such that $c' = a . c$ is a cube; so our goal is to show $T = C^{s+1}(\cD_s(A))$. Then we claim
\begin{enumerate}
\item for $a \in T$, and for $F \subseteq \{0,1\}^{s+1}$ a face of dimension $1$ and $b \in A$, we have $a + [b]_F \in T$; \footnote{Recall (Definition \ref{def:cube-group}) that $[b]_F$ denotes a configuration that is equal to $b$ on $F$ and zero elsewhere.}
\item if $a \in T$ and $a(\omega) = 0$ for all $\omega \ne \vec{1}$, then $a = 0$.
\end{enumerate}
To see (i), note that we can think of any cube $c'$ of dimension $(s+1)$ as a cube of dimension $s$ in the edge cubespace $\cE(X)$. Let $[x,y]$ denote the edge corresponding to $F$ in $c'$. Then $[x,b.x] \sim [y,b.y] \sim b$ in $\cM$ and so $[x,y] \sim_{s-1} [b.x,b.y]$ in $\cE$. By the universal replacement property of $\sim_{s-1}$ (Proposition \ref{relative-canonical-factor}) applied to $\cE(X)$ (which certainly has the glueing property), replacing the vertex $[x,y]$ of the $s$-cube $c'$ of $\cE(X)$ by $[b.x,b.y]$ yields another $s$-cube, as required.
Part (ii) is immediate from relative $(s+1)$-uniqueness.
Now, from (i) it follows that $T$ is a union of cosets of $C^{s+1}(\cD_s(A))$, since such $[b]_F$ generate $C^{s+1}(\cD_s(A))$. In particular this includes $C^{s+1}(\cD_s(A))$ itself as trivially $0 \in T$. Now if $a \in T$, by completing the $(s+1)$-corner $a|_{\{0,1\}\setminus \{\vec{1}\}}$ in the nilspace $\cD_s(A)$, we can find an element $a' \in \cD_s(A)$ such that $a'(\omega) = a(\omega)$ for all $\omega \ne \vec{1}$. So $a - a' \in T$; but by (ii) this means $a = a'$ and hence $a \in C^{s+1}(\cD_s(A))$.
\end{case}
\begin{case}
$k > s+1$. Given $c$, $c'$, $a$ as in the statement, we note that $c'$ is a cube if and only if every face of dimension $(s+1)$ is a cube (by Proposition \ref{prop:relative-boring-high-dimension}), if and only if every face of $a$ of dimension $(s+1)$ is a cube (by the previous case), if and only if $a$ is a cube of $\cD_s(A)$ (by Proposition \ref{prop:high-cubes-boring}).
\end{case}
\end{proof}
This concludes the proof of Theorem \ref{relative-structure-thm}.
\end{proof}
|
\section{Introduction}
Pulsars are fast-spinning neutron stars, with a typical mass of 1.4 - 2.0$M_{\odot}$ and radius of $10^6$cm. The rotational period of a newborn
pulsar ranges from 0.01 to 10s. Pulsars are highly magnetized
with magnetic field strength of the order of $10^{12}$G, and emit electromagnetic radiation
as beams. This beam of radiation can be observed in a pulsation which that has the same period as the rotational period of a pulsar. For an isolated pulsar, the radiation is powered by the rotation. As the rotational energy is eventually consumed, the spin period increases steadily. This spin-down rate is very small. The typical value is $10^{-13}$ s s$^{-1}$ for a young pulsar (e.g. the Crab pulsar), corresponding to an increase of 0.001s after one million years.
Although every pulsar decelerates through radiation, it is not rare to observe pulsars that experience spin-up. Glitches are internal processes that generate a sudden jump in the rotational period of a pulsar. The exact causes and processes of a glitch are not yet fully understood. Currently, there are two common proposed models: superfluidity and starquake. The superfluid model, first predicted by Packard (1972), suggests that the structure of the pulsar is composed of normal matter and superfluid components. Occasionally, the superfluid component may couple with the normal matter, causing a transfer of angular momentum to the surface of the pulsar, which increases the observed frequency immediately. Another explanation for the glitch is the starquake model, which was first discussed by Ruderman (1969) after the first glitches in the Crab and Vela pulsars were observed. In this model, it is predicted that when a pulsar spins down, the geometrical shape deforms from oblate toward spherical. Since the crust is solid, such deformation accumulates stress in the crust. When the stress is beyond the maximum shear strain that can be supported, the crust cracks and causes a slight change in the shape of the pulsar, and hence the moment of inertia. As the angular momentum is conserved, the pulsar spins up when the moment of inertia is reduced. Espinoza (2011) reported their analysis on 315 glitches observed in the rotation of 102 pulsars. They found that glitches are most extensive in young pulsars with a characteristic age $\tau_c \sim 10$kyr. The Crab pulsar, which is the youngest pulsar in their study, experienced a relatively small glitch size ($\Delta \nu / \nu < 200 \times 10^{-9}$) but the largest $\left| \dot{\nu} \right|$ among the samples. This small glitch size is more likely to be explained by the starquake model, while for pulsars with larger glitch sizes like the Vela pulsar, having $\Delta \nu / \nu > 1000 \times 10^{-9}$, the explanation has to be obtained with the superfluid model.
In this study, we analyzed the high-energy gamma-ray emission from PSR~J2021+4026. PSR~J2021+4026 is a bright source in gamma rays first discovered by the \textit{Fermi} Gamma-ray Space Telescope (\textit{Fermi}), which is a space observatory launched into a low Earth orbit in 2008. The pulsar is also the first variable gamma-ray seen by the \textit{Fermi} Large Area Telescope (\textit{Fermi}-LAT). There are spectral and timing analyses that indicate the presence of a glitch
in this pulsar. This glitch has several simultaneous effects on the pulsar emissions. We attempt to use a theoretical model and glitch scenario to explain these observed effects. The data analysis of the gamma-ray observations on PSR~J2021+4026 will be presented in Section~\ref{section:observation}. Then, the details of the model used and the model results will be presented in Section~\ref{section:modeling}. Finally, the study is summarized and concluded in Section~\ref{section:conclusion}.
\section{Gamma-ray Observation on PSR~J2021+4026}
\label{section:observation}
\subsection{Overview of PSR~J2021+4026}
PSR~J2021+4026 is a gamma-ray bright pulsar located in the Cygnus region of the Milky Way. Its gamma-ray detection was first discovered by \textit{Fermi}-LAT in the first year of the \textit{Fermi} mission (Abdo et al. 2009a) as 0FGL~J2021.5+4026 with a test-statistic (TS) value of $\sim 4800$, corresponding to a significance of $\sim 70$-$\sigma$, in the energy range from 200 MeV to 100 GeV. The blind frequency search on the LAT data performed by Abdo et al. (2009b) reported that 0FGL~J2021.5+4026 has a rotational period $P=265$ms and a spin-down rate $\dot{P}=5.48 \times 10^{-14}$ s s$^{-1}$. Thus, the source was confirmed to be a pulsar. It has a spin-down characteristic age of $\tau_{c}\sim 77$kyr, an estimated surface magnetic field of $B \sim 4 \times 10^{12}$G and an estimated spin-down luminosity of $\dot{E} \sim 10^{35}$ erg s$^{-1}$.
A detailed study done by Allafort et al. (2013) on the gamma-ray emission from PSR~J2021+4026 revealed a sudden change in the pulsar behavior that is observed by the \textit{Fermi}-LAT. They found that the glitch happened near MJD 55850 (2011 October 16) with a timescale smaller than a week. The change occurred in four aspects (flux, spin-down rate, pulse profile, and spectrum) simultaneously. First, the flux was decreased by $18\%$, from $(8.33 \pm 0.08) \times 10^{-10}$ erg cm$^{-2}$ s$^{-1}$ to $(6.86 \pm 0.13) \times 10^{-10}$ erg cm$^{-2}$ s$^{-1}$. Second, the magnitude of the spin-down rate was increased from $(7.8 \pm 0.1) \times 10^{-13}$ Hz s$^{-1}$ to $(8.1 \pm 0.1) \times 10^{-13}$ Hz s$^{-1}$ ($\sim 4\%$ increase). Third, the pulse profile was changed significantly ($>5\sigma$). Fourth, the emission spectrum experienced marginal change ($<3\sigma$). In the phase-averaged spectral analysis, the emission from the pulsar is modeled by a power law with exponential cut-off, which has the functional form
\begin{equation}
\dfrac{\textrm{d}N}{\textrm{d}E}=N_0 \left(\dfrac{E}{E_0}\right)^{-\Gamma} \textrm{exp} \left(-\dfrac{E}{E_C}\right),
\label{equation:plsec}
\end{equation}
where $N_0$ is the normalization constant, $E_0$ is the scale factor of energy, $\Gamma$ is the spectral power-law index, and $E_C$ is the cut-off energy. It is found that the cut-off energy ($E_C$) decreases by $\sim 13\%$ in the glitch. There is also a slight decrease in the integrated energy flux in the energy range between 100 MeV and 300 GeV. The phase-resolved spectral analysis also revealed that the major variation in the emission spectrum of PSR~J2021+4026 results from the first major peak (P1) of the pulsed signal. The phase-resolved spectrum in the region of P1 experienced a significant drop in the cut-off energy ($E_C$) of $\sim 27\%$ after the glitch. The authors hypothesized that these sudden changes originate from the reconfiguration of magnetic field on the pulsar surface.
In our study, the gamma-ray emission of PSR~J2021+4026 was re-analyzed, using data of longer length and an updated version of the \textit{Fermi} Science Tools. The phase-averaged and phase-resolved spectral results are consistent with the study by Allafort et al. (2013). From the long-term tracking on the flux emitted from PSR~J2021+4026, we found that the drop caused by the glitch is a permanent effect. In particular, we present the findings from the flux evolution throughout the seven years after the launch of the \textit{Fermi}.
\subsection{Flux Evolution over Seven Years}
\label{section:dataanalysis}
The gamma-ray flux emitted by PSR~J2021+4026 over seven years since the start of \textit{Fermi} mission is analyzed. The data between 2008 August 04 and 2015 August 12 were used. The events in the Pass 8 “source” class were selected and the corresponding instrumental response functions for this event class is the P8R2\_SOURCE\_V6. The region of interest (ROI) is a $20^\circ \times 20^\circ$ square region centered at the epoch J2000 position $(\textrm{R.A.}, \textrm{Dec})=(20^{\textrm{h}} 21^{\textrm{m}} 34.08^{\textrm{s}},40^\circ 26^\prime 27.6^{\prime\prime})$. All photons with energy between 100 MeV and 100 GeV in this ROI were used. To avoid the contamination from the Earth's albedo, we excluded the time intervals with zenith angles $>90^\circ$ and those with the rocking angle of the LAT being $>52^\circ$.
The light curve (photon flux against time) of PSR~J2021+4026 was calculated in two steps. In the first step, the background emission model is obtained from the analysis of the data over
seven years. In the second step, the seven-year time interval is binned and the flux from the target pulsar in each time bin is computed. Both steps involve the binned likelihood analysis performed using the \textit{Fermi} Science Tools version v10r0p5\footnote{Available at \url{http://fermi.gsfc.nasa.gov/ssc/data/analysis/software/}}.
In the first step, the background emission was analyzed using the data over the
entire period ($\sim 7$ yr). The input model contains all 3FGL catalog sources (gll\_psc\_v16.fit; Acero et al. 2015) within $20^\circ$ from the center of ROI. All parameters are fixed
to the 3FGL catalog values for the sources that are $>10^\circ$ away from the ROI center. The galactic diffuse emission (gll\_iem\_v06) and the isotropic diffuse emission (iso\_P8R2\_SOURCE\_V6\_v06), available from the \textit{Fermi} Science Support Center (FSSC)\footnote{\url{http://fermi.gsfc.nasa.gov/ssc/ }}, were also included in the input model. An extended source positioned within the ROI, known as the Cygnus Loop, was modeled with the spatial template (CygnusLoop.fits) provided by the FSSC. Using the gtlike tool in the Science Tools, the best-fit model was obtained. Insignificant sources ($<3$-$\sigma$) were eliminated from the model. All spectral indices in the model were then fixed to the best-fit values. This model is then used as the input model in the second step.
In the second step, the data in seven-years were binned into intervals of 7 days or 30 days.
The photons in two energy ranges, $>100$ MeV and $>1$ GeV, were considered separately. The input model is the best-fit model with fixed spectral indices from the analysis of the entire
seven-years data in the first step. In each time bin, binned likelihood analysis was performed to obtain the normalization constants in the model and to estimate the photon flux of each source included in the model. The photon flux of PSR~J2021+4026 was then obtained. Figure~\ref{fig:lc7day} and figure~\ref{fig:lc30day} show the light curves with energies $>100$ MeV and $>1$ GeV in 7days and 30days bins, respectively. The red dashed line represents the mean flux
value throughout the seven years. The blue dashed lines represent the mean fluxes within three separate intervals: (1) before MJD 55850, (2) between MJD 55850 and 57000, and (3) beyond MJD 57000. As indicated in the light curves, there is a sudden and significant drop ($\sim 20\%$) in the photon flux around MJD 55850. There could also be hints of a second glitch around MJD 57000 in which the flux has experienced another jump with a smaller glitch size. Although we speculate that this could be a second glitch for this pulsar, we cannot confirm it at this moment because we have not obtained the timing parameters beyond MJD 56580. Also, the phase-averaged spectral analysis using data over eight months, which produces a relatively large uncertainty, does not guarantee a change. Therefore, in this paper, we only discuss the details of the first glitch. As it can be seen from figure~\ref{fig:lc30day}, the flux level between the two jumps is steady and shows no hint of gradual recovery. Therefore, we regarded the first jump in flux as a permanent effect by the glitch in MJD 55850.
\section{Modeling}
\label{section:modeling}
In this section, we will use a theoretical model to simulate the gamma-ray emissions from PSR~J2021+4026 and explain the cause and the consequences of the glitch that occurred around 2011 October 16. From the \textit{Fermi}-LAT observations reported in Allafort et al. (2013), the emission of PSR~J2021+4026 exhibits widely separated double peaks connected with a strong bridge, a spectral cut-off at high energy ($\sim 2$ GeV) and a high integrated energy flux. These characteristic properties favor the outer gap model (Cheng et al., 1986a, 1986b). Hence, this study
adopts the outer gap model in simulating the radiation of PSR~J2021+4026. In particular,
we will use the three-dimensional two-layer outer gap model, which is a more realistic outer gap model that considers the development and the closure of the outer gap by photon-photon pair creation (Wang et al. 2011). The details of the three-dimensional two-layer outer gap model are described in Wang et al. (2010, 2011). We will then use the crust cracking scenario to understand the glitch behavior of PSR~J2021+4026 in Section~\ref{section:crustcrack}.
\subsection{Glitch of PSR~J2021+4026: the Crust Cracking Scenario}
\label{section:crustcrack}
In this section, the possible cause of the glitch in PSR~J2021+4026 around 2011 October 16 will be discussed. From the gamma-ray observation described in Section~\ref{section:observation},
the 7 years light curve shows that the photon flux of the pulsar experienced a permanent decrease after the glitch. To produce a permanent change, the crust crack mechanism is preferred to the superfluid coupling. This is because the pulsar will gradually recover to its original condition after the glitch caused by superfluid. We propose that the crust cracking scenario leads to a change in the magnetic inclination angle, and thus the observed consequences of the glitch in PSR~J2021+4026.
\subsubsection{Crust Cracking and the Inclination Angle}
When Link et al. (1992) studied the glitch of the Crab pulsar that occurred in 1975, they found that the superfluid model could not explain the persistent change in the spin-down rate. They proposed that there must be an excess external torque that is a result of the rearrangement of the surface magnetic field. The surface magnetic field is rearranged through the plate tectonic activity (Ruderman 1991). In the plate tectonic model, the structure of a pulsar consists of an upper crust, a lower crust, and a core. The thickness of the solid crust is $\sim 10^5$ cm. Neutron superfluid vortex lines build up shear stresses in the crust lattices of the spinning pulsar. This stress will break the crust and move the crustal plates toward the equator. Since the surface magnetic field lines are localized and pinned to the plates, they move along with the plates. As a result, the dipole moment of the pulsar is more deviated from the rotational axis, thus increasing the magnetic inclination angle. In the case study of the Crab pulsar by Link et al. (1992), they found that a change in the inclination angle $\Delta \alpha \sim 10^{-4} \tan \alpha$ could reproduce the observed offset in the spin-down rate $\Delta \dot{\Omega}$.
\subsubsection{Effect on Spin-down Rate}
\label{subsection:spindownangle}
One main approach to describe the magnetosphere of pulsars is by solving equations for magnetohydrodynamics (MHD) in the force-free limit, where the forces in the plasma are balanced (i.e. $\rho \vec{E}+\frac{1}{c}\vec{j}\times\vec{B}=0$, where $\rho$ and $\vec{j}$ are the charge and current densities, respectively). The equations for MHD are derived from the Maxwell equations in special relativity with the imposed force-free conditions. Spitkovsky (2006) solved the MHD equations numerically using a finite-difference time-domain (FDTD) approach to investigate the evolution of pulsar magnetospheres. From the numerical solutions, it is found that for a pulsar having an inclination angle between the magnetic and the rotational axes, the spin-down luminosity can be formulated by
\begin{equation}
L_{pulsar}=k_1\dfrac{\mu^2 \Omega^4}{c^3}(1+k_2\sin^2\alpha),
\label{equation:spindownluminosity}
\end{equation}
where $k_1=1\pm 0.05$ and $k_2=1\pm 0.1$ are the best-fit coefficients from the numerical results. Considering the outer gap model as a small perturbation of the force-free model, we adopted this relationship. Using $L_{pulsar}=I\Omega\dot{\Omega}$, where $I$ is the moment of pulsar's
inertia, we found that the relative change in the spin-down rate ($\Delta\dot{\Omega}/\dot{\Omega}$) can be expressed as a function of the inclination angle ($\alpha$):
\begin{equation}
\dfrac{\Delta\dot{\Omega}}{\dot{\Omega}}=\dfrac{\sin 2\alpha\Delta\alpha}{1+\sin^2\alpha}.
\label{equation:sdinclination}
\end{equation}
In this way, a shifting in the inclination angle can lead to a change in the spin-down rate of the pulsar. In the case of PSR~J2021+4026, since the spin-down rate is increased after the
glitch, we expect there to be an increase in the inclination angle.
\subsubsection{Effect on Pulse Profile and Cut-off Energy}
Because of the rearrangement of surface magnetic fields due to crust plate activities, it is expected that the emission geometry is affected. The observed pulse profile can be interpreted as the pattern of magnetic field lines viewed at a fixed viewing angle $\beta$, which is the angle between the rotational axis of the pulsar and the observer. Higher flux is observed at the positions with higher magnetic field line densities. This is because the density of the magnetic
field line corresponds to the field strength. If the magnetic field is stronger, the charged particles are accelerated to higher speeds and emit more energetic curvature photons. The secondary pair creation and acceleration also contribute to the high flux observed. When the inclination angle is changed, leading to an adjustment in the magnetic field line pattern, the observed pulse profile will be modified.
Another consequence is the shift in the spectral cut-off energy. As discussed in Section~\ref{subsection:spindownangle}, the inclination angle of PSR~J2021+4026 is expected to increase after the glitch. When the inclination angle increases, the line of sight from the observer will cut the magnetic field lines at an angle closer to the polar region. Therefore, the inner magnetosphere will be viewed. Since the charged particles are accelerated through the outer gap toward the light cylinder, the curvature photons from the inner part have less energy than the outer region. As a result, the observed emission with increased inclination angle generally has lower energy. Thus, the spectral cut-off energy is shifted to a smaller value.
\subsection{Results}
We used the three-dimensional two-layer outer gap model described in Wang et al. (2011) to simulate the high-energy emission from PSR~J2021+4026. It is found that the model could not reproduce the widely separated peaks and high energy flux shown in the observation simultaneously. This is because a larger peak separation requires a thinner gap but a thinner gap will result in a lower spectral energy flux. To preserve both properties, we introduced an extra parameter to increase the energy in the emission spectrum such that a thinner gap can be applied. Takata \& Chang (2009) modeled the dependency between the maximum accelerating electric field in the outer gap structure and the size of the polar cap. It is found that the typical strength of the accelerating electric field is proportional to the square of the polar cap size. Here, we used an enlargement factor to amplify the size of the polar cap. In this way, both the peak separation and the high-energy emission spectrum can be retained. We chose the factor to be $1.6$.
The observation on PSR~J2021+4026 shows that the pulsar releases strong off-pulse emission, which is difficult to explain by traditional outer gap model. In order to produce the off-pulse emission, the outer gap is extended towards the stellar surface, until a limit called the inner boundary (Dyks \& Rudak 2003; Dyks et al. 2004 and Takata et al. 2004). Also, the viewing angle is required to be close to $90^\circ$ in order to observe the off-pulse emissions. In this study, the inner boundary was chosen to be located at the height of $(1-f)$ times the distance between the stellar surface and the null charge surface, where $f$ is the outer gap fraction defined in Wang et al. (2011).
In this section, the results will be presented. In particular, the pulse profile (in the full energy band and in separate energy bins) and spectra (phase-averaged and phase-resolved) before and after the glitch will be discussed.
\subsubsection{Pulse Profiles and Phase-averaged Spectra}
In the model, the pulsar PSR~J2021+4026 has a surface magnetic field of $1 \times 10^{13}$G at the poles, a rotational period of $0.265$~s
and a distance of $1.0$ kpc away from the Earth. The model parameters that characterize the geometry of the pulsar and the distributions of the properties of the two-layer outer gap are presented in table~\ref{table:modelpara}. $\alpha$ and $\beta$ are the inclination angle and viewing angle, respectively. $C$ is the normalization constant for the distribution of the gap fraction $f$. $B_1$ and $B_2$ are the normalization constants for the distribution of the ratio, $h_1/h_2$, between the sizes of the main acceleration ($h_1$) and screening ($h_2$) regions. $F$ is the normalization parameter for the drift motion of charged particles. $A$ is the offset parameter for the distribution of the average charge density. The expressions of these distributions are described in Wang et al. (2011). $\alpha$ and $\beta$ are determined from the pulse profile that
resembles the geometry of the magnetic field structure of a rotating dipole. Other parameters characterize the strength of the accelerating electric potential in the outer gap and thus the energy of the radiated curvature photons.
To test the crust cracking scenario on this pulsar glitch, we first obtained the set of parameters based on the observed pulse profile and spectrum. Then, we increased the value of the inclination angle $\alpha$ until there is no bump in the bridge emission of the pulse. The value of the viewing angle $\beta$ does not change after glitch because it is the angle between the rotational axis and the observer, which is fixed with time. Since we intend to explain the glitch by the change in the inclination angle, other parameters involved in the distributions of the gap structure are assumed to be unchanged to reduce the number of free parameters and the complexity of the scenario. From the relation between the change in the spin-down rate and the inclination angle discussed in Section~\ref{subsection:spindownangle}, we can estimate $\Delta \alpha$ from the observed $\Delta \dot{\Omega}$. Using equation~\ref{equation:sdinclination}, the observed value of $\dfrac{\Delta \dot{\Omega}}{\dot{\Omega}} \sim 4\%$ and the modeled value of $\alpha \sim 63^\circ$ before the glitch, we obtained the expected change in the inclination angle during the glitch to be $\Delta \alpha \sim 3^\circ$. This expectation of a $\sim 3^\circ$ increase in the inclination angle is consistent with the modeled value of $4.2^\circ \pm 0.3^\circ$, which is the difference between the values before and after glitch listed in table~\ref{table:modelpara}.
The model results are shown in figure~\ref{fig:distdiff}-\ref{fig:lcspecbeforeafter}. The structural distributions in the outer gap before and after the glitch are shown
in figure~\ref{fig:distdiff}. The plots include the distributions of the polar cap
radius $r_p$, the gap fraction $f$, the size ratio of the two layers $h_1/h_2$,
the radial distance to null charge surface $r_{null}$, the average charge density $\bar{\rho}$,
and the charge density in the acceleration region. The red and blue solid lines represent the distributions before and after the glitch, respectively. The black dashed lines are the
relative changes in the distributions after the glitch calculated by $(x_f-x_i)/x_i$, where $x$ is the concerned quantity.
The model pulse profile and spectrum before the glitch is shown in the left panel of figure~\ref{fig:lcspecbeforeafter}. The model pulse profile is drawn by a blue solid line and the observed pulse profile is represented by a gray histogram for comparison. The observed pulse profile is a photon-weighted folded curve using the data events between 2008 August 11 and 2011 December 16
on an ROI of $1^\circ$ centered at PSR~J2021+4026. The pulsar ephemeris for PSR~J2021+4026 is obtained from the \textit{Fermi}-LAT Multiwavelength Coordinating Group \footnote{\url{https://confluence.slac.stanford.edu/display/GLAMCOG/LAT+Gamma-ray+Pulsar+Timing+Models}} (Ray et al. 2011). The weighting was done by the gtsrcprob tool in the Science Tools, which calculates the probability that a photon originates from the pulsar, based on the maximum likelihood model in the same time period. The model was computed by the same method as described in Section~\ref{section:dataanalysis}. As seen in the pulse profile, the bump between the double peaks is a result of the outer gap geometry in the model, which states there is a local maximum in the magnetic field line density between the two peak emission regions. The peak separation and the off-pulse emission are also generally consistent with the observed data. The emission spectrum before the glitch is shown in the bottom left panel of figure~\ref{fig:lcspecbeforeafter}. The data points are from the spectral analysis in the same period of time as the pulse profile. An upper limit is reported when the detection significance is less than $3\sigma$.
The right panel in figure~\ref{fig:lcspecbeforeafter} shows the model pulse profile and spectrum after the glitch. The observed pulse profile uses the data between 2011 December 16 and 2013
October 14 limited by the valid time range of the timing model. In the case of post-glitch, there is no bump between the double peaks in the model. This corresponds to the vanish of the bump as seen in the gamma-ray observation. The cause of the bump disappearance is mainly the change of the inclination angle $\alpha$. The off-pulse emission is generally consistent with the observed data. However, the peak separation is less comparable to the data than in the case of pre-glitch. This is due to the fact that we used the inclination angle $\alpha$ as the only parameter in this simple model. $\alpha$ affects the geometry of the magnetic field lines. When it is increased, the inner layer of the magnetosphere is observed at the line of sight. Therefore, the span of the outer gap emission during a pulse period is shorter and the peak separation becomes smaller. For the spectrum, the energy flux at energy $>10$GeV is not well modeled and is underestimated when compared to the observation. It is the model's limitation to produce $>10$ GeV photons because the model describes a static outer gap thickness (Takata et al. 2016).
\subsubsection{Energy-dependent Pulse Profiles and Phase-resolved Spectra}
Through selecting different energy ranges for integrating the flux, the energy-dependent pulse profile is obtained. Figure~\ref{fig:edlcbeforeafter} illustrates the energy-dependent pulse profile of PSR~J2021+4026 before and after the glitch, respectively. The same energy bands used in data analysis were applied to the model pulse profile: (1) $>0.1$ GeV, (2) $>1$ GeV, (3) $0.3-1.0$ GeV, and (4) $0.1-0.3$ GeV.
The phase-resolved spectra are obtained by selecting the pulse intervals: first major peak (P1), bridge (BR), second major peak (P2), and off-pulse emission (OP), as labeled in figure~\ref{fig:edlcbeforeafter}. The results are shown in figure~\ref{fig:phaseresolvedspec}. Both spectra before and after glitch are plotted on the same figure with the red line representing the results before the glitch and the blue line representing the results after glitch. As discussed in Section~\ref{section:observation}, the decrease in the cut-off energy in the spectrum of phase P1 is much more significant among the phase-resolved spectra. Here, the three-dimensional two-layer outer gap model reproduced this special feature. We can see that the cut-off energy in the P1 spectrum in figure~\ref{fig:phaseresolvedspec} experienced a decrease after the glitch, while in other phases, the spectral cut-off energies are approximately maintained at the same values. The current model did not give a well-fit spectrum for the bridge emission. This is because the phase interval for BR is the narrowest among the four stages so that errors are accumulated most easily.
The sharp decrease of the spectral cut-off energy in P1 can be understood by considering the electrical properties in the two layers during the glitch. The relative changes of the azimuthal distributions for these properties of PSR~J2021+4026 are plotted in figure~\ref{fig:distdiff}. Phase intervals (P1, BR, P2, and OP) are indicated on the figure. The gap fraction $f$ shows a $\sim 15\%$ decline during P1 after the glitch. A smaller gap fraction means a smaller outer gap
region. The decrease in the electric potential difference across the gap establishes a weaker electric field in the acceleration region. The charged particles are accelerated to a slower speed and emit less energetic curvature photons, which will eventually suppress the cut-off energy. In another view, the ratio between the sizes of the two layers $h_1/h_2$ in the outer gap also shows a $\sim 4\%$ decrease in P1. This means that the ratio between the amount of $\sim$ GeV and $\sim 10^2$ MeV photons emitted in the acceleration and screen regions is lowered. Assuming
that the amount of GeV photons is subject to a heavier reduction than the amount of hundred MeV photons, the resulting drop in the cut-off energy shown in the phase-resolved spectrum of P1 can be explained.
On the other hand, the spectral cut-off in P2 shows no significant change after the glitch.
This can also be explained by the gap structure as indicated by figure~\ref{fig:distdiff}. The gap fraction $f$ and the ratio $h_1/h_2$ are increased by $\sim 5\%$
and $\sim 1\%$, respectively. The magnitude of these changes is much less than those that occurred in P1. Indeed, the energy flux of P2 at higher energy $(>5\rm{GeV})$ is slightly enhanced.
\section{Conclusion}
\label{section:conclusion}
We re-analyzed the \textit{Fermi}-LAT data by using a data length of seven years and an energy range from $100$ MeV to $100$ GeV. It is found that the flux of PSR~J2021+4026 experienced two jumps separated by more than three years. In the first jump, the flux was decreased by $\sim 20\%$ and it shows no hint of recovery. Therefore, we regarded the flux change as a permanent effect of the glitch. Although we speculate the second jump to be caused by another glitch, it cannot be confirmed at this moment. Concerning the first glitch, we proposed the crust cracking scenario to explain the glitch event, in which we assumed that the glitch was caused by plate tectonic activity. The magnetic field lines frozen on the surfaces were carried by the movement of these plates, resulting in the change of the magnetic inclination angle. This variation in the inclination angle will have an effect on the spin-down rate of the pulsar, the pulse pattern as seen by the observer, and the characteristic energy of the photons viewed at a constant line of sight.
We modeled the pulse profile and spectra of PSR~J2021+4026 using the three-dimensional two-layer outer gap model and simulate the glitch effect by imposing an increase in the inclination angle. In the model parameters, the inclination angle is increased by $\sim 4^\circ$, which is consistent with what we expect from the relation to the observed increase in the spin-down rate. Other parameters are fixed for the purpose of model simplification and control. We generated the pulse profiles, in both energy integrated and energy dependent, and the spectra, in both phase-averaged and phase-resolved, for PSR~J2021+4026. We found from the pulse profiles that the bump disappearance can be explained. We also found that the spectral cut-off is decreased most significantly in the pulse phase interval P1. We argued from the azimuthal distributions of the electrical properties in the interval of P1 that the size of the outer gap in the region of P1 is significantly reduced after the glitch, leading to a weaker electrical acceleration of the charged particles. Moreover, the ratio between the sizes of the acceleration and the screening regions is also reduced, which results in the relatively greater reduction in the high-energy photons at GeV. These conditions together reproduced the observed drop in the spectral cut-off energy.
We express our appreciation to an anonymous referee for useful comments
and suggestions. We thank A. H. Kong, C. Y. Hui, P. H. T. Tam, M. Ruderman,
and S. Shibata for the useful discussions. C.W.Ng N.G. and K.S.Cheng
are supported by a GRF grant from the Hong Kong Government under HKU17300814P. J.Takata is supported by the NSFC grants of China under 11573010.
|
\section{Introduction}
Non-perturbative structure of the vacuum and
confinement phenomenon represent two most important and closely related problems
in foundations of QCD which is supposed to be a basic fundamental theory
of strong interactions. Despite on significant progress
in lattice studies \cite{Phil2008}
there is still no deep knowledge about the microscopic vacuum structure
and origin of the color confinement from the first principles of QCD
(see \cite{diGiacomo2014} and references therein).
One of the attractive mechanisms of quark confinement is based on the
Meissner effect in dual color superconductor which assumes
generation of the monopole condensation
\cite{nambu74,mandelstam76,polyakov77,thooft81,Nair1985}.
The old known Savvidy-Nielsen-Olesen QCD vacuum
based on homogeneous magnetic field configuration
\cite{savv,N-O} suffers instability against the quantum fluctuations.
The most popular Copenhagen ``spaghetti'' vacuum
model \cite{niel-nino,niel-oles,amb-oles1} based on vortex domain
structure does not provide a consistent microscopic description of
the vacuum.
Numerous attempts to construct a rigorous true vacuum of QCD
using various vacuum field configurations
(instantons \cite{shuryak1997}, monopoles,
dyons \cite{1995simonov,kumar2010}, center vortices \cite{engelhardt2000},
monopoles \cite{choprl80,chopakprd2002,diGiacomo2015,pak05},
etc. \cite{schan82,bordag}) show that the existence of a stable vacuum
represents a most serious long-standing problem
in QCD. One of principal obstacles in resolving that problem was absence of a
proper regular classical solution which
must be stable against vacuum fluctuations at microscopic space scale and which
can serve as a structure element in formation of the vacuum.
Note that all previous vacuum models employ static vacuum field configurations
none of which do not possess quantum stability at microscopic scale. For instance,
a single static vortex field in the ``spaghetti'' vacuum model is
unstable, and the vacuum stability is restored only due to averaging
procedure over the statistical ensemble of the vortex domains.
An advantage of the ``spaghetti'' vacuum is that it provides approximate
qualitative description of the vacuum structure, in particular,
it was conjectured from quantum
mechanical consideration that color magnetic vortices should vibrate
at small space-time scale.
This gives a hint that time dependent field configurations might play
an important role in vacuum structure.
In the present paper we explore an idea that stationary generalized
monopole like solution can serve as an initial structure element
in formation of the QCD vacuum.
We consider a recently proposed classical stationary spherically
symmetric monopole solution which can be treated as a system
of a static Wu-Yang monopole interacting to time-dependent
off-diagonal gluon field \cite{pakp1}.
We have proved that such a solution provides a monopole vacuum field
which is stable against gluon fluctuations in one-loop approximation.
This opens a new way towards construction of a microscopic theory
of the QCD vacuum through the condensation of monopole and/or
monopole-antimonopole pairs.
\section{Stationary generalized Wu-Yang monopole solution}
Let us describe the main properties of the spherically symmetric
stationary monopole solution proposed in \cite{pakp1}.
Such a solution will be used in the subsequent sections
as a background field in the effective action functional,
and its non-trivial properties will provide the stability under
quantum gluon fluctuations.
We start with a standard classical Lagrangian of a pure $SU(3)$ QCD
\begin{equation}
{\cal L}=-\dfrac{1}{4}F_{a\mu\nu}F^{a\mu\nu}. \label{Lagr0}
\end{equation}
We introduce the following spherically symmetric ansatz
for non-vanishing components of the gauge potential
in spherical coordinates $(r, \theta, \varphi)$:
\begin{equation}
\begin{array}{rcl}
A_\varphi^1\!\!&=&\!\! -\dfrac{1}{r}\psi_1(t, r), \
A_\theta^2=\dfrac{1}{r}\psi_1 (t, r), \
A_\varphi^3= \dfrac{1}{gr} \cot \theta,
\\
A_\varphi^4\!\!&=&\!\!\dfrac{1}{r}\psi_2(t, r), \
A_\theta^5=\dfrac{1}{r}\psi_2 (t, r), \
A_\varphi^8=-\dfrac{\sqrt 3}{gr} \cot \theta,
\end{array}
\label{spherwav}
\end{equation}
where Abelian gauge potentials
$A_{\varphi}^3$ and $A_{\varphi}^8$ describe a static Wu-Yang monopole
with a total color magnetic charge two, $g_m^{\rm tot}=2$, \cite{choprl80},
and the functions $\psi_{1}(t,r)$ and $\psi_{2}(t,r)$ correspond to
dynamical degrees of freedom
of the off-diagonal components of the gluon field $A_\mu^a$.
The ansatz (\ref{spherwav}) describes two coupled monopole field
configurations corresponding to $I$- and $U$-type subgroups $SU(2)$.
One can verify that the ansatz (\ref{spherwav}) is consistent with all equations of motion
of the pure $SU(3)$ QCD,
and substitution of the ansatz into the equations of motion
results in only two non-trivial independent partial differential equations
\begin{equation}
\begin{array}{rcl}
&&\partial^2_{t} \psi_1 -\partial^2_r \psi_1+\dfrac{g^2}{2 r^2} \psi_1
\Big(2 \psi_1^2-\psi_2^2-\dfrac{2}{g^2}\Big)=0, \\
&&\partial^2_{t} \psi_2 -\partial^2_r \psi_2+\dfrac{g^2}{2 r^2} \psi_2
\Big(2
\psi_2^2-\psi_1^2-\dfrac{2}{g^2}\Big)=0.
\end{array}
\label{eqsu3}
\end{equation}
In a special case when $A_\theta^5=A_\varphi^4=A_\varphi^8= 0$
the ansatz describes a $SU(2)$ embedded field configuration which
corresponds to a system of the Wu-Yang monopole with a
unit monopole charge, $g_m^{\rm tot}=1$, interacting to off-diagonal gluon field.
With $\psi_1(t, r)\equiv \psi(t, r),~ \psi_2(t,r)=0$ the equations (\ref{eqsu3})
reduce to one differential equation
\begin{equation}
\partial^2_{t} \psi -\partial^2_r \psi+\dfrac{g^2}{r^2} \psi
\Big(\psi^2-\dfrac{1}{g^2}\Big)=0.
\label{eqsu2}
\end{equation}
The equation (\ref{eqsu2}) admits a wide class of time dependent
solutions including non-stationary soliton like propagating solutions
in the effective two-dimensional space-time
$(r,t)$ \cite{p14,p15,p16,p17,p18}.
We show that there is a subclass of generalized stationary Wu-Yang monopole
solutions which possess a finite energy density. The most important
issue is that such solutions are stable against quantum gluon
fluctuations
in pure QCD.
For simplicity we consider first the vacuum stability problem
in the case of $SU(2)$ Yang-Mills theory.
In that case by performing an appropriate gauge transformation
\cite{choprd80} one can rewrite
the $SU(2)$ part of the ansatz (\ref{spherwav}) in a regular gauge as follows
($a=1,2,3$)
\begin{equation}
A_m^a= -\epsilon^{abc}\hat n^b \partial_m \hat n^c \Big (\dfrac{1}{g}-\psi(t,r) \Big), \label{wuyang}
\end{equation}
where $\hat n=\vec r/r$.
It is clear that the ansatz (\ref{wuyang})
describes a generalized time dependent Wu-Yang monopole field configuration.
A known static Wu-Yang monopole corresponds to the limiting case $\psi(t,r)=0$,
and a trivial pure gauge vacuum configuration is described
by $\psi(t,r)= \pm 1/g$.
We prefer the ansatz written in the so-called singular
Abelian gauge \cite{choprd80}, (\ref{spherwav}),
since such a notation is more suitable for description of stationary
monopole solutions in $SU(N)$ Yang-Mills theory and in description of
multimonopole configurations.
For simplicity we consider first the vacuum stability problem
in the case of $SU(2)$ embedded solution.
For arbitrary function $\psi(t,r)$ one has
the following non-vanishing field strength components
\begin{equation}
\begin{array}{rcl}
F_{t\varphi}^1\!\!&=&\!\!-\dfrac{1}{r}\partial_t\psi, \ \
F_{r\varphi}^1=-\dfrac{1}{r}\partial_r \psi, \ \
\\
F_{t\theta}^2\!\!&=&\!\!\dfrac{1}{r}\partial_t \psi, \ \
F_{r\theta}^2=\dfrac{1}{r}\partial_r \psi, \ \
F_{\theta\varphi}^3=\dfrac{1}{gr^2}(g^2\psi^2-1).
\end{array}
\end{equation}
The radial magnetic field component $F_{\theta\varphi}^3$
generates a non-zero magnetic flux through a sphere with a center
at the origin, $r=0$.
So that, the color magnetic charge of the monopole depends
on time and distance from the center.
Note that various generalized static Wu-Yang monopoles have been
considered before, all of them have singularities in agreement with the
Derrick's theorem \cite{derr}.
Note that presence of singularities in the expressions for
the gauge potential (\ref{spherwav}) represents coordinate
singularities related to the
chosen singular gauge. Such coordinate singularities disappear in the
regular gauge (\ref{wuyang}).
By substitution of the ansatz (\ref{spherwav}) into the energy functional
one can verify that the energy density is finite everywhere
\begin{equation}
E=
4\pi\!\!\int\! {\rm d}r\,
\Big ((\partial_t \psi)^2+(\partial_r \psi)^2
+\dfrac{1}{2g^2r^2}(g^2 \psi^2-1)^2 \Big ) .
\label{entotspher}
\end{equation}
The equation (\ref{eqsu2}) admits a local non-static solution near the origin
which removes the singularity of the monopole at the center
\begin{eqnarray}
\psi(t,r)&=&\dfrac{1}{g}+\sum_{n=1} c_{2 n}(t) r^{2 n}, \nonumber \\
c_4(t)&=& \dfrac{1}{10} \big (3 g c_2^2(t) +c_2''(t)\big ), \nonumber \\
c_6(t)&=&\frac{1}{28}\big(c_4''(t)+6gc_2(t)c_4(t)+g^2c_2^3(t)\big), \nonumber \\
&\vdots& \label{locsol}
\end{eqnarray}
where the coefficient functions $c_{2n}(t)$ ($n\geq 2$) are determined in
terms of one arbitrary function $c_2(t)$.
In asymptotic region, $r\simeq \infty$,
the non-linear
equation (\ref{eqsu2}) reduces to a free D'Alembert
equation which has a standing spherical wave solution
\begin{eqnarray}
\psi(t, r) &\simeq& a_0+A_0 \cos (M r) \sin(M t)
+{\mathcal O}\Big(\dfrac{1}{r}\Big), \label{asymsol}
\end{eqnarray}
where $a_0, A_0$ are integration constants,
the mass scale $M$ appears due to scaling invariance
in the theory under dilatations $r \rightarrow Mr, t\rightarrow Mt $.
To solve numerically the Eq. (\ref{eqsu2})
we use the local solution (\ref{locsol})
to impose initial Dirichlet conditions
along the boundary $r=L_0$ in the numeric domain
$(L_0\leq r \leq L,~ 0\leq t \leq L)$.
The initial profile function $c_2(t)$ can be chosen arbitrarily as any regular
periodic function, we set
\begin{equation}
c_2(t)=c_0+c_1 \sin (M t), \label{locsol10}
\end{equation}
where $c_0, \, c_1$ are numeric constants.
Note that only one of two parameters in the local solution (\ref{locsol10})
is free, the other is fixed by the requirement that a
numeric solution matches the asymptotic solution (\ref{asymsol}).
Dimensional analysis implies that the energy of the solutions
is proportional to $M$ so that the energy vanishes in the limit $M\rightarrow 0$.
This might cause some doubts on stability of the solution.
However, one should stress that standard arguments
on existence of static solitonic solutions based on the Derrick's theorem
\cite{derr} are not applicable to the case of stationary solutions
which satisfy a condition of extremum value of the classical action,
not the energy functional. In the case of Yang-Mills theory
the action is conformal invariant
and its first variational derivative with respect to the scale parameter
$M$ vanishes identically.
So the parameter $M$ determines the scale of the space-time coordinates
and can be set to one without loss of generality.
With this one can solve numerically the equation (\ref{eqsu2}), and
the corresponding solution is depicted in Fig.\ref{Fig1}.
\begin{figure}[!h]
\centering
\includegraphics[width=70mm,height=52mm, bb=0 0 562 377]{gr1nov17c.jpg}
\caption[fig1]{Stationary monopole solution;
$(L_0\leq r,t \leq L)$, $L_0=10^{-6},~L=8 \pi$, $c_0=-0.041$,~$c_1=-0.523,~g=1$
.}\label{Fig1}
\end{figure}
Note that one has a stiffness numeric problem near the origin, so that
we have checked the regular structure and convergence of the numeric solution
in close vicinity of the origin up to $L_0=10^{-6}$ while keeping the
radial size of the numeric domain up to $L=64 \pi$.
The solution has several surprising features.
The field configuration of the solution includes a static
Wu-Yang monopole immersed in a standing spherical wave made of
off-diagonal gluons.
The standing wave does not screen completely the color monopole
charge at large distances.
One can find that in the asymptotic region the function $\psi(t,r)$
oscillates around the value $a_0\simeq 0.65$.
So the radial component of the color magnetic field
$F_{\theta \varphi}^1$ has a non-zero averaged value
which provides a non-vanishing total color magnetic charge.
Another interesting feature of the solution is that a corresponding
canonical spin density
vanishes identically. This gives a hint that such a solution,
treated as a quantum mechanical wave function, can lead to
a stable condensate of massive spinless particles.
An idea that particles (or pseudo-particles) can be described by
stationary solutions was sounded long time ago \cite{derr, jackiw77,jackiwRMP}.
The main obstacle towards practical realization of this idea
was absence of known regular stationary solutions in real QCD.
Even though our solution has an infinite total energy, the energy
density falls down as $1/r^2$ with increasing the distance
from the center of the monopole.
Note that our solution differs from the known non-linear spherical
wave type solutions which
have singularity at the center and light speed velocity.
The most important question related to properties of the solutions
is whether the monopole solution is stable against the quantum
gluon fluctuations.
\section{Vacuum stability of the monopole field in $SU(2)$ QCD}
\subsection{Effective action approach to vacuum stability problem}
In order to prove the quantum stability of the monopole field
we study the structure of the effective action of QCD using
the standard functional integral methods. We will consider the stability of the monopole
solution under small quantum fluctuations of the gluon field.
An initial gauge potential $A_\mu^a$ is split into a classical $\mathcal{B}_\mu^a$,
and a quantum $Q_\mu^a$ parts
\begin{equation}
A_\mu^a=\mathcal{B}_\mu^a+Q_\mu^a.\label{splitpot}
\end{equation}
We choose a temporal gauge for the classical
and the quantum field, $\mathcal{B}_t^a=Q^a_t=0$. The temporal gauge
has a residual symmetry which can be fixed by
imposing an additional covariant Coulomb gauge condition, ${\mathcal{D}}_i
Q^{ia}=0$,
with a covariant derivative ${\mathcal{D}}_i$ containing the background
field $\mathcal{B}_i^a$.
Hereafter we use the Latin indices for the spacial components.
A one-loop correction to the classical action
is obtained by functional integration over
the quantum fields $Q_i^a$, and it can be expressed
in terms of functional logarithms in a compact form
\cite{yildiz80,claudson80,adler81,dittrich83,flory83,blau91,
reuter97,chopakprd2002}
\begin{eqnarray}
S^{\rm 1 loop}_{\rm eff}&=&-\dfrac{1}{2} {\rm Tr} \ln [K_{ij}^{ab}]+
{\rm Tr}\ln [M_{\rm FP}^{ab}], \nonumber \\
K_{ij}^{ab}&=&-
\delta^{ab} \delta_{ij} \partial^2_t
-\delta_{ij}({\mathcal{D}}_k {\mathcal{D}}^k)^{ab} -2 f^{acb}{\mathcal
F}_{ijc}, \quad \label{Koper}\\
M_{\rm FP}^{ab}&=&-({\mathcal{D}}_k{\mathcal{D}}^k)^{ab}, \nonumber
\end{eqnarray}
where the operators $K_{ij}^{ab}$ and
$M^{ab}_{\rm FP}$ correspond to the contributions of
gluons and Faddeev-Popov ghosts, respectively.
The operator $K_{ij}^{ab}$ represents a
well-defined second order differential operator
of elliptic type in four-dimensional Euclidean space $(t, r,\theta,\varphi)$.
The question of vacuum stability is reduced to the question of
whether the operator $K_{ij}^{ab}$ has negative
eigenvalues for a given classical background field $\mathcal{B}_i^a$.
To find the eigenvalues one has to solve the Schr\"{o}dinger type
equation with the operator $K_{ij}^{ab}$ treated as a Hamiltonian
operator for a quantum mechanical system
\begin{equation}
K_{ij}^{ab} \Psi_j^b=\lambda \Psi_i^a, \label{schr3}
\end{equation}
where the ``wave functions'' $\Psi_i^a(t, r,\theta,\varphi)$ describe quantum
gluon fluctuations.
Note that the ghost operator originates from the interaction of
spinless ghost fields with the color magnetic field. Such interaction
does not produce negative tachyon modes \cite{N-O}, so it is sufficient
to study the eigenvalue spectrum of the operator $K_{ij}^{ab}$ only.
One should note that expression (\ref{Koper}) for the operator $K_{ij}^{ab}$
is valid for an arbitrary space-time dependent background field $\mathcal{B}_\mu^a$.
Moreover, the equation (\ref{Koper}) is manifestly gauge invariant.
This follows from the gauge invariant background field method
applied to one-loop quantization \cite{abbott,siegel}.
A key idea in the background field quantization scheme is that splitting
the initial gauge potential into a sum of the classical, $\mathcal{B}_\mu^a$,
and quantum, $Q_\mu^a$, fields implies two types of symmetries
originating from the original $SU(2)$ gauge transformation.
The first one, (I), is defined by a quantum gauge transformation
\begin{eqnarray}
\delta Q_\mu^a&=& (D_\mu \vec \alpha)^a, \nonumber \\
\delta \mathcal{B}_\mu^a&=&0, \label{typeIgtr}
\end{eqnarray}
where $D_\mu=\partial_\mu +\vec A_\mu$ is a covariant derivative containing the
full gauge potential, and $\alpha^a(x)$ is an infinitesimal gauge parameter.
The second type of symmetry, (II), is defined by a background gauge transformation
\begin{eqnarray}
\delta Q_\mu^a&=& f^{abc} Q_\mu^b \alpha^c, \nonumber \\
\delta \mathcal{B}_\mu^a&=& ({\mathcal{D}}_\mu \vec \alpha )^a. \label{typeIIgtr}
\end{eqnarray}
As a result within the framework of the background quantization
the classical field $\mathcal{B}_\mu^a$ appears in the expression for the
operator $K_{ij}^{ab}$ only in terms of background covariant derivatives and
gauge field strength ${\mathcal F}_\mu^{ab}$. It is clear that
operator $K_{ij}^{ab}$ transforms as an adjoint vector
under $SU(2)$ type (II) transformations,
and the fluon fluctuation field $\Psi_i^a$ transforms as a vector
in fundamental representation.
Taking in account that $SU(2)$ group is locally isomorphic
to the orthogonal group $SO(3)$ one concludes that the eigenvalue
spectrum is gauge invariant and does not depend on a
chosen gauge of the background field.
\subsection{Qualitative analysis of stability of the monopole field}
Before solving the Schr\"{o}dinger equation (\ref{schr3}) numerically,
let us perform a qualitative estimate of ground state eigenvalues
to trace the origin of positiveness of the operator $K_{ij}^{ab}$.
The operator $K_{ij}^{ab}$ does not have explicit dependence
on azimuthal angle $\varphi$, so that the ground state
eigenfunctions $\Psi_i^a$ depend only on three coordinates $(t, r,\theta)$.
First, we apply the variational method to reduce the three-dimensional
equation (\ref{schr3}) to an effective equation
in two-dimensional space-time $(t,r)$.
Within the variational approach one has to minimize
the following ``Hamiltonian'' functional
\begin{equation}
\langle \mathcal{H}\rangle
=\int {\rm d}r\, {\rm d}\theta\,
{\rm d}\varphi\, r^2 \sin \theta\, \Psi_i^a K_{ij}^{ab} \Psi_j^b. \label{funct}
\end{equation}
One can make qualitative estimates assuming
that all ground state eigenfunctions $\Psi_i^a$
includes angle dependence which guarantee the finiteness
of the Hamiltonian
\begin{equation}
\Psi_i^a(t, r,\theta,\varphi)=f_i^a(t, r) \sin \theta. \label{sufcond2}
\end{equation}
With this one can perform integration over the angle variables $(\theta, \varphi)$
in (\ref{funct}) and obtain an effective Schr\"{o}dinger equation for the ground state
\begin{equation}
\tilde K_{ij}^{ab} f_j^b(t,r)=\lambda f_i^a(t,r), \label{schreff}
\end{equation}
where the operator $\tilde K_{ij}^{ab}$
includes dependence only on two coordinates $(t, r)$,
\begin{equation}
\tilde K_{ij}^{ab}=
\delta_{ij}\delta^{ab}\Big(-\partial^2_t-\partial^2_r-\dfrac{2}{r}\partial_r\Big)
+V_{ij}^{ab}(t,r).
\end{equation}
The quadratic form corresponding to the potential $V_{ij}^{ab}$
can be written in the form
\begin{equation}
f_i^a V_{ij}^{ab} f_j^b = \dfrac{1}{r^2} V_0[f]+\dfrac{1}{r^2} V_1[f]
+\dfrac{\psi^2}{r^2}V_2[f]+\dfrac{\psi}{r^2}V_3[f]+\dfrac{\partial_r
\psi}{r}V_4[f],
\quad \label{quadV}
\end{equation}
where the first term includes contribution from a free vector Laplace operator,
$V_1[f]$ corresponds to contribution of the Wu-Yang monopole
and $V_{2,3,4}[f]$ contains
contributions proportional to $\psi^2, \psi$ and $\partial_r \psi$, respectively.
One can verify by using variational minimization procedure
that a total coefficient function in front of the centrifugal
potential $1/r^2$ is positively defined for arbitrary
fluctuating functions $f_i^a$. So that the effective Schr\"{o}dinger
equation (\ref{schreff}) contains a positive centrifugal potential
and a Coulomb type potential with
the oscillating coefficient function $\partial_r \psi V_4[f]$.
It is known that a quantum mechanical problem for a particle in
the potential well with small enough depth does not admit bound states
(in the space of dimension $d\geq 3$).
Therefore, there should be a critical value
of the amplitude $A_0$ of the monopole solution, (\ref{asymsol}),
below which the eigenvalue spectrum becomes positive.
Indeed, numerical study of solutions to the Eq.(\ref{schreff})
implies a critical value $a_{\rm 1cr}^{\rm bound}=1.3$.
The monopole field $\psi(r,t)$ is
approximated by a simple interpolating function
\begin{equation}
\psi^{\rm int}=1-\dfrac{(1-a_0)r^2}{1 + r^2}
+A_0 (1 -{\rm e}^{-d_0 r^2}) \cos (M r+b_0) \sin(M t), \quad \label{interpolfun}
\end{equation}
where $d_0$ and $b_0$ are fitting parameters.
A typical field configuration of $f_i^a$ corresponding
to the background monopole field $\psi(r,t)$
with the asymptotic parameters $a_0=0.895, A_0=0.615$
and $g=1, M=1$ is shown in Fig.2.
\begin{figure}[h!]
\centering
\includegraphics[width=77mm,height=62mm,bb=0 0 473 364]
{Jul1SPJun30PDEeq211332u13f.jpg}
\caption[fig2]{Wave function $f_1^3$ for the ground state with eigenvalue
$\lambda_0=0.0293$, $L=6 \pi$.}\label{Fig2}
\end{figure}
\subsection{Exact numeric solution to eigenvalue problem}
The qualitative estimates obtained in the previous subsection provide
only an upper bound for the critical parameter $a_{\rm 1cr}$.
To prove rigorously that there is no negative eigenvalues
one should solve numerically the original Schr\"{o}dinger type
equation (\ref{schr3}) without any approximation
on functional dependence of the eigenfunctions describing the gluon fluctuations.
The equation (\ref{schr3}) contains nine non-linear partial
differential equations which should be solved on three-dimensional
numeric domain with a sufficient numeric accuracy.
An additional technical difficulty in numeric calculation
on three-dimensional space-time is that one must solve
the equations with changing
the size of the numeric domain in radial direction in the limit $r \rightarrow \infty$
to verify that all eigenvalues remain positive. Fortunately, the numeric analysis
of the solutions corresponding to the lowest eigenvalue is simplified drastically
due to the factorization property of the original equation (\ref{schr3})
and special features of the ground state solutions as we will see below.
The eigenvalue equation (\ref{schr3}) written in the component
form admits factorization,
and it can be rewritten as two decoupled systems of partial differential equations
as follows (for brevity of notation we set $g=1$ since the coupling constant
can be absorbed by the monopole function $\psi(r,t)$)\\
(I):
\vspace*{-3mm}
\begin{eqnarray}
&&(\hat\Delta \Psi)_2^2-\dfrac{2}{r^2}\partial_\theta \Psi_1^2
+ \dfrac{1}{r^2}\Big ((\psi^2-1)\Psi_2^2-2\psi^2 \Psi_3^1+
2 \csc^2\theta (\Psi_2^2+\Psi_3^1)+2\cot \theta \psi \Psi_3^3
\Big )
=\lambda \Psi_2^2 , \nonumber \\
&&(\hat\Delta \Psi)_3^1-\dfrac{2}{r^2}\psi \partial_\theta \Psi_3^3
+
\dfrac{1}{r^2}\Big (\psi^2(-2\Psi_2^2+\Psi_3^1)-\Psi_3^1+2\csc^2\theta(\Psi_2^2+\Psi_3^1)+2\cot \theta\Psi_1^2
\Big ) =\lambda \Psi_3^1 , \nonumber \\
&&(\hat\Delta \Psi)_1^2+\dfrac{2}{r^2}\partial_\theta \Psi_2^2
+\dfrac{1}{r^2}\Big ((\cot^2 \theta + \psi^2)\Psi_1^2
+2\cot \theta (\Psi_2^2+\Psi_3^1) +2 \psi \Psi_3^3 +2\Psi_1^2 \Big )
\nonumber \\
&&\qquad-\dfrac{2}{r} \partial_r \psi \Psi_3^3=\lambda \Psi_1^2 ,
\label{eqI}
\\
&&(\hat\Delta \Psi)_3^3+\dfrac{2}{r^2}\psi \partial_\theta \Psi_3^1
+
\dfrac{1}{r^2}\Big (2\psi \Psi_1^2+2 \cot \theta \psi (\Psi_2^2+\Psi_3^1)
+2 \psi^2 \Psi_3^3 +\csc^2\theta \Psi_3^3\Big )
\nonumber \\
&&
\qquad -\dfrac{2}{r}\partial_r \psi \Psi_1^2=\lambda \Psi_3^3 ,\nonumber
\end{eqnarray}
(II):
\vspace*{-3mm}
\begin{eqnarray}
&& (\hat\Delta \Psi)_1^1+\dfrac{2}{r^2}\partial_\theta \Psi_2^1-\dfrac{2}{r^2}\psi \partial_\theta \Psi_1^3
+\dfrac{1}{r^2}\Big ((2+\cot^2 \theta +\psi^2) \Psi_1^1+2\psi \Psi_2^3
-2\cot \theta (\Psi_3^2-\Psi_2^1) \Big )\nonumber \\
&&\qquad -\dfrac{2}{r}\partial_r \psi \Psi_2^3
=\lambda \Psi_1^1 , \nonumber \\
&&(\hat\Delta \Psi)_2^3-\dfrac{2}{r^2}\partial_\theta \Psi_1^3+\dfrac{2}{r^2}\psi\partial_\theta \Psi_2^1
+\dfrac{1}{r^2}\Big (2\psi \Psi_1^1
+2\cot \theta \psi (\Psi_2^1-\Psi_3^2)+(2 \psi^2 +\csc^2\theta) \Psi_2^3 \Big )\nonumber \\
&&\qquad -\dfrac{2}{r}\partial_r \psi \Psi_1^1=\lambda \Psi_2^3 , \nonumber \\
&&(\hat\Delta \Psi)_2^1-\dfrac{2}{r^2}\partial_\theta \Psi_1^1-\dfrac{2}{r^2}\psi\partial_\theta \Psi_2^3
+\dfrac{1}{r^2}\Big (-2\psi\Psi_1^3
+ \psi^2(\Psi_2^1+2 \Psi_3^2) +2\csc^2\theta (\Psi_2^1-\Psi_3^2)-\Psi_2^1 \Big )\nonumber\\
&&\qquad +\dfrac{2}{r}\partial_r \psi \Psi_1^3
=\lambda \Psi_2^1 ,
\label{eqII}
\\
&&(\hat\Delta \Psi)_1^3+\dfrac{2}{r^2}\partial_\theta \Psi_2^3+\dfrac{2}{r^2}\psi\partial_\theta \Psi_1^1
+\dfrac{1}{r^2}\Big ( 2\cot \theta\psi \Psi_1^1+2(1+\psi^2) \Psi_1^3
-2 \psi(\Psi_2^1+\Psi_3^2) \nonumber \\
&&\qquad +2\cot\theta\Psi_2^3 \Big ) +
\dfrac{2}{r}\partial_r \psi(\Psi_2^1+ \Psi_3^2)
=\lambda \Psi_1^3 , \nonumber \\
&&(\hat\Delta \Psi)_3^2
+\dfrac{1}{r^2}\Big (2\cot \theta (\psi \Psi_2^3-\Psi_1^1)
-2\psi\Psi_1^3+\psi^2(2\Psi_2^1+\Psi_3^2)-2\csc^2\theta (\Psi_2^1-\Psi_3^2)
- \Psi_3^2 \Big )\nonumber \\
&&\qquad +\dfrac{2}{r}\partial_r \psi \Psi_1^3=\lambda \Psi_3^2 , \nonumber
\end{eqnarray}
where
\begin{eqnarray}
\hat \Delta \Psi_m^a \equiv -(\partial^2_{t}+\partial^2_r+\dfrac{2}{r}\partial_r+\dfrac{1}{r^2}\partial^2_\theta
+\dfrac{\cot\theta}{r^2}\partial_\theta)\Psi_m^a.
\end{eqnarray}
To solve numerically the systems of equations (I), (II)
we choose a rectangular three-dimensional domain
$(0\leq t \leq 2 \pi, r_0\leq r \leq L, 0\leq \theta \leq \pi )$
and use the same interpolating function for the monopole solution $\psi(r,t)$
as before. An obtained numeric solution to the system of equations
(I), $(\ref{eqI})$,
implies that the lowest eigenvalue is positive,
$\lambda_{\rm I}=0.0531$, and the corresponding
eigenfunctions have the following properties:
the functions $\Psi_1^2$ and $\Psi_3^3$ vanish identically,
and the remaining two functions
are related by the constraint $\Psi_3^1=-\Psi_2^2$. So that there is only one
independent non-vanishing function which can be chosen as $\Psi_2^2$.
An important feature of the solution corresponding to the lowest eigenvalue
is that the functions $\Psi_3^1,\Psi_2^2$ do not depend on the polar angle, Fig.3.
\begin{figure}[!h]
\centering
\includegraphics[width=77mm,height=62mm,bb=0 0 473 364]{Psi22f.jpg}
\caption[fig3]{Eigenfunction
$\Psi_2^2$ for the ground state with the lowest eigenvalue
$\lambda_I=0.0531$, $a_0=0.895,~A_0=0.615$, $ 0\leq r \leq 6 \pi$
,~ $0\leq t \leq 2 \pi $,
$0\leq\theta \leq \pi$.}\label{Fig3}
\end{figure}
This allows to simplify the system of equations (I)
in the case of ground state solutions with the lowest eigenvalues.
One can easily verify that the system of equations (I), (\ref{eqI}),
reduces to one partial differential equation on two-dimensional space-time
\begin{equation}
\Big (-\partial^2_{t}-\partial^2_{r}-\dfrac{2}{r}\partial_r
+\dfrac{1}{r^2}(3\psi^2-1) \Big) \Psi_2^2
=\lambda \Psi_2^2.
\end{equation}
The last equation represents a simple Schr\"{o}dinger type equation
for a quantum mechanical problem.
The equation does not admit negative eigenvalues if the parameter
$a_0$ of the monopole solution satisfies
the condition $a_0\geq 1/\sqrt 3\simeq 0.577\cdots$
which provides a totally repulsive quantum mechanical potential in this equation.
A structure of the system of equations (II), (\ref{eqII}), admits
a similar factorization property
on the space of ground state solutions.
We have solved numerically the equations (II)
with the same background monopole function $\psi(r,t)$ for various values
of the parameters $a_0, A_0, M$. In a special case,
$a_0=0.895,~A_0=0.615$, $g=1,M=1$, $ 0\leq r \leq 6 \pi$
the obtained numeric solution for the ground state has a lowest
eigenvalue $\lambda_{\rm II}=0.0142$ which is less than $\lambda_{\rm I}$.
All components of the solution do not have dependence on the polar angle and
satisfy the following relationships: $\Psi_2^1=\Psi_3^2$ and
$\Psi_1^1=\Psi_2^3=0$. There are two independent non-vanishing functions
which can be chosen as $\Psi_1^3$ and $\Psi_3^2$.
One can check that on the space of solutions corresponding to
the lowest eigenvalue the system of equations (II) is reduced
to two coupled partial differential equations for two
functions $\Psi_1^3(r,t)$ and $\Psi_3^2(r,t)$,
\begin{equation}
\renewcommand{\arraystretch}{1.8}
\begin{array}{rcl}
(-\partial^2_{t}-\partial^2_{r}-\dfrac{2}{r}\partial_r ) \Psi_1^3
+\dfrac{2}{r^2}\Big( (1+\psi^2)\Psi_1^3-2 \psi \Psi_3^2
\Big )
+\dfrac{4}{r} \partial_r\psi \Psi_3^2&=&\lambda \Psi_1^3, \\
(-\partial^2_{t}-\partial^2_{r}-\dfrac{2}{r}\partial_r ) \Psi_3^2
+\dfrac{1}{r^2} \Big ((3 \psi^2-1) \Psi_3^2-2 \psi \Psi_1^3
\Big )
+\dfrac{2}{r}\partial_r \psi \Psi_1^3&=&\lambda \Psi_3^2.
\end{array}
\label{sys2}
\renewcommand{\arraystretch}{1.0}
\end{equation}
Exact numeric solution profiles for the functions $\Psi_1^3, \Psi_3^2$ are shown in
Fig.4.
\begin{figure*}[!h]
\subfigure[~ $\Psi_1^3$]
{\includegraphics[scale=0.38,bb=0 0 498 359]{PDEred2113u13f.jpg}}
\qquad\qquad\qquad
\subfigure[~ $\Psi_3^2$]
{\includegraphics[scale=0.38,bb=0 0 498 359]{PDEred2113u21f.jpg}}
\caption[fig4]{Solution profiles for the functions
$\Psi_1^3(r,t)$, (a), $\Psi_3^2(r,t)$,
(b), corresponding to the lowest eigenvalue $\lambda =0.014218$,
$a_0=0.895,~A_0=0.615$,
$ 0\leq r \leq 6 \pi$,~ $0\leq t \leq 2 \pi $.}\label{Fig4}
\end{figure*}
We have obtained that the lowest eigenvalue is positive
when asymptotic monopole amplitude is less than
a critical value $a_{\rm 1cr} \simeq 0.625$.
We conclude that since the system of equations (II) has ground state solution
with the lower eigenvalue than ground state solution to the system (I),
it is enough to study the properties of the couple of
equations (\ref{sys2}) defined on two-dimensional space-time.
Note that numeric solving of the original eigenvalue equations (\ref{schr3})
on a three-dimensional domain does not provide high numeric accuracy,
especially in the limit of large values of the size $L$ of the numeric
domain along the radial direction.
Remind that one should prove the positiveness of the ground state
eigenvalues in the limit of infinite space
when the eigenvalues become very close to zero.
Solving the reduced equations (\ref{sys2}) defined on two-dimensional space-time
can be performed easily using standard numeric packages
with sufficient accuracy.
The obtained numeric accuracy for the eigenvalues $\lambda(L)$
in solving the two-dimensional equations (\ref{sys2})
is $1.0 \times 10^{-5}$ which is enough to construct the
eigenvalue dependence on the size $L$ of the box in
the range $6\pi\leq L \leq 64 \pi$.
\section{Vacuum stability of the monopole field in $SU(3)$ QCD}
The analysis of quantum stability
of the stationary monopole solution performed in the previous section
can be generalized straightforward to the case of
a pure $SU(3)$ QCD.
In general the system of two partial differential equations (\ref{eqsu3})
admits non-stationary and quasi-stationary solutions.
We are interested in stationary monopole solutions
which can be obtained in a full analogy with the case of $SU(2)$
QCD by using a constraint $\psi_1=\psi_2 \equiv \tilde \psi$.
The Schr\"{o}dinger type equation (\ref{schr3}) for possible
unstable modes corresponding to quantum gluon fluctuations contains twenty four
partial differential equations. Since the initial ansatz (\ref{spherwav})
describes monopole solution corresponding to $I$- and $U$-type subgroups
$SU(2)$ one expects that factorization property and
reduction of the equations on the space of ground state solutions
will take place in a similar manner as in the case of $SU(2)$ QCD.
Indeed, the numeric analysis of ground state solutions
of the full $SU(3)$ Schr\"{o}dinger equation (\ref{schr3})
with charge two monopole background
field $\tilde \psi$ implies factorization of the equations in each
sector of $I,U$ subgroups $SU(2)$.
The obtained numeric solution with the lowest eigenvalue
has a basis which contains six non-vanishing functions ${\tilde \Psi}_1^3,{\tilde \Psi}_3^2,
{\tilde \Psi}_2^1= {\tilde \Psi}_2^6 ={\tilde \Psi}_3^7={\tilde \Psi}_3^2,{\tilde \Psi}_1^8=\sqrt 3 {\tilde \Psi}_1^3$,
all other functions are vanished identically.
An important feature of the numeric solution is that
all solution profile functions do not depend on the polar angle $\theta$.
One can choose the functions
${\tilde \Psi}_1^3,{\tilde \Psi}_3^2$ as independent ones and verify that twenty four
equations in the Schr\"{o}dinger type equation (\ref{schr3}) reduce to a
simple system of two independent partial differential equations
on two-dimensional space-time $(r,t)$ in a consistent manner
\begin{equation}
\renewcommand{\arraystretch}{1.8}
\begin{array}{rcl}
\partial^2_r {\tilde \Psi}_{1}^3+\dfrac{2}{r} \partial_r {\tilde \Psi}_{1}^3+\partial_t^2{\tilde \Psi}_{1}^3
-\dfrac{1}{r^2}\Big ((2 {\tilde \psi}^2+1){\tilde \Psi}_1^3-2 ({\tilde \psi}-r{\tilde \psi}_r)
{\tilde \Psi}_3^2\Big )
&=&0, \\
\partial^2_r{\tilde \Psi}_{3}^2+\dfrac{2}{r}\partial_r {\tilde \Psi}_{3}^2+\partial^2_t{\tilde \Psi}_{3}^2
-\dfrac{1}{2 r^2}\Big ((3{\tilde \psi}^2-2){\tilde \Psi}_3^2-4({\tilde \psi}-r{\tilde \psi}_r) {\tilde \Psi}_1^3\Big
)&=&0.
\end{array}
\label{su3reduction}
\renewcommand{\arraystretch}{1.0}
\end{equation}
One can show that the structure of the reduced equations
is equivalent to the structure of the
equations (II), (\ref{sys2}), obtained by reduction in the case of $SU(2)$ QCD.
Simple rescaling of the function ${\tilde \psi}= \sqrt 2 \psi$
representing the monopole background field
and function renormalization ${\tilde \Psi}_3^2=\sqrt 2 \Psi_3^2$
(${\tilde \Psi}_1^3=\Psi_1^3$) in the Eqs.(\ref{su3reduction}) lead exactly
to the system of equations (II), (\ref{sys2}).
We have verified that with increasing the size of
the space-time numeric domain in the radial direction,
$L\rightarrow \infty$, the corresponding eigenvalue $\lambda(L)$
vanishes asymptotically from positive values.
The eigenvalue dependencies $\lambda(L)$ obtained
by solving the approximate, (\ref{schreff}),
and exact Schr\"{o}dinger equations, (\ref{schr3}),
in the cases of $SU(2)$ and $SU(3$) QCD are presented in Fig.5.
\begin{figure}[h!]
\centering
\includegraphics[width=72mm,height=52mm,bb=0 0 324 201]
{jul30PRLfininterpolB.jpg}
\caption[fig5]{Lowest eigenvalue dependence, $\lambda(L)$,
on the radial size $L$ of the numeric domain:
(a) approximate results in $SU(2)$ QCD,
(b) exact numeric results in $SU(2)$ QCD,
(c) exact numeric results in $SU(3)$ QCD.}\label{Fig5}
\end{figure}
Note that in $SU(3)$ QCD one has a stable monopole field configuration
for both stationary monopole solutions,
for the embedded $SU(2)$ monopole with a unit magnetic charge,
and for $SU(3)$ monopole with the magnetic charge two.
The $SU(3)$ symmetric monopole solution corresponding to two subgroups $SU(2)$
is preferable since for constant valued functions $\psi_1, \psi_2$ the
corresponding classical potential has an absolute minimum at
$\psi_1\!\!=\!\!\psi_2\!\!=\!\!\sqrt 2$.
This is similar to the behavior of one-loop effective potential
for homogeneous color magnetic fields where the potential has an absolute minimum for
non-vanishing values
of both magnetic fields $H_3$ and $H_8$
corresponding to Cartan subalgebra of $\mathfrak{su}(3)$
\cite{fly}.
\section{Conclusion}
In our consideration of the stationary monopole solutions
we set the mass scale parameter $M$ to unit for simplicity.
We would like to stress the importance of the presence of such a parameter
and its relation to microscopic structure of the vacuum.
One of the main characteristics of
QCD vacuum is the vacuum gluon condensate, $\langle \vec F_{\mu\nu}^2 \rangle$.
In first approximation the vacuum gluon condensate represents a constant parameter
which describes a mass gap in the confinement phase.
Within the framework of the confinement mechanism
based on the monopole condensation and dual Meissner effect it is assumed that
the vacuum gluon condensate is generated due to condensation
of monopoles, i.e., the dominant contribution to
the gauge invariant quantity
$\langle\vec F_{\mu\nu}^2\rangle$ is made of color magnetic
field.
For the present moment
an exact vacuum field configuration corresponding to the vacuum
monopole condensate is unknown. We expect that QCD vacuum
is formed through the condensation of monopoles and/or monopole
pairs since only these
objects possess quantum stability locally in small vicinity of each
space point before any averaging procedure over the whole space-time region.
Simple consideration shows that the mass scale parameter
$M$ determines the microscopic scale of the confinement
phase.
Let us consider the structure of the one-loop effective potential of
a pure $SU(N)$ QCD with a color magnetic background field
which is treated as a constant field
\cite{savv,yildiz80,claudson80,adler81,dittrich83,flory83,blau91,reuter97}
\begin{equation}
V_{\rm eff}
=\dfrac{1}{4} H^2
+\dfrac{11 N g^2(\mu)}{48 \pi^2} H^2\Big(\ln \dfrac{g(\mu) }{\mu^2}
-c\, \Big),
\end{equation}
where $g(\mu)$ is a renormalized coupling constant
defined at the subtraction point $\mu^2\simeq \Lambda_{\rm QCD}$.
The gauge invariant quantity $H^2 \equiv \langle\vec{F}^2_{\mu\nu}\rangle$
represents a vacuum
averaged value of the gluon field operator which describes the monopole condensate.
Note that in the standard QCD the notion of the vacuum gluon condensate
differs from the notion of the electron pair condensate in the superconductor which
is described by a wave function of the Cooper electron pair
(we do not consider phenomenological
approach to QCD based on Ginsburg-Landau type models).
The effective potential has a non-trivial minimum
at non-zero value of the vacuum magnetic field
$\langle H\rangle \simeq 0.14 \mu^2$ which fixes the scale $M$
of the microscopic structure
of the vacuum monopole condensate.
In the confinement phase one
assumes that the periodic structure of the classical
monopole field configuration
is characterized by the length parameter $\lambda_M=2 \pi/M$
which should be less than the inverse deconfinement temperature parameter
$kT_{\rm dec}$, i.e.,
$2 \pi/M\leq \beta_{\rm dec}=1/kT_{\rm dec}$.
Under this condition one can perform space-time averaging of the
classical monopole configuration and estimate a lower bound
$M^2_{\rm bound}\simeq 1.2 \mu^2$ which is of the same order
as the parameter $\Lambda_{\rm QCD}$.
In the confinement phase, $T\simeq 0$, the vacuum averaging value of the gluon field
operator $\langle0|A_\mu^a|0\rangle$ vanishes since the effective
size $\beta = 1/kT$ of the time interval in the Euclidean functional integral
of the effective action becomes much larger than $\lambda_M$.
With increasing temperature the value of the parameter $\beta$
becomes comparable and less than the periodic scale $\lambda_M$
of the vacuum monopole field configuration. This implies that
the vacuum averaging value of the gluon field operator,
$\langle0|A_\mu^a|0\rangle$, becomes
non-vanishing, which leads to the deconfinement phase with the spontaneous
symmetry breaking and gluon can be observed as a color object.
In conclusion,
we have demonstrated that the stationary spherically symmetric monopole
solution represents a stable quasi-classical vacuum field background
under small gluon fluctuations.
Certainly, our consideration is restricted by one-loop consideration
within the formalism of the effective action.
Note that the classical monopole solution is non-perturbative
and it is valid for arbitrary values of the coupling constant.
Besides, the one-loop effective action includes non-perturbative
contribution due to summation over all one-loop Feynman diagrams
corresponding to interaction of the background monopole field with virtual gluons.
This gives a hope that a modified monopole solution to quantum
equations of motion obtained beyond one-loop approximation
will admit the quantum stability as well.
Recently it has been shown that the spherically symmetric
monopole solution is rather classically unstable under
axially-symmetric deformations of the initial spherically-symmetric
ansatz \cite{pakp1}.
This might cause some doubts whether spherically symmetric monopoles can serve
as a structure element of a true vacuum.
One should note that classical stability of the multi-monopole system
represents a non-trivial problem due to presence of
interaction between the monopoles.
Another candidate for a stable structure element in formation of
a stable vacuum has been proposed in \cite{P4}
where it has been shown that a monopole-antimonopole pair solution
is stable under quantum gluon fluctuations.
We expect that monopole and/or monopole-antimonopole pair
condensation can be realized in QCD in analogy
with the Cooper electron pair condensation in ordinary superconductor,
as it was conjectured in the seminal papers long
time ago \cite{nambu74, mandelstam76, polyakov77, thooft81}.
This issue will be considered in a separate paper.
\section*{Acknowledgment}
One of authors (DGP) thanks Prof. C. M. Bai for warm hospitality
during his staying at the Chern Institute of Mathematics.
The work is supported by National Research Foundation
of Korea, grants NRF-2014R1A2A1A01002306 and NRF-2017R1D1A1B03028310;
Korean Federation of Science and Technology, Brain Pool Program,
and grant OT-$\Phi$2-10.
|
\section{Introduction}
\label{sec:introduction}
The goal of sequential data assimilation is to estimate the true state of a dynamical system $\x^\textnormal{true} \in \Re^{\Nstate \times 1}$ using information from numerical models, priors, and observations. A numerical model captures (with some approximation) the physical laws of the system and evolves its state forward in time \cite{Cheng2010}:
\begin{eqnarray}
\label{eq:intro-numerical-model}
\x_{k} = \M_{t_{k-1} \rightarrow t_k} \lp \x_{k-1} \rp \in \Re^{\Nstate \times 1},\, \text{ for $\x \in \Re^{\Nstate \times 1}$} ,\,
\end{eqnarray}
where $\Nstate$ is the dimension of the model state, $k$ denotes time index, and $\M$ can represent, for example, the dynamics of the ocean and/or atmosphere. A prior estimation $\x_k^\textnormal{b} \in \Re^{\Nstate \times 1}$ of $\x_k^\textnormal{true}$ is available, and the prior error $\errbac$ is usually assumed to be normally distributed:
\begin{eqnarray}
\label{eq:background-errors}
\displaystyle
\x_k^\textnormal{b} - \x^\textnormal{true} = \errbac_k \sim \Nor \lp {\bf 0},\, \B_k \rp \in \Re^{\Nstate \times 1},
\end{eqnarray}
where $\B_k \in \Re^{\Nstate \times \Nstate}$ is the background error covariance matrix. Noisy observations (measurements) of the true state $\y_k \in \Re^{\Nobs \times 1}$ are taken, and the observation errors $\errobs$ are usually assumed to be normally distributed:
\begin{eqnarray}
\label{eq:noisy-observations}
\displaystyle
\y_k -\Ho \lp \x_k^\textnormal{true} \rp = \errobs_k \sim \Nor \lp {\bf 0},\, \R_k \rp\, \in \Re^{\Nobs \times 1} ,\,
\end{eqnarray}
where $\Nobs$ is the number of observed components, $\Ho: \Re^{\Nstate\times 1} \rightarrow \Re^{\Nobs \times 1}$ is the observation operator, and $\R_k \in \Re^{\Nobs \times \Nobs}$ is the data error covariance matrix.
Making use of Bayesian statistics and matrix identities, the assimilation of the observation \eqref{eq:noisy-observations} is performed as follows:
\begin{equation}
\label{eq:analysis-state}
\begin{split}
\x_k^\textnormal{a} &= \x_k^\textnormal{b} + \B_k \cdot \H_k^T \cdot \lb \H_k \cdot {\B_k} \cdot \H_k^T +\R_k \rb^{-1} \cdot \lb \y_k - \Ho(\x_k^\textnormal{b}) \rb \in \Re^{\Nstate \times 1} ,\,\\
{\bf A}_k &= \lb \I - \B_k \cdot \H_k^T \cdot \lb \R_k+\H_k \cdot \B_k \cdot \H_k^T \rb^{-1} \cdot \H_k \rb \cdot \B_k \in \Re^{\Nstate \times \Nstate},\,
\end{split}
\end{equation}
where $ \H_k \approx \Ho'(\x_k^\textnormal{b}) \in \Re^{\Nobs \times \Nstate}$ is a linear approximation of the observational operator, and $\x_k^\textnormal{a} \in \Re^{\Nstate \times 1}$ is the analysis (posterior) state.
According to equation \eqref{eq:analysis-state} the elements of $\B_k$ determine how the information about the observed model components contained in the innovations $ \y_k - \Ho(\x_k^\textnormal{b}) \in \Re^{\Nobs \times 1}$ is distributed to properly adjust all model components, including the unobserved ones. Thus, the successful assimilation of the observation \eqref{eq:noisy-observations} will rely, in part, on how well the background error statistics are approximated.
In the context of ensemble based methods, an ensemble of model realizations
\begin{eqnarray}
\label{eq:background-ensemble}
\X^\textnormal{b}_k = \lb \x^{\textnormal{b}[1]}_k,\,\x^{\textnormal{b}[2]}_k,\,\ldots,\, \x^{\textnormal{b}[\Nens]}_k\rb \in \Re^{\Nstate \times \Nens} ,\,
\end{eqnarray}
is used in order to estimate the unknown moments of the background error distribution:
\begin{subequations}
\label{eq:moments-ensemble}
\begin{eqnarray}
\label{eq:ensemble-mean}
\displaystyle
\xm^\textnormal{b}_k &=& \frac{1}{\Nens} \cdot \sum_{i=1}^{\Nens} \x_k^{\textnormal{b}[i]} \in \Re^{\Nstate \times 1} ,\\
\label{eq:covariance-ensemble}
\displaystyle
\B_k \approx \P^\textnormal{b} &=& \frac{1}{\Nens-1} \cdot {\bf U}_k^\textnormal{b} \cdot \left({\bf U}_k^\textnormal{b}\right)^T \in \Re^{\Nstate \times \Nstate} ,\,
\end{eqnarray}
where $\Nens$ is the number of ensemble members, $\x_k^{\textnormal{b}[i]} \in \Re^{\Nstate \times 1}$ is the $i$-th ensemble member, $\xm_k^\textnormal{b} \in \Re^{\Nstate \times 1}$ is the background ensemble mean, $\P_k^\textnormal{b}$ is the background ensemble covariance matrix, and ${\bf U}_k \in \Re^{\Nstate \times \Nens}$ is the matrix of member deviations:
\begin{eqnarray}
\label{eq:matrix-of-member-deviations}
\displaystyle
{\bf U}_k^\textnormal{b} = \X_k^\textnormal{b} - \xm_k^\textnormal{b} \cdot {\bf 1}_{\Nens}^T \in \Re^{\Nstate \times \Nens}.
\end{eqnarray}
\end{subequations}
One attractive feature of $\P_k^\textnormal{b}$ is its flow-dependency which allows to approximate the background error correlations based on the dynamics of the numerical model \eqref{eq:intro-numerical-model}. However, in operational data assimilation, the number of model components is much larger than the number of model realizations $\Nstate \gg \Nens$ and therefore $\P_k^\textnormal{b}$ is rank-deficient. Spurious correlations (e.g., correlations between distant model components in space) can degenerate the quality of the analysis corrections. One of the most succesful EnKF formulations is the local ensemble transform Kalman filter (LETKF) in which the impact of spurious analysis corrections is avoided by making use of local domain analyses. In this context, every model component is surrounded by a box of a prescribed radius, and then the assimilation is performed within every local box. In this case the background error correlations are provided by the local ensemble covariance matrix. The local analyses are mapped back onto the global domain to obtain the global analysis state. Nevertheless, when sparse observational networks are considered many boxes can contain no observations, in which case the local analyses coincide with the background. The local box sizes can be increased in order to include observations within the local domains, in which case local analysis corrections can be impacted by spurious correlations. Moreover, in practice, the size of local boxes can be still larger than the number of ensemble members and therefore, the local sample covariance matrix can be rank-deficient.
In order to address the above issues this paper proposes a better estimation of the inverse background error covariance matrix $\B^{-1}$ obtained via a modified Cholesky decomposition. By imposing conditional independence between errors in remote model components we obtain sparse approximations of $\B^{-1}$.
This paper is organized as follows. In Section \ref{sec:background} ensemble based methods and the modified Cholesky decomposition are introduced. Section \ref{sec:EnKF-MC} discusses the proposed ensemble Kalman filter based on a modified Cholesky decomposition for inverse covariance matrix estimation; a theoretical convergence of the estimator in the context of data assimilation as well as its computational effort are discussed. Section \ref{sec:experimental-settings} presents numerical experiments using the Atmospheric General Circulation Model SPEEDY; the results of the new filter are compared against those obtained by the local ensemble transform Kalman filter. Future work is discussed in Section \ref{sec:future-work} and conclusions are drawn in Section \ref{sec:conclusions}.
\section{Background}
\label{sec:background}
The ensemble Kalman filter is a sequential Monte Carlo method for state and parameter estimation of non-linear models such as those found in atmospheric and oceanic sciences \cite{TELA:TELA299,EnKFEvensen,EnKF1657419}. The EnKF popularity is due to its basic theoretical formulation and its relative ease of implementation \cite{EnKFEvensen}. Given the \textit{background} ensemble \eqref{eq:background-ensemble} EnKF builds the \textit{analysis} ensemble as follows:
\begin{subequations}
\begin{eqnarray}
\label{eq:EnKF-analysis-ensemble}
\displaystyle
\X^\textnormal{a} = \X^\textnormal{b} + \P^\textnormal{b} \cdot \H^T \cdot \lb \R + \H \cdot \P^\textnormal{b} \cdot \H^T \rb \cdot \boldsymbol{\Delta} \in \Re^{\Nstate \times \Nens} ,\,
\end{eqnarray}
where:
\begin{eqnarray}
\label{eq:EnKF-innvo-observations}
\displaystyle
\boldsymbol{\Delta} = \Y^\textnormal{s} - \Ho (\X^\textnormal{b}) \in \Re^{\Nobs \times \Nens} \,,
\end{eqnarray}
and the matrix of perturbed observations $\Y^\textnormal{s} \in \Re^{\Nobs \times \Nens}$ is:
\begin{equation}
\label{eq:EnKF-synthetic-data}
\begin{split}
\Y^\textnormal{s} &= \lb \y+\errobs^{[1]},\, \y+\errobs^{[2]},\, \ldots, \, \y+\errobs^{[\Nens]} \rb \in \Re^{\Nobs \times \Nens} \,,\\
\errobs^{[i]} &\sim \Nor \lp {\bf 0},\, \R \rp, \quad 1 \le i \le \Nens \,.
\end{split}
\end{equation}
\end{subequations}
For ease of notation we have omitted the time index superscripts.
The use of perturbed observations \eqref{eq:EnKF-synthetic-data} during the assimilation provides asymptotically correct analysis-error covariance estimates for large ensemble sizes and makes the formulation of the EnKF statistically consistent \cite{Thomas2002}. However, it also has been shown that the inclusion of perturbed observations introduces sampling errors in the assimilation \cite{SamplingErrors1,SamplingErrors2}.
One of the important problems faced by current ensemble based methods is that spurious correlations between distant components in the physical space lead to spurious analysis corrections. Better approximations of the background error covariance matrix are proposed in the literature in order to alleviate this problem. A traditional approximation of $\B$ is the Hollingworth and Lonnberg method \cite{TELA:TELA460} in which the difference between observations and background states are treated as a combination of background and observations errors. However, this method provides statistics of background errors in observation space, and requires dense observing networks (not the case in practice). Another method has been proposed by Benedetti and Fisher \cite{QJ:QJ37} based on forecast differences in which the spatial correlations of background errors are assumed to be similar at 24 and 48 hours forecasts. This method can be efficiently implemented in practice, however, it does not perform well in data-sparse regions, and the statistics provided are a mixture of analysis and background errors. Another way to reduce the impact of spurious correlations is based on adaptive modeling \cite{AdaptativeModeling}. In this context, the model learns and changes with regard to the data collected (i.e., parameters values and model structures). This allows to calibrate, in time, the error subspace rank (i.e., number of empirical orthogonal functions used in the assimilation process), the tapering parameter (i.e., local domain sizes), and the ensemble size, among others. Yet another method based on error subspace statistical estimation is proposed in \cite{LemusDA}. This approach develops an evolving error subspace, of variable size, that targets the processes where the dominant errors occur. Then, the dominant errors are minimized in order to estimate the best model state trajectory with regard to the observations. We proposed approximations based on autoregressive error models \cite{Sandu_2007_ARMA} and using hybrid subspace techniques.\cite{Sandu_2010_hybridCovariance}.
Covariance matrix localization artificially reduces correlations between distant model components via a Schur product with a localization matrix $\boldsymbol{\Pi} \in \Re^{\Nstate \times \Nstate}$:
\begin{eqnarray}
\label{eq:EnKF-localized-ensemble-covariance}
\displaystyle
\widehat{\P}^\textnormal{b} = \boldsymbol{\Pi}\circ \P^\textnormal{b} \in \Re^{\Nstate \times \Nstate}
\end{eqnarray}
and then $\P^\textnormal{b}$ is replaced by $\widehat{\P}^\textnormal{b} \in \Re^{\Nstate \times \Nstate}$ in the EnKF analysis equation \eqref{eq:EnKF-analysis-ensemble}. The entries of $\boldsymbol{\Pi}$ decrease with the distance between model components depending on the radius of influence ${\zeta}$:
\begin{eqnarray}
\label{eq:EnKF-covariance-localisation}
\lle \boldsymbol{\Pi} \rle_{i,j} = \exp \lp -\frac{\pi \lp m_i,\,m_j \rp}{f({\zeta})} \rp \,, \text{ for $1 \le i \le j \le \Nstate$}\,,
\end{eqnarray}
where $\pi \lp m_i,\,m_j \rp$ represents the physical distance between the model components $m_i$ and $m_j$ while, $f({\zeta})$ is a function of ${\zeta}$ (e.g., $f({\zeta}) = 2 \cdot {\zeta}^2$). The exponential decay allows to reduce the impact of innovations between distant model components. The use of covariance matrix localization alleviates the impact of sampling errors. However, the explicit computation of $\boldsymbol{\Pi}$ (and even $\P^\textnormal{b}$) is prohibitive owing to numerical model dimensions. Thus, domain localization methods \cite{spatial_localization,local_analysis_1} are commonly used in the context of operational data assimilation. One of the best EnKF implementations based on domain localization is the local ensemble transform Kalman filter (LETKF) \cite{LETKFHunt}. In the LETKF the analysis increments are computed in the space spanned by the ensemble perturbations ${\bf U}^\textnormal{b}$ defined in \eqref{eq:matrix-of-member-deviations}. An approximation of the analysis covariance matrix in this space reads:
\begin{subequations}
\label{eq:LETKF-method}
\begin{eqnarray}
\label{eq:LETKF-analysis-covariance-ensemble-space}
\displaystyle
\widehat{\P}^\textnormal{a} = \lb \lp\Nens-1\rp \cdot \I + \Q^T \cdot \R^{-1} \cdot \Q\rb^{-1} \in \Re^{\Nens \times \Nens},\,
\end{eqnarray}
where $\Q = \H \cdot {\bf U}^\textnormal{b} \in \Re^{\Nobs \times \Nens}$ and $\I$ is the identity matrix consistent with the dimension. The analysis increments in the subspace are:
\begin{eqnarray}
\W^\textnormal{a} = \widehat{\P}^\textnormal{a} \cdot \Q^T \cdot \R^{-1} \cdot \lb \y - \Ho (\xm^\textnormal{b}) \rb \in \Re^{\Nens \times 1} ,\,
\end{eqnarray}
from which an estimation of the analysis mean in the model space can be obtained:
\begin{eqnarray}
\label{eq:LETKF-analysis-mean}
\displaystyle
\xm^\textnormal{a} = \xm^\textnormal{b} + {\bf U}^\textnormal{b} \cdot \W^\textnormal{a} \in \Re^{\Nstate \times 1} .\,
\end{eqnarray}
Finally, the analysis ensemble reads:
\begin{eqnarray}
\label{eq:LETKF-analysis-ensemble}
\x^\textnormal{a} = \xm^\textnormal{a} \cdot {\bf 1}_{\Nens}^T + {\bf U}^\textnormal{b} \cdot \lb \lp \Nens-1\rp \cdot \widehat{\P}^\textnormal{a} \rb^{1/2} \in \Re^{\Nstate \times \Nens} .\,
\end{eqnarray}
\end{subequations}
The domain localization in the LETKF is performed as follows: each model component is surrounded by a local box of radius ${\zeta}$. Within each local domain the analysis equations \eqref{eq:LETKF-method} are applied, and therefore a local analysis component is obtained. All local analysis components are mapped back onto the model space to obtain the global analysis state. Local boxes for different radii are shown in Figure \ref{fig:LETKF-radius}. The local sample covariance matrix \eqref{eq:covariance-ensemble} is utilized as the covariance estimator of the local $\B$. This can perform well when small radii ${\zeta}$ are considered during the assimilation step. However, for large values of ${\zeta}$, the analysis corrections can be impacted by spurious correlations since the local sample covariance matrix can be rank deficient. Consequently, the local analysis increments can perform poorly.
\begin{figure}[H]
\centering
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/grid_1-eps-converted-to.pdf}
\caption{${\zeta}=1$}
\end{subfigure}%
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/grid_2-eps-converted-to.pdf}
\caption{${\zeta}=2$}
\label{subfig:grid-2}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/grid_3-eps-converted-to.pdf}
\caption{${\zeta}=3$}
\label{subfig:grid-3}
\end{subfigure}
\caption{Local domains for different radii of influence ${\zeta}$. The red dot is the model component to be assimilated, blue components are within the scope of ${\zeta}$, and black model components are unused during the local assimilation process.}
\label{fig:LETKF-radius}
\end{figure}
There is an opportunity to reduce the impact of sampling errors by improving the background error covariance estimation. We achieve this by making use of the modified Cholesky decomposition for inverse covariance matrix estimation \cite{bickel2008}. Consider a sample of $\Nens$ Gaussian random vectors:
\begin{eqnarray*}
\label{eq:intro-samples}
\displaystyle
\S = \lb {\bf s}^{[1]},\, {\bf s}^{[2]},\, \ldots ,\, {\bf s}^{[\Nens]} \rb \in \Re^{\Nstate \times \Nens} ,\,
\end{eqnarray*}
with statistical moments:
\begin{eqnarray*}
{\bf s}^{[j]} \sim \Nor \lp {\bf 0}_{\Nstate} ,\, \Q \rp, \, \text{for $1 \le j \le \Nens$} ,\,
\end{eqnarray*}
where ${\bf s}^{[j]} \in \Re^{\Nstate \times 1}$ denotes the $j$-th sample. Denote by $\x_{[i]} \in \Re^{\Nens \times 1}$ the vector holding the $i$-th component across all the samples (the $i$-th row of $\S$, transposed). The modified Cholesky decomposition arises from regressing each component on his predecessors according to some component ordering:
\begin{eqnarray}
\label{eq:intro-modified-Chokesky-decomposition}
\displaystyle
\x_{[i]} = \sum_{j=1}^{i-1} \x_{[j]} \cdot \beta_{i,j} + \boldsymbol{\varepsilon}_{[i]} \in \Re^{\Nens \times 1} ,\,\quad 2 \le i \le \Nstate,
\end{eqnarray}
where $\x_{[j]}$ is the $j$-th model component which precedes $\x_{[i]}$ for $1 \le j \le i-1$, $\boldsymbol{\varepsilon}_{[1]} = \x_{[1]}$, and $\boldsymbol{\varepsilon}_{[i]} \in \Re^{\Nens \times 1}$ is the error in the $i$-th component regression for $i \ge 2$. Likewise, the coefficients $\beta_{i,j}$ in \eqref{eq:intro-modified-Chokesky-decomposition} can be computed by solving the optimization problem:
\begin{eqnarray}
\label{eq:intro-coefficient-computation}
\displaystyle
{\boldsymbol \beta}_{[i]} = \underset{{\boldsymbol \beta}}{\arg\,\min} \ln \x_{[i]} - \Z_{[i]} \cdot {\boldsymbol \beta} \ \big\|^2_2
\end{eqnarray}
where
\begin{eqnarray*}
\displaystyle
\Z_{[i]} &=& \lb \x^{[1]},\, \x^{[2]},\, \ldots,\, \x^{[i-1]}\rb^T \in \Re^{(i-1) \times \Nens} ,\quad 2 \le i \le \Nstate, \\
\displaystyle
{\boldsymbol \beta}_{[i]} &=& \lb \beta_{i,1},\, \beta_{i,2},\, \ldots ,\, \beta_{i,i-1} \rb^{T} \in \Re^{(i-1) \times 1} .\,
\end{eqnarray*}
The regression coefficients form the lower triangular matrix
\begin{subequations}
\label{eq:intro-empirical-T-D}
\begin{eqnarray}
\label{eq:intro-lower-T}
\big\{ \widehat{\bf T} \big\}_{i,j} = \left\{
\begin{aligned}
-\beta_{i,j} & \text{ for }1 \le j < i, \\
1 & \text{ for } j=i, \\
0 & \text{ for } j>i,
\end{aligned} \right. \quad 1 \le i \le \Nstate,
\end{eqnarray}
where $\big\{ \widehat{\bf T} \big\}_{i,j}$ denotes the $(i,j)$-th component of matrix $\widehat{\bf T} \in \mathbb{R}^{\Nstate \times \Nstate}$. The empirical variances $\widehat{\bf cov}$ of the residuals $\boldsymbol{\varepsilon}_{[i]}$ form the diagonal matrix:
\begin{eqnarray}
\label{eq:intro-diagonal-D}
\widehat{\bf D} = \underset{1 \le i \le \Nstate}{\textnormal{diag}}\left( \widehat{\bf cov} ( \boldsymbol{\varepsilon}_{[i]}) \right) = \underset{1 \le i \le \Nstate}{\textnormal{diag}}\left( \frac{1}{\Nens-1} \sum_{j=1}^{\Nens} \big\{\boldsymbol{\varepsilon}_{[i]} \big\}^2_{j} \right) \in \mathbb{R}^{\Nstate \times \Nstate}\,.
\end{eqnarray}
\end{subequations}
where $\lle \widehat{\bf D} \rle_{1,1} = \widehat{\bf cov}\lp \x_{[1]}\rp$. Then an estimate of $\Q^{-1}$ can be computed as follows:
\begin{subequations}
\label{eq:intro-Q-approximations}
\begin{eqnarray}
\label{eq:intro-Q-inverse}
\displaystyle
{\bf \widehat{\bf Q}}^{-1} = \widehat{\bf T}^T \cdot \widehat{\bf D}^{-1} \cdot \widehat{\bf T} \in \Re^{\Nstate \times \Nstate} ,\,
\end{eqnarray}
or, by basic matrix algebra identities the estimate of $\Q$ reads:
\begin{eqnarray}
\displaystyle
{\bf \widehat{\bf Q}} = \widehat{\bf T}^{-1} \cdot \widehat{\bf D} \cdot \widehat{\bf T}^{-T} \in \Re^{\Nstate \times \Nstate} \,.
\end{eqnarray}
\end{subequations}
Note that the structure of ${\bf \widehat{\bf Q}}^{-1}$ is strictly related to the structure of $\widehat{\bf T}$. This can be exploited in order to obtain sparse estimators of $\Q^{-1}$ by imposing that some entries of $\widehat{\bf T}$ are zero. This is important for high dimensional probability distributions where the explicit computation of ${\bf \widehat{\bf Q}}$ or ${\bf \widehat{\bf Q}}^{-1}$ is prohibitive. The zero components in $\widehat{\bf T}$ can be justified as follows: when two components are {\it conditionally independent} their corresponding entry in ${\bf \widehat{\bf Q}}^{-1}$ is zero. In the context of data assimilation, the conditional independence of background errors between different model components can be achieved by making use of domain localization. We can consider zero correlations between background errors corresponding to model components located at distances that exceed a radius of influence ${\zeta}$. In the next section we present an ensemble Kalman filter implementation based on modified Cholesky decomposition for inverse covariance matrix estimation.
\section{Ensemble Kalman Filter Based On Modified Cholesky Decomposition}
\label{sec:EnKF-MC}
In this section we discuss the new ensemble Kalman filter based on modified Cholesky decomposition for inverse covariance matrix estimation ( EnKF-MC).
\subsection{Estimation of the inverse background covariance}
The columns of matrix \eqref{eq:matrix-of-member-deviations}
\begin{eqnarray*}
{\bf U}^\textnormal{b} = \lb {\bf u}^{\textnormal{b}[1]},\,{\bf u}^{\textnormal{b}[2]},\, \ldots ,\, {\bf u}^{\textnormal{b}[\Nens]} \rb \in \Re^{\Nstate \times \Nens}
\end{eqnarray*}
can be seen as samples of the (approximately normal) distribution:
\begin{eqnarray*}
\x^{\textnormal{b}[j]} - \xm^\textnormal{b} = {\bf u}^{\textnormal{b}[j]} \sim \Nor \lp {\bf 0},\, \B \rp,\, \text{ for $1 \le j \le \Nens$} \,,
\end{eqnarray*}
and therefore, if we let $\x_{[i]} \in \Re^{\Nens \times 1}$ in \eqref{eq:intro-modified-Chokesky-decomposition} to be the vector formed by the $i$-th row of matrix \eqref{eq:matrix-of-member-deviations}, for $1 \le i \le \Nstate$, according to equations \eqref{eq:intro-Q-approximations}, an estimate of the inverse background error covariance matrix reads:
\begin{subequations}
\label{eq:EnKF-MC-background-cov}
\begin{eqnarray}
\label{eq:EnKF-MC-inverse-background-cov}
\displaystyle \B^{-1} \approx \widehat{\bf B}^{-1} = \widehat{\bf T}^T \cdot \widehat{\bf D}^{-1} \cdot \widehat{\bf T} \in \Re^{\Nstate \times \Nstate} ,\,
\end{eqnarray}
and therefore:
\begin{eqnarray}
\label{eq:EnKF-MC-background-cov}
\displaystyle
\B \approx \widehat{\bf B} = \widehat{\bf T}^{-1} \cdot \widehat{\bf D} \cdot \widehat{\bf T}^{-T} \in \Re^{\Nstate \times \Nstate} \,.
\end{eqnarray}
\end{subequations}
As we mentioned before, the structure of $\widehat{\bf B}^{-1}$ depends on that of $\widehat{\bf T}$. If we assume that the correlations between model components are local, and there are no correlations outside a radius of influence ${\zeta}$, we obtain lower-triangular sparse estimators of $\widehat{\bf T}$. Consequently, the resulting $\widehat{\bf B}^{-1}$ will also be sparse, and $\widehat{\bf B}$ will be localized. Since the regression \eqref{eq:intro-modified-Chokesky-decomposition} is performed only on the predecessors of each model component, an ordering (labeling) must be set on the model components prior the computation of $\widehat{\bf T}$. Since we work with gridded models we consider column-major and row-major orders. They are illustrated in Figure \ref{fig:ordering} for a two-dimensional domain. Figure \ref{fig:predecessors} shows the local domain and the predecessors of the model component 6 when column-major order is utilized.
\begin{figure}[H]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=0.6\textwidth,height=0.6\textwidth]{figures/Ordering-Column.png}
\caption{Column-major order}
\end{subfigure}%
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=0.6\textwidth,height=0.6\textwidth]{figures/Ordering-Row.png}
\caption{Row-major order}
\label{subfig:row-major-order}
\end{subfigure}
\caption{Row-major and column-major ordering for a $4 \times 4$ domain. The total number of model components is 16. }
\label{fig:ordering}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=0.6\textwidth,height=0.6\textwidth]{figures/Ordering_radius.png}
\caption{In blue, local box for the model component 6 when ${\zeta}=2$.}
\end{subfigure} \hspace{1em}
\begin{subfigure}{0.4\textwidth}
\centering
\includegraphics[width=0.6\textwidth,height=0.6\textwidth]{figures/Ordering_regression.png}
\caption{In blue, predecessors of the model component 6 for ${\zeta}=2$.}
\end{subfigure}
\caption{Local model components (local box) and local predecessors for the model component 6 when ${\zeta}=2$. Column-major ordering is utilized to label the model components.}
\label{fig:predecessors}
\end{figure}
The estimation of $\widehat{\bf B}^{-1}$ proceeds as follows:
\begin{enumerate}
\item Form the matrix $\Z_{[i]} \in \Re^{p_i \times \Nens}$ with the predecessors of the $i$-th model component:
\begin{eqnarray}
\label{eq:EnKF-MC-predecessors}
\Z_{[i]} = \lb \x^{[q(i,1)]},\,\x^{[q(i,2)]},\, \ldots ,\, \x^{[q(i,p_i)]} \rb^T \in \Re^{p_i \times \Nens} \,,
\end{eqnarray}
where $\x^{[e]}$ is the $e$-th row of matrix \eqref{eq:matrix-of-member-deviations}, $p_i$ is the number of predecessors of component $i$, and $1 \le q(i,j) \le \Nstate$ is the index (row of matrix \eqref{eq:matrix-of-member-deviations}) of the $j$-th predecessor of the $i$-th model component.
\item For the $i$-th model components the regression coefficients are obtained as follows:
\begin{eqnarray*}
\x_{[i]} = \sum_{j=1}^{p_i} \beta_{i,j} \cdot \x^{[q(i,j)]} + \boldsymbol{\varepsilon}_{[i]} \in \Re^{\Nens \times 1} \,.
\end{eqnarray*}
For $2 \le i \le \Nstate$, compute ${\boldsymbol \beta}_{[i]} = [\beta_{i,1},\,\beta_{i,2},\,\ldots,\,\beta_{i,p_i}] \in \Re^{p_i \times 1}$ by solving the optimization problem \eqref{eq:intro-coefficient-computation} with $\Z_{[i]}$ given by \eqref{eq:EnKF-MC-predecessors}.
\item Build the matrices
\begin{eqnarray*}
\big\{ \widehat{\bf T} \big\}_{i,q(i,j)}= -\beta_{i,j} ~~ \text{ for}~~1 \le j \le p_i, ~~1 < i \le \Nstate \,; \quad \big\{ \widehat{\bf T} \big\}_{i,i}=1,
\end{eqnarray*}
and $\widehat{\bf D}$ according to equation \eqref{eq:intro-diagonal-D}. Note that the number of non-zero elements in the $i$-th row of $\widehat{\bf T}$ equals the number of predecessors $p_i$.
\end{enumerate}
Note that the solution of the optimization problem \eqref{eq:intro-coefficient-computation} can be obtained as follows:
\begin{eqnarray}
\label{eq:EnKF-MC-solution-of-optimization-problem}
{\boldsymbol \beta}_{[i]} = \lb \Z_{[i]} \cdot { \Z_{[i]} }^T \rb^{-1} \cdot \Z_{[i]} \cdot \x_{[i]}
\end{eqnarray}
and since the ensemble size can be smaller than the number of model components, $\Z_{[i]} \cdot { \Z_{[i]} }^T \in \Re^{p_i \times p_i} $ can be rank deficient. To overcome this situation, regularization of the zero singular values of $\Z_{[i]} \cdot {\Z_{[i]}}^T$ can be used. One possibility is Tikhonov regularization \cite{Tik1,Tik2,Tik3}:
\begin{eqnarray}
\label{eq:EnKF-MC-optimization-tik}
{\boldsymbol \beta}_{[i]} = \underset{{\boldsymbol \beta}}{\arg\,\min} \lle \ln \x_{[i]} - \Z_{[i]} \cdot {\boldsymbol \beta} \big\|^2_2 + \lambda^2 \cdot \ln {\boldsymbol \beta} \big\|_2^2 \rle
\end{eqnarray}
where $\lambda \in \Re$. In our context the best choice for $\lambda$ relies on prior knowledge of the background and the observational errors \cite{Tik4}. Another approach to regularization is to use a truncated singular value decomposition (SVD) of $\Z_{[i]}$:
\begin{eqnarray*}
\Z_{[i]} = {\bf U}^{\Z_{[i]}} \cdot \boldsymbol{\Sigma}^{\Z_{[i]}} \cdot {{\bf V}^{\Z_{[i]}}}^T \in \Re^{p_i \times \Nens},
\end{eqnarray*}
where ${\bf U}^{\Z_{[i]}} \in \Re^{p_i \times p_i}$ and ${\bf V}^{\Z_{[i]}} \in \Re^{\Nens \times \Nens}$ are the right and the left singular vectors of $\Z_{[i]}$, respectively. Likewise, $\boldsymbol{\Sigma}^{\Z_{[i]}} \in \Re^{p_i \times \Nens}$ is a diagonal matrix whose diagonal entries are the singular values of $\Z_{[i]}$ in descending order. The solution of \eqref{eq:intro-coefficient-computation} can be computed as follows \cite{Jiang2000137,VANHUFFEL1991675,Per1}:
\begin{eqnarray}
\label{eq:EnKF-MC-truncated-SVD}
{\boldsymbol \beta}_{[i]} =\sum_{j=1}^{k_i} \frac{1}{\tau_j} \cdot {\bf u}^{\Z_{[i]}}_j \cdot {{\bf v}_j^{\Z_{[i]}}}^T \cdot \x_{[i]} \quad \text{with}\,\frac{\tau_{j}}{\tau_{\rm max}} \ge \sigma_r,\,
\end{eqnarray}
where $\tau_j$ is the $j$-th singular value with corresponding right and left singular vectors ${\bf u}^{\Z_{[i]}}_j \in \Re^{p_i \times 1}$ and ${\bf v}^{\Z_{[i]}}_j \in \Re^{p_i \times 1}$, respectively, $\sigma_r \in (0,1)$ is a predefined threshold, and $\tau_{\rm max} = \max\lle \tau_1,\, \tau_2,\, \ldots,\, \tau_{\Nens-1}\rle$. Since small singular values are more sensitive to the noise in $\x_{[i]}$, the threshold $\tau_j > \tau_{\max} \cdot \sigma_r$ seeks to neglect their contributions.
\subsection{Formulation of EnKF-MC}
Once $\widehat{\bf B}^{-1}$ is estimated, the EnKF based on modified Cholesky decomposition (EnKF-MC) computes the analysis using Kalman's formula:
\begin{subequations}
\label{eq:EnKF-MC-formulations}
\begin{eqnarray}
\label{eq:EnKF-MC-primal-incremental}
\x^\textnormal{a} &=& \x^\textnormal{b} + \AE \cdot \H^T \cdot \R^{-1} \cdot \boldsymbol{\Delta} \in \Re^{\Nstate \times \Nens},
\end{eqnarray}
where $\AE \in \Re^{\Nstate \times \Nstate}$ is the estimated analysis covariance matrix
\begin{eqnarray*}
\displaystyle
\AE = \lb \widehat{\bf B}^{-1} +\H^T \cdot \R^{-1} \cdot \H \rb^{-1}\,,
\end{eqnarray*}
and $\boldsymbol{\Delta} \in \Re^{\Nobs \times \Nens}$ is the innovation matrix on the perturbed observations given in \eqref{eq:EnKF-innvo-observations}.
Computationally-friendlier alternatives to \eqref{eq:EnKF-MC-dual} can be obtained by making use of elementary matrix identities:
\begin{eqnarray}
\label{eq:EnKF-MC-primal}
\displaystyle
\x^\textnormal{a} &=& \AE \cdot \lb \widehat{\bf B}^{-1} \cdot \x^\textnormal{b} + \H^T \cdot \R^{-1} \cdot \Y^\textnormal{s} \rb \in \Re^{\Nstate \times \Nens},\, \\
\label{eq:EnKF-MC-dual}
\displaystyle
\x^\textnormal{a} &=& \x^\textnormal{b} + \widehat{\bf T}^{-1} \cdot \widehat{\bf D}^{1/2} \cdot {\bf V}_{\widehat{\bf B}}^T \cdot \lb \R + {\bf V}_{\widehat{\bf B}} \cdot {\bf V}_{\widehat{\bf B}}^T \rb^{-1} \cdot \boldsymbol{\Delta},\, \\
\notag
{\bf V}_{\widehat{\bf B}} &=& \H \cdot \widehat{\bf T}^{-1} \cdot \widehat{\bf D}^{1/2} \in \Re^{\Nstate \times \Nobs},
\end{eqnarray}
\end{subequations}
where $\Y^\textnormal{s}$ are the perturbed observations. The formulation \eqref{eq:EnKF-MC-dual} is well-known as the EnKF dual formulation, \eqref{eq:EnKF-MC-primal} is known as the EnKF primal formulation, and the equation \eqref{eq:EnKF-MC-primal-incremental} is the incremental form of the primal formulation. In the next subsection, we discuss the computational effort of the EnKF-MC implementations \eqref{eq:EnKF-MC-formulations}.
\subsection{Computational effort of EnKF-MC implementations}
The computational cost of the different EnKF-MC implementations depend, in general, on the model state dimension $\Nstate$, the number of observed components $\Nobs$, the radius of influence ${\zeta}$, and the ensemble size $\Nens$. Typically \cite{Tippett2003} the data error covariance matrix $\R$ has a simple structure (e.g., block diagonal), the ensemble size is much smaller than the model dimension ($\Nstate \gg \Nens$), and the observation operator $\H$ is sparse or can be applied efficiently. We analyze the computational effort of the formulation \eqref{eq:EnKF-MC-primal-incremental}; similar analyses can be carried out for the other formulations. The incremental formulation can be written as follows:
\begin{eqnarray*}
\x^\textnormal{a} = \x^\textnormal{b} + {{\boldsymbol \delta} {\bf X}}^\textnormal{a} \,,
\end{eqnarray*}
where the analysis increments ${{\boldsymbol \delta} {\bf X}}^\textnormal{a} \in \Re^{\Nstate \times \Nens}$ are given by the solution of the linear system:
\begin{eqnarray*}
\lb \widehat{\bf B}^{-1} + \R_{\H} \cdot \R_{\H}^T\rb \cdot {{\boldsymbol \delta} {\bf X}}^\textnormal{a} = \boldsymbol{\Delta}_{\H} \,.
\end{eqnarray*}
with $\R_{\H} = \H^T \cdot \R^{-1/2} \in \Re^{\Nstate \times \Nobs}$, $\boldsymbol{\Delta}_{\H} = \H^T \cdot \R^{-1} \cdot \boldsymbol{\Delta} \in \Re^{\Nstate \times \Nens}$, and $\boldsymbol{\Delta}$ is given in \eqref{eq:EnKF-innvo-observations}. This linear system can be solved making use of the iterative Sherman Morrison formula \cite{Nino2014} as follows:
\begin{enumerate}
\item Compute:
\begin{subequations}
\label{eq:EnKF-MC-linear-system-primal}
\begin{eqnarray}
{\bf W}^{(0)[i]}_{\Z} &=& \lb \widehat{\bf T}^T \cdot {\bf D}^{-1} \cdot \widehat{\bf T} \rb^{-1} \cdot \boldsymbol{\Delta}_{\H}^{[i]} ,\, \text{ for $1 \le i \le \Nens$},\, \\
{\bf W}^{(0)[j]}_{{\bf U}} &=& \lb \widehat{\bf T}^T \cdot {\bf D}^{-1} \cdot \widehat{\bf T} \rb^{-1} \cdot \R_{\H}^{[j]} ,\, \text{ for $1 \le j \le \Nobs$} \,.
\end{eqnarray}
\end{subequations}
where $\boldsymbol{\Delta}_{\H}^{[i]} \in \Re^{\Nstate \times 1}$ and $\R_{\H}^{[j]} \in \Re^{\Nstate \times 1}$ denote the $i$ and $j$ columns of matrices $\boldsymbol{\Delta}_{\H}$ and $\R_{\H}$, respectively. Since $\widehat{\bf T}$ is a sparse unitary lower triangular matrix, the direct solution of the linear system \eqref{eq:EnKF-MC-linear-system-primal} can be obtained by making use of forward and backward substitutions. Hence, this step can be performed with:
\begin{eqnarray}
\label{eq:EnKF-MC-computational-effort-step-1}
\displaystyle
\BO{\Nstate_{nz} \cdot \Nstate \cdot \Nens + \Nstate_{nz} \cdot \Nstate \cdot \Nobs }
\end{eqnarray}
long computations, where $\Nstate_{nz}$ denotes the maximum number of non-zero elements across all rows of $\widehat{\bf T}$, this is
\begin{eqnarray*}
\Nstate_{nz} = \max \lle p_1,\,p_2,\,\ldots,\,p_{\Nstate} \rle
\end{eqnarray*}
where $p_i$ is the number of predecessors of model component $i$, for $1 \le i \le \Nstate$.
\item For $1 \le i \le \Nobs$ compute:
\begin{eqnarray*}
{\bf h}^{(i)} &=& \frac{1}{\gamma^{(i)}} \cdot {\bf W}_{{\bf U}}^{(i-1)[i]},\, \text{with } \gamma^{(i)} =\lb 1+{\R_{\H}^{[i]}}^T \cdot {\bf W}_{{\bf U}}^{(i-1)[i]} \rb^{-1} \,, \\
{\bf W}_{\Z}^{(i)[j]} &=& {\bf W}_{\Z}^{(i-1)[j]} - {\bf h}^{(i)} \cdot \lb {\R_{\H}^{[i]}}^T \cdot {\bf W}_{\Z}^{(i-1)[j]} \rb ,\, \text{ for $1 \le j \le \Nens$}\,, \\
{\bf W}_{{\bf U}}^{(i)[k]} &=& {\bf W}_{{\bf U}}^{(i-1)[k]} - {\bf h}^{(i)} \cdot \lb {\R_{\H}^{[i]}}^T \cdot {\bf W}_{{\bf U}}^{(i-1)[k]} \rb ,\, \text{ for $i+1 \le k \le \Nobs$}\,. \\
\end{eqnarray*}
Note that, at each step, ${\bf h}^{(i)}$ can be computed with $\Nstate$ long computations, while ${\bf W}_{\Z}$ and ${\bf W}_{{\bf U}}$ can be obtained with $\Nstate \cdot \Nens$ and $\Nstate \cdot \Nobs$ long computations, respectively. This leads to the next bound for the number of long computations:
\begin{eqnarray*}
\BO{\Nobs \cdot \Nstate+\Nobs \cdot \Nstate \cdot \Nens + \Nobs^2 \cdot \Nstate} \,.
\end{eqnarray*}
\end{enumerate}
Hence, the computational effort involved during the assimilation step of formulation \eqref{eq:EnKF-MC-primal-incremental} can be bounded by:
\begin{eqnarray*}
\BO{\Nobs \cdot \Nstate+\Nobs \cdot \Nstate \cdot \Nens + \Nobs^2 \cdot \Nstate+\Nstate_{nz} \cdot \Nstate \cdot \Nens + \Nstate_{nz} \cdot \Nstate \cdot \Nobs},\,
\end{eqnarray*}
which is linear with respect to the number of model components. For dense observational networks, when local observational operators can be approximated, domain decomposition can be exploited in order to reduce the computational effort during the assimilation cycle. This can be done as follows:
\begin{enumerate}
\item The domain is split in certain number of sub-domains (typically matching a given number of processors).
\item Background error correlations are estimated locally.
\item The assimilation is performed on each local domain.
\item The analysis sub-domains are mapped back onto the model domain from which the global analysis state is obtained.
\end{enumerate}
Figure \ref{fig:sub-domain-splitting} shows the global domain splitting for different sub-domain sizes. In Figure \ref{subfig:boundary-information} the boundary information needed during the assimilation step for two particular sub-domains is shown in dashed blue lines. Note that each sub-domain can be assimilated independently. Note that we only use domain decomposition in order to reduce the computational effort of the proposed implementation (and its derivations) and not in order to reduce the impact of spurious correlations.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.625\textwidth]{figures/grid12.png}
\caption{Number of sub-domains 12}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.625\textwidth]{figures/grid80.png}
\caption{Number of sub-domains 80}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/grid16.png}
\caption{Number of sub-domains 16.}
\label{subfig:boundary-information}
\end{subfigure}%
\caption{Global domain splitting in different sub-domain sizes. Blue local boxes reflects the boundary information utilized in order to perform local data assimilation.}
\label{fig:sub-domain-splitting}
\end{figure}
\subsection{Convergence of the covariance inverse estimator}
\label{subsec:proof-of-convergence}
In this section we prove the convergence of the $\widehat{\bf B}^{-1}$ estimator in the context of data assimilation.
\begin{comment}[Sparse Cholesky factors and localization]
The modified Cholesky decomposition for inverse covariance matrix estimation can be seen as a form of covariance matrix localization method in which the resulting matrix approximates the inverse of a localized ensemble covariance matrix. This process is implicit in the resulting estimator when only a local neighborhood for each model component is utilized in order to perform the local regression and to estimate $\widehat{\bf T}$ and $\widehat{\bf D}$. Figure \ref{fig:theo-Cholesky-factors} shows an example for the Lorenz 96 \cite{LorenzModel}:
\begin{eqnarray}
\label{eqch4:Lorenz-model}
\displaystyle
\frac{dx_j}{dt} = \begin{cases}
\lp x_2-x_{\Nstate-1}\rp \cdot x_{\Nstate} -x_1 +F & \text{ for $j=1$} \\
\lp x_{j+1}-x_{j-2}\rp \cdot x_{j-1} -x_j +F & \text{ for $2 \le j \le \Nstate-1$} \\
\lp x_1-x_{\Nstate-2}\rp \cdot x_{\Nstate-1} -x_{\Nstate} +F & \text{ for $j=\Nstate$}
\end{cases}
\end{eqnarray}
where $F$ is usually set to $8$ to exhibit chaotic behavior and the number of model components is $\Nstate=40$. We assume $\B$ to be a sample covariance matrix based on $10^5$ samples, the localized ensemble covariance matrix $\P^\textnormal{b}$ and the estimator $\widehat{\bf B}^{-1}$ are based on 80 samples. The radius of influence is ${\zeta} = 7$. The similarities among the different Cholesky factors is evident. Even more, along the main diagonal, the correlations decay with respect to the distance of the model components. This is reflected in the resulting estimator of $\B^{-1}$ for each case. Definition \ref{theo-def} of covariance matrices relies on this assumpation for the Cholesky factors ${\bf T}$ and $\widehat{\bf T}$.
\end{comment}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth,height=0.8\textwidth]{figures/theo_BE.png}
\caption{Exact $\B^{-1} \approx {\P^b}^{-1}$ for $\Nens={10^5}$}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth,height=0.8\textwidth]{figures/theo_BE_T.png}
\caption{${\bf T}$, $\B^{-1} = {\bf T}^T \cdot {\bf D}^{-1} \cdot {\bf T}$}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth,height=0.8\textwidth]{figures/theo_PBLI.png}
\caption{Localized ensemble estimate $\widehat{\P^\textnormal{b}}^{-1}$}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth,height=0.8\textwidth]{figures/theo_PBLI_T.png}
\caption{${\bf T}_{\bf L}$, $\widehat{\P^\textnormal{b}}^{-1} = {\bf T}_{\bf L}^T \cdot {\bf D}_{\bf L} \cdot {\bf T}_{\bf L}$}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth,height=0.8\textwidth]{figures/theo_PBLI.png}
\caption{Cholesky estimate $\widehat{\bf B}^{-1}$}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth,height=0.8\textwidth]{figures/theo_PBLI_T.png}
\caption{$\widehat{\bf T}$, $\widehat{\bf B}^{-1} = \widehat{\bf T}^T \cdot \widehat{\bf D}^{-1} \cdot \widehat{\bf T}$}
\end{subfigure}
\caption{Decay of correlations in the Cholesky factors for different approximations of $\B^{-1}$. }
\label{fig:theo-Cholesky-factors}
\end{figure}
We consider a two-dimensional square domain with $s \times s$ grid points. Our proof below can be extended immediately to non-square domains, as well as to three-dimensional domains. In our domain each space point is described by two indices $(i,\,j)$, a zonal component $i$ and a meridional component $j$, for $1 \le i,j \le s$. A particular case for $s = 4$ is shown in Figure \ref{subfig:grid-distribution}. We make use of row-major order in order to map model grid components to the one dimensional ``index space'':
\begin{eqnarray*}
k = f(i,j) = (j-1) \cdot s+i,\,\quad \text{for $1 \le k \le \Nstate$}.\,
\end{eqnarray*}
where here, $\Nstate = s^2$. For a particular grid component $(i,\,j)$, the resulting $k = f(i,j)$ denotes the row index in $\widehat{\bf B}^{-1}$. The results of labeling each model component in this manner can be seen in Figure \ref{subfig:row-major-order}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\fbox{\includegraphics[width=0.8\textwidth]{figures/HIGH_RES_GRID_2.png}}
\caption{Grid components $(i,j)$}
\label{subfig:grid-distribution}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\fbox{\includegraphics[width=0.45\textwidth]{figures/HIGH_RES_ORDER.png}}
\caption{Index space $f(i,j)$}
\label{fig:corresponding-index-B}
\end{subfigure}
\caption{Grid distribution of model components and corresponding index terms in $\widehat{\bf B}^{-1}$. }
\label{fig:ordering-example}
\end{figure}
To start our proof, the inverse of the (exact) background error covariance matrix $\B^{-1}$ and of the its estimator $\widehat{\bf B}^{-1}$ can be written as
\begin{subequations}
\begin{eqnarray}
\label{eq:proof-estimator-written}
\displaystyle
\widehat{\bf B}^{-1} = \lb \I- \widehat{\bf C} \rb^T \cdot \widehat{\bf D}^{-1} \cdot \lb \I - \widehat{\bf C}\rb \in \Re^{\Nstate \times \Nstate}
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:proof-estimator-written}
\displaystyle
\B^{-1} = \lb \I - {\bf C}\rb^T \cdot {\bf D}^{-1} \cdot \lb \I - {\bf C}\rb \in \Re^{\Nstate \times \Nstate} ,\,
\end{eqnarray}
\end{subequations}
respectively, where $\widehat{\bf C} = \I - \widehat{\bf T} \in \Re^{\Nstate \times \Nstate}$ and ${\bf C} = \I - {\bf T} \in \Re^{\Nstate \times \Nstate}$. Moreover, ${\bf D}$ and $\widehat{\bf D}$ are diagonal matrices:
\begin{eqnarray*}
{\bf D} &=& {\bf diag} \lle {d}_{1}^2,\, {d}_{2}^2,\, \ldots,\, {d}_{\Nstate}^2\rle \\
\widehat{\bf D} &=& {\bf diag} \lle \widehat{d}_{1}^2,\, \widehat{d}_{2}^2,\, \ldots,\, \widehat{d}_{\Nstate}^2\rle
\end{eqnarray*}
where $\lle {\bf D} \rle_{i,i} = d_i^2$ and $\lle \widehat{\bf D} \rle_{i,i} = \widehat{d}_i^2$, for $1 \le i \le \Nstate$. In what follows we denote by $\widehat{\bf c}^{\{j\}} \in \Re^{\Nstate \times 1}$ and ${\bf c}^{\{j\}} \in \Re^{\Nstate \times 1}$ the $j$-th columns of matrices $\widehat{\bf C}$ and ${\bf C}$, respectively, for $1 \le j \le \Nstate$.
\begin{subequations}
\label{eq:Proof-theorem-1}
\begin{definition}[Class of matrices under consideration.]
\label{theo-def}
We consider the class of covariance matrices matrices with correlations decreasing quickly:
\begin{eqnarray}
\label{eq:Proof-class-of-matrices}
\displaystyle
\mathcal{U}^{-1} \lp \varepsilon_0, C, \alpha \rp &=& \Bigg \{ \B: 0<\varepsilon_0 \le \lambda_{min} \lp \B \rp \le \lambda_{max} \lp \B \rp \le \varepsilon_0^{-1},\, \\ \nonumber
& & \underset{k}{\max} \sum_{\ell=1}^{\Nstate} \left | \gamma_{k,\ell} \cdot \lle {\bf T} \rle_{k,\ell} \right | \le C \cdot {{\zeta}}^{-\alpha},\, \text{ for ${\zeta} \le s-1$} \Bigg \}
\end{eqnarray}
where $\B^{-1} = {\bf T}^T \, {\bf D}^{-1} \, {\bf T}$, $\alpha$ is the decay rate (related to the dynamics of the numerical model),
\begin{eqnarray*}
\gamma_{k_{(i,j)},\ell_{(p,q)}} &=& \begin{cases}
0 & j-{\zeta} \le q \le j-1 \text{ and } i-{\zeta} \le p \le i+{\zeta} \\
0 & q = j \text{ and } i-{\zeta} \le p \le i \\
1 & \text{otherwise}
\end{cases} \,,
\end{eqnarray*}
and the grid components $(i,j)$ and $(p,q)$, for $1 \le i,j,p,q \le s$ are related to the $(k_{(i,j)},\ell_{(p,q)})$ matrix entry by $k_{(i,j)} = f(i,j)$ and $\ell = f(p,q)$.
\end{definition}
\begin{comment}
The factors $\gamma_{k,\ell}$ for the grid component $(i,j)$ in Definition \eqref{theo-def} are zero inside the scope of ${\zeta}$.
\end{comment}
\begin{theorem}[Error in the covariance inverse estimation]
\label{theo-main}
Uniformly for $\B \in \mathcal{U}^{-1} \lp \varepsilon_0, C, \alpha \rp$, if ${\zeta} \approx \lb \Nens^{-1} \cdot \log \Nstate \rb^{-1/2(\alpha+1)}$ and $\Nens^{-1} \cdot \log \Nstate = o(1)$,
\begin{eqnarray}
\label{eq:Proof-theorem}
\displaystyle
\ln \widehat{\bf B}^{-1} - \B^{-1}\big\|_{\infty} = \mathcal{O} \lp \lb \frac{\log(\Nstate)}{\Nens}\rb^{\alpha(\alpha+1)/2}\rp
\end{eqnarray}
where $\ln \cdot \big\|_{\infty}$ denotes the infinity norm (matrix or vector)
\end{theorem}
\end{subequations}
\begin{comment}
The factors $\gamma_{k,\ell}$ in Theorem \eqref{theo-main} are zero for the predecessors of the grid component $(i,j)$ inside the scope of ${\zeta}$.
\end{comment}
In order to prove Theorem \ref{theo-main}, we need the following result.
\begin{lemma}
\label{lemma:differences-emp-true}
Under the conditions of Theorem \ref{theo-main}, uniformly on $\mathcal{U}^{-1}$
\begin{subequations}
\label{eq:proof-differences-empirical}
\begin{eqnarray}
\label{eq:proof-max-T}
\displaystyle
&&\max \lle \ln \widehat{\bf c}^{\{j\}}-{\bf c}^{\{j\}} \big\|_{\infty} : 1\le j \le \Nstate \rle = \BO{\Nens^{-1/2} \log^{1/2} \Nstate},\, \\
\label{eq:proof-max-D}
&&\max \lle \left| \widehat{d}_{j}^2 - d_{j}^2\right |: 1\le j \le \Nstate \rle = \BO{ \lb \Nens^{-1} \log \Nstate \rb^{\alpha/(2(\alpha+1))}} ,\,
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:proof-norm1}
\displaystyle
\ln {\bf C} \big\|_{\infty} = \ln {\bf D}^{-1} \big\|_{\infty} = \BO{1}.
\end{eqnarray}
\end{subequations}
\end{lemma}
The proof of Lemma \ref{lemma:differences-emp-true} is based on the following results of Bickel and Levina in \cite{bickel2008}.
\begin{lemma}
\label{lemma:lemma-Bickel}[\cite[Lemma A.2]{bickel2008}]
Let $\errbac^{[k]} \sim \Nor \lp {\bf 0},\, \B\rp$ and $\lambda_{\max} \lp \B \rp \le \varepsilon_0^{-1} < \infty$, for $1 \le k \le \Nens$. Then, if $\lle \B \rle_{i,j}$ denotes the $(i,\,j)$-th component of $\B$, for $1 \le i \le j \le \Nstate$,
\begin{eqnarray}
\label{eq:lemma-Bickel}
\displaystyle
&& \textnormal{Prob} \lb \sum_{k=1}^{\Nens} \lb \lle \errbac^{[k]} \rle_i \cdot \lle \errbac^{[k]} \rle_j - \lle \B\rle_{i,j} \rb \ge \Nens \cdot \nu \rb \\
\nonumber
&& \qquad \le C_1 \cdot \exp \lp -C_2 \cdot \Nens \cdot \nu^2 \rp,\,
\end{eqnarray}
for $|\nu| \le \delta$, where $\lle \errbac^{[k]} \rle_i$ is the $i$-th component of the sample $\errbac^{[k]}$, for $1 \le k \le \Nens$, and $1 \le i \le \Nstate$. Likewise, $C_1$, $C_2$ and $\delta$ depend on $\varepsilon_0$ only.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lemma:differences-emp-true}]
In what follows we denote by ${\bf cov}$ and $\widehat{\bf cov}$ denote the true and the empirical covariances, respectively. In the context of EnKF we have that ${\bf cov} \lp {\bf U}^\textnormal{b} \rp = \B$.
Recall that
\begin{eqnarray*}
\widehat{\bf cov} \lp {\bf U}^\textnormal{b} \rp = \P^\textnormal{b} = \frac{1}{\Nens-1} \cdot {\bf U}^\textnormal{b} \cdot {{\bf U}^\textnormal{b}}^T = \frac{1}{\Nens-1} \cdot \sum_{k=1}^{\Nens} {\bf u}^{\textnormal{b}[k]} \cdot {{\bf u}^{\textnormal{b}[k]}}^T ,\,
\end{eqnarray*}
and therefore
\begin{eqnarray*}
\lle \widehat{\bf cov} \lp {\bf U}^\textnormal{b} \rp \rle_{i,j} = \frac{1}{\Nens-1} \cdot \sum_{k=1}^{\Nens} \lle {\bf u}^{\textnormal{b}[k]} \rle_{i} \cdot \lle {\bf u}^{\textnormal{b}[k]} \rle_{j}.
\end{eqnarray*}
For $\nu>0$, $\lle \errbac^{[k]} \rle_i \cdot \lle \errbac^{[k]} \rle_j - \lle \B\rle_{i,j} \ge \Nens \cdot \nu$ implies $\lle \errbac^{[k]} \rle_i \cdot \lle \errbac^{[k]} \rle_j - \lle \B\rle_{i,j} \ge (\Nens-1 )\cdot \nu$, and therefore by Lemma \ref{lemma:lemma-Bickel} we have:
\begin{subequations}
\begin{eqnarray}
\label{eq:proof-bound-U}
\ln {\bf cov} \lp {\bf U}^\textnormal{b} \rp - \widehat{\bf cov} \lp {\bf U}^\textnormal{b} \rp \big\|_{\infty} = \BO {\Nens^{-1/2} \cdot \log^{1/2} \Nstate },\,
\end{eqnarray}
since the entries of ${\bf cov} \lp {\bf U}^\textnormal{b} \rp - \widehat{\bf cov} \lp {\bf U}^\textnormal{b} \rp$ can be bounded by:
\begin{eqnarray*}
\left | \lle {\bf cov} \lp {\bf U}^\textnormal{b} \rp - \widehat{\bf cov} \lp {\bf U}^\textnormal{b} \rp \rle_{i,j} \right | & \le & \Nens^{-1} \cdot \sum_{k=1}^\Nens \left | \lle {\bf u}^{\textnormal{b}[k]} \rle_i \cdot \lle {\bf u}^{\textnormal{b}[k]} \rle_j - \lle \B \rle_{i,j} \right |.
\end{eqnarray*}
Lemma \ref{lemma:lemma-Bickel} ensures that:
\begin{eqnarray*}
&& \textnormal{Prob} \lb \underset{i,j}{\max} \left | \Nens^{-1} \cdot \sum_{k=1}^\Nens \lle {\bf u}^{\textnormal{b}[k]} \rle_i \cdot \lle {\bf u}^{\textnormal{b}[k]} \rle_j - \lle \B \rle_{i,j} \right | \ge \nu \rb \\
&& \qquad \le C_1 \cdot \Nstate^2 \cdot \exp \lp -C_2 \cdot \Nens \cdot \nu^2 \rp,\,
\end{eqnarray*}
for $|\nu| \le \delta$. Let $\nu = \lp \frac{\log \Nstate^2}{ \Nens \cdot C_2} \rp^{1/2}\cdot M$, for $M$ arbitrary.
Since $\Z_{[i]}$ stores the columns of ${\bf U}^\textnormal{b}$ corresponding to the predecessors of model component $i$, an immediate consequence of \eqref{eq:proof-bound-U} is
\begin{eqnarray}
\label{eq:proof-bound-Z}
\underset{i}{\max} \ln {\bf cov} \lp \Z_{[i]} \rp -\widehat{\bf cov} \lp \Z_{[i]} \rp \big\|_{\infty} = \BO{ \Nens^{-1/2} \cdot \log^{1/2} \Nstate} \,.
\end{eqnarray}
\end{subequations}
Also,
\begin{eqnarray*}
\ln \B^{-1} \big\|_{\infty} = \ln {\bf cov} \lp {\bf U}^\textnormal{b} \rp^{-1} \big\|_{\infty} \le \varepsilon_{0}^{-1} \,.
\end{eqnarray*}
According to equation \eqref{eq:EnKF-MC-solution-of-optimization-problem},
\begin{eqnarray*}
\lle {\bf c}^{[i]} \rle_{j} &=& \lle {\bf cov} \lp \Z_{[i]} \rp^{-1} \cdot \Z_{[i]} \cdot \x_{[i]} \rle_j \,,\\
\lle \widehat{\bf c}^{[i]} \rle_{j} &=& \lle \widehat{\bf cov} \lp \Z_{[i]} \rp^{-1} \cdot \Z_{[i]} \cdot \x_{[i]} \rle_j \,,
\end{eqnarray*}
therefore:
\begin{eqnarray} \nonumber
&&\underset{k}{\max} \left | \lle {\bf c}^{[i]} \rle_{k}-\lle \widehat{\bf c}^{[i]} \rle_{k} \right | \\
&=& \underset{k}{\max} \left | \lle {\bf cov} \lp \Z_{[i]} \rp^{-1} \cdot \Z_{[i]} \cdot \x_{[i]} \rle_k - \lle \widehat{\bf cov} \lp \Z_{[i]} \rp^{-1} \cdot \Z_{[i]} \cdot \x_{[i]} \rle_k \right | \\ \nonumber
&=& \underset{k}{\max} \left | \lle \lb {\bf cov} \lp \Z_{[i]} \rp^{-1} - \widehat{\bf cov} \lp \Z_{[i]} \rp^{-1} \rb \cdot \Z_{[i]} \cdot \x_{[i]} \rle_k \right | \\ \label{eq:proof-parta}
&=& \mathcal{O} \lp \Nens^{-1/2} \cdot \log^{1/2} \Nstate \rp
\end{eqnarray}
from which \eqref{eq:proof-max-T} follows. Note that:
\begin{eqnarray*}
&& \x_{[i]} = \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]} \rle_j \cdot \x_{[j]} + \widehat{\boldsymbol{\varepsilon}}^{[i]} \\
&\Leftrightarrow & \widehat{\bf cov} \lp \x_{[i]} \rp = \widehat{\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]} \rle_j \cdot \x_{[j]} + \widehat{\boldsymbol{\varepsilon}}^{[i]} \rp \\
&\Leftrightarrow & \widehat{\bf cov} \lp \x_{[i]} \rp = \widehat{\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]} \rle_j \cdot \x_{[j]} \rp + \widehat{\bf cov} \lp \widehat{\boldsymbol{\varepsilon}}^{[i]} \rp \\
&\Leftrightarrow & \widehat{d}^2_i = \widehat{\bf cov} \lp \x_{[i]}\rp - \widehat{\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]} \rle_j \cdot \x_{[j]} \rp ,\,
\end{eqnarray*}
and similarly
\begin{eqnarray*}
d^2_i = {\bf cov} \lp \x_{[i]}\rp - {\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle {\bf c}^{[i]} \rle_j \cdot \x_{[j]} \rp \,.
\end{eqnarray*}
The claim \eqref{eq:proof-max-D} and the first part of \eqref{eq:proof-norm1} follow from \eqref{eq:proof-bound-U}, \eqref{eq:proof-bound-Z} and \eqref{eq:proof-parta}. Since
\begin{eqnarray} \nonumber
\left | \widehat{d}_{i}^2 - d_{i}^2 \right | &\le & \left | {\bf cov} \lp \x_{[i]}\rp - \widehat{\bf cov} \lp \x_{[i]}\rp \right | \\
&+& \left | \widehat{\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lb \lle {\widehat{\bf c}}^{[i]} \rle_j - \lle {{\bf c}}^{[i]} \rle_j \rb \cdot \x_{[j]} \rp \right | \\ \nonumber
&+& \left | \widehat{\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]} \rle_j \cdot \x_{[j]} \rp - {\bf cov} \lp \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]}\rle_j \cdot \x_{[j]} \rp \right |
\end{eqnarray}
where $\tilde{\gamma}_{i,j} = 1-{\gamma}_{i,j}$. By Lemma \ref{lemma:lemma-Bickel} the maximum over $i$ of the first term is:
\begin{eqnarray*}
\displaystyle
\underset{i}{\max} \left | {\bf cov} \lp \x_{[i]}\rp - \widehat{\bf cov} \lp \x_{[i]}\rp \right | = \BO{\Nens^{-1/2} \cdot \log^{1/2} \Nstate }.\,
\end{eqnarray*}
The second term can be bounded as follows:
\begin{eqnarray*}
&& {\tiny \left | \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j}^2 \cdot \lb \lle {\widehat{\bf c}}^{[i]} \rle_j - \lle {{\bf c}}^{[i]} \rle_j \rb^2 \cdot \widehat{\bf cov}\lp \x_{[j]} \rp \right |} \\
& \le & \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j}^2 \cdot \lb \lle {\widehat{\bf c}}^{[i]} \rle_j - \lle {{\bf c}}^{[i]} \rle_j \rb^2 \cdot \left | \widehat{\bf cov}\lp \x_{[j]} \rp \right | \\
& \le & \underset{k}{\max} \lb \lle {\widehat{\bf c}}^{[i]} \rle_k - \lle {{\bf c}}^{[i]} \rle_k \rb^2 \cdot \underset{i}{\max}\left | \widehat{\bf cov}\lp \x_{[i]} \rp \right | \cdot \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j}^2 \\
&=&\BO{ {\zeta}^2 \cdot \Nens^{-1} \cdot \log \Nstate } \\ \nonumber
&=& \BO{ \lb \Nens^{-1} \cdot \log \Nstate \rb^{\alpha/2 \cdot (\alpha+1)} }
\end{eqnarray*}
by \eqref{eq:proof-max-T} and $\ln \B\big\| \le \varepsilon_0^{-1}$. Recall that ${\zeta} = \lb \Nens^{-1} \cdot \log \Nstate \rb^{1/2 \cdot(\alpha+1)}$ and even more, note that:
\begin{eqnarray*}
\sum_{j=1}^{\Nstate} \tilde{\gamma}^2_{i,j} = \frac{\lb {\zeta}+1 \rb^2}{2} = \frac{{\zeta}^2}{2}+{\zeta}+\frac{1}{2} = \BO{{\zeta}^2} \,.
\end{eqnarray*}
The third term can be bounded similarly. Thus \eqref{eq:proof-max-D} follows. Furthermore,
\begin{eqnarray*}
\displaystyle
d^2_{i} = {\bf cov} \lp \x_{[i]} - \sum_{j=1}^{\Nstate} \tilde{\gamma}_{i,j} \cdot \lle \widehat{\bf c}^{[i]}\rle_j \cdot \x_{[j]} \rp \ge \varepsilon_0 \cdot \lp 1 + \sum_{i=1}^{\Nstate} \lb \widehat{\bf c}^{[i]}_j \rb^2 \rp \ge \varepsilon_0 \,,
\end{eqnarray*}
and the lemma follows.
\end{proof}
We now are ready to prove Theorem \ref{theo-main}.
\begin{proof}[Proof of Theorem \ref{theo-main}]
We need only check that:
\begin{subequations}
\label{eq:Proof-to-check}
\begin{eqnarray}
\label{eq:Proof-difference-of-inverses}
\displaystyle
\ln \widehat{\bf B}^{-1} - \B^{-1} \big\|_{\infty} = \BO{ \Nens^{-1/2} \cdot \log^{1/2} \lp \Nstate \rp}
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:Proof-rate-radius}
\displaystyle
\ln \B^{-1} - \Phi_{{\zeta}} \lp \B^{-1} \rp\big\|_{\infty} = \BO{{\zeta}^{-\alpha}}
\end{eqnarray}
where the entries of $\Phi_{{\zeta}} \lp \B^{-1} \rp$ are given by:
\begin{eqnarray}
\lle \Phi_{{\zeta}} \lp \B^{-1} \rp \rle_{k,\ell} = \delta_{k,\ell} \cdot \lle \B^{-1} \rle_{k,\ell} ,\, \text{ for $1 \le k,\ell \le \Nstate$}
\end{eqnarray}
where $k = f(i,j)$ and $\ell = f(q,p)$ for $1 \le i,j,p,q \le s$, and
\begin{eqnarray*}
\delta_{k,\ell} &=& \begin{cases}
1 & j-{\zeta} \le q \le j+{\zeta} \text{ and } i-{\zeta} \le p \le i+{\zeta} \\
0 & \text{otherwise}
\end{cases}
\end{eqnarray*}
\end{subequations}
We first prove \eqref{eq:Proof-difference-of-inverses}. By definition,
\begin{eqnarray}
\widehat{\bf B}^{-1} - \B^{-1} = \widehat{\bf T}^T \cdot \widehat{\bf D}^{-1} \cdot \widehat{\bf T} - {\bf T}^T \cdot {\bf D}^{-1} \cdot {\bf T}.
\end{eqnarray}
Applying the standard inequality:
\begin{eqnarray*}
\ln {\bf T}^T \cdot {\bf D}^{-1} \cdot {\bf T} -\widehat{\bf T}^T \cdot \widehat{\bf D}^{-1} \cdot \widehat{\bf T}^T\big\| &\le & \ln {\bf T}^T- \widehat{\bf T}^T \big\| \cdot \ln \widehat{\bf D} \big\| \cdot \ln \widehat{\bf T} \big\| \\
&+& \ln {\bf D}- \widehat{\bf D} \big\| \cdot \ln \widehat{\bf T}^T \big\| \cdot \ln \widehat{\bf T} \big\| \\
&+& \ln {\bf T}- \widehat{\bf T} \big\| \cdot \ln \widehat{\bf T} \big\| \cdot \ln \widehat{\bf D} \big\| \\
&+& \ln \widehat{\bf T} \big\| \cdot \ln {\bf D}-\widehat{\bf D} \big\| \cdot \ln \widehat{\bf T}^T-{\bf T}^T \big\| \\
&+& \ln \widehat{\bf D} \big\| \cdot \ln {\bf T}-\widehat{\bf T} \big\| \cdot \ln \widehat{\bf T}^T-{\bf T}^T \big\| \\
&+& \ln \widehat{\bf T}^T \big\| \cdot \ln {\bf D}-\widehat{\bf D} \big\| \cdot \ln \widehat{\bf T}-{\bf T} \big\| \\
&+& \ln {\bf D}-\widehat{\bf D} \big\| \cdot \ln {\bf T}-\widehat{\bf T} \big\| \cdot \ln {\bf T}^T-\widehat{\bf T}^T \big\|
\end{eqnarray*}
all previous terms can be bounded making use of Lemma \ref{lemma:differences-emp-true} and therefore, \eqref{eq:Proof-difference-of-inverses} follows. Likewise, for \eqref{eq:Proof-rate-radius}, we need to note that for any matrix ${\bf M}$,
\begin{eqnarray*}
\ln {\bf M} \cdot {\bf M}^T - \Phi_{{\zeta}} \lp {\bf M} \rp \cdot \Phi_{{\zeta}} \lp {\bf M} \rp^T \big\|_{\infty} & \le & 2 \cdot \ln {\bf M} \big\|_{\infty} \cdot \ln \Phi_{{\zeta}} \lp {\bf M} \rp-{\bf M}^{-1} \big\|_{\infty} \\
&+&\ln \Phi_{{\zeta}} \lp {\bf M} \rp-{\bf M} \big\|_{\infty}^2
\end{eqnarray*}
and by letting ${\bf M} = {\bf T}^T \cdot {\bf D}^{-1/2}$, the theorem follows from Definition \ref{theo-def}.
\end{proof}
\section{Numerical Experiments}
\label{sec:experimental-settings}
In this section we study the performance of the proposed EnKF-MC implementation. The experiments are performed using the atmospheric general circulation model SPEEDY \cite{Speedy1,Speedy2}. SPEEDY is a hydrostatic, spectral coordinate, spectral transform model in the vorticity-divergence form, with semi-implicit treatment of gravity waves. The number of layers in the SPEEDY model is 8 and the T-63 model resolution ($192 \times 96$ grids) is used for the horizontal space discretization of each layer. Four model variables are part of the assimilation process: the temperature ($K$), the zonal and the meridional wind components ($m/s$), and the specific humidity ($g/kg$). The total number of model components is $\Nstate = 589,824$. The number of ensemble members is $\Nens=94$ for all the scenarios. The model state space is approximately 6,274 times larger than the number of ensemble members ($\Nstate \gg \Nens$).
Starting with the state of the system $\x^\textnormal{ref}_{-3}$ at time $t_{-3}$, the model solution $\x^\textnormal{ref}_{-3}$ is propagated in time over one year:
\begin{eqnarray*}
\x^\textnormal{ref}_{-2} = \M_{t_{-3} \rightarrow t_{-2}} \lp \x^\textnormal{ref}_{-3}\rp.
\end{eqnarray*}
The reference solution $\x^\textnormal{ref}_{-2}$ is used to build a perturbed background solution:
\begin{eqnarray}
\label{eq:exp-perturbed-background}
\displaystyle
\widehat{\x}^\textnormal{b}_{-2} = \x^\textnormal{ref}_{-2} + \errobs^\textnormal{b}_{-2}, \quad \errobs^\textnormal{b}_{-2} \sim \Nor \lp {\bf 0}_{\Nstate} ,\, \underset{i}{\textnormal{diag}} \left\{ (0.05\, \{\x^\textnormal{ref}_{-2}\}_i)^2 \right\} \rp.
\end{eqnarray}
The perturbed background solution is propagated over another year to obtain the background solution at time $t_{-1}$:
\begin{eqnarray}
\label{eq:exp-background-state-1}
\x^\textnormal{b}_{-1} = \M_{t_{-2} \rightarrow t_{-1}} \lp \widehat{\x}^\textnormal{b}_{-2}\rp.
\end{eqnarray}
This model propagation attenuates the random noise introduced in \eqref{eq:exp-perturbed-background} and makes the background state \eqref{eq:exp-background-state-1} consistent with the physics of the SPEEDY model. Then, the background state \eqref{eq:exp-background-state-1} is utilized in order to build an ensemble of perturbed background states:
\begin{eqnarray}
\label{eq:exp-perturbed-ensemble}
\displaystyle
\widehat{\x}^{\textnormal{b}[i]}_{-1} = \x^\textnormal{b}_{-1} + \errobs^\textnormal{b}_{-1},\quad \errobs^\textnormal{b}_{-1} \sim \Nor \lp {\bf 0}_{\Nstate} ,\, \underset{i}{\textnormal{diag}} \left\{ (0.05\, \{\x^\textnormal{b}_{-1}\}_i)^2 \right\} \rp,
\quad 1 \le i \le \Nens,
\end{eqnarray}
from which, after three months of model propagation, the initial ensemble is obtained at time $t_0$:
\begin{eqnarray*}
\x^{\textnormal{b}[i]}_0 = \M_{t_{-1} \rightarrow t_0} \lp \widehat{\x}^{\textnormal{b}[i]}_{-1}\rp \,.
\end{eqnarray*}
Again, the model propagation of the perturbed ensemble ensures that the ensemble members are consistent with the physics of the numerical model.
The experiments are performed over a period of 24 days, where observations are taken every 2 days ($\N=12$). At time $k$ synthetic observations are built as follows:
\begin{eqnarray*}
\y_k = \H_k \cdot \x^\textnormal{ref}_k + \errobs_k, \quad \errobs_k \sim \Nor \lp {\bf 0}_{\Nobs},\, \R_k \rp,\,
\quad \R_k = \textnormal{diag}_{i}\left\{ (0.01\, \{\H_k \, \x^\textnormal{ref}_k\}_i )^2 \right\}.
\end{eqnarray*}
The observation operators $\H_k$ are fixed throughout the time interval. We perform experiments with several operators characterized by different proportions $p$ of observed components from the model state $\x^\textnormal{ref}_k$ ($\Nobs \approx p \cdot \Nstate$). We consider four different values for $p$: 0.50, 0.12, 0.06 and 0.04 which represent 50\%, 12 \%, 6 \% and 4 \% of the total number of model components, respectively. Some of the observational networks used during the experiments are shown in Figure \ref{fig:exp-observational-grids} with their corresponding percentage of observed components from the model state.
The analyses of the EnKF-MC are compared against those obtained making use of the LETKF implementation proposed by Hunt et al in \cite{LETKFHunt,TELA:TELA076,application_letkf_1} . The analysis accuracy is measured by the root mean square error (RMSE)
\begin{eqnarray}
\label{eq:ER-RMSE-formula}
\displaystyle
\text{RMSE} = \sqrt{\frac{1}{\N} \cdot \sum_{k=1}^\N \lb \x^\textnormal{ref}_k -\x^\textnormal{a}_k \rb^T \cdot \lb \x^\textnormal{ref}_k -\x^\textnormal{a}_k \rb }
\end{eqnarray}
where $\x^\textnormal{ref} \in \Re^{\Nstate \times 1}$ and $\x^\textnormal{a}_{k} \in \Re^{\Nstate \times 1}$ are the reference and the analysis solutions at time $k$, respectively, and $\N$ is the number of assimilation times.
The threshold used in \eqref{eq:EnKF-MC-truncated-SVD} during the computation of $\widehat{\bf B}^{-1}$ is $\sigma_{r} = 0.10$. During the assimilation steps, the data error covariance matrices $\R_k$ are used (no representativeness errors are involved during the assimilations) and therefore. The different EnKF implementations are performed making use of FORTRAN and specialized libraries such as BLAS and LAPACK are used in order to perform the algebraic computations.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.5\textwidth]{figures/h_3-eps-converted-to.pdf}
\caption{$p=12\%$ }
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.5\textwidth]{figures/h_5-eps-converted-to.pdf}
\caption{$p=4\%$ }
\end{subfigure}%
\caption{Observational networks for different values of $p$. Dark dots denote the location of the observed components. The observed model variables are the zonal and the meridional wind components, the specific humidity, and the temperature.}
\label{fig:exp-observational-grids}
\end{figure}
\subsection{Results with dense observation networks}
We first consider dense observational networks in which 100\% and 50\% of the model components are observed. We vary the radius of influence ${\zeta}$ from 1 to 5 grid points.
Figure \ref{fig:exp-LETKF-RMSE-different-radii-dense-network} shows the RMSE values for the LETKF and EnKF-MC analyses for different values of ${\zeta}$ for the specific humidity when $50\%$ of model components are observed. When the radius of influence is increased the quality of the LETKF results degrades due to spurious correlations. This is expected since the local estimation of correlations in the context of LETKF is the sample covariance matrix. For instance, for a radius of influence of 1, the total number of local components for each local box is 36 which matches the dimension of the local background error distribution. Now, when we compare it against the ensemble size (96 ensemble members), sufficient degrees of freedom (95 degrees of freedom) are available in order to estimate the local background error distribution onto the ensemble space, and consequently all directions of the local probability error distribution are accounted during the estimation and posterior assimilation. On the other hand, when the radius of influence is 5, the local box sizes have dimension 484 (model components) which is approximately 5 times larger than the ensemble size. Thus, when the analysis increments are computed onto the ensemble space, just part of the local background error distribution is accounted during the assimilation. Consequently, the larger the local box, the more local background error information cannot be represented in the ensemble space.
Figure \ref{fig:exp-LETKF-RMSE-different-radii-dense-network} shows that EnKF-MC analyses improve with increasing radius of influence ${\zeta}$. Since a dense observational network is considered during the assimilation, when the radius of influence is increased, a better estimation of the state of the system is obtained by the EnKF-MC. This can be seen clearly in Figure \ref{fig:exp-LETKF-RMSE-different-radii-dense}, where the RMSE values within the assimilation window are shown for the LETKF and the EnKF-MC solutions for the specific humidity variable and different values of ${\zeta}$ and $p$. The quality of the EnKF-MC analysis for ${\zeta}=5$ is better than that of the LETKF with ${\zeta}=1$. Likewise, when a full observational network is considered ($p=100\%$), the proposed implementation outperforms the LETKF implementation. EnKF-MC is able to exploit the large amount of information contained in dense observational networks by properly estimating the local background error correlations. The RMSE values for all model variables and different values for ${\zeta}$ and $p$ are summarized in Table \ref{tab:exp-RMSE-values-all-dense}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_1_2_SPH_SC.png}
\caption{${\zeta} = 1$.}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_2_2_SPH_SC.png}
\caption{${\zeta} = 2$.}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_3_2_SPH_SC.png}
\caption{${\zeta} = 3$.}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_4_2_SPH_SC.png}
\caption{${\zeta} = 4$.}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_5_2_SPH_SC.png}
\caption{${\zeta} = 5$.}
\end{subfigure}
\caption{RMSE of specific humidity analyses with a dense observational network. When the radius of influence ${\zeta}$ is increased the performance of LETKF degrades.}
\label{fig:exp-LETKF-RMSE-different-radii-dense-network}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_1_SPH_RMSE_MC-eps-converted-to.pdf}
\caption{$p = 100\%$.}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_2_SPH_RMSE_MC-eps-converted-to.pdf}
\caption{$p = 50\%$.}
\end{subfigure}%
\caption{Analysis RMSE for the specific humidity variable. The RMSE values of the assimilation window are shown for different values of ${\zeta}$ and percentage of observed components $p$. When the local domain sizes are increased the accuracy of the LETKF analysis degrades, while the accuracy of EnKF-MC analysis improves.}
\label{fig:exp-LETKF-RMSE-different-radii-dense}
\end{figure}
\begin{table}[H]
\centering
{\footnotesize
\begin{tabular}{|c|c|c|c|c|} \hline
Variable (units) &${\zeta}$ & $p$ & EnKF-MC & LETKF \\ \hline
\multirow{10}{*}{Zonal Wind Component ($u$), ($m/s$)} & \multirow{2}{*}{1} & $100 \%$ & $ 6.012 \times 10^{1}$ & $ 6.394 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 4.264 \times 10^{2}$ & $ 9.825 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{2} & $100 \%$ & $ 6.078 \times 10^{1}$ & $ 6.820 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 2.255 \times 10^{2}$ & $ 1.330 \times 10^{3}$\\ \cline{2-5} & \multirow{2}{*}{3} & $100 \%$ & $ 6.080 \times 10^{1}$ & $ 7.969 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 2.341 \times 10^{2}$ & $ 1.124 \times 10^{3}$\\ \cline{2-5} & \multirow{2}{*}{4} & $100 \%$ & $ 6.088 \times 10^{1}$ & $ 9.687 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 2.418 \times 10^{2}$ & $ 1.072 \times 10^{3}$\\ \cline{2-5} & \multirow{2}{*}{5} & $100 \%$ & $ 6.092 \times 10^{1}$ & $ 1.190 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 2.673 \times 10^{2}$ & $ 1.017 \times 10^{3}$\\ \cline{1-5} \multirow{10}{*}{Meridional Wind Component ($v$) ($m/s$)} & \multirow{2}{*}{1} & $100 \%$ & $ 3.031 \times 10^{1}$ & $ 6.418 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 2.632 \times 10^{2}$ & $ 3.247 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{2} & $100 \%$ & $ 3.046 \times 10^{1}$ & $ 6.597 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 1.641 \times 10^{2}$ & $ 4.138 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{3} & $100 \%$ & $ 3.047 \times 10^{1}$ & $ 7.565 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 1.964 \times 10^{2}$ & $ 4.418 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{4} & $100 \%$ & $ 3.052 \times 10^{1}$ & $ 9.332 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 2.084 \times 10^{2}$ & $ 4.832 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{5} & $100 \%$ & $ 3.054 \times 10^{1}$ & $ 1.151 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 2.428 \times 10^{2}$ & $ 5.029 \times 10^{2}$\\ \cline{1-5} \multirow{10}{*}{Temperature ($K$)} & \multirow{2}{*}{1} & $100 \%$ & $ 9.404 \times 10^{2}$ & $ 5.078 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 6.644 \times 10^{2}$ & $ 7.059 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{2} & $100 \%$ & $ 9.416 \times 10^{2}$ & $ 4.112 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 6.129 \times 10^{2}$ & $ 1.138 \times 10^{3}$\\ \cline{2-5} & \multirow{2}{*}{3} & $100 \%$ & $ 9.425 \times 10^{2}$ & $ 3.447 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 5.815 \times 10^{2}$ & $ 1.389 \times 10^{3}$\\ \cline{2-5} & \multirow{2}{*}{4} & $100 \%$ & $ 9.432 \times 10^{2}$ & $ 2.939 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 5.585 \times 10^{2}$ & $ 1.355 \times 10^{3}$\\ \cline{2-5} & \multirow{2}{*}{5} & $100 \%$ & $ 9.432 \times 10^{2}$ & $ 2.554 \times 10^{2}$\\ \cline{3-5} & & $50 \%$ & $ 5.500 \times 10^{2}$ & $ 1.104 \times 10^{3}$\\ \cline{1-5} \multirow{10}{*}{Specific Humidity ($g/Kg$)} & \multirow{2}{*}{1} & $100 \%$ & $ 1.733 \times 10^{1}$ & $ 5.427 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 8.680 \times 10^{1}$ & $ 7.602 \times 10^{1}$\\ \cline{2-5} & \multirow{2}{*}{2} & $100 \%$ & $ 1.712 \times 10^{1}$ & $ 5.669 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 8.204 \times 10^{1}$ & $ 1.045 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{3} & $100 \%$ & $ 1.705 \times 10^{1}$ & $ 6.630 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 8.089 \times 10^{1}$ & $ 1.298 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{4} & $100 \%$ & $ 1.699 \times 10^{1}$ & $ 7.344 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 7.525 \times 10^{1}$ & $ 1.431 \times 10^{2}$\\ \cline{2-5} & \multirow{2}{*}{5} & $100 \%$ & $ 1.694 \times 10^{1}$ & $ 7.617 \times 10^{1}$\\ \cline{3-5} & & $50 \%$ & $ 7.642 \times 10^{1}$ & $ 1.458 \times 10^{2}$\\ \cline{1-5}
\end{tabular}
}
\caption{RMSE values for the EnKF-MC and the LETKF analyses with the SPEEDY model and for different values for ${\zeta}$ and $p$. Dense observational networks are considered in this experimental setting.}
\label{tab:exp-RMSE-values-all-dense}
\end{table}
\subsection{Results with sparse observation networks}
For sparse observational networks, in general, the results obtained by the EnKF-MC are more accurate than those obtained by the LETKF, as reported in the Tables \ref{tab:exp-RMSE-values-wind-components-sparse} and \ref{tab:exp-RMSE-values-others-sparse}. We vary the values of ${\zeta}$ from 1 to 5. Three sparse observational networks with $p=12\%$, $6\%$, and $4\%$, respectively are considered.
Figure \ref{fig:exp-LETKF-RMSE-different-radii-sparse-network} shows the RMSE values of the specific humidity analyses for different radii of influence and $4\%$ of the model components being observed. The best performance of the LETKF analyses is obtained when the radius of influence is set to 2. Note that for ${\zeta}=1$ the LETKF performs poorly, which is expected since during the assimilation most of model components will not have observations in their local boxes. For ${\zeta}\ge 3$ the effects of spurious correlations degrade the quality of the LETKF analysis. On the other hand, the background error correlations estimated by the modified Cholesky decomposition allows the EnKF-MC formulation to obtain good analyses even for largest radius of influence ${\zeta}=5$.
Figure \ref{fig:exp-LETKF-RMSE-different-radii-sparse} shows the RMSE values of the LETKF and the EnKF-MC implementations for different radii of influences and two sparse observational networks. Clearly, when the radius of influence is increased, in the LETKF context, the analysis corrections are impacted by spurious correlations. On the other hand, the quality of the results in the EnKF-MC case is considerably better. When data errors components are uncorrelated ${\zeta}$ can be seen as a free parameter and the choice can be based on the ``optimal performance of the filter''. For the largest radius of influence ${\zeta}=5$ the RMSE values of the ENKF-MC and the LETKF implementations differ by one order of magnitude.
Figure \ref{fig:exp-model-variables} reports the RMSE values for the zonal and the meridional wind component analyses, and for different values of $p$ and ${\zeta}$. As can be seen, the estimation of background errors via $\widehat{\bf B}$ can reduce the impact of spurious correlations; the RMSE values of the EnKF-MC analyses remain small at all assimilation times, from which we infer that the background error correlations are properly estimated. On the other hand, the impact of spurious correlations is evident in the context of LETKF. Since most of the model components are unobserved, the background error correlations drive the quality of the analysis, and spurious correlations lead to a poor performance of the filter at many assimilation times.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_1_5_SPH_NEW.png}
\caption{${\zeta} = 1$.}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_2_5_SPH_NEW.png}
\caption{${\zeta} = 2$.}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_3_5_SPH_NEW.png}
\caption{${\zeta} = 3$.}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_4_5_SPH_NEW.png}
\caption{${\zeta} = 4$.}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_5_5_SPH_NEW.png}
\caption{${\zeta} = 5$.}
\end{subfigure}
\caption{RMSE of specific humidity analyses with a sparse observational network ($p \sim 4\%$) and different values of ${\zeta}$. }
\label{fig:exp-LETKF-RMSE-different-radii-sparse-network}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_4_SPH_RMSE_MC-eps-converted-to.pdf}
\caption{$p = 6\%$.}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/out_5_SPH_RMSE_MC-eps-converted-to.pdf}
\caption{$p = 4\%$.}
\end{subfigure}%
\caption{Analysis RMSE for the specific humidity variable with sparse observation networks. RMSE values are shown for different values of ${\zeta}$ and percentage of observed components $p$.}
\label{fig:exp-LETKF-RMSE-different-radii-sparse}
\end{figure}
Figures \ref{fig:exp-snapshot-meridional-wind-component} and \ref{fig:exp-snapshot-zonal-wind-component} provide snapshots of the meridional and the zonal wind components, respectively, at the first assimilation time. For this particular case the percentage of observed model components is $p=4\%$. At this step, only the initial observation has been assimilated in order to compute the analysis corrections by the EnKF-MC and the LETKF methods. The background solution contains erroneous waves for the zonal and the meridional wind components. For instance, for the $u$ model variable, such waves are clearly present near the poles. After the first assimilation step, the LETKF analysis solution dissipates the erroneous waves but, the numerical values of the wind components are slightly greater than those of the reference solutions. This numerical difference increases at later times due to the highly-nonlinear dynamics of SPEEDY, as can bee seen in Figure \ref{fig:exp-model-variables}. On the other hand, the EnKF-MC implementation recovers the reference shape, and the analysis values of the numerical model components are close to that of the reference solution. This shows again that the use of the modified Cholesky decomposition as the estimator of the background error correlations can mitigate the impact of spurious error correlations.
\begin{table}[H]
\centering
{\footnotesize
\begin{tabular}{|c|c|c|c|c|} \hline
Variable (units) &${\zeta}$ & $p$ & EnKF-MC & LETKF \\ \hline
\multirow{15}{*}{Zonal Wind Component ($u$), ($m/s$)} & \multirow{3}{*}{1} & $12\%$ & $ 5.514 \times 10^{2}$ & $ 5.471 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 6.972 \times 10^{2}$ & $ 1.168 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 9.393 \times 10^{2}$ & $ 1.737 \times 10^{3}$\\ \cline{2-5} & \multirow{3}{*}{2} & $12\%$ & $ 4.187 \times 10^{2}$ & $ 1.275 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 6.090 \times 10^{2}$ & $ 7.591 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 7.853 \times 10^{2}$ & $ 8.569 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{3} & $12\%$ & $ 4.388 \times 10^{2}$ & $ 1.661 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 6.146 \times 10^{2}$ & $ 1.237 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 7.438 \times 10^{2}$ & $ 9.997 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{4} & $12\%$ & $ 4.323 \times 10^{2}$ & $ 1.752 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 5.990 \times 10^{2}$ & $ 1.608 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 7.124 \times 10^{2}$ & $ 1.258 \times 10^{3}$\\ \cline{2-5} & \multirow{3}{*}{5} & $12\%$ & $ 4.456 \times 10^{2}$ & $ 1.862 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 6.106 \times 10^{2}$ & $ 1.983 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 7.160 \times 10^{2}$ & $ 1.602 \times 10^{3}$\\ \cline{1-5} \multirow{15}{*}{Meridional Wind Component ($v$) ($m/s$)} & \multirow{3}{*}{1} & $12\%$ & $ 3.540 \times 10^{2}$ & $ 4.496 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 5.165 \times 10^{2}$ & $ 1.158 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 7.770 \times 10^{2}$ & $ 1.749 \times 10^{3}$\\ \cline{2-5} & \multirow{3}{*}{2} & $12\%$ & $ 3.009 \times 10^{2}$ & $ 7.285 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 4.605 \times 10^{2}$ & $ 5.520 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 6.217 \times 10^{2}$ & $ 7.420 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{3} & $12\%$ & $ 3.172 \times 10^{2}$ & $ 9.510 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 4.735 \times 10^{2}$ & $ 8.334 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 6.014 \times 10^{2}$ & $ 7.455 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{4} & $12\%$ & $ 3.399 \times 10^{2}$ & $ 1.048 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 4.812 \times 10^{2}$ & $ 1.146 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 5.913 \times 10^{2}$ & $ 9.026 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{5} & $12\%$ & $ 3.626 \times 10^{2}$ & $ 1.101 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 5.107 \times 10^{2}$ & $ 1.575 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 6.122 \times 10^{2}$ & $ 1.102 \times 10^{3}$\\ \cline{1-5}
\end{tabular}
}
\caption{RMSE values of the wind-components for the EnKF-MC and LETKF making use of the SPEEDY model.}
\label{tab:exp-RMSE-values-wind-components-sparse}
\end{table}
\begin{table}[H]
\centering
{\footnotesize
\begin{tabular}{|c|c|c|c|c|} \hline
Variable (units) &${\zeta}$ & $p$ & EnKF-MC & LETKF \\ \hline
\multirow{15}{*}{Temperature ($K$)} & \multirow{3}{*}{1} & $12\%$ & $ 6.054 \times 10^{2}$ & $ 6.033 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 5.692 \times 10^{2}$ & $ 6.704 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 6.522 \times 10^{2}$ & $ 8.073 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{2} & $12\%$ & $ 5.680 \times 10^{2}$ & $ 6.693 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 5.193 \times 10^{2}$ & $ 5.556 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 5.299 \times 10^{2}$ & $ 5.529 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{3} & $12\%$ & $ 5.279 \times 10^{2}$ & $ 1.217 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 4.982 \times 10^{2}$ & $ 6.458 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 4.926 \times 10^{2}$ & $ 6.073 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{4} & $12\%$ & $ 5.023 \times 10^{2}$ & $ 1.817 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 4.757 \times 10^{2}$ & $ 1.030 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 4.766 \times 10^{2}$ & $ 7.464 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{5} & $12\%$ & $ 4.898 \times 10^{2}$ & $ 1.600 \times 10^{3}$\\ \cline{3-5} & & $ 6 \%$ & $ 4.644 \times 10^{2}$ & $ 1.473 \times 10^{3}$\\ \cline{3-5} & & $4\%$ & $ 4.684 \times 10^{2}$ & $ 1.172 \times 10^{3}$\\ \cline{1-5} \multirow{15}{*}{Specific Humidity ($g/Kg$)} & \multirow{3}{*}{1} & $12\%$ & $ 9.862 \times 10^{1}$ & $ 9.026 \times 10^{1}$\\ \cline{3-5} & & $ 6 \%$ & $ 1.133 \times 10^{2}$ & $ 1.449 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 1.405 \times 10^{2}$ & $ 1.941 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{2} & $12\%$ & $ 1.029 \times 10^{2}$ & $ 1.125 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 1.146 \times 10^{2}$ & $ 1.137 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 1.270 \times 10^{2}$ & $ 1.321 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{3} & $12\%$ & $ 1.068 \times 10^{2}$ & $ 1.341 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 1.205 \times 10^{2}$ & $ 1.418 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 1.317 \times 10^{2}$ & $ 1.458 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{4} & $12\%$ & $ 1.065 \times 10^{2}$ & $ 1.640 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 1.246 \times 10^{2}$ & $ 1.652 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 1.324 \times 10^{2}$ & $ 1.739 \times 10^{2}$\\ \cline{2-5} & \multirow{3}{*}{5} & $12\%$ & $ 1.089 \times 10^{2}$ & $ 2.078 \times 10^{2}$\\ \cline{3-5} & & $ 6 \%$ & $ 1.301 \times 10^{2}$ & $ 1.950 \times 10^{2}$\\ \cline{3-5} & & $4\%$ & $ 1.373 \times 10^{2}$ & $ 2.068 \times 10^{2}$\\ \cline{1-5}
\end{tabular}
}
\caption{RMSE values for the EnKF-MC and LETKF making use of the SPEEDY model.}
\label{tab:exp-RMSE-values-others-sparse}
\end{table}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/out_3_3_UWC-eps-converted-to.pdf}
\caption{${\zeta} = 3$ and $p=12\%$ }
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/out_4_3_VWC-eps-converted-to.pdf}
\caption{${\zeta} = 4$ and $p=12\%$ }
\end{subfigure}%
\caption{RMSE of the LETKF and EnKF-MC implementations for different model variables, radii of influence and observational networks. }
\label{fig:exp-model-variables}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/true_1_57_5_5_VWC.png}
\caption{Reference}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/back_1_57_5_5_VWC.png}
\caption{Background}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/enkfB_1_57_5_5_VWC.png}
\caption{EnKF-MC}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/LETKF_1_57_5_5_VWC.png}
\caption{LETKF}
\end{subfigure}
\caption{Snapshots of the reference solution, background state, and analysis fields from the EnKF-MC and LETKF for the fifth layer of the meridional wind component ($v$).}
\label{fig:exp-snapshot-meridional-wind-component}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/true_1_42_5_5_UWC.png}
\caption{Reference}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/back_1_42_5_5_UWC.png}
\caption{Background}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/enkfB_1_42_5_5_UWC.png}
\caption{EnKF-MC}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.8\textwidth]{figures/LETKF_1_42_5_5_UWC.png}
\caption{LETKF}
\end{subfigure}
\caption{Snapshots of the reference solution, background state, and analysis fields from the EnKF-MC and LETKF for the second layer of the zonal wind component ($u$).}
\label{fig:exp-snapshot-zonal-wind-component}
\end{figure}
\subsection{Statistics of the ensemble}
In this section, we briefly discuss the spread of the ensemble making use of rank histograms. Of course, we do not claim this to be a verification procedure but, it provides useful insights about the dispersion of the members and the level of uncertainty about the ensemble mean. The plots are based on the 5-th numerical layer of the atmosphere. We collect information across all model variables and the plots are shown in figures \ref{fig:bin-sph}, \ref{fig:bin-tem}, \ref{fig:bin-uwc}, and \ref{fig:bin-vwc}. Based on the results, the proposed implementation seems to be lesser sensitive to the intrinsic need of inflation than the LETKF formulation. For instance, after the assimilation, the ensemble members from the EnKF-MC are spread almost uniformly across different observation times. On the other hand, the spread in the context of the LETKF is impacted by the constant inflation factor used during the experiments (1.04) In practice, the inflation factor is set up according to historical information and/or heuristically with regard to some properties of the dynamics of the numerical model. This implies that, the dispersion of the LETKF members after the analysis will rely in how-well we estimate the optimal inflation factor for such filter. In operational data assimilation, an answer to this question can be hard to find. We think that inflation methodologies such as adaptive inflation can lead to better spread of the ensemble members in the context of the LETKF. For the proposed method, based on the experimental results, such methodology is not needed.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_EnKFMC_SPH.png}
\caption{EnKF-MC}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_LETKF_SPH.png}
\caption{LETKF}
\end{subfigure}%
\caption{Rank-histograms for the Specific Humidity model variable. The information is collected from the 5-th model layer.}
\label{fig:bin-sph}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_EnKFMC_UWC.png}
\caption{EnKF-MC}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_LETKF_UWC.png}
\caption{LETKF}
\end{subfigure}%
\caption{Rank-histograms for the Zonal Wind Component model variable. The information is collected from the 5-th model layer.}
\label{fig:bin-uwc}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_EnKFMC_VWC.png}
\caption{EnKF-MC}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_LETKF_VWC.png}
\caption{LETKF}
\end{subfigure}%
\caption{Rank-histograms for the Meridional Wind Component model variable. The information is collected from the 5-th model layer.}
\label{fig:bin-vwc}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_EnKFMC_TEM.png}
\caption{EnKF-MC}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{Corrections/BIN_LETKF_TEM.png}
\caption{LETKF}
\end{subfigure}%
\caption{Rank-histograms for the Temperature model variable. The information is collected from the 5-th model layer.}
\label{fig:bin-tem}
\end{figure}
\subsection{The impact of SVD truncation threshold}
\label{sec:future-work}
An important question arising from this research is the number of singular values/vectors to be used in \eqref{eq:EnKF-MC-truncated-SVD}. To study this question we use the same experimental setting and the sparse observational network where only $4\%$ of the model components are observed. We apply EnKF-MC algorithm and truncate the summation \eqref{eq:EnKF-MC-truncated-SVD} based on different thresholds $\sigma_r$.
The results are reported in Figure \ref{fig:exp-RMSE-different-thresholds}. Different thresholds lead to different levels of accuracy for the EnKF-MC analyses. There is no unique value of $\sigma_r$ that provides the best ensemble trajectory in general; for instance, the best performance at the beginning of the assimilation window is obtained for $\sigma_r = 0.05$, but, at the end the best solution is obtained with $\sigma_r = 0.2$. This indicates that the results can be improved when $\sigma_r$ is dynamically and optimally chosen. Note that, on average, the results obtained by the EnKF-MC with $\sigma_r \in \lle 0.15,\, 0.20,\, 0.25 \rle$ are much better than those when $\sigma_r = 0.10$ (and therefore much better than the results obtained by the LETKF). In Figure \ref{fig:exp-RMSE-snapshots-thresholds} snapshots of the specific humidity for different $\sigma_r$ are shown. It can be seen that the spurious errors can be quickly decreased when $\sigma_r$ is chosen accordingly.
In order to understand the optimal truncation level note that the summation \eqref{eq:EnKF-MC-truncated-SVD} can be written as follows:
\begin{eqnarray}
\label{eq:alpha_j}
{\boldsymbol \beta}_{[i]} &=& \sum_{j=1}^{\Nens} \alpha_j \cdot {\bf u}^{\Z_{[i]}}_j,\, \\
\nonumber
\alpha_j &=& \frac{1}{\tau_j} \cdot {{\bf v}^{\Z_{[i]}}_j}^T \cdot \x_{[i]} = \frac{1}{\tau_j} \cdot {{\bf v}^{\Z_{[i]}}_j}^T \cdot \lb \widetilde{\x}_{[i]} + \boldsymbol{\theta}_{[i]} \rb \\
\nonumber
&=& \underbrace{ \frac{1}{\tau_j} \cdot {{\bf v}^{\Z_{[i]}}_j}^T \cdot \widetilde{\x}_{[i]} }_{\text{Uncorrupted data}} + \underbrace{\frac{1}{\tau_j} \cdot {{\bf v}^{\Z_{[i]}}_j}^T \cdot \boldsymbol{\theta}_{[i]}}_{\text{Error}}
\end{eqnarray}
where $\widetilde{\x}_{[i]}$ is the perfect data ($\x_{[i]} = \widetilde{\x}_{[i]}+\boldsymbol{\theta}_{[i]}$). The components with small singular values $\tau_j$ will amplify the error more. The threshold should be large enough to include useful information from $\widetilde{\x}_{[i]}$, but small enough in order to prune out the components with large error amplification.
We expect that model components with large variances will need more basis vectors from \eqref{eq:EnKF-MC-truncated-SVD} than those with lesser variance. An upper bound for the number of basis vectors (and therefore the threshold $\sigma_r$) can be obtained by inspection of the values $\alpha_j$ in \eqref{eq:alpha_j}. Figure \ref{fig:exp-random-noise-effect} shows the weights $\alpha_j$ for different singular values for the 500-th model component of the SPEEDY model. The large zig-zag behaviors are evidence of error amplifications and therefore, we can truncate the summation \eqref{eq:alpha_j} before this pattern starts to take place in the values of $\alpha_j$.
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/VWC-eps-converted-to.pdf}
\caption{Meridional wind component ($m/s$)}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/UWC-eps-converted-to.pdf}
\caption{Zonal wind component ($m/s$)}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/TEM-eps-converted-to.pdf}
\caption{Temperature ($K$)}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.75\textwidth]{figures/SPH-eps-converted-to.pdf}
\caption{Specific humidity ($g/kg$)}
\end{subfigure}
\caption{RMSE for the SPEEDY analyses obtained using different SVD truncation levels based on the $\sigma_r$ values.}
\label{fig:exp-RMSE-different-thresholds}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/true_SPH_8.png}
\caption{Reference}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/{xa_SPH_8_0.05}.png}
\caption{$\sigma_r = 0.05$}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/{xa_SPH_8_0.1}.png}
\caption{$\sigma_r = 0.10$}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/{xa_SPH_8_0.15}.png}
\caption{$\sigma_r = 0.15$}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/{xa_SPH_8_0.2}.png}
\caption{$\sigma_r = 0.20$}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/{xa_SPH_8_0.3}.png}
\caption{$\sigma_r = 0.30$}
\end{subfigure}
\caption{Snapshots at the final assimilation time (day 22) of the EnKF-MC analysis making use of different thresholds $\sigma_r$ for ${\zeta} = 5$ and $p = 4 \%$. }
\label{fig:exp-RMSE-snapshots-thresholds}
\end{figure}
\begin{figure}[H]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/MWC_SV-eps-converted-to.pdf}
\caption{Meridional wind component ($m/s$)}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/ZWC_SV-eps-converted-to.pdf}
\caption{Zonal wind component ($m/s$)}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/TEM_SV-eps-converted-to.pdf}
\caption{Temperature ($K$)}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth,height=0.9\textwidth]{figures/SH_SV-eps-converted-to.pdf}
\caption{Specific humidity ($g/kg$)}
\end{subfigure}
\caption{The effect of $\boldsymbol{\theta}$ on the weights $\alpha_j$ for some model component $i$ of the SPEEDY model when ${\zeta}=5$ and $p=4\%$.}
\label{fig:exp-random-noise-effect}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
This paper develops an efficient implementation of the ensemble Kalman filter, named EnKF-MC, that is based on a modified Cholesky decomposition to estimate the inverse background covariance matrix. This new approach has several advantages over classical formulations. First, a predefined sparsity structure can be built into the factors of the inverse covariance. This reflects the fact that if two distant model components are uncorrelated then the corresponding entry in the inverse covariance matrix is zero; the only nonzero entries in the Cholesky factors correspond to components of the model that are located in each other's proximity. Therefore, imposing a sparsity structure on the inverse background covariance matrix is a form of covariance localization. Second, the formulation allows for a rigorous theoretical analysis; we prove the convergence of the covariance estimator for a number of ensemble members that is proportional to the logarithm of the number of states of the model therefore, when $\Nens \approx \log \Nstate$, the background error correlations can be well-estimated making use of the modified Cholesky decomposition.
We discuss different implementations of the new EnKF-MC, and asses their computational effort. We show that domain decomposition can be used in order to decrease even more the computational effort of the proposed implementation. Numerical experiments are carried out using the Atmospheric General Circulation Model SPEEDY reveal that the analyses obtained by EnKF-MC are better than those of the LETKF in the root mean square sense when sparse observations are used in the analysis. For dense observation grids the EnKF-MC solutions are improved when the radius of influence increases, while the opposite holds true for LETKF analyses. (We stress the fact that these conclusions are true for our implementation of the basic LETKF; other implementations may incorporate advances that could make the filter perform considerably better). The use of modified Cholesky decomposition can mitigate the impact of spurious correlation during the assimilation of observations.
\section*{Acknowledgements}
This work was supported in part by awards NSF CCF--1218454,
AFOSR FA9550--12--1--0293--DEF, and by the Computational Science Laboratory at Virginia Tech.
\input{Main.bbl}
\end{document}
|
\section{Introduction}
Let $V(f)$ be a nonsingular algebraic surface defined by a homogeneous polynomial $f(x,y,z,t)$ over an algebraically closed field $k$
of degree $d\geq 3$. The characteristic of the field is assumed to be zero.
The projective automorphism group $\Paut(V(f))$ coincides with the automorphism group $\Aut(V(f))$ unless $d=4$ \cite{mat}. Moreover, there exists a constant
$B_d>0$ such that $|\Paut(V(f))|\leq B_d$ for any nonsingular homogeneous polynomial $f$ of degree $d$.
Let $G$ be a group. If $\Paut(V(f))$ contains a subgroup isomorphic to $G$, $V(f)$ is said to be $G$-invariant.
When $G$ is a subgroup of $PGL_4(k)$, then $V(f)$ is said to be $G$-invariant if $\Paut(V(f))$ contains a subgroup conjugate to
$G$. The most symmetric nonsingular cubic surface is projectively equivalent to $V(x^3+y^3+z^3+t^3)$. Meanwhile the secondly most symmetric
nonsingular cubic surface $V(f)$ can be characterized in two ways: either $f$ is projectively equivalent to $x^2y+y^2z+z^2t+t^2x$ or $V(f)$ is
$\SSS{5}$-invariant \cite{dol}. Indeed, $\Aut(V(x^2y+y^2z+z^2t+t^2x))$ is isomorphic to $\SSS{5}$. Burnside conjectured that the most symmetric nonsingular
quartic surface is projectively equivalent to $V(h)$, where $h=x^4+y^4+z^4+t^4+12xyzt$ so that $|\Paut(V(h))|=1920$ \cite[\S272]{bur}.
In this paper $V(h)$ is shown to be $\SSS{5}$-invariant, and all $\SSS{5}$-invariant nonsingular quartic surfaces are given, up to projective equivalence. |
\section{Introduction}
Throughout this paper $G$ is a finite group, $d(n)$ denotes the
number of divisors of positive integer $n$, $Irr(G)$ stands for
the set of all ordinary irreducible character of $G$ and $Con(G)$
is the set of all conjugacy classes of $G$. For other notations
and terminology concerning character theory, we refer to the
famous book of Isaacs \cite{11}. Following Diaconis and Isaacs
\cite{7}, a pair ($\mathcal{X}$, $\mathcal{K}$) together with the
choices of characters $\chi_X$ is called a \textbf{supercharacter
theory} of $G$ if the following conditions are satisfied:
\begin{enumerate}
\item $\mathcal{X}$ is a partition of $Irr(G)$ and $\mathcal{K}$ is a partition of $Con(G)$;
\item $\{ 1\} \in \mathcal{K}$;
\item the characters $\chi_X$, $X \in \mathcal{X}$, are constant on the members of $\mathcal{K}$;
\item $|\mathcal{X}|$ = $|\mathcal{K}|$.
\end{enumerate}
The elements of $\mathcal{X}$ and $\mathcal{K}$ are called
\textbf{supercharacters} and \textbf{superclasses} of $G$,
respectively. It is easy to see that $m(G) = (Irr(G),Con(G))$ and
$M(G) = (\mathcal{X}, \mathcal{K})$, where $\mathcal{X}$ = $\{ \{
1\}, Irr(G) \setminus \{ 1\}\}$ and $\mathcal{K}$ = $\{
\{1\},Con(G) \setminus\{ 1\}\}$ are supercharacter theories of
$G$ which are called the trivial supercharacter theories of $G$.
The set of all supercharacter theories of a finite group $G$ is
denoted by $Sup(G)$ and set $s(G) = |Sup(G)|$. These notions were
first introduced by Andre for finite unitriangular groups using
polynomial equations defining certain algebraic varieties
\cite{0,1,2}.
Following Hendrickson \cite{9}, we assume that $Part(S)$ denotes
the set of all partitions of a set $S$. If $\mathcal{X}$ and
$\mathcal{Y}$ are two elements of $Part(S)$ then we say that
$\mathcal{X}$ is a \textbf{refinement} of $\mathcal{Y}$ or
$\mathcal{Y}$ is \textbf{coarser than} $\mathcal{X}$ and we
write ``$\mathcal{X}$ $\preceq$ $\mathcal{Y}$", if
$[a]_{\mathcal{X}} \subseteq [a]_{\mathcal{Y}}$ for all $a \in
S$. For two supercharacter theories $(\mathcal{X},\mathcal{K})$
and $(\mathcal{Y},\mathcal{L})$, we define
$(\mathcal{X},\mathcal{K}) \vee (\mathcal{Y},\mathcal{L})$ =
$(\mathcal{X} \vee \mathcal{Y}, \mathcal{K} \vee \mathcal{L})$.
It is well-know that $(Part(S),\preceq)$ is a lattice which is
called the \textbf{partition lattice} of $S$. By
\cite[Proposition 2.16]{10}, the join of two supercharacter
theories of a group $G$ is again a supercharacter theory for $G$,
but it is possible to find a pair of supercharacter theories such
that their meet is not a supercharacter theory. This shows that
the set of all supercharacter theories of a finite group with
usual join and meet don't constitute a lattice in general.
Burkett et al. \cite{6} gave a classification of finite groups
with exactly two supercharacter theories. They proved that a
finite group $G$ has exactly two supercharacter theories if and
only if $G$ is isomorphic to the cyclic group $Z_3$, the
symmetric group $S_3$ or the simple group $Sp(6, 2)$. The aim of
this paper is to continue this work towards a classification of
finite simple groups with exactly three or four supercharacter
theories.
If $p$ and $2p + 1$ are primes then $p$ is called a
\textbf{Sophie Germain prime} and $2p + 1$ is said to be a
\textbf{safe prime}. The safe primes are recorded in on-line
encyclopedia of integer sequences as A005385, see \cite{13} for
details. The first few members of this sequence is $5,$ $7$,
$11$, $23$, $47$, $59$, $83$, $107$, $167$, $179$, $227$, $263$,
$347$, $359$, $383$, $467$, $479$, $503$, $563$, $587$, $719$,
$839$, $863$, $887$, $983$, $1019$, $1187$, $1283$, $1307$,
$1319$, $1367$, $1439$, $1487$, $1523$, $1619$, $1823$, $1907$.
These two sequences of prime numbers have several applications in
public key cryptography and primality testing and it has been
conjectured that there are infinitely many Sophie Germain primes,
but this remains unproven \cite{65}.
\section{Supercharacter Theory Construction for Sporadic Groups}
The aim of this section is to compute some supercharacter
theories for sporadic simple groups. We use the following simple
lemma in our calculations:
\begin{lem}\label{11}
Let $G$ be a finite group, $\chi$ be a non-real valued irreducible
character of $G$ and $x \in G$ such that $\chi(x)$ is a non-real
number. We also assume that each row and column of $G$ has at most
two non-real numbers. Define:
\begin{eqnarray*}
\mathcal{X} &=& \{\{\chi_{1}\}, \{\chi, \overline{\chi}\}, Irr(G) - \{\chi_{1}, \chi, \overline{\chi}\}\}\\
\mathcal{K} &=& \{\{e\}, \{x^{G}, (x^{-1})^{G}\}, Con(G) - \{e,
x^{G}, (x^{-1})^{G}\}\}.
\end{eqnarray*}
Then {$ (\mathcal{X}, \mathcal{K}) $} is a supercharacter theory
for $G$.
\end{lem}
\begin{proof}
To prove, it is enough to investigate the main condition of
supercharacter theory for the conjugacy classes $x^{G}$, $
(x^{-1})^{G} $ and irreducible characters $\chi$,
$\overline{\chi}$. By definition of ${\sigma}_{X}$ on $ X =
\{x^{G}, (x^{-1})^{G}\} $,
\begin{eqnarray*}
\sigma_{X}(x^{-1}) &=& \chi(1)\chi(x^{-1}) + \overline{\chi}(1)\overline{\chi}(x^{-1})\\
&=& \chi(1) \overline{\chi(x)} + \chi(1)\chi(x)\\
&=& \overline{\chi(1)}~\overline{\chi(x)} + \chi(1)\chi(x)\\
&=& \sigma_{X}(x).
\end{eqnarray*}
Therefore, {$ \sigma_{X} $} is constant on the part {$ \{x^{G},
(x^{-1})^{G}\} $} of $\mathcal{K}$, as desired.
\end{proof}
The following lemma is important for constructing supercharacter theories on simple groups.
\begin{lem}\label{12}
Suppose $G$ is a finite group, $A = \{ \chi(x) \ | \ \chi \in Irr(G)\}$ and $\mathbb{Q}(A)$ denotes the filed generated by $\mathbb{Q}$ and $A$. Then the following holds:
\begin{enumerate}
\item If $\mathcal{X}(G) = \{ \{ \chi, \overline{\chi} \} \ | \ \chi \in Irr(G) \}$ and $\mathcal{K}(G) = \{ \{ x^G, (x^{-1})^G\} \ | \ x \in G\}$
then $(\mathcal{X}, \mathcal{K})$ is a supercharacter theory of $G$.
\item If $\Gamma = Gal(\frac{\mathbb{Q(A)}}{\mathbb{Q}})$, $\mathcal{X}(G)$ is the set of all orbits of $\Gamma$ on $Irr(G)$ and $\mathcal{K}(G)$ is
the set of all orbits of $\Gamma$ on $Con(G)$ then $(\mathcal{X},
\mathcal{K})$ is a supercharacter theory of $G$.
\end{enumerate}
\end{lem}
\begin{proof}
The proof follows from \cite[p. 2360]{7} and \cite{12}.
\end{proof}
To calculate the supercharacter of a finite group $G$, we first
sort the character table of $G$ by the following GAP commands \cite{19}:
\begin{alltt}
u:=CharacterTable(G);
t:=CharacterTableWithSortedCharacters(u);
\end{alltt}
Then we prepare a GAP program to check whether or not a given pair
$(\mathcal{X},\mathcal{K})$ of a partition $\mathcal{K}$ for
conjugacy classes and another partition $\mathcal{X}$ for
irreducible characters constitutes a supercharacter theory. To
find this pair of partitions, we usually apply Lemma \ref{11}.
\begin{thm}
The Mathieu groups $M_{11}, M_{12}, M_{22}, M_{23}$ and $M_{24}$ have at least five supercharacter theories.
\end{thm}
\begin{proof}
We will present five supercharacter theories for each Mathieu
group as follows:
\begin{itemize}
\item \textit{The Mathieu group $M_{11}$}. Suppose the irreducible characters and conjugacy classes of the Mathieu group $M_{11}$ are
$Irr(M_{11})$ = $\{ \chi_1, \chi_2, \ldots, \chi_{10}\}$ and
$Con(M_{11})$ = $\{ x_1^{M_{11}}, x_2^{M_{11}}, \ldots,
x_{10}^{M_{11}}\}$, respectively. We now define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{M_{11}}\} (2 \leq i \leq 8), \{x_{9}^{M_{11}},x_{10}^{M_{11}}\} \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 5), \{\chi_{6}, \chi_{7}\},\{\chi_{i}\} (8 \leq i \leq
10)\right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{M_{11}}\} (2 \leq i \leq 6), \{x_{7}^{M_{11}},x_{8}^{M_{11}}\}, \{x_{i}^{M_{11}}\} (9 \leq i \leq 10) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{2}\}, \{\chi_{3},
\chi_{4}\},\{\chi_{i}\} (5 \leq i \leq 10)\right\rbrace.
\end{eqnarray*}
\end{small}
By Lemma \ref{11} and \cite[Proposition 2.16]{10},
$\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ are
supercharacter theories of $M_{11}$. Since $\mathcal{C}_{1}$,
$\mathcal{C}_{2}$, $\mathcal{C}_{3}$, $m(M_{11})$ and $M(M_{11})$
are distinct, $s(M_{11}) \geq 5$.
\item \textit{The Mathieu group $M_{12}$}. We assume that the irreducible characters and conjugacy classes of the Mathieu group $M_{12}$
are $Irr(M_{12})$ = $\{ \chi_1, \chi_2, \ldots, \chi_{15}\}$ and
$Con(M_{12})$ = $\{ x_1^{M_{12}}, \ldots, x_{15}^{M_{12}} \}$,
respectively. Define:
\begin{footnotesize}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{M_{12}}\} (2 \leq i \leq 13), \{x_{14}^{M_{12}},x_{15}^{M_{12}}\} \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 3), \{\chi_{4}, \chi_{5}\},\{\chi_{i}\} (6 \leq i \leq
15)\right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\},
\{x_{i}^{M_{12}}\} (2 \leq i \leq 5),
\{x_{6}^{M_{12}},x_{7}^{M_{12}}\}, \{x_{i}^{M_{12}}\} (8 \leq i
\leq 10),
\{x_{11}^{M_{12}},x_{12}^{M_{12}}\}, \{x_{i}^{M_{12}}\} (13 \leq i \leq 15) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3}\},
\{\chi_{i}\} (4 \leq i \leq 8), \{\chi_{9},
\chi_{10}\},\{\chi_{i}\} (11 \leq i \leq 15)\right\rbrace .
\end{eqnarray*}
\end{footnotesize}
Then by Lemma \ref{11} and \cite[Proposition 2.16]{10}, the pairs
$\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ are
three supercharacter theories of $M_{12}$ different from
$m(M_{12})$ and $M(M_{12})$. This proves that $s(M_{12}) \geq 5$,
as desired.
\item \textit{The Mathieu group $M_{22}$}. Suppose $Irr(M_{22})$ = $\{ \chi_1, \ldots,
\chi_{12}\}$ and $Con(M_{22})$ = $\{ x_1^{M_{22}}, \ldots,
x_{12}^{M_{22}}\}$. We define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{M_{22}}\} (2 \leq i \leq 10), \{x_{11}^{M_{22}},x_{12}^{M_{22}}\} \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 9), \{\chi_{10}, \chi_{11}\},\{\chi_{12}\} \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{M_{22}}\} (2 \leq i \leq 7), \{x_{8}^{M_{22}},x_{9}^{M_{22}}\}, \{x_{i}^{M_{22}}\} (10 \leq i \leq 12) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{2}\}, \{\chi_{3},
\chi_{4}\},\{\chi_{i}\} (5 \leq i \leq 12)\right\rbrace.
\end{eqnarray*}
\end{small}
Then by Lemma \ref{11} the pairs $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$ and $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ are supercharacter theories of
$M_{22}$. We now apply \cite[Proposition 2.16]{10} to prove that
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ is
another supercharacter theory for $M_{22}$ which shows that
$s(M_{22}) \geq 5$.
\item \textit{The Mathieu group $M_{23}$}. Suppose $Irr(M_{23})$ = $\{ \chi_1, \ldots, \chi_{17}\}$ and
$Con(M_{23})$ = $\{ x_1^{M_{23}}, \ldots, x_{17}^{M_{23}}\}$
are the irreducible characters and conjugacy classes of the
Mathieu group $M_{23}$, respectively. If we can present three
supercharacter theories for $M_{23}$ different from $m(M_{23})$
and $M(M_{23})$ then it can be easily proved that $s(M_{23}) \geq
5$, as desired. Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{M_{23}}\} (2 \leq i \leq 15), \{x_{16}^{M_{23}},x_{17}^{M_{23}}\} \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 9), \{\chi_{10}, \chi_{11}\}, \{\chi_{i}\} (12 \leq i \leq
17) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{M_{23}}\} (2 \leq i \leq 9), \{x_{10}^{M_{23}},x_{11}^{M_{23}}\}, \{x_{i}^{M_{23}}\} (12 \leq i \leq 17) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 11), \{\chi_{12}, \chi_{13}\}, \{\chi_{i}\} (14 \leq i \leq
17) \right\rbrace.
\end{eqnarray*}
\end{small}
To complete the proof, it is enough to apply Lemma \ref{11} and
\cite[Proposition 2.16]{10} for proving that $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are supercharacter theories
of $M_{23}$.
\item \textit{The Mathieu group $M_{24}$}. We now assume that $Irr(M_{24})$ = $\{ \chi_1, \chi_2, \ldots, \chi_{26}\}$ and
$Con(M_{24})$ = $\{ x_1^{M_{24}}, x_2^{M_{24}}, \ldots,
x_{26}^{M_{24}}\}$. If we define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{M_{24}}\} (2 \leq i \leq 20), \{x_{21}^{M_{24}},x_{22}^{M_{24}}\}, \{x_{i}^{M_{24}}\} (23 \leq i \leq 26) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 4), \{\chi_{5}, \chi_{6}\}, \{\chi_{i}\} (7 \leq i \leq 26)
\right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{M_{24}}\} (2 \leq i \leq 24), \{x_{25}^{M_{24}},x_{26}^{M_{24}}\} \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 9), \{\chi_{10}, \chi_{11}\}, \{\chi_{i}\} (12 \leq i \leq
26) \right\rbrace.
\end{eqnarray*}
\end{small}
then by a similar calculation as other cases, we can prove that
$\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are three supercharacter
theories different from $m(M_{24})$ and $M(M_{24})$, proving this
case.
\end{itemize}
This completes the proof.
\end{proof}
\begin{thm}
The Leech lattice groups have at least five supercharacter theories.
\end{thm}
\begin{proof}
There are seven Leech lattice simple groups. These are $HS$,
$J_2$, $Co_1, Co_2, Co_3$, $McL$ and $Suz$. Our main proof will
consider seven cases as follows:
\begin{itemize}
\item \textit{The Higman-Sims group HS}. To establish five supercharacter theories for $HS$, we assume that
$Irr(HS) = \{ \chi_i\}_{1 \leq i \leq 24}$ and $Con(HS) = \{
x_i^{HS}\}_{1 \leq i \leq 24}$. Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{HS}\} (2 \leq i \leq 22), \{x_{23}^{HS},x_{24}^{HS}\} \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i \leq 10), \{\chi_{11}, \chi_{12}\}, \{\chi_{i}\} (13 \leq i \leq 24) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{HS}\} (2 \leq i \leq 18), \{x_{19}^{HS},x_{20}^{HS}\}, \{x_{i}^{HS}\} (21 \leq i \leq 24) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 13), \{\chi_{14}, \chi_{15}\}, \{\chi_{i}\} (16 \leq i \leq
24) \right\rbrace.
\end{eqnarray*}
\end{small}
Then by Lemma \ref{11}, one can see that $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$ and $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ are supercharacter theories of
$HS$. We now apply \cite[Proposition 2.16]{10}, to prove that
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ is
another supercharacter theory of $HS$. These supercharacter
theories are different from $m(HS)$ and $M(HS)$ which concludes
that $s(HS) \geq 5$.
\item \textit{The Conway group $Co_{1}$}. Suppose $Irr(Co_{1}) = \{ \chi_i\}_{1 \leq i \leq 101}$ and $Con(Co_{1}) = \{ x_i^{Co_{1}}\}_{1 \leq i \leq 101}$.
Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Co_{1}}\} (2 \leq i \leq 96), \{x_{97}^{Co_{1}},x_{98}^{Co_{1}}\}, \{x_{i}^{Co_{1}}\} (99 \leq i \leq 101) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 26), \{\chi_{27}, \chi_{28}\}, \{\chi_{i}\} (29 \leq i \leq
101) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Co_{1}}\} (2 \leq i \leq 77), \{x_{78}^{Co_{1}},x_{79}^{Co_{1}}\}, \{x_{i}^{Co_{1}}\} (80 \leq i \leq 101) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 16), \{\chi_{17}, \chi_{18}\}, \{\chi_{i}\} (19 \leq i \leq
101) \right\rbrace.
\end{eqnarray*}
\end{small}
By Lemma \ref{11}, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$ and $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ are supercharacter theories of
$Co_1$ and by \cite[Proposition 2.16]{10}, $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$. Since these supercharacter
theories are different from $m(Co_{1})$ and $M(Co_{1})$,
$s(Co_1) \geq 5$. Hence the result follows.
\item \textit{The second Conway group $Co_{2}$}. Let
$Irr(Co_{2}) = \{ \chi_i\}_{1 \leq i \leq 60}$ and $Con(Co_{2})
= \{ x_i^{Co_{2}}\}_{1 \leq i \leq 60}$. Define:
{\small\begin{eqnarray*} \mathcal{K}_{1} &=& \left\lbrace \{1\},
\{x_{i}^{Co_{2}}\} (2 \leq i \leq 45),
\{x_{46}^{Co_{2}},x_{47}^{Co_{2}}\}, \{x_{i}^{Co_{2}}\} (48 \leq
i \leq 58),
\{x_{59}^{Co_{2}},x_{60}^{Co_{2}}\}\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 11), \{\chi_{12}, \chi_{13}\}, \{\chi_{i}\} (14 \leq i \leq
30), \{\chi_{31}, \chi_{32}\}, \{\chi_{i}\} (33 \leq i \leq 60)
\right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Co_{2}}\} (2 \leq i \leq 52), \{x_{53}^{Co_{2}},x_{54}^{Co_{2}}\}, \{x_{i}^{Co_{2}}\} (55 \leq i \leq 60)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 9), \{\chi_{10}, \chi_{11}\}, \{\chi_{i}\} (12 \leq i \leq
60) \right\rbrace.
\end{eqnarray*}}
By Lemma \ref{11} and \cite[Proposition 2.16]{10},
$\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ are
supercharacter theories of the Conway group $Co_2$. Since these
supercharacter theories are different from $m(Co_{2})$ and
$M(Co_{2})$, $s(Co_2) \geq 5$.
\item \textit{The third Conway group $Co_{3}$}. We assume that
$Irr(Co_{3})$ = $\{ \chi_i\}_{1 \leq i \leq 42}$ and
$Con(Co_{3})$ = $\{ x_i^{Co_{3}}\}_{1 \leq i \leq 42}$. Define:
\begin{footnotesize}
\begin{eqnarray*} \mathcal{K}_{1} &=& \left\lbrace \{1\},
\{x_{i}^{Co_{3}}\} (2 \leq i \leq 23),
\{x_{24}^{Co_{3}},x_{25}^{Co_{3}}\}, \{x_{i}^{Co_{3}}\} (26 \leq
i \leq 35),
\{x_{36}^{Co_{3}},x_{37}^{Co_{3}}\}, \{x_{i}^{Co_{3}}\} (38 \leq i \leq 42)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 5), \{\chi_{6}, \chi_{7}\}, \{\chi_{i}\} (8 \leq i \leq 17),
\{\chi_{18}, \chi_{19}\}, \{\chi_{i}\} (20 \leq i \leq 42)
\right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Co_{3}}\} (2 \leq i \leq 37), \{x_{38}^{Co_{3}},x_{39}^{Co_{3}}\}, \{x_{i}^{Co_{3}}\} (40 \leq i \leq 42)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 15), \{\chi_{16}, \chi_{17}\}, \{\chi_{i}\} (18 \leq i \leq
42) \right\rbrace.
\end{eqnarray*}
\end{footnotesize}
Then $m(Co_{3})$, $M(Co_{3})$, $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are five supercharacter
theories for $Co_3$. Thus $s(Co_3) \geq 5$, as required.
\item \textit{The McLaughlin group $McL$}. Suppose $Irr(McL) = \{ \chi_i\}_{1 \leq i \leq 24}$ and
$Con(McL) = \{ x_i^{McL}\}_{1 \leq i \leq 24}$. We now define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{McL}\} (2 \leq i \leq 15), \{x_{16}^{McL},x_{17}^{McL}\}, \{x_{i}^{McL}\} (18 \leq i \leq 24) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 6), \{\chi_{7}, \chi_{8}\}, \{\chi_{i}\} (9 \leq i \leq 24)
\right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{McL}\} (2 \leq i \leq 12), \{x_{13}^{McL},x_{14}^{McL}\}, \{x_{i}^{McL}\} (15 \leq i \leq 24) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 20), \{\chi_{21}, \chi_{22}\}, \{\chi_{i}\} (23 \leq i \leq
24) \right\rbrace.
\end{eqnarray*}
\end{small}
Since $m(McL)$, $M(McL)$, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are five supercharacter
theories of $McL$, $s(McL) \geq 5$.
\item {\textit{The Suzuki group $Suz$}}. The irreducible characters and conjugacy classes for the Suzuki group $Suz$ are
$Irr(Suz) = \{ \chi_i\}_{1 \leq i \leq 43}$ and $Con(Suz) = \{
x_i^{Suz}\}_{1 \leq i \leq 43}$, respectively. Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Suz}\} (2 \leq i \leq 40), \{x_{41}^{Suz},x_{42}^{Suz}\}, \{x_{43}^{Suz}\} \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 24), \{\chi_{25}, \chi_{26}\}, \{\chi_{i}\} (27 \leq i \leq
43) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Suz}\} (2 \leq i \leq 34), \{x_{35}^{Suz},x_{36}^{Suz}\}, \{x_{i}^{Suz}\} (37 \leq i \leq 43) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 12), \{\chi_{13}, \chi_{14}\}, \{\chi_{i}\} (15 \leq i \leq
43) \right\rbrace.
\end{eqnarray*}
\end{small}
Since this group has five supercharacter theories $m(Suz)$,
$M(Suz)$, $\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$, we
conclude that $s(Suz) \geq 5$.
\end{itemize}
This completes the proof.
\end{proof}
\begin{thm}
The Monster sections have at least five supercharacter theories.
\end{thm}
\begin{proof}
The Monster sections are eight simple groups $He$, $HN$, $Th$,
$Fi_{22}$, $Fi_{23}$, $Fi_{24}^\prime$, $B$ and $M$. We will
present five supercharacter theories in each case as follows:
\begin{itemize}
\item {\textit{The Held group $He$}}. Suppose $Irr(He) = \{ \chi_i\}_{1 \leq i \leq 33}$ and
$Con(He) = \{ x_i^{He}\}_{1 \leq i \leq 33}$ are irreducible
characters and conjugacy classes of the group $He$,
respectively. Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{He}\} (2 \leq i \leq 27), \{x_{28}^{He},x_{29}^{He}\}, \{x_{i}^{He}\} (30 \leq i \leq 33) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 29), \{\chi_{30}, \chi_{31}\}, \{\chi_{i}\} (32 \leq i \leq
33) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{He}\} (2 \leq i \leq 25), \{x_{26}^{He},x_{27}^{He}\}, \{x_{i}^{He}\} (28 \leq i \leq 33) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 6), \{\chi_{7}, \chi_{8}\}, \{\chi_{i}\} (9 \leq i \leq 33)
\right\rbrace.
\end{eqnarray*}
\end{small}
Since $m(He)$, $M(He)$, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are five supercharacter
theories for $He$, $s(He) \geq 5$.
\item {\textit{The Harada-Norton group $HN$}}. This group has exactly 54 conjugacy classes and irreducible characters. Suppose $Irr(HN) = \{ \chi_i\}_{1 \leq i \leq 54}$ and
$Con(HN) = \{ x_i^{HN}\}_{1 \leq i \leq 54}$. We also define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{HN}\} (2 \leq i \leq 38), \{x_{39}^{HN},x_{40}^{HN}\}, \{x_{i}^{HN}\} (41 \leq i \leq 54) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i \leq 50), \{\chi_{51}, \chi_{52}\}, \{\chi_{i}\} (53 \leq i \leq 54) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{HN}\} (2 \leq i \leq 52), \{x_{53}^{HN},x_{54}^{HN}\}\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 34), \{\chi_{35}, \chi_{36}\}, \{\chi_{i}\} (37 \leq i \leq
54) \right\rbrace.
\end{eqnarray*}
\end{small}
By Lemma \ref{11} and \cite[Proposition 2.16]{10},
$\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ are
supercharacter theories of $HN$ and apart from supercharacter
theories $m(HN)$ and $M(HN)$, it concludes that $m(HN) \geq 5$.
\item {\textit{The Thompson group $Th$}}. The Thompson group $Th$ has exactly 48 conjugacy classes and irreducible characters. We assume that
$Irr(Th) = \{ \chi_i\}_{1 \leq i \leq 48}$ and $Con(Th) = \{
x_i^{Th}\}_{1 \leq i \leq 48}$ and define:
{\small\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Th}\} (2 \leq i \leq 24), \{x_{25}^{Th}, x_{26}^{Th}\}, \{x_{i}^{Th}\} (27 \leq i \leq 39), \{x_{40}^{Th}, x_{41}^{Th}\}, \{x_{i}^{Th}\} (42 \leq i \leq 48) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 8), \{\chi_{9}, \chi_{10}\}, \{\chi_{i}\} (11 \leq i \leq
34), \{\chi_{35}, \chi_{36}\}, \{\chi_{i}\} (37
\leq i \leq 48) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Th}\} (2 \leq i \leq 36), \{x_{37}^{Th}, x_{38}^{Th}\}, \{x_{i}^{Th}\} (39 \leq i \leq 48) \right\rbrace ,\\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 21), \{\chi_{22}, \chi_{23}\}, \{\chi_{i}\} (24 \leq i \leq
48) \right\rbrace.
\end{eqnarray*}}
Again apply Lemma \ref{11} and \cite[Proposition 2.16]{10} to
deduce that $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and finally $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are supercharacter theories
for $Th$. Thus $s(Th) \geq 5$.
\item {\textit{The Fischer group $Fi_{22}$}}. The Fischer group $Fi_{22}$
has exactly $65$ conjugacy classes and irreducible characters. Set
$Irr(Fi_{22}) = \{ \chi_i\}_{1 \leq i \leq 65}$ and $Con(Fi_{22})
= \{ x_i^{Fi_{22}}\}_{1 \leq i \leq 65}$. We also define:
\begin{scriptsize}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Fi_{22}}\} (2
\leq i \leq 35), \{x_{36}^{Fi_{22}}, x_{37}^{Fi_{22}}\},
\{x_{i}^{Fi_{22}}\} (38 \leq i \leq 60),
\{x_{61}^{Fi_{22}}, x_{62}^{Fi_{22}}\}, \{x_{i}^{Fi_{22}}\} (63 \leq i \leq 65) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i \leq 39), \{\chi_{40}, \chi_{41}\}, \{\chi_{i}\} (42 \leq i \leq 50), \{\chi_{51}, \chi_{52}\}, \{\chi_{i}\} (53 \leq i \leq 65) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Fi_{22}}\} (2 \leq i \leq 54), \{x_{55}^{Fi_{22}}, x_{56}^{Fi_{22}}\}, \{x_{i}^{Fi_{22}}\} (57 \leq i \leq 65)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 42), \{\chi_{43}, \chi_{44}\}, \{\chi_{i}\} (45 \leq i \leq
65) \right\rbrace.
\end{eqnarray*}
\end{scriptsize}
Since $\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and finally
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$, are
three supercharacter theories of $Fi_{22}$ different from
$m(Fi_{22})$ and $M(Fi_{22})$, $s(Fi_{22}) \geq 5$.
\item {\textit{The Fischer group $Fi_{23}$}}. This group has exactly $98$ conjugacy classes and irreducible characters.
Set $Irr(Fi_{23}) = \{ \chi_i\}_{1 \leq i \leq 98}$ and
$Con(Fi_{23}) = \{ x_i^{Fi_{23}}\}_{1 \leq i \leq 98}$. Define:
\begin{footnotesize}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Fi_{23}}\} (2 \leq i \leq 79), \{x_{80}^{Fi_{23}}, x_{81}^{Fi_{23}}\}, \{x_{i}^{Fi_{23}}\} (82 \leq i \leq 98)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i \leq 16), \{\chi_{17}, \chi_{18}\}, \{\chi_{i}\} (19 \leq i \leq 98) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Fi_{23}}\} (2 \leq i \leq 62), \{x_{63}^{Fi_{23}}, x_{64}^{Fi_{23}}\}, \{x_{i}^{Fi_{23}}\} (65 \leq i \leq 79), \{x_{80}^{Fi_{23}}, x_{81}^{Fi_{23}}\}, \{x_{i}^{Fi_{23}}\} (82 \leq i \leq 98)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 14), \{\chi_{15}, \chi_{16}\}, \{\chi_{17}, \chi_{18}\},
\{\chi_{i}\} (19 \leq i \leq 98) \right\rbrace,\\
\mathcal{C}_{1} &=& (\mathcal{X}_{1}, \mathcal{K}_{1}),\ \
\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2}),\ \
\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}.
\end{eqnarray*}
\end{footnotesize}
Since $\{ m(Fi_{23}), M(Fi_{23}), \mathcal{C}_{1},
\mathcal{C}_{2}, \mathcal{C}_{3}\} \subseteq Sup(Fi_{23})$,
$s(Fi_{23}) \geq 5$.
\item {\textit{The Fischer group $Fi_{24}^{\prime}$}}. The largest Fischer group has exactly $108$ conjugacy classes and irreducible characters. Set
$Irr(Fi_{24}^{\prime}) = \{ \chi_i\}_{1 \leq i \leq 108}$ and
$Con(Fi_{24}^{\prime}) = \{ x_i^{Fi_{24}^{\prime}}\}_{1 \leq i
\leq 108}$. Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\},
\{x_{i}^{Fi_{24}^{\prime}}\} (2 \leq i \leq 105),
\{x_{106}^{Fi_{24}^{\prime}}, x_{107}^{Fi_{24}^{\prime}}\},
\{x_{108}^{Fi_{24}^{\prime}}\}\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i \leq 98), \{\chi_{99}, \chi_{100}\}, \{\chi_{i}\} (101 \leq i \leq 108) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\},
\{x_{i}^{Fi_{24}^{\prime}}\} (2 \leq i \leq 80),
\{x_{81}^{Fi_{24}^{\prime}}, x_{82}^{Fi_{24}^{\prime}}\},
\{x_{i}^{Fi_{24}^{\prime}}\} (83 \leq i \leq 108)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 100), \{\chi_{101}, \chi_{102}\}, \{\chi_{i}\} (103 \leq i
\leq 108) \right\rbrace.
\end{eqnarray*}
\end{small}
Since $m(Fi_{24}^{\prime})$, $M(Fi_{24}^{\prime})$,
$\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ are
supercharacter theories of $Fi_{24}$, $s(Fi_{24}) \geq 5$.
\item {\textit{The Baby Monster group $B$}}. This group has exactly $184$ conjugacy classes and irreducible characters.
Suppose $Irr(B) = \{ \chi_i\}_{1 \leq i \leq 184}$, $Con(B) = \{
x_i^{B}\}_{1 \leq i \leq 184}$ and define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{B}\} (2 \leq i \leq 177), \{x_{178}^{B}, x_{179}^{B}\}, \{x_{i}^{B}\} (180 \leq i \leq 184)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 177), \{\chi_{178}, \chi_{179}\}, \{\chi_{i}\} (180 \leq i
\leq 184) \right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{B}\} (2 \leq i \leq 171), \{x_{172}^{B}, x_{173}^{B}\}, \{x_{i}^{B}\} (174 \leq i \leq 184)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 20), \{\chi_{21}, \chi_{22}\}, \{\chi_{i}\} (23 \leq i \leq
184) \right\rbrace.
\end{eqnarray*}
\end{small}
Since $m(B)$, $M(B)$, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are supercharacter theories
of $B$, $s(B) \geq 5$, as desired.
\item {\textit{The Monster group $M$}}. The largest sporadic group
$M$ has exactly $198$ conjugacy classes and irreducible characters.
Set $Irr(M) = \{ \chi_i\}_{1 \leq i \leq 194}$ and $Con(M) = \{
x_i^{M}\}_{1 \leq i \leq 194}$.
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{M}\} (2 \leq i \leq 192), \{x_{193}^{M}, x_{194}^{M}\}\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 46), \{\chi_{47}, \chi_{48}\}, \{\chi_{i}\} (49 \leq i \leq
194) \right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{M}\} (2 \leq i \leq 188), \{x_{189}^{M}, x_{190}^{M}\}, \{x_{i}^{M}\} (191 \leq i \leq 194)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i \leq 123), \{\chi_{124}, \chi_{125}\}, \{\chi_{i}\} (126 \leq i \leq 194) \right\rbrace.
\end{eqnarray*}
\end{small}
We can see that the group $M$ has at least $5$ supercharacter
theories as $m(M)$, $M(M)$, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$. This shows that $s(M) \geq
5$.
\end{itemize}
This completes the proof.
\end{proof}
\begin{thm}
The Pariahs have at least five supercharacter theories.
\end{thm}
\begin{proof}
The Pariahs are six sporadic groups $J_1$, $O^\prime{N}$, $J_3$,
$Ru$, $J_4$ and $Ly$. Our main proof will consider six separate
cases as follows:
\begin{itemize}
\item \textit{Janko group $J_{1}$}. The first Janko group $J_1$ has exactly $15$ conjugacy classes and irreducible characters. Set $Irr(J_{1}) = \{ \chi_i\}_{1 \leq i \leq 15}$ and
$Con(J_{1}) = \{ x_i^{J_{1}}\}_{1 \leq i \leq 15}$.
\begin{footnotesize}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{J_{1}}\} (2 \leq
i \leq 3), \{x_{4}^{J_{1}}, x_{5}^{J_{1}}\}, \{x_{i}^{J_{1}}\} (6
\leq i \leq 7), \{x_{8}^{J_{1}}, x_{9}^{J_{1}}\},
\{x_{10}^{J_{1}}\},
\{x_{11}^{J_{1}}, x_{12}^{J_{1}}\}, \{x_{i}^{J_{1}}\} (13 \leq i \leq 15)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3}\},
\{\chi_{i}\} (4 \leq i \leq 6), \{\chi_{7}, \chi_{8}\},
\{\chi_{i}\} (9 \leq i \leq 12), \{\chi_{13}, \chi_{14}\},
\{\chi_{15}\} \right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{J_{1}}\} (2 \leq i \leq 12), \{x_{13}^{J_{1}}, x_{14}^{J_{1}}, x_{15}^{J_{1}}\} \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 8), \{\chi_{9}, \chi_{10}, \chi_{11}\}, \{\chi_{i}\} (12
\leq i \leq 15) \right\rbrace.
\end{eqnarray*}
\end{footnotesize}
Since $m(J_{1})$, $M(J_{1})$, $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are supercharacter theories
of $J_1$, $s(J_1) \geq 5$.
\item {\textit{The O$^{\prime}$Nan group $O^{\prime}N$}}. The O$^{\prime}$Nan group $O^{\prime}N$ has exactly $30$ conjugacy classes and irreducible characters.
Set $Irr(O^{\prime}N) = \{ \chi_i\}_{1 \leq i \leq 30}$ and
$Con(O^{\prime}N) = \{ x_i^{O^{\prime}N}\}_{1 \leq i \leq 30}$.
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{O^{\prime}N}\}
(2 \leq i \leq 21), \{x_{22}^{O^{\prime}N}, x_{23}^{O^{\prime}N},
x_{24}^{O^{\prime}N}\},
\{x_{i}^{O^{\prime}N}\} (25 \leq i \leq 30) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 25), \{\chi_{26}, \chi_{27}, \chi_{28}\}, \{\chi_{i}\} (29
\leq i \leq 30) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{O^{\prime}N}\} (2 \leq i \leq 26), \{x_{27}^{O^{\prime}N}, x_{28}^{O^{\prime}N}\}, \{x_{i}^{O^{\prime}N}\}
(29 \leq i \leq 30) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 28), \{\chi_{29}, \chi_{30}\}\right\rbrace.
\end{eqnarray*}
\end{small}
Apply Lemma \ref{11} and \cite[Proposition 2.16]{10} to deduce
that $\mathcal{C}_{1} = (\mathcal{X}_{1}, \mathcal{K}_{1})$,
$\mathcal{C}_{2} = (\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ are
supercharacter theories of $O^{\prime}N$. Hence $s(O^{\prime}N)
\geq 5$, as desired.
\item {\textit{The Janko group $J_{3}$}}. We assume that
$Irr(J_{3}) = \{ \chi_i\}_{1 \leq i \leq 21}$ and $Con(J_{3}) =
\{ x_i^{J_{3}}\}_{1 \leq i \leq 21}$. Define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{J_{3}}\} (2 \leq i \leq 9), \{x_{10}^{J_{3}}, x_{11}^{J_{3}}, x_{12}^{J_{3}}\}, \{x_{i}^{J_{3}}\} (13 \leq i \leq 21) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 13), \{\chi_{14}, \chi_{15}, \chi_{16}\}, \{\chi_{i}\} (17
\leq i \leq 21) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{J_{3}}\} (2 \leq i \leq 19), \{x_{20}^{J_{3}}, x_{21}^{J_{3}}\} \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3}\},
\{\chi_{i}\} (4 \leq i \leq 21) \right\rbrace.
\end{eqnarray*}
\end{small}
Since $m(J_{3})$, $M(J_{3})$, $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$ are supercharacter theories
of $J_3$, $s(J_3) \geq 5$.
\item {\textit{The Rudvalis group $Ru$}}. Set $Irr(Ru) = \{ \chi_i\}_{1 \leq i \leq 36}$ and $Con(Ru) = \{ x_i^{Ru}\}_{1 \leq i \leq 36}$ and
define:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Ru}\} (2 \leq i \leq 34), \{x_{35}^{Ru}, x_{36}^{Ru}\}\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 33), \{\chi_{34}, \chi_{35}\}, \{\chi_{36}\} \right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Ru}\} (2 \leq i \leq 31), \{x_{32}^{Ru}, x_{33}^{Ru}, x_{34}^{Ru}\}, \{x_{i}^{Ru}\} (35 \leq i \leq 36) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 16), \{\chi_{17}, \chi_{18}, \chi_{19}\}, \{\chi_{i}\} (20
\leq i \leq 36) \right\rbrace.
\end{eqnarray*}
\end{small}
Apply again Lemma \ref{11} and \cite[Proposition 2.16]{10} to
deduce that the Rudvalis group $Ru$ has at least five
supercharacter theories as $m(Ru)$, $M(Ru)$, $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$. Hence $s(Ru) \geq 5$, as
required.
\item {\textit{The Janko group $J_{4}$}}. This group has exactly $62$ conjugacy classes and irreducible characters. Define
$Irr(J_{4}) = \{ \chi_i\}_{1 \leq i \leq 62}$, $Con(J_{4}) = \{
x_i^{J_{4}}\}_{1 \leq i \leq 62}$ and define:
\begin{footnotesize}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{J_{4}}\} (2 \leq i \leq 49), \{x_{50}^{J_{4}}, x_{51}^{J_{4}}, x_{52}^{J_{4}}\}, \{x_{i}^{J_{4}}\} (53 \leq i \leq 62)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 52), \{\chi_{53}, \chi_{54}, \chi_{55}\}, \{\chi_{i}\} ( 56
\leq i \leq 62) \right\rbrace, \\
\mathcal{K}_{2} &=&
\left\lbrace \{1\}, \{x_{i}^{J_{4}}\} (2 \leq i \leq 42),
\{x_{43}^{J_{4}}, x_{44}^{J_{4}}, x_{45}^{J_{4}}\},
\{x_{i}^{J_{4}}\} (46 \leq i \leq 49),
\{x_{50}^{J_{4}}, x_{51}^{J_{4}}, x_{52}^{J_{4}}\}, \{x_{i}^{J_{4}}\} (53 \leq i \leq 62) \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 52), \{\chi_{53}, \chi_{54}, \chi_{55}\}, \{\chi_{56},
\chi_{57}, \chi_{58}\}, \{\chi_{i}\} (59 \leq i \leq 62)
\right\rbrace.
\end{eqnarray*}
\end{footnotesize}
The Janko group $J_{4}$ has at least five supercharacter theories
as $m(J_{4})$, $M(J_{4})$, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$. Therefore, $s(J_4) \geq 5$.
\item \textit{The Lyons group $Ly$}. The Lyons group $Ly$ has exactly $53$ irreducible characters and conjugacy classes. Set $Irr(Ly) = \{ \chi_i\}_{1 \leq i \leq 53}$
and $Con(Ly) = \{ x_i^{Ly}\}_{1 \leq i \leq 53}$ and define:
\begin{footnotesize}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{Ly}\} (2 \leq i
\leq 37), \{x_{38}^{Ly}, x_{39}^{Ly}, x_{40}^{Ly}, x_{41}^{Ly},
x_{42}^{Ly}\},
\{x_{i}^{Ly}\} (43 \leq i \leq 53)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 38), \{\chi_{39}, \chi_{40}, \chi_{41}, \chi_{42},
\chi_{43}\}, \{\chi_{i}\} ( 44 \leq i \leq 53) \right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{Ly}\} (2 \leq i \leq 37), \{x_{38}^{Ly}, x_{39}^{Ly}, x_{40}^{Ly}, x_{41}^{Ly}, x_{42}^{Ly}\}, \{x_{i}^{Ly}\} (43 \leq i \leq 50), \{x_{51}^{Ly}, x_{52}^{Ly}, x_{53}^{Ly}\}\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 25), \{\chi_{26}, \chi_{27}, \chi_{28}\}, \{\chi_{i}\} (29
\leq i \leq 38),
\{\chi_{39}, \chi_{40}, \chi_{41}, \chi_{42}, \chi_{43}\}, \{\chi_{i}\} (44 \leq i \leq 53) \right\rbrace.
\end{eqnarray*}
\end{footnotesize}
Since the group $Ly$ has at least five supercharacter theories as
$m(Ly)$, $M(Ly)$, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$, $s(Ly) \geq 5$.
\end{itemize}
This completes the proof.
\end{proof}
\begin{lem}
The Janko group $J_{2}$ has at least three supercharacter
theories.
\end{lem}
\begin{proof}
A simple investigation shows that the supercharacter theories of
the Janko group $J_2$ are $m(J_{2})$, $M(J_{2})$, $\mathcal{C} =
(\mathcal{X}, \mathcal{K})$ such that
\begin{footnotesize}
\begin{eqnarray*}
\mathcal{K} &=& \left\lbrace \{1\}, \{x_{i}^{J_{2}}\} (2 \leq i
\leq 6), \{x_{7}^{J_{2}},x_{8}^{J_{2}}\},
\{x_{9}^{J_{2}},x_{10}^{J_{2}}\}, \{x_{i}^{J_{2}}\}
(11 \leq i \leq 14), \{x_{15}^{J_{2}},x_{16}^{J_{2}}\}, \{x_{17}^{J_{2}},x_{18}^{J_{2}}\}, \{x_{19}^{J_{2}}\}, \{x_{20}^{J_{2}},x_{21}^{J_{2}}\} \right\rbrace, \\
\mathcal{X} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3}\},
\{\chi_{4}, \chi_{5}\}, \{\chi_{i}\} (6 \leq i \leq 7),
\{\chi_{8}, \chi_{9}\},\{\chi_{i}\}
(10 \leq i \leq 13), \{\chi_{14}, \chi_{15}\}, \{\chi_{16}, \chi_{17}\}, \{\chi_{i}\} (18 \leq i \leq 20)\right\rbrace,
\end{eqnarray*}
\end{footnotesize}
proving the lemma.
\end{proof}
We run a GAP program to be sure that $s(J_2) = 3$, but our
program after some days stopped, because it needs a huge amount
of RAM.
\begin{conj}
$s(J_2) = 3$.
\end{conj}
\section{Supercharacter Theory Construction for Alternating and Suzuki Groups}
The intention of this section is to move a step towards a
classification of the finite simple groups with exactly three
and four supercharacter theories. We start with the cyclic group
of order $p$.
\begin{lem}\label{12}
Suppose $p$ is prime. Then,
\begin{enumerate}
\item $s(Z_{p}) = 3$ if and only if $p = 5$,
\item $s(Z_{p}) = 4$ if and only if $p$ is a Sophie Germain prime.
\end{enumerate}
\end{lem}
\begin{proof}
It is clear that $p$ is an odd prime. By \cite[Theorem 6.32 and
Table 1]{9}, $s(Z_p) = d(p-1)$.
\begin{enumerate}
\item \textit{$s(Z_{p}) = 3$.} In this case,
$s(Z_{p}) = d(p-1)$ and so $p \geq 5$. Since $p-1$ is an even
integer, the case of $p > 5$ cannot be occurred and so $p = 5$, as
desired.
\item \textit{$s(Z_{p}) = 4$.} Since $d(p-1) = 4$, $p-1 = q^3$, $q$ is prime, or $p-1 = 2r$,
where $r$ is prime. If $p-1 = q^3$ then $q = 2$ and $p = 9$, a
contradiction. So, $p-1 = 2r$, where $r$ is prime. This shows
that $p$ is a Sophie Germain prime.
\end{enumerate}
This completes the proof.
\end{proof}
The following well-known results are crucial in the
classification of alternating simple groups with exactly three or
four supercharacter theories.
\begin{thm}\label{4.1} The following are hold:
\begin{enumerate}
\item {\rm(Berggren \cite{5})} Every irreducible characters of the
alternating group $A_n$ are real valued if and only if $n \in \{
1, 2, 5, 6, 10, 14\}$.
\item {\rm(Grove \cite[Proposition 8.2.1]{85})} If $K$ is a conjugacy
class in $S_{n}$, $K \subset A_{n}$, and $\sigma \in K$, then $K$
is a conjugacy class in $A_{n}$ if and only if some odd elements
of $S_{n}$ commutes with $\sigma$; if that is not the case, then
the conjugacy class $K$ splits as the union of two
$A_{n}-$classes, each of size $|K|/2$. If $\lambda$ is the
(partition) type of $\sigma$ then $K$ splits if and only if the
parts of $\lambda$ are all odd and all different from each other.
\item Suppose $x \in A_n$ is a product of $r$ pair-wise disjoint
cycles including all fixed points as singleton cycles. Then $x^{A_{n}}$
is non-real if and only if $\sum_{j=1}^{m}\frac{r_j-1}{2}$ is
odd.
\end{enumerate}
\end{thm}
\begin{thm}
The simple alternating group $A_{n}$ has exactly three
supercharacter theories if and only if $n = 5$ or $7$. There is
no simple alternating groups with exactly four supercharacter
theories.
\end{thm}
\begin{proof}
Suppose $n \geq 5$. It is easy to see that the alternating groups
$A_5$ and $A_7$ have exactly three supercharacter theories. Our
main proof will consider three separate cases as follows:
\begin{enumerate}
\item \textit{All character values of $A_n$ are real}. Since $n \geq 5$, Theorem
\ref{4.1}(1) implies that $n = 5, 6, 10$ or $14.$ In what follows
three non-trivial supercharacter theories for the alternating
groups $A_6, A_{10}$ and $A_{14}$ are presented.
\begin{enumerate}
\item By a GAP program, one can see that the group $A_{5}$ has exactly $3$ supercharacter theories
$m(A_{5})$, $M(A_{5})$ and $\mathcal{C} = (\mathcal{X},
\mathcal{K})$ such that
\begin{small}
\begin{eqnarray*}
\mathcal{K} &=& \left\lbrace \{1\}, \{x_{2}^{A_{5}}\}, \{x_{3}^{A_{5}}\}, \{x_{4}^{A_{5}}, x_{5}^{A_{5}}\}\right\rbrace, \\
\mathcal{X} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3}\},
\{\chi_{i}\} (4 \leq i \leq 5) \right\rbrace.
\end{eqnarray*}
\end{small}
Here, $Irr(A_{5})$ = $\{ \chi_1, \chi_2, \chi_3, \chi_4,
\chi_{5}\}$ and $Con(A_{5})$ = $\{ x_1^{A_{5}}, x_2^{A_{5}},
x_3^{A_{5}}, x_4^{A_{5}}, x_{5}^{A_{5}}\}$.
\item Suppose $Irr(A_{6})$ = $\{ \chi_1, \chi_2, \ldots,
\chi_{7}\}$ and $Con(A_{6})$ = $\{ x_1^{A_{6}}, x_2^{A_{6}},
\ldots, x_{7}^{A_{6}}\}$. Define $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$ and $\mathcal{C}_{2} =
(\mathcal{X}_{2}, \mathcal{K}_{2})$ as follows:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{2}^{A_{6}}\}, \{x_{3}^{A_{6}}, x_{4}^{A_{6}}\}, \{x_{5}^{A_{6}}\}, \{x_{6}^{A_{6}}, x_{7}^{A_{6}}\}\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3}\},
\{\chi_{4}, \chi_{5}\}, \{\chi_{6}\}, \{\chi_{7}\}\right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{2}^{A_{6}}, x_{5}^{A_{6}}\}, \{x_{3}^{A_{6}}, x_{4}^{A_{6}}\}, \{x_{6}^{A_{6}}\}, \{x_{7}^{A_{6}}\}\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{2}, \chi_{3},
\chi_{7}\}, \{\chi_{4}\}, \{\chi_{5}\}, \{\chi_{6}\}\right\rbrace,
\end{eqnarray*}
\end{small}
If $\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ then
$\mathcal{C}_{3} $ is a supercharacter theory for $A_6$
different from $\mathcal{C}_1$, $\mathcal{C}_2$, $m(A_{6})$ and
$M(A_{6})$. Therefore, $A_{6}$ has at least five supercharacter
theories.
\item Suppose $Irr(A_{10})$ = $\{ \chi_1, \chi_2, \ldots, \chi_{24}\}$, $Con(A_{10}) = \{ x_1^{A_{10}}, x_2^{A_{10}}, \ldots,
x_{24}^{A_{10}}\}$, $\mathcal{C}_{1} =
(\mathcal{X}_{1},\mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$, where
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{A_{10}}\} (2 \leq i \leq 19), \{x_{20}^{A_{10}}, x_{21}^{A_{10}}\}, \{x_{i}^{A_{10}}\} (22 \leq i \leq 24)\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 19), \{\chi_{20}, \chi_{21}\}, \{\chi_{i}\} (22 \leq i \leq
24) \right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{A_{10}}\} (2 \leq i \leq 22), \{x_{23}^{A_{10}}, x_{24}^{A_{10}}\}\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 11), \{\chi_{12}, \chi_{13}\}, \{\chi_{i}\} (14 \leq i \leq
24) \right\rbrace.
\end{eqnarray*}
\end{small}
Then $\mathcal{C}_{i} = (\mathcal{X}_{i},\mathcal{K}_{i})$, $1
\leq i \leq 3$, are three supercharacter theories of $A_{10}$
different from $m(A_{10})$ and $M(A_{10})$.
\item We claim that the alternating group $A_{14}$ has at least $5$ supercharacter
theories. These are $m(A_{14})$, $M(A_{14})$, $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2},\mathcal{K}_{2})$ and
$\mathcal{C}_{3} = \mathcal{C}_{1} \vee \mathcal{C}_{2}$ such that
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{A_{14}}\} (2 \leq i \leq 70), \{x_{71}^{A_{14}}, x_{72}^{A_{14}}\}\right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 19), \{\chi_{20}, \chi_{21}\}, \{\chi_{i}\} (22 \leq i \leq
72) \right\rbrace,\\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{A_{14}}\} (2 \leq i \leq 67), \{x_{68}^{A_{14}}, x_{69}^{A_{14}}\}, \{x_{i}^{A_{14}}\} (70 \leq i \leq 72)\right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 56), \{\chi_{57}, \chi_{58}\}, \{\chi_{i}\} (59 \leq i \leq
72) \right\rbrace.
\end{eqnarray*}
\end{small}\end{enumerate}
\item \textit{The alternating group $A_{n}$ has exactly one pair of non-real valued
irreducible character.} In this case, we will prove $ n \in \{7,
9, 11, 15, 18, 19, 23\}$. Since the number of non-real irreducible
characters is equal to the number of conjugacy classes that are
not preserved by the inversion mapping, we need those $n$ for
which $A_n$ has exactly two such conjugacy classes. Suppose $x
\in A_n$. By \cite[Proposition 12.17(2)]{110}, $x^{S_n} = x^{A_n}
\cup (1 2)x(1 2)^{A_n}$ if and only if $C_{S_n}(x) = C_{A_n}(x)$.
By Theorem \ref{4.1}(2), the last one is satisfied if and only if
$\dfrac{n - r}{2}$ is odd, where $r$ is the number of cycles in
decomposition of $x$. Our aim is to find all natural numbers $n$
such that $A_n$ has exactly one non-real class. Our main proof
will consider the following separate cases:
\begin{enumerate}
\item $n \equiv 0 ~(mod ~4)$. If $n \geq 8$ then $[1, n - 1]$ and $[3, n - 3]$ are two partitions with given
properties. So, $n = 4$ which contradicts by our main assumption
that $n \geq 5$.
\item $n \equiv 1~ (mod~ 4)$. If $ n \geq 13 $ then $ [1, 3, n - 4] $ and $ [1, 5, n - 6]
$ are two partitions with this property that all parts are odd and
$\frac{n-3}{2}$ ($r = 3$) is odd. Thus $n = 5$ or $9$. By Theorem
\ref{4.1}(1), the alternating group $A_5$ does not have non-real
class and the alternating group $A_9$ has exactly a unique pair of
non-real class.
\item $n \equiv 2~ (mod~ 4)$. If $ n \geq 22 $ then $ [1, 3, 5, n - 9] $ and $ [1, 3, 7, n - 11]
$ are two partitions with given properties and so $n = 6, 10,
14$ or $18$. By Theorem \ref{4.1}(1), the alternating groups
$A_6, A_{10}$ and $A_{14}$ don't have non-real conjugacy class,
but the alternating group $A_{18}$ has a unique pair of conjugate
non-real characters.
\item $n \equiv 3~ (mod~ 4)$. If $ n \geq 31 $ then $ [1, 3, 5, 7, n - 16] $ and $ [1, 3, 5, 9, n - 18]
$ are two partitions with this property that all parts are odd and
$\frac{n-5}{2}$ ($r = 5$) is odd. Thus $n = 7, 11, 15, 19, 23$ or
$27$. Since $[27]$ and $[1,5,7,9,11]$ are two partitions with
mentioned properties, the case of $n = 27$ cannot be happened.
Other cases are solution of our problem.
\end{enumerate}
Hence, the alternating group $A_n$, $n \geq 5$, has a unique
non-real conjugacy class if and only if $n \in \{ 7, 9, 11, 15,
18, 19, 23\}$. A simple calculations by GAP shows that the
alternating group $A_7$ has exactly three supercharacter theories
$m(A_{7})$, $M(A_{7})$ and $\mathcal{C} = (\mathcal{X},
\mathcal{K})$, where
\begin{small}
\begin{eqnarray*}
\mathcal{K} &=& \left\lbrace \{1\}, \{x_{i}^{A_{7}}\} (2 \leq i \leq 7), \{x_{8}^{A_{7}}, x_{9}^{A_{7}}\}\right\rbrace, \\
\mathcal{X} &=& \left\lbrace \{1\}, \chi_{2}, \{\chi_{3},
\chi_{4}\}, \{\chi_{i}\} (5 \leq i \leq 9) \right\rbrace.
\end{eqnarray*}
\end{small}
The alternating group $A_{9}$ has at least $5$ supercharacter
theories $m(A_{9})$, $M(A_{9})$, $\mathcal{C}_{1} =
(\mathcal{X}_{1}, \mathcal{K}_{1})$, $\mathcal{C}_{2} =
(\mathcal{X}_{2}, \mathcal{K}_{2})$ and $\mathcal{C}_{3} =
\mathcal{C}_{1} \vee \mathcal{C}_{2}$. The partitions of
conjugacy classes and irreducible characters of $\mathcal{C}_1$
and $\mathcal{C}_2$ are defined as follows:
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \left\lbrace \{1\}, \{x_{i}^{A_{9}}\} (2 \leq i \leq 12), \{x_{13}^{A_{9}},x_{14}^{A_{9}}\}, \{x_{i}^{A_{9}}\} (15 \leq i \leq 18) \right\rbrace, \\
\mathcal{X}_{1} &=& \left\lbrace \{1\}, \{\chi_{2}\}, \{\chi_{3},
\chi_{4}\}, \{\chi_{i}\} (5 \leq i \leq 18)\right\rbrace, \\
\mathcal{K}_{2} &=& \left\lbrace \{1\}, \{x_{i}^{A_{9}}\} (2 \leq i \leq 16), \{x_{17}^{A_{9}},x_{18}^{A_{9}}\} \right\rbrace, \\
\mathcal{X}_{2} &=& \left\lbrace \{1\}, \{\chi_{i}\} (2 \leq i
\leq 6), \{\chi_{7}, \chi_{8}\}, \{\chi_{i}\} (9 \leq i \leq
18)\right\rbrace.
\end{eqnarray*}
\end{small}
The alternating groups $A_{11}, A_{15}, A_{18}$, $A_{19}$ and
$A_{23}$ has at least one non-real irreducible character and one
non-rational irreducible real character. So, by Lemma \ref{11} it
has at least five supercharacter theories $m$, $M$,
$\mathcal{C}_1$, $\mathcal{C}_2$ and $\mathcal{C}_3 =
\mathcal{C}_1 \vee \mathcal{C}_2$, proving this case.
\item \textit{The alternating group $A_{n}$ has at least two pairs of non-real valued
irreducible characters.} We first assume that $n > 24$. If $A =
\{ \chi(x) \ | \ \chi \in Irr(G)\}$ and $\mathbb{Q}(A)$ denotes
the filed generated by $\mathbb{Q}$ and $A$ then by
\cite[Theorem]{112}, the character table of $A_n$ has both
irrational and non-real character values. On the other hand, by
\cite[Theorem 2.5.13]{111}, each row or column of the character
table of $A_n$ contains at most one pair of irrational numbers.
Now by Lemma \ref{11}, we have at least five supercharacter
theories, as required.
Next we assume that $n \leq 24$. Set $\Gamma_1 = \{ 5, 6, 7, 9,
10, 11, 14, 15, 18, 19, 23\}$ and $\Gamma_2 = \{ 8, 12, 13, 16,
17, 20, 21, 22, 24\}$. If $n \in \Gamma_1$ then the number of
supercharacter theories of $A_n$ are investigated in Cases (1)
and (2). So, we have to prove that $s(A_n) \geq 5$, when $n \in
\Gamma_2$. By an easy calculation with GAP, one can see that if
$n \in \Gamma_2$ then $A_n$ has at least two pairs of non-real
valued irreducible characters and by Lemma \ref{11}, $s(A_n) \geq
5$.
\end{enumerate}
Hence the result.
\end{proof}
In the end of this paper we prove that the simple Suzuki group
$Sz(q), q = 2^{2n+1}$ has at least six super character theories.
\begin{thm}
The Suzuki group $Sz(q)$ has at least $6$ supercharacter theories
as: $m(Sz(q))$, $ M(Sz(q)) $, $\mathcal{C}_{1} = (\mathcal{X}_{1},
\mathcal{K}_{1})$, $\mathcal{C}_{2} = (\mathcal{X}_{2},
\mathcal{K}_{2})$, $\mathcal{C}_{3} = (\mathcal{X}_{3},
\mathcal{K}_{3})$ and $\mathcal{C}_{4} = (\mathcal{X}_{4},
\mathcal{K}_{4})$, where
\begin{small}
\begin{eqnarray*}
\mathcal{K}_{1} &=& \{\{1\},\{\rho, \rho^{-1}\}, Con(G(q)) - \{\rho, \rho^{-1}\}\},\\
\mathcal{X}_{1} &=& \{\{1\},\{W_{1}, W_{2}\}, Irr(G(q))-\{W_{1}, W_{2}\}\},\\
\mathcal{K}_{2} &=& \{\{1\},\{\pi_{0}\}, Con(G(q)) - \{\pi_{0}\}\},\\
\mathcal{X}_{2} &=& \{\{1\},\{X_{i}\}, Irr(G(q))-\{X_{i}\}\},\\
\mathcal{K}_{3} &=& \{\{1\},\{\pi_{1}\}, Con(G(q)) - \{\pi_{1}\}\},\\
\mathcal{X}_{3} &=& \{\{1\},\{Y_{j}\}, Irr(G(q))-\{Y_{j}\}\},\\
\mathcal{K}_{4} &=& \{\{1\},\{\pi_{2}\}, Con(G(q)) - \{\pi_{2}\}\},\\
\mathcal{X}_{4} &=& \{\{1\},\{Z_{k}\}, Irr(G(q))-\{Z_{k}\}\}.
\end{eqnarray*}
\end{small}\end{thm}
\begin{proof}
The schematic form of the character table of $Sz(q), q =
2^{2n+1}$, is shown in Table \ref{abc}, see \cite{15,16} for
details.
\begin{table}[htp]
\caption{The Schematic Form of the Character Table of
$Sz(q)$.}\label{abc}
\begin{center}
\begin{tabular}{ccccccc}
Irreducible Characters & Degrees & $\#$Irreducible Characters\\
\hline
$X$ & $q^{2}$ & $1$ \\
$X_{i}$ & $q^{2} + 1$ & $q/2 - 1$ \\
$Y_{j}$ & $(q - r + 1)(q - 1)$ & $(q + r)/4$\\
$Z_{k}$ & $(q + r + 1)(q - 1)$ & $(q - r)/4$\\
$W_{l}$ & $r(q - 1)/2$ & $2$\\ \hline
\end{tabular}
\end{center}
\end{table}
In this table, the first column designates the characters, the
second column indicates the degrees, and the last one is the
number of characters of each degree.
Suppose $2q = r^{2}$. The Suzuki group $Sz(q)$ contains cyclic
groups of order $ q - 1 $, $q + r + 1$ and $ q - r + 1 $. These
subgroups are denoted by $A_{0}$, $ A_{1} $ and $ A_{2} $,
respectively. We also assume that $ \pi_{i} $ is a typical
non-identity element of $ A_{i} $, $i = 0, 1, 2 $, $ \sigma =
(0, 1)$ and $ \rho = (1, 0) $.
Let $ \varepsilon_{0} $, $ \varepsilon_{1} $ and $
\varepsilon_{2} $ be a $ (q - 1)^{th} $, a primitive $ (q + r +
1)^{th} $ and a $ (q - r + 1)^{th} $ root of unity. We also
assume that $ \xi_{0} $, $ \xi_{1} $ and $ \xi_{2} $ are
generators of $ A_{0} $, $ A_{1} $ and $ A_{2} $, respectively.
Define $ \varepsilon_{0}^{i} $, $ \varepsilon_{1}^{i} $ and $
\varepsilon_{2}^{i} $ as follows:
\begin{eqnarray*}
\varepsilon_{0}^{i}(\xi_{0}^{j}) &=& \varepsilon_{0}^{ij} +
\varepsilon_{0}^{-ij};~~~~(~i = 1, \ldots, q/2 - 1),\\
\varepsilon_{1}^{i}(\xi_{1}^{k}) &=& \varepsilon_{1}^{ik} +
\varepsilon_{1}^{ikq} + \varepsilon_{1}^{-ik} +
\varepsilon_{1}^{-ikq};~~~~(~i = 1, \ldots, q + r),\\
\varepsilon_{2}^{i}(\xi_{2}^{k}) &=& \varepsilon_{2}^{ik} +
\varepsilon_{2}^{ikq} + \varepsilon_{2}^{-ik} +
\varepsilon_{2}^{-ikq}.
\end{eqnarray*}
The functions $ \varepsilon_{0}^{i} $, $ \varepsilon_{1}^{i} $
and $ \varepsilon_{2}^{i} $ are characters of $ A_{0} $, $ A_{1}
$ and $ A_{2} $, respectively. Following Suzuki \cite{15}, the
character table of $Sz(q)$, is computed in Table \ref{bcd}.
\begin{table}[htp]
\caption{The character table of $Sz(q)$.} \label{bcd}
\centering
\begin{tabular}{ccccccc}
$$ & $1$ & $\sigma$ & $\rho, \rho^{-1}$ & $\pi_{0}$ & $\pi_{1}$ & $\pi_{2}$ \\
$X$ & $q^{2}$ & $0$ & $0$ & $1$ & $-1$ & $-1$ \\
$X_{i}$ & $q^{2} + 1$ & $1$ & $1$ & $\varepsilon_{0}^{i}(\pi_{0})$ & $0$ & $0$ \\
$Y_{j}$ & $(q - r + 1)(q - 1)$ & $r - 1$ & $-1$ & $0$ & $-\varepsilon_{1}^{j}(\pi_{0})$ & $0$ \\
$Z_{k}$ & $(q + r + 1)(q - 1)$ & $-r - 1$ & $-1$ & $0$ & $0$ & $-\varepsilon_{2}^{k}(\pi_{0})$ \\
$W_{l}$ & $r(q - 1)/2$ & $-r/2$ & $\pm r\sqrt{-1}/2$ & $0$ & $1$ & $-1$ \\
\end{tabular}
\end{table}
We now apply Lemma \ref{11} to construct four supercharacter
theories given the statement of this theorem. By considering
$m(Sz(q))$ and $M(Sz(q))$, it can be proved that $s(Sz(q)) \geq 6$.
\end{proof}
\noindent{\bf Acknowledgement.} The authors are indebted to
professors Marston Conder, Geoffrey R. Robinson and Jeremy
Rickard for some critical discussion through Group Pub Forum and
MathOverFlow on Theorem 3.3. The research of the authors are
partially supported by the University of Kashan under grant no
572760/1.
|
\section{Introduction}
This papers addresses the mathematical connection between two classical models of phase transition phenomena
describing different stages of the growth of clusters (or polymers, or aggregates). The first one is the
Becker-D\"oring model (BD), first introduced in \cite{BD}, that describes the earlier stages of cluster growth,
at a small scale. Cluster of particles may increase or decrease their size one-by-one, capturing (aggregation process)
or shedding (fragmentation process) one particle, according to the set of chemical reactions
\begin{equation*
\displaystyle C_1 + C_i \displaystyle \xrightleftharpoons \displaystyle C_{i+1} \, \quad i\geq 1\,,
\end{equation*}
where $C_i$ stands for the clusters consisting of $i$ particles, $C_1$ being the single {\it free} particle. In
its mean-field version, the BD model is an infinite set of ordinary differential equations for the time
evolution of each concentrations (numbers per unit of volume) of clusters made of $i$ particles. In this works we focus
on a dimensionless BD model that involves a small parameter $\varepsilon>0$. The standard scaling procedure is detailed in
Appendix \ref{annex:adimensionalixation}. We denote by $c^\varepsilon_i(t)$ the concentration at time $t\geq 0$ of
clusters consisting of $i\geq 2$ particles and $u^\varepsilon$ for the concentration of {\it free} particles (clusters of
size $1$), where we make explicit the dependence on $\varepsilon>0$. The dimensionless system reads
\begin{equation}\label{sys:BD_rescaled}
\begin{array}{rclr}
\displaystyle \frac{d}{dt} u^\varepsilon & = & \displaystyle - \varepsilon J_{1}^\varepsilon - \varepsilon \sum_{i \geq {1}} J_i^\varepsilon \,, & t\geq 0\,,\\[0.8em]
\displaystyle \frac{d}{dt} c_{i}^\varepsilon & = & \displaystyle \frac{1}{\varepsilon} \Big{[} J_{i-1}^\varepsilon - J_i^\varepsilon \Big{]}\,, & i\geq 2
\,, \ t\geq 0\,, \\[0.8em]
\end{array}
\end{equation}
where the fluxes are defined by
\begin{equation}\label{sys:BD_rescaled_flux}
J_{1}^\varepsilon = \alpha^\varepsilon (u^\varepsilon)^2- \varepsilon^{\eta} \beta^\varepsilon c_{2}^\varepsilon\,,\ \text{ and } \ J_{i}^\varepsilon =
a_i^\varepsilon u^\varepsilon c_i^\varepsilon - b_{i+1}^\varepsilon c_{i+1}^\varepsilon\,, \quad i\geq 2\,.
\end{equation}
Here, coefficients $a_i^\varepsilon$ and $b_{i+1}^\varepsilon$, for $i\geq 2$, denote respectively the rate of aggregation and
fragmentation ($\varepsilon$-dependent), while $\alpha^\varepsilon$ and $\beta^\varepsilon$ denote respectively the
first rate of aggregation ($i=1$) and the first rate of fragmentation ($i=2$). Finally, $\eta$ is an exponent that
stands for the strength of the first fragmentation rate, on which strongly depends our results (see also Section
\ref{sec:disc} for discussions). Observe that such model (at least formally) preserves the total number of particles
(no source nor sink), that is
\begin{equation}\label{eq:mass}
u^\varepsilon(t) + \sum_{i\geq 2} \varepsilon^2 i c_i^\varepsilon(t) = m^\varepsilon \,, \quad \forall t\geq 0\,.
\end{equation}
The constant $m^\varepsilon$ is entirely determined by the initial conditions at $t=0$ given by $u^{\rm in, \varepsilon}$ and
$(c^{\rm in,\varepsilon}_i)_{i\geq 2}$, non-negatives and $\varepsilon$-dependent.
For theoretical studies on the well-posedness and long-time behaviour of the deterministic Becker-D\"oring model (with
$\varepsilon=1$), we
refer the interested reader to \cite{Penrose2001,Wattis2008,Laurencot2002a} among many others.
\smallskip
The second model of phase transition is the Lifshitz-Slyozov model (LS) introduced in \cite{LS}. It classically
describes the late phase of cluster growth, at a ``macroscopic scale''. The LS model consists in a partial
differential equation (of nonlinear transport type) for the time evolution of the size distribution function $f(t,x)$
of clusters of (continuous) size $x>0$ at time $t\geq 0$, together with an equation stating the conservation of matter,
\begin{equation}\label{eq:LS}
\begin{array}{ll}
\displaystyle \frac{\partial f}{\partial t} + \frac{\partial [ (a(x)u(t) - b(x))f(t,x) ] }{\partial x}=0 \,, & \quad t\geq0\,, \ x>0\,, \\[1.5em]
\displaystyle u(t) + \int_0^\infty xf(t,x) = m \, , & \quad t\geq 0\,,
\end{array}
\end{equation}
where $a$ and $b$ are functions of the size, respectively for the aggregation and fragmentation rates. The constant $m$
plays the same role as in the BD model. Various authors studied this equation when the flux point outward at $x=0$
(when small clusters tends to fragment), namely if a condition like $a(0)u(t) - b(0) < 0$ holds, see
\cite{Laurenccot2002,Collet2002a,Niethammer2008} among other for theoretical studies and technical assumptions. Indeed,
in that case, uniqueness of weak solution to the limit system \eqref{eq:LS} holds. But, recent applications in biology
have raised this problem to include {\it nucleation} in this equation (small clusters tends to aggregate), for instance
in \cite{Prigent2012,Helal2013,Banks2014}. These cases consider fluxes that point inward at $x=0$, and it lacks a
\textit{boundary condition} to \eqref{eq:LS} to be well-defined. Remark, some boundary conditions was conjectured {\it
e.g.} in \cite{Collet2002,COLLET2004,Prigent2012} but never rigorously proved.
\medskip
In this works we aim to recover a solution of the LS equation and construct proper boundary condition,
departing from the BD equation \eqref{sys:BD_rescaled} as the parameter $\varepsilon$ goes to $0$. This connection has been
proved in \cite{Collet2002,Laurencot2002a} for the classical case of outgoing characteristic. The authors represent the
dynamics of the BD model by a density function on a continuous size space. Accordingly, the
size of the clusters are represented by a continuous variable $x>0$, and we let, for all $\varepsilon>0$,
\begin{equation} \label{eq:def_feps}
f^\varepsilon(t,x) = \sum_{i\geq 2} c_i^\varepsilon(t)\indic{\Lambda_i^\varepsilon}(x)\,, \quad x\geq 0\,, \ t\geq 0\,,
\end{equation}
where for each $i\geq 2$, $\Lambda_i^\varepsilon=[(i-1/2)\varepsilon, (i+1/2)\varepsilon)$. We denote for the remainder $f^{\rm in,\,
\varepsilon}:= f^\varepsilon(0,x)$. Hence, each cluster of (discrete) size
initially $i\geq2$ is seen as a cluster of size roughly $i\varepsilon\in\mathbb R_+$. This scaling consists in an acceleration of
the fluxes (by $1/\varepsilon$) in Eq.~\eqref{sys:BD_rescaled} so that it can reach an (asymptotically) {\it infinite} size $i = x/\varepsilon$ in
finite time. Then, an appropriate scaling of the initial conditions, with a large excess of particles, together with
the rate functions entails that $\{f^\varepsilon\}$ converges to a solution of the LS model, Eq. \eqref{eq:LS}. Here we use the
same strategy to construct solutions to \eqref{eq:LS} and we derive appropriate flux conditions at $x=0$ when
the reaction rates behave near $0$ as a power-law, that is
\begin{equation*}
a(x) \sim_{0^+} \overline{a}x^{r_a}\, \text{ and }\ b(x) \sim_{0^+} \overline{b}x^{r_b}\,,
\end{equation*}
with $\overline{a}$ and $\overline{b}$ positives, and the exponents $0\leq r_a <1$, $r_a \leq r_b$ which corresponds to
entrant characteristic. Note, if $r_a=r_b$ we suppose moreover that $u(t)>\overline{b}/\overline{a}$.
\medskip
\begin{remk}
Another scaling approach considers the large time behavior of the Becker-D\"oring model,
and relates the dynamics of large clusters to solutions of various version of Lifshitz-Slyozov equations. It is the
so-called theory of Ostwald ripening, see \cite{Penrose1997,Niethammer2005,Velazquez1998}.
\end{remk}
\medskip
We emphasize that the novelty of our work resides in the rigorous derivation of a boundary condition at $x=0$ for the LS model, Eq. \eqref{eq:LS}, which is needed in the case of entrant characteristic. Thanks to new estimates on the BD model (Proposition \ref{prop:bound_laplace}), we identify the
limit of quantities related to the (finite size) $c_i^\varepsilon$'s by a quasi steady state approximation. From this, we were able to
found various possible boundary conditions depending on different scaling hypotheses on the first fragmentation rate,
{\it i.e.} according to the value of $\eta$ in \eqref{sys:BD_rescaled_flux}, with respect to $r_a$ and $r_b$. Namely,
we found three distinct cases for {\it slow} de-nucleation rate ($\eta>r_a$) in Theorem \ref{thm:LS_slow}, {\it compensated} one ($\eta=r_a$)
in Theorem \ref{thm:LS_compensated} and {\it fast} one ($\eta<r_a$) in Theorem \ref{thm:LS_fast}. We
obtained these main results for measure-valued solution to the LS equation, in Section \ref{sec:results}. But in
Section \ref{sec:density}, we improve this result to obtain density solution when $a$ and $b$ are exact power law. Let us give an example of our result to illustrate it.
\medskip
\noindent \textsc{Illustrating result.} \textit{Assume, for all $x\geq 0$, $a(x) = \overline a x^{r_a}$ and
$b(x)=\overline b
x^{r_b}$ with $r_a < r_b$ and $\eta =
r_b$. We found the limit of $\{f^\varepsilon\}$ is a solution of Eq. \eqref{eq:LS}, with the boundary value given by, for
all
$t\geq 0$ where $u(t)>0$,
\[ \lim_{x\to 0^+} (a(x)u(t)-b(x))f(t,x) = \alpha u(t)^2\,, \]
where $\alpha$ is the limit of $\alpha^\varepsilon$ in \eqref{sys:BD_rescaled_flux}. In other terms we recover the behavior of $f$ near $x=0$ with the free
particles concentration through the limit
\[ \lim_{x\to 0^+} x^{r_a} f(t,x) = \frac{\alpha}{\overline a} u(t)\,. \]
}
\medskip
\noindent {\it Organization of the paper.} In the next Section \ref{sec:scaling} we introduce the main assumptions
together with some properties of the BD model. Then, in Section \ref{sec:results} we state our main result on
measure-valued solution to LS with boundary term. To do so we improved previous compactness arguments on the re-scaled
density \eqref{eq:def_feps}, so that the boundary term can be taken into account in Section \ref{sec:compact}. It is
achieved thanks to a new estimate on the growth of the ``small'' sized clusters (point-wise estimates of the density
approximation, see Proposition \ref{prop:bound_laplace}). The identification of the boundary term in Section
\ref{sec:identif} follows from a rigorous quasi-steady-state approximation of the small-sized clusters, in analogy with
slow-fast systems, and allow proving the main theorems. Finally, we extend some results to a convergence in density,
see Section \ref{sec:density}. We conclude by a discussion and further directions in Section \ref{sec:disc}.
\medskip
\noindent \textit{Notations.} For any $U\subseteq \mathbb R$, we denote by $\mathcal C(U)$, respectively $\mathcal C_c(U)$ and $\mathcal C_b(U)$,
the space of continuous function on $U$, respectively with compact support on $U$, and bounded on $U$. We
denote by $\mathcal M_f(U)$ the set of non-negative and finite
regular Borel measures on $U$. We will
use the classical $weak-*$ convergence (sometimes called {\it vague}) on $\mathcal M_f(U)$, {\it i.e.}~ the topology given by
pointwise convergence for test functions $\varphi\in \mathcal C_c(U)$, {\it i.e.} for $\{\nu^\varepsilon\}$ and $\nu$ in $\mathcal M_f(U)$,
we say $\nu^\varepsilon$ converge to $\nu$ in $\mathcal M_f(U)$ (in the $weak-*$ topology) if and only if for all $\varphi\in
\mathcal C_c(U)$
$$ \displaystyle \int_0^\infty \varphi(x) \nu^\varepsilon(dx) \to \int_0^\infty \varphi(x) \nu(dx) \,. $$
\section{Preliminaries and Assumptions}\label{sec:scaling}
In this section we recall some known results on the BD system together with assumptions for the main
results of this paper. First of all, we refer the reader to Theorem 2.1 in \cite{Laurencot2002a} for existence and
uniqueness of (non-negative) global solution to \eqref{sys:BD_rescaled} satisfying the balance of mass
\eqref{eq:mass} at fixed $\varepsilon>0$. Well-posedness follows from growth conditions on the kinetic rates, namely we
assume
\medskip
\begin{hyp}\label{assumption_1}
The rates $\alpha^\varepsilon$, $\beta^\varepsilon$, $(a_i^\varepsilon)_{i\geq2}$ and $(b_i^\varepsilon)_{i\geq3}$ are positives and, for each
$\varepsilon>0$, there exists a constant $K(\varepsilon)>0$ such that
\begin{equation*}
\begin{array}{ll}
\displaystyle a_{i+1}^\varepsilon - a_i^\varepsilon \leq K(\varepsilon)\,, & i\geq 2\,, \\[0.8em]
\displaystyle b_i^\varepsilon - b_{i+1}^\varepsilon \leq K(\varepsilon)\,, & i\geq 3\,.
\end{array}
\end{equation*}
\end{hyp}
From now, for each $\varepsilon>0$, we assume $u^\varepsilon$ and $(c_i^\varepsilon)_{i\geq 2}$ are non-negatives and define a solution to
\eqref{sys:BD_rescaled}, that belongs (each) to $\mathcal C([0,+\infty))$.
\medskip
\noindent We construct aggregation and fragmentation rates as functions on $\mathbb R_+$ (similarly to $f^\varepsilon$), namely, for
each $\varepsilon>0$ we define, for all $x$ in $\mathbb R_+$,
\begin{equation*}
a^\varepsilon(x) := \displaystyle \sum_{i \geq 2} a_i^\varepsilon \indic{\Lambda_i^\varepsilon}(x)\,, \text{ and } b^\varepsilon(x) := \displaystyle \sum_{i
\geq 3} b_i^\varepsilon \indic{\Lambda_i^\varepsilon}(x)\,.
\end{equation*}
Now, we are able to derive a weak equation on the density approximation $f^\varepsilon$, for each $\varepsilon>0$, in which we will
pass to the limit to recover weak solutions to Eq. \eqref{eq:LS}.
\medskip
\begin{propo}\label{prop:weak_eps}
Under Assumption \ref{assumption_1}, let $\{f^\varepsilon\}$ constructed by
Eq.~\eqref{eq:def_feps}. For each $\varepsilon>0$, and all $\varphi\in W^{1,\infty}_{loc}(\mathbb R_+)$ such that
$\partial_x\varphi \in L^\infty(\mathbb R_+)$, we have, for all $t\geq 0$,
\begin{multline}\label{eq:weak_form_eps}
\int_0^{+\infty} f^\varepsilon(t,x) \varphi(x)\, dx \\ =\int_0^{+\infty} f^{in,\varepsilon}(x) \varphi(x)\, dx + \int_0^t [
\alpha^\varepsilon u^\varepsilon(s)^2 - \beta^\varepsilon \varepsilon^{\eta} c_2^\varepsilon(s)] \left(\frac{1}{\varepsilon} \int_{\Lambda_2^\varepsilon}
\varphi(x)\, dx\right) ds \\
+ \int_0^t \int_0^{+\infty} \left[ a^\varepsilon(x) u^\varepsilon(s) f^\varepsilon(s,x) \Delta_{\varepsilon}\varphi(x) - b^\varepsilon(x)
f^\varepsilon(s,x) \Delta_{-\varepsilon}\varphi(x) \right]\, dx\, ds\,,
\end{multline}
where $\Delta_h \varphi(x) = (\varphi(x+h) - \varphi(x))/h$, for $h\in\mathbb R$, and
\begin{equation} \label{eq:masscons_eps}
u^\varepsilon(t) + \int_0^\infty x f^\varepsilon(t,x) \, dx = m^\varepsilon.
\end{equation}
\end{propo}
This result follows from \cite[Lemma 4.1]{Laurencot2002a}, which allows taking $\varphi(x)=x$ in the equation. In the next
assumption we assume standard hypotheses on the convergence of the rate functions and their sub-linear control, see
also \cite{Collet2002,Laurencot2002a}.
\medskip
\begin{hyp}
{\normalfont Convergence of the rate functions.}
let $\alpha$ and $\beta$ be two positive numbers, and let $a$ and $b$ be two non-negative continuous functions on
$[0,+\infty)$ that are positive on $x\in(0,+\infty)$. Then, as $\varepsilon\to 0 $, we suppose that
\begin{align}
& \{ \alpha^\varepsilon \} \text{ converges towards } \alpha\,. \tag{H1} \label{H1}\\[0.5em]
& \{ \beta^\varepsilon \} \text{ converges towards } \beta\,. \label{H2} \tag{H2} \\[0.5em]
& \{a^\varepsilon(\ldotp)\} \text{ converges uniformly on any compact set of } [0,+\infty) \text{ towards } a(\ldotp) \text{
and } \nonumber \\
& \hspace{3em}\exists K_a>0 \text{ s.t. } a^\varepsilon(x) \leq K_a (1+x), \ \forall x\in \mathbb R_+ \text{ and } \forall
\varepsilon>0\,. \label{H3} \tag{H3} \\[0.5em]
& \{b^\varepsilon(\ldotp)\} \text{ converges uniformly on any compact set of } [0,+\infty) \text{ towards } b(\ldotp) \text{
and } \nonumber \\
& \hspace{3em}\exists K_b>0 \text{ s.t. } b^\varepsilon(x) \leq K_b (1+x), \ \forall x\in \mathbb R_+ \text{ and } \forall
\varepsilon>0\,. \label{H4} \tag{H4}
\end{align}
\end{hyp}
We recall a discussion on the scaling of the coefficients is differed to Section \ref{sec:disc}. The next assumption details the behaviour of the rate functions around $0$. This is the essential assumption which
allow us to identify the limit of $\varepsilon^\eta c_2^\varepsilon$ in the second integral in the right hand side of
\eqref{eq:weak_form_eps}.
\medskip
\begin{hyp} \label{hyp:coef_BD} {\normalfont Behavior of the rate functions near $0$.}
%
We suppose there exist $r_a \in [0,1)$, $r_b\geq r_a$, $\overline{a}>0$, $\overline{b}>0$ such that
%
\begin{equation} \label{H5} \tag{H5}
\begin{array}{lc|cl}
a(x) \sim_{0^+} \overline{a}x^{r_a}\,, & & & b(x) \sim_{0^+} \overline{b}x^{r_b}\,, \\[0.8em]
a^{\varepsilon}(\varepsilon i) = a(\varepsilon i) + o((\varepsilon i)^{r_a})\,, & & &
b^{\varepsilon}(\varepsilon i) = b(\varepsilon i) + o((\varepsilon
i)^{r_b})\,,
\end{array}
\end{equation}
%
where $o$ is the Landau notation, \textit{i.e.} $o(x)/x\to 0$ as $x \to 0$.
\end{hyp}
\medskip
\noindent Note, if $0 \leq r_b < r_a$ or $r_a\geq 1$, the kinetic rates $a$ and $b$ are related to outgoing
characteristics for which the theory already exists, see \cite{Laurencot2002a,Collet2002}. Finally, we assume some
control on the initial conditions. For convenience, we define the quantity
\begin{equation} \label{def:rho}
\rho := \lim_{x\to 0^+} \frac{b(x)}{a(x)} = \lim_{x\to 0^+} \frac{\overline b}{\overline a} x^{r_b-r_a} \in
[0,+\infty)\,.
\end{equation}
It determines whether the characteristic at $x=0$ is ongoing or outgoing, according to whether $u(t)$ is greater or less than
$\rho$ in \eqref{eq:mass}.
Then, we introduce a set of functions which shall play a key role. We denote by $\mathcal U$ the set of
non-negative convex functions $\Phi$ belonging to $\mathcal C^1([0,+\infty))$ and piecewise
$\mathcal C^{2}([0,+\infty))$ such that $\Phi(0)=0$, $\Phi'$ is concave, $\Phi'(0)\geq 0$, and
\[ \lim_{x\to +\infty} \frac{\Phi(x)}{x} = + \infty\,. \]
Note that $\Phi$ is increasing. These functions have remarkable properties when conjugate to the structure of the
Becker-D\"oring system and provide important estimates, see for instance \cite{Laurenccot2002}.
\medskip
\begin{hyp} \label{hyp:initial} {\normalfont Initial conditions.}
We assume there exists $u^{\rm in}>\rho$ and a non-negative measure $\mu^{\rm in}\in\mathcal M_f([0,+\infty))$ such
that $u^{\rm in,\, \varepsilon}$ converges to $u^{\rm in}$ in $\mathbb R_+$ and $\{f^{\rm in,\, \varepsilon} \}$ converges
to $\mu^{\rm in}$, in the $weak-*$ topology of $\mathcal M_f([0,+\infty))$. Moreover, we assume there exists $\Phi\in\mathcal U$ such
that
%
\begin{equation} \label{H6}\tag{H6}
\displaystyle \sup_{\varepsilon>0} \int_0^\infty \Phi(x) f^{\rm in,\, \varepsilon}(x)dx < +\infty \,.
\end{equation}
%
En particular, we can define
%
\[m := u^{\rm in} + \int_0^\infty x \mu^{\rm in}(dx)\,.\]
%
Moreover, we suppose that for all $z\in(0,1)$,
%
\begin{equation} \label{H7} \tag{H7}
\sup_{\varepsilon>0} \, \sum_{i\geq 2} \varepsilon^{r_a} c_i^{\rm in,\,\varepsilon} e^{-iz} < +\infty\,. \\
\end{equation}
%
\end{hyp}
\medskip
\begin{remk}
$m$ is well-defined since $weak-*$ convergence plus the extra-moment in \eqref{H6} give the limit
\[ \int_0^\infty x f^{\rm in,\, \varepsilon}(dx)\to \int_0^\infty x \mu^{\rm in}(dx)\,. \]
See for instance \cite[Proof of Theorem 2.3]{Collet2002}.
\end{remk}
\medskip
\begin{remk}
In fact, we could obtain freely this $\Phi$ assuming a {\it stronger} weak convergence (against $(1+x)\varphi(x)$ for
$\varphi$ bounded and continuous). See for instance \cite{Chauhoan1977} for the construction of such a $\Phi$.
\end{remk}
\medskip
\begin{remk}
We highlight that condition \eqref{H7} is not restrictive. For example, consider
$f^{\rm in}(x)=x^{-r}$ on $(0,1)$ and $0$ elsewhere, with $r\leq r_a$. Then, consider $c_i^{\rm in,\, \varepsilon} =
(i\varepsilon)^{-r}$ for $i\leq 1/\varepsilon$, and $0$
elsewhere. We have that $\{f^{\rm in, \, \varepsilon}\}$ trivially converges to $f^{\rm in}$ in the sense of \eqref{H6} and
it satisfies \eqref{H7}. Note that we do not necessarily require the initial condition is composed of `` very large''
clusters (of size $i \gg 1/\varepsilon$).
\end{remk}
\section{Main results}\label{sec:results}
For the remainder of the paper, we always assume that $\{f^\varepsilon\}$ is constructed by \eqref{eq:def_feps}, that
$\{u^\varepsilon\}$ is given by the balance \eqref{eq:masscons_eps}, and Assumption \ref{assumption_1} to Assumption
\ref{hyp:initial} hold true. The next definition extends the notion of a solution to the LS model, Eq. \eqref{eq:LS},
with a general boundary condition, or {\it nucleation rate}.
\medskip
\begin{defn}{\normalfont N-solution.}
Let $T>0$, a function $N \in L^\infty_{loc}(\mathbb R_+)$ called nucleation rate, $u^{\rm in}>\rho$, a measure $\mu^{\rm
in}\in\mathcal M_f([0,+\infty))$, and a measure-valued function $\mu\in L^\infty([0,T];\mathcal M_f([0,+\infty))$. We say that $\mu$
is a $N$-solution of the LS equation (in measure) on $[0,T]$ with mass $m$, when:
\smallskip
\noindent i) There exists a non-negative $u\in\mathcal C([0,T])$, such that $u(0)=u^{\rm in}$,
\begin{equation*}
\inf_{t\in[0,T]} u(t)>\rho\,, \text{ and } \ \forall t\in[0,T]\,,\ u(t) + \int_0^\infty x \mu_t(dx) = m\,.
\end{equation*}
\smallskip
\noindent ii) For all $\varphi \in \mathcal C_c^1([0,T)\times[0,+\infty))$ and $t\in[0,T]$
\begin{multline}\label{eq:weak_LS_fast2}
\int_{0}^T \int_{0}^{\infty} \big[ \partial_t \varphi(t,x) + (a(x)u(t)-b(x))\partial_x \varphi(t,x)\big]\mu(t,dx)\, dt \\
+ \int_0^\infty \varphi(0,x)\mu^{\rm in}(dx) + \int_0^T \varphi(s,0) N(u(s)) \,ds= 0 \,,
\end{multline}
\end{defn}
We now state our main results. The first theorem, when $\eta>r_a$, corresponds to the case where the
first fragmentation rate is too slow and does not contribute to the boundary value. Thus the nucleation rate is proportional to the
number of encounter of free particles, namely $u(t)^2$ at time $t$.
\medskip
\begin{thm} \label{thm:LS_slow}{\normalfont The {\it slow} de-nucleation case.}
Assume $\eta > r_a$ and let a sequence $\{\varepsilon_n\}$ converging to $0$. There exists $T>0$, a sub-sequence
$\{\varepsilon_{n'}\}$ of $\{\varepsilon_n\}$, and $\mu$ a $N$-solution of LS
with mass $m$, such that
\[f^{\varepsilon_{n'}} \xrightharpoonup[n'\to+\infty]{} \mu\]
in $\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty))$, and, for all $u\geq 0$,
\[ N(u) = \alpha u^2\,. \]
\end{thm}
\begin{remk}
The space $\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty)))$ has to be understand as measure-valued
function that are continuous in time for the $weak-*$ topology on $\mathcal M_f([0,+\infty))$, {\it i.e.} for $\{\nu_t\} \in
\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty)))$, we have, for all $t\in[0,T]$ and $\varphi\in\mathcal C_c([0,+\infty))$,
\[t\mapsto \int_0^\infty \varphi(x)\nu_t(dx)\]
is continuous.
\end{remk}
\medskip
\noindent The second theorem holds in the limit case when $\eta=r_a$, {\it i.e.}~ the first fragmentation rate has the same order of
magnitude than the aggregation rate ($i\geq2$). Compared to the first case, the nucleation rate is balanced by a function varying between $0$
and $1$.
\medskip
\begin{thm} \label{thm:LS_compensated} {\normalfont The {\it compensated} de-nucleation case.}
Assume $\eta =r_a$ and let a sequence $\{\varepsilon_n\}$ converging to $0$. There exists $T>0$, a sub-sequence
$\{\varepsilon_{n'}\}$ of $\{\varepsilon_n\}$, and $\mu$ a $N$-solution of LS
with mass $m$, such that
\[f^{\varepsilon_{n'}} \xrightharpoonup[n'\to+\infty]{} \mu\]
in $\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty))$, and, for all $u\geq 0$,
\begin{equation*}
N(u) = \begin{cases}
\displaystyle \alpha u^2 \frac{ u}{ u+\beta/(\bar a 2^{\eta})}\,, & if \ \eta=r_a <r_b\,, \\[0.8em]
\displaystyle \alpha u^2 \frac{\overline{a}u-\overline{b}}{\overline{a}u-\overline{b}+\beta/2^\eta}\,, & if \
\eta=r_a=r_b\,,
\end{cases}
\end{equation*}
\end{thm}
\begin{remk}
In the pure aggregation case, with $\beta^\varepsilon=b_i^\varepsilon=0$, then $b=0$ and $\beta=\overline b=0$. Our results in
Theorem \ref{thm:LS_slow} and Theorem \ref{thm:LS_compensated} are consistent and remain true.
\end{remk}
\medskip
\noindent Finally, the last theorem considers the case of a fast de-nucleation rate so that the flux at
the boundary vanished, and the solution can reveal fast oscillation near $x=0$.
\medskip
\begin{thm} \label{thm:LS_fast} {\normalfont The {\it fast} de-nucleation rate.}
Assume $\eta <r_a$ and let a sequence $\{\varepsilon_n\}$ converging to $0$. There exists $T>0$, a sub-sequence
$\{\varepsilon_{n'}\}$ of $\{\varepsilon_n\}$, and $\mu$ a $N$-solution of LS
with mass $m$, such that
\[f^{\varepsilon_{n'}} \xrightharpoonup[n'\to+\infty]{} \mu\]
in $weak-*-L^\infty(0,T;\mathcal M_f([0,+\infty))$, and, for all $u\geq 0$,
\[N(u)=0\,.\]
\end{thm}
\begin{remk}
In this case were not able to prove equicontinuity of the density approximation in $\mathcal M_f([0,+\infty))$. For this
case, in fact, it is true for $\mathcal M_f((0,+\infty))$ (open in $x=0$). Also, we use the $weak-*$ topology on
$L^\infty(0,T;\mathcal M_f([0,+\infty))$ which is the topology of the point-wise convergence against test functions in
$L^1(0,T;\mathcal C_c([0,+\infty))$.
\end{remk}
\medskip
\begin{remk}
These limit theorems provide local in time existence and could be extended to a maximal time interval $[0,T)$ where
$T =\sup \{ \tau \, :\, \inf_{t\in[0,\tau]} u(t) > \rho \}$. Also, uniqueness is not investigate here, but
and appropriate result would entails convergence of the whole sequence without extraction.
\end{remk}
\section{The compactness estimates}\label{sec:compact}
In this section we provide the main estimates to obtain sufficient compactness arguments to pass to the limit in
\eqref{eq:weak_form_eps}-\eqref{eq:masscons_eps}. Remark for further estimations, under \eqref{H1} and \eqref{H2},
there exists a positive $K_{\alpha,\beta}$ such that, for all $\varepsilon>0$,
\begin{equation}\label{bound_a1_b2}
\alpha^{\varepsilon}, \beta^{\varepsilon}, \alpha, \beta \in (0,K_{\alpha,\beta}]\,,
\end{equation}
and \eqref{H3}-\eqref{H4} imply the limit functions also satisfy
\begin{align}\label{linear_bound}
a(x) \leq K_a (1+x)\text{ and } b(x) \leq K_b (1+x), \ \forall x\in [0,+\infty).
\end{align}
We fix these constants for the remainder.
\subsection{Uniform bound for the density approximation}
The first lemma gives basic estimates. In particular, it constructs the compact set of $\mathcal M_f([0,+\infty))$ in which the sequence of
solutions remains.
\medskip
\begin{lem}\label{lem:u1xphi1}
For all $T>0$,
\begin{align}
& \sup_{\varepsilon>0} \, \sup_{t\in[0,T]} \, \int_0^{+\infty} (1+x+\Phi(x))f^\varepsilon(t,x)\, dx < +\infty\,,
\label{eq:phi_1} \\
& \sup_{\varepsilon>0} \, \sup_{t\in[0,T]} \, u^\varepsilon(t) < +\infty\,, \vphantom{\int_0^t } \label{bound_u} \\
& \sup_{\varepsilon>0} \, \int_0^T \varepsilon^\eta c_2^\varepsilon(t) \, dt < +\infty\,. \label{bound_c_2_measure}
\end{align}
\end{lem}
\begin{remk}
Similar estimates can be found in \cite{Laurencot2002a} for a different scaling. For sake of completeness we recall the proof below. Note that estimate \eqref{bound_c_2_measure}, although trivial, seems to have not been reported elsewhere, and will be important for the next.
\end{remk}
\medskip
\begin{proof}
By Assumption \ref{hyp:initial}, the convergence of $\{f^{\rm in,\,\varepsilon}\}$ implies that the sequence lies in
a $weak-*$ compact set of $\mathcal M_f([0+\infty))$, and with \eqref{H6} we have
%
\begin{equation}
\sup_{\varepsilon>0} \int_{\mathbb R_+}f^{in,\varepsilon}(x)(1+x+\Phi(x)) dx < +\infty\,. \label{eq:cond_phi1}
\end{equation}
Let us start now with estimate \eqref{bound_u}. By the mass conservation relationship \eqref{eq:masscons_eps},
$u^\varepsilon(t) \leq m^\varepsilon$, for any $t\geq0$, and thanks to Assumption \ref{hyp:initial}, $(m^\varepsilon)$ converges as $\varepsilon\to 0$, thus it is
bounded by a constant $K_m>0$. Then estimate \eqref{bound_u} directly follows. Similarly, we obtain
\[\sup_{\varepsilon>0} \sup_{t\in[0,T]} \int_0^{+\infty} xf^\varepsilon(t,x)\, dx < +\infty\,.\]
%
Then, taking $\varphi = \indic{}$ in Eq. \eqref{eq:weak_form_eps}, it immediately yields by re-arranging the non-positive
term
%
\begin{equation*}
0\leq \int_0^{+\infty} f^\varepsilon(t,x) \, dx + \int_0^t \beta^\varepsilon \varepsilon^\eta c_2^\varepsilon(s) \, ds \leq
\int_0^{+\infty} f^{in,\varepsilon}(x)\, dx + \int_0^t \alpha^\varepsilon u^\varepsilon(s)^2 \, ds.
\end{equation*}
Using the bounds \eqref{bound_a1_b2}, \eqref{bound_u} and \eqref{eq:cond_phi1}, we obtain the inequality
\eqref{bound_c_2_measure} together with the first part of estimate \eqref{eq:phi_1}.
\noindent Finally, we put $\varphi=\Phi$ in \eqref{eq:weak_form_eps}. Remark that the derivative $\Phi'$ is not uniformly
bounded, thus we cannot use Proposition \ref{prop:weak_eps} straightforwardly. However, with a classical regularizing
argument, one can show that the next computations hold true {\it a posteriori}, see for instance \cite[proof of Lemma
4.2]{Laurencot2002a}. We remark that
\[ 0\leq \Delta_{\varepsilon}\Phi(x) \leq \Phi'(x+\varepsilon), \quad -\Delta_{-\varepsilon}\Phi(x) \leq -\Phi'(x) \leq 0. \]
Moreover, $\Phi'(x+\varepsilon) \leq \Phi'(x) + \varepsilon \Phi''(0)$. Thus, dropping the non-positive term, using
\eqref{H3} and again that $u^\varepsilon(t) \leq K_m$,
\begin{multline} \label{eq:weak_form_phi1}
\int_0^{+\infty} f^\varepsilon(t,x) \Phi(x)\, dx \leq \int_0^{+\infty} f^{in,\varepsilon}(x) \Phi(x)\, dx + \int_0^t
\alpha^\varepsilon u^\varepsilon(s)^2 \left(\frac{1}{\varepsilon}\int_{\Lambda_2^\varepsilon} \Phi(x)dx\right) \, ds \\
+ K_m K_a\int_0^t \int_0^{+\infty} (1+x) f^\varepsilon(s,x) (\Phi'(x) + \varepsilon \Phi_{1,r}''(0)) \, dx\, ds,
\end{multline}
Let $\delta>0$. Note that $x\Phi'(x)\leq2\Phi(x)$ (see \cite[Lemma A.1]{Laurencot2001}), we get
\begin{multline*}
\int_0^{+\infty} (1+x) f^\varepsilon(s,x) \Phi'(x) \, dx \leq \int_0^{\delta} f^\varepsilon(s,x) \Phi'(x) \, dx + \left(
\frac 1 \delta +1\right) \int_0^{+\infty} x f^\varepsilon(s,x) \Phi'(x) \, dx \\
\leq \big(\sup_{(0,\delta)} \Phi' \big)\int_0^{\infty} f^\varepsilon(s,x) \, dx + 2\left( \frac 1 \delta +1\right)
\int_0^{+\infty} f^\varepsilon(s,x) \Phi(x) \, dx.
\end{multline*}
We introduce this last estimation into Eq.~\eqref{eq:weak_form_phi1} and we conclude using the previous bounds and
Gr\"onwall lemma.
\end{proof}
\subsection{Pointwise estimations on the density}
We turn now to the main estimate of this paper. Indeed, to obtain equicontinuity for the density $\{f^\varepsilon\}$ (in a
measure space), and then identify the boundary condition, we need to control the behaviour of the small-sized clusters,
particularly because of the term $\varepsilon^{\eta } c_2^{\varepsilon}$ in the weak equation \eqref{eq:weak_form_eps}. Remark that
we already have a weak bound (in time) given by Eq.~\eqref{bound_c_2_measure}. In the next Proposition
\ref{prop:bound_laplace} we improve this estimate by a control on exponential moments which depends on $\rho$ (defined
in Eq. \eqref{def:rho}). Moments are classical tools and
play a key role in the well-posedness of BD theory. More recently, exponential moments were also used
\cite{Jabin2003,Canizo2013} to study long time behavior of BD solutions. Here, let us define the
discrete Laplace transform
\begin{equation} \label{def_laplace}
F^\varepsilon(t,z) = \sum_{j\geq 2} \varepsilon^{r_a} c_j^\varepsilon(t) e^{-jz}\,, \ z\in(0,1)\,.
\end{equation}
From the re-scaled system \eqref{sys:BD_rescaled}, the sequence $(d_i^\varepsilon)_{i \geq 2}$ defined by $d_i^\varepsilon
:=\varepsilon^{r_a} c_i^\varepsilon$, for $i\geq 2$, satisfies, for each $\varepsilon>0$, the following equations
\begin{equation} \label{eq:BD_system_H}
\varepsilon^{1-r_a}\frac{d}{dt}d_i^\varepsilon(t) = H_{i-1}^\varepsilon - H_i^\varepsilon\,, \quad i\geq 2\,,
\end{equation}
where the fluxes are
\begin{equation*}
\displaystyle H_1^\varepsilon =\alpha^\varepsilon u^\varepsilon(t)^2 - \beta^\varepsilon \varepsilon^{\eta-r_a} d_{2}^\varepsilon(t)\,, \text{ and }\displaystyle H_i^\varepsilon =
\overline a_i^\varepsilon u^\varepsilon(t) d_i^\varepsilon(t) - \varepsilon^{r_b-r_a} \overline b_{i+1}^\varepsilon
d_{i+1}^\varepsilon(t)\,, \quad i\geq 2\,,
\end{equation*}
with, for all $i\geq2$,
\begin{equation*}
\overline a_i^\varepsilon = \frac{a_i^\varepsilon}{\varepsilon^{r_a}}\,, \quad \text{and }\quad\overline b_{i+1}^\varepsilon =
\frac{b_{i+1}^\varepsilon}{\varepsilon^{r_b}}\,.
\end{equation*}
Note that, under hypotheses \eqref{H3}, \eqref{H4} and \eqref{H5}, the kinetic coefficients $\alpha^\varepsilon$,
$\beta^\varepsilon$ and $\overline a_i^\varepsilon$, $\overline b_i^\varepsilon$, $i\geq 2$, are convergent sequences toward a positive
value (resp. $\alpha$, $\beta$, $\overline a i^{r_a}$, $\overline{b} i^{r_b}$).
\medskip
\begin{propo}\label{prop:bound_laplace}
Let $T>0$ and $\{\varepsilon_n\}$ a sequence converging to $0$ such that $\{u^{\varepsilon_n}\}$ converges toward $u$
uniformly on $[0,T]$, with $\inf_{t\in [0,T]} u(t) > \rho$. There exists $z_0>0$ such that for all
$z\in(0,z_0)$
%
\begin{equation}\label{bound_ci_caseII}
\sup_{n\geq0} \, \sup_{t\in[0,T]}F^{\varepsilon_{n}}(t,z)<\infty\,.
\end{equation}
%
In particular, for all $r \geq r_a$ and $i\geq 2$, we have
%
\begin{equation}\label{eq:bound_ci_r}
\sup_{n\geq 0} \, \sup_{t\in[0,T]} \, \varepsilon^r c_i^{\varepsilon_{n}}(t) < +\infty \,.
\end{equation}
\end{propo}
\medskip
\begin{remk}
It is immediate from estimate \eqref{eq:bound_ci_r} that we can obtain compactness in $w-*-L^\infty(0,T)$ for any
finite size cluster $\varepsilon^{r}c_i^\varepsilon$, which will be used to prove theorem \ref{thm:LS_slow}
and \ref{thm:LS_compensated}.
\end{remk}
\medskip
\begin{remk}
We cannot prove that the pseudo-moment $F^\varepsilon$ is propagated along limit solution for which $u(t) \leq \rho$ on some
time interval. This is important in the case $r_a=r_b$ since $\rho>0$ and $u$ can eventually cross this threshold
(which is, up to our knowledge, an open problem).
\end{remk}
\medskip
\begin{proof}
Let $z>0$ and $\varepsilon>0$. First, note the discrete Laplace transform define in Eq. \eqref{def_laplace} is finite for each
$\varepsilon>0$ and for all $t$ in
$[0,T]$, since
\[ F^\varepsilon(t,x) \leq \varepsilon^{r_a-1} \int_0^\infty x f^\varepsilon(t,x)\, dx\,.\]
Let us derive $F^\varepsilon$ with respect to time (derivation under the sum is justified by similar bound). For all
$t\in[0,T]$, we get
\begin{equation*}
\varepsilon^{1-r_a} \partial_t F^\varepsilon(t,z) = \sum_{j\geq 2} e^{-jz} \big[ H_{j-1}^\varepsilon - H_j^\varepsilon\big] = e^{-2z}
H_1^\varepsilon - (1-e^{-z}) \sum_{j\geq 2} e^{-jz} H_j^\varepsilon\,.
\end{equation*}
Thus, developing the fluxes we get
\begin{multline*}
\varepsilon^{1-r_a} \partial_t F^\varepsilon(t,z) = e^{-2z} H_1^\varepsilon - (1-e^{-z}) \sum_{j\geq 2} e^{-jz} \overline a_j^\varepsilon
u^\varepsilon(t) d_j^\varepsilon(t)\\
+ (1-e^{-z}) \sum_{j\geq 2} e^{-jz} \varepsilon^{r_b-r_a} \overline b_{j+1}^\varepsilon d_{j+1}^\varepsilon(t) \,.
\end{multline*}
Then, re-indexing the second sum on the right hand side, we obtain
\begin{multline} \label{dt_Feps_intermediate}
\varepsilon^{1-r_a} \partial_t F^\varepsilon(t,z) = e^{-2z} H_1^\varepsilon - (1-e^{-z}) e^{-2z} \overline a_2^\varepsilon u^\varepsilon(t)
d_2^\varepsilon(t) \\
- (1-e^{-z}) \sum_{j\geq 3} e^{-jz} \overline a_j^\varepsilon \left[ u^\varepsilon(t) - \frac{b_{j}^\varepsilon}{a_j^\varepsilon } e^z \right]
d_j^\varepsilon(t) \,.
\end{multline}
Since $\inf_{t\in [0,T]} u(t) > \rho$, we can find a constant $c$ such that $\inf_{t\in [0,T]} u(t)\geq c>\rho$. Then,
by uniform convergence of $\{u^{\varepsilon_n}\}$, there exists $\tilde \varepsilon>0$ small enough, such that for all $n$ with
$\varepsilon_n\leq \tilde \varepsilon$, $\inf_{t\in[0,T]}\, u^{\varepsilon_n}(t) \geq c>\rho$. Also, we can choose $\delta>0$ and $z_0>0$, both small
enough, such that for all $t\in[0,T]$ we have $c > \rho e^{z_0} + 2\delta$. Then, there exists $N>0$ such that, for all $z\in(0,z_0)$
\[ \inf_{n\geq N}\, \inf_{t\in[0,T]}\, u^{\varepsilon_n}(t) > \rho e^{z} + 2\delta \,.\]
Then, by hypothesis \eqref{H5}, for all $3\leq j \leq 1/\sqrt{\varepsilon}$,
\[\frac{b_{j}^\varepsilon}{a_j^\varepsilon } = \frac{\overline b}{\overline a} \frac{(\varepsilon j)^{r_b}+o((\varepsilon j)^{r_b})}{(\varepsilon
j)^{r_a}+o((\varepsilon j)^{r_a})} = \frac{\overline b}{\overline a} (\varepsilon j)^{r_b-r_a} (1+o(1)) \,, \]
so that, we have, for $N$ large enough,
\[\sup_{n\geq N}\, \sup_{j\in[3,\ldots,\lfloor1/\sqrt{\varepsilon_n}\rfloor-1]} \left\lvert
\rho-\frac{b_{j}^{\varepsilon_n}}{a_j^{\varepsilon_n}
}\right\rvert < \delta e^{-z}\,. \]
The latter gives a uniform control in $j$ for the relatively ``small'' sizes $j\leq 1/\sqrt{\varepsilon}$. We separate the
sum in Eq.~ \eqref{dt_Feps_intermediate} in two parts, the small-sized clusters for $j\in
(3,\ldots,\lfloor1/\sqrt{\varepsilon_n}\rfloor-1)$ in one side, for which (for $n\geq N$)
\[ u^{\varepsilon_n}(t) - \frac{b_{j}^{\varepsilon_n}}{a_j^{\varepsilon_n}} e^z = u^{\varepsilon_n}(t) - \rho e^{z} + e^z
\left(\rho - \frac{b_{j}^{\varepsilon_n}}{a_j^{\varepsilon_n}} \right) \geq 2\delta - \delta = \delta\,,\]
and the large-sized clusters in another side. Hence, for all $t\in[0,T]$,
\begin{multline} \label{dt_Feps_intermediate_1}
\sum_{j\geq 3} e^{-jz} \overline a_j^{\varepsilon_n} \left[ u^{\varepsilon_n}(t) -
\frac{b_{j}^{\varepsilon_n}}{a_j^{\varepsilon_n} } e^z \right]
d_j^{\varepsilon_n}(t) \\
\geq \delta \sum_{j=3}^{\lfloor1/\sqrt{\varepsilon_n}\rfloor-1} e^{-jz} \overline a_j^{\varepsilon_n} d_j^{\varepsilon_n}(t)
+ \sum_{j\geq
\lfloor 1/\sqrt{\varepsilon_n}\rfloor} e^{-jz} \overline a_j^{\varepsilon_n} \left[ u^{\varepsilon_n}(t) -
\frac{b_{j}^{\varepsilon_n}}{a_j^{\varepsilon_n} } e^z
\right] d_j^{\varepsilon_n}(t)\,.
\end{multline}
Using hypothesis \eqref{H5}, there exists $x_0$ such that for all $x\in(0,x_0)$, $a(x)/x^{r_a}>3\overline a/4$. Thus,
there exists $\tilde N$ such that for all $n\geq \tilde N$ and for all $2\leq i\leq 1/\sqrt{\varepsilon_n}$ we have $\varepsilon_n i
\leq \sqrt{\varepsilon_n}< x_0$ and $a(\varepsilon i)/(\varepsilon_ni)^{r_a}\geq 3\overline a/4$. Still with hypothesis \eqref{H5}, we can choose $\tilde N$ such that for all $n>\tilde N$, and for all $2\leq i\leq 1/\sqrt{\varepsilon_{n}}$, we have $a^\varepsilon(\varepsilon i)/(\varepsilon_ni)^{r_a}\geq \overline a/2$. Hence, from the rank $\tilde N$, there exists a
constant $\widetilde K_a>0$ such that for all $n\geq \tilde N$ and for all $2\leq j \leq 1/\sqrt{\varepsilon_n}$, we have
\[\overline a_j^{\varepsilon_n} = \frac{a_j^{\varepsilon_n}}{\varepsilon_n^{r_a}} \geq \widetilde K_a := \frac 1 2 \overline
a\, 2^{r_a}\,.\]
Accordingly, the rest of the proof has to be understood for $n$ large enough. Using the equation on $H_1^\varepsilon$ and plugging inequality \eqref{dt_Feps_intermediate_1} into
Eq.~\eqref{dt_Feps_intermediate} we obtain
\begin{multline*}
\varepsilon_n^{1-r_a} \partial_t F^{\varepsilon_n}(t,z) \leq e^{-2z} [\alpha^{\varepsilon_n} u^{\varepsilon_n}(t)^2
-\varepsilon_n^{\eta-r_a} \beta^{\varepsilon_n }
d_{2}^{\varepsilon_n}(t)]
\\ - (1-e^{-z}) e^{-2z} [ \overline a_2^{\varepsilon_n} u^{\varepsilon_n}(t) - \delta \tilde K_a] d_2^{\varepsilon_n}(t) -
(1-e^{-z}) \delta
\widetilde K_a \sum_{j= 2}^{\lfloor1/\sqrt{\varepsilon_n}\rfloor-1} e^{-jz} d_j^{\varepsilon_n}(t) \\
- (1-e^{-z}) \sum_{j\geq \lfloor1/\sqrt{\varepsilon_n}\rfloor} e^{-jz} \overline a_j^{\varepsilon_n} \left[
u^{\varepsilon_n}(t) -
\frac{b_{j}^{\varepsilon_n}}{a_j^{\varepsilon_n}} e^z \right] d_j^{\varepsilon_n}(t)\,.\\
\end{multline*}
Remark that $ \overline a_2^{\varepsilon_n} u^{\varepsilon_n}(t) - \delta \widetilde K_a \geq \tilde K_a( \rho e^z + 2 \delta
-\delta) \geq \widetilde K_a \rho \geq 0$. Using the moment estimates \eqref{bound_u} and hypothesis \eqref{H3}, we
have $\sup_{t\in[0,T]} \alpha^\varepsilon u^\varepsilon(t)^2 \leq K_0$ uniformly in $\varepsilon>0$. Thus, dropping also some negative
terms, we have
\begin{multline*}
{\varepsilon_n}^{1-r_a} \partial_t F^{\varepsilon_n}(t,z) \leq K_0 e^{-2z} - (1-e^{-z}) \delta \widetilde K_a \sum_{j=
2}^{\lfloor1/\sqrt{{\varepsilon_n}}\rfloor-1} e^{-jz} d_j^{\varepsilon_n}(t) \\
+ (1-e^{-z}) \sum_{j\geq \lfloor1/\sqrt{{\varepsilon_n}}\rfloor} e^{-jz}\frac{b_{j}^{\varepsilon_n}}{{\varepsilon_n}^{r_a}}
d_j^{\varepsilon_n}(t) \,.
\end{multline*}
Now using that
\[\sum_{j= 2}^{ \lfloor1/\sqrt{{\varepsilon_n}}\rfloor-1} e^{-jz} d_j^{\varepsilon_n}(t) =F^{\varepsilon_n}(t,z)-\sum_{j \geq
\lfloor1/\sqrt{{\varepsilon_n}}\rfloor } e^{-jz} d_j^{\varepsilon_n}(t)\,,\]
we obtain
\begin{multline*}
{\varepsilon_n}^{1-r_a} \partial_t F^{\varepsilon_n}(t,z) \leq K_0 e^{-2z} - (1-e^{-z}) \delta \widetilde K_a
F^{\varepsilon_n}(t,z) \\
+ (1-e^{-z}) \delta \sum_{j \geq \lfloor1/\sqrt{{\varepsilon_n}}\rfloor } e^{-jz} \widetilde K_a d_j^{\varepsilon_n}(t)
+ (1-e^{-z})e^z \sum_{j\geq \lfloor1/\sqrt{{\varepsilon_n}}\rfloor} e^{-jz}\frac{b_{j}^{\varepsilon_n}}{{\varepsilon_n}^{r_a}}
d_j^{\varepsilon_n}(t) \,.
\end{multline*}
At this step, we recall that by definition we have, for all $j\geq 2$, $d_j^\varepsilon/\varepsilon^{r_a} = c_j^\varepsilon$, and
$\widetilde K_a< a_j^\varepsilon/\varepsilon^{r_a}$, so that, with $K = \max(\delta,e^z)$,
\begin{multline*}
{\varepsilon_n}^{1-r_a} \partial_t F^{\varepsilon_n}(t,z) \leq K_0 e^{-2z} - (1-e^{-z}) \delta \widetilde K_a
F^{\varepsilon_n}(t,z) \\
+ (1-e^{-z}) K \sum_{j\geq \lfloor1/\sqrt{{\varepsilon_n}}\rfloor} e^{-jz} (a_j^{\varepsilon_n} + b_{j}^{\varepsilon_n})
c_j^{\varepsilon_n}(t) \,.
\end{multline*}
Finally, by hypotheses \eqref{H3}-\eqref{H4}, we have, for all $j\geq \lfloor1/\sqrt{\varepsilon}\rfloor$ (and $\varepsilon$ small enough)
\[ e^{-jz} (a_j^\varepsilon + b_{j}^\varepsilon) \leq (K_a+K_b)(1+\varepsilon j) e^{-jz} \leq (K_a+K_b) \varepsilon\,.\]
Thus,
\begin{multline*}
{\varepsilon_n}^{1-r_a} \partial_t F^{\varepsilon_n}(t,z) \leq K_0 e^{-2z} - (1-e^{-z}) \delta \widetilde K_a
F^{\varepsilon_n}(t,z) \\
+ (1-e^{-z}) K (K_a+K_b) \int_0^{+\infty} f^{\varepsilon_n}(t,x)dx \,.
\end{multline*}
By the moment estimates \eqref{eq:phi_1}, there exists $\tilde K$ independent from $\varepsilon_n$ such that
\begin{equation}\label{ineq:F}
{\varepsilon_n}^{1-r_a} \partial_t F^{\varepsilon_n}(t,z) \leq - (1-e^{-z}) \delta \widetilde K_a F^{\varepsilon_n}(t,z) +
\tilde K \,.
\end{equation}
We can conclude that
\begin{equation*}
F^{\varepsilon_n}(t,z) \leq F^{\varepsilon_n}(0,z) + \frac{ \widetilde K}{\delta \widetilde K_a (1-e^{-z})}\,,
\end{equation*}
and the result \eqref{bound_ci_caseII} follows thanks to the initial bound on $ F^\varepsilon(0,z)$ given by hypothesis
\eqref{H7}. Note that \eqref{eq:bound_ci_r} directly follows from the previous bound \eqref{bound_ci_caseII} and the
definition of the discrete Laplace transform \eqref{def_laplace}.
\end{proof}
\medskip
\begin{remk}
Estimate \eqref{ineq:F} on $F^\varepsilon$ can be easily generalized for any exponent $r$ instead of $r_a$.
Writing
$G^{^\varepsilon}(t,z)= \sum_{j\geq 2} \varepsilon^{r} c_j^\varepsilon(t) e^{-jz}$, and following the same steps, we find
\begin{equation*}
\varepsilon^{1-r_a} \partial_t G^\varepsilon(t,z) \leq - (1-e^{-z}) \delta \widetilde K_a G^\varepsilon(t,z) + \varepsilon^{r-r_a}\tilde
K \,.
\end{equation*}
Thus, this inequality provides valuable information if $r\geq r_a$.
\end{remk}
\subsection{Equicontinuity lemmas}
We now turn to the equicontinuity of the density approximation, as a measure valued time-dependent function. The new
result here is to provide equicontinuity in a measure space on $[0,\infty)$ (lemma \ref{lem:equicontinuity}) . The
first lemma is independent on $\eta$ and similar to \cite{Laurencot2002a,Collet2002}.
\medskip
\begin{lem}\label{equicontinuity_u}
Let $T>0$. The family $\{u^{\varepsilon}\}$ is equicontinuous on $[0,T]$.
\end{lem}
\medskip
\begin{proof}
Let us fix $T>0$. From the mass conservation \eqref{eq:masscons_eps}, we can deduce that the equicontinuity of
$\{u^{\varepsilon}\}$ directly follows from the one of the sequence $\{\int_0^{+\infty}xf^{\varepsilon} (\cdot,x)\,dx\}$. Thus, we
focus on this latter. We have, from Eq.~\eqref{eq:weak_form_eps} with $\varphi(x)=x$, for all $t\in [0,T-h]$ and $s \in
[0,h]$ with $0 < h < T$,
\begin{multline} \label{eq:equicontinuity_u1}
\left| \int_0^{+\infty} [f^\varepsilon(t+s,x)-f^\varepsilon(t,x) ] x\, dx \right|\leq \left(\frac 1 \varepsilon
\int_{\Lambda_2^\varepsilon} x\, dx \right) \int_t^{t+s} ( \alpha^\varepsilon u^\varepsilon(\sigma)^2 + \beta^\varepsilon\veps^{\eta}
c_2^\varepsilon(\sigma)) \,d\sigma \\
+ \int_t^{t+s} \int_0^{+\infty} \vert a^\varepsilon(x) u^\varepsilon(\sigma) f^\varepsilon(\sigma,x) - b^\varepsilon(x)
f^\varepsilon(\sigma,x)\vert\, dx\, d\sigma\,.
\end{multline}
The first term in the r.h.s of \eqref{eq:equicontinuity_u1} can be bounded, thanks to the bound \eqref{bound_a1_b2}, by
\begin{multline*}
\left(\frac 1 \varepsilon \int_{\Lambda_2^\varepsilon} x\, dx \right)\int_t^{t+s} ( \alpha^\varepsilon u^\varepsilon(\sigma)^2 +
\beta^\varepsilon\veps^{\eta} c_2^\varepsilon(\sigma)) \,d\sigma \\
\leq 2K_{\alpha,\beta} \left[\varepsilon\sup_{t\in[0,T]} u^\varepsilon(t)^2 + \sup_{t\in[0,T]} \varepsilon^{\eta+1} c_2^{\varepsilon}(t)
\right] h\,.
\end{multline*}
Then, since $\eta\geq 0$ and remarking that $\varepsilon c_2^\varepsilon$ is obviously bounded by the $L^1$ norm of $f^\varepsilon$, we
can use the moment estimates in Eqs.~\eqref{eq:phi_1} and \eqref{bound_u}, so that for $\varepsilon$ sufficiently small,
there exists $K$ independent of $t$ and $\varepsilon$ such that
\begin{equation}\label{equicontinuity_u_first}
\left(\frac 1 \varepsilon \int_{\Lambda_2^\varepsilon} x\, dx \right)\int_t^{t+s} ( \alpha^\varepsilon u^\varepsilon(\sigma)^2 + \beta^\varepsilon
\varepsilon^{\eta} c_2^\varepsilon(\sigma)) \,d\sigma \leq K h\,.
\end{equation}
Let us now focus on the second term on the right-hand side of Eq.~\eqref{eq:equicontinuity_u1}. Using hypotheses
\eqref{H3}-\eqref{H4} and the moment estimates in Eq.~\eqref{eq:phi_1}, we get
\begin{multline*} \int_t^{t+s} \int_0^{+\infty} \vert a^\varepsilon(x) u^\varepsilon(\sigma) f^\varepsilon(\sigma,x) - b^\varepsilon(x)
f^\varepsilon(\sigma,x)\vert\, dx\, d\sigma\\
\leq \left(K_a\sup_{\varepsilon>0}\sup_{t\in[0,T]} u^\varepsilon(t) + K_b \right) \int_t^{t+s} \int_0^{+\infty} f^\varepsilon(\sigma,x)
(1+x)\, dxd\sigma\,.
\end{multline*}
Hence, there is a constant $K>0$ such that
\begin{multline} \label{equicontinuity_u_second} \int_t^{t+s} \int_0^{+\infty} \vert a^\varepsilon(x) u^\varepsilon(\sigma)
f^\varepsilon(\sigma,x) - b^\varepsilon(x) f^\varepsilon(\sigma,x)\vert\, dx\, d\sigma\\
\leq h K \left( \sup_{\varepsilon>0} \, \sup_{t\in[0,T]} \int_0^{+\infty} (1+x)f^\varepsilon(t,x)\, dx\right)\,.
\end{multline}
Combining both inequalities \eqref{equicontinuity_u_first}-\eqref{equicontinuity_u_second}, it follows that for all
$\delta>0$, for all $h\in(0,T)$ ,
\begin{equation*}
\sup_{\varepsilon>0} \sup_{t \in[0,T-h]} \sup_{s \in[0,h]} \left| \int_0^{+\infty} [f^\varepsilon(t+s,x)-f^\varepsilon(t,x) ]
x\, dx \right|\leq \delta\,,
\end{equation*}
which gives the equicontinuity property for $\{u^\varepsilon\}$.
\end{proof}
\medskip
The next lemma improves the equicontinuity of $\{f^\varepsilon\}$ around $x=0$.
\medskip
\begin{lem}\label{lem:equicontinuity}
Assume $\eta \geq r_a$ and $T>0$. Let $\{\varepsilon_n\}$ a sequence converging to $0$ such that $\{u^{\varepsilon_n}\}$
converges toward $u$ uniformly on $[0,T]$ satisfying $\inf_{t\in [0,T]} u(t) > \rho$. Then the sequence $\{f^{\varepsilon_{n}}\}$
is equicontinuous in $\mathcal M_f([0,+\infty))$.
\end{lem}
\medskip
\begin{proof} Let us fix $T>0$. Let $h \geq 0 \in(0,T)$, $t\in[0,T-h]$ and
$s\in[0,h]$ we have, for all $\psi\in\mathcal C^\infty_c([0,+\infty))$ and $\varepsilon>0$
\begin{multline} \label{eq:equicontinuity_1}
\left| \int_0^{+\infty} [f^\varepsilon(t+s,x)-f^\varepsilon(t,x)] \psi(x)\, dx \right| \\
\leq \int_t^{t+s} (\alpha^\varepsilon u^\varepsilon(\sigma)^2 + \beta^\varepsilon \varepsilon^{\eta} c_2^\varepsilon(\sigma))
\left(\frac{1}{\varepsilon}\int_{\Lambda_2^\varepsilon} \left\vert \psi(x) \right\vert \, dx\right) \,d\sigma \\
+ \int_t^{t+s} \int_0^{+\infty} \vert a^\varepsilon(x) u^\varepsilon(\sigma) f^\varepsilon(\sigma,x) \Delta_{\varepsilon}\psi(x) -
b^\varepsilon(x) f^\varepsilon(\sigma,x) \Delta_{-\varepsilon}\psi(x) \vert\, dx\, d\sigma\,.
\end{multline}
The first integral in the right-hand side can be bounded as follows
\begin{multline*}
\int_t^{t+s} ( \alpha^\varepsilon u^\varepsilon(\sigma)^2 + \beta^\varepsilon \varepsilon^{\eta}c_2^\varepsilon(\sigma)) \left(\frac{1}{\varepsilon}
\int_{\Lambda_2^\varepsilon} \left\vert \psi(x) \right\vert\, dx\right) \,d\sigma \\
\leq h \| \psi \|_{\infty} \sup_{t\in[0,T]} \left[ \alpha^\varepsilon u^\varepsilon(t)^2 + \beta^\varepsilon \varepsilon^{\eta}
c_2^{\varepsilon}(t) \right]\,.
\end{multline*}
Using Eqs.~\eqref{bound_a1_b2}, \eqref{bound_u} and by Proposition \ref{prop:bound_laplace}, Eq.~\eqref{eq:bound_ci_r},
both terms in the supremum are uniformly bounded in time and along $\{\varepsilon_{n}\}$. Hence, there exists $K$ independent of $T$ and
$\varepsilon$ such that, for all $t\leq T-h$, $s\in[0,h]$,
\begin{align}\label{bound_first}
\int_t^{t+s} ( \alpha^{\varepsilon_n} u^{\varepsilon_n}(\sigma)^2 + \beta^{\varepsilon_n} {\varepsilon_n}^{\eta} c_2^{\varepsilon_n}(\sigma))
\left(\frac{1}{{\varepsilon_n}}
\int_{\Lambda_2^{\varepsilon_n}} \left\vert \psi(x) \right\vert \, dx\right)\,d\sigma \leq K \| \psi|_{\infty} h\,.
\end{align}
We now focus on the second integral in the right hand side of \eqref{eq:equicontinuity_1}. Using upper bounds
\eqref{linear_bound} and \eqref{bound_u}, we can find a constant $K$ such that for all $\varepsilon >0$
\begin{multline*}
\int_t^{t+s} \int_0^{+\infty} \vert a^\varepsilon(x) u^\varepsilon(\sigma) f^\varepsilon(\sigma,x) \Delta_{\varepsilon}\psi(x) -
b^\varepsilon(x) f^\varepsilon(\sigma,x) \Delta_{-\varepsilon}\psi(x) \vert\, dx\, d\sigma\\
\leq K \| \psi'\|_{\infty} \int_t^{t+s} \int_0^{+\infty} f^\varepsilon(\sigma,x) (1+x)\, dxd\sigma\,.
\end{multline*}
Combining this last inequality with the moment estimate \eqref{eq:phi_1} and the inequality \eqref{bound_first}, there
exists a constant $K$ (not depending on $\psi$, $h$ and $\varepsilon$), such that for all $h \in(0,T)$, $t\in[0,T-h]$,
$s\in[0,h]$, $\psi\in\mathcal C_c^\infty([0,+\infty))$ and $n\geq0$
\begin{equation*} \label{eq:equi_h}
\left| \int_0^{+\infty} [f^{\varepsilon_n}(t+s,x)-f^{\varepsilon_n}(t,x)
]\psi(x) \,
dx \right|\leq K (\| \psi \|_{\infty} +\| \psi'\|_{\infty} ) h\,.
\end{equation*}
Let $\{\varphi_i\}_{i\geq 1} \subset\mathcal C_c^\infty([0,+\infty))$ a dense subset of $\mathcal C_c([0,+\infty))$ for the uniform
norm. The metric $d$ defined by, for all $\mu$ and $\nu$ belonging to $\mathcal M_f([0,+\infty))$,
\[d(\mu,\nu) = \sum_{i} \frac{2^{-i}}{\|\varphi_i\|_\infty+\|\varphi_i'\|_\infty} \left|\int_0^\infty \varphi_i \mu -
\int_0^\infty \varphi_i
\nu\right|\,,\]
is equivalent to the $weak-*$ topology (on bounded subset), see for instance similar construction in
\cite[Theorem III.25]{Brezis}. Thus, for all $h \geq 0 \in(0,T)$, we have
\begin{equation*}
\sup_{t\in[0,T-h]}\, \sup_{s\in[0,h]}\, \sup_{n\geq0} \ d(f^{\varepsilon_n}(t+s),f^{\varepsilon_n}(t)) \leq K h \,.
\end{equation*}
This concludes the proof.
\end{proof}
\subsection{Compactness and limit}
Here we give some technical lemmas which prepare the proof of the main results.
\medskip
\begin{lem} \label{lem:weak_form_tx_eps}
For all $T>0$ and all $\varphi\in \mathcal C^1_c([0,T)\times[0,+\infty))$, we have, for all $\varepsilon>0$,
\begin{multline}\label{eq:weak_form_tx_eps}
\int_0^T \int_0^{+\infty} \left[ \partial_t \varphi(t,x) + a^\varepsilon(x) u^\varepsilon(s) \Delta_{\varepsilon}\varphi(t,x) -
b^\varepsilon(x) \Delta_{-\varepsilon}\varphi(t,x) \right] f^\varepsilon(t,x)\, dx\, dt \\
+ \int_0^{+\infty} f^{in,\varepsilon}(x) \varphi(0,x)\, dx +
\int_0^T [
\alpha^\varepsilon u^\varepsilon(t)^2 - \beta^\varepsilon \varepsilon^{\eta} c_2^\varepsilon(t)] \left(\frac{1}{\varepsilon} \int_{\Lambda_2^\varepsilon}
\varphi(t,x)\, dx\right) dt = 0
\end{multline}
where $\Delta_h \varphi(t,x) = (\varphi(t,x+h) - \varphi(t,x))/h$, for $h\in\mathbb R$, and
\begin{equation} \label{eq:massss}
u^\varepsilon(t) + \int_0^\infty x f^\varepsilon(t,x) \, dx = m^\varepsilon.
\end{equation}
\end{lem}
\medskip
\begin{proof}
The proof remains on multiplying each equation of the Becker-D\"oring system \ref{sys:BD_rescaled} by $\varphi_i =
\int_{\Lambda_i^\varepsilon} \varphi(t,x)\, dx$ for $\varphi\in \mathcal C^1_c([0,T)\times[0,+\infty))$ and using the definition of
$f^\varepsilon$ in Eq.~\eqref{eq:def_feps}. It is similar to Proposition \ref{prop:weak_eps}.
\end{proof}
\medskip
\begin{lem}\label{lem:compactness_weak_star}
Let $T>0$. The family $\{f^{\varepsilon}\}$ is relatively $weak-*$ compact in $L^\infty(0,T;\mathcal M_f([0,+\infty))$. If $\mu$ is
an accumulation point of $\{f^{\varepsilon}\}$, then there exists a sequence $\{\varepsilon_n\}$ converging to $0$ and a
non-negative function
$u\in C([0,T])$ such that $u^{\varepsilon_n}$ converges to $u$ uniformly on $[0,T]$, with $u(0)=u^{\rm in}$ and
\[ u(t) + \int_0^\infty x \mu_t(dx) = m.\]
Moreover, for all $\varphi\in \mathcal C^1_c([0,T)\times[0,+\infty))$
\begin{multline*}
\int_0^T \int_0^{+\infty} \left[ \partial_t \varphi(t,x) + a^{\varepsilon_n}(x) u^{\varepsilon_n}(s) \Delta_{{\varepsilon_n}}\varphi(t,x)
- b^{\varepsilon_n}(x) \Delta_{-{\varepsilon_n}}\varphi(t,x) \right] f^{\varepsilon_n}(t,x)\, dx\, dt \\
\to \int_0^T \int_0^{+\infty} \left[ \partial_t \varphi(t,x) + (a(x) u(s) -
b(x)) \partial_x\varphi(t,x) \right] \mu_t(dx)\, dx
\end{multline*}
\begin{equation*}
\int_0^T \alpha^{\varepsilon_n} u^{\varepsilon_n}(t)^2 \left(\frac{1}{{\varepsilon_n}} \int_{\Lambda_2^{\varepsilon_n}}
\varphi(t,x)\, dx\right) dt \to \int_0^T \alpha u(t)^2 \varphi(t,0) dt \,,
\end{equation*}
and
\begin{equation*}
\int_0^{+\infty} \varphi(0,x) f^{in,{\varepsilon_n}}(x) \, dx \to \int_0^{+\infty} \varphi(0,x)\, \mu^{\rm in}(dx)\,
\end{equation*}
as $n\to +\infty$.
\end{lem}
\medskip
\begin{proof}
First, remark the bound against $1$ in \eqref{eq:phi_1} yields to the relative compactness in
$L^\infty(0,T;\mathcal M_f([0,+\infty))$. Let $\mu$ an accumulation point. By Lemma \ref{equicontinuity_u} and bound
\eqref{bound_u} with Arzel\'a-Ascoli Theorem, entails there exists a sequence $\{\varepsilon_n\}$ and $u\in\mathcal C([0,T])$ such
that $u^{\varepsilon_n}$ converge to $u$ uniformly on $[0,T]$ and $\{f^{\varepsilon_n}\}$ to $\mu$. It remains to note that for
any $\psi^\varepsilon \in \mathcal C_c([0,T)\times[0,+\infty))$ which converge uniformly to some $\psi$, we have
\[ \int_0^T\int_0^\infty \psi^{\varepsilon_n}(t,x) f^{\varepsilon_n}(t,x)\, dx dt \to \int_0^T\int_0^\infty \psi(t,x) \mu_t(dx)
dt\,,\]
as $n\to \infty$, to obtain the desired limit, see also \cite{Laurencot2002a,Collet2002}. In fact, using similar arguments as
in Lemma \ref{lem:equicontinuity} with function in $\mathcal C_c((0,+\infty))$, we can obtain equicontinuity in
$\mathcal M_f((0,+\infty))$ for the $weak-*$ topology (open in $x=0$). Such result has been obtained for instance in
\cite{Collet2002}. Thus, we could improve the compactness of $f^{\varepsilon_n}$ in $\mathcal C([0,T];\mathcal M_f((0,+\infty))$ by Arzel\'a-Ascoli
Theorem. Finally we obtain Eq. \eqref{eq:massss}, using the bound \eqref{eq:phi_1} with $\Phi$, and after
regularization of the identity function, we have for
all $t\in[0,T]$
\[\int_0^\infty x f^{\varepsilon_n}(t,x)\, dx \to \int_0^\infty x \mu_t( dx)\,. \]
See \cite[Proof of Theorem 2.3]{Collet2002} for details.
\end{proof}
\medskip
\begin{lem} \label{lem:compactness_weak_star_2}
Assume $\eta\geq r_a$ and let a sequence $\{\varepsilon_n\}$ converging to $0$. There exists $T>0$ and
a sub-sequence $\{\varepsilon_{n'}\}$ of $\{\varepsilon_n\}$ such that $\{f^{\varepsilon_{n'}}\}$ is relatively compact in
$\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty))$ and $u^{\varepsilon_{n'}}$ converge to some $u$ uniformly on $[0,T]$ with
$\inf_{t\in[0,T]} u(t)>\rho$.
\end{lem}
\begin{proof}
Let $\widetilde T>0$ and $\{\varepsilon_n\}$ a sequence converging to $0$. Thanks to
Lemma \ref{equicontinuity_u} and the bound \eqref{bound_u} we apply Arzel\'a-Ascoli Theorem, and there exists
$u\in\mathcal C([0,\widetilde T])$ and a sub-sequence still denoted by $\{\varepsilon_n\}$ such that $u^\varepsilon$ converge uniformly
to $u$ on $[0,\widetilde T]$. By Assumption \ref{hyp:initial} we have $u(0)>\rho$, thus there exists $T\in
(0,\widetilde T]$ such that we have $\inf_{t\in[0,T]} u(t)>\rho$. We can apply Lemma
\ref{lem:equicontinuity} so that $\{f^{\varepsilon_n}\}$ is
equicontinuous in $\mathcal M_f([0,+\infty))$. By the bound \eqref{eq:phi_1} (against $1$), we have for each $t\in[0,T]$ that
$\{ f^{\varepsilon_n}(t) \, : \, \varepsilon>0 \}$ belongs to a $weak-*$ compact set of $\mathcal M_f([0,+\infty))$. Thus, again by
Arzel\'a-Ascoli Theorem, the sequence $\{f^{\varepsilon_n}\}$ is relatively compact in $\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty))$.
\end{proof}
\medskip
\begin{remk}
Convergence in $\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty))$ entails convergence in $L^\infty(0,T;\mathcal M_f([0,+\infty))$ for the
$weak-*$ topology.
\end{remk}
\section{Identification of the boundary term}\label{sec:identif}
This section is devoted to the proof of Theorems \ref{thm:LS_slow} to \ref{thm:LS_fast}. In view of Lemmas
\ref{lem:weak_form_tx_eps} to \ref{lem:compactness_weak_star_2} it remains to identify the limit of $\varepsilon^\eta
c_2^\varepsilon$ so that we can pass to the limit in the term
\[\int_0^T \beta^\varepsilon \varepsilon^{\eta} c_2^\varepsilon(t) \left(\frac{1}{\varepsilon} \int_{\Lambda_2^\varepsilon}
\varphi(t,x)\, dx\right) dt \]
arising in \eqref{eq:weak_form_tx_eps}.
\medskip
We separate the following in 3 subsections corresponding to the 3 theorems. Thanks to Proposition
\ref{prop:bound_laplace}, the compactness of the term $\varepsilon^\eta c_2^\varepsilon$ has been already obtained in
$w-*-L^\infty(0,T)$ for the first two case, that are $\eta > r_a$ and $\eta = r_a$, and in $\mathcal M_f([0,T])$ by Eq.
\eqref{bound_c_2_measure} for $\eta<r_a$. The identification of the limit relies on arguments similar to the
Fenichel-Tikhonov theory on singularly perturbed dynamical systems \cite{kuehn_multiple_2015}. Multiplying the
re-scaled BD equations \eqref{sys:BD_rescaled} by $\varepsilon$, at least formally, we have for all $t>0$ and $i\geq 2$,
\[ \lim_{\varepsilon \to 0} \varepsilon \frac d {dt} c_i^\varepsilon = \lim_{\varepsilon \to 0} (J_{i-1}^\varepsilon(t) - J_i^\varepsilon(t) ) = 0\,.\]
Hence, at each time $t>0$, the underlying BD model for the discrete sizes $i\geq 2$ has to reach instantaneously the
equilibrium of the BD model with a constant monomer concentration $u=u(t)$. Such version of the BD model has been well
studied in \cite{Penrose1989,WATTIS}.
\subsection{Proof of Theorem \ref{thm:LS_slow} -- The slow de-nucleation case}
Let $\{\varepsilon_n\}$ a sequence converging to $0$. By Lemma \ref{lem:compactness_weak_star_2}, there
exists $T>0$, a sub-sequence, still denoted by $\{\varepsilon_n\}$ for simplicity, $\mu\in\mathcal C([0,T];w-*-\mathcal M_f([0,+\infty)))$ and $u\in\mathcal C([0,T])$ with
$\inf_{t\in[0,T]} u(t) >\rho$ such that $\{f^{\varepsilon_{n}}\}$ converges to $\mu$ in $\mathcal C([0,T];\mathcal M_f([0,+\infty)))$ and $u^{\varepsilon_n}$ converges to $u$ uniformly on $[0,T]$. Now, applying Proposition \ref{prop:bound_laplace}, we get
\[ \sup_{t\in[0,T]} \varepsilon_{n}^\eta c_2^{\varepsilon_{n}} (t) = \varepsilon_{n}^{\eta-r_a} \sup_{t\in[0,T]} \varepsilon^{r_a} c_2^{\varepsilon_{n}} (t) \to 0 \,,\]
as $\eta>r_a$. Thus, combining this result with Lemma \ref{lem:compactness_weak_star} we can pass to the limit in
\eqref{eq:weak_form_tx_eps} to obtain Eq.~\eqref{eq:weak_LS_fast2} with $N(u)=\alpha u^2$, and Theorem \ref{thm:LS_slow} is proved.
\subsection{Proof of Theorem \ref{thm:LS_compensated} -- The compensated nucleation case}
Let $\{\varepsilon_n\}$ a sequence converging to $0$. We proceed similarly as above with Lemma~\ref{lem:compactness_weak_star_2} and Proposition~\ref{prop:bound_laplace}. As for all $i\geq 2$, $d_i^{\varepsilon_{n}}=\varepsilon_n^{r_a}c_i^{\varepsilon_n}$ satisfies $d_i^{\varepsilon_{n}} e^{-iz}\leq F^{\varepsilon_{n}}(t,z)$, thanks to the estimate \eqref{bound_ci_caseII}, there exists $z>0$ such that
\[ \sup_{n\geq0} \, \sup_{t\in[0,T]} \, \sup_{i\geq 2} \, d_i^{\varepsilon_{n}}e^{-iz} < +\infty\,.\]
Hence, by a Cantor diagonal process, we can extract another sub-sequence, still denoted by $\{\varepsilon_n\}$, such that
for all $i\geq 2$,
\[ d_i^{\varepsilon_{n}} \rightharpoonup d_i \,, \quad w-*-L^\infty(0,T)\,,\]
and
\begin{equation}\label{eq:bound_di_ra}
0\leq \sup_{t\in[0,T]}\, \sup_{i\geq2} \, d_i(t) e^{-iz}< K_z \,.
\end{equation}
We recall, from the rescaled BD system \eqref{sys:BD_rescaled}, that the sequence $(d_i^{\varepsilon_{n}})_{i \geq 2}$
satisfies for each $n\geq 0$ Eq. \eqref{eq:BD_system_H}.
Hence, for all $\varphi\in \mathcal C^1([0,T])$,
\begin{multline}\label{eq:fast_case1}
{\varepsilon_{n}}^{1-r_a} d_{i}^{\varepsilon_{n}}(t)\varphi(t) - {\varepsilon_{n}}^{1-r_a} d_{i}^{in,{\varepsilon_{n}}} \varphi(0) -
{\varepsilon_{n}}^{1-r_a}\int_0^t
d_i^{\varepsilon_{n}}(s)\varphi'(s)\, ds \\
= \int_0^t \varphi(s) \left[ H_{i-1}^{\varepsilon_{n}}(s) - H_i^{\varepsilon_{n}}(s) \right] \, ds\,.
\end{multline}
As $r_a<1$, passing to the limit ${\varepsilon_{n}}\to 0$, the left hand-side in Eq.~\eqref{eq:fast_case1} vanishes, and,
with Assumption \ref{hyp:coef_BD} on the kinetic rates, we have, for all $\varphi\in \mathcal C^1([0,T])$,
\[ \int_0^T \varphi(t) \left[ H_{i-1}(t) - H_i(t) \right] \, ds = 0 \,,\]
where $H_1 = \alpha u(t)^2 - \beta d_2 $, and for each $i\geq 2$,
\[ H_i = \begin{cases}
\bar a i^\eta u d_i \,,& \text{if } \eta=r_a < r_b \,, \\[0.8em]
\bar a i^\eta u d_i - \bar b (i+1)^\eta d_{i+1}\,, & \text{if } \eta=r_a=r_b \,.
\end{cases}
\]
Thus, for all $i\geq 2$, we have {\it a.e.} $t\in(0,T)$ that $H_i(t) = H_1(t)$. In the sequel, we will distinguish two
cases, $r_a<r_b$ and $r_a=r_b$.
\subsubsection{The case $\eta=r_a<r_b$} In this case, $H_1 = H_2$ for {\it a.e. } $t\in(0,T)$ yields
\[ d_2(t) = \frac{\alpha u^2(t)}{\overline a 2^\eta u(t)+\beta} \,.\]
Hence, the limit $d_2$ is uniquely identified (and by recurrence, all $d_i$, $i\geq 2$, using $H_i=H_1$) as a function
of the limit $u$. Thus, combining this result with Lemma \ref{lem:compactness_weak_star} we can pass
to the limit in \eqref{eq:weak_form_tx_eps} to obtain Eq.~\eqref{eq:weak_LS_fast2} with $N(u)=\alpha u^2 \frac{ u}{ u+\beta/(\bar a 2^{\eta})}$, and the case $r_a<r_b$ in Theorem \ref{thm:LS_compensated} is proved.
\subsubsection{The case $\eta=r_a=r_b$} In this case, the limit $(d_i)_{i\geq 2}$ must satisfy $H_i\equiv H$, $i\geq
1$, for a given constant $H$. We classically (in the study of the equilibrium states of BD equations \cite{Ball1986})
define $ Q_1=1$ and for all $i\geq
2$,
\begin{equation*}
Q_i=\frac{\alpha}{\beta}\prod_{k=2}^{i-1}\frac{\overline{a}k^{r_a}}{\overline{b}(k+1)^{r_a}},\ i\geq 2\,.
\end{equation*}
The solutions that satisfy $H_i\equiv H$ for all $i\geq 1$, are given by, after some algebraic manipulation (see
\cite[lemma 1]{Penrose1989}),
\begin{equation*
d_i=Q_i u^i\Big{(}1-H\frac{1}{\alpha u^2}-H\sum_{k=2}^{i-1}\frac{1}{\overline{a}k^{r_a}Q_ku^{k+1}}\Big{)}\,,\quad
i\geq2\,.
\end{equation*}
Thus, for all $i\geq 2$,
\begin{equation*
d_i=\frac{\alpha
u^2}{\beta}\frac{2^{r_a}}{i^{r_a}}\Big{(}\frac{\overline{a}u}{\overline{b}}\Big{)}^{i-2}\left[1-\frac{H}{\alpha
u^2}\left(1+\frac{\beta}{2^{r_a}}\frac{1}{\overline{a}u-\overline{b}}\right)+\frac{H \beta}{\alpha u^2
2^{r_a}}\frac{\left(\overline{b}/(\overline{a}u)\right)^{i-2}}{\overline{a}u-\overline{b}}\right]\,.
\end{equation*}
However, for $u(t)>\rho = \overline b / \overline a$, there exists a unique $H$ such that the bound
\eqref{eq:bound_di_ra} is satisfied, given by
\begin{equation*}
H= \frac{\alpha u^2}{\left(1+\frac{\beta}{2^\eta}\frac{1}{\overline{a}u-\overline{b}}\right)}= \frac{\alpha u^2
(\overline{a}u-\overline{b})}{\overline{a}u+\frac{\beta}{2^\eta}-\overline{b}}\,.
\end{equation*}
For this value, we have {\it a.e.} $t\in[0,T]$
\begin{equation*}
d_2(t)=\frac{\alpha u(t)^2}{2^\eta(\overline{a}u-\overline{b})+\beta}=\frac{\alpha
u(t)^2}{\beta}\left[1-\frac{\overline{a}u-\overline{b}}{\overline{a}u-\overline{b}+\beta/2^\eta}\right]\,.
\end{equation*}
Hence, proceeding as before we recover the second part of Theorem \ref{thm:LS_compensated}.
\subsection{Proof of Theorem \ref{thm:LS_fast} -- The fast de-nucleation} In the case $\eta<r_a$ we have no $L^\infty$
bound over $\varepsilon^\eta c_2^\varepsilon$, and no equicontinuity property on $\{f^\varepsilon\}$ in $\mathcal M_f([0,+\infty))$.
Nevertheless, we can apply Lemma \ref{lem:compactness_weak_star}. Thus, let $\widetilde T>0$ and $\{\varepsilon_n\}$ a
sequence converging to $0$, there exists a sub-sequence of $\{\varepsilon_n\}$ (not relabeled), $\mu \in
L^\infty([0, \widetilde T];\mathcal M_f([0,+\infty)))$ and $u\in \mathcal C([0,\widetilde T])$ such that $f^{\varepsilon_n}
\rightharpoonup\mu$ in $w-*-L^\infty([0,\widetilde T];\mathcal M_f([0,+\infty)))$ and $u^{\varepsilon_n}$ converges uniformly to $u$
on $[0,\widetilde T]$. Since $u^{\rm in}>\rho$ by Assumption \ref{hyp:initial}, there exists $T\in(0,\widetilde T]$
such that $\inf_{t\in[0,T]} u(t)>\rho$. Moreover, by the bound \eqref{bound_c_2_measure} we can extract another
sub-sequence of $\{\varepsilon_n\}$ (not relabeled) such that $d_2^{\varepsilon_{n}}:=\varepsilon_{n}^\eta c_2^\varepsilon$ converges to a non-negative
finite measure $\Gamma_2$ on $[0,T]$, where the
convergence holds in $\mathcal M_f([0,T])$ endowed with the $weak-*$ topology. Also, for all $\varphi \in \mathcal C^1([0,T])$, the
equation \eqref{sys:BD_rescaled} for $i=2$ yields
\begin{multline}\label{limit_mesure_gamma2_temp}
{\varepsilon_{n}}^{1-r_a} {\varepsilon_{n}}^{r_a} c_{2}^{\varepsilon_{n}}(T)\varphi(T) - {\varepsilon_{n}}^{1-r_a} {\varepsilon_{n}}^{r_a}
c_{2}^{in,\varepsilon_n} \varphi(0) -
{\varepsilon_{n}}^{1-r_a} \int_0^T \varphi'(t) {\varepsilon_{n}}^{r_a} c_{2}^{\varepsilon_{n}}(t) dt \\
= \int_0^T \varphi(t) [ \alpha^{\varepsilon_{n}} u^{\varepsilon_{n}}(t)^2 - \beta^{\varepsilon_{n}} d_2^{\varepsilon_{n}}(t) ]dt \\ -
\int_0^T \varphi(t) [\overline
a_2^{\varepsilon_{n}} {\varepsilon_{n}}^{r_a-\eta}
u^{\varepsilon_{n}}(t) d_2^{\varepsilon_{n}}(t) - \overline b_3^{\varepsilon_{n}} {\varepsilon_{n}}^{r_b} c_3^{\varepsilon_{n}}(t)] dt\,.
\end{multline}
By Proposition \ref{prop:bound_laplace}, ${\varepsilon_{n}}^{r_a} c_2^{\varepsilon_{n}}(t)$ is uniformly bounded with respect to
both time $t\in[0,T]$ and $n$, so that the left hand side of Eq.~\eqref{limit_mesure_gamma2_temp} goes to $0$ as
${\varepsilon_{n}}\to0$. Hence, with the bound \eqref{bound_c_2_measure} and since $\eta < r_a$, we have
\begin{equation} \label{limit_mesure_gamma2}
\lim_{{\varepsilon_{n}}\to 0} \int_0^T \varphi(t) {\varepsilon_{n}}^{r_b} c_3^{\varepsilon_{n}}(t) dt = \frac{1}{\overline b_3}
\left( \int_0^T \varphi(t) \beta
\Gamma_2(dt) - \int_0^T \varphi(t) \alpha u(t)^2 dt\right)\,.
\end{equation}
Here again two cases have to be considered, $r_a<r_b$ and $r_a=r_b$.
\subsubsection{The case $r_a<r_b$} In this case we use again Proposition \ref{prop:bound_laplace} for the left
hand-side of Eq.~\eqref{limit_mesure_gamma2} and use that $\varepsilon^{r_b-r_a} \to 0$ as ${\varepsilon_{n}}\to0$. Thus, we are
led with the following equality in measure
\[ \Gamma_2(dt) = \frac \alpha \beta u(t)^2 dt\,. \]
Thus, combining this result with Lemma \ref{lem:compactness_weak_star} we can pass
to the limit in \eqref{eq:weak_form_tx_eps} and we obtain the first case of Theorem \ref{thm:LS_fast}.
\subsubsection{The case $r_a=r_b$}
In this case, we use again the fact that by Proposition~\ref{prop:bound_laplace}, up to a sub-sequence of $\{\varepsilon_n\}$ (not relabeled), for all
$i\geq 2$, there exists $d_i \in L^\infty(0,T)$ and $z_0>0$ such that
\[ {\varepsilon_{n}}^{r_b} c_i^{\varepsilon_{n}} \rightharpoonup d_i \, \quad w-*-L^\infty(0,T)\,,\]
and for all $z<z_0$, there exists $K_z>0$ such that
\begin{equation}\label{eq:bound_di_ra_fastdenucl}
0 \leq \sup_{t\in[0,T]}\, \sup_{i\geq 2} \, d_i(t) e^{-iz}< K_z \,.
\end{equation}
From Eq.~\eqref{limit_mesure_gamma2}, we obtain the equality in measure
\[ \overline b_3 d_3 \, dt = \beta \Gamma_2(dt) - \alpha u(t)^2 \, dt \,.\]
Then, iterating the procedure, from equation \eqref{sys:BD_rescaled}, we get that, for all $i\geq 3$ and $\varphi \in
\mathcal C^1([0,T])$
\begin{multline*
{\varepsilon_{n}}^{1-r_a} {\varepsilon_{n}}^{r_a} c_{i}^{\varepsilon_{n}}(T)\varphi(T) - {\varepsilon_{n}}^{1-r_a} {\varepsilon_{n}}^{r_a}
c_{i}^{in,{\varepsilon_{n}}} \varphi(0) -
{\varepsilon_{n}}^{1-r_a} \int_0^T \varphi'(t){\varepsilon_{n}}^{r_a} c_i^{\varepsilon_{n}}(t) dt \\
= \int_0^T \varphi(t) [ \overline a_{i-1}^{\varepsilon_{n}} u^{\varepsilon_{n}}(t){\varepsilon_{n}}^{r_a}c_{i-1}^{\varepsilon_{n}}(t) -
\overline b_i^{\varepsilon_{n}} {\varepsilon_{n}}^{r_a}
c_i^{\varepsilon_{n}}(t)] \, dt \\
- \int_0^T \varphi(t) [\overline a_i^{\varepsilon_{n}} u^{\varepsilon_{n}}(t) {\varepsilon_{n}}^{r_a} c_i^{\varepsilon_{n}}(t) -
\overline b_{i+1}^{\varepsilon_{n}} {\varepsilon_{n}}^{r_a}
c_{i+1}^{\varepsilon_{n}}(t)] dt\,.
\end{multline*}
Hence, for $i=3$, writing $ {\varepsilon_{n}}^{r_a}c_{2}^{\varepsilon_{n}}(t)= \varepsilon_n^{r_a-\eta} d_{2}^{\varepsilon_{n}}(t)\to 0$ (in
$\mathcal M_f([0,T])$), we
obtain
\begin{equation*}
0= \int_0^T \varphi(t)[ - \overline b_3 d_3(t) - \overline a_3 u(t) d_3(t) + \overline b_{4}d_{4}(t)] dt\,.
\end{equation*}
And for all $i\geq 4$,
\begin{equation*}
0= \int_0^T \varphi(t)[\overline a_{i-1} u(t) d_{i-1}(t) - \overline b_i d_i(t) - \overline a_i u(t) d_i(t) + \overline
b_{i+1}d_{i+1}(t)] dt\,.
\end{equation*}
With $H_2=-\overline b_3 d_3$, $H_i=\overline a_i u^\varepsilon d_i(t) - \overline b_{i+1}d_{i+1}$, $i\geq 3$, then we must
have a.e. $H_i=H_2=:H$, for all $i\geq 2$. Then we get, for all $i\geq 3$,
\begin{equation*}
d_i(t)=-\frac{H}{\overline{b}_i}\sum_{j=3}^i\left( \prod_{k=j}^{i-1}
\frac{\overline{a}_k}{\overline{b}_{k}}\right)u^{(i-j)}=-\frac{H}{\overline{b}_i}\sum_{j=3}^i
\left(\frac{\overline{a}u}{\overline{b}}\right)^{i-j}\,.
\end{equation*}
In order to fulfil the bound \eqref{eq:bound_di_ra_fastdenucl},
we must get $H=0$, so that $d_3=0$ and the following equality in measure holds
\[ \Gamma_2(dt) = \frac \alpha \beta u(t)^2 dt \,.\]
This ends the proof of Theorem \ref{thm:LS_fast}.
\section{Extension to a density}\label{sec:density}
In this section, we make an extra-assumption in order to obtain a convergence result in $L^1$ functional space, so that the limit measure has a density with respect to the Lebesgue measure:
\begin{hyp}\label{assumption_5}
There is $\delta\in(0,1/r_a-1)$ such that, for the function $\Psi(y)=y^{1+\delta}$,
\begin{equation}\label{hyp:psi_init} \tag{H8}
\sup_{\varepsilon>0}\int_0^{\infty}\Psi(f^{in,\varepsilon}(x))dx<\infty\,.
\end{equation}
Moreover, the kinetic rates are given by exact power law functions, \textit{i.e.},
\begin{equation}\label{hyp:powerlow} \tag{H9}
\begin{array}{ll}
\displaystyle a_i^\varepsilon = \overline{a}(\varepsilon i)^{r_a}\,,& i\geq 2\,, \\[0.8em]
\displaystyle b_i^\varepsilon = \overline{b}(\varepsilon i)^{r_b}\,, & i\geq 3\,.
\end{array}
\end{equation}
\end{hyp}
\begin{remk}
The first hypothesis \eqref{hyp:psi_init} is slightly stronger than a compactness hypothesis in $L^1(dx)$, where a more
general (and not explicit) $\Psi$ can be obtained, see \cite{Chauhoan1977}. However, having an explicit power low
function for $\Psi$ will simplify the following calculus. The same is valid for the extra
hypothesis \eqref{hyp:powerlow} on the kinetic rates (which is in agreement with hypothesis \eqref{H5}).
\end{remk}
\medskip
\noindent Assuming Assumption \ref{assumption_1}-\ref{assumption_5} hold true, we can now prove the last result.
\medskip
\begin{thm}\label{thm_density}
Assume $\eta \geq r_a$ and $r_a=r_b$. Let a sequence $\{\varepsilon_n\}$ converging to $0$. There
exists $T>0$, a sub-sequence
$\{\varepsilon_{n'}\}$ of $\{\varepsilon_n\}$, and $f\in \mathcal C([0,T],w-L^1(\mathbb R_+,x^{r_a\delta}dx))\cap
L^\infty(0,T;L^1(\mathbb R_+,(1+x)dx))$
such that the measure $f(t,x)dx$ is a $N$-solution of LS with mass $m$ and
\[f^{\varepsilon_{n'}} \xrightharpoonup[n'\to+\infty]{} f\]
in $\mathcal C([0,T];w-L^1(\mathbb R_+,x^{r_a\delta}dx))$. $N$ is given in Theorem
\ref{thm:LS_slow}-\ref{thm:LS_compensated} according to the value of $\eta$.
\end{thm}
\medskip
\noindent The proof of this theorem is based on the following lemma which proof is postponed below
\medskip
\begin{lem} \label{lem:psi_bound}
Assume $\eta \geq r_a$ and $r_a=r_b$. Let a sequence $\{\varepsilon_n\}$ converging to $0$. There exist $T>0$
and a sub-sequence $\{\varepsilon_{n'}\}$ of $\{\varepsilon_n\}$ such that
\begin{equation*
\sup_{n'\geq0} \, \sup_{t\in[0,T]} \, \int_0^\infty \min(1,x^{r_a\delta}) \Psi(f^{\varepsilon_{n'}}(t,x))\, dx
< +\infty\,.
\end{equation*}
\end{lem}
\medskip
{ {\it Proof of Theorem \ref{thm_density}.}
We reproduce the same proof as for Theorem \ref{thm:LS_slow} and \ref{thm:LS_compensated} and obtain a sub-sequence $f^{\varepsilon_{n'}}$
that converges in measure. We now remark that, combining the estimates \eqref{eq:phi_1} in Lemma \ref{lem:u1xphi1}
and the last Lemma \ref{lem:psi_bound} we can apply the Dunford-Pettis theorem and we have a weak compact subset
$\mathcal K$ of $L^1(\mathbb R_+,x^{r\delta}dx)$ such that for all $t\in[0,T]$ and $n'\geq0$, $f^{\varepsilon_{n'}}(t)\in\mathcal
K$. We are
now in position to prove that along another subsequence, still denoted by $\{\varepsilon_{n'}\}$, the
sequence converges to some $f$ in $C([0,T],w-L^1(\mathbb R_+,x^{r\delta}dx))$. Moreover, $f$ belongs to
$L^\infty(0,T,L^1(\mathbb R_+,(1+x)dx)$.
The proof follows similar arguments as in \cite[Proof of Theorem 2.2, p. 981]{Laurencot2002a} which consists in proving
the equicontinuity of
\[t \to \int_{0}^R f^\varepsilon(t,x)\varphi(x)x^{r\delta}dx\,,\]
for all $\varphi\in L^\infty(0,R)$ and $R>0$. Indeed, by Eq.~\eqref{eq:equi_h} we have for any $\varphi \in \mathcal C^1$ with
compact support in $(0,R)$ that (see also the proof of lemma \ref{lem:equicontinuity})
\[ \lim_{h\to0} \, \sup_{t\in[0,T-h]} \, \sup_{s\in(0,h)}\, \left| \int_0^\infty
(f^\varepsilon(t+s,x)-f^\varepsilon(t,x) )\varphi(x)\, x^{r\delta} dx \right| = 0 \,. \]
Then taking a pointwise convergent sequence $\{\varphi^n\}$ in $\mathcal C_c([0,R])$ of $\varphi\in L^\infty(0,R)$ and using
Egorov's theorem we get the desire results. Finally, we apply a variant of Arzela-Ascoli theorem for weak topology,
see \cite[Theorem 1.3.2]{Vrabie}, so that for each $R>0$, the sequence is relatively compact in
$C([0,T],w-L^1((0,R),r^{r\delta}dx)$.
By the compact containment we improve this results on $\mathbb R_+$.
\subsection*{Technical results}
\medskip
\noindent Before proving Lemma \ref{lem:psi_bound}, we start by some technical lemmas.
\medskip
\begin{lem}\label{lem1}
Let $\varphi \in \mathcal C_b\left([0,\infty)\right)$ non-negative. Then, for any $I\geq 3$,
\begin{multline} \label{eq:vphi_Psi}
\int_0^\infty \varphi(x) \left[ \Psi(f^\varepsilon(t,x)) - \Psi(f^{in,\varepsilon}(x)) \right] \, dx \\ \leq \varepsilon \sum_{i=2}^{I-1}
\varphi_i^\varepsilon \Psi(c_i^\varepsilon(t)) + \int_0^t \left[\varphi_{I}^\varepsilon a_{I-1}^\varepsilon u^\varepsilon(s) \Psi(c_{I-1}^\varepsilon(s)) -
\varphi_{I-1}^\varepsilon b_I^\varepsilon \Psi(c_{I}^\varepsilon(s))\right]ds \\
+\int_0^t\int_{(I-1/2)\varepsilon}^\infty \Big{[} a^\varepsilon(x) u^\varepsilon(s) \Delta_\varepsilon\varphi(x) - b^\varepsilon(x)
\Delta_{-\varepsilon}\varphi(x) \\ - \delta \left( u^\varepsilon(s) \Delta_{-\varepsilon} a^\varepsilon(x) - \Delta_{\varepsilon} b^\varepsilon(x)
\varphi(x)\right) \Big{]}\Psi(f^\varepsilon(x,s))dxds \,.
\end{multline}
%
where $\varphi_i^\varepsilon = 1/{\varepsilon} \int_{\Lambda_i^\varepsilon} \varphi(x) \, dx$.
\end{lem}
\medskip
\begin{proof}
The proof follows similar lines as in \cite[Lemma 4.1]{Laurencot2002a}, but we take profit of the
explicit form of $\Psi$ to obtain a necessary finer estimate. We sketch it briefly below. From the (BD) system
\eqref{sys:BD_rescaled}, it comes
\begin{multline*}
\int_0^\infty \varphi(x) \left[ \Psi(f^\varepsilon(t,x)) - \Psi(f^{in,\varepsilon}(x)) \right] \, dx= \sum_{i\geq 2}
\int_{\Lambda_i^\varepsilon} \varphi(x) \left[\Psi(c_i^\varepsilon(t)) - \Psi(c_i^\varepsilon(0))\right] \, dx \\
= \varepsilon \sum_{2\leq i\leq I-1} \varphi_i^\varepsilon \left[\Psi(c_i^\varepsilon(t)) - \Psi(c_i^\varepsilon(0))\right] + \sum_{i\geq I} \varphi_i^\varepsilon \int_0^t [J_{i-1}^\varepsilon(s) -
J_i^\varepsilon(s)] \Psi'(c_i^\varepsilon(s))\, ds \,.
\end{multline*}
We can decompose the latter in three parts,
\begin{equation*}
\begin{array}{rcl}
\displaystyle \int_0^\infty \varphi(x) \left[ \Psi(f^\varepsilon(t,x)) - \Psi(f^{in,\varepsilon}(x)) \right] \, dx & = & \displaystyle N^\varepsilon(t)
+ \int_0^t [A^\varepsilon(s)+B^\varepsilon(s)] \, ds \,,
\end{array}
\end{equation*}
where
\begin{equation*}
\begin{array}{rcl}
\displaystyle N^\varepsilon(t) & := & \displaystyle \varepsilon \sum_{2\leq i\leq I-1} \varphi_i^\varepsilon \left[\Psi(c_i^\varepsilon(t)) - \Psi(c_i^\varepsilon(0))\right]\,, \\[0.8em]
\displaystyle A^\varepsilon(t) & := & \displaystyle \sum_{i\geq I} \varphi_i^\varepsilon u^\varepsilon(t) [a_{i-1}^\varepsilon c_{i-1}^\varepsilon(t) - a_{i}^\varepsilon
c_{i}^\varepsilon(t)] \Psi'(c_i^\varepsilon(t))\,,\\[0.8em]
\displaystyle B^\varepsilon(t) & := & \displaystyle \sum_{i\geq I} \varphi_i^\varepsilon [ b_{i+1}^\varepsilon c_{i+1}^\varepsilon(t)-b_{i}^\varepsilon c_{i}^\varepsilon (t)]
\Psi'(c_i^\varepsilon (t))\,.
\end{array}
\end{equation*}
Then, in $A^\varepsilon$ we can re-write, using the convexity of $\Psi$, for all $i\geq I$,
\begin{multline*}
[a_{i-1}^\varepsilon c_{i-1}^\varepsilon(t) - a_{i}^\varepsilon c_{i}^\varepsilon(t)] \Psi'(c_i^\varepsilon(t)) \\
= a_{i-1}^\varepsilon [c_{i-1}^\varepsilon(t) - c_{i}^\varepsilon(t)] \Psi'(c_i^\varepsilon(t)) + (a_{i-1}^\varepsilon-a_i^\varepsilon)
c_{i}\Psi'(c_i^\varepsilon(t)) \\
\leq a_{i-1}^\varepsilon \left( \Psi(c_{i-1}^\varepsilon(t)) - \Psi(c_{i}^\varepsilon(t)) \right) +(a_{i-1}^\varepsilon-a_i^\varepsilon)
c_i^\varepsilon(t)\Psi'(c_i^\varepsilon(t)) \,.
\end{multline*}
Then, reordering the term in the last inequality and then using that $x\Psi'(x) - \Psi(x) = \delta \Psi(x)$,
\begin{multline*}
[a_{i-1}^\varepsilon c_{i-1}^\varepsilon(t) - a_{i}^\varepsilon c_{i}^\varepsilon(t)] \Psi'(c_i^\varepsilon(t))\\
\leq a_{i-1}^\varepsilon \Psi(c_{i-1}^\varepsilon(t)) - a_i^\varepsilon \Psi(c_{i}^\varepsilon(t)) +(a_{i-1}^\varepsilon-a_i^\varepsilon) [
c_i^\varepsilon(t)\Psi'(c_i^\varepsilon(t))
-\Psi(c_i^\varepsilon(t))] \\
= a_{i-1}^\varepsilon \Psi(c_{i-1}^\varepsilon(t)) - a_i^\varepsilon \Psi(c_{i}^\varepsilon(t)) - \delta (a_{i}^\varepsilon-a_{i-1}^\varepsilon)
\Psi(c_i^\varepsilon(t))\,.
\end{multline*}
Thus, we obtain for $A$ the following estimation,
\begin{multline*}
A^\varepsilon(t) \leq \sum_{i\geq I} a_i^\varepsilon u^\varepsilon (\varphi_{i+1}^\varepsilon-\varphi_i^\varepsilon) \Psi(c_i^\varepsilon(t)) +
\varphi_{I}^\varepsilon a_{I-1}^\varepsilon u^\varepsilon(t)
\Psi(c_{I-1}^\varepsilon(t))\\- \delta u^\varepsilon \sum_{i\geq I}\varphi_i^\varepsilon (a_i^\varepsilon-a_{i-1}^\varepsilon) \Psi(c_i^\varepsilon) \,.
\end{multline*}
We estimate $B$, by similar argument, to get,
\begin{multline*}
B^\varepsilon(t) \leq \sum_{i\geq I} \varphi_i^\varepsilon [b_{i+1}^\varepsilon \Psi(c_{i+1}^\varepsilon)- b_i^\varepsilon \Psi(c_i^\varepsilon)] +\delta \sum_{i\geq I} \varphi_i^\varepsilon (b_{i+1}^\varepsilon -b_i^\varepsilon) \Psi(c_i^\varepsilon) \\
\leq \sum_{i\geq I} (\varphi_{i-1}^\varepsilon - \varphi_{i}^\varepsilon) b_{i}^\varepsilon \Psi(c_{i}^\varepsilon) - \varphi_{I-1}^\varepsilon b_I^\varepsilon \Psi(c_{I}^\varepsilon) +\delta \sum_{i\geq I} \varphi_i^\varepsilon (b_{i+1}^\varepsilon -b_i^\varepsilon) \Psi(c_i^\varepsilon)\,.
\end{multline*}
Both estimates on $A^\varepsilon$ and $B^\varepsilon$ directly give \eqref{eq:vphi_Psi}.
\end{proof}
\medskip
\begin{lem}\label{lem:estimequitue}
For all $0\leq r<1$, and for all $0<\delta<\frac{1}{r}-1$, there exists $I_0$ such that for all $i\geq I_0$, and all
$x\in[0,1]$,
\begin{equation*}
\Big{[}i^{r}\left((i+1/2+x)^{r\delta}-(i-1/2+x)^{r\delta}\right)-\delta(i^{r}-(i-1)^{r}) (i-1/2+x)^{r\delta}\Big{]}\leq
0\,,
\end{equation*}
\end{lem}
\medskip
\begin{proof}
Doing an expansion as $i\to \infty$, we easily obtain
\begin{multline*}
\Big{[}i^{r}\left((i+1/2+x)^{r\delta}-(i-1/2+x)^{r\delta}\right)-\delta(i^{r}-(i-1)^{r}) (i-1/2+x)^{r\delta}\Big{]}\\
=r\delta \frac{i^r(i-\frac{1}{2}+x)^{r\delta}}{i^2}\Big{[}\frac{r(1+\delta)-1}{2}-x+O(\frac{1}{i})\Big{]}\,.
\end{multline*}
We conclude straightforwardly as $r(1+\delta)-1<0$.
\end{proof}
\medskip
\subsection*{Proof of Lemma \ref{lem:psi_bound}}
In the following, let $r=r_a=r_b$ and $I=I_0$ given by Lemma \ref{lem:estimequitue}. We want to
bound each term of Eq. \eqref{eq:vphi_Psi} with $\varphi(x)=\min(1,x^{r\delta})$.
Remark the term $-\varphi_{I_0-1}^\varepsilon b_{I_0}^\varepsilon \Psi(c_{I_0}^\varepsilon(t))$ can be easily
dropped in Eq. \eqref{eq:vphi_Psi} since it is non-positive. Also, note that, for $2\leq i\leq I_0$,
\begin{equation*}
\varepsilon \varphi_i^\varepsilon \Psi(c_i^\varepsilon(t)) \leq \varepsilon^{1-r(1+\delta )} \varphi_i^\varepsilon
\left(\varepsilon^{r}c_i^\varepsilon(t)\right)^{1+\delta}\,.
\end{equation*}
Thus, since $\varphi_i^\varepsilon$ is bounded and $\delta \leq 1/r-1$, we apply Lemma
\ref{lem:compactness_weak_star_2} and Proposition \ref{prop:bound_laplace} to obtain $T>0$ and a sub-sequence, still denoted by $\{\varepsilon_n\}$, such that
\begin{equation}\label{begin}
\sup_{n\geq 0}\, \sup_{t\in[0,T]} \, \left(\varepsilon_n \varphi_i^{\varepsilon_n} \Psi(c_i^{\varepsilon_n}(t))\right)<\infty.
\end{equation}
Similarly, using that $u^\varepsilon(t)\leq K_m$, we have
\begin{multline}
\varphi_{I_0}^\varepsilon a_{I_0-1}^\varepsilon u^\varepsilon(t)\Psi(c_{I_0-1}^\varepsilon(t)) =
\overline{a}(I_0-1)^{r}u^\varepsilon(t)\left(\int_{I_0-1/2}^{I_0+1/2}y^{r \delta}dy\right) \left(\varepsilon^{r}
c_{I_0-1}^\varepsilon(t)\right)^{1+\delta}\\
\leq K_m \overline{a}(I_0-1)^{r}\left(\int_{I_0-1/2}^{I_0+1/2}y^{r \delta}dy\right)
\sup_{\varepsilon>0} \, \sup_{t\in[0,T]}\, \left(\varepsilon^{r} c_{I_0-1}^\varepsilon(t)\right)^{1+\delta} <\infty\,,
\end{multline}
By these estimates, the boundary terms in Eq.~\eqref{eq:vphi_Psi} are uniformly bounded. We are lead
with the remaining integral term on $\left((I_0-1/2)\varepsilon,\infty\right)$. Denote, for all $\varepsilon>0$ and $x>0$,
\[ D^\varepsilon(x) = a^\varepsilon(x) u^\varepsilon(t) \Delta_\varepsilon\varphi(x)- b^\varepsilon(x) \Delta_{-\varepsilon}\varphi(x)
-\delta \left( u^\varepsilon \Delta_{-\varepsilon} a^\varepsilon(x) - \Delta_{\varepsilon} b^\varepsilon(x) \right)\varphi(x)\,. \]
Thus,
\begin{multline*}
\int_{(I_0-1/2)\varepsilon}^{1} D^\varepsilon(x) \Psi(f^\varepsilon(x,t))dx\\
=\sum_{i= I_0}^{1/\varepsilon} \frac{1}{\varepsilon}\int_{\Lambda_i^\varepsilon}\Big{[} \left( a_i^\varepsilon u^\varepsilon(t)
(\varphi(x+\varepsilon)-\varphi(x)) - b_i^\varepsilon (\varphi(x)-\varphi(x-\varepsilon)) \right) \\-\delta \left(
u^\varepsilon(a_{i}^\varepsilon-a_{i-1}^\varepsilon) - (b_{i+1}^\varepsilon-b_i^\varepsilon) \right)\varphi(x) \Big{]} \Psi(c_i^\varepsilon(t))dx\,.
\end{multline*}
Then, on $x\in (0,1)$, we have that $\varphi(x)=x^{r\delta}$, and letting $\Gamma_i=[i-1/2,i+1/2)$ and changing variable $\varepsilon y=x$, we obtain
\begin{multline*}
\int_{(I_0-1/2)\varepsilon}^{1} D^\varepsilon(x) \Psi(f^\varepsilon(x,t))dx\\
= \sum_{i= I_0}^{1/\varepsilon} \varepsilon^{r(1+\delta)}\int_{\Gamma_i}\Big{[} \left( \overline{a} i^r u^\varepsilon(t)
((y+1)^{r\delta}-y^{r\delta}) - \overline{b}i^r (y^{r\delta}-(y-1)^{r\delta}) \right)\\-\delta \left( u^\varepsilon
\overline{a}(i^r-(i-1)^r) - \overline{b} ((i+1)^r-i^r) \right)y^{r\delta} \Big{]} \Psi(c_i^\varepsilon(t))dy\,.
\end{multline*}
Finally, rearranging the term we have
\begin{multline*}
\int_{(I_0-1/2)\varepsilon}^{1} D^\varepsilon(x) \Psi(f^\varepsilon(x,t))dx\\
=\sum_{i= I_0}^{1/\varepsilon} \varepsilon^{r(1+\delta)}\int_{\Gamma_i}\Big{[} \left( \overline{a} u^\varepsilon(t)-
\overline{b}\right) \left( i^r ((y+1)^{r\delta}-y^{r\delta}) -\delta (i^r-(i-1)^r)y^{r\delta} \right) \\
+\overline{b}i^r \left((y+1)^{r\delta}-2y^{r\delta}+(y-1)^{r\delta}\right)\\
+\delta \overline{b} \left((i+1)^r-2i^r
+(i-1)^r\right) y^{r\delta} \Big{]} \Psi(c_i^\varepsilon(t))dy\,.
\end{multline*}
Then, as the second discrete derivative are negative, that is, for all $s<1$ and all $x>1$,
\begin{equation*}
\left((x+1)^{s}-2x^{s}+(x-1)^{s}\right)\leq 0\,,
\end{equation*}
we obtain
\begin{multline*}
\int_{(I_0-1/2)\varepsilon}^{1} D^\varepsilon(x) \Psi(f^\varepsilon(x,t))dx\\
\leq \varepsilon^{r(1+\delta)} \left( \overline{a}
u^\varepsilon(t)- \overline{b}\right) \sum_{i= I_0}^{1/\varepsilon} \int_{\Lambda_i}\Big{[} i^r ((y+1)^{r\delta}-y^{r\delta})\\
-\delta (i^r-(i-1)^r)y^{r\delta} \Big{]} \Psi(c_i^\varepsilon(t))dy\,.
\end{multline*}
The term under the integral is negative by Lemma \ref{lem:estimequitue}. We now fix $T>0$ and extract a
sub-sequence $\{\varepsilon_{n'}\}$ given by Lemma \ref{lem:compactness_weak_star_2} such that $\overline{a}
u^\varepsilon(t)- \overline{b} > 0$ on $[0,T]$. Thus,
\begin{equation}
\int_{(I_0-1/2)\varepsilon}^{1} D^\varepsilon(x) \Psi(f^\varepsilon(x,t))dx \leq 0 \,.
\end{equation}
On the other hand we have, since $\Delta_\varepsilon \varphi = 0$ on $(1,+\infty)$,
\begin{multline} \label{final}
\int_{1}^{\infty} \Big{[} D^\varepsilon(x) \Psi(f^\varepsilon(x,t))dx\\
\leq \delta ( K_m \sup_{x\geq 1}\mid a'(x) \mid + \sup_{x\geq 1}\mid b'(x) \mid)
\int_{1}^{\infty} \varphi(x) \Psi(f^\varepsilon(x,t))dx\,,
\end{multline}
and we conclude by estimates \eqref{begin} to \eqref{final} that, for some constant $K>0$ and all $t\in[0,T]$,
\begin{equation*}
\int_0^\infty \varphi(x) \Psi(f^{\varepsilon_n}(t,x)) \leq K + \int_0^\infty \Psi(f^{in,\varepsilon_n}(x))\, dx
+ K \int_0^t \int_0^\infty \varphi(x) \Psi(f^{\varepsilon_n}(t,x))\,.
\end{equation*}
We conclude the proof with the Gronwall Lemma.
\subsection*{The general case}
The main difficulty to treat the case $r_a<r_b$ is to find a test function $\varphi$ in Eq. \eqref{eq:vphi_Psi} which
make the term under the integral negative around $0$, but which also keep the boundary terms bounded. We believe that a
good function would be
\[\varphi(x) = \min( x^{r\delta}e^{-Kx^{r_b-r_a}},c)\,,\]
for some $c>0$ small and $K>0$ large enough. It recovers the case $r_a=r_b$ (with $c=1$). Computations are not presented here because too fastidious. Just let us show that, at the limit $\varepsilon\to 0$,
\begin{multline*}
[ \overline a x^{r_a} u(t) - \overline b x^{r_b}] \varphi'(x) - \delta \left[ r_a \overline a x^{r_a-1} u(t) -
r_b \overline b x^{r_b-1} \right] \varphi(x) \\
= \frac{\varphi(x)}{x} (r_b-r_a)\left[ \delta \overline b x^{r_b} - K x^{r_b-r_a} (\overline a x^{r_a} u(t) -\overline
b x^{r_b}) \right] \,.
\end{multline*}
But since $u(t)> \rho$, it exists $x_0>0$ small and $\gamma>0$ such that the flux is bounded from below by $ \overline
a x^{r_a}u(t) - \overline b x^{r_b} \geq \gamma \overline a x^{r_a}$ on
$[0,x_0]$, thus
\begin{multline*}
[ \overline a x^{r_a} u(t) - \overline b x^{r_b}] \varphi'(x) - \delta \left[ r_a \overline a x^{r_a-1} u(t) -
r_b \overline b x^{r_b-1} \right] \varphi(x) \\
\leq \frac{\varphi(x)}{x}(r_b-r_a) \left[ \delta \overline b - K \gamma \right] x^{r_b} \,.
\end{multline*}
Hence, for $K$ large enough the term is negative around $0$, which was the essential ingredient of the proof of Theorem \ref{thm_density}.
\section{Discussion}\label{sec:disc}
In this work, we obtained limit theorems to derive rigorously the link between a discrete-size
coagulation-fragmentation model, the Becker-D\"oring (BD) model, and a continuous-size model, the Lifshitz-Slyozov (LS)
model. We used weak-convergence in measure, to prove that a sequence of discrete stepwise functions
associated to the BD model converges towards a measure solution of the LS model. The novelty of our work, compared to
previous work in \cite{Laurencot2002a,Collet2002}, consists of being able to rigorously defined a boundary flux
condition for the limit non-linear transport partial differential equation of the LS model. This boundary condition has
been obtained thanks to an averaging procedure for the smaller-sized cluster, namely the one of size $i=2$. It is
classical when passing from a discrete to a continuous model (think of a random walk converging to a Brownian motion)
to accelerate the rates (or equivalently, the time) between each discrete transition. Hence, each individual
discrete-size cluster evolves in the re-scaled BD model~\eqref{sys:BD_rescaled} at a faster time scale than the
continuous density function $f^\varepsilon$ in Eq.~\eqref{eq:weak_form_eps}. Although the fast-motion involves a dynamical
system of infinite dimension, we could obtain appropriate $L^\infty$-bounds on the time trajectories of each
discrete-sized cluster, and proves that, in the limit when the scaling parameter $\varepsilon\to 0$, each discrete-sized
cluster is the unique solution of an algebraic equation, which appears to be the same as the steady-state condition of a constant monomer BD model.
Let us now discuss in more details what were the scaling assumptions that lead to the study of the
system~\eqref{sys:BD_rescaled} (for the mathematical derivation, see the appendix~\ref{annex:adimensionalixation}).
Roughly, the system~\eqref{sys:BD_rescaled} is obtained when we consider that the clusters have very large sizes but
are present in a low quantity compared to a large excess of free particles. The rescaled equations are obtained in a
large volume hypothesis, and the scaling of the macroscopic reaction rates accounts for the volume-dependence of the
aggregation (so that aggregation and fragmentation occur at the same time scale).
\noindent However, importantly enough, the first aggregation (nucleation) rate is scaled differently from the other
aggregation rates (see Appendix \ref{annex:adimensionalixation}) and this comes from the special role played by the
free particles.
Despite the large excess of free particles, in this framework, the nucleation occurs at the same time scale than the
aggregation of large-sized clusters, and has for consequence to prevent the formation of too many clusters. A
different choice at this step would lead to a rapid depletion of free particles, and would result in different mass
conservation where free particles are not present as a distinct entity any more-- see the work \cite{Laurencot2002a} on
the Lifshitz-Slyozov-Wagner equation.
\noindent Finally, we allowed a flexibility in the scaling of the first fragmentation (de-nucleation), quantified by
the exponent $\eta$. We found (see Theorems \ref{thm:LS_slow}-\ref{thm:LS_compensated}-\ref{thm:LS_fast} ) that
different values of $\eta$ give rise to distinct boundary condition at the limit when $\varepsilon$ goes to $0$. The most
natural case, $\eta=r_b$, corresponds to the case where the clusters of size $2$ dissociate at the same speed than the
small-sized clusters of size $i$, $i\geq 3$. Then, the case $\eta>r_b$ corresponds to an asymptotically irreversible
nucleation (and leads to a macroscopic flux $N(t)= \alpha u(t)^2$, which corresponds to the microscopic nucleation rate
-- this conclusion actually holds for all $\eta>r_a$). And the case $\eta\leq r_a<r_b$ corresponds to a strongly
reversible de-nucleation (and leads to $0 \leq N(t)<\alpha u(t)^2$ according to the value $r_a$).
Hence, our work shed lights on which appropriate boundary condition should be used for the LS equation (or
similar continuous coagulation models) according to specific microscopic hypotheses (unfavorable, balanced or
irreversible nucleation). We believe that our procedure could be applied to several related models (for instance, the
Lifshitz-Slyozov-Wagner equation mentioned above, or the prion equation \cite{doumic}) and should help to build reduced structured population models while
taking into account of their intrinsic multi-scale nature (see \cite{Yvinec2012,Yvinec2016} for applications).
|
\section{Introduction}
In \cite{M_K} we have discovered that the $K_0$ group for a $C^*$-algebra $A$ can be defined using more general elements than projections. Namely. one can consider homotopy classes of {\it pairs} $(a,b)$ of selfadjoint matrices with entries in $A$ that satisfy the relations $p(a)=p(b)$ for polynomials $p(t)=t(1-t)$ and $t^2(1-t)$ (or, equivalently, for all polynomials satisfying $p(0)=p(1)=0$). Genuine projections satisfy $p(a)=p(b)=0$, but this property is too restrictive. It suffices to require only that the defect from being a projection should be the same for $a$ and for $b$, but it doesn't need to vanish.
In this paper we give a similar description for the $K_1$ group. It is generated by {\it balanced} pairs $(a,b)$, where $a,b$ need not to be unitaries, but their defect from being unitary should be the same.
Then we consider pairs $(D_1,D_2)$ of order zero pseudodifferential operators on a manifold $M$, such that their symbols $\sigma_1$, $\sigma_2$ are a balanced pair of matrix-valued functions on the cospherical bundle $S^*M$ over $M$. As $(\sigma_1,\sigma_2)$ represents an element in $K^1(S^*M)$, so its topological index is defined. We show that one can decompose $L^2(M)$ as an orthogonal direct sum $L^2(M)=H_1\oplus H_2$ in such a way that the restrictions of $D_1$ and of $D_2$ onto $H_2$ are almost the same, and the restrictions of $D_1$, $D_1^*$, $D_2$, $D_2^*$ onto $H_1$ are left-invertible modulo compact operators. The latter property allows to define a relative analytical index for the pair $((D_1)|_{H_1},(D_2)|_{H_1})$ and to show that it is equal to the topological index determined by the pair of symbols. We need hardly mention that neither the symbols $\sigma_1$, $\sigma_2$ have to be invertible, nor the operators $D_1$, $D_2$ have to be elliptic.
\section{$K_1$ group --- another description}
Let $A$ be a $C^*$-algebra. For two contractions $a,b\in A$, consider the following sets of relations
\begin{equation}\label{rel1}
a^*a=b^*b;\quad aa^*=bb^*;\quad a(1-a^*a)=b(1-b^*b);\quad (1-aa^*)a=(1-bb^*)b
\end{equation}
and
\begin{equation}\label{rel2}
(a-b)c=(a^*-b^*)c=0 \mbox{ for\ } c \mbox{ being\ one\ of\ } 1-a^*a, 1-aa^*, 1-b^*b,1-bb^*.
\end{equation}
Although we use the unit in these relations for convenience of notation, these relations make sense for non-unital $C^*$-algebras as well.
\begin{lem}
Under the assumption that $a$ and $b$ are contractions, the sets of relations (\ref{rel1}) and (\ref{rel2}) are equivalent.
\end{lem}
\begin{proof}
(\ref{rel2}) easily (algebraically) follows from (\ref{rel1}). To prove the opposite, one needs to use the uniqueness of the positive square root in $C^*$-algebras. It follows from (\ref{rel2}) that
$$
(1-a^*a)a^*a=a^*(1-aa^*)a=a^*(1-aa^*)b=(1-a^*a)a^*b=(1-a^*a)b^*b,
$$
therefore
$$
(1-a^*a)(1-b^*b)=(1-a^*a)-(1-a^*a)b^*b=(1-a^*a)-(1-a^*a)a^*a=(1-a^*a)^2.
$$
Passing to adjoints, we get
$$
(1-b^*b)(1-a^*a)=(1-a^*a)^2.
$$
Interchanging $a$ and $b$, we get
$$
(1-a^*a)(1-b^*b)=(1-b^*b)^2.
$$
Thus, $(1-a^*a)^2=(1-b^*b)^2$, hence $1-a^*a=1-b^*b$, and $a^*a=b^*b$. Similarly one can prove that $aa^*=bb^*$. The two other relations in (\ref{rel1}) can be shown algebraically.
\end{proof}
\begin{defn}
Pairs $(a,b)$ of contractions satisfying the relations (\ref{rel1}) or (\ref{rel2}) are called {\it balanced}.
\end{defn}
Two pairs, $(a_0,b_0)$ and $(a_1,b_1)$ of elements in $A$, are {\it homotopy equivalent} if there are paths $a=(a_t),b=(b_t):[0,1]\to A$, connecting $a_0$ with $a_1$ and $b_0$ with $b_1$ respectively, such that the pair $(a_t,b_t)$ is balanced for each $t\in[0,1]$.
Note that the pair $(0,0)$ is balanced. A pair $(a,b)$ is {\it homotopy trivial} if it is homotopy equivalent to $(0,0)$.
\begin{lem}\label{L1}
The pair $(a,a)$ is homotopy trivial for any $a\in A$.
\end{lem}
\begin{proof}
The linear homotopy $a_t=t\cdot a$ would do.
\end{proof}
\begin{lem}
If $\|a\|<1$, $\|b\|<1$ and the pair $(a,b)$ is balanced then it is homotopy trivial.
\end{lem}
\begin{proof}
The assumption implies that $a=b$.
\end{proof}
Let $M_n(A)$ denote the $n{\times n}$ matrix algebra over $A$. Two balanced pairs, $(a_0,b_0)$ and $(a_1,b_1)$, where $a_0,a_1,b_0,b_1\in M_n(A)$, are {\it equivalent} if there is a homotopy trivial pair $(a,b)$, $a,b\in M_m(A)$ for some integer $m$, such that the balanced pairs $(a_0\oplus a,b_0\oplus b)$ and $(a_1\oplus a,b_1\oplus b)$ are homotopy equivalent in $M_{n+m}(A)$. Using the standard inclusion $M_n(A)\subset M_{n+k}(A)$ (as the upper left corner) we may speak about equivalence of pairs of different matrix size.
Let $[(a,b)]$ denote the equivalence class of the pair $(a,b)$, $a,b\in M_n(A)$.
For two pairs, $(a,b)$, $a,b\in M_n(A)$, and $(c,d)$, $c,d\in M_m(A)$, set
$$
[(a,b)]+[(c,d)]=[(a\oplus c,b\oplus d)].
$$
The result obviously doesn't depend on a choice of representatives. Also $[(a,b)]+[(c,d)]=[(a,b)]$ when $(c,d)$ is homotopy trivial.
\begin{lem}\label{commut}
The addition is commutative and associative.
\end{lem}
\begin{proof}
If $(u_t)_{t\in[0,1]}$ is a path of unitaries in $A$, $u_1=1$, $u_0=u$, then $[(u^*au,u^*bu)]=[(a,b)]$ for any $a,b\in A$, as the relations (\ref{rel1}) are not affected by unitary equivalence. The standard argument with a unitary path connecting $\left(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\right)$ and $\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)$ proves commutativity. A similar argument proves associativity.
\end{proof}
\begin{lem}
$[(a,b)]+[(b,a)]=[(0,0)]$ for any balanced pair $(a,b)$.
\end{lem}
\begin{proof}
By definition, $[(a,b)]+[(b,a)]=\left[\left(\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right),\left(\begin{smallmatrix}b&0\\0&a\end{smallmatrix}\right)\right)\right]$.
Set
\begin{equation}\label{U}
U_t=\left(\begin{smallmatrix}\cos t&-\sin t\\ \sin t&\cos t\end{smallmatrix}\right),
\end{equation}
$$
A=\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right),\quad B_t=U_t^*\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right)U_t.
$$
We claim that if the pair $(a,b)$ is balanced then the pair $(A,B_t)$ is balanced for any $t$. If true, this implies that
$$
\left[\left(\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right),\left(\begin{smallmatrix}b&0\\0&a\end{smallmatrix}\right)\right)\right]=[(A,B_{\pi/2})]=[(A,B_0)]=\left[\left(\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right),\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right)\right)\right]=[0,0]
$$
by Lemma \ref{L1}. So, it remains to check that the relations (\ref{rel1}) hold for $(A,B_t)$. Note that if the pair $(a,b)$ is balanced then $\left(\begin{smallmatrix}a^*a&0\\0&b^*b\end{smallmatrix}\right)=a^*a\left(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\right)$. So,
$$
B_t^*B_t=U_t^*\left(\begin{smallmatrix}a^*&0\\0&b^*\end{smallmatrix}\right)U_t U_t^*\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right)U_t=U_t^*a^*aU_t=A^*A;
$$
$$
B_t(1-B_t^*B_t)=U_t^*\left(\begin{smallmatrix}a&0\\0&b\end{smallmatrix}\right)(1-a^*a)U_t=U_t^*\left(\begin{smallmatrix}a(1-a^*a)&0\\0&a(1-a^*a)\end{smallmatrix}\right)U_t=A(1-A^*A).
$$
The two other relations in (\ref{rel1}) are proved in the same way.
\end{proof}
\begin{lem}
$[(a,b)]+[(a^*,b^*)]=[(0,0)]$ for any balanced pair $(a,b)$.
\end{lem}
\begin{proof}
Set
$$
A_t=\left(\begin{smallmatrix}a&0\\0&1\end{smallmatrix}\right) U_t^*\left(\begin{smallmatrix}1&0\\0&a^*\end{smallmatrix}\right) U_t,\quad
B_t=\left(\begin{smallmatrix}b&0\\0&1\end{smallmatrix}\right) U_t^*\left(\begin{smallmatrix}1&0\\0&b^*\end{smallmatrix}\right) U_t,
$$
where $U_t$ is given by (\ref{U}). We claim that the pair $(A_t,B_t)$ is balanced for any $t$. Let us check the first relation in (\ref{rel1}) (other relations are checked similarly).
\begin{eqnarray*}
A_t^*A_t&=&C_t^*\left(\begin{smallmatrix}1&0\\0&a\end{smallmatrix}\right)C_t\left(\begin{smallmatrix}a^*a&0\\0&1\end{smallmatrix}\right)C_t^*\left(\begin{smallmatrix}1&0\\0&a^*\end{smallmatrix}\right)C_t\\
&=&C_t^*\left(\begin{smallmatrix}1&0\\0&a\end{smallmatrix}\right)\Bigl(\left(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\right)+(1-a^*a)\left(\begin{smallmatrix}-\cos^2 t&\cos t\sin t\\ \cos t\sin t&-\sin^2 t\end{smallmatrix}\right)\Bigr)\left(\begin{smallmatrix}1&0\\0&a^*\end{smallmatrix}\right)C_t\\
&=&C_t^*\left(\begin{smallmatrix}1&0\\0&a\end{smallmatrix}\right)\left(\begin{smallmatrix}1&0\\0&a^*\end{smallmatrix}\right)C_t+C_t^*\left(\begin{smallmatrix}1&0\\0&a\end{smallmatrix}\right)(1-b^*b)\left(\begin{smallmatrix}-\cos^2 t&\cos t\sin t\\ \cos t\sin t&-\sin^2 t\end{smallmatrix}\right)\left(\begin{smallmatrix}1&0\\0&a^*\end{smallmatrix}\right)C_t\\
&=&C_t^*\left(\begin{smallmatrix}1&0\\0&bb^*\end{smallmatrix}\right)C_t+C_t^*\left(\begin{smallmatrix}1&0\\0&b\end{smallmatrix}\right)(1-b^*b)\left(\begin{smallmatrix}-\cos^2 t&\cos t\sin t\\ \cos t\sin t&-\sin^2 t\end{smallmatrix}\right)\left(\begin{smallmatrix}1&0\\0&b^*\end{smallmatrix}\right)C_t= B_t^*B_t.
\end{eqnarray*}
One has $A_0=\left(\begin{smallmatrix}a&0\\0&a^*\end{smallmatrix}\right)$, $B_0=\left(\begin{smallmatrix}b&0\\0&b^*\end{smallmatrix}\right)$; $A_{\pi/2}=\left(\begin{smallmatrix}a^*a&0\\0&1\end{smallmatrix}\right)=B_{\pi/2}$. By Lemma \ref{L1}, we are done.
\end{proof}
We see that the equivalence classes of balanced pairs in matrix algebras over $A$ form an abelian group for any $C^*$-algebra $A$. Let us denote this group by $L_1(A)$.
Let now $A$ be unital.
Note that the pairs $(u,v)$, where $u,v\in A$ are unitaries, are patently balanced. The map
$$
\iota([u])=[(u,1)]
$$
gives rise to a homomorphism $\iota:K_1(A)\to L_1(A)$.
Set
\begin{equation}\label{unitary}
c=c(a,b)=1+b^*(a-b).
\end{equation}
\begin{lem}
Let $(a,b)$ be a balanced pair. Then
\begin{enumerate}
\item
$c(a,b)$ is unitary; similarly, $1+(a-b)b^*$ is unitary;
\item
$bc=a$;
\item
$b^*b$ commutes with $c$ and with $c^*$, hence with $f(c)$ for any continuous function $f$ on the spectrum of $c$;
\item
$(1-b^*b)(c-1)=0=(c-1)(1-b^*b)$, hence $(1-b^*b)g(c)=0$ for any continuous function $g$ on the spectrum of $c$ with $g(1)=0$
\end{enumerate}
\end{lem}
\begin{proof}
(1) $c^*c=(1+(a-b)^*b)(1+b^*(a-b))=1+a^*b-b^*b+b^*a-b^*b+(a-b)^*bb^*(a-b)=1+a^*b+b^*a-2b^*b+(a-b)^*(bb^*-1)(a-b)+(a-b)^*(a-b)=
1+a^*b+b^*a-2b^*b+(a-b)^*(a-b)=1+a^*b+b^*a-2b^*b+a^*a+b^*b-a^*b-b^*a=1$, similarly one gets $cc^*=1$. The case of $1+(a-b)b^*$ can be checked in the same way.
(2)
$bc=b(1+b^*(a-b))=b+bb^*(a-b)=b-(1-bb^*)(a-b)+(a-b)=b+(a-b)=a$;
(3)
$b^*bc=b^*b+b^*bb^*a-(b^*b)^2$; $cb^*b=b^*b+b^*ab^*b-(b^*b)^2$, so it remains to show that $b^*bb^*a=b^*ab^*b$, which holds true: $b^*(bb^*)a=b^*(aa^*)a=b^*a(a^*a)=b^*a(b^*b)$. Also $b^*bc^*=(cb^*b)^*=(b^*bc)^*=c^*b^*b$.
(4)
$(1-b^*b)(c-1)=(1-b^*b)b^*(a-b)=b^*(1-bb^*)(a-b)=0$; $(c-1)(1-b^*b)=b^*(a-b)(1-b^*b)=0$.
\end{proof}
\begin{thm}\label{Thm}
The map $\iota:K_1(A)\to L_1(A)$ is a natural isomorphism for any unital $C^*$-algebra $A$.
\end{thm}
\begin{proof}
For a balanced pair $(a,b)$, set $\kappa(a,b)=c(a,b)$. Then it gives rise to a homomorphism $\kappa:L_1(A)\to K_1(A)$. We shall show that $\kappa$ is the inverse map for $\iota$.
Let $u$ be a unitary. Then $c(u,1)=1+(u-1)=u$, hence $\kappa\circ\iota([u])=[c(u,1)]=[u]$, so $\kappa\circ\iota=\id_{K_1(A)}$.
Let $(a,b)$ be a balanced pair. Then $\iota\circ\kappa([(a,b)])=[(c(a,b),1)]$. We have to check that $[(c(a,b),1)]=[(a,b)]$. Equivalently, we check that the pair $\left(\left(\begin{smallmatrix}c&\\0&b\end{smallmatrix}\right),\left(\begin{smallmatrix}1&\\&a\end{smallmatrix}\right)\right)$ is homotopy trivial, where $c=c(a,b)$, $a=bc$.
Set $A=\left(\begin{smallmatrix}c&0\\0&b\end{smallmatrix}\right)$, $B_t=\left(\begin{smallmatrix}1&0\\0&b\end{smallmatrix}\right)U_t^*\left(\begin{smallmatrix}1&0\\0&c\end{smallmatrix}\right)U_t$, where $U_t$ is given by (\ref{U}).
As $B_0=\left(\begin{smallmatrix}1&0\\0&a\end{smallmatrix}\right)$, $B_{\pi/2}=A$, and as the pair $(A,A)$ is trivial, so it remains to check that the pair $(A,B_t)$ is balanced for any $t$.
\begin{eqnarray*}
B_t^*B_t&=&U_t^*\left(\begin{matrix}1&0\\0&c^*\end{matrix}\right)U_t\left(\begin{matrix}1&0\\0&b^*b\end{matrix}\right)U_t^*
\left(\begin{matrix}1&0\\0&c\end{matrix}\right)U_t\\
&=&U_t^*\left(\begin{matrix}1&0\\0&c^*\end{matrix}\right)U_t\left(\left(\begin{matrix}1&0\\0&1\end{matrix}\right)-\left(\begin{matrix}0&0\\0&1-b^*b\end{matrix}\right)\right)U_t^*
\left(\begin{matrix}1&0\\0&c\end{matrix}\right)U_t\\
&=&\left(\begin{matrix}1&0\\0&1\end{matrix}\right)-U_t^*\left(\begin{matrix}1&0\\0&c^*\end{matrix}\right)U_t\left(\begin{matrix}0&0\\0&1-b^*b\end{matrix}\right)U_t^*
\left(\begin{matrix}1&0\\0&c\end{matrix}\right)U_t.
\end{eqnarray*}
Note that
$$
C=
U_t^*\left(\begin{matrix}1&0\\0&c\end{matrix}\right)U_t=\left(\begin{matrix}1&0\\0&1\end{matrix}\right)+
\left(\begin{matrix}\sin^2t&\sin t\cos t\\\sin t\cos t&\cos^2 t\end{matrix}\right)(c-1),
$$
and, as $(1-b^*b)(c-1)=0$, so we have
$$
C^*\left(\begin{matrix}0&0\\0&1-b^*b\end{matrix}\right)C=\left(\begin{matrix}0&0\\0&1-b^*b\end{matrix}\right),
$$
so
$$
B_t^*B_t=\left(\begin{matrix}1&0\\0&1\end{matrix}\right)-\left(\begin{matrix}0&0\\0&1-b^*b\end{matrix}\right)=
\left(\begin{matrix}1&0\\0&b^*b\end{matrix}\right)=A^*A.
$$
The other relations are checked in the same way.
\end{proof}
\section{Nonunital case}
\begin{lem}
Let $A$ be a non-unital $C^*$-algebra, $A^+$ its unitalization. Then the inclusion $A\subset A^+$ induces an isomorphism $L_1(A)\to L_1(A^+)$.
\end{lem}
\begin{proof}
Let us prove surjectivity first. Since $L_1(A^+)\cong K_1(A^+)$, any element of $L_1(A^+)$ is, for some $n$, of the form $(u,1)$, where $u\in M_n(A^+)$ is unitary, $1=1_n\in M_n(A^+)$ is the unit, and without loss of generality we may assume that $u-1\in M_n(A)$. Set $A_n=M_n(A)$. Then $u,1\in A_n^+$ and $u-1\in A_n$. Fix $\delta\in(0,1/3)$ and let $f\in C(\mathbb S^1)$ be a continuous function with the following properties:
\begin{enumerate}
\item
$|f(z)|=1$ for any $z\in\mathbb S^1$;
\item
$f(z)=1$ for any $z\in\mathbb S^1$, for which $|z-1|<\delta$;
\item
$|f(z)-z|<\delta$ for any $z\in\mathbb S^1$.
\end{enumerate}
Then $f(u)\in A_n^+$ is unitary, $f(u)-1\in A_n$, and $\|f(u)-u\|<\delta$, hence $[f(u)]=[u]$ in $K_1(A^+)$ and $[(f(u),1)]=[(u,1)]$ in $L_1(A^+)$.
Let $g\in C(\mathbb S^1)$ satisfy the properties
\begin{enumerate}
\item
$g(z)=1$ for any $z\in\mathbb S^1$, for which $|z-1|>\delta$;
\item
$|g(z)|\leq 1$ for any $z\in\mathbb S^1$;
\item
$g(1)=0$,
\end{enumerate}
and let $g_t(z)=t+(1-t)g(z)$, $t\in[0,1]$. Then the pair $(f(u)g(u),g(u))$ is balanced for any $t\in[0,1]$ (all calculations are in $C(X)$: $|(fg)(z)|=|g(z)|=1$ when $|z-1|>\delta$, and $(fg)(z)=g(z)$ when $|z-1|<\delta$), hence $[(u,1)]=[(f(u)g(u),g(u))]$ in $L_1(A^+)$. But $g(u)\in A_n$, so $[(f(u)g(u),g(u))]\in L_1(A)$.
Now let us prove injectivity. Take $[(a,b)]\in L_1(A)$. Let $c$ be as in Lemma \ref{c}, and set $\tilde a=bf(c)$. Then $[(a,b)]=[(\tilde a,b)]$.
Set $A=\left(\begin{matrix}\tilde a&0\\0&g(c)\end{matrix}\right)$, $B_t=\left(\begin{matrix}b&0\\0&g(c)\end{matrix}\right)U_t^*\left(\begin{matrix}f(c)&0\\0&1\end{matrix}\right)U_t$, where $U_t$ is given by (\ref{U}). Direct calculation allows to check that
\begin{enumerate}
\item
the pair $(A,B_t)$ is balanced for any $t$;
\item
$B_0=A$; $B_{\pi/2}=\left(\begin{matrix}b&0\\0&f(c)g(c)\end{matrix}\right)$;
\item
$B_t$ is a matrix with coefficients in $A$ (not in $A^+$).
\end{enumerate}
The first claim can be proved as in Theorem \ref{Thm}, and the other claims are trivial.
It follows that $[(A,B_{\pi/2})]=0$ in $L_1(A)$. As $A=\tilde a\oplus g(c)$, $B_{\pi/2}=b\oplus f(c)g(c)$, so $[(\tilde a,b)]+[(g(c),f(c)g(c))]=0$, i.e. $[(\tilde a,b)]=[(f(c)g(c),g(c))]$ in $L_1(A)$.
Now suppose that $[c]=0$ in $K_1(A)\cong L_1(A^+)$. Then, passing to matrices of greater size if necessary, we can connect $c$ and 1 by a path of unitaries $c_t$, $t\in[0,1]$. The pair $(f(c_t)g(c_t),g(c_t))$ is balanced for each $t\in[0,1]$, and, by definition, $(fg)(1)=g(1)=0$, hence $[(a,b)]=[(0,0)]$, which proves injectivity.
\end{proof}
\begin{cor}
The groups $K_1(A)$ and $L_1(A)$ are naturally isomorphic for any $C^*$-algebra $A$.
\end{cor}
\section{Some examples}
Let us consider the case $A=\mathbb C$. It is easy to see that the following holds.
\begin{lem}
Let $(a,b)$ be a balanced pair of matrices, i.e. of automorphisms of a finitedimensional space $V$. There exists a decomposition $V=V_1\oplus V_2$, $V_2=V_1^\perp$, such that $1-a^*a$ and $1-aa^*$ have the form $\left(\begin{smallmatrix}\ast&0\\0&0\end{smallmatrix}\right)$, and $a-b$ has the form $\left(\begin{smallmatrix}0&0\\0&\ast\end{smallmatrix}\right)$ with respect to this decomposition.
\end{lem}
But even in this case it is not so easy to describe all balanced pairs of numerical matrices. In general case, the structure of balanced pairs may be even more complicated. Here we give two examples for the case when $A$ is commutative.
\begin{example}\label{nontrivial}
Let $A=M_2(C(\mathbb S^1))$ be the algebra of $2{\times}2$-matrix-valued functions on a circle with the coordinate $t$, $t\in[0,\pi/2]$. Set $U(t)=\left(\begin{matrix}\cos t&-\sin t\\\sin t&\cos t\end{matrix}\right)$, $\alpha,\beta,\gamma:[0,\frac{\pi}{2}]\to\mathbb C$ with $\alpha(0)=\alpha(\frac{\pi}{2})=\beta(0)=\beta(\frac{\pi}{2})=\gamma(0)=\gamma(\frac{\pi}{2})=1$, $|\alpha(t)|=|\beta(t)|=1$, $|\gamma(t)|< 1$ for all $t\in(0,\frac{\pi}{2})$,
$$
a(t)=U(t)^*\left(\begin{matrix}\alpha(t)&0\\0&\gamma(t)\end{matrix}\right)U(t), \qquad b(t)=U(t)^*\left(\begin{matrix}\beta(t)&0\\0&\gamma(t)\end{matrix}\right)U(t).
$$
The pair $(a,b)$ is balanced. Note that although $a(t)$ and $b(t)$ are diagonal at each $t$, $a$ and $b$ cannot be diagonalized as elements of $M_2(C(\mathbb S^1))$ (the eigenvectors of $a(t)$ and $b(t)$ cannot be continuous at $0$).
\end{example}
\begin{example}
Let $X$ be a $2n$-dimensional manifold with boundary $\partial X=\mathbb S^{2n-1}$ and let $Y=X\cup \mathbb D^{2n}$ be $X$ glued with the $2n$-dimensional disc over the common boundary. Let $a,b:X\to \mathbb U_N$ be two maps into the unitary group of order $N$, such that $c=a|_{\partial X}=b|_{\partial X}:\mathbb S^{2n-1}\to\mathbb U_N$ represents a non-trivial element of $\pi_{2n-1}(\mathbb U_N)$. Then $c$ does not extend to a map $\mathbb D^{2n}\to\mathbb U_N$, but easily extends to a map $\bar c$ from $\mathbb D^{2n}$ to the set of $N$-dimensional matrices of norm $\leq 1$. Set
$$
\bar a(x)=\left\lbrace \begin{array}{rl}a(x)&\mbox{if\ }x\in X;\\
\bar c(x)&\mbox{if\ }x\in\mathbb D^{2n}\end{array}\right.;
\qquad
\bar b(x)=\left\lbrace \begin{array}{rl}b(x)&\mbox{if\ }x\in X;\\
\bar c(x)&\mbox{if\ }x\in\mathbb D^{2n}.\end{array}\right.;
$$
Then the pair $(\bar a,\bar b)$ is balanced in $M_N(C(Y))$.
\end{example}
\section{$\mathbb K$-balanced pairs of operators}
For an operator $A$ on a Hilbert space $H$, let $\dot A$ denote the class of $A$ in the Calkin algebra.
Let $A$, $B$ be operators on a Hilbert space $H$. We call the pair $(A,B)$ balanced modulo compacts ($\mathbb K$-balanced) if the pair $(\dot A,\dot B)$ is balanced in the Calkin algebra. This means that the relations (\ref{rel1}) and (\ref{rel2}) hold modulo compacts.
To study properties of $\mathbb K$-balanced pairs of operators we need the following corollaryof theKasparov's technical theorem \cite{KaspTT}. A set $X_1,\ldots,X_n$ of operators is symmetric if for every $i=1,\ldots,n$, $X_i^*$ is contained in this set.
\begin{lem}\label{Kasparov}
Let $\overline{X}=\{X_1,\ldots,X_n\}$ and $\overline{Y}=\{Y_1,\ldots,Y_m\}$ be two symmetric sets of contractions on a Hilbert space $H$. Suppose that $X_iY_j$ is compact for any $i$ and $j$. Then, for any $\varepsilon>0$, there exists a projection $P$ in $H$ such that
\begin{enumerate}
\item
$\|PX_i-X_i\|<\varepsilon$ for $i=1,\ldots,n$;
\item
$\|\dot P\dot Y_j\|<\varepsilon$ for $j=1,\ldots,m$.
\end{enumerate}
\end{lem}
\begin{proof}
Let $\mathcal A$ and $\mathcal B$ be the $C^*$-subalgebras in the Calkin algebra $\mathbb Q(H)$ generated by the sets $\dot X_1,\ldots,\dot X_n$ and $\dot Y_1,\ldots,\dot Y_m$ respectively. By the Kasparov's technical theorem, there exists $m\in\mathbb Q(H)$ such that $ma=a$, $mb=0$ for any $a\in\mathcal A$ and any $b\in\mathcal B$. Let $M\in\mathbb B(H)$ be a lift for $m$, i.e. an operator on $H$ such that $\dot M=m$. Without loss of generality we may assume that $M$ satisfies $0\leq M\leq 1$.
Let $Q=E_{(1-\varepsilon,1]}(M)$ be the spectral projection of $M$ corresponding to the set $(1-\varepsilon,1]$.
As $Q\leq M+\varepsilon 1$, it follows from $\dot M\dot Y=0$ that $\|\dot Q\dot Y\|<\varepsilon$. The latter estimate will not change if we replace $Q$ by its compact perturbation $P$.
Let us write operators on $H$ as matrices with respect to the decomposition $H=QH\oplus(1-Q)H$. Then $M=\left(\begin{matrix}M_1&0\\0&M_2\end{matrix}\right)$ with $(1-\varepsilon)1\leq M_1\leq 1$ and $0\leq M_2\leq (1-\varepsilon)1$.
By Kasparov's technical theorem, $MX-X$ is compact for any $X$ from $\overline{X}$. Write $X$ as $X=\left(\begin{matrix}x_{11}&x_{12}\\x_{21}&x_{22}\end{matrix}\right)$. Then
$$
X-MX=\left(\begin{matrix}1-M_1&0\\0&1-M_2\end{matrix}\right)\left(\begin{matrix}x_{11}&x_{12}\\x_{21}&x_{22}\end{matrix}\right)
$$
is compact. Since $1-M_2$ is invertible, $x_{21}$ and $x_{22}$ are compact. It follows from symmetricity of $\overline{X}$ that $x_{12}$ is compact as well. So,
$$
QX-X=-\left(\begin{matrix}0&x_{12}\\x_{21}&x_{22}\end{matrix}\right),
$$
therefore, for any $\varepsilon>0$ there is a projection $Q_0$ onto a finitedimensional subspace $H_0\subset (1-Q)H$ such that $\|Q_0x_{21}-x_{21}\|<\varepsilon/3$ and $\|Q_0x_{22}-x_{22}\|<\varepsilon/3$ for all $X$ from $\overline{X}$. Set
$P=Q+Q_0$. Then $P$ is a compact perturbation of $Q$, and $\|PX-X\|<\varepsilon$ for each $X\in\overline{X}$.
\end{proof}
\begin{thm}\label{H}
Let $(A,B)$ be a $\mathbb K$-balanced pair of operators on a Hilbert space $H$. For any $\varepsilon>0$ there exists a decomposition $H=H_1\oplus H_2$ with the following properties, where we write operators on $H$ as matrices with respect to this decomposition.
\begin{enumerate}
\item
$\|A_{ij}-B_{ij}\|<\varepsilon$ for any $(i,j)\neq (1,1)$;
\item
$C_{ij}$ is of the form $D+K$ with $\|D\|<\varepsilon$ and $K$ compact for any $(i,j)\neq(2,2)$, where $C$ is one of the four operators: $1-A^*A$, $1-AA^*$, $1-B^*B$, $1-BB^*$.
\end{enumerate}
\end{thm}
\begin{proof}
Apply Lemma \ref{Kasparov} to the sets $\overline{X}=\{A-B,A^*-B^*\}$ and $\overline{Y}=\{1-A^*A,1-AA^*,1-B^*B,1-BB^*\}$, and set $H_1=PH$, $H_2=(1-P)H$.
\end{proof}
\begin{lem}
Let $(A,B)$ be a $\mathbb K$-balanced pair of operators on a Hilbert space $H=H_1\oplus H_2$, where $H_1$ and $H_2$ satisfy the conclusion of Theorem \ref{H} for some $\varepsilon$. Then
\begin{enumerate}
\item
$\|\dot A_{11}^*\dot A_{11}-\dot B_{11}^*\dot B_{11}\|<2\varepsilon$, $\|\dot A_{11}\dot A_{11}^*-\dot B_{11}\dot B_{11}^*\|<2\varepsilon$;
\item
$\|(\dot B_{11}-\dot A_{11})(1-\dot A_{11}^*\dot A_{11})\|<4\varepsilon$, $\|(\dot B_{11}-\dot A_{11})^*(1-\dot A_{11}\dot A_{11}^*)\|<4\varepsilon$.
\end{enumerate}
\end{lem}
\begin{proof}
Both claims follow from $\dot A^*\dot A=\dot B^*\dot B$.
(1) As $(\dot A^*\dot A)_{11}=\dot A_{11}^*\dot A_{11}+\dot A_{21}^*\dot A_{21}$, $(\dot B^*\dot B)_{11}=\dot B_{11}^*\dot B_{11}+\dot B_{21}^*\dot B_{21}$, so $\|\dot A_{11}^*\dot A_{11}-\dot B_{11}^*\dot B_{11}\|=\|\dot A_{21}^*\dot A_{21}-\dot B_{21}^*\dot B_{21}\|<2\varepsilon$. The second estimate in (1) is proved similarly.
(2) As $(\dot A^*\dot A)_{12}=\dot A_{11}^*\dot A_{12}+\dot A_{12}^*\dot A_{22}$, $(\dot B^*\dot B)_{12}=\dot B_{11}^*\dot B_{12}+\dot B_{12}^*\dot B_{22}$, so $\|A_{11}^*\dot A_{12}-\dot B_{11}^*\dot B_{12}\|<2\varepsilon$. This implies that $\|(A_{11}^*-\dot B_{11}^*)\dot A_{12}\|<3\varepsilon$, hence $\|(A_{11}^*-\dot B_{11}^*)\dot A_{12}\dot A_{12}^*\|<3\varepsilon$.
As $(\dot B-\dot A)^*(1-\dot A\dot A^*)=0$, so $\|(\dot B_{11}-\dot A_{11})^*(1-\dot A\dot A^*)_{11}\|<\varepsilon$. Then
$\|(\dot B_{11}-\dot A_{11})^*(1-\dot A_{11}\dot A_{11}^*)\|<\|(\dot B_{11}-\dot A_{11})^*\dot A_{12}\dot A_{12}^*\|+\varepsilon<4\varepsilon$.
The second estimate in (2) is proved similarly.
\end{proof}
\section{Relative index}
Let $(A,B)$ be a $\mathbb K$-balanced pair, and let the decomposition $H=H_1\oplus H_2$ satisfies the conclusion of Theorem \ref{H} for some $\varepsilon\in(0,1/20)$, i.e.
\begin{enumerate}
\item
$\|A_{ij}-B_{ij}\|<\varepsilon$ for any $(i,j)\neq (1,1)$;
\item
$C_{ij}$ is of the form $D+K$ with $\|D\|<\varepsilon$ and $K$ compact for any $(i,j)\neq(2,2)$, where $C$ is one of the four operators: $1-A^*A$, $1-AA^*$, $1-B^*B$, $1-BB^*$.
\end{enumerate}
In particular, this means that
\begin{enumerate}
\item
$\|A|_{H_2}-B|_{H_2}\|<\varepsilon$;
\item
$A|_{H_1}$ and $B|_{H_1}$ are isometries up to $\varepsilon$ modulo compacts, i.e. $\|\dot X^*\dot X-\dot 1_{H_1}\|<\varepsilon$, where $X$ is either $A|_{H_1}$ or $B|_{H_1}$.
\end{enumerate}
Then we can define a relative index of the pair $(A,B)$ as follows. Let us write operators as matrices with respect to the direct sum $H_1\oplus H_2$. Then $X|_{H_1}=\left(\begin{smallmatrix}X_1\\X_2\end{smallmatrix}\right)$. For convenience we write here $X_1$ instead of $X_{11}$ and $X_2$ instead of $X_{21}$.
Note that $A|_{H_1}$ and $B|_{H_1}$ behave like Fredholm operators, but their ranges may be completely different. To compare them, take one more operator $C=\left(\begin{smallmatrix}C_1\\C_2\end{smallmatrix}\right):H_1\to H$ with the following properties (where $X$ is either $A$ or $B$):
\begin{enumerate}
\item[(C1)]
$\|C_2-X_2\|<\varepsilon$;
\item[(C2)]
$\|\dot C_1^*\dot C_1-\dot X_1^*\dot X_1\|<2\varepsilon$; $\|\dot C_1\dot C_1^*-\dot X_1\dot X_1^*\|<2\varepsilon$;
\item[(C3)]
$\|(\dot C_1-\dot X_1)(1-\dot X_1^*\dot X_1)\|<4\varepsilon$ and $\|(\dot C_1-\dot X_1)^*(1-\dot X_1\dot X_1^*)\|<4\varepsilon$.
\end{enumerate}
Note that such operators $C$ exist. For example, one may take $C=A|_{H_1}$ or $C=B|_{H_1}$.
As $C^*=(C_1^*,C_2^*):H\to H_1$ has range $H_1$, so the compositions $C^*\circ A|_{H_1}$, $C^*\circ B|_{H_1}$ are operators on $H_1$.
\begin{lem}\label{Lemma1F}
Operators $C^*\circ A|_{H_1}$ and $C^*\circ B|_{H_1}$ are Fredholm.
\end{lem}
\begin{proof}
We have $F=C^*\circ A|_{H_1}=C_1^*A_1+C_2^*A_2$. Set $G=C_1^*A_1+A_2^*A_2$, then $\|F-G\|<\varepsilon$. We have $\dot G=\dot C_1^*\dot A_1+1-\dot A_1^*\dot A_1=1+(\dot C_1^*-\dot A_1^*)\dot A_1$. Then, using $\|\dot C_1(1-\dot A_1^*\dot A_1)-\dot A_1(1-\dot A_1^*\dot A_1)\|<4\varepsilon$ and $\|\dot C_1\dot C_1^*-\dot A_1\dot A_1^*\|<2\varepsilon$, we get
\begin{eqnarray*}
\|\dot G^*\dot G-1\|&=&\|\dot A_1^*\dot C_1+\dot C_1^*\dot A_1-2\dot A_1^*\dot A_1+\dot A_1^*\dot C_1\dot C_1^*\dot A_1-\dot A_1^*\dot A_1\dot C_1^*\dot A_1-\dot A_1^*\dot C_1\dot A_1^*\dot A_1+\dot A_1^*\dot A_1\dot A_1^*\dot A_1\|\\
&<&\|\dot A_1^*\dot C_1(1-\dot A_1^*\dot A_1)+(1-\dot A_1^*\dot A_1)\dot C_1^*\dot A_1-2\dot A_1^*\dot A_1+2\dot A_1^*\dot A_1\dot A_1^*\dot A_1\|+2\varepsilon\\
&<&\|\dot A_1^*\dot A_1(1-\dot A_1^*\dot A_1)+(1-\dot A_1^*\dot A_1)\dot A_1^*\dot A_1-2\dot A_1^*\dot A_1+2\dot A_1^*\dot A_1\dot A_1^*\dot A_1\|+10\varepsilon=10\varepsilon.
\end{eqnarray*}
Hence, $\|\dot F^*\dot F-1\|<11\varepsilon$. Similarly, $\|\dot F\dot F^*-1\|<11\varepsilon$, so $F$ is Fredholm. The same proof works for the second operator.
\end{proof}
Define $\ind(A,B)$ by $\ind(A,B)=\ind (C^*\circ A|_{H_1})-\ind (C^*\circ B|_{H_1})$.
\begin{lem}\label{Lemma2F}
$\ind(A,B)$ does not depend on $C$ when $C$ satisfies (C1)-(C3).
\end{lem}
\begin{proof}
Without loss of generality we may take $C_2=A_2$. Then $(C^*\circ A|_{H_1})^\cdot=1+(\dot C_1^*-\dot A_1^*)\dot A_1$ and
$\|(C^*\circ B|_{H_1})^\cdot-(1+\dot C_1^*\dot B_1-\dot A_1^*\dot A_1)\|<\varepsilon$. Let us check that
$$
\ind (C^*\circ A|_{H_1})-\ind (C^*\circ B|_{H_1})=\ind ((A|_{H_1})^*\circ A|_{H_1})-\ind ((A|_{H_1})^*\circ B|_{H_1}).
$$
By multiplicativity of index, this is equivalent to
$$
\ind(C^*\circ A|_{H_1})((A|_{H_1})^*\circ B|_{H_1})=\ind(C^*\circ B|_{H_1})((A|_{H_1})^*\circ A|_{H_1}),
$$
or, as $\|\dot A_1^*\dot A_1+\dot A_2^*\dot A_2-1\|<\varepsilon$, to
\begin{equation}\label{2indices}
\ind(1+(C_1^*-A_1^*)A_1)(1+A_1^*(B_1-A_1))=\ind(1+C_1^*B_1-A_1^*A_1).
\end{equation}
Direct calculation shows that
$$
(1+(\dot C_1^*-\dot A_1^*)\dot A_1)(1+\dot A_1^*(\dot B_1-\dot A_1))=1+\dot C_1^*\dot A_1-\dot A_1^*\dot A_1+\dot A_1^*\dot B_1-\dot A_1^*\dot A_1+(\dot C_1^*-\dot A_1^*)\dot A_1\dot A_1^*(\dot B_1-\dot A_1);
$$
\begin{eqnarray*}
&&\|(1+(\dot C_1^*-\dot A_1^*)\dot A_1)(1+\dot A_1^*(\dot B_1-\dot A_1))-(1+\dot C_1^*\dot B_1-\dot A_1^*\dot A_1)\|\\
&=&\|\dot C_1^*\dot A_1-2\dot A_1^*\dot A_1+\dot A_1^*\dot B_1+(\dot C_1^*-\dot A_1^*)\dot A_1\dot A_1^*(\dot B_1-\dot A_1)-\dot C_1^*\dot B_1+\dot A_1^*\dot A_1\|\\
&\leq&\|\dot C_1^*\dot A_1-2\dot A_1^*\dot A_1+\dot A_1^*\dot B_1+(\dot C_1^*-\dot A_1^*)(\dot B_1-\dot A_1)-\dot C_1^*\dot B_1+\dot A_1^*\dot A_1\|+4\varepsilon=4\varepsilon.
\end{eqnarray*}
Therefore, the operators in the both sides of (\ref{2indices}) coinside up to $4\varepsilon$ modulo compacts, thus they have the same index.
\end{proof}
\begin{cor}
$\ind(A,B)=\ind(1+B_1^*(A_1-B_1))$.
\end{cor}
\begin{proof}
Take $C_1=A_1$. Then
\begin{eqnarray*}
\ind(A,B)&=&\ind((A|_{H_1})^*\circ A|_{H_1})-\ind((A|_{H_1})^*\circ B|_{H1})\\
&=&-\ind(1+A_1^*(B_1-A_1))=\ind(1+(B_1-A_1)^*A_1)\\
&=&\ind(1+B_1^*A_1-A_1^*A_1)=\ind(1+B_1^*A_1-B_1*B_1)\\
&=&\ind(1+B_1^*(A_1-B_1)).
\end{eqnarray*}
\end{proof}
Now let us show that $\ind(A,B)$ does not depend also on the decomposition $H_1\oplus H_2$. Let $F=c(A,B)=1+B^*(A-B)$ be the operator on $H\oplus H$ defined by the formula (\ref{unitary}). Independence on $H_1$ is implied by the following theorem.
\begin{thm}
$\ind(A,B)=\ind(1+B^*(A-B))$.
\end{thm}
\begin{proof}
The proof follows from the estimate
$$
\left\|B^*(A-B)-\left(\begin{matrix}B_1^*(A_1-B_1)&0\\0&0\end{matrix}\right)\right\|<6\varepsilon.
$$
\end{proof}
\begin{remark}
Note that if, by some reason, $A_{21}$ is compact (or small plus compact) then $A|_{H_1}$ and $B|_{H_1}$ are Fredholm even without compositions with $C$, and there is no need in Lemmas \ref{Lemma1F} and \ref{Lemma2F}.
\end{remark}
\section{Application to pseudodifferential operators}
Let $D_1$, $D_2$ be two order zero pseudodifferential operators on a manifold $M$, with symbols $\sigma(D_1)$, $\sigma(D_2)$. Let $\sigma_1$, $\sigma_2$ be the restrictions of $\sigma(D_1)$ and $\sigma(D_2)$ onto the cospherical bundle $S^*M$. If $(\sigma_1,\sigma_2)$ is a balanced pair in matrices over $C(S^*M)$ then $(D_1,D_2)$ (more exactly, their compact perturbations) is a $\mathbb K$-balanced pair of operators. As the pair $(\sigma_1,\sigma_2)$ is balanced, it determines a class in $K^1(S^*M)$. The standard construction allows to define an integer-valued topological index for such pairs.
\begin{thm}
If $(\sigma_1,\sigma_2)$ is balanced in matrices over $C(S^*M)$ then $\ind(D_1,D_2)=\ind(\sigma_1,\sigma_2)$.
\end{thm}
\begin{proof}
If $(\sigma_1,\sigma_2)$ is balanced then compact perturbations of $D_1$ and $D_2$ give a $\mathbb K$-balanced pair of operators ($D_1$ and $D_2$ satisfy the relations (\ref{rel1}) or (\ref{rel2}) modulo compacts, but may not be contractions; the latter can be remedied by compact perturbations). Then the operator $U(D_1,D_2)$ (\ref{unitary}) is pseudodifferential with the symbol $U(\sigma_1,\sigma_2)$, hence $\ind U(D_1,D_2)=\ind U(\sigma_1,\sigma_2)$.
\end{proof}
Let us consider some simple examples.
\begin{example}
Let $M_1$, $M_2$ be two manifolds with the same boundary $\partial M_1=\partial M_2=N$, and let $M=M_1\cup_N M_2$. Let $(\sigma_1,\sigma_2)$ be a balanced pair in $M_n(C(S^*M))$ such that $\sigma_1=\sigma_2$ on $M_2$, and $\sigma_1,\sigma_2$ unitary on $M_1$. Then one can take $H_1=L^2(M_1)$, $H_2=L^2(M_2)$ (if necessary, one can take a finite codimension subspace in $L^2(M_2)$ as $H_2$). The relative index $\ind(D_1,D_2)$ in this case equals the relative index of the restrictions of $D_1$ and $D_2$ on $M_1$.
\end{example}
\begin{example}
Let $p\in M_n(S^*M)$ be a projection, $P$ a pseudodifferential operator with the symbol $p$. Suppose that $p$ commutes with $\sigma_1$ and with $\sigma_2$, and that $(1-p)(\sigma_1-\sigma_2)=0$. Set $\sigma'_1=p\sigma_1 p$, $\sigma'_2=p\sigma_2 p$, and let $D'_1$, $D'_2$ be pseudoddifferential operators with symbols $\sigma'_1$ and $\sigma'_2$ respectively. Then we are in the setting of operators in subspaces \cite{SSS}.
\end{example}
\begin{example}
Let $x_0\in M$, and let $\sigma_1|_{x=x_0}=\sigma_1|_{x=x_0}=0$. Then we are in the setting of elliptic operators on a noncompact manifold, or on a manifold with a singularity at $x_0$.
\end{example}
The examples above reduce to known cases, but the next example seems to be new.
\begin{example}
Let $a(t)$, $b(t)$, $t\in[0,\pi/2]$, be as in Example \ref{nontrivial}. Let $M=\mathbb S^1$ (with the points $0$ and $\pi/2$ glued together), then $S^*M=\{(t,i):t\in\mathbb S^1, i=\pm 1\}$. Set
$$
\sigma_1(t,1)=a(t),\quad \sigma_2(t,1)=b(t);\qquad \sigma_1(t,-1)=\sigma_2(t,-1)=\left(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\right).
$$
Then the relative index for $D_1$, $D_2$ having symbols $\sigma_1$, $\sigma_2$ respectively, is well defined and can be evaluated from the functions $\alpha$ and $\beta$.
\end{example}
|
\section{Introduction}
Cold molecules are intriguing quantum objects with a complex energy-level structure, yet amenable to excellent control over their motional and internal degrees of freedom~\cite{Carr2009,Lemeshko2013}. This renders them particularly suitable for studying low-energy collision dynamics~\cite{Gilijamse2006,Parazzoli2011,Henson2012,Kirste2012,Strebel2012,Hall2012,Tariq2013,Hauser2015} and enabling controlled chemistry~\cite{Tscherbul2006a,Ni2010}. Cold molecules are also beneficial for precision spectroscopy, serving as an important tool for exploring fundamental physics~\cite{Hudson2011,ACME2014,Dietiker2015}. For various applications of cold polar molecules, achieving high purity of and control over their internal states~\cite{Viteau2008,Manai2012,Gloeckner2015}, in addition to having the ability to manipulate their motional behaviour~\cite{Gupta1999,Bethlem1999,Fulton2004,Narevicius2008b,Hogan2009,Sommer2010,Shuman2010,Zeppenfeld2012,Chervenkov2014}, is of paramount importance. In pursuit of this goal, cryogenic buffer-gas cooling has proven to be a very general and powerful method to produce internally and translationally cold molecules~\cite{Weinstein1998,Maxwell2005,vanBuuren2009}.\\
Here we demonstrate the full characterization of the properties of buffer-gas-cooled molecular beams, and, on this basis, elucidate various aspects of buffer-gas cooling. Our characterization relies on the combination of four components. First, we extract cold molecules from a buffer-gas cell by electrostatic guiding in a quadrupole guide~\cite{vanBuuren2009}. Second, time-of-flight measurements are used to derive the longitudinal velocity distribution of the molecules at the end of the guide~\cite{Sommer2009}. Third, resonant radio-frequency (RF) depletion spectroscopy is employed to characterize the internal state distribution of the molecules. Fourth, extensive Monte-Carlo trajectory simulations of the electrostatic guiding allow for the properties of the molecules directly after the buffer-gas cell to be inferred from the signal at the end of the guide.\\
With these tools, we are able to characterize the buffer-gas cooling in various aspects. We demonstrate addressing of all significantly populated internal states of the guided molecules emerging from our cryogenic buffer-gas source, operated in different regimes, boosted~\cite{Maxwell2005,MotschBoosting} or supersonic~\cite{Hutzler2011}. We applied this method to different species of polyatomic symmetric-top polar molecules, fluoromethane, $\rm{CH_3F}$, and 3,3,3-trifluoropropyne, $\rm{CF_3CCH}$, and made comparison of their cooling processes. We provide clear evidence of the efficient control over the internal cooling of the guided molecular beams through varying the buffer-gas-cell temperature and the buffer-gas density, demonstrating the possibility of cooling far below the cell temperature in the supersonic regime. The detailed study of the buffer-gas cooling reveals two interesting phenomena. First, comparing the rotational with the translational temperature, we provide evidence that rotational cooling takes place more efficiently than translational cooling for the CH$_3$F-He system in the low He density regime. Second, the measurements provide indications of dependence of the collisional relaxation rate on the rotational states of the molecules.\\
Our paper is structured as follows. In Section~\ref{Section_Method} we present a brief review of symmetric-top molecules, and then describe the principle of our method for rotational-state detection of guided molecules. Section~\ref{Section_ExpSetup} describes the experimental set-up. The characterization and analysis of the radio-frequency depletion measurements are described in Section~\ref{Section_ExpResults}. The Monte Carlo trajectory simulations used to calculate the guiding efficiencies for all relevant rotational states are described in Section~\ref{Section_MonteCarlo}. Finally, in Section~\ref{Section_Discussion} we present the results from the comprehensive characterization of our buffer-gas source. There we analyze the cell's output for different operating regimes and for different molecular species, and draw inferences on the cooling mechanisms in the buffer-gas cell.
\section{Method for rotational-state detection}
\label{Section_Method}
One of the key features in this work is the implementation of a new method for state detection, which is a pivotal tool for the characterization of cold molecular beams from cryogenic buffer-gas sources. Several techniques for state detection of cold molecules have been used so far~\cite{vanVeldhoven2002,MotschDepletion,Bertsche2010,Barry2011,Patterson2012,Twyman2014,Gloeckner2015a}. Among them, depletion methods, as adopted in previous experiments carried out by our group~\cite{MotschDepletion,Gloeckner2015a}, are particularly advantageous since they are applicable to a large range of molecules and avoid the difficulties of the light-induced fluorescence (LIF) and resonance-enhanced multi-photon ionization (REMPI) detection techniques. Along this line, we have extended the depletion technique by implementing rotational-state detection adapted to the cold guided molecular beams from a cryogenic buffer-gas source, using resonance RF depletion spectroscopy in a parallel-plate capacitor.\\
\subsection{The symmetric-top molecule}
\label{Subsection_SymmTopMol}
To understand our method for rotational-state detection, first we briefly review the relevant properties of symmetric-top molecules, which have been used in the current study. Their rotational states can be fully described by three quantum numbers~\cite{Wollrab1967}, the total angular momentum, $J$, its projection on the molecule's symmetry axis, $K$, ($K=-J, ..., J$), and its projection on a laboratory-fixed axis, $M$, ($M=-J, ..., J$). Hereinafter the rotational states will be designated as $|JKM\rangle$ or $|JK\rangle$, depending on the need to specify the quantum number $M$.\\
In the absence of external electric fields, the energy of a symmetric-top molecule in the rigid-rotor approximation is expressed as $E_{J,K}=h[BJ(J+1)+(A-B)K^2]$, where $h$ is Planck's constant, and $A$ and $B$ are the rotational constants of a symmetric-top molecule. The presence of an external electric field splits every $|JK\rangle$ state (Stark effect) into $2J+1$ $M$-sublevels, corresponding to all possible projections of $\rm{\mathbf{J}}$ on the electric-field axis. The first-order Stark splitting is given by the expression $E^{(1)}=-\mu\mathcal{E}\frac{KM}{J(J+1)}$, where $\mu$ and $\mathcal{E}$ stand for the permanent electric dipole moment of the molecule and the electric-field strength, respectively. Sublevels with a positive Stark shift are referred to as low-field-seeking states (states that are guided). Since states $|J\,-KM\rangle$ are degenerate with states $|JK\,-M\rangle$ under inversion symmetry, we ignore hereinafter the sign of $K$ and adopt the convention that states with positive $M$ are low-field-seeking. An example of the Stark effect is presented in Figure~\ref{Fig_ExpSetup}(a) for the $|1,1\rangle$ and $|2,1\rangle$ states of $\rm{CH_3F}$. For $|JK\rangle$ states with $K=0$, the first-order Stark splitting is zero. In this case, low-field-seeking states due to the second-order (quadratic) Stark shift exist, if they satisfy the condition $J(J+1)>3M^2$, which derives from the expression for the quadratic Stark shift $E^{(2)}=-\frac{\mu^2\mathcal{E}^2}{2hB}\left[\frac{3M^2-J(J+1)}{J(J+1)(2J-1)(2J+3)}\right]$ for $K=0$~\cite{Wollrab1967}. The expression of $E^{(2)}$ for the case $M=0$ is identical to the above one, except the quantum number $M$ in the expression is replaced by $K$. Thus states with $M=0$ can also be low-field-seeking states.\\
\subsection{Radio frequency depletion spectroscopy}
\label{Subsecton_RFDSpectroscopy}
Our method for rotational-state detection is based on state-selective addressing and elimination of molecules from a guided population by applying an RF field resonant to the DC Stark splitting in a homogeneous field. This leads to a depletion of the measured signal and reveals the relative state population.\\
Molecules emerging from the cryogenic source populate different rotational states $|JK\rangle$. Only molecules in low-field-seeking $M$-sublevels, however, are confined in the guide. Subsequently, the guide is interrupted by a parallel-plate capacitor creating a homogeneous offset electric field. At low field, this causes equidistant Stark splitting of the $M$-sublevels for a given $|JK\rangle, K\neq0$ state. The magnitude of the splitting is unique for most of the low-lying $|JK\rangle$ states for a given offset field (see Figure~\ref{Fig_ExpSetup}(a)). This allows for $|JK\rangle$ states to be addressed individually. An RF field which is resonant to the Stark splitting between the $M$-sublevels of the rotational state $|JK\rangle$ we want to address is applied to the capacitor, and transfers the molecules into different $M$-sublevels. Those molecules which land in non-guidable states are eliminated from the guiding after the capacitor, and this leads to a depletion signal on the detector.\\
To achieve a good control over the resulting $M$-substate distribution, we broaden the RF signal with white noise and hence eliminate coherent effects in the population transfer process. The molecules are thereby equally redistributed among all the $M$-sublevels within the given $|JK\rangle$ rotational state (Figure~\ref{Fig_ExpSetup}(a)), when the transitions are power-saturated. To a first approximation, the depletion ratio in this case is given by the fraction of molecules converted from guidable to non-guidable states, and this is the ratio between the number of non-guidable $M$-sublevels and the total number of $M$-sublevels, i.e., $\frac{J+1}{2J+1}$ for a linear Stark shift. The fraction of guided molecules populating the probed rotational state is therefore obtained by dividing the magnitude of the saturated depletion by this ratio. In addition, a decrease in signal results also from transferring molecules from a guidable state with a large Stark shift to one with a small Stark shift, as the latter are more prone to losses in the guide. Thus the simplified depletion ratio needs to be corrected including the subtle effects of guiding efficiency of molecules in different $M$-sublevels, which are described in detail in Section~\ref{Section_MonteCarlo}.
\section{Experimental Set-up}
\label{Section_ExpSetup}
The scheme of the experimental set-up is shown in Figure~\ref{Fig_ExpSetup}(b). In our experiment, buffer-gas cooling takes place in a cryogenic cell where a continuous flow of polyatomic polar molecules is mixed with a continuous flow of helium or neon at cryogenic temperatures of $5\,\rm{K}$ or $18\,\rm{K}$, respectively. The exact cell temperature can be monitored by diode sensors and controlled by an electric heating block. The gas inflow is monitored by pressure gauges and regulated by needle valves along the corresponding gas lines. After the cold molecules leave the cell, they are captured by a quadrupole guide, which delivers them to a detector or to further experiments. The experimental set-up is an upgraded version of the set-up described previously~\cite{Sommer2009} to enable RF depletion spectroscopy on the guided molecules.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig1_StarkSplittingExamples_CryoRFSchematic_InitSubmission}
\caption{(a) Stark splittings as a function of the applied electric field for two rotational states of $\rm{CH_3F}$, $|1,1\rangle$ and $|2,1\rangle$: red lines show guidable sublevels, and blue lines show non-guidable ones. The $|2,1,0\rangle$ state is weakly guidable due to its second-order Stark shift. The vertical dashed line designates the applied homogeneous electric field ($220\,\rm{V\,cm^{-1}}$) used in the radio-frequency scans shown in Figure~\ref{Fig_RFSpectroscopy}(a). The arrows show the magnitude of the Stark splitting corresponding to the applied electric field. This splitting is unique for most low-lying $|JK\rangle$ states. (b) Scheme of the experimental set-up.}
\label{Fig_ExpSetup}
\end{figure}
Our buffer-gas set-up can be operated in two different regimes, boosted and supersonic. The supersonic regime is achieved for buffer-gas densities typically one to two orders of magnitude higher compared to those used in the boosted regime, and is characterized by a large number of collisions at the cell output orifice and within a short distance downstream from it. The high buffer-gas densities result in a high pressure gradient across the nozzle, and consequently, in an adiabatic expansion, leading to further cooling of the gas emerging from the cell~\cite{Hutzler2011}. The boosted regime is an intermediate between the supersonic and the effusive (i.e., collisionless) regime, where a few collisions occur near the nozzle, and thereby the slowest molecules are eliminated from the beam. The two regimes require different geometries. For the boosted regime, the guide is distanced about $2\,\rm{mm}$ from the cell nozzle. In the case of the supersonic regime, that spacing is {\it ca.} $20\,\rm{mm}$ to allow enough distance for the supersonic beam formation. The cell exit aperture used in our experiments has a diameter of $2\,\rm{mm}$. The maximum electric field that can be achieved between the guiding electrodes is about $100\,\rm{kV}\,\rm{cm^{-1}}$. We use a quadrupole mass spectrometer (QMS) to detect the molecules after they leave the guide.\\
The capacitor for driving RF transitions is built of two parallel ($2\,\rm{cm}\times2\,\rm{cm}$) plates with a $2.75\,\rm{mm}$ spacing between them. A bias voltage from a few tens of volts up to a few thousand volts is applied to the capacitor, providing a homogeneous electrostatic field between the plates. The capacitor is separated from the two guide segments by gaps of $1\,\rm{mm}$. The plates are made from circuit boards. One of the plates is a plain board, whereas the other one features a T-shaped microstrip used to apply the RF field. As the flat end of the T-shaped microstrip is aligned with the molecules' flow axis, the RF field is expected to be predominantly perpendicular to the homogeneous field in the molecules' transit region. This configuration enables driving of $\Delta M=\pm1$ transitions.\\
\section{Radio frequency depletion measurements for state detection}
\label{Section_ExpResults}
In this Section we present the characterization and the analysis of radio-frequency depletion measurements of buffer-gas-cooled and electrically guided molecules, and the derivation of their rotational-state distribution based on these measurements. Figure~\ref{Fig_RFSpectroscopy}(a) shows the relative signal of guided helium-buffer-gas-cooled fluoromethane, $\rm{CH_3F}$, as a function of the applied RF frequency, for different buffer-gas densities in the cell, $3.7\times10^{14}\rm{cm^{-3}}$, $1.3\times10^{15}\rm{cm^{-3}}$, and $3.2\times10^{15}\rm{cm^{-3}}$, and for a cell temperature of $6.7\,\rm{K}$. The applied homogeneous electric field of $220\,\rm{Vcm^{-1}}$ was the same for all three scans. The linear Stark splittings for different $|JK\rangle$ rotational states of fluoromethane have been calculated for the applied homogeneous electric field, and, on this basis, the observed features in the depletion spectrum have been assigned. The assignments of the most prominent dips are shown in the figure. The broadening of the lines is attributed to the inhomogeneity of the electric field, which leads to a linewidth proportional to the scanning frequency.\\
Figure~\ref{Fig_RFSpectroscopy}(a) shows a clear change in the dip pattern and in the relative dip depth as the buffer-gas density is changed. Increasing the buffer gas density in the cell increases the number of collisions and therefore leads to a lower rotational temperature. This effect is clearly visible, as the depletion signal corresponding to the low-lying rotational states $|1,1\rangle$ and $|2,1\rangle$ increases, while the signal from the other states of higher rotational energy decreases, and eventually vanishes.\\
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig2_RFSpectroscopy_PowerScan_TwoSubfigures_InitSubmission}
\caption{(a) RF depletion spectra of $\rm{CH_3F}$ for different He densities in the buffer-gas cell ($T_{\rm{cell}}=6.7\,\rm{K}$). The most prominent dips are assigned to the corresponding $|JK\rangle$ rotational states. (b) Depletion saturation measurement for the $|1,1\rangle$ state of $\rm{CH_3F}$. The fit to the experimental points is a sum of an exponential and a linear power dependence (For details, see text).}
\label{Fig_RFSpectroscopy}
\end{figure}
The magnitude of the depletion signal depends not only on the thermal population of the probed state, but also on the applied RF power. To eliminate this effect, we performed RF power scans for each $|JK\rangle$ rotational state to obtain its saturated depletion level. Coherent effects, which lead to Rabi oscillations between the $M$-sublevels and therefore hinder an equal redistribution of population among all $M$-sublevels, were eliminated by broadening the RF signal to $5\,\rm{MHz}$ with white noise (the white noise is not applied in the frequency scan in Figure~\ref{Fig_RFSpectroscopy}(a)).
An example of an RF power scan is shown in Figure~\ref{Fig_RFSpectroscopy}(b) for the $|1,1\rangle$ state of $\rm{CH_3F}$, where the relative signal is plotted as a function of the applied RF power spectral density. The data were fitted with a sum of an exponential and a linear function. The exponential term is attributed to resonant driving of the transition of interest, while the linear decay term is attributed to off-resonant driving of transitions in other states taking place in the non-homogeneous-field regions near the quadrupole guide, as well as to transitions driven by higher harmonics produced by the RF amplifier that are not perfectly filtered out. The sought value of the saturated depletion for each of the considered rotational states is taken to be the amplitude of the exponential function. For $\rm{CH_3F}$ we have performed saturation scans only for states with $J\leq4$ under the reasonable assumption that states with higher $J$ rotational number have a negligible thermal population below $10\,\rm{K}$.\\
Results from the RF depletion measurements of $\rm{CH_3F}$ for $T_{\rm{cell}}=6.4\,\rm{K}$ and $n_{\rm{He}}=1.6\times10^{15}\,\rm{cm^{-3}}$ are shown in Table~\ref{Table_Populations}. The values listed in the second column show the measured depletion signal for a given rotational state $|JK\rangle$ from the RF depletion scans. The nominal depletion ratio, $\frac{J+1}{2J+1}$ (discussed in Section~\ref{Subsecton_RFDSpectroscopy}), is listed in the third column. The more accurate simulated depletion ratio obtained from the trajectory simulations described in detail in the next Section is listed in the fourth column of Table~\ref{Table_Populations}. The relative population of each of the rotational states $|JK\rangle$ present in the guided beam can be deduced from the ratio of the measured depletion signal and the simulated depletion ratio. Note that the sum of the state populations is $\sim 100\,\%$, indicating our ability to account for the entire population of rotational states of $\rm{CH_3F}$ present in the guided beam.\\
\begin{table}
\centering
\newcolumntype{d}[1]{D{.}{\cdot}{#1} }
\begin{tabular}{ld{2}lld{1}}
\toprule \toprule
\multicolumn{1}{c}{State}&\multicolumn{1}{c}{Depletion}&\multicolumn{1}{c}{Nominal DR}&\multicolumn{1}{c}{Simulated DR}&\multicolumn{1}{c}{Population}\\
\multicolumn{1}{c}{$|JK\rangle$}&\multicolumn{1}{c}{(\%)}&&\multicolumn{1}{c}{(\%)}&\multicolumn{1}{c}{(\%)}\\
\midrule
$|1,1\rangle$&$16.96\,(0.14)$&$2/3$& $66.7$ &$25.4\,(0.2)$\\
$|1,1\rangle$&$16.96\,(0.14)$&$2/3$& $66.7$ &$25.4\,(0.2)$\\
$|2,1\rangle\,[|3,2\rangle$]&$19.07\,(0.11)$&$3/5\,[4/7]$& $57.3\,[58.8]$ &$32.8\,(0.2)$\\
$|3,1\rangle$&$5.08\,(0.12)$&$4/7$& $54.3$ &$9.4\,(0.2)$\\
$|4,1\rangle$&$1.02\,(0.13)$&$5/9$& $55.6$ &$1.8\,(0.2)$\\
$|2,2\rangle$&$1.88\,(0.11)$&$3/5$& $62.2$ &$3.0\,(0.2)$\\
$|1,0\rangle$&$12.32\,(0.15)$&$2/3$& $66.7$ &$18.5\,(0.2)$\\
$|2,0\rangle$&$5.6 \,(0.4)$&$2/3$& $66.0$ &$8.5\,(0.6)$\\
\midrule
Total&$62.6 \,(0.5)$& & &$\bf{99.4\,(0.8)}$\\
\bottomrule \bottomrule
\end{tabular}
\caption{Experimental relative populations of the $|JK\rangle$ rotational states of $\rm{CH_3F}$ derived from saturated depletion signals measured for each of the rotational states at $T_{\rm{cell}}=6.4\,\rm{K}$ and $n_{\rm{He}}=1.6\times10^{15}\,\rm{cm^{-3}}$. Nominal and simulated depletion ratios (DR), are defined and explained in the text. The values in parentheses of the depletion and the population columns are the statistical errors. The $|3,2\rangle$ state in the square brackets in the second row shares the same Stark splitting with the $|2,1\rangle$ state, and its contribution to the depletion signal is also taken into account. The corresponding depletion ratios for the $|3,2\rangle$ state are also given in square brackets in the third and fourth column.}
\label{Table_Populations}
\end{table}
\section{Monte Carlo trajectory simulations and guiding efficiencies}
\label{Section_MonteCarlo}
To characterize the buffer-gas source and to quantify the effect of buffer-gas cooling for different cell operating regimes, it is necessary to determine the rotational-state and velocity distributions of the molecules upon emerging from the cryogenic buffer-gas cell. To retrieve this information from the measurements at the end of the guide, we perform comprehensive Monte Carlo trajectory simulations, which yield the guiding efficiencies for the molecules in the quadrupole guide. In this Section, we describe the basic principle and some crucial elements of the simulations.\\
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig3_StarkCurves_BarPlot_InitSubmission}
\caption{(a) Normalized exact Stark-shift curves as a function of the electric field for states with different types of Stark-shift curvatures: convex ($|1,1,1\rangle$ rotational state of $\rm{CH_3F}$) and concave ($|2,0,0\rangle$ rotational state of $\rm{CH_3F}$). They were obtained based on diagonalizing the rigid-rotor Hamiltonian. The linear approximation to the Stark-shift curve is shown for comparison (black dashed line). (b) Demonstration of the necessity of using the exact Stark-shift curves: comparison between the relative depletion signals for simulation I , simulation II, and the measurement. The simulated depletion signals are taken at the fitted $T_{\rm{rot}}$, with the fitting procedure explained in details in Section~\ref{Subsection_TrotTcell}. The major difference between simulations I and II is that the former uses only linear approximations to the Stark curves while the latter uses the exact Stark curves. Very good agreement with the experimental results is obtained only for simulation II.}
\label{Fig_StarkCurves}
\end{figure}
The electric-field distribution along the entire molecule guide is calculated using the SIMION software package, and, on this basis, the molecular trajectories from the cell output to the QMS are simulated. In particular, this includes the two 20~cm-radius 45-degree bends, the capacitor plates, and the divergence of the molecular beam between the end of the guide and the QMS, as shown in Figure~\ref{Fig_ExpSetup}(b). To determine the longitudinal velocity distributions, experiments and simulations are also performed for the case of a continuous guide, in which the capacitor plates are replaced by a straight piece of guide, as here the capacitor plates are not necessary. The effect of applying an RF field to the capacitor plates is simulated by evenly redistributing the population of a given $|JK\rangle$ state across all of its $M$-sublevels in the middle of the capacitor. The ratio of the simulated signals with and without reshuffling the population across the M-sublevels gives the simulated depletion ratio.\\
In the following, we describe the initial conditions and the method for the trajectory simulations. The transverse spatial distribution of the molecules at the guide input is taken to be flat, and the transverse energy distribution is taken as the Boltzmann distribution ($\sim T_{\rm{cell}}$) truncated by the transverse acceptance of the guide ($\sim1K$ trap depth). The initial longitudinal velocity, $v_{\parallel}$, is varied, allowing a guiding efficiency as a function of $v_{\parallel}$ to be extracted from the simulation. Dividing the measured $v_{\parallel}$-distribution after the guide by the corresponding guiding efficiencies yields the initial $v_{\parallel}$-distribution before the guide. Assuming this initial $v_{\parallel}$-distribution is the same for all internal states, we simulate the guiding efficiency for every relevant $|JKM\rangle$ state with the above-determined initial $v_{\parallel}$-distribution taken as an input. Once these state-dependent guiding efficiencies become available, we could recover the initial rotational-state distribution before the guide (see Figure~\ref{Fig_LongVeloDistrib}(a)) from the depletion measurements described in Section~\ref{Section_ExpResults}. One subtlety here is that the simulations for the $v_{\parallel}$ distribution and for the rotational-state distribution need each other's results as their inputs. This interdependence can be resolved by realizing that the $v_{\parallel}$-dependent guiding efficiency is only weakly influenced by the state composition, and the two simulations can be performed iteratively. In practice, we start with a state distribution corresponding to the Boltzmann distribution at $T_{\rm{cell}}$ in order to determine the initial $v_{\parallel}$-distribution at the guide input. This in turn helps us calculate the theoretical depletion values for various states between the capacitor plates, assuming a Boltzmann distribution in the cell. By fitting the measured RF depletion data to the so-obtained theoretical values, we can determine the rotational temperature of the molecules exiting the cell. This last assumption of reaching a Boltzmann distribution in the cell at a fitted temperature is validated by the results shown in Section~\ref{Subsection_TrotTcell}.\\
To achieve a good agreement between the simulated and the experimental results, a key requirement is to take into account the exact dependence of the Stark shift on the electric field, rather than using a linear or quadratic approximation. As shown in Figure~\ref{Fig_StarkCurves}(a), the Stark shift for certain states deviates strongly from a linear approximation for fields up to 100~kV/cm, as present in the experiment. Employing the exact Stark-shift curve dramatically improves the match between the measured and simulated depletion signal, as shown in Figure~\ref{Fig_StarkCurves}(b). In particular, including the exact Stark curve decreases the simulated depletion signal for states with a convex Stark shift and increases the simulated depletion signal for states with a concave one, with a less pronounced effect for states with a Stark shift more closely following a linear behaviour, as might be expected.\\
\section{Analysis and discussion}
\label{Section_Discussion}
Sections~\ref{Section_ExpResults} and \ref{Section_MonteCarlo} have described the measurement and the simulation methods. In this Section we apply these tools for a comprehensive characterization of the buffer-gas-cooling process in the cryogenic cell. First, we explore in Sections~\ref{Subsection_TrotTcell} and \ref{Subsection_TrotBGdensity} the influence of cell parameters on the buffer-gas cooling. Next, we investigate in Sections~\ref{Subsection_Trifluropropyne} and \ref{Subsection_Supersonic} the cooling of different molecular species and in the supersonic operating regime of the cell. At the end, in Sections~\ref{Subsection_TrotTtrans} and \ref{Subsection_RotStateDependence} we discuss and draw conclusions on the cooling effect for different degrees of freedom and for different rotational states.\\
\subsection{Dependence of the rotational temperature on the cell temperature}
\label{Subsection_TrotTcell}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig4_FittedCurvesCorrected_TrotTcell_InitSubmission}
\caption{(a) Measured depletion signals (data points) for different rotational states $|JK\rangle$ of $\rm{CH_3F}$ and different buffer-gas cell temperatures at a fixed He flux of $\Phi_{\rm{He}}=2\times10^{17}\,\rm{s^{-1}}$ and simulated depletion signals (lines) for different rotational temperatures for different rotational states. The grey columns group sets of measurements performed at the same $T_{\rm{cell}}$. Measurements at five different cell temperatures have been performed. The rotational temperature of the molecules leaving the buffer-gas cell was determined by columnwise simultaneous fitting of the measured depletion signals to the calculated curves. (b) Relation between the measured cell temperature, $T_{\rm{cell}}$, and the fitted rotational temperature, $T_{\rm{rot}}$, for $\rm{CH_3F}$.}
\label{Fig_TrotFit}
\end{figure}
The first aspect we consider is the dependence of the rotational temperature on the buffer-gas cell temperature. Here, we would like to address two questions: first, whether the rotational states of molecules obey a Boltzmann distribution in the cell; second, how $T_{\rm{rot}}$ compares to $T_{\rm{cell}}$. Figure~\ref{Fig_TrotFit}(a) shows the measured depletion (in percentage of the total signal) for different rotational states $|JK\rangle$ of $\rm{CH_3F}$ for different cell temperatures, $T_{\rm{cell}}$, along with their theoretical predictions (solid curves). The measurements were done at a fixed He flux of $\Phi_{\rm{He}}=2\times10^{17}\,\rm{s^{-1}}$ (corresponding to a density of $n_{\rm{He}}=1.6\times10^{15}\,\rm{cm^{-3}}$ at $T_{\rm{cell}}=6.4\,\rm{K}$ for our cell geometry). The data for each $T_{\rm{cell}}$ are grouped in a separate vertical column in Figure~\ref{Fig_TrotFit}(a). The theory curves are based on the assumption of a Boltzmann distribution for the rotational states at the cell (as mentioned in Section~\ref{Section_MonteCarlo}), with $T_{\rm{rot}}$ being an independent variable on the horizontal axis. The guiding efficiencies are simulated using the method described in Section~\ref{Section_MonteCarlo}. In particular, we have included all 64 guidable $|JKM\rangle$ states of $\rm{CH_3F}$ (rotational constants $A=155\,\rm{GHz}$ and $B=26\,\rm{GHz}$), with rotational energies below $72\,\rm{K}$ ($50\,\rm{cm^{-1}}$). For a source at $10\,\rm{K}$ those states account for $99.7\,\%$ of the thermal population. Note that $T_{\rm{rot}}$ is the only fit parameter in Figure~\ref{Fig_TrotFit}(a), which sets the horizontal position of each column via a least-$\chi^2$ fit to the theory curves. The good agreement between measurement and theory confirms the assumption that a Boltzmann distribution is reached for the rotational states in the cell, at the He flux applied in this particular measurement. Moreover, the plot in Figure~\ref{Fig_TrotFit}(b) also shows a relatively good agreement between $T_{\rm{rot}}$ and $T_{\rm{cell}}$. This, however, is not a universal effect, but rather a consequence of the applied He flux. In fact, as will be shown in the next Section, by varying the He density in the cell, we also vary $T_{\rm{rot}}$.
\subsection{Dependence of the rotational temperature on the buffer-gas density}
\label{Subsection_TrotBGdensity}
As already pointed out qualitatively (see Section~\ref{Section_ExpResults}), increasing the buffer-gas density in the cell increases the cooling capacity, and, consequently, results in a better cooling of the molecules. To quantify this effect, we determined the rotational temperature of $\rm{CH_3F}$ for different He densities ranging from $2.2\times10^{14}\rm{cm^{-3}}$ to $7.1\times10^{15}\rm{cm^{-3}}$ at a fixed cell temperature of $T_{\rm{cell}}=6.4\,\rm{K}$. The dependence is plotted in Figure~\ref{Fig_HeDensity}(a) (red squares). We observe cooling of rotational states to near the cell temperature for minimal He density below $1\times10^{15}\rm{cm^{-3}}$ for our cell geometry. It is also particularly interesting to note that for buffer-gas densities above $2\times{10^{15}\rm{cm^{-3}}}$, the rotational temperature of the $\rm{CH_3F}$ molecules leaving the cell is lower than the cell temperature. This evidences further cooling as a result of the adiabatic expansion of the gas upon streaming out of the cell~\cite{Hutzler2011,Barry2011}. The higher the buffer-gas density, the stronger the effect of the adiabatic cooling. For a He density of $7.1\times10^{15}\rm{cm^{-3}}$, the rotational temperature is determined to be $(4.2\pm0.1)\,\rm{K}$.\\
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig5_Trot_vs_BGdensity_FlowmeterCalibrated_AllBgDensities_2Panels_InitSubmission}
\caption{(a) Rotational temperatures of two molecular species, $\rm{CH_3F}$ and $\rm{CF_3CCH}$, as a function of the He buffer-gas density in a cell operated in the boosted regime. The cell temperatures were kept constant, $6.4\,\rm{K}$ and $8.5\,\rm{K}$, for $\rm{CH_3F}$ and $\rm{CF_3CCH}$, respectively. (b) Rotational temperature of $\rm{CH_3F}$ as a function of the Ne buffer gas in a cell operated in the supersonic regime. The cell temperature was kept fixed at $18\,\rm{K}$.}
\label{Fig_HeDensity}
\end{figure}
\subsection{Cooling of a heavier molecule}
\label{Subsection_Trifluropropyne}
In addition to varying the cell parameters, we also investigated the effect of the molecular species on the cooling process. Here we cool the rotational degrees of freedom of a heavier molecule, 3,3,3-trifluoropropyne, $\rm{CF_3CCH}$ (mass $m=94\,\rm{u}$, rotational constants $A=5.7\,\rm{GHz}$ and $B=2.8\,\rm{GHz}$), and compare the results with those for $\rm{CH_3F}$ (mass $m=34\,\rm{u}$). The at least 10-fold decrease in energy spacing ($\propto A,\,B$) for $\rm{CF_3CCH}$ compared to that of $\rm{CH_3F}$ not only greatly reduces the signal per state in the measurement, but also requires more energy levels to be included in the simulation. In this case we have performed trajectory simulations including the exact Stark curves for all 2869 guidable $|JKM\rangle$ levels below 72~K (corresponding to 99.7\% thermal population at 10~K). We have derived the rotational temperature of $\rm{CF_3CCH}$ for different He densities in a similar range as for $\rm{CH_3F}$, from $1.6\times10^{15}\rm{cm^{-3}}$ to $8.9\times10^{15}\rm{cm^{-3}}$, and for a fixed buffer-gas cell temperature of $T_{\rm{cell}}=8.5\,\rm{K}$, as shown in Figure~\ref{Fig_HeDensity}(a) (blue circles). The higher cell temperature in this case results from the maintained higher temperature of the molecule feed line, as $\rm{CF_3CCH}$ has a considerably higher freezing point than $\rm{CH_3F}$. The trend of better cooling for higher buffer-gas densities is clearly observed. Nevertheless, even for the largest He density applied here, the rotational temperature remains slightly above the cell temperature, i.e., the molecules are not fully thermalized. Various effects play a role in determining the thermalization rate. A possible reason for the worse cooling of the rotational degrees of freedom for $\rm{CF_3CCH}$ compared to $\rm{CH_3F}$ for similar buffer-gas densities includes the larger density of rotational states populated by $\rm{CF_3CCH}$ requiring more collisions to thermalize.\\
\subsection{Rotational temperatures in the supersonic regime}
\label{Subsection_Supersonic}
So far we have only discussed the buffer-gas cooling in the boosted regime. We have also measured the rotational temperature of the $\rm{CH_3F}$ molecules leaving the buffer-gas cell operated in the supersonic regime. The supersonic beams are formed when a large number of collisions take place near the cell nozzle during expansion into a vacuum. During this process, the particles convert their internal energy into a kinetic one whereby they get colder~\cite{Levy1980}. In this case, we used Ne as a buffer gas since at $T_{\rm{cell}}=18\,\rm{K}$ it has a supersonic flow speed below $200\,\rm{m\,s^{-1}}$, lower compared to the supersonic flow velocity of He at $6\,\rm{K}$, which makes guiding more feasible. The other benefit of using Ne is the better pumping of the background gas. The dependence of the measured rotational temperature, $T_{\rm{rot}}$, as a function of the buffer-gas density is shown in Figure~\ref{Fig_HeDensity}(b). Here we demonstrate that the supersonic expansion leads to a reduction of the rotational temperature by almost an order of magnitude below the cell temperature, allowing us to cool molecules down to $(2.1\pm0.1)\,\rm{K}$.\\
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{Fig6_LongVelos_BufferGasScan_InitSubmission}
\caption{(a) Derived longitudinal velocity distribution for $\rm{CH_3F}$ at the input of the guide for $T_{\rm{cell}}=6.4\,\rm{K}$ and different He densities, along with their fits. (b) Mean longitudinal velocity, $\bar{v}_{\parallel}$, from the fits as a function of the corresponding $n_{\rm{He}}$. The solid curve is based on a momentum transfer model described in the text.}
\label{Fig_LongVeloDistrib}
\end{figure}
\subsection{Comparison between rotational and translational temperatures}
\label{Subsection_TrotTtrans}
In addition to characterizing the rotational-state distribution and measuring $T_{\rm{rot}}$ of the molecules emerging from the cell, it is also important to retrieve their longitudinal velocity distribution ($v_{\parallel}$-distribution), determine the translational temperature, and compare it to the rotational one. The procedure of deducing the $v_{\parallel}$-distributions at the cell output is explained in Section~\ref{Section_MonteCarlo}. Figure~\ref{Fig_LongVeloDistrib}(a) shows the $v_{\parallel}$-distributions of $\rm{CH_3F}$ for the same cell parameters as those used in the $T_{\rm{rot}}$ measurement in Section~\ref{Subsection_TrotBGdensity}. The solid curves in Figure~\ref{Fig_LongVeloDistrib}(a) are fits to the distributions. Gaussian fits centred at peak velocities have been applied to all but the lowest-density distribution ($n_{\rm{He}}=7.8\times10^{14}\,\rm{cm}^{-3}$). The latter has been fitted with a Maxwell distribution corrected for the boosting effect~\cite{MotschBoosting}. The use of a different fit function is substantiated by the quantitative model below.\\
To further analyze the changes in the $v_{\parallel}$-distribution, we plot in Figure~\ref{Fig_LongVeloDistrib}(b) the mean longitudinal velocity, $\bar{v}_{\parallel}$, as a function of $n_{\rm{He}}$. The value of $\bar{v}_{\parallel}$ is extracted from the fits in Figure~\ref{Fig_LongVeloDistrib}(a). Its increase at higher $n_{\rm{He}}$ is understood to result from collisions between molecules and He atoms in the vicinity of the nozzle~\cite{Maxwell2005}. The lighter He atoms move at a higher speed and kick the heavier molecules from behind whereby the molecules get accelerated in the forward direction. To quantify this effect, we have built the following model based on the momentum transfer for elastic collisions. From conservation of energy and momentum, the change in the molecules' velocity per head-on collision is $\Delta v_{\parallel}=\beta(v_{\rm{He}}-v_{\parallel})$, where $\beta=\frac{2m_{\rm{He}}}{m_{\rm{mol}}+m_{\rm{He}}}$ with $m_{\rm{mol(He)}}$ being the mass of the molecule (He atom), and $v_{\rm{He}}$ is the longitudinal velocity of the He atom. As the He density in the nozzle region is typically much higher than the molecule density, we assume $v_{\rm{He}}$ to be constant and approaching the supersonic velocity, which sets the upper bound for the mean velocity of He in the limit of high density~\cite{Levy1980}. This leads to $v_{\parallel}(l)=\bar{v}_{\rm{He}}-(\bar{v}_{\rm{He}}-v_{0})exp(-\beta l)$, where $l$ is the number of head-on collisions in the molecular-beam direction, $\bar{v}_{\rm{He}}$ is the centre velocity of a He supersonic beam at $T_{\rm{cell}}$, and $v_{0}$ is the initial longitudinal velocity of the molecules. $l$ is related to $n_{\rm{He}}$ by $l=dg\sigma n_{\rm{He}}$, where $d=2$mm is the nozzle diameter setting a rough estimate of the distance where collisions take place, $\sigma$ is the $\rm{CH_3F}$-He elastic collision cross-section, and $g$ is a proportionality factor on the order of unity taking into account the fact that only collisions in the direction normal to the nozzle plane contribute to the change in $v_{\parallel}$. Assuming the molecules have translationally thermalized to $T_{\rm{cell}}$ before arriving at the nozzle, and hence $v_{0}$ is approximately their thermal mean velocity at $T_{\rm{cell}}$, our model (the solid curve in Figure~\ref{Fig_LongVeloDistrib}(b)) describes very well the dependence of $\bar{v}_{\parallel}$ with respect to $n_{\rm{He}}$ with the exception of the lowest-density point. The only fit parameter in the model is $g\sigma=2.0\times10^{-15}\,\rm{cm}^2$, which agrees within one order-of-magnitude with the typical $\rm{CH_3F}$-He collision cross-section at this temperature range~\cite{Willey1988}. Moreover, the fact that the lowest-density data stays above the model curve shows that the translational degree of freedom of molecules is not yet thermalized to $T_{\rm{cell}}$ at this point, in line with the finding in the $T_{\rm{rot}}$ measurement at low He densities (see Section~\ref{Subsection_TrotBGdensity}). This also justifies using a different fit function for the lowest-density $v_{\parallel}$-distribution (Figure~\ref{Fig_LongVeloDistrib}(a)).\\
In the following we discuss the translational temperature of molecules, $T_{\rm{tr}}$, and compare it to the rotational one. While the width of a Gaussian fit to the $v_{\parallel}$-distribution defines $T_{\rm{tr}}$ in the high-density limit, where molecular beams approach the supersonic regime, it provides only a lower bound to $T_{\rm{tr}}$ in the low-density limit where a Maxwellian distribution describes $v_{\parallel}$ better, as shown in Figure~\ref{Fig_LongVeloDistrib}(a). Thus we assign a lower bound to $T_{\rm{tr}}$ of $15.7\pm1.7\,\rm{K}$ for the distribution at $n_{\rm{He}}=7.8\times10^{14}\,\rm{cm}^{-3}$ based on its width, while its mean kinetic energy corresponds to $28\pm3\,\rm{K}$. In comparison, $T_{\rm{rot}}$ at this density is found to be $9.5\pm0.1\,\rm{K}$ (the second point in Figure~\ref{Fig_HeDensity}), which is clearly lower than the value of $T_{\rm{tr}}$. Having a colder $T_{\rm{rot}}$ than $T_{\rm{tr}}$ might not be surprising for the given system because of the large mass imbalance between $\rm{CH_3F}$ and He, which makes cooling of the centre of mass motion inefficient. On the other hand, the $v_{\parallel}$-distribution for the highest density in Figure~\ref{Fig_LongVeloDistrib}(a) ($n_{\rm{He}}=7.1\times10^{15}\,\rm{cm}^{-3}$) has a width corresponding to $4.7\pm0.4\,\rm{K}$, which is very similar to $T_{\rm{rot}}=4.2\pm0.1\,\rm{K}$ found in Section~\ref{Subsection_TrotBGdensity}. Presumably, in the high-density limit, $T_{\rm{tr}}$ and $T_{\rm{rot}}$ should converge due to the sufficient number of collisions.\\
\subsection{Rotational-state-dependent thermalization rates}
\label{Subsection_RotStateDependence}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{Fig7_JK-Thermalization_CombinedPlot_InitialSubmission}
\caption{Measured depletion signals (data points) for different rotational states of $\rm{CH_3F}$ and different He densities, (a) $n_{\rm{He}}=2.2\times10^{14}\rm{cm^{-3}}$, (b) $n_{\rm{He}}=7.8\times10^{14}\rm{cm^{-3}}$, and (c) $n_{\rm{He}}=1.6\times10^{15}\rm{cm^{-3}}$, at a fixed cell temperature $T_{\rm{cell}}=6.4\,\rm{K}$, and simulated depletion signals for different rotational states and different rotational temperatures (lines) for each of the applied He densities.}
\label{Fig_JKThermalization}
\end{figure}
In Section~\ref{Subsection_TrotBGdensity} we have investigated the dependence of fitted $T_{\rm{rot}}$ on the He density in the cell. In this Section we take a closer look at the state distribution in the low He density limit. Figure~\ref{Fig_JKThermalization} shows a comparison between the measured depletion signals and the corresponding theory curves in increasing order of He density, $n_{\rm{He}}=2.2\times10^{14}\rm{cm^{-3}}$, $7.8\times10^{14}\rm{cm^{-3}}$, and $1.6\times10^{15}\rm{cm^{-3}}$ and at $T_{\rm{cell}}=6.4\,\rm{K}$, corresponding to the first three points of the $\rm{CH_3F}$ data in Figure~\ref{Fig_HeDensity}(a). Similarly to Figure~\ref{Fig_TrotFit}(a), the data are grouped in columns, and the only fit parameter is the horizontal position of the column, which determines $T_{\rm{rot}}$. The results clearly demonstrate that with increasing He density, the rotational states are better thermalized, and the fitted $T_{\rm{rot}}$ approaches $T_{\rm{cell}}$. Moreover, at lower He density, the larger deviation between the measured depletion signals and the theoretical curves also indicates that the internal states are not fully thermalized. In particular, we would like to point out that the depletion signals for the $|33\rangle$ state (orange diamonds in Figure~\ref{Fig_JKThermalization}) appear far above their theoretical curves, and would correspond to a much higher rotational temperature if the $|33\rangle$ state were fitted alone. This is an indication that the $|33\rangle$ state thermalizes more slowly than the other states. Based on nuclear spin statistics, states with $K$ being a multiple of 3 (ortho-$\rm{CH_3F}$) are not interconvertible with states of $K$ not being a multiple of 3 (para-$\rm{CH_3F}$)~\cite{Wollrab1967} via buffer-gas collisions. Hence, $\rm{CH_3F}$ from the $|33\rangle$ state can only relax to states with $K=0$. Since the rotational energy scales as $K^2$, a transition of the type $K=0\leftarrow K=3$ requires 3 times more energy to be taken away than a transition of the type $K=1\leftarrow K=2$ does. Thus we attribute the above-observed effect to the larger energy difference between the initial and the final state, which makes the $|33\rangle$ state thermalize at a lower rate than the other probed states.\\
\section{Summary and outlook}
We have developed a versatile and robust method for rotational and translational thermometry of cold molecules emerging from a cryogenic buffer-gas source. The method consists of three independent tools, time-of-flight measurements yielding the longitudinal velocity distribution, resonant radio-frequency rotational-state-selective detection of guided molecules, capable of addressing $100\,\%$ of their population, and a complete Monte-Carlo simulation package resolving the guiding efficiencies for all present states.\\
The obtained results make possible the full characterization of buffer-gas sources and provide insight into the cooling processes in a cryogenic cell operated in different regimes. Studying those mechanisms, we have established an efficient control over the rotational and translational cooling of guided molecular beams by tuning the buffer-gas-cell temperature and the buffer-gas density. We have demonstrated rotational cooling below the cell temperature for the strongly boosted and supersonic regimes, achieving rotational temperatures more than an order of magnitude lower than the cell temperature. Interesting physical phenomena have also been observed. Comparing the rotational with the translational temperature, we have shown evidence of faster thermalization of the rotational degrees of freedom for the $\rm{CH_3F}$-He collisions at the limit of low $n_{\rm{He}}$. We have also observed manifestations of rotational-state-dependent cooling rates, and we have provided an explanation of this effect.\\
The presented thermometry method is a very general one, which can be applied to a vast variety of molecules. It is particularly useful for probing cold molecules that are not amenable to other techniques, e.g., LIF or REMPI. Moreover, the thorough understanding of the cooling mechanisms in a cryogenic buffer-gas cell, provided by the new method, is the key to optimizing the production of internally and translationally cold molecules.\\
\textbf{Acknowledgement} We would like to thank Ferdinand Jarisch for the technical support in the lab.
\section{References}
\bibliographystyle{unsrt}
|
\section{Introduction}
Heat stored in neutron stars after their birth is gradually lost to neutrino emission
and surface radiation. As a result, a kyr-old neutron star is expected to have an internal
temperature $T\approx 10^8$~K and a surface temperature $T_s\approx 10^6$~K
\citep{2004ARA&A..42..169Y}.
This expectation is violated by magnetars, a special class of neutron stars with
ultrastrong magnetic fields, $B\sim 10^{14}-10^{16}$~G.
The ages of observed magnetars are $\sim 1-10$~kyr
and their persistent surface temperatures reach $5\times 10^6$~K, making them much
more luminous than ordinary, passively cooling, neutron stars of the same age
(e.g. \citealp{2013MNRAS.434..123V}).
Persistent active magnetars show a remarkably narrow range of surface luminosities
around $\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$ \citep{2006ApJ...650.1070D}. For
a neutron star of radius $R\approx 10-13$~km, this luminosity
corresponds to effective surface temperature $T_s\approx 4\times 10^6$~K,
which is consistent with the temperatures estimated from the shape
of the observed soft X-ray emission.\footnote{Gravitational redshift reduces the
observed temperature by the factor of
$(1-2GM/c^2R)^{1/2}\approx 0.8$. On the other hand, radiation emerging from the
magnetar atmosphere is not exactly Planckian, which tends to somewhat increase the
observed temperature.
In contrast to a normal blackbody, magnetar surface radiation is dominated by
one of the two polarization states (e.g. \citealp{2006RPPh...69.2631H}).}
By definition of magnetars, their luminosities are fed by magnetic energy stored in
the neutron star \citep{1992ApJ...392L...9D,1992AcA....42..145P}. How can
magnetic energy be converted to heat?
(1) One dissipative process is provided by ambipolar diffusion of the magnetized
electron-proton fluid through the liquid neutron core
\citep[][hereafter TD96]{1992ApJ...395..250G,1996ApJ...473..322T}.
The rate of this process scales as $B^2$, which suggests its efficiency
in magnetars. Ambipolar diffusion could keep the core hot for some time,
and the heat flux from the core could sustain the observed surface temperature.
The challenge faced by this scenario is the enormous neutrino cooling that hinders
the heating of the core.
(2) Strong magnetic stresses deform the solid crust beyond the elastic limit, resulting
in mechanical dissipation. Mechanical heating was envisioned in the starquake
picture of magnetar activity \citep{1995MNRAS.275..255T,1996ApJ...473..322T};
its more plausible version is a plastic flow
\citep{2002ApJ...574..332T,2003ApJ...595..342J,2014ApJ...794L..24B}.
Mechanical heating can only occur in the solid crust below the melted ocean,
at depths $z\mathrel{\copy\simgreatbox} 100$~m below the stellar surface.
(3) Magnetic fields in neutron stars gradually decay due to ohmic dissipation.
This mechanism is usually considered to be inefficient on the kyr timescales of interest,
because of a high electric conductivity of the crust (and a huge conductivity of the core).
Ohmic heating could become important in the presence of strong gradients of the
magnetic field, which are sustained by strong electric currents. It was proposed that
ohmic dissipation is assisted by the Hall drift, which can transport magnetic energy to
the shallow subsurface layers \citep{1988MNRAS.233..875J}, where conductivity is
lowest, and could develop a ``Hall cascade'' \citep{1992ApJ...395..250G}.
In addition, it was proposed that deformations of the crust by the magnetic stresses
could create current sheets where strong localized heating could
occur \citep{2001ApJ...561..980T,2002ApJ...580L..69L}.
(4) The bombardment of the magnetar surface by magnetospheric
particles results in its external heating. Evidence for high-energy particles is
provided by persistent nonthermal emission from magnetars, which
is associated with continual electron-positron discharge in
the twisted magnetosphere \citep{2007ApJ...657..967B,2013ApJ...777..114B}.
In this paper we examine the efficiencies of the heating mechanisms (1)-(4)
using simple estimates and illustrating with sample numerical models.
\section{Cooling of a hot core}
The heat capacity of a core with non-superfluid neutrons determines the maximum
thermal energy that could be stored in a neutron star (e.g. \citealp{2004ApJS..155..623P}),
\begin{equation}
E_{\rm th}\sim 10^{48} T_9^2 {\rm ~erg}.
\end{equation}
Without heating, most of $E_{\rm th}$ is lost to neutrino emission
on a timescale shorter than the typical magnetar age $t\sim 10^{11}$~s,
and the core temperature decreases to $T_{\rm core}\sim 10^8$~K while its
surface luminosity $\mathscr L_s$ drops well below $10^{35}$~erg~s$^{-1}$
\citep{2004ARA&A..42..169Y,2009ASSL..357..247P}.
In this section, we discuss what core temperature would be sufficient to sustain the
observed $\mathscr L_s$ of active magnetars. Then we estimate the required heating that
must offset the neutrino cooling to keep the core hot. Section~\ref{ambipolar} will address how the
high temperature could be sustained by ambipolar diffusion.
\subsection{Core temperature capable of sustaining $\mathscr L_s$}
The surface luminosity of persistent magnetars $\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$
approximately corresponds to the
{\it average} surface flux
\begin{equation}
F_s=\sigma_{\rm SB} T_s^4=\frac{\mathscr L_s}{A}
=10^{22} A_{13}^{-1} {\rm ~erg~s}^{-1}~{\rm cm}^{-2},
\end{equation}
where $\sigma_{\rm SB}\approx 5.67\times 10^{-5}$~erg~cm$^{-2}$~s$^{-1}$~K$^{-4}$ is the
Stephan-Boltzmann constant, $T_s\approx 4\times 10^6$~K is
the effective surface temperature, and $A$ is the emission area, which may be
smaller than the stellar surface area
$4\pi R^2\approx 1.5\times 10^{13}$~cm$^2$.
Such a high $T_s$ can be sustained if the interior temperature is comparable to
$10^9$~K \citep{2004ARA&A..42..169Y}.
The interior region here includes
not only the core ($\rho>1.4\times 10^{14}$~g~cm$^{-3}$) but also the lower
crust ($\rho\gg 10^{11}$~g~cm$^{-3}$);
this region is nearly isothermal
due to its high thermal conductivity.
A strong temperature gradient is sustained in the blanketing envelope
in the upper crust, especially where $\rho<10^{9}$~g~cm$^{-3}$,
because this region has a lower thermal conductivity.
A steady heat flux $F_s$ is established on the timescale of heat
conduction across the crust, $t_c\sim 1-10$~yr.
The relation between $T_{\rm core}$ and $T_s$ depends on the strength of the magnetic field
${\,\mathbf B}$
in the blanketing envelope and its angle with respect to the radial
direction, $\Theta_B$, because both affect heat conduction \citep{1999A&A...351..787P}. A strong radial magnetic field ($\Theta_B=0$)
increases the heat flow to the surface. This is the result of Landau quantization of
electron motion in the envelope (electrons can only move along ${\,\mathbf B}$ at low densities
where the electron Fermi energy is below
the Landau energy $\hbar\omega_B$).
In contrast, a horizontal field ($\Theta_B=\pi/2$) impedes the heat flow by the factor of
$(\tau\omega_B)^{-2}\ll 1$, where $\tau$ is the collisional free path time of electrons.
The $T_{\rm core}$-$T_s$ relation also depends on the chemical composition, which must be
iron in the lower envelope $\rho\mathrel{\copy\simgreatbox} 10^9$~g~cm$^{-3}$
but may be lighter elements in the
surface layers.
\citet{2003ApJ...594..404P} calculated the $T_{\rm core}$-$T_s$ relation and gave its analytical approximation
for various $B$ and $\Theta_B$, for both iron and light element envelopes.
Their calculations assumed that neutrino cooling of the envelope is negligible
and the envelope has a gaseous atmosphere. The latter assumption may be invalid,
as the magnetar surface is likely condensed
\citep{2006PhRvA..74f2508M}, although only
approximate calculations are available for the phase transition to the condensed state.
\citet{2007Ap&SS.308..353P} included neutrino cooling
and studied heat conduction in stars with condensed surfaces.
They found that replacing the atmosphere with a condensed surface
weakly affects the $T_{\rm core}$-$T_s$ relation, and
neutrino losses in the envelope become important when $T_{\rm core}\mathrel{\copy\simgreatbox} 10^9$~K.
The losses effectively impose a ceiling for the surface luminosity: $\mathscr L_s$ reaches
its maximum value $\sim 10^{35}$~erg~s$^{-1}$ when $T_{\rm core}\sim 10^9$~K and does
not respond to further increase of $T_{\rm core}$, because the heat flux is lost to neutrino
emission on its way through the crust.
Below this ceiling the $T_{\rm core}$-$T_s$ relation from \citet{2003ApJ...594..404P},
with neglected neutrino losses, may be used.
\begin{figure}[t]
\begin{tabular}{c}
\includegraphics[width=0.5\textwidth]{core_l_tc.eps}
\end{tabular}
\caption{Surface luminosity emitted by a neutron star with a hot core, as
observed at infinity. Each symbol shows a calculated model of steady heat
transfer from the core to the stellar surface. The star is assumed to have
a dipole magnetic field near the surface, in the heat blanketing envelope.
Two cases are considered: the iron envelope and the maximal light element
envelope, which is called ``fully accreted" in \citet{2003ApJ...594..404P}.
The luminosity is shown for two values of the polar magnetic
field: $B_p=3\times 10^{13}$~G and a more typical for magnetars $B_p=10^{15}$~G.
As $T_{\rm core}$ approaches $10^9$~K, $\mathscr L_s^\infty$ approaches the ceiling imposed by
neutrino cooling \citep{2007Ap&SS.308..353P}; heating the core to higher
temperatures would not significantly increase the surface luminosity. }
\label{fig_Tc}
\end{figure}
The dependence of the surface luminosity on $T_{\rm core}$ is shown in Figure~\ref{fig_Tc}
for a neutron star of mass $M=1.4M_\odot$ and radius $R=11.7$~km.
The envelope is assumed to have an approximately dipole magnetic field ${\,\mathbf B}$;
then the angle between ${\,\mathbf B}$ and the radial direction is
\begin{equation}
\tan \Theta_B=\frac{\sin\theta}{2\cos\theta},
\end{equation}
where co-latitude $\theta$ is measured from the magnetic pole
and general relativistic corrections have been neglected.
The surface luminosity of the star is given by
\begin{equation}
\mathscr L_s=4\pi R^2\int_{0}^{1} F_s(\Theta_B,B)\,d\cos\theta,
\end{equation}
and the observed luminosity at infinity $\mathscr L_s^\infty=(1-2GM/c^2R)\mathscr L_s$ is reduced
by the factor of 1.5. One can see
from Figure~\ref{fig_Tc} that the strong magnetic field assists heat conduction to the surface, however in any case
a high core temperature is required to sustain
$\mathscr L_s^\infty=10^{35}$~erg~s$^{-1}$.
In particular, $T_{\rm core}\approx 10^9$~K is required if the star
has an iron envelope.
In the case of the maximum light element envelope,
the required $T_{\rm core}$ is reduced to $\approx 6\times 10^8$~K
(see also Figure~4 in \citealp{2009MNRAS.395.2257K}).
\subsection{Neutrino cooling of the core}
The high $T_{\rm core}\mathrel{\copy\simgreatbox} 6\times 10^8$~K implies a high cooling rate due to neutrino
emission.
Direct urca cooling (hereafter Durca) can provide a huge sink of energy
in the center of the core,
\begin{equation}
\dot{q}_\nu^D\sim 10^{27}\,T_9^6\, {\cal R}_D {\rm~erg~s}^{-1}{\rm cm}^{-3}
\:\:\: (\rho\mathrel{\copy\simgreatbox} 10^{15} {\rm~g~cm}^{-3}),
\end{equation}
where ${\cal R}_D\leq 1$ is a suppression factor that appears in the presence of
superfluidity \citep{2001PhR...354....1Y}. No reasonable heating mechanism can
compete with Durca cooling at temperatures $T_{\rm core}\sim 10^9$~K.
However, it is activated only if the separation between the Fermi levels of protons and
neutrons is sufficiently small, which requires a minimum density comparable to
$10^{15}$~g~cm$^{-3}$, and hence a minimum mass of the neutron star
\citep{1991PhRvL..66.2701L}. The exact threshold mass for the onset of Durca,
$M_D$, depends on the equation of state of the core matter
\citep{1998PhRvC..58.1804A,2011PhRvC..84f2802C,2013A&A...560A..48P}
and can significantly exceed the canonical neutron star mass $M=1.4M_\sun$.
Stars with masses $M<M_D$
do not activate Durca, and the cooling occurs with a lower rate due to
the modified urca reactions (hereafter Murca), which involve a spectator nucleon
taking the excess momentum.
Murca occurs everywhere in the core with the cooling rate given by
(\citealp{1979ApJ...232..541F}),
\begin{equation}
\label{eq:Murca}
\dot{q}_{\nu}^M\sim 7\times 10^{20}
\,T_9^8 \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{2/3} {\cal R}_M
{\rm ~erg~s}^{-1}{\rm ~cm}^{-3},
\end{equation}
where $\rho_{\rm nuc}=2.8\times 10^{14}$~g~cm$^{-3}$ is the nuclear saturation density.
With the onset of proton or neutron superfluidity the Murca rate is suppressed by the
factor ${\cal R}_M<1$, and the main cooling process becomes``Cooper pair cooling'' ---
neutrino emission that accompanies the formation and breaking of Cooper
pairs \citep{1976ApJ...205..541F,2006MNRAS.365.1300K,2009ApJ...707.1131P}.
Its rate is given by
\begin{equation}
\dot{q}_\nu^{CP}\sim 10^{21} \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{1/3} T_9^7
\; f\left(\frac{T_{\rm core}}{T_{\rm crit}}\right) {\rm ~erg~s}^{-1}{\rm ~cm}^{-3},
\end{equation}
where
$T_{\rm crit}$ is the critical temperature for the transition to superfluidity, and
the numerical factor $f(T_{\rm core}/T_{\rm crit})$ describes
the temperature dependence of the Cooper pair cooling;
$f=0$ at $T_{\rm core}>T_{\rm crit}$, $f$ steeply reaches a
maximum at $T_{\rm core}\approx 0.8T_{\rm crit}$ and steeply declines at $T_{\rm core}<0.5 T_{\rm crit}$.
\begin{figure}[t]
\hspace*{-8mm}
\begin{tabular}{c}
\includegraphics[width=0.55\textwidth]{qv.eps}
\end{tabular}
\caption{Neutrino cooling rate as a function of temperature in the core at density
$\rho_{\rm nuc}=2.8\times 10^{14}$~g~cm$^{-3}$. Black curve shows Murca cooling
assuming no superfluidity ($T_{\rm crit}<10^8$~K). Colored curves show the cooling
of matter with non-superfluid protons and superfluid neutrons, for two cases:
$T_{\rm crit}=10^9$~K (blue curves) and $T_{\rm crit}=3\times 10^9$~K (red curves).
Dashed curve shows the Murca contribution and dash-dotted curve shows the
Cooper pair contribution; the net cooling rate is shown by the solid curve.
The triplet-state neutron pairing is assumed
(model B in \citealp{2001PhR...354....1Y}).
}
\label{fig_qv}
\end{figure}
The putative internal magnetic fields $B\mathrel{\copy\simgreatbox} 10^{16}$~G are sufficiently strong to quench
Cooper pairing of protons in most of the core volume, except perhaps its center
\citep{1969Natur.224..673B}. Therefore, we assume that protons are normal, not
superfluid. The onset of neutron superfluidity is theoretically expected in the core at a
temperature $T_{\rm crit}\sim 10^8-10^9$~K (see e.g. Figure~5 in \citealp{2015SSRv..191..239P}).
There is some observational evidence for this transition from the observed surface
temperatures of isolated neutron stars, however this is not settled
and there remains a significant uncertainty in $T_{\rm crit}$
\citep{2004ARA&A..42..169Y,2009ApJ...707.1131P,2012MNRAS.422.2632H,2015PhRvC..91a5806H}.
Based on the detailed calculations summarized by \citet{2001PhR...354....1Y},
Figure~\ref{fig_qv} shows the temperature dependence of the neutrino cooling rate
$\dot{q}_\nu=\dot{q}_\nu^M+\dot{q}_\nu^{CP}$ at the characteristic nuclear density
$\rho_{\rm nuc}=2.8\times 10^{14}$~g~cm$^{-3}$.
One can see that the onset of superfluidity {\it increases} neutrino cooling in the
temperature range of main interest $T>6\times 10^8$~K.
Using the conservative (non-superfluid Murca) cooling rate one can estimate the minimum
neutrino luminosity as $\mathscr L_\nu=V_c\, \dot{q}_\nu\sim 10^{39}\, T_9^8$~erg~s$^{-1}$,
where $V_c\sim 10^{18}$~cm$^{3}$ is the volume of the core.
Sustaining a hot core over the typical magnetar age $t\sim 10^{11}$~s
requires deposition of energy
\begin{equation}
\label{eq:Ecool}
E\approx \mathscr L_\nu t \sim 10^{50}\,T_9^8\, {\rm~erg}.
\end{equation}
This rough, conservative estimate should be compared with the magnetic energy stored
in the neutron star,
\begin{equation}
\label{eq:Emag_int}
E_{\rm mag}\approx \frac{4\pi}{3}\,R^3\, \frac{B^2}{8\pi} \sim 10^{49}\,B_{16}^2 {\rm ~erg}.
\end{equation}
Comparison of Equations~(\ref{eq:Ecool}) and (\ref{eq:Emag_int}) shows that
internal magnetic fields $B\sim 10^{16}$~G are required to provide energy
for interesting heating capable of sustaining
$T_{\rm core}>6\times 10^8$~K and surface luminosity $\mathscr L_s\sim 10^{35}$~erg~s$^{-1}$.
The reservoir of magnetic energy $E_{\rm mag}\sim 10^{48}-10^{49}$~erg is consistent with the
observed radiation output of magnetars.
After three decades of observations of SGR~1806-20, a giant flare
radiated $\sim 2\times 10^{46}$~erg \citep{2005Natur.434.1107P},
and the total energy output over the magnetar lifetime is likely to approach
$E_{\rm rad}\sim 10^{48}$~erg. Assuming a reasonable
efficiency $E_{\rm rad}/E_{\rm mag}\mathrel{\copy\simlessbox} 0.1$, the inferred magnetic energy is
$E_{\rm mag}\sim 10^{49}$~erg.
\section{Magnetic dissipation in the core}\label{ambipolar}
\subsection{Ambipolar drift}
The main
process capable of dissipating magnetic energy in the core is ambipolar diffusion
(\citealt{1992ApJ...395..250G}, \citetalias{1996ApJ...473..322T}).
Ambipolar drift is the motion of the e-p plasma through the (approximately static)
neutron fluid.
The drift is driven by the Lorentz force ${\,\mathbf j}\times{\,\mathbf B}/c=(\nabla\times{\,\mathbf B})\times{\,\mathbf B}/4\pi$
and tends to relieve the magnetic stresses that drive it. Below we summarize
the standard description of ambipolar diffusion in a neutron star core and then examine
its role in magnetars.
The drift is opposed by two factors: pressure perturbations it
induces and friction against the neutron fluid. Friction results from nuclear
collisions between protons and neutrons \citep{1990SvAL...16...86Y,2001A&A...374..151B};
electron-neutron collisions are negligible. The rate of p-n collisions per proton is given by
\begin{equation}
\tau_{pn}^{-1}\approx 5\times 10^{18}\,T_9^2
\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{-1/3}{\cal R}_{pn} \; {\rm s}^{-1},
\end{equation}
where $\rho_{\rm nuc}\approx 2.8\times 10^{14}$~g~cm$^{-3}$, and ${\cal R}_{pn}=1$ if both
protons and neutrons are non-superfluid.
In the presence of superfluidity, ${\cal R}_{pn}<1$ describes the strong (asymptotically exponential)
suppression of the collision rate \citep{2001A&A...374..151B}.
Pressure perturbations
are induced if $\nabla\cdot(n_e {\,\mathbf v})\neq 0$, where $n_e=n_p$ is the electron/proton
number density and ${\,\mathbf v}$ is the proton drift velocity. Such a ``compressive'' drift generates
a change in $n_e$, and hence changes the electron and proton pressures, which
are related to the chemical potentials $\mu_e$ and $\mu_p$ (Fermi energy levels).
The resulting pressure gradient may be written as $-n_e\nabla(\Delta\mu)$
where
\begin{equation}
\Delta\mu=\mu_e+\mu_p-\mu_n,
\end{equation}
which also describes a local deviation from chemical $\beta$-equilibrium
$e,p\leftrightarrow n$.\footnote{For simplicity, our discussion here assumes
the $n,p,e$ composition of the core. A more detailed model will need to include the
muon component that appears where the electron chemical potential exceeds the
muon rest-mass energy.}
The chemical potentials $\mu_e$, $\mu_p$, $\mu_n$ include the rest-mass
energies of the species.
The pressure perturbation $\Delta P\sim n_e\Delta\mu$ cannot exceed the magnetic
stresses that drive the compressive drift --- the drift is chocked when
$n_e|\Delta\mu|\sim B^2/8\pi$. For $B<10^{17}$~G, the magnetic stresses
are small compared with the hydrostatic pressure in the core. Therefore, possible
deviations $\Delta\mu$ are small compared with the neutron chemical potential
$\tilde{\mu}_n=\mu_n-m_nc^2$.
The latter may be approximated as $\tilde{\mu}_n\approx 100\,(\rho/\rho_{\rm nuc})^{2/3}$~MeV with a moderate accuracy of tens of percent, depending on the core equation of state.
Note that $\tilde{\mu}_n\gg (m_n-m_p-m_e)c^2$.
When evaluating quantities weakly affected by the small $\Delta\mu$, such as
plasma density $n_e$, one can use the approximate chemical balance
$\tilde{\mu}_e+\tilde{\mu}_p\approx \tilde{\mu}_n$,
where the chemical potentials with tilde do not include the rest-mass energies.
Note also that $\mu_e\approx\tilde{\mu}_e\gg \tilde{\mu}_p$,
because the degenerate electrons are ultra-relativistic and the
degenerate protons are non-relativistic (while their number densities are equal).
Therefore the approximate equilibrium implies $\tilde{\mu}_e\approx\tilde{\mu}_n$.
The equation of ambipolar diffusion driven by the Lorentz force and opposed by p-n friction and pressure gradients reads \citep{1992ApJ...395..250G},
\begin{equation}
\label{eq:fbalance}
\frac{(\nabla\times{\,\mathbf B})\times{\,\mathbf B}}{4\pi}=n_e\nabla(\Delta\mu) +\frac{n_e m_p^\star {\,\mathbf v}}{\tau_{pn}},
\end{equation}
where $v$ is the proton velocity and $m_p^\star\approx 10^{-24}$~g is the effective
proton mass. This equation takes into account that the drift is slow and one can neglect
the $dv/dt$ term in the dynamic equation, i.e. the force balance is satisfied.
The charged-current weak interactions (in particular the Murca reactions)
tend to restore $\beta$-equilibrium, i.e. to erase $\Delta\mu$.
The reaction rate may be written as $\dot{n}_e=-\lambda|\Delta\mu|$,
where $\lambda$ is related to the compressibility of the plasma \citep{1989PhRvD..39.3804S}.
The low ``ceiling'' $|\Delta\mu|\ll \mu_e$ implies that significant
compression or expansion can only proceed as allowed by the
Murca reactions, i.e. there is an approximate balance,
\begin{equation}
\label{eq:compr}
\nabla\cdot(n_e{\,\mathbf v})\approx -\lambda\,\Delta\mu,
\qquad \left|\frac{\partial n_e}{\partial t}\right| \ll \left|\lambda\,\Delta\mu\right|.
\end{equation}
The value of $\lambda$ is given by
\begin{equation}
\label{eq:lambda}
\lambda\approx 5\times 10^{33}\,T_9^6\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{2/3}
H\,{\cal R}_\lambda
{\rm ~erg}^{-1}{\rm ~cm}^{-3}{\rm ~s}^{-1}.
\end{equation}
This expression takes into account the possible suppression of $\lambda$ due to
neutron superfluidity (factor ${\cal R}_\lambda\leq 1$) and the enhancement due to the
deviation from $\beta$-equilibrium (factor $H\geq 1$), see \citet{2001PhR...354....1Y}.
The $H$-factor is significantly above unity when $\xi\equiv |\Delta\mu|/kT\gg 1$,
\begin{eqnarray}
H(\xi)=\left\{ \begin{array}{ll}
1 & \quad \xi\ll 1 \\
(0.11\xi)^6 & \quad \xi\gg 10.
\end{array}
\right.
\end{eqnarray}
In the regime $\xi\gg 10$ the Murca rate is independent of temperature.
An explicit analytical expression for $H(\xi)$ is given by \citet{1995ApJ...442..749R}
and Appendix~\ref{app}. The factor ${\cal R}_\lambda(T/T_{\rm crit})$
was calculated by \citet{2001A&A...372..130H}, where $T_{\rm crit}$ is the temperature of
the superfluid transition (it appears that they mislabeled the curves in their Figure~2).
The basic picture of ambipolar diffusion may be summarized as follows.
Let $L$ be a characteristic scale of the field variation $\Delta B$. Estimating
$(\nabla\times{\,\mathbf B})\times{\,\mathbf B}\sim B\,\Delta B /L$, $\nabla(\Delta\mu)\sim\Delta\mu/L$, and
$\nabla\cdot(n_e{\,\mathbf v})\sim n_e v/L$, one finds
\begin{equation}
\label{eq:force}
\frac{m_p^\star v}{\tau_{pn}}\sim \frac{B\,\Delta B}{4\pi L\, n_e}-\frac{|\Delta\mu|}{L},
\end{equation}
\begin{equation}
\frac{n_ev}{L}\sim \lambda |\Delta\mu|.
\end{equation}
These two equations can be solved for $|\Delta\mu|$ and $v$,
\begin{equation}
|\Delta\mu|\sim \frac{B\,\Delta B}{4\pi n_e(1+L^2/a^2)},
\end{equation}
\begin{equation}
\label{eq:v}
v\sim \frac{B\,\Delta B\, \tau_{pn}}{4\pi \rho_p L\,(a^2/L^2 + 1)},
\end{equation}
where $\rho_p=n_em_p^\star\sim \rho/20$ is the mass density of the plasma, and
\begin{equation}
a=\left(\frac{\tau_{pn} n_e}{\lambda\,m_p^\star}\right)^{1/2}
\end{equation}
is a characteristic length introduced by \citet{1992ApJ...395..250G}.
Its dependence on the electron density $n_e$ is not strong, and we will use
a crude estimate of $n_e$ obtained from the approximate relation
$\tilde{\mu}_e\approx\tilde{\mu}_n\sim 100(\rho/\rho_{\rm nuc})^{2/3}$~MeV, where
$\tilde{\mu}_e= c\hbar(3\pi^2 n_e)^{1/3}$. This gives
\begin{equation}
\label{eq:ne}
n_e\approx 10^{37}\left(\frac{\rho}{\rho_{\rm nuc}}\right)^2 {\rm ~cm}^{-3},
\end{equation}
\begin{equation}
\label{eq:a}
a\approx
10^4 \,T_9^{-4} \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{5/6}
\left({\cal R}_{pn}{\cal R}_\lambda H\right)^{-1/2} {\rm ~cm}.
\end{equation}
Two regimes are possible: (1) Friction-dominated regime $L\gg a$. The pressure gradient
is sufficiently quickly erased so that p-n friction is the main factor limiting the drift speed.
(2) Pressure ``pillow'' regime $L\ll a$. Friction is negligible and the magnetic force is
nearly balanced by the gradient of the local pressure enhancement (``pillow'').
Then the drift speed is controlled by Murca reactions, which tend to deflate the pillow
and allow slow compression or expansion of the e-p plasma.
We end this brief review of ambipolar diffusion with the following remark.
As pointed out by \citet{1992ApJ...395..250G}, solenodial plasma
motions $\nabla\cdot(n_e{\,\mathbf v})=0$ are not accompanied by any compression
and hence do not perturb $\mu_e$ or $\mu_p$. Such motions
are only limited by the p-n friction, so in this case the term $a^2/L^2$ in
Equation~(\ref{eq:v}) should be removed. The neutron fluid could, in principle, be pulled
into motion with velocity ${\,\mathbf v}_n\neq 0$
without perturbing neutron density or pressure if $\nabla\cdot(n_n{\,\mathbf v}_n)=0$.
However, since $n_n(r)\neq n_e(r)$, the incompressible motion could only occur with
${\,\mathbf v}\neq {\,\mathbf v}_n$, i.e. neutrons cannot move with the plasma and the p-n friction is
inevitable. Moreover, the large density of neutrons $n_n\gg n_e$ implies that their
allowed motions are generally slow compared with those of the plasma, $v_n\ll v$.
Therefore, neutrons are treated as a static background in Equation~(\ref{eq:fbalance}).
\subsection{Magnetic field evolution equation}\label{sec:eqn}
The magnetic field evolution is governed by the Maxwell equation
$\partial{\,\mathbf B}/\partial t=-c\nabla\times{\,\mathbf E}$. The electric field can be expressed from
the force balance for the electron fluid (omitting the small resistive term),
\begin{equation}
\label{eq:el_bal}
-e\left({\,\mathbf E}+\frac{{\,\mathbf v}_e\times{\,\mathbf B}}{c}\right)-\frac{\nabla P_e}{n_e}+m_e^\star{\mathbf g}=0.
\end{equation}
Here $e$ is the absolute value of the electron charge, $n_e\approx Y_e\rho/m_p^\star$
is the electron density, ${\,\mathbf v}_e$ is the velocity of the electron fluid, $P_e\propto n_e^{4/3}$
is the electron pressure, ${\mathbf g}=-\nabla \Phi_g$ is the gravitational
acceleration (in the Newtonian approximation), and $m_e^\star$ is the effective
inertial mass of the relativistic electron.
After taking curl of Equation~(\ref{eq:el_bal}), the two last terms disappear, taking
into account that $(\nabla P_e)/n_e=\nabla(4 P_e/n_e)$ and
$\nabla m_e^\star\parallel \nabla \Phi_g$. Then one finds
\begin{equation}
\label{eq:el1}
\frac{\partial{\,\mathbf B}}{\partial t}=\nabla\times \left({\,\mathbf v}_e\times{\,\mathbf B}\right),
\end{equation}
which states that the magnetic field is frozen in the electron fluid.
An alternative form of $\partial{\,\mathbf B}/\partial t$ is obtained if ${\,\mathbf E}$ is expressed
from the force balance for the proton fluid,
\begin{equation}
\label{eq:ions0}
e\left({\,\mathbf E}+\frac{{\,\mathbf v}\times {\,\mathbf B}}{c}\right)+m_p^\star{\mathbf g}-\frac{\nabla P_p}{n_e}
-\frac{m_p^\star {\,\mathbf v}}{\,\tau_{pn}}=0,
\end{equation}
where $P_p$ is the pressure of the degenerate protons, $P_p\propto n_e^{\gamma}$
with $\gamma\approx 5/3$, and we have used the neutrality condition ($en_e$ equals
the proton charge density). This gives
\begin{eqnarray}
\label{eq:ions1}
\nonumber
\frac{\partial{\,\mathbf B}}{\partial t} & = & \nabla\times \left({\,\mathbf v}\times{\,\mathbf B}
-\frac{c\,m_p^\star{\,\mathbf v}}{e\,\tau_{pn}}\right) \\
& = & \nabla\times\left({\,\mathbf v}\times{\,\mathbf B}+{\mathbf v}_{\rm H}\times{\,\mathbf B}\right),
\end{eqnarray}
where ${\mathbf v}_{\rm H}=-{\,\mathbf j}/en_e={\,\mathbf v}_e-{\,\mathbf v}$ is the Hall velocity (the velocity of the electron fluid
relative to the protons) and in the second equality we have used Equation~(\ref{eq:fbalance}).
Equation~(\ref{eq:ions1}) is equivalent to Equation~(\ref{eq:el1}).
Note that
(1) ${\mathbf v}_{\rm H}$ is perpendicular to the ambipolar drift velocity ${\,\mathbf v}$, and
(2) Hall drift conserves magnetic energy \citep{1992ApJ...395..250G}
while ambipolar drift dissipates it. The ratio of the two drift speeds is given by
\begin{equation}
\frac{v}{v_{\rm H}}=\frac{\tau_{pn} eB}{m_p^\star c}
\sim 30\, B_{16}T_9^{-2}\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{1/3}{\cal R}_{pn}^{-1}.
\end{equation}
In the parameter range of main interest $v_{\rm H}\ll v$.
Note also that the term ${\mathbf v}_{\rm H}\times{\,\mathbf B}$ in Equation~(\ref{eq:ions1}) is proportional to the
Lorentz force applied to the e-p plasma, and its solenoidal component
can only be balanced by the friction force, so Hall drift in the core requires
p-n friction. The proton fluid itself cannot offset the force associated with
Hall drift, because the proton stress tensor $\sigma_{ik}=P_p\delta_{ik}$ is only capable
of sustaining a curl-free force, which corresponds to a curl-free contribution to the electric
field and makes no contribution to $\partial{\,\mathbf B}/\partial t$.
\subsection{Plateau in the thermal evolution}
A nascent neutron star with its initial temperature $\sim 10^{11}$~K is quickly cooled
by neutrino emission until heating due to ambipolar drift offsets cooling.
Below we show that the drift of magnetic fields $B\mathrel{\copy\simgreatbox} 10^{16}$~G can sustain
a high temperature $T_{\rm core}>6\times 10^8$~K for $\sim 1$~kyr.
We first consider the core with normal (non-superfluid) matter, i.e. assume that
$T_{\rm crit}$ is below the temperature range of interest.
The core heat capacity is dominated by neutrons and given by
\begin{equation}
C_V \approx \frac{\pi^2}{2} n_n k\left(\frac{kT}{\tilde{\mu}_n}\right)
\approx 2\times 10^{20}\,T_9 \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{1/3}
\frac{\rm erg}{{\rm K~cm}^3},
\end{equation}
where $k$ is the Boltzmann constant, $n_n\approx \rho/m_n^\star$, and
$m_n^\star\approx 10^{-24}$~g.
The thermal conductivity of the core is very high, orders of magnitude
higher than in the crust (e.g. \citealp{2001A&A...374..151B,2001MNRAS.324..725G}).
Any locally generated heat is quickly shared by the entire core at approximately uniform
temperature $T_{\rm core}$.\footnote{In a star with mass $M>M_D$, activation of Durca cooling
at the center could create a temperature gradient, however such stars are not
considered here --- Durca cooling would steal too much energy and make the
core uninteresting as a heat source for the surface luminosity.}
The evolution of $T_{\rm core}$ is approximately described by the volume-averaged equation,
\begin{equation}
C_V\frac{dT_{\rm core}}{dt}=-\dot{q}_\nu+\dot{q}_h,
\end{equation}
where $\dot{q}_h$ is the volume-averaged heating rate.
At early times (when $T_{\rm core}\gg 10^9$~K) the cooling term strongly dominates,
$\dot{q}_\nu\gg\dot{q}_h$, and the temperature follows a power-law,
\begin{equation}
\label{eq:cool}
T_{\rm core}\approx 10^9\left(\frac{t}{\rm yr}\right)^{-1/6} {\rm K} \qquad (\dot{q}_\nu\gg\dot{q}_h).
\end{equation}
The core cools to $10^9$~K in about 1~yr and then the ambipolar drift
$v\propto T_{\rm core}^{-2}$ becomes fast enough to provide strong heating and offset the
neutrino cooling.
Indeed, consider magnetic field $B$ that varies by $\delta B$ on a scale $L$.
The scale should not exceed a few km and is certainly smaller than the radius
of the star; it will be normalized below to $10^5$~cm.
In the temperature range of interest the ambipolar drift occurs in the friction-dominated
regime $a<L$ (see Equation~(\ref{eq:a})), in contrast to the opposite assumption in TD96
and \citet{2004ApJ...608L..49A}.
The heating rate is the product of the friction force $\rho_p v/\tau_{pn}$ and the drift speed $v$,
\begin{equation}
\dot{q}_h=\frac{\rho_p v^2}{\tau_{pn}}
\sim \frac{\tau_{pn}}{\rho_p}\left(\frac{B\,\delta B}{4\pi L}\right)^2.
\end{equation}
The heating balances Murca cooling, $\dot{q}_h\approx\dot{q}_\nu$, when the core temperature
decreases to
\begin{equation}
\label{eq:Tbal}
T_{\rm bal}\approx
8 \times 10^8 \left(\frac{B_{16}\,\delta B_{16}}{L_5}\right)^{0.2}
\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{-7/30} {\rm~K}.
\end{equation}
The characteristic timescale for dissipating the available magnetic energy is given by
\begin{eqnarray}
\nonumber
t_{\rm diss} &\sim& \frac{(\delta B)^2}{8\pi \dot{q}_h} \sim \frac{2\pi L^2\rho_p}{B^2 \tau_{pn}} \\
& \approx & 6\times 10^2\,
\frac{L_5^{1.6} (\delta B_{16})^{0.4}}{B_{16}^{1.6}} \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{6/5} {\rm ~yr}.
\label{eq:tdiss}
\end{eqnarray}
Comparing with the neutrino cooling timescale $t_\nu=C_VT/2\dot{q}_\nu$, one finds
\begin{equation}
\frac{t_{\rm diss}}{t_\nu} \sim 10\, \frac{L_5^{0.4}(\delta B_{16})^{1.6}}{B_{16}^{0.4}}
\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{2/15}\gg 1.
\end{equation}
Thus, strong fields $\delta B\sim B\mathrel{\copy\simgreatbox} 10^{16}$~G imply that $T_{\rm core}$ stays near
$T_{\rm bal}$ for a relatively long time $t_{\rm diss}$, much longer than it takes to reach the balance.
This picture may be extended to allow a spectrum of magnetic field variations in
a nascent magnetar,
\begin{equation}
(\delta B)^2\propto L^{-\alpha}, \qquad L<R.
\end{equation}
As the core cools and ambipolar diffusion develops, $\delta B$ may be first damped on small
scales $L$ and then on progressively larger $L(t)$. Equating $t_{\rm diss}$ to the stellar age $t$,
one finds from the above equations $L\propto t^{5/(8-\alpha)}$ and
$T_{\rm bal}\propto L^{-(2+\alpha)/10}$, which gives
\begin{equation}
T_{\rm bal}\propto t^{\frac{-(2+\alpha)}{2(8-\alpha)}}.
\end{equation}
Eventually ambipolar diffusion becomes efficient on the largest scale $L_{\max}<R$.
Once $\delta B$ is damped on this scale, heating is extinguished and $T_{\rm core}$ quickly drops.
With the end of ambipolar diffusion one may expect a decline in magnetar activity.
\subsection{One-dimensional model}
Ambipolar diffusion may be illustrated by the following model.
Consider an approximately uniform background $\rho\approx const$ and a magnetic
field in Cartesian coordinates $x,y,z$ of the form,
\begin{equation}
\label{eq:B}
{\,\mathbf B}=(0,0,B), \qquad B=B_0\sin \mathbb{k} x.
\end{equation}
Ambipolar diffusion will tend to flatten the profile of $B$.
However, the ``null points'' $x=0,\pi/\mathbb{k}$ where $B=0$ do not move, as
the magnetic force $\partial/\partial x(B^2/8\pi)$ vanishes at these points.
As a result, the initial sine profile will relax to the final step-like shape,
\begin{equation}
\label{eq:B1}
B(x)=\pm B_1, \qquad B_1=\frac{2}{\pi}\,B_0,
\end{equation}
with the jumps at the null points.
Note that a large free energy remains stored in the magnetic field after ambipolar
diffusion has done its work. As the sine profile relaxes
to the top-hat $B(x)=\pm B_1$, only a fraction $1-8/\pi^2\approx 19$\% of the
initial magnetic energy is dissipated. The average dissipated
energy density is $U_{\rm diss}\approx 3.8\times 10^{29}\,B_{0,16}^2 {\rm ~erg~cm}^{-3}$.
This value should be compared with the minimum neutrino losses (non-superfluid Murca
cooling) for the desired temperature $T>6\times 10^8$~K over the magnetar age
$t\sim 10^{11}$~s:
$U_{\rm lost}\mathrel{\copy\simgreatbox} 10^{30} (T/6\times 10^8{\rm ~K})^8\, t_{11}$~erg~cm$^{-3}$.
One can see that models with $B_0\mathrel{\copy\simgreatbox} 2\times 10^{16}$~G are of main interest for
the hot-core scenarios. Then the main stage of ambipolar diffusion must occur
in the friction-dominated regime.
Evolution of the magnetic field is described by Equation~(\ref{eq:ions1}).
It is easy to see that ${\mathbf v}_{\rm H}=-(c/4\pi en_e)\nabla\times{\,\mathbf B}$ is in the $y$-direction,
${\mathbf v}_{\rm H}\times{\,\mathbf B}$ is in the $x$-direction, and the Hall term $\nabla\times({\mathbf v}_{\rm H}\times{\,\mathbf B})$
vanishes. The ambipolar drift velocity of protons (which is along the $x$-axis) leads to
the evolution of $B=B_z$ according to the equation,
\begin{equation}
\label{eq:Bdot}
\frac{\partial B}{\partial t}=-\frac{\partial}{\partial x} (vB).
\end{equation}
In the friction-dominated regime, the drift velocity ${\,\mathbf v}=(v,0,0)$ is given by
\begin{equation}
\label{eq:vfric}
v= -\frac{\tau_{pn}(T)}{\rho_p}\frac{\partial}{\partial x}\frac{B^2}{8\pi}.
\end{equation}
This yields a nonlinear diffusion equation for $B$.
The diffusion is accompanied by heating with rate
$\dot{q}_h=[\partial_x (B^2/8\pi)]^2 \tau_{pn}/\rho_p$, and the temperature evolution is described by
\begin{equation}
C_V\frac{dT}{d t}=\frac{\mathbb{k}\,\tau_{pn}}{2\pi\rho_p}\int_0^{2\pi/\mathbb{k}}
\left(\frac{\partial}{\partial x}\frac{B^2}{8\pi}\right)^2 dx - \dot{q}_\nu.
\label{eq:T}
\end{equation}
The coupled Equations~(\ref{eq:B})-(\ref{eq:T}) can be solved numerically for
$B(t,x)$and $T(t)$. Note, however that these equations assume $\Delta\mu\approx 0$
due to efficient Murca reactions and do not take into account the possible
build-up of a pressure gradient (pillow) as the flow converges toward the null points.
An approximate solution to the full problem, which includes the pillow
formation, may be obtained as follows. Let us define a characteristic scale
\begin{equation}
L_1=B_1\left(\frac{dB}{dx}\right)^{-1},
\end{equation}
with $dB/dx$ evaluated at the null point $x=0$.
In the region $0<x<L_1$ we have $B(x)\approx B_1 x/L_1$.
The initial $B(x)=B_0\sin\mathbb{k} x$ has $L_1=2/\pi \mathbb{k}$, and later $L_1$ shrinks ---
the profile $B(x)$ steepens near the null point as
it evolves toward the final top-hat shape $B(x)=\pm B_1$.
Using magnetic flux conservation, one can parameterize the state of the system at
any time $t$ with only one degree of freedom $L_1(t)$ (see Appendix), which obeys
the following dynamic equation,
\begin{eqnarray}
\label{eq:Ldot}
\frac{dL_1}{dt}\approx \left\{\begin{array}{ll}
\vspace*{0.2cm}
\displaystyle{-\frac{\tau_{pn} B_1^2}{2\pi\rho_p L_1}} & \quad L_1>L_\star, \\
\displaystyle{-\frac{\lambda B_1^2 L_1}{4\pi n_e^2}} & \quad L_1<L_\star.
\end{array}
\right.
\end{eqnarray}
Here the transition $L_1=L_\star$ corresponds to $L_1/a=\sqrt{2}$. At this moment,
the rate of Murca reactions becomes insufficient to remove $\Delta\mu$ in the compressed
region near $x=0$, and the dynamics near the null point occurs in the pillow-dominated
rather than friction-dominated regime.
The coefficient $\lambda$ is evaluated inside the pillow at $x=0$, where the local
$\Delta\mu$ can exceed $kT$. Therefore, $\lambda$ in Equation~(\ref{eq:Ldot}) must be
calculated using the correction factor $H(\xi)$ (see Equation~(\ref{eq:lambda}) and Appendix~A).
The evolution of the system is described by two coupled differential equations for
$L_1(t)$ and $T(t)$. The temperature $T(t)$ remains approximately uniform across
the domain, because of the high thermal conductivity, and its evolution is described by
\begin{equation}
\label{eq:Tdot}
C_V\, \frac{dT}{dt}=\dot{q}_h-\dot{q}_\nu, \qquad \dot{q}_h \approx -\frac{B_1^2\, \mathbb{k} \dot{L}_1}{12\pi^2}.
\end{equation}
This approximate expression for the volume-averaged
heating rate $\dot{q}_h$ is derived in Appendix~A. It underestimates the heating rate by
a factor of 2 at the initial stage when $L_1=2/\pi\mathbb{k}$. A simple approximate way to
correct this (used in the numerical models below) is to multiply $\dot{q}_h$ by $1+\pi\mathbb{k} L_1/2$.
The volume-average cooling rate $\dot{q}_\nu$ is dominated by the large region
$x>L_1$ where $\Delta\mu<kT$
throughout the evolution; therefore, the standard Murca cooling can be used for $\dot{q}_\nu$
(Equation~\ref{eq:Murca}), neglecting $\Delta\mu$.
\begin{figure}[t]
\vspace*{-0.3cm}
\begin{tabular}{c}
\hspace*{-1.5cm}
\includegraphics[width=0.65\textwidth]{Temp.pdf}
\end{tabular}
\vspace*{-0.8cm}
\caption{Temperature evolution in a non-superfluid core with the simple initial
configuration of the magnetic field given in Equation~(\ref{eq:B}). The curve $T_{\rm core}(t)$
is solid as long as ambipolar drift proceeds in the friction-dominated regime in
the entire domain, and dashed after the formation of pressure pillows at the
null points. Three sample models are shown:
(1) $B_1=10^{16}$~G ($B_0=1.57\times 10^{16}$~G),
$\mathbb{k}=\pi\times 10^{-6}$~cm$^{-1}$ (red curve),
(2) $B_1=1.5\times 10^{16}$~G ($B_0=2.36\times 10^{16}$~G),
$\mathbb{k}=\pi\times 10^{-6}$~cm$^{-1}$ (blue curve), and
(3) $B_1=1.5\times 10^{16}$~G, $\mathbb{k}=10^{-5}$~cm$^{-1}$ (black curve).
All models assume $\rho=\rho_{\rm nuc}$.
The dotted line shows the core cooling in the absence of heating by ambipolar
diffusion (Equation~\ref{eq:cool}).
The region that would explain the observed surface luminosities
$\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$ at ages 1-10~kyr is shaded in green.
}
\label{fig:Tc}
\end{figure}
Figure~\ref{fig:Tc} shows the temperature evolution calculated in a few sample
models with $B_1=10^{16}$~G and $1.5\times 10^{16}$~G.
One can see that the cooling curve $T\approx 10^9\,t_{\rm yr}^{-1/6}$~K
is followed by the heating=cooling plateau with $T=T_{\rm bal}$.
The models with $T_{\rm bal}>6\times 10^8$~K have the plateau duration up to $\sim 1$~kyr.
The higher the plateau temperature $T_{\rm bal}$ the shorter its duration.
We ran many more models with various $B_0$, $\rho$, and $2\pi/\mathbb{k}\leq 20$~km,
and in all cases the core temperature was below $6\times 10^8$~K at the
typical observed magnetar age $t=1-10$~kyr.
This tension between the model and observations would only be alleviated for
smaller $\mathbb{k}$ outside the plausible range $2\pi/\mathbb{k}\leq 20$~km.
The heating stage is followed by a steep drop of temperature back to the cooling curve
$T\approx 10^9\,t_{\rm yr}^{-1/6}$~K.
Note that the temperature evolution shown in Figure~\ref{fig:Tc}
does not take into account the transition to neutron superfluidity, which should
occur when the temperature drops well below $10^9$~K. This transition is accompanied
by enhanced cooling due to Cooper pair formation
\citep{2009ASSL..357..247P}, and the core
temperature will decrease to $10^8$~K much faster than in million years.
The simple one-dimensional model illustrates another interesting feature of
ambipolar diffusion: the creation of current sheets separating the domains of opposite
magnetic fields. It is described in more detail in Appendix~A.
The appearance of current sheets may be viewed as a consequence of magnetic
flux conservation: ambipolar diffusion tends to minimize the magnetic energy
while the magnetic flux remains frozen in the plasma. Current sheets are also expected
in MHD relaxation of more general (less symmetric) magnetic configurations
(e.g. \citealp{2009arXiv0909.1815G,2015MNRAS.450.3201B}) and can have a strong
guide field.
\subsection{Effects of superconductivity and superfluidity}
The models in Figure~\ref{fig:Tc} neglect possible superconductivity near the
null point, where the magnetic field is weak and incapable of suppressing
Cooper pairing of protons. The superconducting region, where $B<B_c$,
has the thickness $L_c\approx (B_c/B_1)L_1$.
Here the magnetic flux becomes quantized into flux tubes, which
reduces the effective magnetic pressure. On the other hand,
superconductivity also suppresses Murca reactions and so
the region $x<L_c$ becomes nearly incompressible.
This will prevent the collapse of the current sheet,
however will not change our conclusions regarding the temperature evolution.
Superconductivity near null points reduces the energy dissipated by ambipolar
diffusion and does not help to achieve $6\times 10^8$~K at ages of 1-10~kyr.
Note also that in a less symmetric configuration, with a guide magnetic field
in the current sheet, $B$ would not go through zero and can be strong
enough to quench superconductivity everywhere.
Next consider the effects of neutron superfluidity.
The critical temperature for Copper pairing of neutrons is lower than that for protons, but may be high enough to interfere the evolution at temperatures
$T\mathrel{\copy\simlessbox} 10^9$~K. Neutron superfluidity brings the following changes:
\noindent
(1) The rate of p-n collisions is reduced by the factor ${\cal R}_{pn}< 1$.
This reduction promotes ambipolar diffusion.
\noindent
(2) Superfluidity suppresses Murca reactions
responsible for erasing $\Delta\mu$ by the factor ${\cal R}_\lambda<1$; this slows down
the compressive ambipolar drift.
\noindent
(3) Although the Murca cooling is suppressed,
a much stronger cooling occurs due to
Cooper pairing at temperatures $0.3<T/T_{\rm crit}<1$.
It implies a cooling phase with $\dot{q}_\nu$ exceeding
$10^{22}T\,_9^8$~erg~s$^{-1}$~cm$^{-3}$ at $T\approx (0.7-0.8)T_{\rm crit}$
(see Figure~\ref{fig_qv} and \citealp{2009ApJ...707.1131P}).
\noindent
(4) Superfluid neutrons lose their heat capacity. The heat capacity of the core
can become dominated by protons, which
are guarded from Cooper pairing by the ultra-strong field $B\mathrel{\copy\simgreatbox} 10^{16}$~G.
As soon as $T$ decreases below $T_{\rm crit}$ the strong Cooper pair cooling switches
on and the heating cannot balance it until $T/T_{\rm crit}\sim 0.3-0.5$.
At these temperatures, the suppression factors ${\cal R}_{pn}$ and ${\cal R}_\lambda$ are
moderate --- both are comparable to 0.2.
Most of the dissipation still occurs in the friction-dominated regime,\footnote{Superfluidity
increases $a$ by the factor $({\cal R}_{pn}{\cal R}_\lambda)^{-1/2}\sim 5$ during the p-n
friction stage. On the other hand, the faster heating implies a higher temperature,
which tends to reduce $a$ as $a\propto T^{-4}$ (Equation~\ref{eq:a}).
}
and the main effect of superfluidity is the increased dissipation rate,
shortening the duration of the main heating by the factor of $\sim 0.2-0.3$.
Superfluidity only makes the final (pillow) stage
slower, as it reduces $\lambda$ and makes the pillow harder,
however the heating at this stage is insufficient to sustain $T_{\rm core}>6\times 10^8$~K.
Therefore, superfluidity does not help the core to become the main heat source
for persistent magnetars.
\subsection{Comparison with previous work}
In contrast to \citetalias{1996ApJ...473..322T}, we find that the plateau phase
(the balance between ambipolar heating and neutrino cooling) does not sustain
$\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$ for 10~kyr.
The main reason for this disagreement is the heating mechanism.
\citetalias{1996ApJ...473..322T} assumed that ambipolar drift occurs in the pillow regime,
i.e. it is limited by the finite rate of Murca reactions, sustaining the pressure pillow
$\Delta\mu\sim B^2/8\pi$. In this case, what \citetalias{1996ApJ...473..322T} call heating
and cooling processes are in fact the same Murca process that converts
$e,p\leftrightarrow n$ while changing temperature and producing
neutrinos.\footnote{To clarify the meaning of the thermal balance in the pillow regime
one should note the following.
Murca reactions are pure cooling when $\Delta\mu\ll kT$ and pure heating when
$\Delta\mu\gg kT$. The latter limit is approached when $\Delta\mu>10 kT$ --- then each
Murca reaction releases energy $\Delta\mu$; 3/8 of this energy is carried away by
neutrinos and 5/8 heats the matter \citep{2006MNRAS.372..276F}. TD96 simply
assumed that in thermal balance $kT\approx \Delta\mu$. However, in the pillow regime
of ambipolar diffusion, there is a strong gradient of $\Delta\mu$ while $T$ is approximately
uniform due to efficient heat conduction. In this situation, heating=cooling means the balance
between Murca heating in the regions of large $\Delta\mu/kT$ and Murca cooling in the
regions of small $\Delta\mu/kT$. As the field evolves, heating tends to concentrate in a
small fraction of the core volume (see the end of \Sect~3.4).}
In contrast, we find that ambipolar diffusion could sustain
$\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$ only when it occurs in the friction-dominated
regime, i.e. when $\Delta\mu$ is unimportant. The heating by p-n friction is capable of offsetting
the neutrino cooling at $T_{\rm core}>6\times 10^8$~K, however this balance has a short lifetime.
The suggestion of \citetalias{1996ApJ...473..322T} that neutron superfluidity would
prolong the hot phase is incorrect; they neglected the Cooper pair cooling.
\citet{2004ApJ...608L..49A} extended the model of \citetalias{1996ApJ...473..322T}
by assuming superconductivity at $T<T_{\rm crit}=5\times 10^9$~K and by including Hall drift
(we find that Hall drift is unimportant in the core, see \Sect~\ref{sec:eqn}).
Superconductivity would suppress the Murca reaction by a factor of $\sim 10^2$
before the core temperature drops to $T\sim 7\times 10^8$~K \citep{2001A&A...372..130H}.
Then it becomes possible to sustain this temperature for a long time,
because neutrino cooling becomes slow:
cooling due to Murca and Cooper pairing of protons are both inefficient at $T\ll T_{\rm crit}$,
and cooling due to Cooper pairing of neutrons may not begin yet at $T\sim 7\times 10^8$~K.
Superconductivity {\it everywhere} in the core is an essential assumption of this picture.
We argued, however, that the energy
budget of magnetars implies that $B\sim 10^{16}$~G somewhere inside the star,
quenching superconductivity. Then neutrino cooling cannot be suppressed at
$T_{\rm core}\sim 10^9$~K. Note that quenching superconductivity in a fraction of the core
volume is sufficient for fast cooling of the entire core. Quenching is particularly easy
in the outer core, as this requires field $B_c<10^{16}$~G.
\citet{2011MNRAS.413.2021G} studied in detail the effect of strong superfluidity
on ambipolar diffusion. They focused on the regime $T_{\rm core}\llT_{\rm crit}$, which permits
simple analytical expressions for the suppression factors ${\cal R}_{pn}$ and ${\cal R}_\lambda$.
This asymptotic description is useful for superfluid
neutrons in a cool core (with normal protons).
However, it is not applicable to the main phase of ambipolar
diffusion that releases most of the energy --- in the temperature range of main interest,
$T>6\times 10^8$~K, $T_{\rm crit}/T$ can hardly exceed 3.
\citet{2012MNRAS.422.2632H} calculated the temperature of a core heated
by the decay of an initial $B=10^{16}$~G on a prescribed timescale of 10~kyr.
This phenomenological heating model gave $T_{\rm core}\approx 7\times 10^8$~K at 1~kyr
and $5\times 10^8$~K at 10~kyr.
They deemed $T_{\rm core}\approx 7\times 10^8$~K insufficient because it gave $T_s$
below the spectroscopically measured surface temperature $T_X$
(after correcting for the gravitational redshift). In fact, $T_s=(F_s/\sigma_{\rm SB})^{1/4}$ is
allowed to be somewhat below $T_X$ as the surface emission
deviates from blackbody due to radiative transfer effects in the surface layers.
\section{Thermal balance for a heated crust}
\label{balance}
We now turn to another possible explanation of the high surface temperature:
a dissipative process in the crust of the neutron star. General requirements to a
successful quasi-steady heater in the crust were investigated by
\citet{2006MNRAS.371..477K,2009MNRAS.395.2257K,2014MNRAS.442.3484K}.
They assumed a cool core and placed a phenomenological heat source at various
depths in the crust without specifying its mechanism. Their detailed simulations of
heat conduction and neutrino cooling demonstrated that a heating rate
$\dot{q}_h\mathrel{\copy\simgreatbox} 3\times 10^{19}$~erg~s$^{-1}$~cm$^{-3}$ is required at depths
$z<300$~m to sustain the surface luminosity $\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$.
Our goal is to assess if physical mechanisms --- mechanical or ohmic dissipation ---
could provide such heating. However, we begin with a simple phenomenological model
similar to that of \citet{2014MNRAS.442.3484K} to check the constraints on the required
heating. Our sample numerical models below assume a neutron star
with a canonical mass $M=1.4M_\odot$ and the BSk20 equation of state $P(\rho)$
\citep{2013A&A...560A..48P}; it has the radius $R=11.7$~km and
surface gravity $g=1.7\times 10^{14}$~cm~s$^{-2}$.
In the presence of steady crustal heating, the heat transfer equation reads,
\begin{equation}
\label{eq:steady}
-\frac{d}{dz}\left( \kappa\frac{dT}{dz}\right) = \dot{q}_h-\dot{q}_\nu.
\end{equation}
It determines the subsurface temperature profile $T(z)$ for a given
heating rate $\dot{q}_h(z)$ and the self-consistently calculated neutrino cooling
rate $\dot{q}_\nu(z,T(z))$.
The crust is approximated as a slab of thickness much smaller than the stellar
radius; then the relativistic metric coefficients may be approximated as constant
and cancelled from the heat transfer equation.
We numerically solve Equation~(\ref{eq:steady}) as described in \citet{2015ApJ...815...25L},
using thermal conductivity $\kappa$ calculated by Potekhin's code
\citep{1999A&A...351..787P} and neutrino emissivities given by \citet{2001PhR...354....1Y}.
The solution with $\dot{q}_h=0$ gives the relation between $T_s$ and temperature
$T_b$ at a chosen depth $z_b$ above the heater. We choose a small
$z_b\approx 60$~m where $\rho_b=10^9$~g~cm$^{-3}$ and use the obtained
$T_b$-$T_s$ relation as a boundary condition in models with heating at $z>z_b$.
Below we examine the ability of crustal heating to power the observed $\mathscr L_s$
and therefore consider models with a relatively cool core $T_{\rm core}\ll 10^9$~K which is not
capable of sustaining $\mathscr L_s=10^{35}$~erg~s$^{-1}$.
In the sample models we assume $T_{\rm core}=2\times 10^8$~K,
which sustains a surface luminosity $\mathscr L_s\sim 3\times 10^{33}$~erg~s$^{-1}$ with
an iron envelope and $\mathscr L_s\sim 10^{34}$~erg~s$^{-1}$ with a light-element envelope
(Figure~\ref{fig_Tc}).
\begin{figure}[t]
\vspace*{-2cm}
\begin{tabular}{c}
\includegraphics[width=0.45\textwidth]{source.eps}
\end{tabular}
\vspace*{-2.5cm}
\caption{A heated layer of thickness $\Delta z$ at depth $z_h$ feeds the heat
flux toward the core $F_{\rm down}$ and toward the stellar surface $F_{\rm up}$.
The net heating rate per unit area is $F_{\rm up}+F_{\rm down}=F_h=\dot{q}_h \Delta z$.
}
\label{fig:source}
\end{figure}
The desired surface flux $F_s=\sigma_{\rm SB} T_s^4$ requires a heating rate per unit area,
\begin{equation}
F_h=\int \dot{q}_h~dz\ggF_s,
\end{equation}
as most of the heat is conducted to the core and lost to neutrino
emission; only a small fraction $\epsilon$ is conducted to the surface
(Figure~\ref{fig:source}). The required $F_h$ depends on
the characteristic depth $z_h$ where heating occurs.
The calculation is simplified if we use the approximation of a thin
heated layer with thickness $\Delta z\ll z_h$,
\begin{equation}
\label{eq:delta}
\dot{q}_h=F_h\, \delta(z-z_h).
\end{equation}
This idealized model gives a reasonable approximation to the required $F_h$,
which is independent of $\Delta z$. The value of $\Delta z\mathrel{\copy\simlessbox} z_h$ is used
to convert the results obtained with the delta-function approximation to a realistic
heating rate, using the relation $\dot{q}_h=F_h/\Delta z$.
The solution of Equation~(\ref{eq:steady}) with the heat source (\ref{eq:delta})
is found as follows. We fix the effective surface temperature $T_s=4\times 10^6$~K
(which corresponds to $F_s=\sigma_{\rm SB} T_s^4\approx 10^{22}$~erg~s$^{-1}$~cm$^{-2}$)
and integrate the heat diffusion equation
with $\dot{q}_h=0$ downward to $z_h$ where the heater is located. Thus we find $T(z_h)$
and the heat flux $F_{\rm up}$ from $z_h$. This flux can be somewhat larger than $F_s$,
because of neutrino losses at $z<z_h$. The heating rate $F_h$ at $z_h$ feeds two fluxes:
toward the surface and toward the core, $F_h=F_{\rm up}+F_{\rm down}$.
We find the downward flux $F_{\rm down}(z_h)$ using iterations:
any trial $F_{\rm down}$ gives a steady solution connecting $T(z_h)$ and $T_{\rm core}$,
and we iterate it until the solution matches $T_{\rm core}=2\times 10^8$~K
at the bottom of the crust, $\rho=1.4\times10^{14}$~g~cm$^{-3}$.
Note that
$\dot{q}_\nu\neq 0$ and part of the heat flux is lost to neutrino emission before
reaching the core.
\begin{figure}[t]
\begin{tabular}{c}
\hspace*{-0.7cm}
\includegraphics[width=0.55\textwidth]{fz1e15.eps}
\end{tabular}
\vspace*{-3mm}
\caption{
The required internal heating rate per unit area of the crust, $F_h$, as a function
of the heater position $z_h$.
Each symbol shows a calculated model of steady heat transfer from the heater
to the stellar surface (and to the core of temperature $T_{\rm core}=2\times 10^8$~K).
In all models, $F_h$ is adjusted to sustain the effective surface temperature
$T_s=4\times 10^6$~K, which corresponds to surface flux
$F_s\approx 10^{22}$~erg~s$^{-1}$~cm$^{-2}$. Magnetic field $B=10^{15}$~G is
assumed and two cases are shown: $\Theta_B=0$ (radial field) and
$\Theta_B=60^{\rm o}$. The calculations are performed for two different chemical
compositions of the envelope --- iron and maximal light element envelope.
The melted region is indicated by the red part of the curve connecting the symbols.
}
\label{fig:Fh}
\end{figure}
The result for $F_h=F_{\rm up}+F_{\rm down}$ is shown in Figure~\ref{fig:Fh}.
One can see that $F_h\gg F_s\approx 10^{22}$~erg~s$^{-1}$~cm$^{-2}$ is
required in all cases except when the heater is very close to the surface
(near or outside the boundary of our computational domain $z_b\approx 60$~m.)
A moderate inclination of the magnetic field significantly reduces the radial heat flow.
Inclination $\Theta_B=60^{\rm o}$ strongly increases the required $F_h$, especially
for the iron envelope, and excludes $z_h\gg 100$~m. The steep increase and runaway
of the required $F_h$ at large $z_h$ is the result of neutrino losses, which
prevent the internal temperature profile $T(z)$ from reaching the values required
to sustain $F_s$.
\section{Mechanical heating}
The ultrastrong magnetic fields of magnetars can stress their crusts beyond the
elastic limit \citepalias{1996ApJ...473..322T}. Then the crustal deformations become
irreversible and are accompanied by heating.
Part of the released magnetic energy is passed to the external
magnetosphere attached to the crust and part is converted locally to heat.
Thermoplastic waves effectively ``burn'' magnetic energy in the crust,
resembling deflagration fronts in combustion.
Large stresses can be created in the crust in three ways:
\\
(1) Magnetic field evolution in the liquid core differs from the field behavior in the solid
crust. This generates a gradient in the field at the crust-core interface. The resulting
magnetic force applied to the crust may be able to deform it beyond the elastic limit.
Then the crust is expected to experience a shear flow, relieving the applied stress.
This shear flow will tend to localize along ``heat lines'' similar to those observed in
laboratory experiments with a torsional Hopkinson bar (e.g. \citealp{2002pmas.book.....W}).
It must, however,
satisfy an important constraint: the crustal shear should not tear magnetic field lines
(as this would generate magnetic energy) and may develop along magnetic flux surfaces.
\\
(2) Magnetic stresses can be fostered by internal processes in the crust itself,
in particular due to Hall drift.
As long as ohmic dissipation is negligible, the magnetic field remains frozen in the
electron fluid drifting through the ion lattice with velocity ${\,\mathbf v}_{\rm H}={\,\mathbf j}/en_e$,
where ${\,\mathbf j}=(c/4\pi)\nabla\times{\,\mathbf B}$ is the electric current density determined
by the magnetic configuration of the star.
The Hall drift ${\mathbf v}_{\rm H}$ deforms the magnetic field lines and is capable of creating
large magnetic stresses
(\citetalias{1996ApJ...473..322T}; \citealp{2011ApJ...741..123P}).
This leads to launching thermoplastic waves \citep{2014ApJ...794L..24B},
which move the crust and relieve the internal magnetic stresses.
\citet{2016arXiv160604895L} further investigate plastic flows fostered by Hall drift and find
that they can occur in avalanches that develop due to the
excitation of short Hall waves by the plastic flows.
\\
(3)
Magnetospheric flares launch strong Alfv\'en waves that are ducted
along the magnetic field lines and impinge on the crust.
The waves carry enormous magnetic stresses that immediately initiate a
strong oscillating plastic flow in the crust until the wave is
damped into heat, which occurs on a timescale of $\sim 10$~ms \citep{2015ApJ...815...25L}.
Below we explore the maximum efficiency of magnetar surface heating by
mechanical dissipation in the crust.
It must satisfy two general constraints:
\noindent
(1) Mechanical dissipation can only occur in the solid phase below the ocean.
At shallow depths $z\mathrel{\copy\simlessbox} 100$~m the crust is melted and forms a
liquid ocean with a negligible shear viscosity.
This fact limits the efficiency of heating the surface, because most of the heat
produced at large depths is conducted to the core and lost to neutrino emission.
\noindent
(2) The mechanical heating rate is proportional to the shear stress of the deformed
crust. There is an upper limit on this stress (maximal strength of the crustal lattice)
which imposes a ceiling on the heating rate.
\subsection{Quasi-steady mechanical heating}\label{5.1}
We first examine whether quasi-steady mechanical dissipation can explain the
surface luminosity of persistent magnetars, $\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$.
Since there is no mechanical heating in the ocean, one can find
its temperature profile from Equation~(\ref{eq:steady}) with $\dot{q}_h=0$ (for a given
$T_s$). This profile determines the melting depth $z_{\rm melt}$ --- the bottom of the ocean
--- where $T$ reaches the melting temperature
$T_{\rm melt}(\rho)\approx 2.4\times 10^9\,\rho_{12}^{1/3}$~K.
For instance, for an iron envelope with a radial magnetic field, the surface
temperature $T_s=4\times 10^6$~K implies $z_{\rm melt}\approx 200$~m.
The ocean is less deep, $z_{\rm melt}<60$~m, for the light-element envelope.
A conservative lower limit on $F_h$ required to sustain
$T_s\approx 4\times 10^6$~K
is obtained by assuming that mechanical heating is
concentrated at the shallowest possible depth, i.e. $\dot{q}_h$ is given by
Equation~(\ref{eq:delta}) with $z_h=z_{\rm melt}$.
A realistic $\dot{q}_h$ must be distributed over a range
of depths $z>z_{\rm melt}$ (and $\dot{q}_h$ is bounded from above, as discussed below), so realistic
mechanical heating will be less efficient in feeding the surface flux $F_s$.
Therefore, the model with $\dot{q}_h=F_h\,\delta(z-z_{\rm melt})$
gives a conservative upper limit on the surface heating
efficiency $\epsilon=F_s/F_h$.
\begin{figure}[t]
\hspace*{-6mm}
\begin{tabular}{c}
\includegraphics[width=0.55\textwidth]{delta_eff_theta.eps}
\end{tabular}
\caption{Fraction of heat conducted to the stellar surface,
$\epsilon=F_s/F_h$, from a steady delta-function heat source located at $z_{\rm melt}$.
All calculated models have $F_s=10^{22}$~erg~s$^{-1}$~cm$^{-2}$; the star is
assumed to have a relatively cool core ($T_{\rm core}=2\times 10^8$~K).
The surface heating efficiency $\epsilon$ is shown as a function of magnetic field angle
$\Theta_B=\arccos (B_r/B)$. The main fraction $1-\epsilon$ of the produced heat is conducted
to the core and lost to neutrino emission.
}
\label{fig:eff}
\end{figure}
The efficiency of the heater localized at $z_{\rm melt}$
is shown in Figure~\ref{fig:eff} for $F_s\approx 10^{22}$~erg~s$^{-1}$~cm$^{-2}$.
Replacing the delta-function with a more realistic heating $\dot{q}_h$ distributed over
$z>z_{\rm melt}$ would significantly reduce $\epsilon$, and a tilt of magnetic field $\Theta_B>0$ would
further reduce $\epsilon$. We conclude that the most optimistic $\epsilon\mathrel{\copy\simlessbox} 0.1$.
It is also useful to estimate the energy budget invoked by the crustal heating scenario.
Using the typical age of magnetars, $t\sim 10^{11}$~s, their emitted energy from the
surface is $E_s\sim \mathscr L_s t\sim 10^{46}$~erg. The modest efficiency of surface
heating implies deposition of significant energy in the crust,
\begin{equation}
\label{eq:Eh}
E_h\sim 10^{48}\left(\frac{\epsilon}{0.01}\right)^{-1} {\rm~erg}.
\end{equation}
Our next goal is to compare the required heating rate with the maximum rate of
mechanical dissipation due to crustal motions. Note that vertical motions are arrested
by the hydrostatic balance between two dominant forces --- gravity and pressure gradient.
The pressure $P$ of the compressed, hydrostatic crust is dominated by
degenerate electrons (or neutrons, near the bottom of the crust).
The lattice Coulomb energy density $U_{\rm Coul}$ is much smaller than $P$
and the crust is relatively fragile to horizontal shear, which leaves pressure unperturbed.
Therefore, we consider below dissipation due to horizontal shear motions.
The dissipative flow of the lattice begins
when its elastic shear stress $\sigma$ reaches
a critical value $\sigma_{\rm cr}$.
The highest possible value of $\sigma_{\rm cr}\sim 0.1\mu$ represents the strength of an ideal
crystal subject to a fast shear deformation,
where $\mu$ is the shear modulus of the lattice.
The flow initiated in response to excessive stress buffers stress growth and satisfies
the condition
\begin{equation}
\label{eq:sigmax}
\sigma<\sigma_{\max}\sim 0.1\mu.
\end{equation}
This is a conservative limit, which may only be approached when the crust is cold and
deformed quickly \citep{2010MNRAS.407L..54C}.
Note that $\mu$ is comparable to $U_{\rm Coul}$ and
the maximum lattice stress is always a fraction of $\mu$, because
there is no agent to carry the stress other than the Coulomb fields.
The rate of mechanical dissipation is given by
\begin{equation}
\dot{q}_h=-\sigma\dot{s},
\end{equation}
where $s$ is the strain of the dissipative (plastic) deformation, and $\dot{s}$ is its
time derivative. The time-averaged $\dot{s}$ driven by magnetic field evolution
inside the star may be estimated as follows. The solid crust serves as a gate
for the energy strored in helical magnetic fields inside the star
\citepalias{1996ApJ...473..322T}.
The stored wound-up field $B$ can significantly
exceed its radial component $B_r$ emerging through the stellar surface, possibly by a
factor up to $\sim 10^2$ (which corresponds to $B\mathrel{\copy\simlessbox} 10^2B_r\sim 10^{16}-10^{17}$~G).
The maximum angle of field unwinding
$B_{\max}/B_r\sim 10^2$~radian gives a net maximum strain flow $s_{\max}\sim 10^2$.
The corresponding maximum {\it average} strain rate over the active lifetime of a
magnetar $t\sim 10^{11}$~s is
\begin{equation}
\label{eq:smax}
\bar{\dot{s}}\sim \frac{s_{\max}}{t}\sim 10^{-9} {\rm ~sec}^{-1}\approx 0.03 {\rm~yr}^{-1}.
\end{equation}
The maximum stress of a plastic flow $\sigma_{\max}\sim 0.1\mu$ gives the maximum
heating rate,
\begin{equation}
\label{eq:dqmax}
\dot{q}_h^{\max}=\sigma_{\max} \,\dot{s}\sim 10^{18}\, \rho_{12}\,
\dot{s}_{-9}
{\rm ~erg~s}^{-1}{\rm cm}^{-3}.
\end{equation}
Here we used $\mu\approx 10^{28}\rho_{12}$~erg~cm$^{-3}$
(e.g. \citealp{1991ApJ...375..679S}); in the numerical models below we use
more detailed approximations for $\mu$ from
\citet{2005ApJ...634L.153P} and \citet{2007MNRAS.375..261S}.
The characteristic scale of density variation with depth is $\Delta z\approx 100$~m
for depths $z$ of interest, including the lower crust.
The heating in a layer around a given $\rho$ may be estimated as
$F_h(\rho)=\dot{q}_h(\rho)\,\Delta z$. This gives the energy release rate per unit area of the crust,
\begin{equation}
\label{eq:Fpl}
F_h
\sim 10^{22}\,\rho_{12}\,
\left(\frac{\sigma}{\sigma_{\max}}\right)
\dot{s}_{-9}
{\rm ~erg~s}^{-1}{\rm cm}^{-2}.
\end{equation}
This shows
that even with the most optimistic assumptions, quasi-steady mechanical
dissipation can hardly provide the needed heat source
$F_h\sim 10^{24}(\epsilon/0.01)^{-1}$~erg~s$^{-1}$~cm$^{-2}$ capable of
powering the observed surface luminosity. The upper bound on $F_h$ is somewhat lifted
to $\sim10^{24}$~erg~s$^{-1}$~cm$^{-2}$ if the plastic flow occurs in the deep crust
where $\rho\sim 10^{14}$~g~cm$^{-3}$. However, this remains insufficient as the
efficiency of surface heating by the deep heat source decreases below $10^{-2}$
(cf. Figure~\ref{fig:Fh}).
We conclude that quasi-steady mechanical dissipation is incapable of powering
the persistent surface luminosity of bright magnetars.
This conclusion is illustrated by the numerical model assuming the maximum possible
mechanical heating (Figure~\ref{fig:mech}). The model makes the most optimistic (and
unrealistic) assumption that the crust flows everywhere with $\sigma=\sigma_{\max}$.
Even in this case, $F_s$ can barely approach $10^{22}$~erg~s$^{-1}$~cm$^{-2}$,
as long as $|\dot{s}|\ll 0.1$~yr$^{-1}$.
\begin{figure}[t]
\hspace*{-3mm}
\begin{tabular}{c}
\includegraphics[width=0.5\textwidth]{12flux.eps} \\
\includegraphics[width=0.5\textwidth]{14flux.eps}
\end{tabular}
\caption{Surface radiation flux generated by the maximum possible mechanical
dissipation $\dot{q}_h^{\max}=\sigma_{\max}|\dot{s}|$ (Equation~\ref{eq:dqmax}).
Upper panel: the maximum dissipation occurs everywhere in the upper crust
$\rho<10^{12}$~g~cm$^{-3}$. Lower panel: the maximum dissipation occurs
in the entire crust $\rho<10^{14}$~g~cm$^{-3}$.
The magnetic field is assumed to be radial, which is the best possible configuration
for maximizing the surface flux.
The shaded region corresponds to shear rates exceeding $s_{\max}/t$
estimated in Equation~(\ref{eq:smax}).
Shear rates in the darker region $\dot{s}>1$~yr$^{-1}$ would be able
to sustain external magnetic twists against their resistive dissipation in the
magnetosphere \citep{2009ApJ...703.1044B}.
}
\label{fig:mech}
\end{figure}
\subsection{Pumping crustal strain by Hall drift}\label{5.2}
Feeding the surface radiation flux by mechanical dissipation would become possible if
the crust experiences an {\it oscillating} plastic flow with an effective $\Delta s\gg 100$.
Large-amplitude oscillating shear could, in principle, be fed by the internal toroidal field
energy without requiring a quick reduction of the internal $B_\phi$.
In particular, Hall waves in the crust is a possible driver of the oscillations \citep{2016arXiv160604895L}.
The magnetic field evolves according to the equation
$\partial{\,\mathbf B}/\partial t=-c\nabla\times{\,\mathbf E}$. The electric field in the crust satisfies the
generalized Ohm's law which expresses the balance of forces applied to the electron fluid,
\begin{equation}
\label{eq:el_bal1}
{\,\mathbf E}+\frac{{\,\mathbf v}_e\times{\,\mathbf B}}{c}+\frac{\nabla P_e}{e\,n_e}-\frac{m_e^\star{\mathbf g}}{e}
=\frac{{\,\mathbf j}}{\tilde{\sigma}},
\end{equation}
where ${\,\mathbf j}=(c/4\pi)\nabla\times{\,\mathbf B}$ and $\tilde{\sigma}$ is the electric conductivity.
In contrast to Equation~(\ref{eq:el_bal}), here we included the resistive term ${\,\mathbf j}/\tilde{\sigma}$.
This gives
\begin{equation}
\label{eq:el}
\frac{\partial{\,\mathbf B}}{\partial t}=\nabla\times \left[({\,\mathbf v}+{\mathbf v}_H)\times{\,\mathbf B}
-\frac{c}{\tilde{\sigma}}\,{\,\mathbf j}\right], \quad
{\mathbf v}_{\rm H}=\frac{\mathbf j}{en_e},
\end{equation}
where ${\,\mathbf v}$ is the velocity of the ion lattice/liquid, and ${\mathbf v}_{\rm H}={\,\mathbf v}_e-{\,\mathbf v}$ is the velocity
of the electrons relative to the ions (the Hall drift). As long as the ohmic term $c{\,\mathbf j}/\tilde{\sigma}$
is negligible, the magnetic field is frozen in the electron fluid moving with
${\,\mathbf v}_e={\,\mathbf v}+{\mathbf v}_{\rm H}$.
The ion motion ${\,\mathbf v}\neq 0$ occurs in response to magnetic forces, not only
in the liquid ocean but also in the solid crust, as a result of elastic or plastic deformations.
This motion can offset Hall drift. Previous numerical simulations of Hall drift
in the crust used Equation~(\ref{eq:el}) with ${\,\mathbf v}$ set to zero, neglecting ion motion
(e.g. \citealp{2009A&A...496..207P,2013MNRAS.434..123V,2016PNAS..113.3944G}).
For our purposes it is instructive to look at the force balance for ions,
\begin{equation}
\label{eq:ions}
{\,\mathbf E}+\frac{{\,\mathbf v}\times {\,\mathbf B}}{c}+\frac{Am_p{\mathbf g}}{Ze}
-\frac{\nabla\cdot {\sigma}}{en_e}=0,
\end{equation}
where $Z$ and $A$ are the ion charge and mass numbers, and
$\sigma$ stands for ${\sigma}_{ik}$ --- the ion stress tensor. Using the expression
for ${\,\mathbf E}$ provided by Equation~(\ref{eq:ions}) and taking into account that
$\nabla (A/Z)\parallel \nabla \Phi_g=-{\mathbf g}$, one finds
\begin{equation}
\frac{\partial{\,\mathbf B}}{\partial t}=\nabla\times \left({\,\mathbf v}\times{\,\mathbf B}
-\frac{c\,\nabla\cdot\sigma}{en_e}\right).
\end{equation}
The second term in parenthesis determines the drift of the magnetic field relative to the ions.
A conservative upper limit on the ion stress tensor components is given by the
Coulomb energy density, and is also comparable to the shear modulus of the crust $\mu$.
Therefore, one can roughly estimate
\begin{equation}
\left|\frac{\partial{\,\mathbf B}}{\partial t}\right|_{\rm H}^{\max}\sim \frac{c \mu}{en_e \ell^2}
\sim \frac{c\,m_p v_{\rm sh}^2}{e Y_e \ell^2},
\end{equation}
where $\ell$ is a characteristic scale of stress variations and
$v_{\rm sh}=(\mu/\rho)^{1/2} \sim 10^8$~cm~s$^{-1}$ is the speed
of shear waves sustained by the ion lattice; this speed
is approximately uniform throughout the solid crust (e.g. \citealp{1991ApJ...375..679S}).
This gives an estimate for the maximum strain rate pumped by the Hall drift,
\begin{equation}
\label{eq:sH}
\dot{s}_{\rm H}\mathrel{\copy\simlessbox} \frac{1}{B}\left|\frac{\partial{\,\mathbf B}}{\partial t}\right|_{\rm H}^{\max}
\sim \frac{10^{-3}{\rm ~yr}^{-1}}{Y_e\, B_{15}\, \ell_4^{2}}.
\end{equation}
The highest rate can be reached in the deep crust where $Y_e$ decreases to $\sim 0.1$.
The rate $\dot{s}_{\rm H}$ can cause plastic flow with a comparable time-averaged strain
rate $\dot{s}\sim\dot{s}_{\rm H}$. It is lower than needed for
mechanical dissipation to keep the magnetar surface at $T_s\approx 4\times 10^6$~K.
Note also that the tension
of magnetic field lines $\mu_B=B^2/8\pi$ exceeds the shear modulus of the upper crust
$\mu\sim 10^{27}\rho_{11}$~erg~cm$^{-3}$, and $\mu$ practically vanishes in the ocean.
This fact alone suggests that Hall drift cannot cause interesting deformations of the
magnetic field in the upper layers. The presence of significant $\dot{s}_{\rm H}$
by itself does not imply significant field deformations, because it can be offset by the
ion motion that limits the growth of shear stress.
\subsection{Intermittent mechanical dissipation}
The main conclusion of \Sects~\ref{5.1} and \ref{5.2} is that
mechanical dissipation driven by internal evolution of the magnetic field in the
star is too weak to sustain the observed persistent surface luminosity of magnetars.
Strong mechanical heating is only possible in an intermittent regime, where
part of magnetic energy is suddenly dissipated due to an instability. The
instability can happen inside the crust (a thermoplastic wave or an avalanche
of failures driven by short Hall waves) or outside the star (a magnetospheric flare).
In general, the efficiency of surface heating by mechanical dissipation is maximized
when the dissipation takes place at a minimum depth, just below the liquid ocean.
This naturally occurs when a strong high-frequency shear wave is launched from
the magnetosphere toward the crust, as expected in a powerful magnetospheric flare.
Therefore, we now focus on this more promising mechanism.
The magnetospheric wave damping somewhat increases the depth of the ocean
by melting the crust, so that the heat deposition has to peak at the transition to the
solid phase \citep{2015ApJ...815...25L}.
This heating occurs very quickly, on a timescale $\sim 10$~ms.
The Alfv\'en waves excited by the flare create a
train of $\sim 10$ strong oscillations of the crust, with a compressed and amplified strain,
and produce a net plastic strain flow $\Delta s$ that can exceed $10$.
Most of the plastic dissipation occurs in a layer of thickness
$\Delta z\sim 100$~m at a depth of a few hundred meters.
This depth is found by balancing the wave energy deposited per unit area
of the crust, $Q$, with the energy it takes to melt the layer,
\begin{equation}
\label{eq:melt}
Q\sim \Delta z \int_0^{T_m} C_V dT,
\end{equation}
where $T_{\rm melt}\approx 10^{9}\rho_{11}^{1/3}$~K is the melting temperature, and $C_V$
is the heat capacity; for instance,
$C_V\sim 4\times 10^{17}$~erg~cm$^{-3}$~K$^{-1}$ at $\rho=10^{11}$~g~cm$^{-3}$
and $T\approx T_{\rm melt}$ (e.g. \citealp{2001MNRAS.324..725G,2015SSRv..191..239P}).
Equation~(\ref{eq:melt}) determines the characteristic density at which the wave is damped;
it is comparable to $10^{11}$~g~cm$^{-3}$ for $Q\sim 10^{30}-10^{31}$~erg~cm$^{-2}$
and grows with $Q$.
\begin{figure}[t]
\begin{tabular}{c}
\hspace*{-6mm}
\includegraphics[width=0.51\textwidth]{transient_fr_q.eps}
\end{tabular}
\caption{Efficiency $\epsilon=Q_{\rm aft}/Q$ of surface heating by plastic damping of
Alfv\'en waves from a magnetospheric flare.
The efficiency is defined as the fraction of the deposited energy that is
radiated from the surface (rather than conducted to the core and lost to neutrino
emission). It is shown as a function of the
deposited energy per unit area of the crust $Q$. A radial magnetic field was assumed
in the calculations, which gives the maximum $\epsilon$.
}
\label{fig:trans}
\end{figure}
\begin{figure}[t]
\begin{tabular}{c}
\hspace*{-6mm}
\includegraphics[width=0.51\textwidth]{transient_fpeak_q.eps}
\end{tabular}
\caption{Peak flux of the surface radiation following the plastic damping of
Alfv\'en waves. The peak lasts the cooling time (comparable to one year).
}
\label{fig:peak}
\end{figure}
A fraction $\epsilon$ of the deposited heat $Q$ is gradually conducted from the deep melted
ocean (where $T\approx T_{\rm melt}$ immediately after the heating event) to the surface.
This fraction is maximum when the magnetic field is approximately
radial (vertical) in the ocean. We have calculated $\epsilon$ for this case using
detailed time-dependent simulations of heat conduction and neutrino cooling.
The method of our calculations is similar to previous simulations of time-dependent
heat transfer in a neutron star crust (e.g. \citealp{2009ApJ...698.1020B,2009A&A...496..207P,2014MNRAS.442.3484K}) and described in \citet{2015ApJ...815...25L}.
Figure~\ref{fig:trans} shows the result. When $B\sim 10^{15}$~G the afterglow efficiency
$\epsilon=Q_{\rm aft}/Q$ can be approximated by the formula,
\begin{equation}
\epsilon\approx \epsilon_0\,(1+2 Q_{31})^{-3/4},
\end{equation}
where $\epsilon_0\approx 0.1$ and 0.2 for iron and light element envelopes,
respectively. A strong wave delivering energy $Q\gg 10^{30}$~erg~cm$^{-2}$
results in deep melting of the crust and deposits energy at large depths, which
reduces $\epsilon$. Therefore, the afterglow energy radiated per unit area of the
crust $Q_{\rm aft} = \epsilon Q$ saturates near a few times $10^{30} {\rm ~erg~cm}^{-2}$,
slowly changing with $Q>10^{31}{\rm ~erg~cm}^{-2}$.
The peak flux of the surface afterglow is shown in Figure~\ref{fig:peak}. It is reached on
the heat conduction timescale of the ocean, $t_c\sim 10^7$~s, and then gradually
decays as the crust cools. The characteristic afterglow flux from the surface is
$F_s\sim \epsilon Q/t_c$. Our calculations assumed a single flare, however, a similar result
would be obtained if $N$ flares occur during time interval $t<t_c$, as long as $Q$
represents their cumulative energy deposition over the time $t_c$. The frequent flares
may have a
slightly higher efficiency of surface heating, because of lower neutrino cooling, as each
individual heating event $Q/N$ is weaker at large $N$ and has a lower peak temperature.
At $N\gg 1$, the heating approaches the quasi-steady regime with the self-consistent
$z_{\rm melt}$ that was considered in Section~\ref{5.1}.
\section{Ohmic dissipation in the crust}\label{ohm}
Magnetars may have strong non-potential magnetic fields stored in the crust and
sustained by electric currents, which satisfy the relation $(4\pi/c){\,\mathbf j}=\nabla\times{\,\mathbf B}$.
Ohmic dissipation tends to convert the stored energy of non-potential field to heat.
The rate of this process is controlled by the electric conductivity.
\subsection{Electric conductivity}
The electric conductivity of the crustal material is related to its thermal conductivity,
as both charge and heat are transported by the electrons. The conductivities are controlled
by the electron interaction with atomic nuclei (which form the lattice in the solid phase or
the strongly coupled Coulomb liquid in the ocean) and by the magnetic field.
The conductivity tensor $\tilde{\sigma}_{ik}$ in the magnetized crust is described by three
components: $\tilde{\sigma}_\parallel$ (conductivity parallel to the magnetic field), $\tilde{\sigma}_\perp$
(perpendicular to the field), and the Hall component $\tilde{\sigma}_{\rm H}$ (the antisymmetric
off-diagonal component of the tensor $\tilde{\sigma}_{ik}$, see e.g. \citealp{1960ecm..book.....L}).
Detailed calculations of $\tilde{\sigma}_{ik}$ for densities, temperatures,
and magnetic fields relevant to neutron starts are found in \citet{1999A&A...351..787P}.
For a given electric current density ${\,\mathbf j}$, the electric field ${\,\mathbf E}$ can be found by
inverting the relation $j_i=\tilde{\sigma}_{ik} E^k$. It is useful to express the electric current
as the sum of components parallel and perpendicular to ${\,\mathbf B}$, ${\,\mathbf j}={\,\mathbf j}_\parallel+{\,\mathbf j}_\perp$.
Then the rate of ohmic heating is given by
\begin{equation}
\dot{q}_{\rm ohm}={\,\mathbf E}\cdot{\,\mathbf j}=\frac{j_\parallel^2}{\tilde{\sigma}_\parallel}
+ \frac{j_\perp^2}{\tilde{\sigma}_\perp^{\rm eff}},
\end{equation}
where
\begin{equation}
\tilde{\sigma}_\perp^{\rm eff}=\tilde{\sigma}_\perp+\frac{\tilde{\sigma}_{\rm H}^2}{\tilde{\sigma}_\perp}
\approx \frac{\tilde{\sigma}_{\rm H}^2}{\tilde{\sigma}_\perp}
\end{equation}
is the effective conductivity perpendicular to ${\,\mathbf B}$. Electron collisions resist
${\,\mathbf j}_\parallel$ and help conduct ${\,\mathbf j}_\perp$ with a non-zero component along ${\,\mathbf E}$.
Without collisions, ${\,\mathbf j}_\perp$ would be the pure drift current proportional to
${\,\mathbf E}\times{\,\mathbf B}$, which does not contribute to ohmic dissipation ${\,\mathbf E}\cdot{\,\mathbf j}$.
\begin{figure}[t]
\begin{tabular}{c}
\hspace*{-6mm}
\includegraphics[width=0.52\textwidth]{sigma.eps}
\end{tabular}
\caption{Components of the conductivity tensor in the crust with a
steady temperature profile that sustains $T_s=4\times 10^6$~K.
The temperature profile was calculated assuming a radial magnetic field
$B=3\times 10^{14}$~G and an iron envelope.
Temperature $T\approx 8.7\times 10^{8}$~K is approximately uniform
in the region of $\rho>10^{10}$~g~cm$^{-3}$ ($T$ steeply decreases toward
the surface in the blanketing envelope). In the presence of a heat source
in the crust at depth $z_h$, the curves can only be used at $z<z_h$.
}
\label{cond}
\end{figure}
The components of the conductivity tensor obey the following relations
(e.g. \citealp{1990A&A...229..133H}),
\begin{equation}
\tilde{\sigma}_\perp=\frac{\tilde{\sigma}_\parallel}{1+a^2}, \qquad \tilde{\sigma}_{\rm H}=a\tilde{\sigma}_\perp,
\end{equation}
where $\tilde{\sigma}_\parallel=(e^2n_e/m_e^\star)\tau_0$ is related to the electron collision
time $\tau_0$, $m_e^\star$ is the electron inertial mass, and $a=\tau_0\, eB/m_e^\star c$
is the magnetization parameter. For magnetar fields $a\gg 1$, and then
$\tilde{\sigma}_\perp^{\rm eff}\approx\tilde{\sigma}_\parallel$. Therefore, one can use the simple
equation,
\begin{equation}
\dot{q}_{\rm ohm}=\frac{j^2}{\tilde{\sigma}}, \qquad \tilde{\sigma}\approx\tilde{\sigma}_{\parallel}\approx\tilde{\sigma}_\perp^{\rm eff}.
\end{equation}
Figure~\ref{cond} shows $\tilde{\sigma}_\parallel$, $\tilde{\sigma}_\perp$, $\tilde{\sigma}_{\rm H}$
for a steady temperature profile with $T_s=4\times 10^6$~K and an iron envelope.
In the main region of interest, where $\rho=10^9-10^{11}$~g~cm$^{-3}$,
$\tilde{\sigma}\sim 10^{22}$~s$^{-1}$. Note also that in the region where heating occurs the
conductivity will be reduced, because of the local increase in temperature.
\subsection{Dissipation of electric currents in the crust}
The timescale for dissipating electric currents that sustain variations $\delta B$
on a scale $\ell$ is
\begin{equation}
\label{eq:tohm}
t_{\rm ohm}= \frac{4\pi\tilde{\sigma} \ell^2}{c^2}
\approx 4\times 10^4\, \tilde{\sigma}_{22}\,\ell_{\rm km}^2 {\rm~yr}.
\end{equation}
This timescale in the upper crust may be comparable to the magnetar age of 1-10~kyr
if the field varies on a scale $\ell\sim 0.3$~km. The corresponding electric current,
\begin{equation}
j\sim \frac{c}{4\pi}\,\frac{\delta B}{\ell},
\end{equation}
produces the heating rate
\begin{equation}
\dot{q}_{\rm ohm}
\sim \frac{(\delta B)^2}{4\pi t_{\rm ohm}}
\sim 6\times 10^{18}\,(\delta B_{16})^2\,\ell_{\rm km}^{-2}\,\tilde{\sigma}_{22}^{-1}
{\rm ~erg~s}^{-1}{\rm cm}^{-3}.
\end{equation}
A minimum heating rate $\sim3\times 10^{19}$~erg~s$^{-1}$~cm$^{-3}$ capable
of sustaining $T_s\sim 4\times 10^6$~K \citep{2014MNRAS.442.3484K}, can be
achieved if the field varies on a small scale $\ell\sim 0.3$~km and these variations
are large, $\delta B\sim 10^{16}$~G, which requires an ultrastrong field, $B>10^{16}$~G.
Such crustal fields were invoked by \citet{2011ApJ...741..123P} to explain the surface
luminosity of magnetars. Their model of AXP~1E~2259+586 assumes
a toroidal magnetic field $B=2.5\times 10^{16}$~G hidden in the middle of the crust,
which drops toward the core and toward the stellar surface on a scale comparable
to 0.3~km. Similar configurations with weaker fields evolving due to the combined effects
of Hall drift and ohmic dissipation were simulated by \citet{2009A&A...496..207P}
and \citet{2013MNRAS.434..123V}.
They argued that the magneto-thermal evolution of crustal fields can explain
the observed properties of a broader class of neutron stars, not only magnetars.
The requirements to the ohmic heating model are illustrated in more detail
by the following calculation.
Let $z_h$ be the characteristic depth where the ohmic heating occurs.
The corresponding heated volume is $V=\Delta z\,A$, where $\Delta z$
is the thickness of the heated layer and $A\mathrel{\copy\simlessbox} 10^{13}$~cm$^2$ is its area.
Suppose this heating sustains the observed surface temperature
$T_s\approx 4\times 10^6$~K. The heat transfer equation determines
the required heating rate $F_h=\Delta z\,\dot{q}_{\rm ohm}$ and $T(z_h)$.
The calculation is simplified if we use the approximation of a thin
heated layer $\Delta z\ll z_h$ (Section~\ref{balance}). Then the required $F_h$
is independent of $\Delta z$, and a realistic $\Delta z\mathrel{\copy\simlessbox} z_h$ only enters
at the final step when evaluating the required $\dot{q}_{\rm ohm}=F_h/\Delta z$.
The obtained temperature $T(z_h)$ determines the conductivity $\tilde{\sigma}(z_h)$,
and one can find $|\nabla\times{\,\mathbf B}|=(4\pi/c)(\tilde{\sigma}\dot{q}_{\rm ohm})^{1/2}$
that is required in the heated region.
\begin{figure}[t]
\begin{tabular}{c}
\hspace*{-6mm}
\includegraphics[width=0.52\textwidth]{delb1e15.eps}
\end{tabular}
\caption{The required $|\nabla\times{\,\mathbf B}|=(4\pi/c)j$ in the ohmically heated layer
if the heating is to sustain $T_s=4\times 10^6$~K. The required $|\nabla\times{\,\mathbf B}|$
depends on the position of the ohmic heater $z_h$. The thickness of the heated
layer was estimated as $\Delta z=z_h/2$. Magnetic field $B=10^{15}$~G is
assumed and two cases are shown: $\Theta_B=0$ (radial field) and $\Theta_B=60^{\rm o}$.
}
\label{fig:ohm}
\end{figure}
The result of this calculation is shown in Figure~\ref{fig:ohm} as a function of $z_h$,
assuming $\Delta z=z_h/2$. One can see that $|\nabla\times{\,\mathbf B}|>10^{12}$~G~cm$^{-1}$
is required by a successful ohmic heating model, which corresponds to
field variations $\delta B\mathrel{\copy\simgreatbox} 10^{16}$~G on a 0.1-km scale.
If the field is dominated by a non-radial component, heat conduction across the envelope
is reduced; then for a heater located deep below the surface it becomes impossible to
sustain $T_s=4\times 10^6$~K regardless of the ohmic power $\dot{q}_{\rm ohm}$.
The required temperature at $z_h$ becomes so high that neutrino losses prevent
from reaching it, leading to the runaway of the required $F_h$ and $|\nabla\times{\,\mathbf B}|$.
The ultrastrong crustal fields invoked by the ohmic heating model imply the following
special feature.
Magnetic energy density $B^2/8\pi\approx 4\times 10^{30}B_{16}^2$~erg~cm$^{-3}$
exceeds the crustal shear modulus $\mu\sim 10^{28}\rho_{12}$~erg~cm$^{-3}$,
and hence the maximum elastic stress $\sigma_{\max}\sim 0.1\mu$
is far below the magnetic stress. In this situation, the
crust should be viewed as an incompressible stratified liquid, with practically zero
tolerance to unbalanced shear stresses. In particular,
in an axisymmetric configuration, the toroidal component
of the Lorentz force cannot develop, ${\mathbf e}_\phi\cdot({\,\mathbf j}\times{\,\mathbf B})/c\approx 0$.
This condition implies that the poloidal current ${\,\mathbf j}_p$
is nearly parallel to the poloidal magnetic field ${\,\mathbf B}_p$,
\begin{equation}
\label{eq:jp}
{\,\mathbf j}_p\times{\,\mathbf B}_p\approx 0.
\end{equation}
As long as the strong currents are confined to the crust,
Equation~(\ref{eq:jp}) requires that the current-carrying field lines are
also closed below the stellar surface.
Another special feature of this configuration is that the effect of Hall drift is limited
(cf. the end of Section~5.2).
Like the magnetized liquid in the ocean, the magnetically dominated solid crust
should follow the field in its relaxation to the lowest MHD equilibrium accessible
through horizontal plastic shear motions
(vertical motions are constrained by the stable stratification of the crust).
The class of such constrained MHD equilibria is rather broad
\citep{2013MNRAS.433.2445A}.
\subsection{Ohmic dissipation in current sheets}
Currents sheets with thickness $\ell\ll 0.1$~km would produce a high local dissipation
rate $\dot{q}_{\rm ohm}=j^2/\tilde{\sigma}$. The immediate result is the growth of thickness $\ell(t)$ on
the timescale $t_{\rm ohm}$ given by Equation~(\ref{eq:tohm}). This limits the energy dissipated
at given $\ell$ before the current sheet doubles its thickness.
The magnetic energy that is released by a current sheet of area $A$ and thickness $\ell$
sustaining a field jump $\delta B$ is
\begin{equation}
E_{\rm diss} \sim A\,\ell\,\frac{(\delta B)^2}{8\pi}
\sim 4\times 10^{45}\,A_{12}\,\ell_{\rm km}\,(\delta B_{15})^2.
\end{equation}
Feeding the magnetar surface luminosity during its lifetime $t\sim 10^{11}$~s
requires large heat $E_h\sim \epsilon^{-1} t \mathscr L_s\sim 10^{46}\epsilon^{-1}$~erg, which
implies tapping into magnetic energy in a large fraction of the crust volume.
Therefore, formation of thin current sheets by itself is insufficient to explain the
surface luminosities of magnetars. The high rate of ohmic dissipation could only
be sustained if some process prevents the current sheet from thickening and also
advects magnetic energy into it, feeding its dissipation power.
Hall drift is a process that could in principle do this.
In particular, consider a horizontal field $B_y$ which varies in the orthogonal horizontal
direction $x$;\footnote{In the presence of other components of the magnetic field,
the current sheet formation is less efficient \citep{2004MNRAS.347.1273H};
therefore we focus here on the simple and most optimistic configuration $B_y(x)$.}
the vertical $z$-axis is chosen along the electron density gradient $\nabla n_e$.
As long as resistivity is neglected, the Hall drift of $B_y$ is described by
\begin{equation}
\frac{4\pi e}{c}\frac{\partial B_y}{\partial t}=\frac{d}{d z}\left(\frac{1}{n_e}\right) B_y\,\frac{\partial B_y}{\partial x}.
\end{equation}
Its behaviour is similar to a non-linear wave described by Burger's equation, as
discussed by \citet{2000PhRvE..61.4422V}. The profile of $B_y(x)$ can continue to
steepen until resistivity becomes important and the magnetic diffusivity offsets the
steepening. Then a current sheet of a small thickness $\ell$ will be sustained.
The resulting energy dissipation rate is controlled by the speed of Hall drift
that advects magnetic energy toward the current sheet. This rate is formed outside
the current sheet and independent of its thickness $\ell$.
Thus, tapping into magnetic energy stored in a large volume anyway relies
on electric currents far from the current sheet.
The large-scale Hall drift transports energy slowly, in particular in the deep
dense crust that takes most of the volume and stores most of the magnetic energy.
The fastest energy transport due to Hall drift would occur in small-scale Hall
waves propagating along the magnetic field lines with the group speed
$v_{\rm gr}=cBk/2\pi en_e$ where $k$ is the wavenumber \citep{1992ApJ...395..250G}.
However, very short waves are ohmically damped. The shortest waves that can
propagate an interesting distance $H\mathrel{\copy\simgreatbox} 10^4$~cm have
\begin{equation}
k_{\max}\sim \frac{\tilde{\sigma} B}{en_eH},
\end{equation}
and their energy transport time is
\begin{equation}
t_{\min}\sim \frac{H}{v_{\rm gr}^{\max}}
\sim 10\, H_4^2\,n_{e,36}^2 B_{15}^{-2}\tilde{\sigma}_{24}^{-1} {\rm~yr}.
\end{equation}
A mechanism generating short Hall waves could lead to fast energy transport across
the crust and assist ohmic or mechanical dissipation; this scenario is investigated
in \citet{2016arXiv160604895L} and also found incapable of sustaining the surface
luminosity of persistent bright magnetars.
Another possibility for creating current sheets was considered by
\citet{2001ApJ...561..980T}. In their scenario, magnetar starquakes produced
crustal fractures with localized shear. Shear localization along a fault surface would
create a jump of the (tangential) magnetic field --- a current sheet.
This could occur if the crust breaks and slides along a magnetic flux surface ---
otherwise the transverse field suppresses such sliding \citep{2012MNRAS.427.1574L}.
It was proposed that the current sheets induced by crustal fractures could quickly
dissipate a large magnetic energy through reconnection \citep{2001ApJ...561..980T,2002ApJ...580L..69L}.
This scenario is however problematic. Strong magnetic fields may exist when they are
rooted in the deep crust, which keeps the field in place.
The current sheet created by localized shear is immersed
in a guide field that is frozen in the lower crust and therefore cannot be moved
out of the sheet, inhibiting reconnection.\footnote{For a similar reason the current
sheet hugging the closed magnetosphere of a rotation-powered pulsar is stable.
Direct plasma simulations of pulsar magnetospheres show fast reconnection
only in the equatorial part of the current sheet outside the light cylinder, where
a guide field is absent \citep{2014ApJ...795L..22C,2015ApJ...801L..19P,2015MNRAS.448..606C}.}
The current sheet will simply thicken
with time due to resistive magnetic diffusion, and ohmic dissipation will
become slow before tapping into the larger reservoir of magnetic energy.
A network of $N\gg 1$ fractures occupying a large region of scale $L$
would reduce the distance between the multiple current sheets to $L_0=L/N$.
However, it would also reduce the field jump $\delta B\sim B/N$
in each sheet, resembling a staircase with many small stairs.
Only a small magnetic energy converts to heat before ohmic
dissipation washes out the ``stairs'' and makes the field profile smooth.
This energy may be estimated as
\begin{equation}
E_h\sim V\,\frac{(\delta B)^2}{8\pi}\sim N^{-2}\,V\,\frac{B^2}{8\pi},
\end{equation}
where volume $V\mathrel{\copy\simlessbox} 10^{18}$~cm$^3$ does not exceed the volume of the crust.
The dissipation timescale for this small energy is short,
$t_{\rm ohm}=4\pi\tilde{\sigma} L_0^2/c^2\sim 4\pi\tilde{\sigma} L^2/N^2c^2$.
However, dissipation of the main magnetic energy can only occur on a long ohmic
timescale that corresponds to scale $L$ comparable to the size of the magnetic
energy reservoir. In summary, we do not find any scenario for
efficient crustal heating by current sheets.
\section{External heating}
Magnetar surface can be heated by relativistic magnetospheric particles.
Clear evidence for magnetospheric activity is provided by hard X-ray observations:
persistent magnetars show a strong nonthermal component in their spectra,
peaking at photon energies $E>100$~keV
\citep{2008arXiv0810.4801K,2010PASJ...62..475E}. The power released in the
magnetosphere exceeds the surface luminosity $\mathscr L_s$,
and partial reprocessing of this power may be sufficient to feed $\mathscr L_s$.
The source of hard X-rays was identified as a decelerating outflow of copious
$e^\pm$ pairs in the closed magnetosphere
\citep{2013ApJ...777..114B,2013ApJ...762...13B,2014ApJ...786L...1H,2015ApJ...807...93A}.
The $e^\pm$ fountain forms near the neutron star and radiates the observed hard
X-rays at several stellar radii before reaching the top of the closed magnetic loop and
annihilating there. The model successfully fitted the variation of the observed spectrum
with rotational phase, and the fits determined the location of the $e^\pm$ fountain,
in particular in 1RXS J1708-4009 and AXP~1E~1841-045. The fountain
typically operates on 1-10\% of magnetic field lines emerging from the star,
which form a twisted bundle carrying electric current ${\,\mathbf j}=(c/4\pi)\nabla\times{\,\mathbf B}$;
the observed activity is the result of electric discharge in this ``j-bundle.''
These results imply
that the hard X-ray emission is
directed away from the star and cannot heat its surface. However, a significant fraction
of the primary {\it particles} created by the discharge near the star are expected to
flow toward the surface.\footnote{The energy flow from the discharge zone toward
the star is carried by relativistic particles rather than photons.
The main radiative process for the particles is resonant scattering of soft X-rays,
and in the ultrastrong field near the star it gives so energetic photons that they
immediately convert to $e^\pm$ pairs \citep{2007ApJ...657..967B}.
In contrast, particles that flow away from the star and reach $B<10^{13}$~G
eventually radiate almost all their energy through resonant scattering.
}
These particles must bombard the surface and heat it, forming a hot spot at the
footprint of the j-bundle.
Strong observational evidence for external heating exists for transient magnetars.
A canonical transient magnetar, e.g. XTE~J1810-197, shows an outburst followed by a decay
of emission on a timescale of months to years, returning to the quiescent state
\citep{2007Ap&SS.308...79G}. The outburst results from a shear motion of
the magnetar surface twisting the external magnetosphere, which is followed by
gradual untwisting on the resistive timescale. The timescale is regulated by the
discharge voltage $\Phi\sim 10^{10}$~V that sustains the magnetospheric current ${\,\mathbf j}$.
Electrodynamics of untwisting requires that the current becomes localized on a fraction
of magnetic field lines, forming the j-bundle, and this fraction slowly shrinks with time
\citep{2009ApJ...703.1044B}. As the j-bundle shrinks so does its hot footprint.
Figure~\ref{fig:spots} summarizes observations of shrinking hot spots
in seven transient magnetars. The observed evolution of the spot area $A$ and
luminosity $\mathscr L$ agrees with the special trend predicted by the untwisting
magnetosphere model:
$A$ and $\mathscr L$ decrease with time. The slope of the $\mathscr L$-$A$ relation
(controlled by the behavior of $\Phi$) varies between 1 and 2, in the theoretically
expected region of the $\mathscr L$-$A$ plane.
The typical timescale of this evolution --- months to years ---
is also consistent with theoretical expectations, although there are outliers that
require a more detailed modeling.
\begin{figure}[t]
\hspace*{-0.5cm}
\includegraphics[width=0.53\textwidth]{spots.eps}
\caption{The evolution of hot spots observed on transient magnetars following
their outbursts. The spot shrinks (its area $A$ decreases) and becomes dimmer
(its luminosity $\mathscr L$ decreases) with time, forming tracks on the $A$-$\mathscr L$ plane.
The theoretical prediction is shown by the strip between the two lines,
$\mathscr L\sim 1.3 \times 10^{33} K\,A_{11}^2$~erg~s$^{-1}$, where $K=B_{14} \Phi_9 \psi$
\citep{2009ApJ...703.1044B}.
The value of $K$ depends on the discharge voltage $\Phi\sim 10^9-10^{10}$~V, the
twist angle $\psi\sim 1$, and the characteristic surface magnetic field $B$.
The strip shown in the figure corresponds to $1<K<20$,
however a broader range is possible, and $K$ may evolve during the outburst.
Data for SGR 1745-2900 are from \citet{2015MNRAS.449.2685C};
CXOU J1647-45 from \citet{2011ApJ...726...37W} and \citet{2013ApJ...763...82A};
Swift J1822.3-1606 from \citet{2012ApJ...754...27R};
SGR 0418+5729 from \citet{2010MNRAS.405.1787E};
SGR 0501+4516 from \citet{2009MNRAS.396.2419R};
XTE J1810-197 from \citet{2007Ap&SS.308...79G};
1E 1547-5408 from \citet{2008ApJ...676.1178H} and \citet{2010PASJ...62..475E}.
The distance to 1E~1547-5408 was changed to 4~kpc following
\citet{2010ApJ...710..227T} and \citet{2007ApJ...667.1111G}.
}
\label{fig:spots}
\end{figure}
The predicted and observed localization of external heating
in transient magnetars suggests that this mechanism does not dominate
$\mathscr L_s$ in persistent magnetars, as most of their surface emission apparently
comes from a large area comparable to $4\pi R^2$.
\section{Discussion}
The observed surface luminosity of persistent magnetars
$\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$ is a challenge to magnetar theory.
Energy transport from the core heated by ambipolar diffusion
is an attractive scenario, which lead \citetalias{1996ApJ...473..322T} to propose
an explanation for $\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$: it corresponds to
the highest core temperature that ambipolar heating could sustain against neutrino cooling.
We find, however, that this scenario faces the following problem. Even in the best
case of a magnetar with a light-element envelope,
$T_{\rm core}\mathrel{\copy\simgreatbox} 6\times 10^8$~K is required (Figure~\ref{fig_Tc}). Although ultrastrong
magnetic fields can drive a fast ambipolar drift that generates a huge heating rate, we
find that such hot cores have lifetimes shorter than the typical magnetar age
(Figure~\ref{fig:Tc}), as long as the typical wavenumbers of the variation of ${\,\mathbf B}$ in the
core satisfy the plausible assumption $2\pi/\mathbb{k}\mathrel{\copy\simlessbox} 20$~km.
The lifetime is short because the ambipolar drift is fast in the hot core. It is not slowed
down by the induced pressure gradients in a compressive drift and is only
limited by the p-n friction, which is modest at high temperatures. Assuming stronger
magnetic fields helps increase the energy reservoir available for dissipation, however it also
accelerates its dissipation, with enormous heat promptly released and lost to neutrino
emission. The hot stage $T_{\rm core}>6\times 10^8$~K becomes particularly short if the core becomes
superfluid at this stage, as the transition to superfluidity both speeds up the ambipolar
drift and enhances neutrino cooling.
The issue of short lifetime could be resolved if ambipolar drift is intermittent,
which would allow the magnetar to enter ``ice ages'' between hot periods.
This would help explain the $1-10$~kyr ages of currently observed hot magnetars.
Objects classified as ``persistent'' after 4 decades of observations may not be truly
persistent on longer timescales; their appearance may dramatically change over centuries.
The surface luminosity would respond to changes in the core heating on the thermal
conduction timescale, which is comparable to a few years.
Note that the reduced duty cycle of magnetar activity would imply a large number of
undetected quiescent objects. Then the inferred magnetar population is increased from
10-20\% to more than half of all neutron stars with age less than 10~kyr.
Evidence for the dormant population is provided by the growing
number of transient magnetars. They are discovered in their outbursts of activity,
which are followed by the decay to the quiescent state.
It is unclear whether heating of the core can become intermittent due to
complicated dynamics of the magnetic field.
The dynamics may be influenced by current sheets, which are naturally
created by ambipolar diffusion (\Sect~\ref{ambipolar} and Appendix~\ref{app}).
Three-dimensional global simulations of ambipolar drift may clarify the
possibilities and limitations for variable core heating.
An alternative location for the internal heat source is the crust surrounding
the liquid core. This possibility became popular in recent years
(e.g. \citealp{2006MNRAS.371..477K,2011ApJ...727L..51P,2014ApJ...794L..24B}),
and we have examined it here in some detail.
Two mechanisms can heat the crust: mechanical dissipation and ohmic dissipation.
The dissipative shear deformations can be triggered by the slowly evolving magnetic
field inside the star. However, we find that even with most optimistic assumptions, this
scenario can hardly sustain the observed surface luminosity of persistent magnetars.
We have calculated upper limits on mechanical heating that result from two
general constraints: (1) the mechanical heating must occur in the solid phase below
the deep melted ocean, and (2) the heating rate is proportional to the shear stress,
which cannot exceed $\sigma_{\max}\sim 0.1\mu$, where $\mu$ is the shear modulus
of the crustal material. Mechanical heating is also proportional to the crustal shear rate
$\dot{s}$. The maximum average $\dot{s}$ over the magnetar lifetime fails to generate
the observed surface luminosity $\mathscr L_s\approx 10^{35}$~erg~s$^{-1}$. Therefore,
we have also considered the possibility of {\it oscillating} plastic shear driven by
crustal Hall waves and have shown that it also obeys an upper limit, which
cannot sustain the observed $\mathscr L_s$ over the magnetar lifetime (\Sect~\ref{5.2}).
This mechanism can, however, explain the intermittent heating observed in transient
magnetars (see \citet{2016arXiv160604895L}).
Ohmic heating approaches the needed rate only for extreme magnetic configurations
with crustal fields $B>10^{16}$~G varying on a scale of 100~m (Figure~\ref{fig:ohm}).
For instance, an ultrastrong toroidal field stored in the crust can be considered as an
ohmic heater \citep{2011ApJ...741..123P}. However, it is unclear how so
energetic magnetic torus could form and
remain confined to the crust of a nascent magnetar; such configurations were
not seen among calculated stable MHD equilibria \citep{2009MNRAS.397..763B}.
We have further explored the possibility of crustal ohmic heating by localized current
sheets envisioned by \citet{2001ApJ...561..980T} and \citet{2002ApJ...580L..69L}.
We found no way for the crustal current sheets to efficiently dissipate the magnetic
energy that would explain the observed surface luminosities.
The difficulties with finding a compelling internal heating mechanism suggest the
possibility that magnetars are heated as a result of their magnetospheric activity.
In particular, magnetospheric flares create strong intermittent dissipation in the crust.
The flares launch powerful Alfv\'en waves \citep{2013ApJ...774...92P}
which induce plastic flow in the crust and dissipate in $\sim 10$~ms
\citep{2015ApJ...815...25L}. This impulsive heating occurs immediately below the
melted ocean, and heat conduction from this region sustains a high surface
temperature for $\sim 1$~yr with a relatively high efficiency $\epsilon$ (Figure~\ref{fig:trans}).
Repeated flares could keep the magnetar surface hot for a longer time.
In this picture, $\mathscr L_s\sim 10^{35}$~erg~s$^{-1}$ requires an average power
released in the magnetospheric flares $\mathscr L_f\sim 10^{36}$~erg~s$^{-1}$.
Curiously, this $\mathscr L_f$ is comparable to the persistent nonthermal luminosity
estimated from the hard X-ray observations of persistent magnetars.
A flare of total energy $E_f$ produces surface afterglow with energy $E_{\rm aft}=\epsilon\, f_{\rm wave} E_f$,
where $f_{\rm wave}$ is the energy fraction given to the Alfv\'en waves damped in the crust.
The fraction $1-f_{\rm wave}$ is promptly radiated away during the flare, and the ratio of the
energies radiated in the prompt phase and its crustal afterglow is
\begin{equation}
\frac{E_{\rm aft}}{E_{\rm prompt}}=\frac{\epsilon\,f_{\rm wave}}{1-f_{\rm wave}}.
\end{equation}
If the magnetospheric flares occur much more frequently than once per year, the
afterglow luminosity becomes quasi-steady. For instance, flares with
$E_f\sim 10^{42}$~erg and a rate of 30~yr$^{-1}$ would sustain a surface
luminosity $\mathscr L_s\sim 10^{35}~(\epsilon/0.1)\, f_{\rm wave}$~erg~s$^{-1}$.
Each flare could involve a strong deformation of a ``flux rope'' carrying a
fraction of the stellar magnetic flux. A large number of such localized
flares could occur in a complicated magnetic field, with many twisted flux ropes.
A problem with this scenario is that the high flare rates are not observed with
current instruments. Most of them would need to be hidden by assuming that their
prompt phase is ``dark'': $1-f_{\rm wave}\ll 1$, so that most of the released magnetic energy
goes to the excitation of Alfv\'en waves.
Finally, magnetars must be heated by relativistic magnetospheric particles bombarding
the stellar surface. This external heating accompanies long-lived twists of the
magnetosphere, $\nabla\times{\,\mathbf B}\neq 0$, which imply long-lived electric currents ${\,\mathbf j}$.
The currents are sustained (and gradually dissipated) through continual electric
discharge that fills the active j-bundle with relativistic $e^\pm$ pairs, and some of
these particles bombard the footprint of the j-bundle.
Figure~\ref{fig:spots} shows observational evidence for this mechanism in
transient magnetars --- the shrinking hots spots predicted by electrodynamics
of resistive ``untwisting'' \citep{2009ApJ...703.1044B}.
Similar localized heating is expected to operate in persistent magnetars, however,
it appears insufficient to explain emission with large surface area $A>10^{12}$~cm$^2$.
A related puzzle of persistent magnetars is
that their magnetospheres stay twisted much longer than in transient magnetars.
In particular, AXP~1E~1841$-$045 has been producing approximately steady
nonthermal emission for at least one decade. Its phase-resolved hard X-ray spectrum
is well reproduced by the model of $e^\pm$ flow in the j-bundle, and
the soft X-ray component may be described as two blackbodies --- the warm stellar
surface + the hot j-bundle footprint \citep{2015ApJ...807...93A}.
At the same time, the nonthermal luminosity implies a short timescale for
ohmic dissipation of the magnetospheric twist $t_{\rm diss} \approx 0.1\, \psi^2$~yr,
which can hardly exceed $\sim 1$~yr (here $\psi\mathrel{\copy\simlessbox} \pi$~radian is the twist amplitude).
To survive a decade, this configuration requires energy supply from the star,
and it is unclear how the system finds a steady state.
If the magnetar surface is indeed heated by the magnetospheric activity (through
damping of Alfv\'en waves or particle bombardment) this still relies on
a primary driver inside the star, regardless of how dissipative or quiet
it may be. In particular, sustaining the magnetospheric twists
against ohmic decay requires continual (or frequent) shear motions of the crust,
which must be driven by the internal fields. The ultimate energy source for both
magnetospheric emission and surface glow must be the
magnetic energy stored inside the star.
Two processes can build up internal stresses that drive crustal motions: Hall drift in
the crust and ambipolar diffusion in the core. Both, however, have their limitations.
The Hall driver obeys a strong upper limit given by Equation~(\ref{eq:sH}). Hall drift can
generate significant transient shear \citep{2016arXiv160604895L}
but not the persistent activity with luminosity exceeding $10^{35}$~erg~s$^{-1}$.
Ambipolar diffusion naturally creates stresses at the bottom of the crust and can force it
to flow, allowing the helical field in the core to unwind \citep{2002ApJ...574..332T}.
The limitation here is the net flow/unwinding angle
$\Delta s\sim (B_\phi/B_r)_{\rm core}<10^2$.
The unwinding motion with $\Delta s\sim 10^2$ could
sustain the magnetospheric activity for only $\sim 10^2$~yr,
if it occurs with the optimal rate $\dot{s}\sim 1$~rad~yr$^{-1}$ that is just sufficient to
offset ohmic decay of the magnetospheric twist.
The external activity would last longer if the internal field has many twisted
domains that unwind at different times, creating a kind of a firework with the overall duration
longer than the output of each individual domain. This could perhaps bring the time-span
of activity to the observed magnetar ages of $\sim 10$~kyr.
There is some observational support for the intermittency of the magnetic flux emerging
from magnetars, consistent with the picture of patches of concentrated flux (flux tubes).
Evidence for an active flux tube with a magnetic field much stronger than the
average (dipole) field was found in SGR~0418+5729 \citep{2013Natur.500..312T}.
\acknowledgements
We thank Ashley Bransgrove and the referee for useful comments on the manuscript.
This work was supported by NASA grant NNX13AI34G
and a grant from the Simons Foundation (\#446228, Andrei Beloborodov).
\begin{appendix}
\section{Approximate model for ambipolar diffusion} \label{app}
The one-dimensional model with the initial magnetic field $B(x)=B_0\sin(\mathbb{k} x)$
is illustrated in the left panel of Figure~\ref{fig:L1}.
The region $0<x<L_1$ is shrinking with rate $\dot{L}_1$ that is twice the
local plasma speed $v_1=v(L_1)$. Note that the magnetic flux in this region
$\Psi_1=\int B\,dx=B_1L_1/2$ is decreasing, which is only possible if the boundary
$L_1$ moves faster than the plasma. The flux transport across the boundary $L_1$
is described by $\dot{\Psi}_1=(-v_1+\dot{L}_1) B_1$, which gives
\begin{equation}
\dot{L}_1=2v_1.
\end{equation}
\begin{figure}[]
\begin{tabular}{c}
\includegraphics[width=0.47\textwidth]{sine.eps}
\includegraphics[width=0.47\textwidth]{L1.eps}
\end{tabular}
\caption{Left: the profile of the magnetic field $B(x)$, in its initial and
final states. The evolution is indicated by the two arrows: $L_1$ shrinks, making
the profile steeper near the null points, while the maximum $B$ decreases,
making the profile flatter between the null points.
The resulting final state is close to a step function, with a steep jump of $B$
near the null point, which is supported by a thin current sheet.
Right: Evolution of the
current sheet half-thickness $L_1$ in the three models shown in Figure~\ref{fig:Tc}.
Solid part of the curve shows the friction-dominated stage and dashed part
shows the pillow stage. The moment of the hydrostatic pillow formation near the
null point corresponds to the peak in temperature seen in Figure~\ref{fig:Tc};
the shrinking of $L_1$ is fastest at this moment.
The curves end when $L_1$ reaches $L_{\min}$ estimated in Equation~(\ref{eq:Lmin}).
}
\label{fig:L1}
\end{figure}
As long as the plasma speed $v$ is regulated by the p-n friction (as in
Equation~(\ref{eq:vfric})), one finds
\begin{equation}
v_1=-\frac{\tau_{pn} B_1^2}{4\pi\rho_pL_1}, \qquad
\frac{dL_1}{dt}=-\frac{\tau_{pn} B_1^2}{2\pi\rho_pL_1}, \qquad {\rm (friction~dominated)}
\end{equation}
which would lead to the singularity $L_1\rightarrow 0$ in a finite time.
This model is, however, incomplete, because it neglects the build up of
pressure near the null point, which can slow down the compression.
The pressure gradient remains negligible as long as Murca reactions sufficiently
quickly convert electrons and protons to
neutrons (which can flow out of the compressed region across the magnetic field).
Eventually this approximation breaks and the finite rate of Murca reactions
becomes an important limitation near the null point. This occurs when $L_1$
becomes smaller than the scale $a$ given in Equation~(\ref{eq:a}). Then a hydrostatic
pressure ``pillow'' is formed at $x=0$ which nearly offsets the surrounding
magnetic pressure $B_1^2/8\pi$,
\begin{equation}
n_e\Delta\mu\sim \frac{B_1^2}{8\pi}, \qquad L_1\ll a.
\end{equation}
Equation~(\ref{eq:compr}) now yields the following compression rate near the null point,
\begin{equation}
\frac{\partial}{\partial x} (n_ev)\approx -\frac{\lambda B_1^2}{8\pi n_e},
\qquad L_1\ll a.
\end{equation}
This gives $v(x)=v_1\,x/L_1$ with $v_1=-\lambda B_1^2 L_1/8\pi n_e^2$.
In summary, the compression rate of the current sheet $\dot{L}_1=2v_1$ is
controlled by p-n friction as long as $L_1\gg a$ and by Murca reactions in the
pillow when $L_1\ll a$. Equation~(\ref{eq:Ldot}) summarizes the two
regimes; the transition between them, $L_1=L_\star$,
is defined by matching the two formulas for $v_1$.
Both p-n friction and the Murca rate depend on temperature,
whose evolution is controlled by heating due to magnetic energy dissipation.
An approximate equation for magnetic dissipation may be derived as follows.
Consider the domain $0<x<x_0=\pi/2\mathbb{k}$ with the (conserved) total magnetic flux,
\begin{equation}
\Psi_0=\int_0^{x_0} B\,dx=\frac{B_0}{\mathbb{k}}=B_1 x_0=const,
\end{equation}
and the (decreasing) magnetic energy
\begin{equation}
E(t)=\int_0^{x_0}\frac{B^2}{8\pi}\,dx.
\end{equation}
We divide the domain into two parts:
\\
(1) In the current sheet $0<x<L_1$, we use the approximation $B(x)=B_1 x/L_1$.
The magnetic flux and energy of this region are given by
\begin{equation}
\Psi_1\approx \frac{B_1L_1}{2}, \qquad E_1\approx \frac{B_1^2L_1}{24\pi}.
\end{equation}
(2) In the region $L_1<x<x_0$, the magnetic flux $\Psi_2$ is
\begin{equation}
\Psi_2=\Psi_0-\Psi_1=B_1x_0-\frac{B_1L_1}{2}.
\end{equation}
A simple expression for the magnetic energy of this region is found
in the linear order of $B-B_1\ll B_1$, neglecting $(B-B_1)^2$,
\begin{equation}
E_2=\int_{L_1}^{x_0} \frac{B^2}{8\pi}\,dx
\approx \frac{B_1\Psi_2}{4\pi}-(x_0-L_1)\frac{B_1^2}{8\pi}=\frac{x_0 B_1^2}{8\pi}.
\end{equation}
It remains constant and equals the final energy of the entire domain $E_{\rm fin}$.
The total magnetic energy is then given by
\begin{equation}
\label{eq:energy}
E=E_1+E_2\approx \frac{B_1^2 L_1}{24\pi}+E_{\rm fin}.
\end{equation}
This equation should provide a good approximation to the magnetic energy
when $\mathbb{k} L_1\ll 1$. The initial state $B(x)=B_0\sin\mathbb{k} x$ has a large $L_1=2/\pi\mathbb{k}$;
in this case, our approximation underestimates the energy available for dissipation,
$E-E_{\rm fin}$, by a factor of 2.
Using the approximate relation between $L_1$ and magnetic energy provided by
Equation~(\ref{eq:energy}) one finds the volume-averaged heating rate $\dot{q}_h=-\dot{E}/x_0$,
which we use in Equation~(\ref{eq:Tdot}).
Inside the pillow ($x=0$), a significant $\Delta\mu$ is built up,
\begin{equation}
\label{eq:xi}
\xi=\frac{\Delta\mu}{kT}\sim \frac{B_1^2}{8\pi n_e kT}
\approx 2.9\, B_{1,16}^2\, T_9^{-1} \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{-2}.
\end{equation}
Therefore, $\lambda$ in Equation~(\ref{eq:Ldot}) must be evaluated using
the correction factor $H(\xi)$ (see Equation~(\ref{eq:lambda}) and \citet{1995ApJ...442..749R}),
\begin{equation}
\label{eq:H}
H(\xi)=1+\frac{189\,\xi^2}{367\pi^2}+\frac{21\,\xi^4}{367\pi^4}+\frac{3\,\xi^6}{1835\,\pi^6}.
\end{equation}
Note also that $\lambda$ and $\dot{q}_\nu$ are related, since both depend on the rate
of Murca reactions. This relation is given by \citep{2001PhR...354....1Y},
\begin{equation}
\label{eq:dqnu_lam}
\lambda_0=\frac{\lambda}{H(\xi)}=\frac{14680}{11513}\,\frac{\dot{q}_\nu^0}{(\pi kT)^2},
\end{equation}
where $\dot{q}_\nu^0$ is the Murca cooling rate at $\Delta\mu\ll kT$, and
$\lambda_0$ describes the rate of $\Delta\mu$ relaxation for $\Delta\mu\ll kT$.
We use $\dot{q}_\nu=\dot{q}_\nu^0$, because most of neutrino losses occur in the region
$x>L_1$ where $\Delta\mu$ remains small.
During the main heating stage there is an approximate balance between heating
and cooling $\dot{q}_h\approx\dot{q}_\nu$, which gives
\begin{eqnarray}
\label{eq:TL1}
L_1\approx \left\{\begin{array}{ll}
\vspace*{0.2cm}
\displaystyle{ \frac{\tau_{pn} B_1^4\,\mathbb{k}}{24\pi^3\rho_p\, \dot{q}_\nu}} & \quad L_1>L_\star, \\
\displaystyle{\frac{12}{\xi^2H(\xi)\,\mathbb{k}}} & \quad L_1<L_\star.
\end{array}
\right.
\end{eqnarray}
This provides a relation between $T$ and $L_1$, and then it is sufficient to solve one
differential equation, e.g. Equation~(\ref{eq:Tdot}) for $T(t)$. In particular, the
transition $L_1=L_\star$ typically occurs in the regime $\dot{q}_h\approx\dot{q}_\nu$.
One can solve for $\xi$ and $T$ at the transition by matching the two expressions
in Equation~(\ref{eq:TL1}) and using Equation~(\ref{eq:xi}),
\begin{equation}
\label{eq:xis}
\xi_\star\approx 4\;\mathbb{k}_{-5}^{-1/6}\,B_{1,16}^{4/3}\,\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{-3/2},
\qquad T_\star\approx 7.2\times 10^8\,\mathbb{k}_{-5}^{1/6}\, B_{1,16}^{2/3}
\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{-1/2} {\rm K},
\end{equation}
where we have used the approximation $\xi H^{1/12}\approx \xi$.
A significant deviation from the balance
$\dot{q}_h\approx\dot{q}_\nu$
develops at later stages; then Equation~(\ref{eq:TL1}) becomes invalid and the evolution is
found from the coupled differential equations for $T(t)$ and $L_1(t)$.
Figure~\ref{fig:L1} shows the evolution of $L_1(t)$ for the sample models presented
in Figure~\ref{fig:Tc}.
The initial evolution on the friction timescale takes less than 1~kyr, then the pillow
forms, however it does not stop the fast collapse of the current sheet.
The compression timescale $L_1/|\dot{L}_1|$ is then controlled by Murca reactions,
\begin{equation}
\label{eq:tl}
t_\lambda=\frac{4\pi n_e^2}{\lambda B_1^2}
\approx \frac{80 {\rm~yr}}{(B_{1,16})^2\, T_9^6\, H(\xi)}\,\left(\frac{\rho}{\rho_{\rm nuc}}\right)^{10/3}.
\end{equation}
An upper limit to this timescale is obtained if $\xi\gg 10$; then
$H(\xi)\approx (0.11\xi)^6$ and
$t_{\lambda}\approx 80\,(B_{1,16})^{-14} (\rho/\rho_{\rm nuc})^{46/3}$kyr.
However, before the regime $\xi\gg 10$ is approached, the effects of
a finite electric conductivity become important and stop the shrinking of $L_1$.
The effective conductivity (associated with ohmic dissipation) across
the magnetic field ${\,\mathbf B}$ is approximately equal to the conductivity along ${\,\mathbf B}$
(see \Sect~\ref{ohm}), which in the core is given by
$\tilde{\sigma}\approx 4.2\times 10^{26}\, T_9^{-2}(\rho/\rho_{\rm nuc})^3$~s$^{-1}$ \citep{1990A&A...229..133H}.
Magnetic diffusivity $\eta=c^2/4\pi\tilde{\sigma}$ stops the compression of $L_1$ when
$L_1v_1\sim c^2/4\pi\tilde{\sigma}$. This gives the minimum thickness of the current sheet,
\begin{equation}
\label{eq:Lmin}
L_{\min}\approx 1\,T_9 \left(\frac{\rho}{\rho_{\rm nuc}}\right)^{-3/2}
\left(\frac{t_\lambda}{1{\rm ~kyr}}\right)^{1/2} {\rm m}.
\end{equation}
\end{appendix}
|
\section{Introduction}\label{se:intro}
Stochastic games with a large number $n$ of players are notoriously intractable. Mean field games were introduced by Lasry and Lions~\cite{LasryLions.06a,LasryLions.06b,LasryLions.07} and Huang, Malham\'e, and Caines~\cite{HuangMalhameCaines.07, HuangMalhameCaines.06} to study Nash equilibria in the limiting regime where $n$ tends to infinity and the players interact symmetrically through the empirical distribution of the private states of all players. Given such a distribution $\mu$, each player typically solves a standard control problem; that is, controls a diffusion while paying some cost of effort. On the other hand, the reward (and possibly the diffusion) depend on $\mu$, which is in turn determined by the actions of all agents. In the analytic theory, such a system is described by a coupled system of nonlinear partial differential equations (PDEs): a Hamilton--Jacobi--Bellman equation describes the optimal control problem when $\mu$ is given, and a Kolmogorov-type equation describes the evolution of~$\mu$ over time as a result of the optimal controls. One of the major difficulties is that the former equation naturally starts from a terminal condition and runs backward in time, whereas the latter runs forward to ensure the consistency of $\mu$; we refer to \cite{Cardaliaguet.13,GamesSaude.14} for background. In a probabilistic version of the theory, the stochastic maximum principle is used and the system of PDEs is replaced by a coupled forward-backward stochastic differential equation; cf.\ \cite{BensoussanFrehseYam.13, CarmonaDelaRue.13, CarmonaDelaRue.14, CarmonaDelaRue.15}. In the simplest case, the agents are exposed to idiosyncratic i.i.d.\ noise (essentially, an independent Brownian motion for each diffusion equation) and thus the equilibria are formulated as deterministic. More recently, the presence of an additional common noise and stochastic equilibria have received considerable attention; see \cite{CarmonaDelaRueLacker.13,CarmonaLacker.15,Fischer.14,Lacker.14,Pham.16}.
A wide range of applications from production models to population dynamics have emerged over the last decade, several of them summarized in \cite{GueantLasryLions.11}; see also \cite{CarmonaFouqueSun.13} for a recent model of systemic risk and \cite{CarmonaLacker.15} for price impact in finance.
While mean field games were introduced as a tractable model for a large stochastic game, they are still rather complex. To the best of our knowledge, the only case that can be solved explicitly is linear-quadratic control (linear dynamics, quadratic cost). This situation has been studied in detail; see \cite{Bardi.12, BardiPriuli.14, BensoussanSungYamYung.16, CarmonaFouqueSun.13}. In other cases, one generally has to settle for an abstract description by a coupled system of nonlinear equations.
The main aim of the present paper is to formulate a tractable game of mean field type where the properties of equilibria can be understood somewhat more directly. In our case, the agents will be solving optimal stopping problems rather than diffusion control\footnote{
The possible interest of such a game was first pointed out to the author by Ren\'e Carmona. Section~2 of \cite{GueantLasryLions.11} can be seen as a predecessor, at least in spirit: in a toy example called ``When Does the Meeting Start?'' the agents indirectly control their arrival times at a prespecified location.
}.
While in a standard mean field game the (spatial) location of the players matters, the state space here becomes binary: each player either has stopped or is still in the game, and the interaction occurs through the number of players that have already stopped. This structure seems appealing due to its simple interpretation and a wide range of possible applications from bank-run models to traffic optimization. On the other hand, it produces an inherent discontinuity in the game: as is well-known in economics (e.g., \cite{DiamondDybvig.83, MorrisShin.04}), games of optimal timing may easily degenerate in that all players stop at the same time. Thus, one of the challenges is to produce a class of models where typical equilibria are non-trivial.
Specifically, we shall study a continuous-time stochastic game with a continuum of players. In equilibrium, each agent $i$ will be solving an optimal stopping problem of the form
$$%
\sup_{\tau} E\bigg[\exp\bigg(\int_{0}^{\tau} r_{s}\,ds\bigg) \mathbf{1}_{\{\theta>\tau\}\cup \{\theta=\infty\}}\bigg];
$$%
it has two competing forces. The process $r$ can be interpreted as a reward or \emph{interest rate} that is accrued as long as the agent does not stop, thus incentivizing the agent to stay in the game. On the other hand, there is a random time $\theta$ of \emph{default} (of the interest-paying institution): the agent will lose everything if $\theta$ happens before she leaves the game. While the default happens as a ``surprise'' to the agent, the distribution of $\theta$ is governed by an intensity process $\gamma^{i}$ that is known to the agent: the larger $\gamma^{i}$, the more likely it is that default happens soon. More precisely, $\theta$ is modeled as the first jump time of a Cox process with intensity $\gamma^{i}$. This leads to a tractable solution of the single-agent optimal stopping problem---we are taking inspiration from the finance literature
(e.g., \cite[Chapter~5]{Lando.09}) where it is well-known that a defaultable bond in a similar setting will be priced just like a non-defaultable one, but with an adjusted interest rate $r-\gamma^{i}$.
The agents are heterogeneous in their views on the distribution of the default---we think of the intensity $\gamma^{i}$ as depending on the subjective probability used by agent $i$. As a result, the players face different optimal stopping problems and may stop at different times. The agents' views on the default intensity will also be influenced by how many players have already stopped; more precisely, the proportion $\rho_{t}\in[0,1]$ of players that have left the game by time $t$. This process is observed by all agents and creates an interaction of mean field type: if $\rho_{t}$ is larger, the intensity of any player will also be larger, meaning that the perceived default will happen sooner. As in bank-run models, this represents that the default of the institution is more likely if more customers have abandoned ship.
While we defer the general formulation of the setting to Section~\ref{se:generalModel}, a typical model may postulate that $\gamma^{i}$ is of the form
\begin{equation}\label{eq:introAdditive}
\gamma^{i}_{t} = X_{t} + Y^{i}_{t} + c\rho_{t}.
\end{equation}
Here $X$ plays the role of a common noise (the same for all agents) whereas $Y^{i}$ is an idiosyncratic noise that will be i.i.d.\ within the population. Depending on the application, one may interpret $X$ and $Y^{i}$ as public and private signals, respectively, or see their sum as a noisy observation of the true signal $X$. Moreover, the constant $c\geq0$ governs the strength of interaction; that is, how much the agents' views are affected by $\rho_{t}$.
Suppose that $\tau^{i}$ is the stopping time chosen by agent $i$, and that the continuum of agents is represented by an atomless probability space $(I,\mathcal{I},\lambda)$. Then,
\begin{equation}\label{eq:introEquilib}
\rho_{t}(\omega) = \lambda\{i:\, \tau^{i}(\omega) \leq t\}
\end{equation}
is the ``proportion'' of players that have stopped prior to time $t$. This can also be seen as the cumulative distribution function (c.d.f.) at time $t$ of the empirical measure that describes the evolution of the system on $I\times \{0,1\}$, recording for each agent $i$ whether stopping has occurred (1) or not (0).
If we start with a given process $\rho$, the intensities $\gamma^{i}$ of the agents are determined. Let us suppose that the associated optimal stopping problems have solutions $(\tau^{i})_{i\in I}$. Tacitly assuming a suitable measurability, we may then consider the process $\lambda\{i:\, \tau^{i}(\omega) \leq t\}$, and if it satisfies~\eqref{eq:introEquilib}, we shall say that $\rho$ and $(\tau^{i})_{i\in I}$ form an equilibrium. Since we are working with a continuum of players, the decision of a single agent does not influence $\rho$, and hence this notion corresponds to a Nash equilibrium: given the strategies of the other players, each player is behaving optimally.
Our main result (Theorem~\ref{th:general}) relates equilibria $\rho_{t}$ to the solution of a finite-dimensional equation. For instance, in the case of~\eqref{eq:introAdditive}, it reads
\begin{equation}\label{eq:masterIntro}
1-u = F_{t}(r-x-cu),\quad u\in [0,1],
\end{equation}
where $F_{t}$ is the c.d.f.\ of the idiosyncratic noise $Y_{t}$ and $r$ is the (constant) interest rate. If $\rho(t,x)$ is the (maximal) solution $u$ at time $t$ and $X_{t}$ is the common noise, then $\rho(t,X_{t})$ describes an equilibrium, and a converse is also established. This simple equation allows us to understand the structure and multiplicity of equilibria in some detail. Two ingredients are important for the tractability of our setting. On the one hand, we use an Exact Law of Large Numbers to completely eliminate idiosyncratic randomness (the associated mathematical setup will be discussed later on). This idea and its many incarnations are well-known in economics; see \cite{Aumann.64, DuffieSun.10, KaratzasShubikSudderth.94, Sun.06} to cite but a few examples. On the other hand, we impose the structural assumption that $\gamma^{i}-r$ is increasing, and as we shall see, that leads to a simple solution of the single-agent problem. In comparison to the coupled forward-backward system of equations that is common in the literature on mean field games, one may say that the Hamilton--Jacobi--Bellman part becomes irrelevant because we know the solution of the single-agent problem (given $\rho$) in feedback form, whereas our equation represents the Kolmogorov forward equation---indeed, we may take time derivatives in~\eqref{eq:masterIntro} to find a PDE for $\rho$, or an ODE in the case without common noise.
While the main aim of the present paper is to formulate a tractable example of a mean field game of optimal stopping, there are important aspects that are not discussed. Three major questions are the passage to the limit from a game with $n$ players, what happens when the monotonicity condition is dropped, and the analysis of applications. Recent results on these can be found in \cite{CarmonaDelaRueLacker.17}, and we would like to emphasize that much of that work was carried out in parallel or before ours. In particular, it is shown that in our model of Section~\ref{se:additiveModel}, equilibrium strategies from the continuum formulation are $\varepsilon$-equilibrium strategies in an $n$-player game with large~$n$. Moreover, a general framework for mean field games of optimal stopping (or ``timing'') is introduced and analyzed, and applications to bank run models are discussed.
The remainder of this paper is organized as follows. In the next section, we describe the game more rigorously and analyze the single-player problem in detail, whereas Section~\ref{se:exactLLN} introduces the mathematical setting that allows for an Exact Law of Large Numbers. Section~\ref{se:toyModel} analyzes an insightful toy model without common noise; we discuss examples of (non-)uniqueness and the impact of noise and strength of interaction on the continuity of equilibria in time. Finally, Section~\ref{se:generalModel} treats the general model with common noise.
\section{Description of the Game}\label{se:basicSetup}
Let $(I,\mathcal{I},\lambda)$ be a probability space; each $i\in I$ will correspond to an agent. Moreover, let $(\Omega,\mathcal{F},P)$ be another probability space, to be used as the sample space. We suppose that $(\Omega,\mathcal{F},P)$ is equipped with right-continuous filtrations $\mathbb{G}^{i}=(\mathcal{G}^{i}_{t})_{t\in\mathbb{R}_{+}}$ and an exponentially distributed random variable $\mathcal{E}$ which is independent of $\mathbb{G}^{i}$ for all $i\in I$. We interpret $\mathbb{G}^{i}$ as the information available to agent $i$. Finally, let $r$ be a real-valued and locally integrable process (i.e., Lebesgue-integrable on bounded intervals) which is $\mathbb{G}^{i}$-progressively measurable for all $i\in I$; that is, observed by all agents.
\subsection{Single-Agent Problem}
We first consider the optimization problem for a fixed agent $i\in I$ and we denote by $\mathcal{T}^{i}$ the set of all $\mathbb{G}^{i}$-stopping times.
Let $\gamma^{i}\geq0$ be a $\mathbb{G}^{i}$-progressively measurable process which is locally integrable %
and consider the random time
$$
\theta^{i} = \inf\bigg\{t:\, \int_{0}^{t} \gamma^{i}_{s}\,ds = \mathcal{E}\bigg\}.
$$
One may think of $\theta^{i}$ as the first jump time of a Cox process with intensity~$\gamma^{i}$. The default time $\theta^{i}$ depends on $i$, which will allow us to write all optimal stopping problems under a common probability measure $P$. Alternately, we could deal with a single random time on a canonical space and endow the agents with subjective probabilities $P^{i}$. We have found the former solution easier to write, and they are equivalent in that the agents' decisions (and the equilibria) only depend on the distribution of the intensity.
For technical reasons, we shall assume that
\begin{equation}\label{eq:intCond}
\left.
\begin{matrix}
\text{$r^{+}$ is integrable on $[0,\infty)$, $P$-a.s.,} \;\\[.1em]
\text{or}\\[.1em]
\text{ $\inf\{t: \, \gamma^{i}_{t}-r_{t}\geq 0\}<\infty$, $P$-a.s.} \;
\end{matrix}
\right\}
\end{equation}
See also Remark~\ref{rk:noOptimizer} below for the necessity of such a condition. We then have the following result on the single-agent problem.
%
\begin{lemma}\label{le:singleAgent}
Suppose that $\gamma^{i}-r$ is increasing\,\footnote{Increase is to be understood in the non-strict sense throughout the paper.} and~\eqref{eq:intCond} holds. Then,
$$
\tau^{i} := \inf\{t: \, \gamma^{i}_{t}-r_{t}\geq 0\} \in\mathcal{T}^{i}
$$
is a solution of the optimal stopping problem\,\footnote{
We use the convention that $\int_{0}^{\infty} r_{s}\,ds:=-\infty$ if $\int_{0}^{\infty} r^{+}_{s}\,ds=\int_{0}^{\infty} r^{-}_{s}\,ds=\infty$.
}
\begin{equation}\label{eq:optStopProblem}
\sup_{\tau\in\mathcal{T}^{i}} E\bigg[\exp\bigg(\int_{0}^{\tau} r_{s}\,ds\bigg) \mathbf{1}_{\{\theta^{i}>\tau\}\cup \{\theta^{i}=\infty\}}\bigg] .
\end{equation}
If the value of \eqref{eq:optStopProblem} is finite, then $\tau^{i}$ is minimal among all solutions, and if, in addition, $\gamma^{i}-r$ is strictly increasing, then $\tau^{i}$ is the unique solution.
\end{lemma}
\begin{proof}
Due to the increase of $\gamma^{i}-r$ and the right-continuity of $\mathbb{G}^{i}$, the right limit process $\zeta$ of $\gamma^{i}-r$ exists and is $\mathbb{G}^{i}$-progressively measurable. As $\{\tau^{i}\leq t\} = \{\zeta_{t}\geq0\}\in\mathcal{G}^{i}_{t}$, we have $\tau^{i}\in\mathcal{T}^{i}$.
Let $\tau\in\mathcal{T}^{i}$ be such that $r^{+}$ is integrable on $[0,\tau)$, $P$-a.s. Using also the independence of $\mathcal{E}$ and $\mathbb{G}^{i}$, we have
\begin{align*}
P\bigg[\{\theta^{i}>\tau\}\cup \{\theta^{i}=\infty\}\bigg|\mathcal{G}^{i}_{\tau}\bigg]
&= P\bigg[\int_{0}^{\tau} \gamma^{i}_{s}\,ds < \mathcal{E}\bigg|\mathcal{G}^{i}_{\tau}\bigg] \\
&= E\bigg[\exp \bigg(-\int_{0}^{\tau} \gamma^{i}_{s}\,ds\bigg)\bigg|\mathcal{G}^{i}_{\tau}\bigg].
\end{align*}
Hence,
\begin{align*}
E\bigg[\exp\bigg(\int_{0}^{\tau} r_{s}\,ds\bigg) \mathbf{1}_{\{\theta^{i}>\tau\}\cup \{\theta^{i}=\infty\}}\bigg|\mathcal{G}^{i}_{\tau}\bigg]
= E\bigg[\exp\bigg( \int_{0}^{\tau} (r_{s}-\gamma^{i}_{s})\,ds \bigg)\bigg|\mathcal{G}^{i}_{\tau}\bigg]
\end{align*}
and finally
\begin{equation}\label{eq:adjustedInterest}
E\bigg[\exp\bigg(\int_{0}^{\tau} r_{s}\,ds\bigg) \mathbf{1}_{\{\theta^{i}>\tau\}\cup \{\theta^{i}=\infty\}}\bigg] = E\bigg[\exp\bigg( \int_{0}^{\tau} (r_{s}-\gamma^{i}_{s})\,ds \bigg)\bigg].
\end{equation}
If we are in the first case of~\eqref{eq:intCond}, our integrability condition holds for all $\tau\in\mathcal{T}^{i}$ and as $r-\gamma^{i}$ is decreasing, the representation on the right-hand side shows that $\tau^{i}$ is optimal. In the second case of~\eqref{eq:intCond}, as $r$ is locally integrable, we still have~\eqref{eq:adjustedInterest} for every finite-valued $\tau\in\mathcal{T}^{i}$, and we deduce that $\tau^{i}$ is optimal among all those stopping times. If $\tau$ is a general stopping time and $N\in\mathbb{N}$, Fatou's lemma and that optimality yield
\begin{align*}
E\bigg[\exp\bigg(\int_{0}^{\tau}\! r_{s}\,ds\bigg) \mathbf{1}_{\{\theta^{i}>\tau\}\cup \{\theta^{i}=\infty\}}\bigg]
& \leq \liminf_{N\to\infty} E\bigg[\exp\bigg(\int_{0}^{\tau\wedge N} \!\! r_{s}\,ds\bigg) \mathbf{1}_{\theta^{i}>\tau\wedge N}\bigg] \\
& \leq E\bigg[\exp\bigg(\int_{0}^{\tau^{i}} \! r_{s}\,ds\bigg) \mathbf{1}_{\theta^{i}>\tau^{i}}\bigg],
\end{align*}
so that $\tau^{i}$ is in fact optimal among all stopping times.
The remaining assertions can also be inferred from~\eqref{eq:adjustedInterest}.
\end{proof}
We see from the proof of Lemma~\ref{le:singleAgent} how the increase of $\gamma^{i}-r$ leads to a simple solution of the optimal stopping problem and that will contribute greatly to the tractability of equilibria. We have little else to say in defense of that condition.
\begin{remark}\label{rk:noOptimizer}
As usual in infinite-horizon stopping problems, an integrability assumption is necessary to ensure existence of an optimal stopping time. In particular, if $r>\gamma^{i}>0$ are constant, then $\tau^{i}=\infty$ which is clearly not optimal as then $P(\{\theta^{i}>\tau^{i}\}\cup\{\theta^{i}=\infty\})=0$.
If we consider the same problem with a horizon $T\in(0,\infty)$, no extra assumption is necessary. %
\end{remark}
\begin{remark}\label{rk:strictIncrease}
In Lemma~\ref{le:singleAgent} and the remainder of this paper, we use strict monotonicity of $\gamma^{i}-r$ as a simple sufficient condition for the uniqueness of~$\tau^{i}$. In specific cases one may want to use a sharper condition; for instance, discrete-time problems can be embedded in our results by using piecewise constant processes, but then the notion of strict monotonicity needs to be adapted.
\end{remark}
\subsection{Interaction}
While the $i$th agent chooses to stop at $\tau^{i}\in\mathcal{T}^{i}$, the agents will interact through the ``proportion'' of agents that have already stopped. Indeed, we shall specify $\gamma^{i}$ as a functional depending on a process $\rho$, and then an \emph{equilibrium} will be a collection of stopping times $\tau^{i}\in\mathcal{T}^{i}$ which solve~\eqref{eq:optStopProblem} for $\lambda$-almost all $i\in I$ and such that
\begin{equation}\label{eq:equilibDef}
\rho_{t} = \lambda\{i:\, \tau^{i}\leq t\}.
\end{equation}
When $\lambda$ is atomless, the decision of a single agent does not influence this quantity and hence we indeed have a Nash equilibrium. Clearly, the process~$\rho$ will necessarily be increasing and $[0,1]$-valued. Moreover, we think of $\rho$ as being observed by all agents, so $\rho$ will be $\mathbb{G}^{i}$-adapted for all $i$.
In~\eqref{eq:equilibDef}, we are tacitly assuming that the set on the right-hand side is $\mathcal{I}$-measurable $P$-a.s., which is highly nontrivial for a continuum of i.i.d.\ random variables. The setup that can guarantee this is discussed in the next section. Before that, however, let us illustrate the concepts introduced thus far by a very simple example where the agents do not use any signals except $\rho$.
\begin{example}[Sunspot]\label{ex:sunspot}
(i) Let $\lambda$ be the Lebesgue measure on $I=[0,1]$, let $r>0$ be constant and let $X$ be a right-continuous, increasing process on $(\Omega,\mathcal{F},P)$, progressively measurable for the common, right-continuous filtration $\mathbb{G}^{i}=\mathbb{G}$ (i.e., the same for all agents) and such that $X_{\infty}>1$. Suppose agent $i\in [0,1]$ believes in the intensity
$$
\gamma^{i}_{t} = (r - i + \rho_{t})\vee 0.
$$
Thus, $i$ acts as an index of ``optimism'' or ``risk tolerance''---agents with higher index believe that $\theta^{i}$ will happen later.
We claim that
$$
\rho_{t}=(X_{t}\wedge 1)\vee 0
$$
yields an equilibrium. Indeed, the optimal stopping times are then given by
$$
\tau^{i} = \inf\{t: \, \gamma^{i}_{t}-r \geq 0\} = \inf \{t:\, \rho_{t}=i\}=\inf \{t:\, X_{t}\geq i\}
$$
and that results in
$$
\lambda \{i:\, \tau^{i}\leq t\} = \lambda \{i:\, X_{t}\geq i\} =(X_{t}\wedge 1)\vee 0 = \rho_{t};
$$
note that the second condition of~\eqref{eq:intCond} is satisfied.
For instance, the choice $X_{t}=t$ gives rise to $\rho_{t}=t\wedge1$ and $\tau^{i}=i$, showing that the agents stop at deterministic times which are uniformly distributed over the time interval $[0,1]$. If $X_{0}$ is strictly positive, we see that some of the agents stop instantaneously at $t=0$, whereas if $X_{0}$ is strictly negative, it will take a while before any agents stop.
(ii) A similar equilibrium exists in a finite player game. Let $n\in\mathbb{N}$ and let $\lambda$ be the normalized counting measure on $I=\{1/n,2/n,\ldots,1\}$; this corresponds to $n$ equally weighted agents. In the same setting as in (i), an equilibrium is described by $\tau^{i}=\inf \{t:\, X_{t}\geq i\}$ and
$$
\rho_{t}=\lfloor (X_{t}\wedge 1)\vee 0 \rfloor,
$$
where $\lfloor x \rfloor := \max \{s\in I:\, s\leq x\}$.
\end{example}
\begin{remark}\label{rk:ex}
(i) In the preceding example, the process $X$ is not part of the functional form of $\gamma^{i}$; essentially, \emph{any} process $X$ gives rise to an equilibrium. The interpretation is that if all agents agree that some commonly observed signal $X$ is relevant, it indeed becomes relevant---the name ``sunspot equilibrium'' suggests itself. We shall see in Example~\ref{ex:uniformInitial} that this situation is a degenerate limit of a model where uniqueness is the typical case.
(ii) If all agents are perfectly identical, the problem will degenerate since they will (typically) all stop at the same time. Thus, in the above example, the agents have been made heterogeneous by varying the risk tolerance. This is not necessary when the agents are already heterogeneous due to private signals, as in the later sections.
\end{remark}
\section{Mathematical Setting and Exact Law of Large Numbers}\label{se:exactLLN}
In this section, we introduce the setting to accommodate a continuum of agents and their private signals.
Let $(I,\mathcal{I},\lambda)$ be an atomless (hence, uncountable) probability space and let $(\Omega,\mathcal{F},P)$ be another probability space.
\begin{definition}\label{de:essentialPairwiseIndep}
A family $(f_{i})_{i\in I}$ of random variables on $(\Omega,\mathcal{F},P)$ is \emph{essentially pairwise independent} if for $\lambda$-almost all $i\in I$, $f_{i}$ is independent of $f_{j}$ for $\lambda$-almost all $j\in I$. The family is \emph{essentially pairwise i.i.d.}\ if, in addition, all $f_{i}$ have the same distribution. Analogously, for a $\sigma$-field $\mathcal{C}\subseteq\mathcal{F}$, the family $(f_{i})_{i\in I}$ is \emph{essentially pairwise conditionally independent} given $\mathcal{C}$, if for $\lambda$-almost all $i\in I$, $f_{i}$ is conditionally independent of $f_{j}$ given $\mathcal{C}$ for $\lambda$-almost all $j\in I$.
\end{definition}
In what follows, we need to work on a probability space that is larger than the usual product\footnote{
Here and below, we use the convention that the product $\sigma$-field $\mathcal{I}\otimes \mathcal{F}$ is completed.
}
$(I\times\Omega,\mathcal{I}\otimes \mathcal{F},\lambda\otimes P)$, because the latter does not support relevant families of i.i.d.\ random variables. More precisely, we have the following fact; see, e.g., \cite[Proposition~2.1]{Sun.06}.
\begin{remark}\label{rk:productDegenerate}
If $f: I\times\Omega\to\mathbb{R}$ is an $\mathcal{I}\otimes \mathcal{F}$-measurable function such that $f(i,\cdot)$, $i\in I$ are essentially pairwise i.i.d., then $f$ is constant $\lambda\otimes P$-a.s.
\end{remark}
Following \cite{Sun.06}, we say that a probability space $(I\times\Omega,\Sigma,\mu)$ is an \emph{extension} of the product $(I\times\Omega,\mathcal{I}\otimes \mathcal{F},\lambda\otimes P)$ if $\Sigma$ contains $\mathcal{I}\otimes \mathcal{F}$ and the restriction of $\mu$ to $\mathcal{I}\otimes \mathcal{F}$ coincides with $\lambda\otimes P$. It is a \emph{Fubini extension} if, in addition, any $\mu$-integrable\footnote{
That is, $f$ is measurable for the $\mu$-completion of $\Sigma$ and $\int |f|\,d\mu<\infty$.
}
function $f: I\times\Omega\to\mathbb{R}$ satisfies the assertion of Fubini's theorem\footnote{
Since $\Sigma$ may be strictly larger than $\mathcal{I}\otimes \mathcal{F}$, this is not automatic.
}; that is,
\begin{enumerate}
\item for $\lambda$-almost all $i\in I$, the function $f(i,\cdot)$ is $P$-integrable,
\item for $P$-almost all $\omega\in \Omega$, the function $f(\cdot,\omega)$ $\lambda$-integrable,
\item $i\mapsto \int f(i,\cdot)\,dP$ is $\lambda$-integrable, $\omega\mapsto \int (\cdot,\omega)\,d\lambda$ is $P$-integrable, and
$$
\int f\,d \mu = \iint f(i,\omega)\,P(d\omega)\,\lambda(di) = \iint f(i,\omega)\,\lambda(di)\,P(d\omega).
$$
\end{enumerate}
Let $(I\times\Omega,\Sigma,\mu)$ be a Fubini extension of $(I\times\Omega,\mathcal{I}\otimes \mathcal{F},\lambda\otimes P)$. Then, essentially pairwise independent families satisfy an exact version of the Law of Large Numbers. The simplest version runs as follows---more generally, an exact version of the Glivenko--Cantelli Theorem holds; cf.\ \cite[Corollary~2.9]{Sun.06}.
\begin{proposition}[Exact Law of Large Numbers]\label{pr:LLN}
Let $f: I\times\Omega\to \mathbb{R}$ be $\mu$-integrable. If $f(i,\cdot)$, $i\in I$ are essentially pairwise i.i.d.\ with a distribution having mean $m$, then $\int f(\cdot,\omega)\,d\lambda=m$ for $P$-almost all $\omega\in\Omega$.
\end{proposition}
We shall also need a conditional version as provided by \cite[Corollary~2]{QiaoSunZhang.14}.
\begin{proposition}[Conditional Exact Law of Large Numbers]\label{pr:condLLN}
Let $\mathcal{C}\subseteq\mathcal{F}$ be a countably generated $\sigma$-field and let $f: I\times\Omega\to \mathbb{R}$ be $\mu$-integrable. If $f(i,\cdot)$, $i\in I$ are essentially pairwise conditionally independent given $\mathcal{C}$, then $\int f(\cdot,\omega)\,d\lambda=\int E^{\mu}[f|\mathcal{I}\otimes\mathcal{C}](\cdot,\omega)\,d\lambda$ for $P$-almost all $\omega\in\Omega$.
\end{proposition}
In view of Remark~\ref{rk:productDegenerate}, it is not obvious that the preceding propositions are not vacuous---that is guaranteed by the next two results.
The space $(I\times\Omega,\Sigma,\mu)$ is called \emph{rich} if there exists a $\Sigma$-measurable function $f: I\times\Omega\to\mathbb{R}$ such that $f(i,\cdot)$, $i\in I$ are essentially pairwise i.i.d.\ with a uniform distribution on $[0,1]$.
Like an atomless probability space supports random variables with any given distribution, a rich Fubini extension supports essentially pairwise i.i.d.\ families with any given distribution; cf.\ \cite[Corollary~5.4]{Sun.06}.
\begin{lemma}\label{le:richSupportsiid}
Let $(I\times\Omega,\Sigma,\mu)$ be a rich Fubini extension of $(I\times\Omega,\mathcal{I}\otimes \mathcal{F},\lambda\otimes P)$, let $S$ be a Polish space and let $\nu$ be a Borel probability measure on $S$. There exists a $\Sigma$-measurable function $f:I\times\Omega\to S$ such that $f(i,\cdot)$, $i\in I$ are essentially pairwise independent and $f(i,\cdot)$ has distribution $\nu$ for all $i\in I$.
\end{lemma}
\begin{lemma}\label{le:richExtensionsExist}
There exist atomless probability spaces $(I,\mathcal{I},\lambda)$ and $(\Omega,\mathcal{F},P)$ such that $(I\times\Omega,\mathcal{I}\otimes \mathcal{F},\lambda\otimes P)$ admits a rich Fubini extension.
\end{lemma}
This is part of the assertion of \cite[Proposition~5.6]{Sun.06} which also shows that one can take $I=[0,1]$ and $\Omega=\mathbb{R}^{[0,1]}$. The main result of \cite{SunZhang.09} shows that, in addition, one can take $\lambda$ to be an extension of the Lebesgue measure (but not the Lebesgue measure itself). A different construction, avoiding nonstandard analysis, is presented in \cite{Podczeck.10}.
\section{A Toy Model}\label{se:toyModel}
In this section, we discuss a simple setting where the agents' signals are i.i.d.; that is, pure idiosyncratic noise. While not suitable for most applications, this will allow us to explain the effect of the Exact Law of Large Numbers in our model and to discuss some finer questions of uniqueness and nondegeneracy without too many distractions.
Consider the setup introduced in Section~\ref{se:basicSetup} with atomless probability spaces $(I,\mathcal{I},\lambda)$ and $(\Omega,\mathcal{F},P)$, and let $(I\times\Omega,\Sigma,\mu)$ be a Fubini extension of their product. For each $i\in I$, let $Y^{i}\geq0$ be a right-continuous, increasing, $\mathbb{G}^{i}$-progressively measurable process. We assume that for each $t\geq0$, $(i,\omega)\mapsto Y^{i}_{t}(\omega)$ is $\Sigma$-measurable and that $Y^{i}_{t}$, $i\in I$ are essentially pairwise i.i.d. Moreover, we assume that the distribution of $Y^{i}_{t}$ has no atoms; that is, its c.d.f.\ $y\mapsto F_{t}(y):=P\{Y^{i}_{t}\leq y\}$ is continuous.
\begin{proposition}\label{pr:privateInfo}
Let $r\in\mathbb{R}$ and $c\in\mathbb{R}_{+}$. The equation
\begin{equation}\label{eq:privateInfo}
1-u = F_{t}(r-cu),\quad u\in [0,1]
\end{equation}
has a maximal solution $\rho(t)\in[0,1]$ for every $t\geq0$, and $t\mapsto \rho(t)$ is right-continuous. Define also
$$
\gamma^{i}_{t} = Y^{i}_{t} + c\rho(t), \quad \tau^{i} = \inf \{t:\, Y^{i}_{t} + c\rho(t) =r\},
$$
and assume that \eqref{eq:intCond} is satisfied for all $i$.
(i) Then, $\rho$ and $(\tau^{i})_{i\in I}$ define an equilibrium: $\tau^{i}\in\mathcal{T}^{i}$ is an optimal stopping time for agent $i$, the mapping $(i,\omega)\mapsto \tau^{i}(\omega)$ is $\Sigma$-measurable, and
$$
\lambda\{i:\,\tau^{i}\leq t\}=\rho(t)\quad P\mbox{-a.s.} \quad \mbox{for all} \quad t\geq0.
$$
(ii) Conversely, let $\bar\rho$ be a right-continuous function corresponding to an equilibrium. If $\gamma^{i}$ is strictly increasing for all $i$, then $\bar\rho(t)$ is a solution of~\eqref{eq:privateInfo} for every $t\geq0$.
\end{proposition}
\begin{proof}[Sketch of Proof.]
The proposition is a special case of Theorem~\ref{th:general} that will be proved later on, so we shall only explain the most important steps.
(a) We first argue that $\rho$ is well-defined, increasing and right-continuous.
Let us consider, for a right-continuous and increasing function $F: \mathbb{R}\to [0,1]$, the zeros of
$$
G(u):=F(r-cu)-1+u, \quad u\in[0,1].
$$
We have $G(0)=F(r)-1\leq0$ and $G(1)=F(r-c)\geq0$. Moreover, $G$ is left-continuous and its jumps satisfy $\Delta G\leq0$. Thus, $G$ must have at least one zero in $[0,1]$. If $u_{n}\uparrow u$ is a maximizing sequence of zeros in $[0,1]$, then $G(u)=0$ by left-continuity and $u$ is the maximal zero.
Next, write $G_{t}(u):=F_{t}(r-cu)-1+u$ and let $\rho(t)$ be the maximal zero for each $t\geq0$. The increase and the right-continuity of $t\mapsto\rho(t)$ can be inferred from the increase of $Y$ and the right-continuity of $Y$ and the continuity of $y\mapsto F_{t}(y)$, respectively---we defer the details.
(b) Next, we verify that $\rho$ and $(\tau^{i})_{i\in I}$ determine an equilibrium. It follows from (a) that $\gamma^{i}=Y^{i} + c\rho$ is increasing and right-continuous; hence, Lemma~\ref{le:singleAgent} yields that $\tau^{i}\in\mathcal{T}^{i}$ is an optimal stopping time for all $i\in I$ and that $\{(i,\omega):\, \tau^{i}(\omega)\leq t\} = \{(i,\omega):\, Y^{i}_{t}(\omega) + c\rho(t)\geq r\}\in \Sigma$. Using the Exact Law of Large Numbers of Proposition~\ref{pr:LLN}, the continuity of $F_{t}$ and the definition of $\rho(t)$, we have $P$-a.s.\ that
\begin{align*}
\bar\rho(t):=\lambda\{i:\, \tau^{i}\leq t\} &=\lambda\{i:\, Y^{i}_{t} + c\rho(t)\geq r\} \\
&= \int \! P\{ Y^{i}_{t} + c\rho(t)\geq r \}\,\lambda(di)= 1-F_{t}(r-c\rho(t))=\rho(t)
\end{align*}
for all $t\geq0$.
(c) Let $\bar\rho: \mathbb{R}_{+}\to\mathbb{R}$ be a right-continuous function corresponding to an equilibrium; that is, $\bar\rho(t)=\lambda\{i:\, \tau^{i}\leq t\}$ for some optimal $\tau^{i}\in\mathcal{T}^{i},$ $i\in I$. Then $\bar\rho$ is clearly increasing and $[0,1]$-valued. Due to the strict increase of~$\gamma^{i}$, we know from Lemma~\ref{le:singleAgent} that $\tau^{i}=\inf\{t: \gamma^{i}_{t}\geq r\}$. Thus,
\begin{align*}
\bar\rho(t)=\lambda\{i:\, \tau^{i}\leq t\} &=\lambda\{i:\, Y^{i}_{t} + c\bar\rho(t)\geq r\} \\
&= \int P\{ Y^{i}_{t} + c\bar\rho(t)\geq r \}\,\lambda(di)= 1-F_{t}(r-c\bar\rho(t));
\end{align*}
that is, $\bar\rho(t)$ is a solution of~\eqref{eq:privateInfo} for all $t\geq0$.
\end{proof}
We begin our discussion with some observations about uniqueness.
\begin{remark}\label{rk:minimalZero}
(i) Equation~\eqref{eq:privateInfo} may have more than one solution; cf.\ Example~\ref{ex:threeSolutions}. If $t\mapsto\rho(t)$ is any right-continuous, increasing solution of \eqref{eq:privateInfo}, not necessarily maximal, then $\rho$ induces an equilibrium, by the same arguments as in the above proof of Proposition~\ref{pr:privateInfo}.
%
(ii) Equation~\eqref{eq:privateInfo} also has a minimal solution, and it is automatically increasing in $t$. However, it is not necessarily right-continuous; see also Example~\ref{ex:uniformInitial}\,(iii) below. Instead, it is left-continuous provided that $Y$ is. If the solution is unique and $Y$ is continuous, the solution is both minimal and maximal, and therefore continuous.
\end{remark}
Next, we analyze a special case of Proposition~\ref{pr:privateInfo} that is explicitly solvable and sheds some light on the impact of the constant $c\geq0$ that parametrizes the strength of interaction. One intuition is that if the interaction between the agents is too strong, some agents' stopping will lead to a domino effect where all others end up stopping immediately after.
\begin{example}\label{ex:uniformInitial}
Let $r\geq1$ and let $U^{i}$, $i\in I$ be essentially pairwise i.i.d.\ with a uniform distribution on $[r-1,r]$. Moreover, let $a: \mathbb{R}_{+}\to\mathbb{R}_{+}$ be a strictly increasing, right-continuous function with $a(0)=0$ and $a(\infty)>1$---the latter will ensure that~\eqref{eq:intCond} holds. We then consider the strictly increasing process
$$
Y^{i}_{t}=U^{i}+a(t)
$$
and note that $F_{t}(y)=F(1+y-a(t)-r)$, where $F$ is the c.d.f.\ of the uniform distribution on $[0,1]$. Thus, Equation~\eqref{eq:privateInfo} becomes
\begin{equation}\label{eq:uniformInitial}
1-u = F(1-cu-a(t)),\quad u\in [0,1].
\end{equation}
\begin{enumerate}
\item \emph{No interaction,} $c=0$. Clearly, the unique solution is $\rho(t)=a(t)\wedge 1$, and this is the unique equilibrium by the last part of Proposition~\ref{pr:privateInfo}.
\item \emph{Moderate interaction,} $c\in(0,1)$. Then, \eqref{eq:uniformInitial} is easily seen to have a unique solution $\rho(t)\in [0,1]$; namely,
$$
\rho(t)=[(1-c)^{-1}a(t)]\wedge 1,
$$
and this is the unique equilibrium. In particular, the population of stopped agents evolves in a nondegenerate, continuous fashion for $c\in(0,1)$. The larger the interaction coefficient $c$, the more agents stop earlier.
\item \emph{Critical interaction,} $c=1$. Using that $a(t)>0$ for $t>0$, we can check that $\rho(t)= 1$ is the unique solution of~\eqref{eq:uniformInitial} for $t>0$. Thus, $\rho\equiv1$ is the unique right-continuous solution; that is, all agents stop at $t=0$. This is also the unique right-continuous equilibrium.
It is worth noting that any $u\in[0,1]$ is a solution of \eqref{eq:uniformInitial} for $t=0$; recall that $a(0)=0$. Intuitively speaking, solutions $\rho(t)=u\mathbf{1}_{\{0\}}+\mathbf{1}_{(0,\infty)}$ that are not right-continuous correspond to equilibria where a fraction $u$ of the agents stop at time zero whereas the rest stop ``immediately after'' zero. This may illustrate why we have imposed right-continuity in our results.
\item \emph{Supercritical interaction,} $c>1$. We see directly that $\rho(t)= 1$ is the unique solution of~\eqref{eq:uniformInitial} for all $t\geq0$.
\end{enumerate}
\end{example}
\subsection{On the Multiplicity of Equilibria}
The following is a fairly well-behaved example of non-uniqueness.
\begin{example}\label{ex:threeSolutions}
We consider again the setting of Example~\ref{ex:uniformInitial}, with $r=1$ and $a(t)=t$, say, but we now replace the uniform distribution of $U^{i}$ with a measure that assigns mass $\varepsilon\in(0,1/4)$ uniformly to $[0,\varepsilon]$ and to $[1-\varepsilon,1]$, and the remaining mass $1-2\varepsilon$ uniformly to $[1/2-\varepsilon,1/2+\varepsilon]$. For small enough $\varepsilon>0$ and $c\in(0,1)$, we see that Equation~\eqref{eq:privateInfo} has three interior solutions for $t$ within a certain interval, whereas all solutions are at the origin for $t=0$. We can select any of these solutions to form an increasing right-continuous process $\rho$ that corresponds to a legitimate equilibrium.
\end{example}
To understand this bifurcation, let us first look at an even simpler situation where $\gamma^{i}_{t}=r-c+c\rho(t)$ for all $i$. At time $t=0$, two obvious equilibria are: No agent stops, then $\rho(t)=0$ and $\gamma^{i}_{t}=r-c<r$, so it is indeed optimal not to stop. Or, all agents stop immediately, then $\rho(t)=1$ and $\gamma^{i}_{t}=r$, so it is indeed optimal to stop. (This coordination problem is very similar to the phenomenon discussed e.g.\ in \cite{DiamondDybvig.83, MorrisShin.04}.) When $\gamma^{i}_{t}$ is random, a similar choice can arise at an intermediate time for a subset of the population corresponding to an atom in the distribution of $\gamma^{i}$. More generally, the bifurcation can also happen in a continuous fashion when the random variable is sufficiently concentrated (relative to the size of $c$) around some point rather than having atoms, and this is what was witnessed in Example~\ref{ex:threeSolutions}.
The following observation is a different view on the same interplay.
\begin{remark}\label{rk:implicitFunctionThm}
Let $(t,y)\mapsto F_{t}(y)$ be $C^{1}$ and write $f_{t}=\partial_{y}F_{t}$ for the probability density at time $t$. Suppose that $\partial_{u} F_{t}(r-cu) \neq -1$; that is,
$$
cf_{t}(r-cu)\neq 1
$$
for $u$ in a neighborhood of a solution $\rho(t)$ of~\eqref{eq:privateInfo}. Then, the Implicit Function Theorem shows that $\rho$ is locally unique and $C^{1}$. For $c>0$, this is true, in particular, if $0\leq f_{t} < c^{-1}$ on $[r-c,r]$. Or, put differently: if $F_{t}$ is not too concentrated or if $c$ is small enough, local uniqueness holds.
\end{remark}
The following provides a broader perspective on Example~\ref{ex:sunspot} and shows that uniqueness may fail even more dramatically in certain regimes.
\begin{example}\label{ex:retrieveSunspot}
We consider again the setting of Example~\ref{ex:uniformInitial}, except that we now take
$$
a(t)=\begin{cases}
0, & t<T\\
2, & t\geq T,
\end{cases}
$$
where $T\in(0,\infty)$ acts as a time horizon. Indeed, this definition implies $\gamma^{i}_{t}\geq r+1$ for all $t\geq T$ and $i\in I$, so that all agents will stop at $T$, if not earlier.
Thus, we are interested in the situation on $[0,T)$, where \eqref{eq:uniformInitial} becomes
$$
1-u = F(1-cu).
$$
(i) For $c\in[0,1)$, the unique solution is $u=0$, and thus
$$
\rho(t) = \mathbf{1}_{[T,\infty)}
$$
is the corresponding equilibrium: at $t=0$, only a nullset of agents stop, and that does not change until $T$. One can check that this equilibrium is unique, even though $\gamma^{i}$ is not strictly increasing.
%
(ii) At the critical value $c=1$, uniqueness is lost and the situation is completely different. Indeed, the equation becomes the tautological $1-u=1-u$. Thus, \emph{any} right-continuous, increasing function $X(t)$ with values in $[0,1]$ determines an equilibrium via
$$
\rho(t) = X(t)\mathbf{1}_{[0,T)} + \mathbf{1}_{[T,\infty)}.
$$
This is the situation we have encountered in Example~\ref{ex:sunspot}: in terms of the equilibrium distribution, there is equivalence between assigning risk aversion $r-i$ to agent $i$ and sampling uniformly from $[r-1,r]$ for every agent as in the present example. The latter basically corresponds to randomly permuting the labels of the agents in Example~\ref{ex:sunspot}.
\end{example}
\section{The General Model}\label{se:generalModel}
In this section, we generalize the model from the previous section by specifying the intensities $\gamma^{i}$ as a possibly nonlinear function of i.i.d.\ signals $Y^{i}$ and a common signal $X$. As a result, the intensities are conditionally independent rather than independent, and the equilibrium becomes a function of $X$.
As above, we consider the setup introduced in Section~\ref{se:basicSetup} with atomless probability spaces $(I,\mathcal{I},\lambda)$ and $(\Omega,\mathcal{F},P)$, and let $(I\times\Omega,\Sigma,\mu)$ be a Fubini extension of their product. For each $i\in I$, let $Y^{i}\geq0$ be a right-continuous, increasing, $\mathbb{G}^{i}$-progressively measurable process. We assume that for each $t\geq0$, $(i,\omega)\mapsto Y^{i}_{t}(\omega)$ is $\Sigma$-measurable and that $Y^{i}_{t}$, $i\in I$ are essentially pairwise i.i.d. Moreover, we assume that the distribution of $Y^{i}_{t}$ has no atoms; that is, its c.d.f.\ $y\mapsto F_{t}(y):=P\{Y^{i}_{t}\leq y\}$ is continuous. In addition, let $X$ be a $d$-dimensional, right-continuous, (componentwise) increasing process which is $\mathbb{G}^{i}$-progressively measurable for all $i$ and such that $X_{t}$ and $Y^{i}_{t}$ are independent for all $t\geq0$. Thus, $X$ is interpreted as public information whereas $Y^{i}$ is a private signal%
\footnote{
Additional idiosyncratic signals could also be included. In particular, a signal at time zero can be used to assign different functional forms of $\gamma^{i}$ to the agents, similarly as at the end of Example~\ref{ex:uniformInitial}.
}
available only to agent $i$. (A slightly different setup and interpretation are discussed in Section~\ref{se:additiveModel}.)
Let $r: \mathbb{R}_{+}\times\mathbb{R}^{d}\to\mathbb{R}$ be a right-continuous, decreasing function. The interest rate process will be assumed to be of the form\footnote{
Since $X$ can be multivariate, this entails no loss of generality relative to introducing yet another stochastic process.
}
$$
r_{t}=r(t,X_{t}).
$$
Finally, the intensity of agent $i$ will be of the form $\gamma^{i}_{t} = g(t,X_{t},Y^{i}_{t},\rho_{t})$, where
$$
g: \mathbb{R}_{+}\times \mathbb{R}^{d}\times \mathbb{R}\times [0,1]\to\mathbb{R}
$$
is a continuous function with $g(t,X_{t},Y^{i}_{t},0)\geq0$, increasing in all its arguments and such that for all $(t,x,u)$, $y\mapsto g(t,x,y,u)$ admits an inverse $y'\mapsto g^{-1}(t,x,y',u)$ on its range which we assume to be $\mathbb{R}$ for simplicity. We suppose that $g^{-1}$ is again continuous.
\begin{theorem}\label{th:general}
The equation
\begin{equation}\label{eq:general}
1-u = F_{t}(g^{-1}(t,x,r,u)),\quad u\in [0,1]
\end{equation}
has a maximal solution $\rho(t,x,r)\in[0,1]$ for every $(t,x,r)\in\mathbb{R}_{+}\times\mathbb{R}^{d}\times\mathbb{R}$, and $\rho_{t}:=\rho(t,X_{t},r(t,X_{t}))$ is a right-continuous, increasing process. Define also
$$
\gamma^{i}_{t} = g(t,X_{t},Y^{i}_{t},\rho_{t}), \quad \tau^{i} = \inf \{t:\, \gamma^{i}_{t} =r_{t}\}
$$
and assume that~\eqref{eq:intCond} is satisfied for all $i$.
(i) Then, $\rho$ and $(\tau^{i})_{i\in I}$ define an equilibrium: $\tau^{i}\in\mathcal{T}^{i}$ is an optimal stopping time for agent $i$, the mapping $(i,\omega)\mapsto \tau^{i}(\omega)$ is $\Sigma$-measurable, and
$$
\lambda\{i:\,\tau^{i}\leq t\}=\rho_{t}\quad P\mbox{-a.s.} \quad \mbox{for all} \quad t\geq0.
$$
More generally, this holds for any measurable solution $\rho(t,x,r)$ of~\eqref{eq:general} such that $\rho(t,X_{t},r(t,X_{t}))$ is right-continuous and increasing.
(ii) Conversely, let $t\mapsto \bar\rho_{t}$ be a right-continuous process corresponding to an equilibrium and suppose that $\bar\rho_{t}=\bar\rho(t,X_{t},r(t,X_{t}))$ for some measurable function $\bar\rho$. If $\gamma^{i}$ is strictly increasing for all $i$, then for every $t\geq0$, $\bar\rho(t,x,r(t,x))$ solves~\eqref{eq:general} for $(P\circ X_{t}^{-1})$-almost all $x\in\mathbb{R}^{d}$.
%
%
%
%
\end{theorem}
\begin{proof}
(a) We claim that $\rho(t,x,r)$ is well-defined, increasing in $(t,x)$ and decreasing in $r$, and (jointly) right-continuous in $(t,x)$ and left-continuous in $r$. Indeed, fix $(t,x,r)$ and consider the function
$$
G_{t,x,r}(u) :=F_{t}(g^{-1}(t,x,r,u))-1+u, \quad u\in[0,1].
$$
Since $F_{t}$ takes values in $[0,1]$, we have $G_{t,x,r}(0)\leq0$ and $G_{t,x,r}(1)\geq0$. As $u\mapsto G_{t,x,r}(u)$ is continuous, it follows that there is least one zero in $[0,1]$, and since the set of all zeros is compact, it has a maximum.
We write $\rho(t,x,r)$ for the maximal zero of $G_{t,x,r}$. As $Y$ is increasing, the function $t\mapsto F_{t}(y)$ is decreasing and then so is $t\mapsto G_{t,x,r}(u)$; note that $g^{-1}$ is decreasing in $(t,x,u)$ and increasing in $r$. Hence, if $s\leq t$, the fact that $G_{t,x,r}>0$ on $(\rho(t,x,r),1]$ implies that $G_{s,x,r}>0$ on $(\rho(t,x,r),1]$ and hence that $\rho(s,x,r)\leq \rho(t,x,r)$. The monotonicity in $x$ and $r$ follows analogously.
Let $t_{n}\downarrow t$ and $x_{n}\downarrow x$ and $r_{n}\uparrow r$. Set $\rho_{n}=\rho(t_{n},x_{n},r_{n})$ and $\rho_{*}=\rho(t,x,r)$. By the above, $\rho_{n}$ is decreasing and $\rho_{n}\geq \rho_{*}$. Thus, we only need to verify that $\rho_{\infty}:=\lim \rho_{n}\leq \rho_{*}$. In view of the definition of $\rho_{*}$ as a maximal zero, it suffices to show that $\rho_{\infty}$ is a zero of $G_{t,x,r}$, and as $G_{t_{n},x_{n},r_{n}}(\rho_{n})=0$, that will follow if
\begin{equation}\label{eq:eq:claimRC}
(t,x,-r,u)\mapsto G_{t,x,r}(u)\quad \mbox{is jointly right-continuous.}
\end{equation}
Indeed, $y\mapsto F_{t}(y)$ is continuous, and together with the right-continuity of $Y$, it follows that $t\mapsto F_{t}(y)$ is right-continuous. %
Using also the continuity of $g^{-1}$, we see that~\eqref{eq:eq:claimRC} holds as desired. This completes the proof of the claim on $\rho$.
(b) Next, we verify the equilibrium conditions. As a result of (a), the processes $t\mapsto\rho_{t}=\rho(t,X_{t},r(t,X_{t}))$ and $\gamma^{i}$ are increasing and right-continuous, and Lemma~\ref{le:singleAgent} yields that $\tau^{i}\in\mathcal{T}^{i}$ is an optimal stopping time. Note that $(i,\omega)\mapsto\gamma^{i}(\omega)$ is $\Sigma$-measurable, and so is $(i,\omega)\mapsto r_{t}(\omega)$. Thus, $\{\tau^{i}\leq t\}=\{\gamma^{i}\geq r_{t}\}\in\Sigma$ for all $t\geq0$. Using the Conditional Exact Law of Large Numbers of Proposition~\ref{pr:condLLN}, the continuity of $y\mapsto F_{t}(y)$ and the definition of~$\rho_{t}$, we have $P$-a.s.\ that
\begin{align*}
\bar\rho_{t}:=\lambda\{i:\, \tau^{i}\leq t\}
&= \lambda\{i:\, g(t,X_{t},Y^{i}_{t},\rho(t,X_{t},r(t,X_{t})))\geq r(t,X_{t})\}\\
&= \int P\{ g(t,X_{t},Y^{i}_{t},\rho(t,X_{t},r(t,X_{t})))\geq r(t,X_{t}) | X_{t}\}\,\lambda(di)\\
&= 1-F_{t}(g^{-1}(t,X_{t},r(t,X_{t}),\rho(t,X_{t},r(t,X_{t}))))\\
&=\rho(t,X_{t},r(t,X_{t}))=\rho_{t}.
\end{align*}
%
(c) Let $\bar\rho$ be a right-continuous process corresponding to an equilibrium; that is, $\bar\rho_{t}=\lambda\{i:\, \tau^{i}\leq t\}$ for some optimal $\tau^{i}\in\mathcal{T}^{i}$. Then $\bar\rho$ is clearly increasing and $[0,1]$-valued. Due to the strict increase of $\gamma^{i}$, we know from Lemma~\ref{le:singleAgent} that $\tau^{i}=\inf\{t: \gamma^{i}_{t}\geq r_{t}\}$, which also ensures that $\{\tau^{i}\leq t\}\in\Sigma$. Since we have assumed that
$\bar\rho_{t}=\bar\rho(t,X_{t},r(t,X_{t}))$, we obtain as in~(b) that
$$
\bar\rho_{t}=\lambda\{i:\, \tau^{i}\leq t\}
= 1-F_{t}(g^{-1}(t,X_{t},r(t,X_{t}),\bar\rho(t,X_{t},r(t,X_{t}))))\quad P\mbox{-a.s.}
$$
for all $t\geq0$.
\end{proof}
\begin{remark}\label{rk:MarkovUniqueness}
The result in Theorem~\ref{th:general}\,(ii) assumes a priori that the equilibrium $\bar\rho_{t}$ is Markovian; that is, a deterministic function of $(t,X_{t})$.
(i) First, let us observe that this is not automatically the case: randomized equilibria may exist. Consider the setting of Example~\ref{ex:threeSolutions} where $X$ is deterministic and Equation~\eqref{eq:general} has several (deterministic, increasing, right-continuous) solutions; in particular, a maximal solution $\rho(t)$ and a minimal solution $\rho'(t)$. Suppose that there is a Poisson process $N$ which is $\mathbb{G}^{i}$-adapted and independent of $Y^{i}$ for all $i$, and let $\sigma$ be its first jump time. Then,
$$
\bar\rho_{t} = \mathbf{1}_{[0,\sigma)}(t)\rho'(t) + \mathbf{1}_{[\sigma,\infty)}(t)\rho(t)
$$
defines another right-continuous, increasing solution of~\eqref{eq:general} which determines an equilibrium. However, $\bar\rho_{t}$ is not of the mentioned Markovian form. Instead, the agents can agree to change their behavior according to the independent randomization $\sigma$. We also refer to \cite{CarmonaDelaRueLacker.13, Lacker.14} for further insights on randomized (or ``weak'') equilibria in the context of standard mean field games.
(ii) Second, let us show that the phenomenon mentioned in (i) cannot occur if uniqueness holds in Equation~\eqref{eq:general}. In the setting of Theorem~\ref{th:general}\,(ii), even if we do not suppose a priori that $\bar\rho$ is a function of $(t,X_{t})$, we have
\begin{align*}
\bar\rho_{t}=\lambda\{i:\, \tau^{i}\leq t\}
&= \lambda\{i:\, g(t,X_{t},Y^{i}_{t},\bar\rho_{t})\geq r(t,X_{t})\}\\
&= \int P\{ g(t,X_{t},Y^{i}_{t},\bar\rho_{t})\geq r(t,X_{t}) | X_{t}\}\,\lambda(di)\\
&= 1-F_{t}(g^{-1}(t,X_{t},r(t,X_{t}),\bar\rho_{t})).
\end{align*}
If $\rho(t,x,r)$ is the maximal solution of \eqref{eq:general} as constructed in the theorem and uniqueness holds for~\eqref{eq:general}, it follows that
$$
\bar\rho_{t} = \rho(t,X_{t},r(t,X_{t}))\quad P\mbox{-a.s.}
$$
and in that sense, $\bar\rho$ is necessarily of the Markovian form.
\end{remark}
\begin{remark}\label{rk:conditioningOnPath}
In the literature on mean field games driven by stochastic differential equations, the private states at time $t$ are usually independent conditionally on the whole path $(X_{s})_{s\leq t}$ of the common noise before time~$t$. In the present setting, we have assumed that the intensities $\gamma^{i}$ depend on~$X$ in a Markovian way, and hence it is sufficient to condition on the current value $X_{t}$. One could envision a similar result where $\gamma^{i}$ depends on $X$ in a path-dependent way, and then one would condition on the whole past of~$X$.
\end{remark}
\subsection{Additive Model and Noisy Observations}\label{se:additiveModel}
In this section, we enhance the toy model from Proposition~\ref{pr:privateInfo} by incorporating a public signal and obtain a tractable specification of the general model from Theorem~\ref{th:general}.
Consider the setup introduced in Section~\ref{se:basicSetup} with atomless probability spaces $(I,\mathcal{I},\lambda)$ and $(\Omega,\mathcal{F},P)$, and let $(I\times\Omega,\Sigma,\mu)$ be a Fubini extension of their product. For each $i\in I$, let $Y^{i}\geq0$ be a right-continuous, increasing, measurable process. We assume that for each $t\geq0$, $(i,\omega)\mapsto Y^{i}_{t}(\omega)$ is $\Sigma$-measurable and that $Y^{i}_{t}$, $i\in I$ are essentially pairwise independent. Moreover, we assume that the distribution of $Y^{i}_{t}(\cdot)$ has no atoms; that is, its c.d.f.\ $F_{t}$ is continuous. Furthermore, let $X\geq0$ be a right-continuous, increasing, measurable process such that $X_{t}$ and $Y^{i}_{t}$ are independent for all $t\geq0$.
We take $r\in\mathbb{R}$ to be constant (for simplicity) and
$$
\gamma^{i}_{t} = X_{t} + Y^{i}_{t} + c\rho_{t},
$$
where $c\geq0$ is a constant governing the strength of interaction; see also Example~\ref{ex:uniformInitial}. For the information structure, we may consider two cases. Either we see $\mathbb{G}^{i}$ as given and assume that
\begin{itemize}
\item $X$ and $Y^{i}$ are $\mathbb{G}^{i}$-progressively measurable for all $i\in I$,
\end{itemize}
which was the point of view taken above. Or, we model that the agents observe only $X+Y^{i}$ and $\rho$, and thus we convene that
\begin{itemize}
\item $\mathbb{G}^{i}$ is the right-continuous filtration generated by $X+Y^{i}$ and $\rho$, for all $i\in I$.
\end{itemize}
This allows for the interpretation of $X$ as a ``true signal,'' whereas agent $i$ can only observe the noisy signal $X+Y^{i}$ with i.i.d.\ noise $Y^{i}$. Although the agents have more information in the first setting, both yield the same equilibria---the form of $\tau^{i}$ stated below shows that the agents only use the observation of $\gamma^{i}$.
Indeed, Theorem~\ref{th:general} yields the following.
\begin{corollary}\label{co:additiveModel}
The equation
$$
1-u = F_{t}(r-x-cu),\quad u\in [0,1]
$$
has a maximal solution $\rho(t,x)\in[0,1]$ for every $(t,x)\in\mathbb{R}_{+}\times\mathbb{R}$, and $\rho_{t}:=\rho(t,X_{t})$ is a right-continuous process. Define also
$$
\gamma^{i}_{t} = X_{t} + Y^{i}_{t} + c\rho_{t}, \quad \tau^{i} = \inf \{t:\, X_{t} + Y^{i}_{t} + c\rho_{t} =r\}
$$
and assume that~\eqref{eq:intCond} is satisfied for all $i$.
(i) Then, $\rho$ and $(\tau^{i})_{i\in I}$ define an equilibrium: $\tau^{i}\in\mathcal{T}^{i}$ is an optimal stopping time for agent $i$, the mapping $(i,\omega)\mapsto \tau^{i}(\omega)$ is $\Sigma$-measurable, and
$$
\lambda\{i:\,\tau^{i}\leq t\}=\rho_{t}\quad P\mbox{-a.s.} \quad \mbox{for all} \quad t\geq0.
$$
(ii) Conversely, let $t\mapsto \bar\rho_{t}$ be a right-continuous process corresponding to an equilibrium and suppose that $\bar\rho_{t}=\bar\rho(t,X_{t})$ for some measurable function $\bar\rho$. If $\gamma^{i}$ is strictly increasing for all $i$, then for every $t\geq0$, $\bar\rho(t,x)$ solves~\eqref{eq:general} for $(P\circ X_{t}^{-1})$-almost all $x\in\mathbb{R}$.
\end{corollary}
A solvable example can be constructed along the lines of Example~\ref{ex:uniformInitial}.
\begin{example}\label{ex:uniformInitialCommonNoise}
Let $r\geq1$ and let $U^{i}$, $i\in I$ be essentially pairwise i.i.d.\ with a uniform distribution on $[r-1,r]$ and such that $U^{i}$ and $X_{t}$ are independent for all $t\geq0$. Moreover, suppose that $X$ is strictly increasing with $X_{0}=0$ and $X_{\infty}>1$. For $c\in(0,1)$, consider the intensity process
$$
\gamma^{i}_{t} = X_{t}+ U^{i} + c\rho_{t}.
$$
Then, the equation has a unique solution $\rho(t,x)$, and
$$
\rho(t,X_{t})=[(1-c)^{-1}X_{t}]\wedge 1
$$
corresponds to the unique (Markovian) equilibrium. In particular, this equilibrium evolves in a nondegenerate way as long as $X$ does.
\end{example}
\newcommand{\dummy}[1]{}
|
\section{Introduction}
\label{intro}
Vitamin B$_{12}$ is a very important organometallic molecule for biological systems \cite{Prada,Harris,Stich}.
In this paper, we study the electronic structure of cyanocobalamin (CNCbl) which is a form of vitamin B$_{12}$. In Fig. 1,
we illustrate the molecular structure of CNCbl.
The cobalt atom neighbours five nitrogen atoms, of which four are located in the corrin ring.
The CN ligand is also attached to Co, making the rare cobalt-carbon bonding. The charge neutral CNCbl molecule has 718 electrons.
For this molecule,
the HOMO (highest occupied molecular orbital) and the LUMO (lowest unoccupied molecular orbital) levels
are separated by an energy gap of $\approx 2.2$ eV \cite{Firth}. The photoabsorption spectrum of CNCbl exhibits
distinct peaks at $\approx 3.5$ eV and $4.5$ eV, of which origin remains elusive \cite{Firth}. In additon,
it is known that CNCbl has a weak diamagnetic response \cite{Grun,Diehl}.
\begin{figure}[!htp]
\includegraphics[width=8cm]{fig1}
\caption{(Color online)
Schematic plot of the molecular structure of cyanocobalamin
(C$_{63}$H$_{88}$CoN$_{14}$O$_{14}$P), which contains 181 atoms.
}
\label{fig1}
\end{figure}
Despite many years of research, there remains questions about the electronic structure and
the functioning of vitamin B$_{12}$ as well as the role of the transition-metal cobalt atom.
In this paper, we study the electronic structure of CNCbl from the perspective of many body physics. In
particular, we use the combined density functional theory (DFT) and quantum Monte Carlo (QMC) approach to study
the electronic structure and magnetic correlations of this molecule within the framework of the Haldane-Anderson
model \cite{Haldane}. This model was initially introduced to describe the electronic state of Au in a semiconductor
Ge host. Here, we use the Haldane-Anderson model because CNCbl exhibits a semiconductor energy gap and contains the
transition-metal cobalt atom.
In the combined DFT+QMC approach, we first determine the parameters of the Anderson Hamiltonian \cite{Anderson}
by the DFT calculations carried out with the Gaussian program \cite{Gaussian}. Then, we study this effective Anderson Hamiltonian
by performing QMC simulations with the Hirsch-Fye \cite{Hirsch} algorithm. In our calculations, we use an orbital independent intra-orbital Coulomb interaction $U$
at the Co($3d_{\nu}$) orbitals and we neglect the inter-orbital Coulomb interactions along with the Hund's coupling.
We take the value of $U$ to be 4 eV. In addition, the Co($3d_{\nu}$) energy
levels are shifted in order to prevent the double counting of the local Coulomb interactions
by both DFT and QMC.
The outline of this paper is as follows. In Section 2, we obtain the one-electron parameters of the effective
Anderson Hamiltonian. We present the QMC data on the electronic structure
and the magnetic correlation functions in Section 3. Here, we find that impurity bound states develop above the LUMO level for CNCbl
instead of being inside the semiconducting energy gap. This is due to the discrete single-particle spectrum
of the molecule.
We observe that magnetic moments develop at the Co($3d_{\nu}$) orbitals and at the host states
which have the strongest hybridization. We also observe that there are antiferromagnetic
correlations between these host magnetic moments and the Co($3d_{\nu}$) moments.
These antiferromagnetic correlations disappear when the impurity bound states are filled.
We note that in Ref. [\onlinecite{Kandemir2}], we combined the Hartree-Fock (HF) approximation with the QMC technique
to study CNCbl. In the HF+QMC calculations, we used only the smaller imidazole part of
the CNCbl instead of the
whole molecule.
In Section 4, we compare the HF+QMC results on Im-[Co$^{\rm III}$(corrin)]-CN$^+$ with the DFT+QMC results on CNCbl.
Section 5 gives the summary and conclusions of the paper.
\section{Effective Anderson Hamiltonian}
\label{sec:1}
The multi-orbital single-impurity Anderson Hamiltonian \cite{Anderson}
is given by
\begin{eqnarray}
H &=& \sum_{m,\sigma} (\varepsilon_m-\mu) c^{\dagger}_{m\sigma} c_{m\sigma} +
\sum_{\nu,\sigma} (\varepsilon_{d\nu}-\mu)
d^{\dagger}_{\nu\sigma} d_{\nu\sigma} \nonumber \\
&+& \sum_{m,\nu,\sigma} ( V_{m\nu} c^{\dagger}_{m\sigma} d_{\nu\sigma} +
V^*_{m\nu} d^{\dagger}_{\nu\sigma} c_{m\sigma} ) \\
&+& \sum_{\nu} U_{\nu}
n_{\nu\uparrow} n_{\nu\downarrow} \nonumber
\label{hamiltonian}
\end{eqnarray}
where $c^{\dagger}_{m\sigma}$ ($c_{m\sigma}$) creates (annihilates) an electron
in host state $m$ with spin $\sigma$,
$d^{\dagger}_{\nu \sigma}$ ($d_{\nu \sigma}$) is the creation
(annihilation) operator for a localized electron with spin $\sigma$ at
the Co($3d_{\nu}$) orbital, and
$n_{\nu\sigma}= d^{\dagger}_{\nu\sigma} d_{\nu\sigma}$.
Here, $\varepsilon_m$ and $\varepsilon_{d\nu}$ are the
energies of the host and the Co($3d_{\nu}$) impurity states, respectively,
The hybridization matrix element
between these states is $V_{m\nu}$. The intra-orbital Coulomb repulsion is $U_{\nu}$.
Finally, a chemical potential $\mu$ is introduced
since the QMC calculations are performed in the grand canonical ensemble.
We obtain the one-electron parameters $\varepsilon_{m}$, $\varepsilon_{d \nu}$ and
$V_{m \nu}$ as explained below.
Within DFT, the one-electron wave function $\psi_{n}({\bf r})$ are determined from
\begin{eqnarray}
F({\bf r})\psi_{n}({\bf r})=E_n\psi_{n}({\bf r}),
\label{dftt}
\end{eqnarray}
where the Kohn-Sham operator \cite{Kohn} is
\begin{eqnarray}
\nonumber F({\bf r}) = \left(-\dfrac{\hbar^{2}}{2m_{e}}\nabla_{{\bf r}}^{2}+V_{ext}({\bf r}) + \int d^{3}r' \dfrac{\rho({\bf r'})}{\lvert {\bf r}-{\bf r'}\rvert}+ V_{xc}({\bf r}) \right)\\
\label{dft}
\end{eqnarray}
and the molecular orbital energy is $E_{n}$.
The molecular orbitals $\psi_{n}$ can be expanded in terms of the $N$ atomic orbitals,
\begin{eqnarray}
| \psi_{n} \rangle = \sum_{i}^{N} C_{ni} | \phi_{i} \rangle,
\label{psi}
\end{eqnarray}
where $C_{ni}$ are the elements of the coefficient matrix $\bf{C}$. Substituting Eq. (4) in
Eq (2), the Roothan equation, $\bf {CF=ECS}$, is obtained. Here, the elements of the Kohn-Sham
matrix $\bf{F}$ are defined as
\begin{eqnarray}
F_{ij} = \int d^{3} r\, \phi^{*}_{i}({\bf r})\, F({\bf r})\, \phi_{j}({\bf r})
\end{eqnarray}
and the overlap matrix $\bf{S}$ has the matrix elements $S_{ij} = \langle \phi_{i} | \phi_{j} \rangle$.
Because the atomic orbitals do not form an orthogonal basis, we use the
natural atomic orbitals (NAO's) \cite{Reed} which form an orthogonal basis. The NAO's form a maximally localized basis set.
Next, we express the Kohn-Sham matrix in the NAO basis. We take the Co($3d_{\nu}$) NAO's as the impurity orbitals and
their energy levels as $\varepsilon_{d \nu}$'s in the Anderson Hamiltonian. Diagonalizing the remaining part of the
Kohn-Sham matrix, we obtain the host eigenstates $|u_m\rangle$ and their energy levels $\varepsilon_{m}$ in additon
to the hybridization matrix elements $V_{m \nu}$
This procedure is explained in more detail in Ref. [\onlinecite{Kandemir2}].
We use the Gaussian program \cite{Gaussian} with the BP86 energy functional \cite{BP86} and the 6-31G basis set with $N=1035$ basis functions
to obtain the DFT solutions.
\begin{figure}
{\includegraphics[width=6.8cm]{fig2a}}
{\includegraphics[width=6.8cm]{fig2b}}
\caption{(Color online)
(a) Density of states $ D(\varepsilon) $ of CNCbl
obtained by the DFT calculations.
(b) Density of states of the host states of the effective Haldane-Anderson model
$D_h(\varepsilon)$. The shifted Co($3d_{\nu}$) natural atomic orbital levels $\tilde{\varepsilon}_{d\nu}$ are also indicated as vertical lines.
Here, the Co($3d_{\nu}$) levels have been shifted by $\mu^{\mathrm {DC}}_{\nu}$, which was obtained
for $U=4$ eV. The vertical solid and dashed
lines denote the HOMO and LUMO levels, respectively.
}
\label{fig2}
\end{figure}
In the QMC calculations, we use $U=4$ eV. In various Co compounds, the intra-orbital
Coulomb interaction at the Co($3d_{\nu}$) orbitals is estimated to be between 4 eV and
5 eV \cite{Sasioglu}. It is important to note that, in the DFT+QMC approach,
the on-site Coulomb interaction $U$ is taken into
account both by the DFT and the QMC calculations. Therefore, in order to prevent this double counting, an orbital-dependent
double-counting term $\mu_{\nu}^{\rm {DC}}$, which is defined as
\begin{eqnarray}
\mu^{\mathrm {DC}}_{\nu} = \frac{U \langle n^{\mathrm{DFT}}_{d\nu} \rangle}{2}
\end{eqnarray}
is substracted from the bare Co($3d_{\nu}$) levels,
$\varepsilon_{d\nu}
\rightarrow \tilde{\varepsilon}_{d\nu} = \varepsilon_{d\nu} - \mu_{\nu}^{\rm {DC}}$
\cite{Anisimov,Czyzyk,Kunes,Karolak}. In the Anderson Hamiltonian, $\tilde{\varepsilon}_{d\nu}$ is used instead of $\varepsilon_{d\nu}$. Here, $\langle n^{\mathrm{DFT}}_{d\nu} \rangle$
is the electron number in the Co($3d_{\nu}$) NAO's obtained by the DFT calculations.
We begin presenting data by showing the density of states $D(\varepsilon) = \sum_{n=1}^N \delta(\varepsilon-E_n)$ in Figure \ref{fig2}(a).
In this figure, the highest occupied molecular orbital (HOMO) is located at -4.9 eV,
and the lowest unoccupied molecular orbital (LUMO) is located at -3.2 eV, which means that
the energy gap is 1.7 eV.
In Fig. \ref{fig2}(b),
the host density of states $D_h(\varepsilon) = \sum_{m=1}^{N-5} \delta(\varepsilon-\varepsilon_m)$
is shown as a function of energy $\varepsilon$. In this figure, vertical lines indicate the
shifted Co($3d_{\nu}$) NAO energy levels $\tilde{\varepsilon}_{d\nu}$.
For these parameters,
while the $\nu=x^2-y^2 $ NAO is located at $ \varepsilon \approx -9.5$ eV,
the $ \nu=yz $, $ xz $ and $ 3z^2-r^2 $ NAO's are located at $ \varepsilon \approx -8.5$ eV and
the $ \nu=xy $ NAO is located at $ \varepsilon \approx -7.5$ eV.
We note that here, we choose a coordinate system in which the x and y axis are located at 45 degrees to the Co-N bond direction
instead of being parallel.
The DFT data on
the square of the hybridization matrix elements $|V_{m\nu}|^2$ between the $m$'th host eigenstate $|u_m\rangle$ and the Co($3d_{\nu}$) NAO's
are shown as a function of $\varepsilon_m$ in Fig. \ref{fig3}.
Here, we observe that the $m=336, 337$ and $340$ host states have the largest hybridization matrix
elements. The $m=336$ and $337$ host states hybridize most strongly with the Co$(3d_{xy})$ NAO, while,
the $m=340$ host state has the strongest hybridization with the Co$(3d_{3z^2-r^2})$ and Co$(3d_{xz})$ NAO's.
In the QMC data, we will see that these host states are strongly influenced by the local Coulomb interaction.
\begin{figure}
\includegraphics[width=6.8cm]{fig3a}
\includegraphics[width=6.8cm]{fig3b}
\caption{(Color online)
DFT results on the
square of the hybridization matrix elements $ |V_{m\nu}|^2 $
between the Co($3d_{\nu}$) natural atomic orbitals and the $m$'th host states
versus the host energy $ \varepsilon_{m} $. In (a)
results are shown for the $3d_{3z^2-r^2} $ and $ 3d_{xy} $ natural atomic orbitals, and
in (b) for the $3d_{xz} $, $ 3d_{x^2-y^2} $ and $ 3d_{yz} $ natural atomic orbitals.
Here, the vertical solid and dashed lines denote the values of the
HOMO and LUMO levels, respectively.
We observe that
$m=336$, $337$ and $340$'th host states have the strongest
hybridization matrix elements.
}
\label{fig3}
\end{figure}
In order to gain insight into these host states, in Fig. \ref{fig4} we illustrate
the $m=336, 337$ and $340$ host states in terms of the NAO's. These host states contain contributions from NAO's around the Co site.
In particular, the $m=340$'th host state contains significant amount of weight from the CN ligand.
\begin{figure}[t]
\centering
\includegraphics[width=5.8cm]{fig4a}
\includegraphics[width=5.8cm]{fig4b}
\includegraphics[width=5.8cm]{fig4c}
\caption{(Color online)
Illustration of the $ m=336$, $337$ and $340$'th host states in terms of the natural atomic orbitals.
}
\label{fig4}
\end{figure}
\section{Quantum Monte Carlo results}
\label{sec:2}
In this section, QMC data on the effective
Haldane-Anderson model for CNCbl are presented. For this model, QMC calculations
were performed by using the Hirsch-Fye QMC algorithm \cite{Hirsch}.
In the QMC calculations, a discrete Matsubara time step of
$ \Delta\tau = 0.13$ eV$^{-1} $ is used. The results are presented for temperature $ T=700 $ K
in the grand canonical ensemble.
Figure \ref{fig5}(a) shows
the electron occupation number $\langle n_{\nu}\rangle = \sum_{\sigma} \langle d^{\dagger}_{\nu \sigma} d_{\nu \sigma} \rangle $ for the Co($3d_{\nu}$) NAO
states as a function of the chemical potential $\mu$.
In Fig. \ref{fig5}(a), we see that
$\langle n_{\nu}\rangle$ becomes finite
at $\mu \approx -12$ eV, and the Co($3d_{\nu}$) NAO's become singly occupied at $\mu \approx -6.5$ eV.
At $\mu \approx -5.5$ eV, the Co($3d_{x^2-y^2}$) NAO becomes doubly occupied. At the HOMO level,
the $3d_{xy}$ NAO is singly occupied, while the remaining orbitals have occupation number
near $1.4$. Between the HOMO and LUMO levels, the Co($3d_{\nu}$) occupations do not change.
The occupation of the Co($3d_{xy}$) NAO exhibits a sudden increase at $\mu \approx -2.5$ eV by
about 0.3 electrons. We think that this sudden increase corresponds to an impurity bound state
located at this energy. When $\mu$ reaches $-1.0$ eV, all of the Co($3d_{\nu}$) orbitals become doubly
occupied.
An interesting observation in Fig. \ref{fig5}(a) is that the impurity bound states are not induced in the semiconducting
energy gap between the HOMO and LUMO levels. They are located above the LUMO level between $-3.0$ eV and $-2.0$ eV. Now, as seen in Fig.
\ref{fig2}, the main features of the host density of states $D_{h}(\varepsilon)$ consist of a continuous
conduction band located below $\approx -5 $ eV, two discrete states located at $-3.3$ eV and $-2.6$ eV,
and a continuum of valence band states above $-2.0$ eV. In Fig. \ref{fig5}(a), we observe that
the new impurity bound states are induced right below the continuum of valence band states located
above $-2.0$ eV. The impurity bound states are located above the LUMO level because the
spectrum of the host states is discrete around the LUMO level.
\begin{figure}
\includegraphics[width=6.8cm]{fig5a}
\includegraphics[width=6.8cm]{fig5b}
\caption{(Color online)
(a) Electron occupation number $ \langle n_{\nu} \rangle $
of the Co($3d_{\nu}$) natural atomic orbitals
versus the chemical potential $ \mu $.
(b) Square of magnetic moment $ \langle \left( M_{\nu}^{z}\right)^{2} \rangle $
for Co($3d_{\nu}$) natural atomic orbitals versus the chemical potential $ \mu $.
Here, the vertical solid and dashed lines denote the values of the HOMO and LUMO levels,
respectively. These results are for $ U=4 $ eV.
}
\label{fig5}
\end{figure}
Figure \ref{fig5}(b) shows the square of the magnetic moment at the Co($3d_{\nu}$) NAO's, $\langle (M^z_{\nu})^2\rangle$,
where $M^z_{\nu} = d^{\dagger}_{\nu \uparrow} d_{\nu \uparrow} - d^{\dagger}_{\nu \downarrow} d_{\nu \downarrow} $,
as a function of $\mu$. We observe that the magnetic moments of the Co($3d_{\nu}$) NAO's increase continuously, as $\mu$ is increased up to
$\mu \approx -6.5$ eV. The magnetic moment of the Co($3d_{x^2-y^2}$) NAO decreases rapidly at $ \mu \approx -5.5$ eV
due to double occupancy. For $ \nu = 3z^2-r^2$, $xz$ and $yz$,
the magnetic moments decrease in the interval $-6.5$ eV $ \protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}{\sim}\:$} \mu \protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <}{\sim}\:$} -5.0$ eV,
however, they do not go to zero. They have finite magnetic moments at the HOMO level. Between
the HOMO and LUMO levels, the magnetic moments do not change. For the Co($3d_{xy}$) NAO,
$\langle (M^{z}_{\nu})^2 \rangle$ exhibits a small increase at $\mu \approx -2.5$ eV. Upon
further increase of $\mu$, for all of the Co($3d_{\nu}$) NAO's, $\langle (M^{z}_{\nu})^2 \rangle $ vanishes
as the orbitals become doubly occupied.
\begin{figure}
\includegraphics[width=6.8cm]{fig6a}
\includegraphics[width=6.8cm]{fig6b}
\includegraphics[width=6.8cm]{fig6c}
\caption{(Color online)
(a) Total electron occupation number $\langle n_{\mathrm{d}} \rangle$ of the Co($3d$) natural atomic orbitals versus chemical potential $\mu$.
(b) Total number of the host electrons $\langle n_{\mathrm{h}} \rangle$ versus $\mu$.
(c) Total number of electrons $\langle n_{\mathrm{T}} \rangle= \langle n_{\mathrm{d}} \rangle + \langle n_{\mathrm{h}} \rangle$ for CNCbl versus $\mu$.
Here,
the vertical solid and dashed lines denote the HOMO and LUMO levels,
respectively. The charge neutral CNCbl molecule contains 718 electrons.
These results are for $U=4$ eV.
}
\label{fig6}
\end{figure}
In Fig. \ref{fig6}(a), we present QMC data on the total electron occupation of the Co($3d_{\nu}$) NAO's
$\langle n_{\mathrm{d}} \rangle = \sum^{5}_{\nu=1} \sum_{\sigma} \langle d_{\nu \sigma}^{\dagger} d_{\nu \sigma} \rangle$
as a function of $\mu$. We see that $\langle n_{\mathrm{d}} \rangle$ increases up to HOMO level, it equals 7.2 at
$\mu \approx -4.8$ eV. We observe that $\langle n_{\mathrm{d}} \rangle$ does not change between the HOMO and
LUMO levels.
The total number of the host electrons
$\langle n_{\mathrm{h}} \rangle = \sum^{N-5}_{m=1} \sum_{\sigma} \langle c_{m \sigma}^{\dagger} c_{m \sigma} \rangle$ is
shown in Fig. \ref{fig6}(b).
Figure \ref{fig6}(c) shows the total electron number for CNCbl $\langle n_{\mathrm{T}} \rangle = \langle n_{\mathrm{d}} \rangle + \langle n_{\mathrm{h}} \rangle$
versus $\mu$. Here, we clearly see that between the HOMO and LUMO levels the total electron number $\langle n_{\mathrm{T}} \rangle = 718$
corresponding to the neutral CNCbl molecule.
\begin{figure}
\includegraphics[width=6.8cm]{fig7a}
\includegraphics[width=6.8cm]{fig7b}
\caption{(Color online)
(a) Electron occupation number of the $m$'th host state $\langle n_m\rangle$ versus $\mu$.
(b) Square of the magnetic moment of the $m$'th host state
$\langle (M_m^z)^2 \rangle$ versus $\mu$.
Here,
the vertical solid and dashed lines denote the HOMO and LUMO levels,
respectively.
These results are for $U=4$ eV.
}
\label{fig7}
\end{figure}
In Fig. \ref{fig7}(a), we present QMC data on the host electron number $\langle n_{m} \rangle = \sum_{\sigma} \langle c^{\dagger}_{m \sigma} c_{m \sigma}
\rangle$ versus $\mu$ for the $m=336, 337$ and $340$ host eigenstates. The bare energy levels $\varepsilon_{m}$ of
these states are located at $-6.48$ eV, $-6.43$ eV and $-6.16$ eV for $m=336, 337$ and $340$, respectively.
Here, we observe that these host states do not become doubly occupied as $\mu$ passes through
the $\varepsilon_{m}$'s. For example, at the HOMO level, $\langle n_{m} \rangle = 1.72$, $1.67$ and
$1.52$ for $m=336, 337$ and $340$ respectively even though they are located deep below the HOMO level. Consequently, these host states have finite
magnetic moments when $\mu$ is at the HOMO level as seen in Fig. \ref{fig7}(b). The magnetic moments vanish
after these host states become doubly occupied for $\mu \protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle >}{\sim}\:$} -1$ eV.
\begin{figure}[!htpb]
\includegraphics[width=6.8cm]{fig8a}
\includegraphics[width=6.8cm]{fig8b}
\includegraphics[width=6.8cm]{fig8c}
\caption{(Color online)
Magnetic correlation function
$\langle M_{\nu}^z M_m^z \rangle$
between the $m$'th host state and
the Co($3d_{\nu}$) natural atomic orbitals.
Here, results are shown for the host states
(a) $m=336$, (b) 337, and (c) 340.
The vertical solid and dashed lines denote the HOMO and LUMO levels,
respectively.
These results are for $U=4$ eV.
}
\label{fig8}
\end{figure}
Next, in Fig. \ref{fig8}(a)-(c), we present QMC data on the magnetic correlation function
$\langle M^{z}_{\nu} M^{z}_{m} \rangle$ between the magnetic moments at the Co($3d_{\nu}$) NAO's and the
$m$'th host states for $m=336, 337$ and $340$. These figures show that the host states with the strongest
hybridization have antiferromagnetic correlations with the moments at the Co($3d_{\nu}$) NAO's.
These antiferromagnetic correlations vanish as the host states become doubly occupied.
\section{Comparison of the HF+QMC and the DFT+QMC results}
In Ref. [\onlinecite{Kandemir2}], numerical results were presented from previous HF+QMC calculations
on Im-[Co$^{\rm III}$(corrin)]-CN$^+$, which is a smaller piece of the CNCbl molecule. Here, we
compare those results with the current DFT+QMC data obtained for the whole CNCbl molecule.
We find various differences between the outcomes of the HF and DFT calculations. In the HF method,
non-interacting electrons are described under the influence of a mean field potential
which consists of the classical Coulomb potential and a non-local exchange potential. On the
other hand, in the DFT calculations a local exchange potential is used \cite{Harrison}. Furthermore,
in the HF+QMC approach, the intra-orbital Coulomb interaction was assumed to be unscreened and the
bare Coulomb matrix elements were used. For cobalt, the intra-orbital $U$ was taken to be about $36$ eV.
On the other hand, in the DFT+QMC approach, the intra-orbital Coulomb interaction is assumed to be renormalized due to long-range
screening effects. Here, $U$ was taken to be $4$ eV.
In the HF calculations, the locations of the Co($3d_{\nu}$) NAO's separate into two groups
corresponding to the $e_{g}$ and $t_{2g}$ symmetries. In the HF+QMC calculations, we have used a constant
energy shift $\mu^{\mathrm{DC}}$ to compensate for the double counting of $U$. The resultant QMC data reflect
this ordering of the Co($3d_{\nu}$) NAO's. However, the DFT calculations yield nearly degenerate Co($3d_{\nu}$)
NAO's. In the DFT+QMC approach, we have used an orbital dependent $\mu_{\nu}^{\mathrm{DC}}$ which led
to the ordering of the Co($3d_{\nu}$) NAO's seen in Fig. \ref{fig2}(b). We find that there are some differences
in the locations of the shifted Co($3d_{\nu}$) NAO's where we compare the HF and DFT results. These differences clearly influence the outcome of the
QMC calculations. Nevertheless, we observe that impurity bound states involving the Co($3d_{xy}$) and
Co($3d_{3z^{2}-r^{2}}$) NAO's are found in both the HF+QMC and DFT+QMC calculations. In both cases, the impurity
bound state for the Co($3d_{3z^{2}-r^{2}}$) NAO is located higher in energy compared to that of the Co($3d_{xy}$).
\section{Summary and conclusions}
In summary, we have studied the electronic structure and magnetic correlations of cyanocobalamin. For this purpose, we have used the
multi-orbital single-impurity Halda\-ne-\-An\-der\-son model of a transition metal impurity embedded in a semiconductor host.
First, we have constructed an effective Haldane-Anderson model by using DFT calculations.
We have obtained the one-electron
parameters of this model from the Kohn-Sham matrix written in the basis of the natural atomic orbitals.
We have taken $U=4$ eV and have shifted the Co$(3d_{\nu})$ levels by $\mu^{\mathrm{DC}}_{\nu}$
to prevent the double-counting of the intra-orbital Coulomb interaction by both DFT and QMC.
The QMC results clearly show how the single-electron spectral weight is distributed in energy.
In our calculations, we see that as the chemical potential increases, the Co($3d_{\nu}$) NAO's become occupied. When the chemical potential
re\-ac\-hes the HOMO level, the total electron number equals 718. This corresponds to the neutral CNCbl molecule.
In this case, the Co($3d_{\nu}$) NAO's are less than doubly occupied and they have finite magnetic moments.
Between the HOMO and LUMO levels, there is no single-particle spectral weight. We observe that
above the LUMO level and between $-3.0$ eV and $-2.0$ eV, there are new states induced by the Coulomb
interaction $U$. We identify these new states as impurity bound states because of the filling
dependence of the antiferromagnetic correlations between the Co($3d_{\nu}$) NAO's and the host magnetic
moments. This identification is similar to that done previously in the analysis of the HF+QMC
results presented in Ref. [\onlinecite{Kandemir2}].
The impurity bound state is most clearly seen for the Co($3d_{xy}$) NAO. Surprisingly according to the DFT+QMC result, the impurity bound
states are located above the LUMO level instead of being in the semiconducting energy gap. We think that this is
because of the discrete energy spectrum of the CNCbl molecule.
It remains to be seen whether the impurity
bound states found in the DFT+QMC calculations are related to the peaks observed in the photoabsorption
spectrum of CNCbl. For a more direct comparison with the experimental data, it would be necessary to include the inter-orbital Coulomb
interactions along with the Hund's coupling. We note that we have perfomed similar DFT+QMC calculations for
hemoglobin, where we also find impurity bound states. Hence, these correlated electronic states appear to be a common feature of
metalloproteins and metalloenzymes. We think that it will be interesting to figure out whether the impurity
bound states have a general role in the functioning of metalloproteins and metalloenzymes.
\begin{acknowledgements}
We thank
Hadi M. Zareie, Tahir \c{C}a\u{g}{\i}n, Mehmet Sar{\i}kaya,
Nuran Elmac{\i}, \"{O}zg\"{u}r \c{C}ak{\i}r,
Devrim G\"{u}\c{c}l\"{u},
Jingyu Gan, Bo Gu, and Sadamichi Maekawa
for valuable discussions and suggestions.
The numerical calculations reported here were performed
in part at the TUBITAK ULAKBIM, High Performance and Grid
Computing Center (TRUBA resources).
Financial support by the Turkish Scientific and Technical Research Council
(TUBITAK grant numbers 110T387 and 113F242) is gratefully acknowledged.
\end{acknowledgements}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Constrained mechanical multibody systems arise in many areas of applications
such as robotics, biomechanics of locomotion of humans and dynamics of
vehicle and machinery \cite{Haug,Truck-model,Knee,Wash-m-1}. The dynamical
behavior of constrained multibody systems is described by the Euler-Lagrange
equation
\begin{equation}
\left\{
\begin{array}{l}
\dfrac{dp}{dt}=v, \\
M(p)\dfrac{dv}{dt}=f\left( p,v\right) -G^{^{\text{\textsf{T}}}}\left(
p\right) \lambda \text{,} \\
0=g\left( p\right) \text{, \ }t\geq 0\text{.
\end{array
\right. \label{M1}
\end{equation
Here $t$ is the time, $p\left( t\right) \in \mathbb{R}^{n_{p}}$ and $v\left(
t\right) \in \mathbb{R}^{n_{p}}$ specify the positions and the orientations
of all bodies and their velocities, respectively. The vector $\lambda \in
\mathbb{R}^{n_{\lambda }}$ is the vector of Lagrange multipliers. The matrix
$M\left( p\right) \in \mathbb{R}^{n_{p}}\times \mathbb{R}^{n_{p}}$ is a mass
matrix. The mapping $f:\mathbb{R}^{n_{p}}\times \mathbb{R
^{n_{p}}\longrightarrow \mathbb{R}^{n_{p}}$\ defines the applied and
internal forces (other than the constraint forces), whereas $g:\mathbb{R
^{n_{p}}\longrightarrow \mathbb{R}^{n_{\lambda }},$ ($n_{\lambda }\leq n_{p}
) defines the constraints. The term $G(p)^{^{\text{\textsf{T}}}}\lambda $
represents the constraint forces, where $G\left( p\right) =\partial g/dp\in
\mathbb{R}^{n_{\lambda }\times n_{p}}$ denotes the Jacobian of $g\left(
p\right) $. \
The Euler-Lagrange equations (\ref{M1}) form a nonlinear system of
differential-algebraic equations (DAEs). Consistent initial conditions
\begin{equation}
p\left( 0\right) =p_{0}\text{, \ }v\left( 0\right) =v_{0}\text{,} \label{M2}
\end{equation
are necessary to uniquely determine a solution. The given vectors $p_{0}\
and $v_{0}$, which specify the initial configuration and initial velocity,
are chosen so that the consistency equations
\begin{eqnarray}
g\left( p_{0}\right) &=&0, \label{Consistency-1} \\
G(p_{0})v_{0} &=&0, \label{Consistency-2}
\end{eqnarray
are satisfied. For the variable $\lambda \left( t\right) ,$ no initial
condition is prescribed because $\lambda \left( 0\right) $ is already
determined by DAE (\ref{M1}) and initial conditions (\ref{M2}).
Throughout this paper, we assume that $M\left( p\right) $, $f\left(
p,v\right) $ and $g\left( p\right) $ are analytical. A standard assumption
on the Jacobian $G\left( p\right) $ is the full row rank conditio
\begin{equation}
\text{rank }G\left( p\right) =n_{\lambda }, \label{Full-rank}
\end{equation
which means that the constraint equations defined by $g(p)$ are linearly
independent.
In addition the matrix $M$ is assumed to be symmetric and positive definite,
that is
\begin{equation}
z^{^{\text{\textsf{T}}}}M\left( p\right) z>0,\text{for all }z\in \text{Ker
G(p). \label{Positive}
\end{equation}
\noindent If assumptions (\ref{Full-rank})-(\ref{Positive}) hold, then the
matrix
\begin{equation}
\left(
\begin{array}{cc}
M\left( p\right) & G^{^{\text{\textsf{T}}}}\left( p\right) \\
G\left( p\right) &
\end{array
\right) , \label{nonsingular}
\end{equation
is nonsingular and DAE (\ref{M1}) is therefore index-three. For any
consistent initial conditions (\ref{M2}), system (\ref{M1}) has a unique
solution \cite{Haug}. The index we consider here is the differential index
which is the minimum number of times that all or a part of the DAE must be
differentiated with respect to time in order to obtain an ordinary
differential equation \cite{Index}. It is well-known that index-three DAEs
present difficulties for numerical integration methods \cite{Brenan}.
Therefore, several techniques have been proposed in the literature to solve
DAE system (\ref{M1}) \cite{Brenan,Baumgarte,Stab-0,Stab-1,
Robot-example,Brahim-SC}. A very popular way is to reduce the index by
differentiating the constraints one or more times with respect to time
before applying numerical integration methods. However, the main problem
with this technique is that the numerical solution of the index-reduced
system may no longer satisfy the constraints of original DAE (\ref{M1}) due
to error propagation. Constraints violation, known also as drift-off
phenomena, leads to non physical solutions. To overcome this difficulty,
some techniques like stabilization or augmented Lagrangian formulation have
been proposed to keep the constraint violations under control during the
numerical integration \cite{Baumgarte,Stab-0,Stab-1,Stab-2}. The most
popular stabilization method is that of Baumgarte \cite{Baumgarte}, but its
drawback is the way of choosing its feedback parameters. The augmented
Lagrangian formulation \cite{Aug-Lag} has the same problem of parameter
selection. The challenge is therefore to construct efficient methods that
provide solutions of the Euler-Lagrange equations which satisfy the
constraints in these equations.
The Adomian decomposition method (ADM) and its modifications \cite{Alge-ADM,
ADM-ode-1, ADM-ode-2, ADM-ode-3, ADM-ode-4, ADM-Mod-1, ADM-ode-Mod-2} are
known to be efficient methods in solving a large variety of linear and
nonlinear problems in science and engineering. Among these problems, we
mention algebraic equations \cite{Alge-ADM}, ordinary differential equations
\cite{ADM-ode-1, ADM-ode-2, ADM-ode-3, ADM-ode-4, ADM-Mod-1, ADM-ode-Mod-2},
partial differential equations \cite{ADM-PDEs} and integral equations \cit
{ADM-Mod-Integrals}.
In this work, we present a new approach to solve the Euler-Lagrange
equations using ADM. The solution by this method satisfies all the DAE
constraints. The ADM\ is first applied directly to the Euler-Lagrange
equations where the nonlinear terms are expanded using the Adomian
polynomials \cite{ADP-1,ADP-2,ADP-3,ADP-4,ADP-5,ADP-6}. Based on the index
of the Euler-Lagrange equations, a nonsingular linear algebraic recursion
system is derived for the expansion components of the solution. Our
technique has the great advantage that it does not use complex
transformations like index reductions before applying the ADM to the
equations. To demonstrate the effectiveness of the proposed method, we solve
an example of the Euler-Lagrange equations that models a two-link planar
robotic system. Further, our technique is based on a simple algorithm that
can be programmed in Maple or Mathematica to simulate real application
problems.
This paper is organized as follows: in section \ref{ADM}, we review the ADM
for solving ordinary differential equations. Next, in section \ref{Multibody
system-ADM}, we present our method for the solution of the Euler-Lagrange
equations. Then, in section \ref{Test problems}, we apply the developed
technique to solve an example of the Euler-Lagrange equations that models a
two-link \ planar robotic system. Finally, a discussion and a conclusion are
given in sections \ref{Discussion} and \ref{Conclusion}, respectively.
\section{Adomian decomposition method}
\label{ADM}In this section, we give a brief review for the Adomian
decomposition method (ADM) \cite{Alge-ADM, ADM-ode-1, ADM-ode-2, ADM-ode-3,
ADM-ode-4, ADM-Mod-1, ADM-ode-Mod-2} to solve ordinary differential
equations. For this purpose, let us consider the following nonlinear
differential equatio
\begin{equation}
Lu+Ru+N(u)=f, \label{ADM-1}
\end{equation
where $L$ is an easily invertible operator (usually taken as the
highest-order derivative), $R$ is an operator grouping the remaining
lower-order derivatives, $N\left( u\right) $ is the nonlinear term and $f$
is a given analytical function.
Solving equation (\ref{ADM-1}) for $Lu$ then applying the inverse operator
L^{-1}$ to both sides, we obtai
\begin{equation}
L^{-1}Lu=L^{-1}f-L^{-1}Ru-L^{-1}N\left( u\right) . \label{ADM-2}
\end{equation
If $Lu=du/dt$ and the initial condition $u\left( t_{0}\right) =u_{0}$ is
given, then $L^{-1}$ represents the integral from $t_{0}$ to $t$ and
L^{-1}Lu=u-u_{0}.$
\begin{equation}
u=u_{0}+L^{-1}f-L^{-1}Ru-L^{-1}N\left( u\right) . \label{ADM-3}
\end{equation
To apply the ADM to equation (\ref{ADM-3}), we first assume that the
solution $u$ of (\ref{ADM-1}) to have the infinite series form
\begin{equation}
u=\sum_{n=0}^{\infty }u^{\left( n\right) }, \label{ADM-4}
\end{equation
where the unknown solution components $u^{\left( n\right) },$
n=0,1,2,\ldots $ are to be determined later by the method.
Second, the nonlinear term $N(u)$ is expanded in an infinite series in terms
of the Adomian polynomials $N^{\left( n\right) }$ \cit
{ADP-1,ADP-2,ADP-3,ADP-4,ADP-5,ADP-6} as
\begin{equation}
N\left( u\right) =\sum_{n=0}^{\infty }N^{\left( n\right) }\left( u^{\left(
0\right) },\ldots ,u^{\left( n\right) }\right) .\text{ } \label{ADM-5}
\end{equation
Substituting (\ref{ADM-4}) and (\ref{ADM-5}) into (\ref{ADM-3}) and choosing
$u^{\left( 0\right) }$ as
\begin{equation}
u^{\left( 0\right) }=u_{0}+L^{-1}f\text{,} \label{ADM-6}
\end{equation
we obtai
\begin{equation}
\sum_{n=0}^{\infty }u^{\left( n\right) }=u^{\left( 0\right)
}-L^{-1}R\sum_{n=0}^{\infty }u^{\left( n\right) }-L^{-1}\sum_{n=0}^{\infty
}N^{\left( n\right) }. \label{ADM-7}
\end{equation
Comparing the general term on the left hand side with that on the right hand
side, we derive the following recursion scheme for the ADM
\begin{equation}
u^{\left( n\right) }=-L^{-1}Ru^{\left( n-1\right) }-L^{-1}N^{\left(
n-1\right) },\text{ \ }n\geq 1. \label{ADM-8}
\end{equation
Since $u^{\left( 0\right) }$ is known, recursion (\ref{ADM-8}) can be used
to generate as many solution components $u^{\left( n\right) }$ as one wants.
Further, if series (\ref{ADM-4}) converges then it gives the exact solution
of (\ref{ADM-1}) and an approximation of order $n_{0}$ to solution can be
obtained from
\begin{equation}
u=\sum_{n=0}^{n_{0}-1}u^{\left( n\right) }. \label{ADM-9}
\end{equation
To compute the Adomian polynomials $N^{\left( n\right) }$, $n=0,1,\ldots $
associated with the nonlinearity $N\left( u\right) ,$ one can use the
following definition for all forms of nonlinearity
\begin{equation}
N^{\left( n\right) }:=N^{\left( n\right) }\left( u^{\left( 0\right) },\ldots
,u^{\left( n\right) }\right) =\frac{1}{n!}\frac{d^{n}}{d\lambda ^{n}}\left(
N\left( \sum_{i=0}^{\infty }\lambda ^{i}u^{\left( i\right) }\right) \right)
_{\lambda =0}\text{, \ }n\geq 0. \label{ADM-10}
\end{equation
Using this formula, we obtain the following first few Adomian polynomials
\begin{eqnarray}
N^{\left( 0\right) } &=&N\left( u^{\left( 0\right) }\right) , \notag \\
N^{\left( 1\right) } &=&u^{\left( 1\right) }N^{\prime }\left( u^{\left(
0\right) }\right) , \notag \\
N^{\left( 2\right) } &=&u^{\left( 2\right) }N^{\prime }\left( u^{\left(
0\right) }\right) +\dfrac{\left( u^{\left( 1\right) }\right) ^{2}}{2!
N^{\prime \prime }\left( u^{\left( 0\right) }\right) , \label{ADM-11} \\
N^{\left( 3\right) } &=&u^{\left( 3\right) }N^{\prime }\left( u^{\left(
0\right) }\right) +u^{\left( 1\right) }u^{\left( 2\right) }N^{\prime \prime
}\left( u^{\left( 0\right) }\right) +\dfrac{\left( u^{\left( 1\right)
}\right) ^{3}}{3!}N^{\prime \prime \prime }\left( u^{\left( 0\right)
}\right) , \notag \\
N^{\left( 4\right) } &=&u^{\left( 4\right) }N^{\prime }\left( u^{\left(
0\right) }\right) +\left( \dfrac{\left( u^{\left( 2\right) }\right) ^{2}}{2!
+u^{\left( 1\right) }u^{\left( 3\right) }\right) N^{\prime \prime }\left(
u^{\left( 0\right) }\right) +\dfrac{\left( u^{\left( 1\right) }\right)
^{2}u^{\left( 2\right) }}{2!}N^{\prime \prime \prime }\left( u^{\left(
0\right) }\right) +\dfrac{\left( u^{\left( 1\right) }\right) ^{4}}{4!
N^{\prime \prime \prime \prime }\left( u^{\left( 0\right) }\right) , \notag
\end{eqnarray
where the dash $(^{\prime })$ represents the differentiation with respect to
$u$.
In a similar manner, one can easily generate the remaining polynomials from
\ref{ADM-10}). In the literature, there are several algorithms for computing
the Adomian polynomials without the need for formula (\ref{ADM-10}), but a
more convenient algorithm for the $m$-variable case is recently proposed in
\cite{ADP-6}
\begin{equation}
N^{\left( n\right) }=\dfrac{1}{n}\sum_{k=1}^{m}\sum_{i=0}^{n-1}\left(
i+1\right) v_{k}^{\left( i+1\right) }\frac{\partial N^{\left( n-1-i\right)
}{\partial v_{k}^{\left( 0\right) }},\text{ \ }n\geq 1. \label{ADM-12}
\end{equation}
\section{The proposed method}
\label{Multibody system-ADM}In this section, we present our method for
solving the Euler-Lagrange equations. These equations are known to be
difficult to treat numerically since they represent an index-three system of
differential-algebraic equations (DAEs). The technique we propose here is
based on the Adomian decomposition method (ADM). To solve these equations,
we first apply the ADM directly to them and expand the nonlinear terms using
the Adomian polynomials. Then, an algebraic linear recursion system for the
solution expansion components is derived. Taking account of the index of the
DAE, this algebraic system is shown to be uniquely solvable for the solution
expansion components. The main advantage of our technique is that it does
not require to transform the equations to lower index DAEs before applying
the ADM to them. We start our technique by the following proposition.
\textbf{Proposition}: Let $u=\left( u_{1},\ldots ,u_{l}\right) ^{^{\text
\textsf{T}}}}\in \mathbb{R}^{l},$ $v=\left( v_{1},\ldots ,v_{n}\right) ^{^
\text{\textsf{T}}}}$ $\in \mathbb{R}^{n}$ be two vectors, $A:=(A\left(
u\right) _{i,j})\in \mathbb{R}^{m}$ $\times \mathbb{R}^{n}$ be a matrix,
u=\sum_{k=0}^{\infty }u^{\left( k\right) },v=\sum_{k=0}^{\infty }v^{\left(
k\right) },$ $u^{\left( k\right) }=$ $\left( u_{1}^{\left( k\right) },\ldots
,u_{l}^{\left( k\right) }\right) ,$ $v^{\left( k\right) }=$ $\left(
v_{1}^{\left( k\right) },\ldots ,v_{n}^{\left( k\right) }\right) .$ Let
A^{\left( k\right) }$ $=\left( A_{i,j}^{\left( k\right) }\right) ,$ where
A_{i,j}^{\left( k\right) }=A_{i,j}^{\left( k\right) }\left( u^{\left(
0\right) },\ldots ,u^{\left( k\right) }\right) $ denotes the $k$-th Adomian
polynomial of the entry $A_{i,j}$. Then, given $A$ and $v$, the Adomian
polynomials of $z=Av$ and $z=A^{-1}v$ are given by
(a) $z^{\left( k\right) }=\sum_{l=0}^{k}A^{\left( l\right) }v^{\left(
k-l\right) }=$ $\sum_{l=0}^{k}A^{\left( k-l\right) }v^{\left( l\right) }$ and
(b) $A^{\left( 0\right) }z^{\left( k\right) }=v^{\left( k\right)
}-\sum_{l=0}^{k-1}A^{\left( k-l\right) }z^{\left( l\right) },$ $k=0,1,\ldots
$
\textbf{Proof}:
(a) Let $z=Av$, then the Adomian polynomials $z^{\left( k\right) }$ of $z$
are given by
\begin{eqnarray*}
z^{\left( k\right) } &=&\left( Av\right) ^{\left( k\right) }=\left(
\begin{array}{c}
\left( \sum_{j=1}^{n}A_{1,j}v_{j}\right) ^{\left( k\right) } \\
\vdots \\
\left( \sum_{j=1}^{n}A_{m,j}v_{j}\right) ^{\left( k\right)
\end{array
\right) =\left(
\begin{array}{c}
\sum_{j=1}^{n}\left( A_{1,j}v_{j}\right) ^{\left( k\right) } \\
\vdots \\
\sum_{j=1}^{n}\left( A_{m,j}v_{j}\right) ^{\left( k\right)
\end{array
\right) , \\
&=&\left(
\begin{array}{c}
\sum_{j=1}^{n}\sum_{l=0}^{k}A_{1,j}^{\left( k-l\right) }v_{j}^{\left(
l\right) } \\
\vdots \\
\sum_{j=1}^{n}\sum_{l=0}^{k}A_{m,j}^{\left( k-l\right) }v_{j}^{\left(
l\right)
\end{array
\right) =\left(
\begin{array}{c}
\sum_{l=0}^{k}\sum_{j=1}^{n}A_{1,j}^{\left( k-l\right) }v_{j}^{\left(
l\right) } \\
\vdots \\
\sum_{l=0}^{k}\sum_{j=1}^{n}A_{m,j}^{\left( k-l\right) }v_{j}^{\left(
l\right)
\end{array
\right) , \\
&=&\sum_{l=0}^{k}A^{\left( k-l\right) }v^{\left( l\right) }.
\end{eqnarray*}
(b) Let $Az=v$ then, using (a), the Adomian polynomials $v^{\left( k\right) }
$ of $v$ are given by
\begin{equation*}
v^{\left( k\right) }=\left( Az\right) ^{\left( k\right)
}=\sum_{l=0}^{k}A^{\left( k-l\right) }z^{\left( l\right) }=A^{\left(
0\right) }z^{\left( k\right) }+\sum_{l=0}^{k-1}A^{\left( k-l\right)
}z^{\left( l\right) }.
\end{equation*
The Adomian polynomials $z^{\left( k\right) }$ of $\ z=A^{-1}v$ are then
given by
\begin{equation*}
A^{\left( 0\right) }z^{\left( k\right) }=v^{\left( k\right)
}-\sum_{l=0}^{k-1}A^{\left( k-l\right) }z^{\left( l\right) }.
\end{equation*}
\bigskip
Now to solve DAE system (\ref{M1}), let $Lp\left( t\right) =dp/dt$ and
L^{-2}p\left( t\right) =L^{-1}(L^{-1}p\left( t\right) ).$ Then
L^{-1}p\left( t\right) =\int_{0}^{t}p\left( t\right) dt$ and $L^{-2}p\left(
t\right) =\int_{0}^{t}\int_{0}^{t}p\left( t\right) dtdt.$
We solve the second equation of (\ref{M1}) for $dv/dt$ as
\begin{equation}
\dfrac{dv}{dt}=M^{-1}\left( p\right) \Big(f\left( p,v\right) -G^{^{\text
\textsf{T}}}}\left( p\right) \lambda \Big)\text{.} \label{M3-new}
\end{equation
Applying the operator $L^{-1}$ to both sides of the first equation of (\re
{M1}) and (\ref{M3-new}) then using initial conditions (\ref{M2}), we get
\begin{equation}
\left\{
\begin{array}{l}
p=p_{0}+L^{-1}v\text{,} \\
v=v_{0}+L^{-1}\Big(M^{-1}\left( p\right) \Big(f\left( p,v\right) -G^{^{\text
\textsf{T}}}}\left( p\right) \lambda \Big)\Big)
\end{array
\right. \label{M4}
\end{equation
Then, we expand the solution components $p,v$ and $\lambda $ as
\begin{equation}
p=\sum_{n=0}^{\infty }p^{\left( n\right) }\text{, \ }v=\sum_{n=0}^{\infty
}v^{\left( n\right) }\text{, \ }\lambda =\sum_{n=0}^{\infty }\lambda
^{\left( n\right) }, \label{M6}
\end{equation
where the unknowns $p^{\left( n\right) },v^{\left( n\right) }$ and $\lambda
^{\left( n\right) },$ $n=0,1,2,\ldots $ will be determined later by our
method. The nonlinear terms $f\left( p,v\right) $ and\ $g(p)$ are also
expanded in infinite series as
\begin{equation}
f\left( p,v\right) =\sum_{n=0}^{\infty }f^{\left( n\right) }\text{, \
g\left( p\right) =\sum_{n=0}^{\infty }g^{\left( n\right) }\text{ ,}
\label{M8}
\end{equation
where $f^{\left( n\right) }:=f^{\left( n\right) }\left( p^{\left( 0\right)
},v^{\left( 0\right) },\ldots ,p^{\left( n\right) },v^{\left( n\right)
}\right) $ and $g^{\left( n\right) }:=g^{\left( n\right) }\left( p^{\left(
0\right) },\ldots ,p^{\left( n\right) }\right) $ denote the Adomian
polynomials. Using (\ref{ADM-12}), the Adomian polynomials $f^{\left(
n\right) }$ and $g^{\left( n\right) },$ $n=0,1,2,\ldots $ can be written a
\begin{equation}
f^{\left( n\right) }=\left\{
\begin{array}{l}
f\left( p^{\left( 0\right) },v^{\left( 0\right) }\right) ,\text{ \ }n=0, \\
\dfrac{1}{n}\sum_{i=1}^{n-1}i\left( \dfrac{\partial f^{\left( n-i\right) }}
\partial p^{\left( 0\right) }}\right) p^{\left( i\right) }+i\left( \dfrac
\partial f^{\left( n-i\right) }}{\partial v^{\left( 0\right) }}\right)
v^{\left( i\right) }+\left( \dfrac{\partial f^{\left( 0\right) }}{\partial
p^{\left( 0\right) }}\right) p^{\left( n\right) }+\left( \dfrac{\partial
f^{\left( 0\right) }}{\partial v^{\left( 0\right) }}\right) v^{\left(
n\right) },\text{ \ }n\geq 1
\end{array
\right. \label{M10}
\end{equation
an
\begin{equation}
g^{\left( n\right) }=\left\{
\begin{array}{l}
g\left( p^{\left( 0\right) }\right) ,\text{ \ }n=0, \\
\dfrac{1}{n}\sum_{i=1}^{n-1}i\left( \dfrac{\partial g^{\left( n-i\right) }}
\partial p^{\left( 0\right) }}\right) p^{\left( i\right) }+\left( \dfrac
\partial g^{\left( 0\right) }}{\partial p^{\left( 0\right) }}\right)
p^{\left( n\right) },\text{ \ }n\geq 1
\end{array
\right. \label{M12}
\end{equation
Substituting expansions (\ref{M6}) into equations (\ref{M4}) and the third
equation of (\ref{M1}), we ge
\begin{equation}
\left\{
\begin{array}{l}
\sum_{n=0}^{\infty }p^{\left( n\right) }=p_{0}+\sum_{n=0}^{\infty
}L^{-1}v^{\left( n\right) }, \\
\sum_{n=0}^{\infty }v^{\left( n\right) }=v_{0}+\sum_{n=0}^{\infty }L^{-1
\Big(M^{-1}\left( p\right) \Big(f\left( u,v\right) -G^{^{\text{\textsf{T}
}}\left( p\right) \lambda \Big)\Big)^{\left( n\right) }, \\
0=\sum_{n=0}^{\infty }g^{\left( n\right) }
\end{array
\right. \label{M14}
\end{equation
Choosing the initial terms $p^{\left( 0\right) }$ and $v^{\left( 0\right) }$
as
\begin{equation}
p^{\left( 0\right) }=p_{0}\text{, \ }v^{\left( 0\right) }=v_{0}, \label{M18}
\end{equation
then comparing the general terms on the left and right hand sides of (\re
{M14}), we obtain the following recursion system
\begin{equation}
\left\{
\begin{array}{l}
p^{\left( n\right) }=L^{-1}v^{\left( n-1\right) }, \\
v^{\left( n\right) }=L^{-1}\Big(M^{-1}\left( p\right) \Big(f\left(
p,v\right) -G^{^{\text{\textsf{T}}}}\left( p\right) \lambda \Big)\Big
^{\left( n-1\right) }, \\
0=g^{\left( n\right) },\text{ \ }n\geq 1
\end{array
\right. \label{M20}
\end{equation
This system leads to the following recursion system
\begin{equation}
\left\{
\begin{array}{l}
p^{\left( n\right) }=L^{-2}\Big(M^{-1}\left( p\right) \Big(f\left(
p,v\right) -G^{T}\left( p\right) \lambda \Big)\Big)^{\left( n-2\right) }, \\
0=g^{\left( n\right) },\text{ \ }n\geq 2
\end{array
\right. \label{M22}
\end{equation
wher
\begin{equation}
v^{\left( n-1\right) }=Lp^{\left( n\right) }. \label{M24}
\end{equation
Using expansions (\ref{M8}) and the previous proposition, we calculate the
right hand side of the first equation of system (\ref{M22}), and obtain the
following linear algebraic system for the unknowns $p^{\left( n\right) }$
and $L^{-2}\lambda ^{\left( n-2\right) }$
\begin{eqnarray}
M\left( p^{\left( 0\right) }\right) p^{\left( n\right) }+G^{^{\text{\textsf{
}}}}(p^{\left( 0\right) })L^{-2}\lambda ^{\left( n-2\right) } &=&R\text{,}
\label{M25} \\
G(p^{\left( 0\right) })p^{\left( n\right) } &=&S,\text{ \ }n\geq 2, \notag
\end{eqnarray
where
\begin{equation*}
R=L^{-2}\Bigg(f^{\left( n-2\right) }-\sum_{k=0}^{n-3}\Big(\left( G^{^{\text
\textsf{T}}}}(p)\right) ^{\left( n-2-k\right) }\lambda ^{\left( k\right)
}+\left( M\left( p\right) \right) ^{\left( n-2-k\right) }z^{\left( k\right)
\Big)\Bigg),
\end{equation*
and
\begin{equation*}
S=-\dfrac{1}{n}\sum_{i=1}^{n-1}i\left( \dfrac{\partial g^{\left( n-i\right)
}{\partial p^{\left( 0\right) }}\right) p^{(i)}.
\end{equation*
The iterates $z^{\left( k\right) },$ $k=2,\ldots ,n-3$ are computed fro
\begin{equation}
M\left( p^{\left( 0\right) }\right) z^{\left( k\right) }=f^{\left( k\right)
}-G^{^{\text{\textsf{T}}}}(p^{\left( 0\right) })\lambda ^{\left( k\right)
}-\sum_{l=0}^{k-1}\Big(\left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left(
k-l\right) }\lambda ^{\left( l\right) }+\left( M\left( p\right) \right)
^{\left( k-l\right) }z^{\left( l\right) }\Big). \label{M26}
\end{equation
In system (\ref{M25}),\ the right hand side depends only on previous
iterations $p^{\left( n-1\right) },\ldots ,p^{\left( 0\right) },\lambda
^{\left( n-3\right) },\ldots ,\lambda ^{\left( 0\right) }$ and $v^{\left(
n-1\right) },\ldots ,v^{\left( 0\right) }$. Since the Jacobian $G$ has full
row rank and the matrix$\ M$ is positive definite, then system (\ref{M25})
determines $p^{\left( n\right) }$ and $L^{-2}\lambda ^{\left( n-2\right) }$
uniquely for $n$ $\geq 2.$
One way to solve system (\ref{M25}) is to multiply the first equation of
this system from left by the matrix $G(p^{\left( 0\right) })M^{-1}(p^{\left(
0\right) }).$ Then substitute $G(p^{\left( 0\right) })p^{\left( n\right) }$
by its expression from the second equation of (\ref{M25}), to obtain the
following nonsingular algebraic system for the unknown $L^{-2}\lambda
^{\left( n-2\right) }
\begin{equation}
G(p^{\left( 0\right) })M^{-1}\left( p^{\left( 0\right) }\right) G^{^{\text
\textsf{T}}}}(p^{\left( 0\right) })L^{-2}\lambda ^{\left( n-2\right)
}=G(p^{\left( 0\right) })M^{-1}\left( p^{\left( 0\right) }\right) R-S.
\label{M27}
\end{equation
Since rank condition (\ref{Full-rank}) holds and $M$ is positive definite,
equation (\ref{M27}) can be solved uniquely for $L^{-2}\lambda ^{\left(
n-2\right) }$ to ge
\begin{equation}
L^{-2}\lambda ^{\left( n-2\right) }=\left( G(p^{\left( 0\right)
})M^{-1}\left( p^{\left( 0\right) }\right) G^{^{\text{\textsf{T}
}}(p^{\left( 0\right) })\right) ^{-1}\left( G(p^{\left( 0\right)
})M^{-1}\left( p^{\left( 0\right) }\right) R-S\right) . \label{M29}
\end{equation
Now applying the operator $L^{2}$ to both sides of equation (\ref{M29}), we
can determine the unknown $\lambda ^{\left( n-2\right) }$
\begin{equation}
\lambda ^{\left( n-2\right) }=\left( G(p^{\left( 0\right) })M^{-1}\left(
p^{\left( 0\right) }\right) G^{^{\text{\textsf{T}}}}(p^{\left( 0\right)
})\right) ^{-1}\left( G(p^{\left( 0\right) })M^{-1}\left( p^{\left( 0\right)
}\right) L^{2}R-L^{2}S\right) . \label{M30}
\end{equation
Then, substituting the expression of $L^{-2}\lambda ^{\left( n-2\right) }$
into the first equation of (\ref{M25}), we determine the unknown $p^{\left(
n\right) }$
\begin{equation}
p^{\left( n\right) }=M^{-1}\left( p^{\left( 0\right) }\right) R-M^{-1}\left(
p^{\left( 0\right) }\right) G^{^{\text{\textsf{T}}}}(p^{\left( 0\right)
})\left( G(p^{\left( 0\right) })M^{-1}\left( p^{\left( 0\right) }\right)
G^{^{\text{\textsf{T}}}}(p^{\left( 0\right) })\right) ^{-1}\left(
G(p^{\left( 0\right) })M^{-1}\left( p^{\left( 0\right) }\right) R-S\right) .
\label{M31}
\end{equation
Now, using equation (\ref{M24}) we can calculate $v^{\left( n-1\right) }.$
Finally, we obtain an approximate solution for DAE initial-value problem
\ref{M1})-(\ref{M2}) as
\begin{equation}
p(t)=\displaystyle{{\sum_{n=0}^{n_{0}-1}}}p^{\left( n\right) }\text{, \
v(t)=\displaystyle{{\sum_{n=0}^{n_{0}-2}}}v^{\left( n\right) }\text{, \
\lambda (t)=\displaystyle{{\sum_{n=0}^{n_{0}-3}}}\lambda ^{\left( n\right) },
\label{MDAE-19}
\end{equation
where $n_{0}$ is the order of approximation of $p(t).$
\section{Application}
\label{Test problems}
In this section, we illustrate and demonstrate the effectiveness of our
technique to solve Euler-Lagrange equations (\ref{M1})-(\ref{M2}) which
describe the motion of constrained mechanical multibody systems. These
equations are known to be difficult to solve numerically because they are
index-three differential-algebraic equations (DAEs). Following the procedure
developed in the previous section, we first apply the Adomian decomposition
method (ADM) directly to these equations without using complex
transformations like index-reductions. Then, we expand the nonlinear terms
using the Adomian polynomials. Taking account of the index-three condition,
we derive a nonsingular linear algebraic recursion system for the expansion
components of the solution. Finally, by solving this algebraic system, we
obtain the solution of the Euler-Lagrange equations. As a test problem, we
consider the following example of constrained multibody system made up from
example 6.4 in \cite{Robot-example} which describes a two-link planar
robotic system, where the mass matrix i
\begin{equation*}
M(\theta _{1},\theta _{2})=\left(
\begin{array}{cc}
m_{1}l_{1}^{2}/3+m_{2}\left( l_{1}^{2}+l_{2}^{2}/3+l_{1}l_{2}\cos \theta
_{2}\right) & m_{2}\left( l_{2}^{2}/3+(1/2)l_{1}l_{2}\cos \theta _{2}\right)
\\
m_{2}\left( l_{2}^{2}/3+(1/2)l_{1}l_{2}\cos \theta _{2}\right) &
m_{2}l_{2}^{2}/
\end{array
\right) .
\end{equation*
The force term is
\begin{equation*}
f\left( \theta _{1},\theta _{2},d\theta _{1}/dt,d\theta _{2}/dt\right)
=\left(
\begin{array}{c}
\Big(l_{1}\cos \theta _{1}+l_{2}\cos \left( \theta _{1}+\theta _{2}\right)
\Big)\left( d\theta _{1}/dt\right) -3\theta _{1} \\
\Big(l_{2}\cos \left( \theta _{1}+\theta _{2}\right) \Big)\left( d\theta
_{1}/dt\right) +\left( 1-\left( 3/2\right) \cos \theta _{2}\right) \theta
_{1
\end{array
\right) \text{, \ }
\end{equation*
and the constraint function is given b
\begin{equation*}
g\left( \theta _{1},\theta _{2}\right) =l_{1}\sin \theta _{1}+l_{2}\sin
\left( \theta _{1}+\theta _{2}\right) .
\end{equation*
Taking $l_{1}=l_{2}=1,m_{1}=m_{2}=3$ and using the notation of (\ref{M1}),
with $n_{p}=2,$ $n_{\lambda }=1,$ $p=\left( p_{1},p_{2}\right) ^{^{\text
\textsf{T}}}}=\left( \theta _{1},\theta _{2}\right) ^{^{\text{\textsf{T}}}},$
$v=dp/dt$, we have
\begin{equation*}
M\left( p\right) =\left(
\begin{array}{cc}
5+3\cos p_{2} & \left( 3/2\right) \cos p_{2} \\
1+\left( 3/2\right) \cos p_{2} &
\end{array
\right) ,
\end{equation*
and
\begin{equation*}
f\left( p,v\right) =\left(
\begin{array}{c}
\Big(\cos p_{1}+\cos \left( p_{1}+p_{2}\right) \Big)v_{1}-3p_{1} \\
\Big(\cos \left( p_{1}+p_{2}\right) \Big)v_{1}+\left( 1-\left( 3/2\right)
\cos p_{2}\right) p_{1
\end{array
\right) \text{. \ }
\end{equation*
The constraint function becomes
\begin{equation*}
g\left( p\right) =\sin p_{1}+\sin \left( p_{1}+p_{2}\right) ,
\end{equation*
and its Jacobian $G\left( p\right) =$ $\left( \cos p_{1}+\cos \left(
p_{1}+p_{2}\right) ,\cos \left( p_{1}+p_{2}\right) \right) $ is full row
rank $n_{\lambda }=1.$
For the consistent initial conditions
\begin{equation}
p\left( 0\right) =\left( 0,0\right) ^{^{\text{\textsf{T}}}}\text{, \
v\left( 0\right) =\left( 1,-2\right) ^{^{\text{\textsf{T}}}},\text{\ }
\label{Ex-1-c}
\end{equation
the exact solution for this example is $p\left( t\right) =\left( \sin
t,-2\sin t\right) ^{^{\text{\textsf{T}}}}$, \ $v\left( t\right) =\left( \cos
t,-2\cos t\right) ^{^{\text{\textsf{T}}}}$ \ and $\lambda \left( t\right)
=\cos t.$
The Euler-Lagrange equations corresponding to this example form an
index-three DAE and therefore difficult to solve numerically. Using the
procedure developed in the previous section, system (\ref{M25}) can be
solved for $n=2,3,\ldots $ to reveal the dynamics of the mechanical system.
For $n=2,$ we ge
\begin{eqnarray}
p^{\left( 2\right) }+M^{-1}\left( p^{\left( 0\right) }\right) G^{^{\text
\textsf{T}}}}(p^{\left( 0\right) })L^{-2}\lambda ^{\left( 0\right) }
&=&L^{-2}M^{-1}\left( p^{\left( 0\right) }\right) f^{\left( 0\right) },
\label{Ex-6-a} \\
G(p^{\left( 0\right) })p^{\left( 2\right) } &=&0. \label{Ex-6-b}
\end{eqnarray
Now, since
\begin{equation*}
M^{-1}\left( p^{\left( 0\right) }\right) =\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) ,\ G(p^{\left( 0\right) })=\left( 2,1\right) ,
\end{equation*
\begin{equation*}
p^{\left( 0\right) }=\left(
\begin{array}{c}
p_{1}^{\left( 0\right) } \\
p_{2}^{\left( 0\right)
\end{array
\right) =\left(
\begin{array}{c}
0 \\
\end{array
\right) ,\ v^{\left( 0\right) }=\left(
\begin{array}{c}
v_{1}^{\left( 0\right) } \\
v_{2}^{\left( 0\right)
\end{array
\right) =\left(
\begin{array}{c}
1 \\
-
\end{array
\right) ,
\end{equation*
an
\begin{equation*}
f^{\left( 0\right) }=\left(
\begin{array}{c}
\Big(\cos p_{1}^{\left( 0\right) }+\cos \left( p_{1}^{\left( 0\right)
}+p_{2}^{\left( 0\right) }\right) \Big)v_{1}^{\left( 0\right)
}-3p_{1}^{\left( 0\right) } \\
\Big(\cos \left( p_{1}^{\left( 0\right) }+p_{2}^{\left( 0\right) }\right)
\Big)v_{1}^{\left( 0\right) }+\left( 1-\left( 3/2\right) \cos p_{2}^{\left(
0\right) }\right) p_{1}^{\left( 0\right)
\end{array
\right) =\left(
\begin{array}{c}
2 \\
\end{array
\right) ,
\end{equation*
equations (\ref{Ex-6-a})-(\ref{Ex-6-b}) reduce t
\begin{eqnarray}
p^{\left( 2\right) }+\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 0\right) } &=&\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( 1\right) , \label{Ex-7-a} \\
\left( 2,1\right) p^{\left( 2\right) } &=&0. \label{Ex-7-b}
\end{eqnarray
Multiplying system (\ref{Ex-7-a}) from left by the matrix $\left( 2,1\right)
$ then using (\ref{Ex-7-b}), we ge
\begin{equation}
\left( 2/7\right) \left( 2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 0\right) }=\left( 2/7\right) \left(
2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( 1\right) , \label{Ex-8}
\end{equation
which leads
\begin{equation}
\lambda ^{\left( 0\right) }=1. \label{Ex-9}
\end{equation
Now substituting the value of $\lambda ^{\left( 0\right) }$ from (\ref{Ex-9
) into (\ref{Ex-7-a}), we obtai
\begin{equation}
p^{\left( 2\right) }+\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( 1\right) =\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( 1\right) , \label{Ex-10}
\end{equation
which gives
\begin{equation}
p^{\left( 2\right) }=\left(
\begin{array}{c}
0 \\
\end{array
\right) , \label{Ex-11}
\end{equation
and using (\ref{M24}), we have
\begin{equation}
v^{\left( 1\right) }=Lp^{\left( 2\right) }=\left(
\begin{array}{c}
0 \\
\end{array
\right) . \label{Ex-12}
\end{equation
For $n=3$, we hav
\begin{eqnarray}
p^{\left( 3\right) }+M^{-1}\left( p^{\left( 0\right) }\right) G^{^{\text
\textsf{T}}}}(p^{\left( 0\right) })L^{-2}\lambda ^{\left( 1\right) }
&=&L^{-2}M^{-1}\left( p^{\left( 0\right) }\right) \Bigg(f^{\left( 1\right) }
\notag \\
&&-\left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 1\right) }\lambda
^{\left( 0\right) }-\left( M\left( p\right) \right) ^{\left( 1\right)
}z^{\left( 0\right) }\Bigg)\text{,} \label{Ex-13} \\
G(p^{\left( 0\right) })p^{\left( 3\right) } &=&0. \notag
\end{eqnarray
Now, since
\begin{equation*}
\left( M\left( p\right) \right) ^{\left( 1\right) }=\left(
\begin{array}{cc}
-3p_{2}^{\left( 1\right) }\sin p_{2}^{\left( 0\right) } & -\left( 3/2\right)
p_{2}^{\left( 1\right) }\sin p_{2}^{\left( 0\right) } \\
-\left( 3/2\right) p_{2}^{\left( 1\right) }\sin p_{2}^{\left( 0\right) } &
\end{array
\right) =0,\ \left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 1\right)
}=0,\
\end{equation*
an
\begin{equation*}
f^{\left( 1\right) }=\left(
\begin{array}{c}
-3 \\
-1/
\end{array
\right) t,
\end{equation*
system (\ref{Ex-13}) reduces t
\begin{eqnarray}
p^{\left( 3\right) }+\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 1\right) } &=&\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
-3 \\
-1/
\end{array
\right) L^{-2}\left( t\right) , \label{Ex-14-a} \\
&&\left( 2,1\right) p^{\left( 3\right) }=0. \label{Ex-14-b}
\end{eqnarray
Multiplying system (\ref{Ex-14-a}) from left by the matrix $(2,1)$ then
using (\ref{Ex-14-b}), we ge
\begin{eqnarray}
\left( 2/7\right) \left( 2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 1\right) } &=&\left( 2/7\right) \left(
2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
-3 \\
-1/
\end{array
\right) L^{-2}\left( t\right) , \notag \\
&=&0, \label{Ex-15}
\end{eqnarray
which gives
\begin{equation}
\lambda ^{\left( 1\right) }=0. \label{Ex-17}
\end{equation
Now substituting the value of $\lambda ^{\left( 1\right) }$ from (\ref{Ex-17
) into (\ref{Ex-14-a}), we obtai
\begin{equation}
p^{\left( 3\right) }=\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
-3 \\
-1/
\end{array
\right) L^{-2}\left( t\right) , \label{Ex-18}
\end{equation
which gives
\begin{equation}
p^{\left( 3\right) }=-\dfrac{t^{3}}{3!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) , \label{Ex-19}
\end{equation
and
\begin{equation}
v^{\left( 2\right) }=Lp^{\left( 3\right) }=-\dfrac{t^{2}}{2!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) . \label{Ex-20}
\end{equation
For $n=4$, we hav
\begin{eqnarray}
p^{\left( 4\right) }+M^{-1}\left( p^{\left( 0\right) }\right) G^{^{\text
\textsf{T}}}}(p^{\left( 0\right) })L^{-2}\lambda ^{\left( 2\right) }
&=&L^{-2}M^{-1}\left( p^{\left( 0\right) }\right) \Bigg(f^{\left( 2\right) }
\notag \\
&&-\left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 2\right) }\lambda
^{\left( 0\right) }-\left( M\left( p\right) \right) ^{\left( 2\right)
}z^{\left( 0\right) } \notag \\
&&-\left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 1\right) }\lambda
^{\left( 1\right) }-\left( M\left( p\right) \right) ^{\left( 1\right)
}z^{\left( 1\right) }\Bigg)\text{,} \label{Ex-21} \\
G(p^{\left( 0\right) })p^{\left( 4\right) } &=&-\dfrac{1}{4
\sum_{i=1}^{3}i\left( \dfrac{\partial g^{\left( 4-i\right) }}{\partial
p^{\left( 0\right) }}\right) p^{(i)}, \notag
\end{eqnarray
and
\begin{equation*}
f^{\left( 2\right) }=\left(
\begin{array}{c}
-2 \\
-
\end{array
\right) t^{2},\ \left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 2\right)
}=\left(
\begin{array}{c}
-1 \\
-1/
\end{array
\right) t^{2}.
\end{equation*
The iterates $z^{\left( 0\right) }$ is calculated fro
\begin{equation*}
M\left( p^{\left( 0\right) }\right) z^{\left( 0\right) }=f^{\left( 0\right)
}-G^{^{\text{\textsf{T}}}}(p^{\left( 0\right) })\lambda ^{\left( 0\right)
}=\left(
\begin{array}{c}
2 \\
\end{array
\right) -\left(
\begin{array}{c}
2 \\
\end{array
\right) \left( 1\right) =0,
\end{equation*
which gives $z^{\left( 0\right) }=\left(
\begin{array}{c}
0 \\
\end{array
\right) .$ Thus (\ref{Ex-21}) becomes
\begin{eqnarray}
p^{\left( 4\right) }+\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 2\right) } &=&\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( -1/2t^{2}\right) , \label{Ex-22-a} \\
(2,1)p^{\left( 4\right) } &=&0. \label{Ex-22-b}
\end{eqnarray
Multiplying system \ (\ref{Ex-22-a}) from left by the matrix $(2,1)$ then
using (\ref{Ex-22-b}), we ge
\begin{equation}
\left( 2/7\right) \left( 2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 2\right) }=\left( 2/7\right) \left(
2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( -1/2t^{2}\right) , \label{Ex-24}
\end{equation
which gives
\begin{equation}
\lambda ^{\left( 2\right) }=-1/2t^{2}. \label{Ex-25}
\end{equation
Now substituting the value of $\lambda ^{\left( 2\right) }$ from (\ref{Ex-25
) into (\ref{Ex-22-a}), we obtai
\begin{equation*}
p^{\left( 4\right) }+\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( -1/2t^{2}\right) =\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\left( -1/2t^{2}\right) ,
\end{equation*
which give
\begin{equation}
p^{\left( 4\right) }=\left(
\begin{array}{c}
0 \\
\end{array
\right) , \label{Ex-26}
\end{equation
and using (\ref{M24}), we have
\begin{equation}
v^{\left( 3\right) }=Lp^{\left( 4\right) }=\left(
\begin{array}{c}
0 \\
\end{array
\right) . \label{Ex-27}
\end{equation
For $n=5$, we hav
\begin{eqnarray}
p^{\left( 5\right) }+M^{-1}\left( p^{\left( 0\right) }\right) G^{^{\text
\textsf{T}}}}(p^{\left( 0\right) })L^{-2}\lambda ^{\left( 3\right) }
&=&L^{-2}M^{-1}\left( p^{\left( 0\right) }\right) \Bigg(f^{\left( 3\right) }
\notag \\
&&-\sum_{k=0}^{2}\Big(\left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left(
3-k\right) }\lambda ^{\left( k\right) }+\left( M\left( p\right) \right)
^{\left( 3-k\right) }z^{\left( k\right) }\Big)\Bigg)\text{,} \label{Ex-28}
\\
G(p^{\left( 0\right) })p^{\left( 5\right) } &=&-\dfrac{1}{5
\sum_{i=1}^{4}i\left( \dfrac{\partial g^{\left( 5-i\right) }}{\partial
p^{\left( 0\right) }}\right) p^{(i)}, \notag
\end{eqnarray
where
\begin{equation*}
f^{\left( 3\right) }=\left(
\begin{array}{c}
1/2 \\
37/1
\end{array
\right) t^{3},\ \left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 3\right)
}=0.
\end{equation*
\begin{equation*}
\left( M(p)\right) ^{\left( 2\right) }=\left(
\begin{array}{cc}
-6t^{2} & -3t^{2} \\
-3t^{2} &
\end{array
\right) .
\end{equation*
The term $z^{\left( 1\right) }$ is calculated from
\begin{eqnarray*}
M\left( p^{\left( 0\right) }\right) z^{\left( 1\right) } &=&f^{\left(
1\right) }-\left( G^{^{\text{\textsf{T}}}}(p)\right) ^{\left( 1\right)
}\lambda ^{\left( 0\right) }-G^{^{\text{\textsf{T}}}}(p^{\left( 0\right)
})\lambda ^{\left( 1\right) }-\left( M\left( p\right) \right) ^{\left(
1\right) }z^{\left( 0\right) } \\
&=&f^{\left( 1\right) }=\left(
\begin{array}{c}
-3 \\
-1/
\end{array
\right) t,
\end{eqnarray*
which gives $z^{\left( 1\right) }=\left(
\begin{array}{c}
-1 \\
\end{array
\right) t.$ Thus system (\ref{Ex-28}) becomes
\begin{eqnarray}
p^{\left( 5\right) }+\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 3\right) } &=&\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
1/2 \\
1/1
\end{array
\right) L^{-2}\left( t^{3}\right) , \label{Ex-29-a} \\
(2,1)p^{\left( 5\right) } &=&0. \label{Ex-29-b}
\end{eqnarray
Multiplying system (\ref{Ex-29-a}) from left by the matrix $(2,1)$ then
using (\ref{Ex-29-b}), we ge
\begin{eqnarray}
\left( 2/7\right) \left( 2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
2 \\
\end{array
\right) L^{-2}\lambda ^{\left( 3\right) } &=&\left( 2/7\right) \left(
2,1\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
1/2 \\
1/1
\end{array
\right) L^{-2}\left( t^{3}\right) , \notag \\
&=&0, \label{Ex-30}
\end{eqnarray
which gives
\begin{equation}
\lambda ^{\left( 3\right) }=0. \label{Ex-31}
\end{equation
Now substituting the value of $\lambda ^{\left( 3\right) }$ from (\ref{Ex-31
) into (\ref{Ex-29-a}), we obtai
\begin{equation}
p^{\left( 5\right) }=\left( 2/7\right) \left(
\begin{array}{cc}
2 & -5 \\
-5 & 1
\end{array
\right) \left(
\begin{array}{c}
1/2 \\
1/1
\end{array
\right) L^{-2}\left( t^{3}\right) , \label{Ex-32}
\end{equation
which give
\begin{equation}
p^{\left( 5\right) }=\dfrac{t^{5}}{5!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) , \label{Ex-33}
\end{equation
and using (\ref{M24}), we have
\begin{equation}
v^{\left( 4\right) }=Lp^{\left( 5\right) }=\dfrac{t^{4}}{4!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) . \label{Ex-34}
\end{equation
Continuing this process until $n=8$, we obtain the following ADM solution
\begin{equation}
p\left( t\right) =t\left(
\begin{array}{c}
1 \\
-
\end{array
\right) -\dfrac{t^{3}}{3!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) +\dfrac{t^{5}}{5!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) -\dfrac{t^{7}}{7!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) , \label{Ex-35}
\end{equation
\begin{equation}
v\left( t\right) =\left(
\begin{array}{c}
1 \\
-
\end{array
\right) -\dfrac{t^{2}}{2!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) +\dfrac{t^{4}}{4!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) -\dfrac{t^{6}}{6!}\left(
\begin{array}{c}
1 \\
-
\end{array
\right) , \label{EX-35a}
\end{equation
and
\begin{equation}
\lambda \left( t\right) =1-\dfrac{t^{2}}{2!}+\dfrac{t^{4}}{4!}-\dfrac{t^{6}}
6!}. \label{Ex-36}
\end{equation
These are the first few terms of Taylor expansions, around $t=0,$ of
\begin{equation}
p\left( t\right) =\left( \sin t,-2\sin t\right) ^{^{\text{\textsf{T}}}}\text
, \ \ }v\left( t\right) =\left( \cos t,-2\cos t\right) ^{^{\text{\textsf{T}
}}\text{, \ \ }\lambda \left( t\right) =\cos t, \label{Ex-37}
\end{equation
which is the exact solution of the problem in this example.
\section{Discussion}
\label{Discussion}The Euler-Lagrange equations are known to be difficult to
solve numerically. The reason is that they form an index-three system of
differential-algebraic equations (DAEs). In this paper, we propose a novel
technique that applies the Adomian decomposition method (ADM) directly to
solve the Euler-Lagrange equations. This technique has successfully handled
these equations without the need for complex transformations like
index-reductions. This method transforms these equations into easily
solvable algebraic systems for the expansion components of the solution. To
illustrate the effectiveness of the proposed technique, an example of the
Euler-Lagrange equations describing a two-link planar robot system is
solved. This example shows that the ADM is a simple powerful tool to obtain
the exact or approximate solutions of the Euler-Lagrange equations.
\section{Conclusion}
\label{Conclusion}
This work presents the analytical solution of the Euler-Lagrange equations
using the ADM. A \ procedure for solving these is presented. The technique
was tested on an example of the Euler-Lagrange equations that describes a
two-link robot system. The results obtained show that the proposed method
can be applied to solve the Euler-Lagrange equations efficiently to obtain
the exact or an approximate solution. On the one hand, it is important to
note that these types of equations are difficult to solve and on the other,
the direct application of the ADM was able to solve the Euler-Lagrange
equations. Also, it is important to note that, our technique does not make
transformations to the equations before applying the ADM to them. The
technique is based on a straightforward procedure that can be programmed in
Maple or Mathematica to simulate real application problems. Finally, further
work is needed to apply a multistage ADM form to solve the Euler-Lagrange
equations and other semi-explicit nonlinear higher-index DAEs. \
\textbf{Conflict of Interests}
The author declares that there is no conflict of interests regarding the
publication of this paper.
|
\section{Introduction}
Magnetic trapping is one of the workhorses for cold atom physics. A commonly used
trap is the Ioffe-Pritchard trap \cite{pritchard1983}, in which the magnetic field is non-zero
at the center to prevent Majorana transitions to scattering states. In most situations,
it is highly desirable to obtain high trapping frequencies and gradients for fast thermalization
and strong confinement of the atoms. One method to reach high frequencies is to miniaturize
the magnetic trap by using microfabricated chips \cite{Fortagh2007}. This far most magnetic
traps operate in a regime where the atomic spin can adiabatically follow local changes in
the magnetic field. \cite{ AnglinSchmiedmayer, Franzosi, Hinds}.
In this article, we analyze the situation in magnetic traps with very high trap frequencies,
where the spin is coupled to the motion of the atom. This allows us to investigate how
the Born-Oppenheimer approximation fails and how effective magnetic and electric fields
appear as a consequence of the non-adiabatic dynamics
\cite{Berry1990,mead1992,dalibard2011,goldman2014}. The results are based on exact numerical diagonalization of the full Hamiltonian describing the coupling between the internal and external degrees of freedom. Using the complex rotation method, the position in energy and the decay rate of the trapping states correspond to the imaginary part of the resonances of this Hamiltonian are computed \cite{Buchl94}.
\section{Magnetic trapping}
\subsection{Atoms in a magnetic field}
The Hamiltonian of an atom with mass $M$ and spin operator ${\bf J}=(J_x,J_y,J_z)$ in a magnetic
field ${\bf B}$, is given by
\begin{eqnarray}
H = \frac{{\bf p}^2}{2M} +
\frac{g \mu_B}{\hbar} {\bf J} \cdot {\bf B} ,
\end{eqnarray}
where $\mu_B$ is the Bohr magneton and $g$ is the $g-$factor. We consider the $J=1$
case. By writing
\begin{eqnarray}
{\bf B} = B (\sin{\alpha}\cos{\beta},\sin{\alpha}\sin{\beta},\cos{\alpha}),
\end{eqnarray}
where $B$, $\alpha$ and $\beta$ values are varying in space, the Hamiltonian reads
\begin{eqnarray}
H & = & \frac{{\bf p}^2}{2M}
\nonumber \\
& + & \frac{ \mu_B B_{tot}}{ \sqrt{2}}\left(
\begin{array}{ccc}
\sqrt{2}\cos\alpha & e^{-i\beta}\sin\alpha & 0 \\
e^{i\beta}\sin\alpha& 0 & e^{-i\beta}\sin\alpha \\
0 & e^{i\beta}\sin\alpha & -\sqrt{2} \cos \alpha
\end{array}
\right)
\label{HS1},
\end{eqnarray}
in the eigenbasis $|-1\rangle, |0\rangle, |1\rangle$ of $J_z$.
When the Larmor frequency $\omega_L=\mu_B B_{tot}/\hbar$ is much larger than the
typical frequency of the atomic motion, the fast spin dynamics decouple from the slow
evolution of the center of mass of the atom. In the Born-Oppenheimer approximation,
the state of the system $\Psi({\bf r})$ is written as
\begin{equation}
|\Psi ({\bf r}) \rangle \approx \psi_j ({\bf r}) |\chi_j ({\bf r})\rangle \label{ket} ,
\end{equation}
where $\psi_j {({\bf r})}$ is the wave function and $|\chi_j ({\bf r}) \rangle$ is an eigenvector
of the position dependent spin-part of the Hamiltonian $H$ in Eq. (\ref{HS1}). The effective
potential seen by the atoms is just the associated eigenvalue $V_j ({\bf r})$, whose minimum,
if existing, creates a trapping potential. In addition, since the local eigenstates
$|\chi_j ({\bf r})\rangle$ are position dependent a vector potential and an additional
scalar potential appear, for each component
$\psi_j{({\bf r})}$:
\begin{eqnarray}
{\bf A}_j ({\bf r}) & = & - i \hbar\langle \chi_j ({\bf r})| \nabla \chi_j ({\bf r})\rangle ,
\nonumber \\
\Phi_j ({\bf r}) & = & \frac{\hbar^2}{2}\left[\bra{\nabla\chi_{j}({\bf r})}
(1-\ket{\chi_{j}({\bf r})}\bra{\chi_{j}({\bf r})}) \cdot \ket{\nabla\chi_{j}({\bf r})}\right].
\nonumber \\
\end{eqnarray}
These quantities are well-known in molecular physics as Berry-Mead and Born-Huang
potentials~\cite{mead1992}. In cold atomic gases, they appear in the situation of position
dependent dark-states \cite{Rus}, allowing experimental realizations of artificial
gauge fields \cite{Lin}. The eigenstates $\ket{\chi_j ({\bf r})}$ of the spin Hamiltonian
are only defined up to a phase factor, which can be position dependent. From the preceding
expression, one readily obtains that the change $|\chi_j ({\bf r})\rangle\rightarrow
e^{if({\bf r})}|\chi_j ({\bf r})\rangle$ amounts
to the gauge transformation:
\begin{eqnarray}
{\bf A}_j ({\bf r}) & \rightarrow & {\bf A}_j ({\bf r})+ \nabla f({\bf r}) ,
\nonumber \\
\Phi_j ({\bf r}) & \rightarrow & \Phi_j ({\bf r}) .
\label{gauget}
\end{eqnarray}
In addition to these adiabatic terms, there are also coupling between the different eigenstates
$|\chi_j ({\bf r})\rangle$, which appears as off-diagonal terms of the Hamiltonian in the
$|\chi_j({\bf r})\rangle$ basis. They are precisely responsible for the Majorana losses.
\subsection{Experimental set-up}
To observe the effects of the effective Berry-Mead and Born-Huang potentials, it is
required to use magnetic traps with high trap frequencies. One efficient way to realize this
is to use microfabricated magnetic traps, which are formed when the magnetic field from
a small current-carrying wire is superimposed with an external bias field. The magnetic field
from a thin and long wire along the $z$- axis is given by ${\bf B}_{\theta}(r) =
\frac{\mu_0}{2 \pi} \frac{I_{\omega}}{r} {\bf e}_\theta$, where $\theta$ is the polar angle in
cylindrical coordinates, and $r$ is the distance from the wire. We assume a homogeneous
external bias field ${\bf B}_{\textrm{bias}}= (\frac{B_0}{\sqrt{2}}, -\frac{B_0}{\sqrt{2}},B_z)$.
Superimposing ${\bf B}_{\theta}(r)$ and ${\bf B}_{\textrm{bias}}$, we obtain, in cartesian coordinates,
the total magnetic field
\begin{eqnarray}
{\bf B} & = & (\frac{B_0}{\sqrt{2}}-\frac{\mu_0}{2\pi}\frac{I_{\omega}y}{x^2+y^2},
-\frac{B_0}{\sqrt{2}}+\frac{\mu_0}{2\pi}\frac{I_{\omega}x}{x^2+y^2},B_z)
\nonumber \\
& = & (\tilde{B}_x,\tilde{B}_y,\tilde{B_z}) .
\end{eqnarray}
The magnitude of ${\bf B}$ is given by
\begin{equation}
B^2 = B_z^2+B_0^2-\frac{\sqrt{2} B_0 \xi (x+y)}{x^2+y^2} + \frac{\xi^2}{x^2+y^2}
\end{equation}
with $\xi=\frac{\mu_0 I_{\omega}}{2 \pi}$. This has a minimum at the position
$x_0 = y_0 = \frac{\xi}{\sqrt{2} B_0}$, where weak-field seeking atoms can be trapped.
Introducing local coordinates $(\tilde{x},\tilde{y})$ around the minimum $(x_0,y_0)$,
the $xy$-components of the magnetic field read:
\begin{eqnarray}
\tilde{B}_x & = & \frac{\frac{B_0}{\sqrt{2}}
((x_0+\tilde{x})^2+(y_0+\tilde{y})^2) - \xi (y_0+\tilde{y})}{(x_0+\tilde{x})^2+(y_0+\tilde{y})^2} ,
\nonumber \\
\tilde{B}_y & = & \frac{-\frac{B_0}{\sqrt{2}}
((x_0+\tilde{x})^2+(y_0+\tilde{y})^2)+\xi (x_0+\tilde{x})}{(x_0+\tilde{x})^2+(y_0+\tilde{y})^2}
\end{eqnarray}
to first order in $\tilde{x}$ and $\tilde{y}$
\begin{equation}
\tilde{B}_x \approx \frac{B_0^2}{\xi} \tilde{x}, \ \ \ \tilde{B}_y \approx -\frac{B_0^2}{\xi} \tilde{y} .
\end{equation}
This approximation can be readily obtained from the linearization of ${\bf B}$ at the
minimum of the potential: ${\bf B}\approx(G\tilde{x},-G\tilde{y},B_z)$, where $G$ is the gradient
of the magnetic field in the $(\tilde{x},\tilde{y})$ plane, i.e.,
\begin{equation}
G = \frac{B_0^2}{\xi} = \frac{2 \pi}{\mu_0}\frac{B_0^2}{I_{\omega}} .
\end{equation}
\subsection{Harmonic units}
As explained just above, weak-field seeking atoms can be trapped around the minimum of the
magnetic field; more precisely, the effective trapping potential is proportional to:
\begin{eqnarray}
\mu_B B =\mu_B B_z \sqrt{1+\frac{G^2}{ B_z^2}\left(\tilde{x}^2+\tilde{y}^2\right)} ,
\end{eqnarray}
where the harmonic approximation leads an effective trap frequency (and harmonic length)
\begin{eqnarray}
\omega_T = \sqrt{\frac{\mu_B B_z}{M}} \frac{G}{B_z} ,
\qquad \ell_T = \sqrt{\frac{\hbar}{M\omega_T}} .
\end{eqnarray}
Using $\ell_T$ and $\omega_T$ as harmonic units for length and energy and by introducing
the Larmor-type frequency $\omega_L= \mu_B B_z/ \hbar$, the Hamiltonian for a spin-$1$
reads
\begin{eqnarray}
& & \frac{H}{\hbar\omega_T}=\frac{{\bf p}^2}{2}
\nonumber \\
& & + \frac{V({\bf r})}{\sqrt{2}}\left( \begin{array}{ccc}
\sqrt{2}\cos\alpha & e^{-i\beta}\sin\alpha & 0 \\
e^{i\beta}\sin\alpha& 0 & e^{-i\beta}\sin\alpha \\
0 & e^{i\beta}\sin\alpha & -\sqrt{2} \cos \alpha
\end{array}\right) ,
\end{eqnarray}
where
\begin{equation}
V({\bf r})=\frac{\mu_B B_{tot}}{\hbar\omega_T}=\frac{\mu_B B_z}{\hbar\omega_T}
\sqrt{1+\frac{G^2\ell_T^2}{B_z^2}(x^2+y^2)} .
\end{equation}
We are now using the notation $x$ and $y$ for the scaled position around the minimum, not
to be confused with the original notation. In this rescaled units, one has ${\bf p}=-i\nabla$.
With the above definition, it follows that
\begin{eqnarray}
\frac{\mu_B B_z}{\hbar\omega_T} & = & \frac{\omega_L}{\omega_T} \equiv \rho^{-2} ,
\nonumber \\
\frac{G^2\ell_T^2}{B_z^2} & = & \frac{\omega_T}{\omega_L} \equiv \rho^2 ,
\end{eqnarray}
in terms of which
\begin{equation}
V({\bf r})=\frac{1}{\rho^2}\sqrt{1+\rho^2(x^2+y^2)} .
\end{equation}
The ratio $\rho$ precisely compares the motional dynamics (frequency trap) to the spin dynamics
(Larmor frequency). As one can see, the full dynamics of the problem depend only on this single
dimensionless parameter. In the usual trapping situation, this is a small parameter, typically
ranging from $10^{-3}$ to $10^{-1}$, telling that, (i) the Born-Oppenheimer approximation is
valid, and (ii) the trap is almost harmonic $V({\bf r})\approx\frac{1}{\rho^2}+\frac{1}{2}(x^2+y^2)$.
On the other hand, for $\rho\approx 1$, i.e., the timescale of the spin and the motional dynamics
are comparable, deviations from the harmonic behavior are marked and all the effects beyond
the Born-Oppenheimer approximation become important, in particular the Majorana's losses.
Finally, as functions of the magnetic field gradient $G$ (in $T/m$) and the bias field $B_z$
(in Gauss), one has for Rb$^{87}$:
\begin{equation}
\begin{aligned}
\omega_T&=\sqrt{\frac{\mu_B}{M}}\frac{G}{\sqrt{B_z}}=8.135\times10^2\frac{G}{\sqrt{B_z}}\\
\nu_T&=\frac{\omega_T}{2\pi}=129.47\frac{G}{\sqrt{B_z}}\\
\frac{\omega_T}{\omega_L}&=\frac{\hbar}{\sqrt{\mu_BM}}\frac{G}{B_z^{3/2}} =
9.25\times10^{-5}\frac{G}{B_z^{3/2}}.
\end{aligned}
\end{equation}
For a given value of $G$ and $B_z$, the dimensionfull
values for frequencies and the decay rates
are related to the numerical ones as follows:
\begin{equation}
\begin{aligned}
\nu_{\text{exp}}&=\nu_T\times E_{\text{num}} , \\
\frac{\Gamma_{\text{exp}}}{2\pi}&=\nu_T\times\Gamma_{\text{num}} ,
\end{aligned}
\end{equation}
see below for application.
\section{Beyond the Born-Oppenheimer approximation}
\subsection{Effective Hamiltonian}
As mentioned above, beside the vector potentials ${\bf A}_j$ and the Born-Huang scalar potentials
$\Phi_j$, off-diagonal couplings between the components $\psi_j$ arise because of the motion
of the atoms. From a mathematical point of view, this is nothing but saying that the operator
$\frac{{\bf p}^2}{2}$ does not commute with the position dependent diagonalization of the
spin Hamiltonian. More precisely, decomposing the state of the system $\ket{\Psi({\bf r})}$
in the eigenstates of the spin Hamiltonian, $\ket{\Psi ({\bf r})} = \sum_j \psi_j ({\bf r})
|\chi_j({\bf r}) \rangle$, one can derive the effective Hamiltonian $H_{\textrm{eff}}$ acting
on the wave functions $\psi_j ({\bf r})$. One has
\begin{equation}
\begin{aligned}
\ket{\chi_{-}({\bf r})} & = \sin^2{\frac{\alpha}{2}}\ket{1} -
\frac{1}{\sqrt{2}}\sin \alpha e^{i \beta}\ket{0} \\
& + \cos^2{\frac{\alpha}{2}} e^{2 i \beta}\ket{-1} , \\
\ket{\chi_0({\bf r})} & = -\frac{1}{\sqrt{2}} \sin \alpha \ket{1}+\cos \alpha
e^{i \beta}\ket{0} \\
& + \frac{1}{\sqrt{2}} \sin \alpha e^{2 i \beta}\ket{-1} , \\
\ket{\chi_+({\bf r})} & = \cos^2{\frac{\alpha}{2}} \ket{1} +
\frac{1}{\sqrt{2}} \sin \alpha e^{i \beta}\ket{0} \\
& + \sin^2{\frac{\alpha}{2}} e^{2 i \beta}\ket{-1},
\end{aligned}
\end{equation}
where the $\ket{m}$ are the Zeeman states. The associated eigenvalues are $-V({\bf r})$,
$0$ and $+V({\bf r})$, respectively.
Assuming that the magnetic field simply reads $(Gx,-Gy,B_z)$, one obtains
\begin{equation}
\begin{aligned}
\tan{\beta}&=-y/x=-\tan{\theta}\\
\tan{\alpha}&=\frac{G}{B_z}\sqrt{\tilde{x}^2+\tilde{y}^2}=\frac{G\ell_T}{B_z}\sqrt{x^2+y^2}=\rho\sqrt{x^2+y^2} .
\end{aligned}
\end{equation}
In the adiabatic limit,
i.e., $\rho\rightarrow 0$, $\alpha\rightarrow 0$, such that the trapping state
$\ket{\chi_+ ({\bf r})} \approx |m=1\rangle$, i.e., the atoms with a $J_z=1$ are
the ones which are trapped.
In harmonic units, $H_{\mathrm{eff}}$ acting on the vector $(\psi_+, \psi_0,\psi_-)$
reads formally as follows:
\begin{equation}
H_{\mathrm{eff}}=\left(
\begin{array}{ccc}
h_{++} & h_{+0} & h_{+-} \\
h_{0+} & h_{00} & h_{0-} \\
h_{-+} & h_{-0} & h_{--}
\end{array}
\right)+
\left(\begin{array}{ccc}
V({\bf r}) & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & -V({\bf r})
\end{array}
\right)
\end{equation}
The diagonal entries take the form
\begin{eqnarray}
h_{jj} = \frac{1}{2} ({\bf p}-{\bf A}_j)^2 + \Phi_j ,
\end{eqnarray}
where
\begin{equation}
\begin{aligned}
{\bf A}_{+}({\bf r})&= \left( 1- \frac{1}{\gamma} \right)\frac{{\bf e}_{\theta}}{r} , \\
{\bf A}_0({\bf r})&= \frac{{\bf e}_{\theta}}{r} , \\
{\bf A}_{-}({\bf r})&= \left( 1+\frac{1}{\gamma} \right)\frac{{\bf e}_{\theta}}{r} , \\
\Phi_{-}({\bf x})&=2\Phi_{0}({\bf x})=\Phi_+({\bf x})=\frac{\rho^2(1+\gamma^2)}{4 \gamma^4}
\end{aligned}
\end{equation}
with $r=\sqrt{x^2+y^2}$ and $\gamma = \gamma (r) = \sqrt{1+\rho^2r^2}$.
The off-diagonal entries of $H_{\mathrm{eff}}$ are
\begin{equation}
\begin{aligned}
h_{+0} &=-\frac{1}{2}\left[ \frac{\sqrt{2} \rho}{r \gamma^2}
\left(- r \frac{\partial}{\partial r}-i \gamma \frac{\partial}{\partial \theta}\right)
-\frac{\sqrt{2} \rho}{r \gamma^4}(\gamma^3-\gamma^2+1)\right] , \\
h_{+-} & = \frac{r^2 \rho^4}{4 \gamma^4} , \\
h_{0+} & = -\frac{1}{2} \left[ \frac{\sqrt{2} \rho}{r \gamma^2}
\left( r \frac{\partial}{\partial r}-i \gamma \frac{\partial}{\partial \theta}\right)
-\frac{\sqrt{2} \rho}{r\gamma^4}(\gamma^3-1)\right] , \\
h_{0-} &=- \frac{1}{2}\left[ -\frac{\sqrt{2} \rho}{r \gamma^2}
\left( r \frac{\partial}{\partial r}+i \gamma \frac{\partial}{\partial \theta}\right)
-\frac{\sqrt{2} \rho}{r \gamma^4}(\gamma^3+1)\right] , \\
h_{-+} & = \frac{r^2 \rho^4}{4 \gamma^4} , \\
h_{-0} &=-\frac{1}{2}\left[ \frac{\sqrt{2} \rho}{r \gamma^2}
\left( r \frac{\partial}{\partial r}-i \gamma \frac{\partial}{\partial \theta}\right)
-\frac{\sqrt{2} \rho}{r \gamma^4}(\gamma^3+\gamma^2-1)\right] .\\
\end{aligned}
\end{equation}
$H_{\mathrm{eff}}$ is Hermitian with respect to the scalar product $\langle f(r,\theta)\ket{g(r,\theta)} =
\iint rdr\,d\theta\, f^*(r,\theta)g(r,\theta)$.
\subsection{Numerical implementation}
The Hamiltonian $H_{\mathrm{eff}}$ is invariant under spatial rotations, therefore the eigenstates
can be written as $e^{im\theta}(\psi_+(r), \psi_0(r),\psi_-(r))$. For each value of $m$, the resulting
Hamiltonian $H_m$ only depends on the radial coordinate $r$. Since the value of the functions
$\psi_j(r)$ at $r=0$ is not fixed by any boundary condition, one uses a discretization scheme that
does not contain the point $r=0$, namely the grid points are $r_n=(n+1/2)\Delta r$, for $n$
ranging from $0$ to a maximum value $N$. $1/\Delta r$ fixes the number of grid points per
harmonic length, whereas $N\Delta r$ corresponds to the size of the system in harmonic length
units. The Hamiltonian $H_m$ is Hermitian for the scalar product $\langle F|G\rangle =
\int r dr \left(f_+^*(r)g_+(r)+f_0^*(r)g_0(r)+f_-^*(r)g_-(r)\right)$, where $F=(f_+(r), f_0(r),f_-(r))$,
such that it is the discretization of the equation $rH_m{\bf \psi}=Er{\bf \psi}$ that leads to
a generalized eigenvalue problem $AX=EBX$, where $A$ and $B$ are $(3N+3)\times(3N+3)$
Hermitian matrices, $B$ being positive definite.
Neglecting the off-diagonal coupling, the effective potentials seen by $\psi_0$ and $\psi_-$
components being $0$ and $-V({\bf r})$, the corresponding eigenstates are not bound states
but scattering states, whereas they are bound states for the $\psi_+$ component. The off-diagonal
coupling allows this bound states to decay to the scattering channels, which, from a mathematical
point of view, becomes resonances, i.e., complex poles of the Green function $G(z)=1/(z-H)$.
The complex rotation method is appropriate to compute directly the properties of these
resonances (energy, width). Its properties rely on mathematical properties of the analytic
continuation of the Green's function in the complex plane \cite{Balslev71,Ho83}. A review of
its application to atomic physics can be found in \cite{Buchl94}.
The method is implemented in our case by making the radial coordinate complex:
$r\rightarrow e^{i\phi}r$ and $\frac{\partial}{\partial r} \rightarrow
e^{-i\phi}\frac{\partial}{\partial r}$, where $\phi$ is a real parameter (the rotation angle).
The matrix representations of the Hamiltonian then become complex symmetric,
but are no more Hermitian. The fundamental properties of the complex spectra are :
\begin{itemize}
\item The bound states, if any, are still on the real axis.
\item The continua are rotated by an angle $2\phi$ on the lower-half complex
plane, around their branching point.
\item Each complex eigenvalue $E_j$ gives the properties of one resonance, i.e.,
the energy is the real part of $E_j$, and the width is two times the negative
of the imaginary part. The complex eigenvalues are independent of $\phi$,
provided that they are not covered by the continua.
\end{itemize}
Note that in the present case, the branching point associated to $\psi_0$ is located at
$E=0$, whereas the one associated to $\psi_-$ is formally at $E=-\infty$, since
$-V(r)\approx -r/\rho$ for large $r$. The scattering states for a linear potential oscillate faster
for large distance, such that they cannot be accurately described within a discretization
scheme. On the other hand, the overlap with the bound states at very large distance is
negligible, so that the exact behavior barely impacts the position and the width of the
resonance. Therefore to avoid numerical artifacts and increase the numerical accuracy,
we replace the antitrapping potential for $\psi_-$ by
\begin{equation}
V_{\mathrm{scatt}}(r)=\frac{1}{\rho^2}\sqrt{1+\rho^2\frac{r^2}{1+h^2r^2}},
\end{equation}
where $h$ is a small parameter, typically $h\approx10^{-2}$. For $hr\ll1$, then
$V_{\mathrm{scatt}}(r)\approx V(r)$, whereas for $hr\gg 1$, then $V_{\mathrm{scatt}} (r)
\approx 1/h\rho$, such that the scattering states have a well defined wavelength for large
$r$ values. We have numerically checked that the results presented here are insensitive to
the actual $h$ value.
\subsection{Results}
The properties of the resonances (position in energy and width) are displayed, as functions of
$\omega_T/\omega_L$, by the Fig.~\ref{m0} for the $m=0$ states and by the Fig.~\ref{m1}
for the $m=\pm1$ states. In both cases, the zero of energy corresponds to the bottom of the
trapping state, i.e., corresponding to a global shift of $-\frac{1}{\rho^2}$ of the eigenergies
of $H_{\mathrm{eff}}$, such that in the adiabatic limit $\rho\rightarrow0$, the energies directly
correspond to the harmonic oscillator levels. This is clearly seen on the left part of the top plots.
On the contrary, for $\rho\approx 1$ (right part), the energy levels differ from the harmonic
one, in particular, the difference in energy $E_{n+1}-E_{n}$ gets smaller with higher $n$ reflecting
the linear behavior of the trapping potential $V(r)$ at large $r$. The behavior in the adiabatic
regime can be obtained from the perturbation expansion of $h_{++}$ with respect to ${\bf A}_{+}$
and $\phi_+$. More precisely, for a fixed value of $m$, $h_{++}$ reads:
\begin{equation}
h_{++}=h_{++}^{(0)}-\frac{m}{r^2}\left(1-\frac{1}{\gamma}\right)+
\frac{1}{4r^2}\left(\frac{3\gamma^4-4\gamma^3-2\gamma^2+1}{\gamma^4}\right).
\end{equation}
The second term arises from the gauge field ${\bf A}_{+}$ and results in a energy
splitting among the $\pm m$ states, which is clearly observed in Fig.~\ref{m1}. A Taylor expansion
of these two terms at small distances $\rho r\ll 1$ leads respectively to:
\begin{equation}
\begin{aligned}
-\frac{m}{r^2}\left(1-\frac{1}{\gamma}\right)&\approx-m\left(\frac{\rho^2}{2}-\frac{3}{8}\rho^4r^2\right)\\
\frac{1}{4r^2}\left(\frac{3\gamma^4-4\gamma^3-2\gamma^2+1}{\gamma^4}\right)&\approx\frac{\rho^2}{2}-\frac{5}{8}\rho^4r^2
\end{aligned}
\end{equation}
The two terms proportional to $\rho^2$ correspond to an energy shift whereas the two terms
proportional to $\rho^4r^2$ correspond to modification of the trap frequency. However, the
preceding formula cannot be compared directly to the numerical results since at large distances
$\rho r\gg 1$ , the two last terms in $h_{++}$ leads to an effective centrifugal potential
\begin{equation}
(-m+\frac{3}{4})\frac{1}{r^2},
\end{equation}
which is independent of $\rho$. This shows that, although ${\bf A}_{+}$ and $\phi_+$ are
small perturbations to the $h_{00}$, the resulting energy shift has to be computed using their
exact expressions, not just their Taylor expansion around $r=0$.
\begin{figure}[thb]
\centerline{\includegraphics[width=7cm]{S_1_m_0.pdf}}
\caption{(color online) Properties of the $m=0$ resonances of the Hamiltonian
$H_{\mathrm{eff}}$: position in energy (top plot) and decay rates (bottom plot) as functions
of $\omega_T/\omega_L$. A given symbol and color correspond to the same state for the
two plots.}
\label{m0}
\end{figure}
\begin{figure}[thb]
\centerline{\includegraphics[width=6cm]{S_1_m_1m1.pdf}}
\caption{(color online) Properties of the $m=\pm1$ resonances of the Hamiltonian
$H_{\mathrm{eff}}$: positions in energy as functions of $\omega_T/\omega_L$ (top plot)
and decay rates as functions of $\omega_L/\omega_T$ (bottom plot).}
\label{m1}
\end{figure}
Furthermore, in the adiabatic regime, the decay rates can be obtained from the Fermi golden
rule. Assuming that one can approximate the scattering states as plane waves $\psi_0(r)
\approx e^{ikr}$, i.e., neglecting the fact that $r$ is a radial coordinate, the decay rates are
proportional to the modulus square of the Fourier transform $| \psi^{(n)}_{+} (k) |^2$ of the
harmonic trap wave functions, taken at the $k$ corresponding to the energy of the bound state,
i.e., such that $\frac{k^2}{2}=E_n+\frac{1}{\rho^2}$. At first order the decay is dominated by
the Gaussian decay of the wave function $e^{-r^2/2}$, such that one has:
\begin{equation}
\Gamma\approx e^{-k^2}=e^{-2E_n-2/\rho^2}.
\end{equation}
Therefore, one predicts a linear behavior of $\ln \Gamma$ as functions of $\rho^{-2} =
\omega_L/\omega_T$. This can be seen for both $m=0$ and $m=\pm 1$ states
(Figs. \ref{m0} and \ref{m1}, bottom plots). The behavior, at fixed $\rho$, is slightly more complicated since one has to
take into account the exact shape of the harmonic functions, in particular to explain that for
small $\rho$, the decay rates increase with the principal quantum number. We have checked on a simpler model that it is indeed a generic behavior: for a fixed value
of $\rho$, the decay rates depict a maximum around an energy (which depends on $\rho$),
see Fig.~\ref{resonances}.
\begin{figure}[thb]
\centerline{\includegraphics[width=7cm]{resonances.pdf}}
\caption{(colour online) Resonances in the complex energy plane for a fixed value
of $\rho^2=\frac{\omega_T}{\omega_L}=0.9$. As one can see, the decay rates are not simply decreasing with increasing energy, but rather exhibit a maximum around an energy value that depends on $\rho$.}
\label{resonances}
\end{figure}
From the numerical point of view, Table~\ref{exptab} summarizes the expected decay rate
of the $m=0$ ground state and the energy splitting between the first two $m=\pm1$ states
for few values of $\omega_T/\omega_L$. From the experimental point of view, for Rb$^{87}$ Fig.~\ref{resonance}
displays the life-time $2\pi/\Gamma$ (solid black line) and the gradient $G$ (red dashed line)
as functions of the bias field along the $z$ axis, for the four different values of
$\omega_T/\omega_L$ depicted in Table~\ref{exptab}. For instance, for a value of
$\omega_T/\omega_L = 0.2$ (top-left plot), for a bias field value of $0.05$ Gauss, the
life-time is $\approx70$ms, whereas $G$ is $\approx25$T/m. For these values, the trap
frequency is $\approx14.5$kHz. The energy splitting between the $\pm 1$ states is $2.2$kHz.
The splitting of the energy levels can be measured using RF-spectroscopy on the magnetic
trap for a cold thermal cloud \cite{Martin1988}.
\begin{table}[!ht]
\begin{ruledtabular}
\begin{tabular}{|c|c|c|c|c|}
$\omega_T/\omega_L$ & 0.2 & 0.25 & 0.31 & 0.4 \\
\hline
$\Gamma_{\text{num}}\times10^{3}$ & $1.0$ & $4.8$ & $16.0$ & $41.8$\\
\hline
$E^-_{\text{num}}-E^+_{\text{num}}$ & 0.15 & 0.18 & 0.21 & 0.25\\
\end{tabular}
\end{ruledtabular}
\caption{Decay rate of the $m=0$ ground state and the energy splitting between the
first two $m=\pm1$ states for four values of $\omega_T/\omega_L$.}
\label{exptab}
\end{table}
\begin{figure}[H]
\centerline{\includegraphics[width=9.1cm]{exp.pdf}}
\caption{(color online) Lifetime $2\pi/\Gamma$ (solid black line, left axis) and the gradient
$G$ (red dashed line, right axis) as functions of the bias field along the $z$ axis, for the four
different values of $\omega_T/\omega_L$ depicted in the table~\ref{exptab}. For instance,
for a value of $\omega_T/\omega_L=0.2$ (top-left plot), for a bias field value of $0.05$ Gauss,
the life-time is $\approx70$ms, whereas $G$ is $\approx25$T/m. For this value, the trap
frequency is $\approx14.5$kHz. The energy splitting between the $\pm 1$ states is
$2.2$kHz.}
\label{resonance}
\end{figure}
\section{Conclusions}
A tight magnetic trap allows you to investigate the break-down of the Born-Oppenheimer
condition. The adiabatic corrections will appear as reduced trap frequencies. The Born-Huang
potential counteracts the adiabatic potential. Losses will appear, but remain small. Situation is
experimentally tolerable. Molecular Aharonov-Bohm effects and geometric phases appear
because of the effective magnetic field associated with the vector potential.
Using experimentally accessible quantities (reasonable $B_z$), Majorana losses
could be measured experimentally. Splitting between $m=\pm 1$ is due to Berry connection
and Born-Huang terms. As a future work dynamical properties for a wave packet (shift of the
trap center) and impact of interactions can be studied.
\section*{Acknowledgments}
The Centre for Quantum Technologies is a Research Centre of Excellence funded by the
Ministry of Education and National Research Foundation of Singapore. E. S. acknowledges financial support from the Swedish
Research Council (VR) through Grant No. D0413201.
|
\section{Introduction}
Multilevel Monte Carlo (MLMC) \citep{Giles} is a technique that has
gained significant popularity over the past decade. It is designed to
produce statistical estimators for discretized random variables at
significantly lower computational costs than their Monte Carlo
counterparts for a fixed error. This is done by using a hierarchy of larger ensembles
using lower accuracy models, and smaller ensembles using higher
accuracy models. For probabilistic forecasting, one can use this
multilevel technique to estimate statistics from a forecast
probability distribution, given some distribution of the initial conditions
and/or random forcing.
In the multilevel Monte Carlo framework,
one usually considers a particular statistic, such as evaluations of the cumulative distribution function (CDF) \citep{Ritter, Wilson, Elfverson}, probability density function (PDF) \citep{Bierig} or expected values \citep{Giles,Cliffe}, selecting the ensemble sizes / finest level of resolution so that the overall multilevel estimator produces an efficient and
accurate approximation.
In the case of ensemble forecasting, one
usually wishes to compute many statistics from the same ensemble.
These approximations can be assessed using suitable verification
techniques. Verification tools used within ensemble forecasting
usually work alongside observations of the process that one is
interested in forecasting and can help verify properties from
calibration to the sharpness of a forecast \citep{Gneiting07}.
Given the multilevel hierarchy of ensembles from different resolutions
that form MLMC estimates of statistics, we would also like to evaluate
/ verify these ensembles in the same way; this is the subject of this
paper. We propose a methodology to take observables of a univariate random variable, or scalar observables of a multidimensional random variable (such as a random field evaluated at a
point in space), from a
multilevel hierarchy of ensembles with varying resolutions and generate an accompanying single
ensemble forecast. Most of the standard techniques in the field of ensemble forecasting are limited to the univariate case; in the context of large dimensional models in weather and climate these are usually applied to scalar observables such as point values or integral quantities.
This single forecast is statistically consistent
with the multilevel estimate. It can then
be used to verify the forecast from the original ensemble
hierarchy using standard methods such as calibration tests.
An alternative approach to this could be approximating each verification or scoring measure, such as the calibration or sharpness, individually and directly from the multilevel hierarchy of ensembles. For example, one could use a MLMC approximation for the CDF \citep{Ritter, Elfverson} to help compute a rank histogram to evaluate the forecast calibration.
Each different MLMC approximation typically comes with a framework to implement it, such as
a smoothing scheme in the former of those two studies.
However, by using the proposed methodology in this paper, one does not need a different multilevel approximation and framework for each individual scoring measure; instead any standard verification technique, such as the calibration or continuous ranked probability score (CRPS) \citep{Gneiting07}, can be employed on this standard single ensemble forecast.
To generate this ensemble forecast, inverse transform sampling is used. The new single ensemble forecast preserves the unbiased approximation to the mean of the forecast distribution from the original multilevel estimator and forms consistent approximations to other statistics, such as higher moments.
This study proceeds as follows; an introduction to MLMC will be given in Section 2, then a simple method to find a consistent ensemble forecast from a MLMC approximation will be given in Section 3 alongside a corresponding verification technique for these ensemble forecasts. Finally, a conclusion follows.
\section{Multilevel Monte Carlo}
Multilevel Monte Carlo \citep{Giles} is primarily used as a computationally cheap alternative to an equivalent accuracy single level Monte Carlo estimator of statistics with respect to a probability distribution. Suppose one wishes to compute estimates to statistics of $f(X_{L,t})$, such as $\mathbb{E}[f(X_{L,t})]$,
where $X_{L,t}$ is a numerical approximation of our `forecast' random variable $X$
(with discretization parameter $h_{L} \propto M^{-L}$, $M>1$) at time $t \geq 0$ and $f$ is some scalar observable function. Let $X^{i}_{L,t}$,
$i=1,\ldots,N$, be $N \geq 1$ i.i.d. samples of the random variable
$X_{L,t}$. Then an empirical approximation to the density of $X_{L,t}$ is
\begin{equation}
\pi_{L,t}^{MC}(x)=\frac{1}{N}\sum^{N}_{i=1}\delta(x-X_{L,t}^{i}),
\label{equation:PDF}
\end{equation}
where $\delta$ is the Dirac delta function. One can then estimate statistics to this empirical distribution via the Monte Carlo method. For example, the standard estimator for $\mathbb{E}[f(X_{L,t})]$ is given by
\begin{equation}
\bar{f}_{L,t}^{MC}=\frac{1}{N}\sum^{N}_{i=1}f(X_{L,t}^{i}).
\label{equation:MonteCarlo}
\end{equation}
Now consider the multilevel framework, using $L+1$ ensembles $\{X_{l-1,t}^{i},X_{l,t}^{i}\}_{l=0,i=1}^{l=L,i=N_l}$ (with $X_{-1}^{i}=0$) of sizes $N_{l}$, to derive the equivalent MLMC approximation to $\mathbb{E}[f(X_{L,t})]$,
\begin{equation}
\bar{f}_{L,t}=\frac{1}{N_{0}}\sum^{N_{0}}_{i=1}f(X_{0,t}^{i})+\sum^{L}_{l=1}\Big(\frac{1}{N_{l}}\sum^{N_{l}}_{i=1}(f(X_{l,t}^{i})-f(X_{l-1,t}^{i}))\Big).
\label{equation:MLestimator}
\end{equation}
Taking the telescoping sum of expectations,
\begin{equation}
\mathbb{E}[f(X_{L,t})]=\mathbb{E}[f(X_{0,t})]+\sum^{L}_{l=1}\mathbb{E}[f(X_{l,t})]-\mathbb{E}[f(X_{l-1,t})],
\label{equation:linearityexpectation}
\end{equation}
and considering
\begin{equation}
\mathbb{E}[\hat{f}_{l,t}]=
\begin{cases}
\mathbb{E}[f(X_{0,t})],& l=0, \\
\mathbb{E}[f(X_{l,t})]-\mathbb{E}[f(X_{l-1,t})],& l>0 ,
\end{cases}
\label{equation:unbiased}
\end{equation}
where
\begin{equation}
\hat{f}_{l,t}=
\begin{cases}
\sum^{N_{0}}_{i=1}\frac{1}{N_{0}}f(X_{0,t}^{i}),& l=0, \\
\sum^{N_{l}}_{i=1}\frac{1}{N_{l}}\big(f(X_{l,t}^{i})-f(X_{l-1,t}^{i})\big),& l>0,
\end{cases}
\label{equation:MLMC}
\end{equation}
one recovers $\bar{f}_{L,t}$ as an unbiased approximation of $\mathbb{E}[f(X_{L,t})]$.
The important thing to note here is that the fine (level $l$) and coarse
(level $l-1$) samples in
the difference estimators, $\hat{f}_{l,t}$, must be positively
correlated for each $i$. This can be achieved by using the same random system input
(e.g. initial conditions/stochastic forcing) for each $i$ on both
levels. On the other hand, the samples in different ensembles must be uncorrelated.
The uses of the above framework are incredibly varied. One can even condition these multilevel estimators on observations using processes such as filtering \citep{Jasra, Gregory, Gregory16}. In addition to this, there have been many other applications of MLMC, some of which are highlighted in the review \cite{GilesReview}.
Given an optimal choice of $L$ and $N_{l}$, one can compute these estimators, with the same accuracy as their standard Monte Carlo counterparts, for significantly less computational expense. This works by noting that due to the correlation between the pairs of
samples in each difference estimator, the sample variance of $f(X_{l,t})-f(X_{l-1,t})$, denoted $V_{l}$, should decrease asymptotically with $l \to \infty$. If one desires the accuracy of
$\bar{f}_{L,t}$ to be
\begin{equation}
\mathbb{E}\left(\left(\bar{f}_{L,t}-\mathbb{E}[f_{L,t}]\right)^{2}\right) < \epsilon^{2},
\end{equation}
then one can follow the algorithm in \cite{Giles} to compute $\bar{f}_{L,t}$ by updating, on-line
(as you add additional samples),
the optimal sample sizes
\begin{equation}
N_{l}=\ceil*{2\epsilon^{-2}\left(V_{l}h_{l}\right)\left(\sum_{n=0}^{L}\sqrt{V_{n}/h_{n}}\right)},
\end{equation}
whilst increasing $L$ until $\abs*{\hat{f}_{L,t}}<\frac{1}{\sqrt{2}}(M-1)\epsilon$. An estimated $V_{l}$ can be used in the optimal sample size formula.
Computational cost reductions occur because, if $V_{l}$ decreases asymptotically with $l \to \infty$, then $N_{l}$ also does, leading to a trade-off between estimator variance and bias in each difference estimator. To conclude, we should have large ensembles for the lower levels, and smaller ensembles on the higher levels, given by asymptotically decreasing values of $V_{l}$.
For the full algorithm and corresponding theory, see \cite{Giles, GilesReview}.
\section{Ensemble Forecasting}
This paper now proposes a method to generate a single ensemble forecast from the hierarchy of ensembles created from the MLMC method. Put simply, one can generate a large ensemble (much larger than the finest level ensemble) that represents the entire MLMC approximation to the forecast distribution. This is more useful for the verification of the hierarchy of ensembles rather than simply using standard verification techniques on the finest ensemble in this hierarchy. As mentioned in the previous section, the sample sizes, $N_{l}$, for the pairs of ensembles
on all levels decrease asymptotically, and thus the finest ensemble is the smallest ensemble in the hierarchy. Using the finest ensemble for the verification of the entire MLMC approximation of the forecast distribution would neglect the majority of samples, on lower levels, from which the approximation was composed.
In addition to this verification, given the statistical consistency of this ensemble forecast with the multilevel ensemble hierarchy, many statistics can be easily estimated via this ensemble.
\subsection{Multilevel Monte Carlo Ensemble Forecasts}
Now assume $X_{l,t} \in \mathbb{R}$, and so if $X_{l,t}$ was multivariate in the section before, $X_{l,t}$ now represents $f(X_{l,t})$ , the scalar observable, such as $X_{l,t}$ evaluated at a point in space. Here we describe how to generate the single ensemble forecast of a scalar observable $X_{F,t}^{i}\in \mathbb{R}$, $i=1,...,N$ from the MLMC hierarchy of ensembles through inverse transform sampling. It is important to note that this ensemble does not contain i.i.d samples from the forecast distribution, instead they will simply be approximations to these samples. However, this single ensemble has the properties to form a consistent empirical estimate to the forecast distribution and associated distribution functions.
From here onwards, we will assume that values of $N_{l}$ and $L$ have been either set or found, and that the
hierarchy of ensembles $\{X_{l-1,t}^{i},X_{l,t}^{i}\}_{l=0,i=1}^{l=L,i=N_l}$, with $X_{-1,t}=0$, has been
generated.
Predominantly, this is because the framework that this paper presents is designed for evaluating any given MLMC
approximation. Each approximation has a hierarchy of ensembles that use values of $N_{l}$ and $L$ that have been optimised around
minimising the cost of that particular approximation. Each approximation typically comes with it's own algorithm to
set-up these values. Thus, by making the aforementioned assumption we can keep this framework general to all approximations.
In addition to this, it is likely that in real forecasting practice
one would pick the desired maximum level $L$ and then set fixed values of $N_{l}$ based on the maximum computational
expense one can use on a particular level.
This way of choosing $N_{l}$ and $L$ is implemented in the numerical example later in the paper.
Inverse transform sampling is the process of evaluating an (approximation to the) inverse CDF, $F^{-1}(u)$, $u \in [0,1]$, also known as the \textit{quantile function}. In the case where the CDF, $F$, of a random variable is strictly increasing and absolutely continuous, there exists a unique value $x \equiv F^{-1}(u)$ for which $F(x)=u$. This distribution must usually be estimated empirically. If the true CDF of the forecast distribution is known to be absolutely continuous and the samples are sorted to form order statistics, then some of these estimates have been shown to be consistent approximations to $F^{-1}(u)$ \citep{Ma}. A very simple consistent estimate for an evaluation to the quantile function, of the distribution with CDF, $F$, using the (ascending) sorted samples $\big\{X^{i}\big\}_{i=1,...,N} \sim F$, $X^{1} < X^{2} < .... < X^{N}$ is,
\begin{equation}
\hat{F}^{-1}(u)=X^{\ceil{N \times u}}.
\label{equation:SampleQuantileFunction}
\end{equation}
Here, the estimate is a consistent one in the sense that it converges in probability to $F^{-1}(u)$ as $N \to \infty$.
One can use linear interpolation and extrapolation to smooth this consistent estimate. Other inconsistent techniques include fitting a parametric distribution to the ensemble, such as a Gaussian, and sampling from a closed form quantile function (e.g. $\Phi$ for a Gaussian distribution) for that distribution. In all cases, when the empirical quantile function is evaluated with i.i.d uniform samples $u \in [0,1]$, approximations to samples of $X$ can be generated.
The use of inverse transform sampling alongside MLMC was first suggested in \cite{GilesBook}. Here it was proposed to be used to minimise the discrete Wasserstein distance between the two paired ensembles in each difference estimator within (\ref{equation:MLestimator}) and thus positively couple them. Instead, here we will use inverse transform sampling in the context of a MLMC approximation to the quantile function of the forecast distribution,
\begin{equation}
\bar{F}_{L,t}^{-1}(u) = R(X)_{0,t}^{\ceil{N_{0} \times u}}+\sum^{L}_{l=1}\big(R(X)_{l,t}^{\ceil{N_{l} \times u}}-R(X)_{l-1,t}^{\ceil{N_{l} \times u}}\big),
\label{equation:MLestimatorquantile}
\end{equation}
where $R(X)^{i}_{l}$ is the $i'th$ order statistic of $X_{l}$, so that $R(X)^{1}_{l}<R(X)^{2}_{l}<...<R(X)^{N_{l}}_{l}$.
Note that there is not an exact cancellation in expected values of the above estimator terms, as in
the telescoping sum of expectations in (\ref{equation:linearityexpectation}), as
the individual approximations on each level are not unbiased, only consistent in the limit of $N_{l} \to \infty$.
The following algorithm demonstrates how to generate an ensemble $\big\{X_{F,t}^{i}\big\}_{i=1,...,N}$ of arbitrary size $N$, approximating samples of $X_{L,t}$.
\BEGINALG
\Procedure{}{}
\For{$l=0,...,L$}
\If{$l=0$}
\State Sort $X_{0,t}^{j}$, $j=1,...,N_{0}$, so that $R(X)_{0,t}^{1}<R(X)_{0,t}^{2}<...<R(X)_{0,t}^{N_{0}}$
\Else
\State Sort $X_{l,t}^{j}$, $X_{l-1,t}^{j}$, $j=1,...,N_{l}$, so that $R(X)_{l,t}^{1}<R(X)_{l,t}^{2}<...<R(X)_{l,t}^{N_{l}}$ and $R(X)_{l-1,t}^{1}<R(X)_{l-1,t}^{2}<...<R(X)_{l-1,t}^{N_{l}}$
\EndIf
\EndFor
\For{$i=1,...,N$}
\State Set $X_{F,t}^{i}=0$
\State Sample $u^{i} \sim U[0,1]$
\For{$l=0,...,L$}
\If{$l=0$}
\State $X_{F,t}^{i}+=R(X)_{0,t}^{\ceil{N_{0} \times u^{i}}}$
\Else
\State $X_{F,t}^{i}+=R(X)_{l,t}^{\ceil{N_{l} \times u^{i}}}-R(X)_{l-1,t}^{\ceil{N_{l}\times u^{i}}}$
\EndIf
\EndFor
\EndFor
\EndProcedure
\end{algorithmic}
Note that these $X_{F,t}^{i}$ are not samples from $X_{L,t}$, they are only consistent approximations to the evaluations of $F^{-1}_{L,t}(u)$ for a particular $u$. More specifically, for a random uniform sample $u \sim U[0,1]$, we have
\begin{equation}
x=\bar{F}^{-1}_{L,t}(u),
\end{equation}
and as $N_{l} \to \infty$ for all $l$,
\begin{equation}
x \xrightarrow{p} F^{-1}_{L,t}(u),
\end{equation}
where $N_{l}$ are the number of samples used in each difference estimator in (\ref{equation:MLestimatorquantile}). Then in this limit, $x$ converges in probability to a sample from the forecast distribution on the finest level, i.e. $x \sim X_{L,t}$. Therefore any statistical estimate using these samples is a consistent one within this limit.
The single ensemble $\{X_{F,t}^{i}\}_{i=1,...,N}$ can form valid and consistent approximations to statistics of the forecast distribution. For example, the empirical, consistent, CDF of this ensemble forecast found from the MLMC approximation to the forecast distribution is,
\begin{equation}
\hat{F}_{X_{F,t}}(x)=\frac{1}{N}\sum^{N}_{i=1}\mathbb{I}_{X_{F,t}^{i} \leq x}
\end{equation}
where $\mathbb{I}$ is the indicator function. Clearly this is non-decreasing for continuous $X_{F,t}$ and has the support of $[0,1]$.
One assumes that in practice the computational effort of evaluating the above function a large number of times to generate the ensemble $\big\{X_{F,t}^{i}\big\}_{i=1,...,N}$ is negligible in comparison to the expense of generating the original samples on all of the different levels. Thus, the method seems likely to be admissible even when $N$ is much larger than $N_{0}$. Having said this, it makes sense here to set $N \propto N_{0}$ so that both aspects of the approximation (inverse CDF estimator and the ensemble forecast) converge in probability simultaneously. We take $N=\alpha N_{0}$ with $\alpha \in \mathbb{Z}$, $\alpha \geq 1$ for simplicity.
The proposed ensemble forecast also preserves the unbiasedness of the approximation to the first moment of the forecast distribution from the original MLMC approximation. To show this let $\bar{X}_{F,t}=\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1} X_{F,t}^{i}$ be the sample mean of the ensemble forecast from the multilevel hierarchy of ensembles. Then,
\begin{equation}
\begin{split}
\bar{X}_{F,t}&=\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1} X_{F,t}^{i}\\
\quad &=\Big(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1}\hat{F}^{-1}_{0,t}(u^{i})\Big) +\\
\qquad &\sum^{L}_{l=1}\bigg(\Big(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1}\hat{F}^{-1}_{l,t}(u^{i})\Big) -\Big(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1}\hat{F}^{-1}_{l-1,t}(u^{i})\Big)\bigg) , \\
\quad &=\Big(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1}X_{0, t}^{\ceil*{N_{0} \times u^{i}}}\Big) +\\
\quad &\sum^{L}_{l=1}\bigg(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1}\Big(X_{l, t}^{\ceil*{N_{l} \times u^{i}}} - X_{l-1, t}^{\ceil*{N_{l} \times u^{i}}}\Big)\bigg),
\end{split}
\end{equation}
and given that $u^{i}$ are i.i.d. draws of the uniform distribution $Unif[0,1]$, $i=1,...,\alpha N_{0}$ then
\begin{equation}
\begin{split}
\mathbb{E}[\bar{X}_{F,t}]&=\Big(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1} \mathbb{E}[X_{0,t}]\Big) +\\
\qquad &\sum^{L}_{l=1}\bigg(\frac{1}{\alpha N_{0}}\sum^{\alpha N_{0}}_{i=1}\Big(\mathbb{E}[X_{l,t}] - \mathbb{E}[X_{l-1,t}]\Big) \bigg)\\
\quad &=\mathbb{E}[X_{0,t}]+\sum^{L}_{l=1}\mathbb{E}[X_{l,t}-X_{l-1,t}]=\mathbb{E}[X_{L,t}].
\end{split}
\end{equation}
\subsection{Assessing the Calibration of Multilevel Monte Carlo Ensemble Forecasts}
Evaluating the ensembles used in ensemble forecasts is very important in checking the predictive value of the forecast. The remainder of this paper concentrates on a method of evaluating the calibration of forecasts from the MLMC approximations to the forecast distribution, directly via the single ensemble forecast found in the previous section: the Probability Integral Transform Histogram. This technique uses observations from the target distribution to evaluate ensemble forecasts. Calibration is the measure of whether the observations are indistinguisable from the samples of the ensemble forecast distribution \citep{Carney}. This is a quality of the empirical forecast distribution that is possibly disregarded if one were to simply study errors of point statistical estimators.
Consider the target distribution, $Y_{obs,t_{k}}$, behind the observed process, where partial observations $y_{obs,t_{k}}$, are taken from a single realisation of this process at times $t_{k}$, $k \in [0,N_{y}]$, $t_{0}=0$, $t_{N_{y}}=T$. Clearly, our aim would be to use a forecast distribution associated with the random variable $X_{t_{k}}=Y_{obs,t_{k}}$, however in many real-world scenarios, $Y_{obs,t_{k}}$ is unknown. Therefore verification techniques are used to rank forecasts on their similarity to the observed process, with the aim of finding the best forecast / model that derived them. The case of $X_{t_{k}}=Y_{obs,t_{k}}$ is known as the random variable with associated forecast distribution from the perfect model.
\subsubsection{Probability Integral Transform Histogram}
The Probability Integral Transform (PIT) histogram is used to determine the uniformity of the observations with respect to the (empirical) CDF of the ensemble, and thus the calibration of the forecast distribution with respect to the target distribution. One can define a random variable $R \sim F_{L,t_{k}}(Y_{obs,t_{k}})$, the Probability Integral Transform. Then samples of $R$ are given by,
\begin{equation}
r_{t_{k}}=F_{L,t_{k}}(y_{obs,t_{k}}).
\end{equation}
The forecast distribution is said to be calibrated with respect to the target distribution if $R \sim Unif[0,1]$, and so a histogram of $r_{t_{k}}$ would be relatively flat. Using the MLMC approximation to the forecast distribution, define the associated multilevel empirical PIT samples,
\begin{equation}
\hat{r}_{t_{k}}=\frac{1}{N}\sum^{N}_{i=1}\mathbb{I}_{X_{F,t_{k}}^{i} \leq y_{obs,t_{k}}}
\end{equation}
where $X_{F,t_{k}}^{i}$ are an arbitrary $N$ members of the ensemble forecast from the multilevel hierarchy of ensembles at time $t_{k}$ using the aforementioned inverse transform sampling method. This is simply the empirical cumulative distribution function of the $N$ ensemble forecast members $X_{F,t_{k}}^{i}$. Here, given that we set $N \propto N_{0}$, then in the limit of $N_{l} \to \infty$, for all $l=0,...,L$,
\begin{equation}
X_{F,t_{k}} \sim F_{L,t_{k}}^{-1},
\end{equation}
and thus $F_{L,t_{k}}(X_{F,t_{k}}) \sim U[0,1]$. By considering this, we have a consistent estimate to the PIT sample $r_{t_{k}}$, when concentrating on the limit of $N \propto N_{0} \to \infty$. One can find the frequency ($H_{i}$, $i=1,...,B$) of $B$ evenly spaced bins in a histogram of these samples by:
\bigskip
\begin{enumerate}
\item Set $H_{i}=0$, for $i=1,...,B$.
\item For each $k=1,...,N_{y}$, find the $i=1,...,B$ in which $\frac{i-1}{B} \leq \hat{r}_{t_{k}} \leq \frac{i}{B}$ and set $H_{i}=H_{i}+1$.
\end{enumerate}
\bigskip
This histogram will be refered to as the multilevel PIT histogram (MLPIT) for the remainder of this paper. The MLMC approximation that derives the ensemble forecast $\{X_{F,t_{k}}^{i}\}_{i=1,...,N}$ can then be described as calibrated with respect to the target distribution if $H_{i} \approx \frac{N_{y}}{B}$ for each $i=1,...,B$. Thus, this can be used to test the variance and biasedness of the ensembles with respect to the target distribution. If the histogram is convex then the ensembles are said to be overdispersed, whereas if it is concave, then the ensembles are said to be underdispersed, and if it is skewed then there exists a bias in the ensembles \citep{Carney}. This is therefore a very appropriate way to clarify if there is any additional bias from the cancellation of intermediate estimators in a MLMC approximation, thus negating the telescoping sum of expectations in (\ref{equation:linearityexpectation}), although this is not demonstrated in this paper.
\bigskip
\textit{\textbf{Example:}}
The following linear mean reverting OU process, $X_{t} \in \mathbb{R}$,
\begin{equation}
\label{eq:OU}
dX_{t}=\alpha(\mu-X_{t})dt + \sigma^{2}d W_{t} ,
\end{equation}
over time time interval $t \in [0,T]$, where $W_{t}$ is a univariate Brownian Motion, will be used alongside pre-defined scenarios of calibration for a MLMC approximation to the forecast distribution to provide a demonstration of the proposed method. We let the observations come from the above model, discretized with timestep $h=2^{-5}$, with $\alpha=0.1$, $\mu=0$ and $\sigma^{2}=0.1$.
In this example, an Euler-Maruyama numerical scheme will be used to discretize
the OU process.
To frame this problem in a likely forecasting setting, we first choose a fixed finest resolution
that we desire, $L=4$ and so $l \in [0, 4]$. A maximum computational expense that we are allowed to use on propagating the entirity of samples in each level of the ensemble hierarchy, $C_{max}=1.536 \times 10^{7}$, is then set. The cost of each sample in $l$'th difference estimator is $\left(Th_{l}^{-1}(1 + 1/2)\right)$ (as all but the first difference estimators in (\ref{equation:MLestimator}) require
coarse and fine time-steps of the discretization) where $h_{l}=2^{-1-l}$ and so $N_{l}$ is given by
\begin{equation}
\begin{split}
N_{l} &= \floor*{\frac{C_{max}}{T\left(h_{l}^{-1}\left(1 + 1/2\right)\right)}}\\
\quad &= \floor*{\left(\frac{2}{3}\right)C_{max}T^{-1}h_{l}}.
\end{split}
\end{equation}
This corresponds to $N_{0} = 2^{7}$.
The arbitrary number of samples $X_{F}^{i}$ to draw from the MLMC approximation to the inverse CDF is set to $N=8N_{0}=2^{10}$.
Pairs of samples from coarse and fine ensembles in each difference estimator in (\ref{equation:MLestimator}) are positively coupled by using the same underlying Brownian Motion, as in \cite{Giles}.
The models are run over times $t \in [0,40000]$ (the long run time is to give the stationary distributions a chance to be simulated), and observations are collected at $t_{k}=k$, $k \in [1,40000]$. At each of these times, a single ensemble forecast is generated from the hierarchy of ensembles that build up the MLMC approximation to the forecast distribution, and is used to verify the calibration of the approximation. Model parameters for four experimental setups are given as follows: $\alpha = 0.1$, $\sigma^{2}=0.1$, $\mu=0$ for the calibrated scenario, $\alpha = 0.1$, $\sigma^{2}=0.02$, $\mu=0$ for the underdispersed scenario, $\alpha = 0.1$, $\sigma^{2}=0.5$, $\mu=0$ for the overdispersed scenario and $\alpha = 0.4$, $\sigma^{2}=0.1$, $\mu=0.2$ for the biased scenario.
\iffalse
\begin{center}
\begin{tabular}{| c | c | c | c |}
\hline
Scenario Name & $\alpha$ & $\sigma^{2}$ & $\mu$ \\
\hline
Calibrated & 0.1 & 0.1 & 0\\
\hline
Underdispersed & 0.1 & 0.02 & 0 \\
\hline
Overdispersed & 0.1 & 0.5 & 0\\
\hline
Biased & 0.4 & 0.1 & 0.2\\
\hline
\end{tabular}
\end{center}
\fi
\begin{figure}[h!]
\centering
\includegraphics[width=94mm]{EnsembleForecastPIT}
\caption{\textit{Multilevel Probability Integral Transform Histograms, using the ensemble $\big\{X_{F}^{i}\big\}_{i=1,...,N}$, of the linear OU process for the four different calibration scenarios. The solid line on the Biased scenario plot shows a smoothed kernel of the PIT histogram generated from the actual stationary forecast and target distributions.}}
\label{figure:MLPIT_Experiments_LinearSystem}
\end{figure}
This setup allows us to establish that the correct calibration behaviour is being shown by the Multilevel PIT histogram for each of the scenarios, however we will also compare this to the PIT histogram using just the finest ensemble, although this isn't the primary goal of the section. Figures \ref{figure:MLPIT_Experiments_LinearSystem} and \ref{figure:PIT_Experiments_LinearSystem} show the MLPIT and PIT histograms respectively for the four scenarios of calibration listed above. Due to the small number of samples in the finest ensemble, the PIT histogram can only represent a very small number of bins of probability. Both show similar general behaviour for the cases above.
We can derive the stationary distribution to both the forecast distribution and the target distribution from the model specifications above from the Fokker-Planck equation corresponding to \eqref{eq:OU}. One notes that the stationary forecast distribution using the Biased scenario model above is given by $f \sim N\big(0.2,\frac{1}{8}\big)$ and the stationary target distribution is given by $y \sim N\big(0,\frac{1}{2}\big)$ (as the general form is $\sim N\big(\mu,\frac{\sigma^{2}}{2\alpha}\big)$). Thus the actual PIT histogram can be generated by taking an arbitrarily large number of samples of $F(y)$, where $F$ is the CDF of $f$. A smoothed density kernel of this histogram is superimposed on the corresponding empirical PIT histograms for the single level and multilevel approximations. The empirical histograms approximately match this, however, due to the lack of samples in the finest ensembles, the single level histogram is not as clear to the type or magnitude of bias as shown by the multilevel PIT histogram. This is due to the lack of probability bins in a small, single finest ensemble ($N_{L}+1$), and one would still suffer from similar problems if using interpolation techniques in between the limited number of samples of this ensemble. The MLMC approximations of the forecast distributions and associated histograms are numerically biased (proportional to the finest timestep) from this exact PIT histogram due to the use of a numerical discretization, and so are expected to be slightly different. Despite this, one can clearly interpret the calibration and identify the extent and type of such bias in the MLMC approximations to forecast distributions with more clarity using the multilevel PIT histogram technique proposed here than using standard methods with the small finest ensemble.
\begin{figure}[h!]
\centering
\includegraphics[width=94mm]{FinestEnsemblePIT}
\caption{\textit{Probability Integral Transform Histograms, using just the finest ensemble $\big\{X_{L}^{i}\big\}_{i=1,...,N_{L}}$, of the linear OU process for the four different calibration scenarios. The solid line on the Biased scenario plot shows a smoothed kernel of the PIT histogram generated from the actual stationary forecast and target distributions.}}
\label{figure:PIT_Experiments_LinearSystem}
\end{figure}
\section{Conclusion and Outlook}
This work has discussed the benefits of generating an ensemble forecast from Multilevel Monte Carlo (MLMC) approximations to statistics of random variables representing forecast distributions. The proposed procedure to do this is simple and easily implemented. The calibration of this ensemble forecast has also been examined. Ensemble forecasts provide a simple methodology of deriving empirical estimates to associated distribution functions. The ensemble hierarchy that forms the computationally efficient MLMC approximations to an arbitrary statistic of the forecast distribution is assumed to have already been generated in preparation for forecasting. It is anticipated that in real forecasting practice, this hierarchy of ensembles would simply be generated by using the maximum ensemble sizes affordable at each level of resolution. The ensemble forecast calibration verification technique takes the entire multilevel hierarchy into account when using the proposed methodology.
Calibration of this ensemble forecast is assessed using the Probability Integral Transform histogram after this single ensemble is generated from the ensemble hierarchy. Thus we have stated what it means for a MLMC approximation to be calibrated with respect to a target distribution. This can be used to evaluate many properties of a MLMC approximation to a forecast distribution including biases (and their type) from intermediate terms in the MLMC telescoping sum of estimators, variances of these approximation and even potentially distribution multimodal feature detection.
\ack Alastair Gregory was supported by the Science and Solutions to a Changing Planet DTP and the Natural Environment Research Council. He was also supported by the Mathematics of Planet Earth CDT. The authors of this paper would like to thank Chris Ferro (Exeter) for the range of helpful conversations and advice he gave in the development of this research. Thanks also goes to the reviewers of this paper for their constructive feedback on the manuscript.
|
\section{Introduction}
This paper introduces a new family of rapidly-evolving time series models, inspired by real data applications, and then develops the appropriate analysis tools for their computationally-efficient and consistent inference. Statistical models for time series observations are usually described by their expectations and covariance structure. Classic families of covariance structure correspond to stationary covariances, governed only by the temporal lags between observed values of the process. The assumption of stationarity greatly simplifies analysis, as it renders the covariance structure homogeneous across time and this motivates averaging for estimation. Unfortunately most often this homogeneous time structure is inadequate as a model for real-world applications, and does not reflect the variability of the observed time series.
In order to analyse nonstationary time series, using the framework of locally stationary time series is standard \citep{priestley1988non, Dahlhaus1997}. The idea is to allow for a time-varying spectral density. Parametric models for the time-varying spectral density can be fitted via the use of local Fourier transforms, usually requiring a spectral smoothness assumption. The concept of infill asymptotics developed by \citet{Dahlhaus1997} is based on the idea that a growing amount of data is obtained locally in time.
Normally, for nonstationary time series analysis, there is a bias-variance trade-off that occurs when selecting the length of an analysis window. Longer windows will decrease variance, but will simultaneously increase bias due to the variation of the covariance function over the analysis window \citep{adak1998time}. In our case we shall eliminate the bias, and this will enable us to use longer time window lengths.
For this purpose we exploit the notion of a modulated process~\citep{Parzen1963, priestley1965evol}. A modulated process is a latent stationary process multiplied pointwise by a modulating function. If we observe the modulating function, this framework allows us to define an averaged autocovariance function, despite the clear nonstationarity of the modulated process. This in turn allows us to introduce the Fourier transform of the averaged autocovariance function of the modulated process, which is equal to its expected periodogram. Through examining the expected periodogram, the properties controlling the latent random process may be inferred even when the modulating function changes very rapidly.
The standard class of modulated processes are asymptotically stationary modulated processes~\citep{Parzen1963, Iacobucci2003, Jiang2004}. Here the autocovariance of the modulating function converges to a fixed function, which is too restrictive for our real-world application. We introduce a more general class, which we call \emph{modulated processes with a significant correlation contribution}. This more flexible model will still allow us to infer the parameters of the driving process using likelihood-based methodologies. An alternative approach might be to simply divide the observed process by the known modulating sequence to recover the latent process, and then perform inference directly on the recovered latent process. However this is not possible in general, as the modulating function may contain zeros, or the observed process may in fact be an aggregation of different processes, as will be the case in the real-world application that motivated us to develop this model class.
Anticipating our application to oceanographic surface flow measurements, we present a novel generalization of modulated processes for isotropic bivariate processes, or equivalently proper complex-valued processes \citep{complexProcesses}. The wealth of possible structure in multivariate processes
is considerable in general. Inherent documented challenges in modelling include producing valid joint representations~\citep{tong1973some,tong1974time,priestley1973analysis}. This problem does not apply here as we shall modulate both processes under consideration simultaneously, thus automatically removing such problems.
Having set up our model of modulated processes with a significant correlation contribution, we show how a modified version of a frequency-domain likelihood allows us to consistently estimate parameters with a high degree of computational efficiency. More specifically, the Whittle likelihood for stationary Gaussian processes is an approximation of the exact likelihood that is consistent and can be computed in $\mathcal{O}(N\log N)$ elementary operations.
We adapt this pseudo-likelihood to our class of models, making use of the expected periodogram, and conserve the $\mathcal{O}_P\left(N^{-1/2}\right)$ convergence rate. We also conserve the $\mathcal{O}(N\log N)$ computational cost in the minimization procedure, except for a pre-computational step of $\mathcal{O}(N^2)$ operations, which must be performed only once per observed time series sample. Exact likelihood for nonstationary time series, on the other hand, will in general require more than $\mathcal{O}(N^2)$ operations, due to the need to manipulate large covariance matrices.
We apply this method to an important dataset measuring ocean currents. There are only a handful of observational platforms capable of providing continuous global coverage of the Earth's oceans and so it is critical that we fully utilize these datasets to advance our understanding of the oceans and their impact on climate.
One of these studies is the Global Drifter Program (GDP, www.aoml.noaa.gov/phod/dac), consisting of freely drifting instruments, or ``drifters" \citep{lumpkin07}.
Fig. 1(a) shows positions from multiple trajectories obtained from drifters at or near the equator. From the positions of the trajectories, we may also calculate the velocities of the instruments, and these velocity time series are useful measurements for understanding ocean dynamics.
Depending on the instrument it may not be reasonable to model the velocity time series as locally stationary, as is assumed in \citet{sykulski2016Lagrangian}.
In particular, for reasons to be discussed, regions near the equator are likelier to yield drifter trajectories with highly nonstationary velocities where locally stationary modelling breaks down. Instead, to capture such rapid time-variability, we use a modulated stochastic process from our class of nonstationary models. This model allows us to capture the rapid frequency modulation of oscillations known to geophysicists as ``inertial oscillations''. An example of a time series with such rapid frequency modulation can be seen in Fig. 1(c).
We organize the paper into the following sections. Section 2 reviews the model family of modulated processes, the standard assumption of asymptotic stationarity associated with such processes, and introduces our generalized class called modulated processes with a significant correlation contribution. This section also includes extensions to bivariate processes. Section 3 describes a computationally consistent pseudo-likelihood estimation procedure. In Section 4 we apply our methods to real-world oceanographic data and various numerical experiments; we also apply our methods to a simulated missing data problem.
We establish consistency of our proposed procedure in Section 5, under the assumption of significant correlation contribution, as well as standard assumptions on the stationary process that is modulated. Finally, concluding remarks can be found in Section 6.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{Arthur.eps}
\caption{(a) The trajectories of the 200 drifters from the Global Drifter Program, analysed in Section~\ref{sec=GDPapp}, that exhibit the greatest change in Coriolis frequency ($f$) across 60 inertial cycles, as described in that section; (b) a segment of data of the meridional (latitudinal) positions over time from Drifter ID\#43594; and (c) a segment of data of the meridional velocities from this drifter in cm/s. This figure is produced using the jLab toolbox~\citep{jlab}.}
\label{Equatorial}
\end{figure}
\section{Modulated time series}
In this paper we review and study modelling and inference methods for univariate and bivariate nonstationary Gaussian processes.
We have that the first moment of a univariate stochastic discrete Gaussian process $\{X_t:t\in\N\}$, with index set $\N = \left\{0,1,2,\cdots \right\}$, is provided pointwise by
\begin{equation*}
\mu_X(t)=\E \left\{X_t\right\},
\end{equation*}
and the second-order structure is given by
\begin{equation*}
c_X(t_1,t_2)=\cov\left\{ X_{t_1},X_{t_2}\right\},
\end{equation*}
where moments are finite as a direct consequence of the joint Gaussianity of $\{X_t\}$.
We shall assume throughout this paper that $\mu_X(t) = 0$. In practice this may require us to subtract the sample mean from the observed series, or more generally remove trends and seasonal components \citep[see][chap. 1]{BrockwellDavisTimeSeries}.
Second-order stationarity implies that the function $c_X(t,t+\tau)$ does not depend on the index $t$ and takes the simplified form $c_X(\tau)$. An alternative way to represent $c_X(\tau)$, assuming it is absolutely summable, is via its Fourier transform $S_X(\cdot)$, also known as the spectral density of $\{X_t\}$,
\[
S_X(\omega) =\frac{1}{2\pi}\sum_{\tau=-\infty}^\infty{c_X(\tau)e^{-i\omega\tau}}, \ \ \omega\in [-\pi,\pi].
\]
The spectral density $S_X(\omega)$ is then a continuous function of $\omega$. The corresponding inversion formulae is given for all integer value $\tau$ by
\[
c_X(\tau) = \int_{-\pi}^\pi{S_X(\omega)e^{i\omega\tau}d\omega}.
\]
A consequence of stationarity is that the quantities in question can be stably estimated by averaging in time \citep{BrockwellDavisTimeSeries}. If $c_X(t,t+\tau)$ is not stationary, but is slowly varying in time, then it can be estimated by dividing the observed data into multiple segments and performing inference on each segment~\citep{adak1998time}.
This does not hold in settings where the time variation is too rapid.
Our goal is to estimate $c_X(t,t+\tau)$ in such settings, in particular when a parametric specification is made for the function.
\subsection{Classes of modulated processes}
Modulation is a natural and simple method of producing a nonstationary process~\citep{Parzen1963}.
A univariate modulated process is defined as follows.
\begin{definition}[Modulated process]
\label{def=modulatedprocesses}
Let $\{X_t:t\in\N\}$ be a Gaussian, real-valued, zero-mean stationary process.
Let $\{g_t:\; t\in {\mathbb{N}}\}$ be a given bounded real-valued deterministic sequence.
Then a modulated process is defined as one taking the form
\begin{equation}
\label{modeqn}
\Y_t=g_t X_t
\end{equation}
at all time points $t\in\N$.
\end{definition}
Herein we treat $\{g_t\}$ as a known deterministic signal.
In our setting the process $\{X_t\}$, which is referred to as the {\em latent} process, is modelled through a finite set of parameters $\boldsymbol{\theta}\in\Theta\subset\R^d$, where $d$ is a positive integer and $\Theta$ is the parameter space. Usually our object of interest is $\boldsymbol{\theta}$, the particular values of parameters that generated the observed realization.
For example, if the latent process is an autoregressive process of order $p$, we then have $d = p+1$ if the mean is known ($p$ regressive parameters and the variance of the innovations).
We denote the autocovariance function of the stationary zero-mean process $\{X_t\}$ by $c_X(\tau)$, or $c_X(\tau;\boldsymbol{\theta})$ when we want to make the dependence on $\boldsymbol{\theta}$ explicit. Its Fourier transform, the spectral density, is denoted $S_X(\omega)$ or $S_X(\omega;\btheta)$, respectively.
The modulation of the latent process $X_t$ is a convenient mechanism to account for a wide range of nonstationary processes. In particular this mechanism has been widely used as a modelling tool for missing data problems, where $g_t$ is assigned values $0$ or $1$ when respectively missing or observing a data point in time \citep{R.H}.
To understand when we can recover the parameters controlling the latent process $X_t$ from observing $\Y_t$, we need to put further conditions in place on $g_t$.
The time series $\Y_t/g_t$ cannot always be formed as $g_t$ may be zero for some time indices, corresponding to missing observations. Another reason is that we may not directly observe $\Y_t$, but instead we may observe an aggregated process $\Y_t + Z_t$, where $Z_t$ is a stationary process (or more generally another modulated process) independent from $\Y_t$, this preventing us from recovering the stationary latent process $X_t$ by division.
We assume that $\Y_t$ satisfies~\eqref{modeqn} for a Gaussian, real-valued, zero-mean stationary $X_t$ with absolutely summable autocovariance sequence. Then $\E\{\Y_t\}=g_t\E\{X_t\}=0$ and the time-varying autocovariance sequence is defined by $c_{\Y}(t, t+\tau;\btheta)=\E\left\{\Y_t\Y_{t+\tau}\right\}$. Given a single length N realization $\Y_0,\cdots,\Y_{N-1}$, we start by computing the usual method of moments estimator according to
\begin{equation}
\label{eq=biasedAutocovEst}
\hat{c}_{\Y}\sN(\tau)=\frac{1}{N}\sum_{t=0}^{N-\tau-1} \Y_t\Y_{t+\tau},
\end{equation}
for $\tau=0,1,...,N-1$, such that $\tau$ is within the range of time offsets that is permissible given the length-$N$ sample.
Equation~\eqref{eq=biasedAutocovEst} is the biased sample autocovariance sequence of the modulated time series, which we define even though the process is nonstationary, as this object will become pivotal in our estimation procedure.
The expectation of this object, which we denote $\overline{c}_{\Y}^{(N)}(\tau;\btheta)$ or simply $\overline{c}_{\Y}^{(N)}(\tau)$, takes the following form,
\begin{equation}
\label{eq=expectedAutocovSequence1}
\overline{c}_{\Y}\sN(\tau)= \E\{\hat{c}_{\Y}\sN(\tau)\} = \E\left\{ \frac{1}{N}\sum_{t=0}^{N-\tau-1} \Y_t\Y_{t+\tau} \right\} = c_X(\tau)\frac{1}{N}\sum_{t=0}^{N-\tau-1} g_t g_{t+\tau} = c_g\sN(\tau)\cdot c_X(\tau),
\end{equation}
where we have introduced the (deterministic) sample autocovariance of the modulating sequence,
\begin{equation}
\label{eq=autocovOfg}
c_g\sN(\tau) = \frac{1}{N}\sum_{t=0}^{N-1-k}{g_tg_{t+\tau}}.
\end{equation}
In the specific case where the modulating sequence $\{g_t\}$ is constant and equal to unity everywhere, which would correspond to observing the latent stationary process directly, we recover the expectation of the biased sample autocovariance for stationary time series, $\left(1-\tau/N\right) c_X(\tau)$, for $\tau=0,\cdots,N-1$.
More generally, a standard assumption is to say that the modulated process $\Y_t$ is an asymptotically stationary process~\citep{parzen1961,Parzen1963}, which arises if for all lags $\tau$, the quantity $c_g\sN(\tau)$ in~\eqref{eq=expectedAutocovSequence1} converges as $N$ tends to infinity. We define this formally as follows.
\begin{definition}[Asymptotically stationary process]\label{asympstat}
Let $\{\Y_t\}$ be a discrete time random process.
We say that
$\{\Y_t\}$ is an asymptotically stationary process if there exists a fixed function $\{\gamma(\tau):\tau\in\N \}$ such that for all $\tau\in\N$,
\begin{equation}
\lim_{N\rightarrow \infty}\E\left\{\frac{1}{N}\sum_{t=0}^{N-\tau-1} \Y_t\Y_{t+\tau}\right\} = \gamma(\tau),
\end{equation}
or specifically if $\Y_t$ is a modulated process as defined in Definition~\ref{def=modulatedprocesses}, $\Y_t$ is asymptotically stationary if,
\begin{equation}
\label{eq=asymptStat}
\lim_{N\rightarrow \infty} \overline{c}_{\Y}\sN(\tau) = \gamma(\tau),
\end{equation}
where $\overline{c}_{\Y}\sN(\tau)$ is defined in~\eqref{eq=expectedAutocovSequence1}.
\end{definition}
An example of a nonstationary but asymptotically stationary process is given by \citet{Parzen1963}, where a stationary process is observed according to a periodically missing data pattern, such that the first $k$ values are observed, the next $l$ values are missed, the next $k$ values are observed, and so on, where $k$ and $l$ are two strictly positive integers.
The class of asymptotically stationary modulated processes \citep{ Parzen1963, Dunsmuir1981, Iacobucci2003, Jiang2004} corresponds to that for which there exists a sequence $\{R_g(\tau):\tau\in\N\}$ such that
\begin{equation}
\label{asymptotic}
\lim_{N\rightarrow \infty} c_g\sN(\tau) = R_g(\tau), \ \ \forall\tau\in\N.
\end{equation}
Indeed we then note that $\overline{c}_{\Y}\sN(\tau) \rightarrow R_g(\tau) c_X(\tau)$ as $N\longrightarrow\infty$, so we could estimate $c_X(\tau)$ by defining
\begin{equation}
\label{estimate}
\hat{c}_X\sN(\tau)=\frac{\hat{c}_{\Y}\sN(\tau)}{R_g(\tau)},
\end{equation}
assuming $R_g(\tau) \neq 0$ for all $\tau\in\N$, and is known.
It is shown in \citet{Parzen1963} that $\hat{c}_X\sN(\tau)$ is a consistent estimator of the autocovariance sequence $c_X(\tau)$ of the latent stationary process, under some rather mild conditions. Further results are found in~\cite{Dunsmuir1981}. Consistent spectral density estimates can be obtained by a Fourier transformation of the sequence $\{k\sN(\tau)\hat{c}_X\sN(\tau):\tau=0,\cdots,N-1\}$, where $k\sN(\tau)$ is chosen suitably for $\tau=0,\cdots,N-1$.
The key feature in Definition~\ref{asympstat} is that in~(\ref{eq=asymptStat}) we average the time-varying autocovariance sequence $c_{\Y}(t,t+\tau)=\E\left\{\Y_{t}\Y_{t+\tau}\right\}$ across a time period $N$ to produce an average autocovariance across the time period, written as $\overline{c}_{\Y}\sN(\tau)$.
If this converges (in $N$) to a function of $\tau$, then by observing the modulated process over a suitably long time interval
we can recover the second-order properties of the stationary latent process.
We now wish to explore a more general assumption than that of asymptotic stationarity for modulated processes.
Specifically, we seek a larger class of models where consistent inference is still achievable. This will be smaller than the full class of models for $g_t$, as using a trivial example, if
$g_t\equiv 0$ always then we would not be able to infer properties of the generating mechanism of $X_t$.
For consistent inference we propose the following class of modulated processes.
\begin{definition}[Modulated process with a significant correlation contribution]
\label{def=univariateSignificantCorrel}
Assume that $\Y_t$ is specified by~\eqref{modeqn}. We say that $\Y_t$ is a \emph{modulated process with a significant correlation contribution} if there exists a finite subset of non-negative lags $\Gamma\subset\N$ such that,
\begin{enumerate}
\item The mapping $\btheta\mapsto \left\{c_X(\tau):\tau\in\Gamma\right\}$ is one-to-one (injective).
\item For all lags $\tau\in\Gamma$,
\begin{equation}
\label{eq=assumption1}
\liminf\limits_{N\rightarrow\infty}\left|c_g\sN(\tau)\right| > 0,
\end{equation}
where $\liminf\limits_{N\rightarrow\infty}$ is the limit inferior.
\end{enumerate}
\end{definition}
Because of the symmetry of autocovariance sequences we do not need to consider $\tau<0$ in this definition.
Point 1 of Definition~\ref{def=univariateSignificantCorrel} means that for any two distinct parameter vectors $\btheta,\btheta'\in\Theta$, there exists at least one lag $\tau$ in the finite set $\Gamma$ such that $c_X(\tau;\btheta)\neq c_X(\tau;\btheta')$. It is therefore an assumption about the latent process model.
The sequence $\left|c_g\sN(\tau)\right|$ is bounded above since the modulating sequence is assumed to be bounded above. Therefore the limit inferior in~\eqref{eq=assumption1} is always finite.
We observe that, for $\tau\in\Gamma$,~\eqref{eq=assumption1} is equivalent to,
\begin{equation}
\label{eq=equivalenceoflimitinf}
\exists\alpha_\tau>0, \exists N_\tau\in\N, \forall N\in\N, N\geq N_\tau \Rightarrow \left|c_g\sN(\tau)\right| \geq \alpha_\tau,
\end{equation}
which we interpret as the fact that the sequence $\left|c_g\sN(\tau)\right|$ is bounded below for $N$ large enough.
For further understanding of Point 1 in Definition \ref{def=univariateSignificantCorrel} we provide the following two simple examples.
\begin{enumerate}
\item Let the latent process $\{X_t\}$ be an autoregressive process of order $p$, denoted AR($p$), with known mean zero and unknown innovation variance, and with the parameter set $\Theta$ that is a subset of $\R^{p+1}$. If the parameter set $\Theta$ is chosen appropriately, i.e. such that the roots of the characteristic equation all lie outside the unit circle, the Yule-Walker equations \citep{BrockwellDavisTimeSeries} show that $\btheta\mapsto \left\{c_X(\tau; \btheta):\tau\in\Gamma\right\}$, where $\Gamma = \{0, \cdots,p\}$, is a one-to-one mapping.
Similarly if $\{X_t\}$ is a moving average process of order $q$, denoted MA($q$), with known mean an unknown innovation variance and if the parameter set $\Theta$ is chosen appropriately~\citep{dzhaparize83}, then the mapping $\btheta\mapsto \left\{c_X(\tau; \btheta):\tau\in\Gamma\right\}$, where $\Gamma = \{0, \cdots,q\}$, is one-to-one.
\item Let the latent process $\{X_t\}$ be the MA(2) process defined by,
\begin{equation}
X_t = \sigma\left(\epsilon_t + \theta_2\epsilon_{t-2}\right),
\end{equation}
where the innovations $\epsilon_t$ are i.i.d and have a standard normal distribution and $\sigma>0$. The parameters of the model are $(\theta_2, \sigma)$, and the parameter set $\Theta = \R\times\R\backslash\{0\}$ ensures that the mapping $\btheta\mapsto \left\{c_X(\tau; \btheta):\tau\in\Gamma\right\}$, where $\Gamma = \{0, 2\}$, is one-to-one. Note that observing lag-$1$ is not required here as we have assumed $\theta_1=0$ in the model.
\end{enumerate}
The definition of a \emph{significant correlation contribution} constrains how much energy adds up for any fixed lag $\tau\in\Gamma$. We see directly from~\eqref{eq=expectedAutocovSequence1} that if we assume a significant correlation contribution, the expectation of the estimated autocovariance of $\Y_t$ does not vanish with the length of the observation $N$, at least for lags in $\Gamma$.
This allows for consistent estimation of the parameter $\btheta$ as we will see in Section \ref{sec=consistency}. As a trivial counterexample, assume for instance that $c_g\sN(\tau)$ goes to zero when $N$ goes to infinity. Then $\hat{c}_{\Y}\sN(\tau)$ in~\eqref{eq=biasedAutocovEst} goes to zero as well, independently of the parameter vector $\btheta$, resulting either in infeasible estimation or requiring a change of estimation approach.
Asymptotically stationary modulated processes are a subclass of modulated processes with a significant correlation contribution.
Specifically, for the class of asymptotically stationary modulated processes,~\eqref{estimate} requires that $c_g\sN(\tau)$ converges to the non-zero quantity $R_g(\tau)$, which is a stronger requirement than~\eqref{eq=assumption1} where we only require an asymptotic positive lower bound rather than convergence.
\subsection{Missing observations}
\label{sec=missing}
A particularly enticing use of modulated processes is to account for missing observations in stationary time series.
Let $\{X_t:t\in\N\}$ be a stationary process. For each time point $t\in\N$, we set \citep{Parzen1963},
\begin{equation}
g_t= \left\{
\begin{array}{ll}
0 & \mbox{if $X_t$ is missing} \\
1 & \mbox{if $X_t$ is observed}
\end{array}
\right.
.
\end{equation}
The process $\Y_t=g_t X_t$ is formed at all time points $t\in\N$, forming a modulated process in the sense of Definition~\ref{def=modulatedprocesses}.
An example where the missing observation pattern is deterministic and leads to an asymptotically stationary modulated process is the case of $(k,l)$-periodically missing data treated by \citet{R.H} and \citet{Parzen1963}. This corresponds to observing the $k$ first values, missing the $l$ next values, observing the $k$ next values, and so on. Note that \citet{Parzen1963} requires $k>l$ for non-parametric estimation of the spectral density of $X_t$ based on ~\eqref{estimate}.
Our model of modulated processes with significant correlation contribution allows for $k\leq l$, as long as we observe the lags in $\Gamma$.
A generalization of this missing data scheme was introduced by \citet{clinger1976} with an application to oceanography.
Missing observations can also occur according to a random mechanism. This can be modelled by a random modulation sequence taking values zero and one
\citep{Schein1965, Bloomfield1970}, when the random mechanism according to which missing points occur is independent from the observed process, which we shall assume.
Conditioning on the observed modulation function, we then return to the deterministic modulating sequence described in this paper.
Most works, to our knowledge, have assumed some sort of stationarity for the random modulation sequence, i.e. that the sample autocovariance of the modulation sequence converges almost surely to a non-zero value at all lags~\citep{Dunsmuir1981b, Dunsmuir1981c}.
Some authors do not require such an assumption but have treated only specific models, usually autoregressive models \citep{Jones1980,Broersen2004}.
The definition of a modulated process with a significant correlation contribution in such a situation needs to be understood in a probabilistic fashion, i.e. we require that Property 2 of Definition \ref{def=univariateSignificantCorrel} be satisfied with probability one. Indeed, if one sees the general random experiment as a two-step experiment, where first the random modulating sequence $\{g_t\}$ is generated and observed and then a stationary process $\{X_t\}$ is modulated by this modulating sequence to produce $\{\Y_t\}$, then with probability one the moduating sequence $\{g_t\}$ in the first step makes $\{\Y_t\}$ a modulated process with significant correlation contribution.
Such a situation will be described by saying that $\{\Y_t\}$ is a modulated process with an almost surely significant correlation contribution.
We shall now give a few examples of cases satisfying the stated conditions.
\begin{enumerate}
\item Let $X_t$ be an AR($p$) Gaussian process with mean zero. If we set $\Gamma=\left\{ 0,\cdots,p\right\}$, and if the missing data occurs deterministically according to a $(k,l)$-periodic pattern,
$k\geq p$ is a sufficient condition for the resulting modulated process to have a significant correlation contribution. This is because we are able to observe an infinite number of time the lags in $\Gamma$. We do not require any additional condition on $l$.
\item \label{ex=missingDataScheme} Let $X_t$ be an AR($p$) process, and consider the missing data scheme treated by \citet{Schein1965}, where the random mechanism is a sequence of Bernoulli i.i.d trials with identical probability of success (to be understood as \emph{observation} here) $0<p\leq 1$.
According to the strong central limit theorem, for all lag $\tau\in\N$, $c_g\sN(\tau)$ converges a.s. to $p^2>0$ and therefore $\liminf\limits_{N\rightarrow\infty}\left|c_g\sN(\tau)\right| > 0$ a.s. Therefore the observed process is a modulated process with an almost surely significant correlation contribution.
\item\label{example=missingDatasch} Consider the random mechanism where the sequence $\{g_t\}$ is generated according to
\begin{equation}
g_t \sim \mathcal{B}(p_t),
\end{equation}
where $\mathcal{B}(p)$ represents the Bernoulli distribution with parameter $p$, and where we set
\begin{equation}
\label{eq:missing3}
p_t = \mathcal{P} + A_p\cos\left(\omega_p t\right),
\end{equation}
with $0<\mathcal{P}<1$, $0\leq A_p<\min\left(\mathcal{P}, 1-\mathcal{P}\right)$ (which ensures $0<\mathcal{P}-A_p\leq p_t\leq1, \forall t\in\N$), and $\omega_p\in[-\pi,\pi]$.
The Bernoulli parameters $p_t$ as given by~\eqref{eq:missing3} will oscillate periodically around their mean value $\mathcal{P}$. This also leads to
$\liminf\limits_{N\rightarrow\infty}\left|c_g\sN(\tau)\right| > 0$ a.s., using the fact that $p_t$ is bounded below by $\mathcal{P}-A_p>0$.
\end{enumerate}
In section \ref{sec=missingSims} we will provide a simulation study based on example \ref{example=missingDatasch}. This is novel in comparison of previsouly studied missing observation schemes as we do not make an assumption of stationarity for the process $g_t$ \citep{Dunsmuir1981}.
\subsection{Sampling properties of modulated processes}
\label{sec=samplingproperties}
In this section we shall review and study some distributional properties of the periodogram of a modulated time series. \citet{Dunsmuir1981} used the periodogram as the basis for designing pseudo-likelihood
methods for asymptotically stationary modulated time series, with an emphasis on treating the problem of missing data.
Similarly, in Section \ref{sec=estimation} we will use the results of this section to formulate a pseudo-likelihood using the periodogram, for our class of modulated processes with significant correlation contribution. Herein we shall denote $\Omega_N$ the set of Fourier frequencies $\frac{2\pi}{N}\cdot\left( -\lceil\frac{N}{2}\rceil+1, \cdots, -1, 0, 1, \cdots, \lfloor\frac{N}{2}\rfloor \right)$.
We denote $\mathbf{\Y}=\{\Y_t:t=0,\cdots,N-1\}$ as a single realization of a length-$N$ sample of a modulated process $\{\Y_t\}$ defined in Definition \ref{def=modulatedprocesses}. The unobserved sample of the latent stationary process is denoted $\mathbf{X}=\{X_t:t=0,\cdots,N-1\}$ accordingly.
The squared modulus of the Fourier transform of the time series $\mathbf{X}$, known as the \emph{periodogram}, is a common statistic in stationary time series analysis~\citep{SpectralAnalysis}, and is given by
\begin{equation}
\label{eq=defPeriodogram}
\hat{S}_X\sN(\omega) = \frac{1}{N}\left|\sum_{t=0}^{N-1}{X_t e^{-i\omega t}}\right|^2,\ \omega\in\R.
\end{equation}
Note that this quantity is $2\pi$-periodic, i.e. $\hat{S}_{X}\sN(\omega+2\pi)=\hat{S}_{X}\sN(\omega)$, $\omega\in\R$. The periodogram of the sample $\bold{X}$ is an asymptotically unbiased estimator of the spectral density of the stationary process $\{X_t\}$, i.e. $\allowbreak\lim_{N\rightarrow\infty}\E\{\hat{S}_X\sN(\omega);\btheta\}=2\pi S_X(\omega;\btheta)$ for all $\omega\in[-\pi,\pi)$~\citep{BrockwellDavisTimeSeries}. However the variance of the periodogram does not decrease to zero as the sample size increases.
A consistent nonparametric estimator of a smooth spectral density $S_X(\omega;\boldsymbol{\theta})$ of the latent process $\{X_t\}$, were it to be directly observed, could be obtained by smoothing the periodogram across frequencies~\citep[p. 235--253]{SpectralAnalysis}, as long as $S_X(\omega;\boldsymbol{\theta})$ is continuous.
For the modulated process $\{\Y_t\}$, the latent time series $\{X_t\}$ is not observed, so we instead compute the periodogram of the modulated (and observed) process itself, $\hat{S}_{\Y}\sN(\omega)$, and we define the expected periodogram to be
\[
\overline{S}_{\Y}^{(N)}(\omega;\boldsymbol{\theta}) = \E\left\{\hat{S}_{\Y}\sN(\omega);\; \boldsymbol{\theta}\right\}, \ \ \omega\in\R.
\]
Note that this quantity is also $2\pi$-periodic.
It is necessary to understand how modulation in the time domain will affect the expected periodogram.
Proposition \ref{prop=freqPropSbar} gives more insight on how $\overline{S}_{\Y}\sN(\omega;\boldsymbol{\theta})$ relates to the modulating sequence $g_t$ and the spectral density $S_X(\omega;\boldsymbol{\theta})$ of the latent stationary process $\{X_t\}$.
\begin{proposition}[Expectation of the periodogram of a modulated time series]
\label{prop=freqPropSbar}
The expectation of the periodogram of the modulated time series takes the form
\begin{equation}
\label{eq=Sbar1}
\overline{S}_{\Y}\sN(\omega;\boldsymbol{\theta})
= 2\pi\int_{-\pi}^\pi{S_X(\omega-\lambda;\boldsymbol{\theta})S_g\sN(\lambda)d\lambda}, \ \forall\omega\in\R,
\end{equation}
which is a periodic convolution.
Here $S_g\sN(\lambda)$ is the squared value of the Fourier Transform
of the finite sequence $\{g_t\}_{t=0,\cdots,N-1}$ i.e.
\[
S_g\sN(\lambda) = \frac{1}{2\pi N}\left|\sum_{t=0}^{N-1}{g_t e^{-i\lambda t}}\right|^2,
\]
defined for $\lambda\in\R$ and which is $2\pi$ periodic.
\end{proposition}
\begin{proof}
The proof for this proposition, which is a well-known result, can be found in~\citet[p.~562]{Dunsmuir1981}.
\end{proof}
When $g_t=1$ everywhere, which corresponds to observing the stationary latent process directly, the quantity $S_g\sN(\lambda)$ is the usual Féjer kernel \citep{Bloomfield2000} defined by,
\begin{equation}
\label{eq=FejerKernel}
\mathcal{F}^{(N)}(\lambda) = \frac{\sin^2\left(\frac{N\lambda}{2}\right)}{2\pi N\sin^2\left(\frac{\lambda}{2}\right)}, \ \ \forall\lambda\in\mathbb{R}\setminus\Omega_N,
\end{equation}
which behaves asymptotically (as $N$ tends to infinity) as a Dirac delta-function centred at zero. This explains why the periodogram is, asymptotically, an unbiased estimator of the spectral density of a stationary process up to a multiplicative factor of $2\pi$~\citep{BrockwellDavisTimeSeries}.
When $g_t$ is such that the modulated process is asymptotically stationary, \citet{Dunsmuir1981b} approximate $\frac{1}{2\pi}\sum_{\tau=-\infty}^\infty{\gamma(\tau)e^{i\omega\tau}}$, where $\gamma(\tau) = R_g(\tau)c_X(\tau)$ using the notation of~\eqref{asymptotic}, for $\omega$ at Fourier frequencies by,
\begin{equation}
\label{eq=DunsmuirSpectrum}
\widetilde{S}_{\Y}^{(D)}(\omega;\btheta) = \frac{2\pi}{N}\sum_{\lambda\in\Omega_N}{S_X(\omega-\lambda;\btheta)S_g^{(N)}(\lambda)}.
\end{equation}
When $g_t$ is such that the modulated process $\Y_t$ has a significant correlation contribution, we derive the exact value of $\overline{S}_{\Y}\sN(\omega;\boldsymbol{\theta})$ by using the theoretical autocovariances of the latent model, in a similar fashion as in \citet{WhittleAdam} for stationary processes. This is the result of Proposition \ref{prop=expectationPeriodogram2}, which follows.
\begin{proposition}[Computation of the expected periodogram]
\label{prop=expectationPeriodogram2}
Let $\omega\in\R$.
We have
\begin{equation}
\label{eq=computationOfPeriodogram}
\overline{S}_{\Y}\sN(\omega;\boldsymbol{\theta}) =
2\mathcal{R}\left\{\sum_{\tau=0}^{N-1}{\overline{c}_{\Y}\sN(\tau;\boldsymbol{\theta})e^{-i\omega \tau}}\right\}
- \overline{c}_{\Y}\sN(0;\boldsymbol{\theta}),
\end{equation}
where $\overline c_{\Y}\sN(\tau;\bm\theta)$ is defined in~(\ref{eq=expectedAutocovSequence1}).
By defining $\overline{c}_{\Y}\sN(-\tau;\boldsymbol{\theta}) = \overline{c}_{\Y}\sN(\tau;\boldsymbol{\theta})$ for $\tau=1, \cdots, N-1$ we can (equivalently) express this relationship as
\[
\overline{S}_{\Y}\sN(\omega;\boldsymbol{\theta}) =
\sum_{\tau=-(N-1)}^{N-1}{\overline{c}_{\Y}\sN(\tau;\boldsymbol{\theta})e^{-i\omega \tau}},
\]
\end{proposition}
\begin{proof}
The proof, which is standard~\citep[page 334]{BrockwellDavisTimeSeries}, follows directly from~\eqref{eq=defPeriodogram} and~\eqref{eq=expectedAutocovSequence1} in a few lines of algebra after aggregating along the diagonal of the covariance matrix.
\end{proof}
Therefore the expectation of the periodogram of $\bold{\Y}$ is the discrete Fourier transform of the expected sample autocovariance sequence.
This is true even though we have not assumed stationarity; it is simply a consequence of the relation between the formal definitions of~\eqref{eq=expectedAutocovSequence1} and~\eqref{eq=defPeriodogram}.
Note that calculating the Fourier transform of the sequence $\overline{c}_{\Y}\sN(\tau;\theta)$ will always give a real-valued positive $\overline{S}_{\Y}\sN(\omega; \boldsymbol{\theta})$ for $\boldsymbol{\theta}\in\Theta$, as the latter is defined as the expectation of the squared modulus of the Fourier transform of the process.
Proposition~\ref{prop=expectationPeriodogram2} can be used to compute the expected periodogram of an asymptotically stationary modulated process. In such cases, the difference between~\eqref{eq=DunsmuirSpectrum} and Proposition~\ref{prop=expectationPeriodogram2} is that~\eqref{eq=DunsmuirSpectrum} is a finite approximation of~\eqref{eq=Sbar1}, whereas Proposition~\ref{prop=expectationPeriodogram2} is exact. The difference occurs because~\eqref{eq=DunsmuirSpectrum} does not account for the bias of the periodogram that results from leakage (see \citet{WhittleAdam}), whereas these effects are naturally accounted for in Proposition~\ref{prop=expectationPeriodogram2}.
To justify the use of the expected periodogram in the setting of modulated processes with a significant correlation contribution,
we now consider what conditions are required for the expected periodogram to \emph{carry enough information} so that the parameter vector is identifiable within the parameter set $\Theta$.
\begin{proposition}[Identifiability of the expected periodogram]
\label{prop=identifiabilityViaPeriodogram}
If the modulated process has a significant correlation contribution, the expected periodogram is a one-to-one (i.e. injective) mapping from the parameter set $\Theta$ to the set of non-negative continuous functions on $[-\pi,\pi]$, for a large enough sample size.
More specifically, for two distinct parameter vectors $\btheta$ and $\btheta'$, the expected periodograms $\overline{S}_{\Y}\sN(\omega;\btheta)$ and $\overline{S}_{\Y}\sN(\omega;\btheta')$ cannot be equal for all Fourier frequencies $\frac{2\pi}{N}\left( -\lceil\frac{N}{2}\rceil+1, \cdots, -1, 0, 1, \cdots, \lfloor\frac{N}{2}\rfloor \right)$.
\end{proposition}
\begin{proof}
Let $\btheta, \widetilde{\btheta}\in\Theta$ be distinct parameter vectors and let $N$ be a positive integer. Let $\Gamma$ be as given by Definition~\ref{def=univariateSignificantCorrel}.
By the assumption of significant correlation contribution, the finite sequences $\left\{c_X(\tau; \btheta):\tau\in\Gamma\right\}$ and $\{c_X(\tau; \widetilde{\btheta}):\tau\in\Gamma\}$ are not equal. Since $\overline{c}_{\Y}\sN(\tau;\btheta) = c_g\sN(\tau)c_X(\tau;\btheta)$ for $\tau\in\Gamma$, and according to~\eqref{eq=equivalenceoflimitinf}, for $N$ large enough the sequences $\{\overline{c}\sN_{\Y}(\tau; \btheta):\tau\in\Gamma\}$ and $\{\overline{c}\sN_{\Y}(\tau; \widetilde{\btheta}):\tau\in\Gamma\}$ are not equal. Hence for $N$ large enough the sequences $\{\overline{c}\sN_{\Y}(\tau; \btheta):\tau=-(N-1),\cdots,N-1)\}$ and $\{\overline{c}\sN_{\Y}(\tau; \widetilde{\btheta}):\tau=-(N-1),\cdots,N-1)\}$ are not equal. Their finite Fourier transforms $\{\overline{S}_{\Y}\sN(\omega;\btheta):\omega\in\Omega_N\}$ and $\{\overline{S}_{\Y}\sN(\omega;\widetilde{\btheta}):\omega\in\Omega_N\}$, are by the bijective nature of the Fourier transform, not equal either.
\end{proof}
This means that for two distinct parameters vectors $\btheta,\btheta'\in\Theta$, we will have two distinct expected periodograms. This is a necessary condition for an estimation procedure based on the expected periodogram.
We will propose such an estimation procedure in Section~\ref{sec=estimation}, and derive its consistency and convergence rate in Section~\ref{sec=consistency}.
\subsection{Bivariate modulated processes}
\label{sec=bivariate}
It is common in pratical applications to observe more than one time series at any time, and to analyse a set together. Often the series in the set are related via phase-shifts and other small temporal inhomogeneities, see e.g.~\citet{allen1996distinguishing,runstler2004modelling,allefeld2009mental,lilly2012analysis}.
Bivariate nonstationary processes can be challenging to model, as they may not be representable in the same nonstationary oscillatory family~\citep{tong1973some,tong1974time}.
To explore the nature of multivariate modulation, we shall investigate the representation
of bivariate processes. For ease of exposition we shall represent such series using complex-valued time series, see~\citet{walker1993complex}. We shall continue to assume that the latent process, now denoted $Z_t$ for complex-valued processes, is Gaussian and zero-mean, leaving only the second order structure to be modelled.
For complex-valued processes both the autocovariance $c_Z(\tau)=\E\{Z_t^*Z_{t+\tau}\}$ (the star denotes conjugation) and the relation $r_Z(\tau)=\E\{Z_t Z_{t+\tau}\}$ sequences need to be modelled~\citep{walden2013rotary}. Complex-valued processes, unlike real-valued, no longer have a spectrum that needs to satisfy Hermitian symmetry, and if the series represents motion in the plane, the positive and negative frequencies represent clockwise and anti-clockwise rotations respectively.
Following the classical modelling framework~\citep{miller1969complex} for complex-valued processes we shall assume that the relation sequence takes the value zero for all lags. The complex-valued process is then said to be \emph{proper}. The assumption of propriety has the consequence of directly extending equation~\eqref{modeqn} to the complex-valued case from the real-valued case. Specifically, let $Z_t$ be a complex-valued Gaussian proper zero-mean process, a complex-valued modulated process is defined as one taking the form,
\begin{equation}
\label{eq=complexMod}
\widetilde{Z}_t = g_tZ_t,
\end{equation}
at all times $t\in\N$, where $g_t=\rho_t e^{i\phi_t}$ is a bounded modulation sequence.
We note that for complex-valued time series the modulation sequence is complex-valued.
With this definition, the modulation series $g_t$ accomplishes a time-dependent rescaling or expansion/dilation, from $\rho_t$, together with a time-dependent rotation, from $e^{i\phi_t}$.
The autocovariance of the complex-valued modulated process $\widetilde{Z}_t$ at times $t_1$ and $t_2$ is given by the conveniently simple form,
\begin{equation}
\label{autocovY1}
\nonumber
c_{\widetilde{Z}}(t_1,t_2;\boldsymbol{\theta})=\E\left\{\widetilde{Z}_{t_1}^*\widetilde{Z}_{t_2};\btheta\right\}=g_{t_1}^*g_{t_2}c_Z(t_2-t_1;\boldsymbol{\theta}) = \rho_{t_1}\rho_{t_2} e^{i(\phi_{t_2}-\phi_{t_1})}c_Z(t_2-t_1;\boldsymbol{\theta}),
\end{equation}
and $c_{\widetilde{Z}}(t_1,t_2;\boldsymbol{\theta})$ fully characterizes the process.
Note that this quantity is not only a function of the lag $t_2-t_1$ as the process is no longer stationary.
Similarly to the univariate case cf.~\eqref{eq=autocovOfg}, let $N$ be any positive integer, we define for $\tau=0,\cdots,N-1$,
\begin{equation}
\label{eq=cgcomplex}
c_g\sN(\tau) = \frac{1}{N}\sum_{t=0}^{N-\tau-1}{g_t^*g_{t+\tau}}.
\end{equation}
Note that when $g_t$ is real-valued (\ref{eq=cgcomplex}) and (\ref{eq=autocovOfg}) are the same.
We also extend the notion of a significant correlation contribution for complex-valued modulated processes, which naturally mimics Definition~\ref{def=univariateSignificantCorrel}.
We define the expected periodogram of a complex-valued modulated time series as
$\overline{S}_{\widetilde{Z}}\sN(\omega) = \E\left\{
\hat{S}_{\widetilde{Z}}\sN(\omega) ; \btheta\right\}$, which can be computed efficiently similarly to Proposition~\ref{prop=expectationPeriodogram2} for the univariate case, by replacing $\Y_t$ by $\widetilde{Z}_t$ in~\eqref{eq=expectedAutocovSequence1} and~\eqref{eq=computationOfPeriodogram}.
A univariate real-valued modulated process is stationary if and only if the modulating sequence is a constant.
A necessary and sufficient condition on the modulating sequence for the complex-valued modulated process~\eqref{eq=complexMod} to be stationary is more complicated to obtain, and is determined in the following proposition.
\begin{proposition}[Stationary bivariate modulated processes]
\label{prop=stationaryModulatedProcesses}
Let $\widetilde{Z}_t$ be the complex-valued modulated process defined in~\eqref{eq=complexMod}.
First, assume the latent process $\{Z_t\}$ is a white noise process.
Then the modulated process $\{\widetilde{Z}_t\}$ is stationary if and only if the modulating sequence $g_t = \rho_te^{i\phi_t}$ is
of constant modulus, i.e. $\rho_t=a\geq0$. In such case the modulated process is a white noise process with variance $a^2 \E\{|Z_0|^2\}$.
More generally, assume the stationary latent process $\{Z_t\}$ is not a white noise process, and let
$\mu = \gcd\{\tau\neq0\in\N :|c_Z(\tau;\boldsymbol{\theta})|>0\}$ where $\gcd$ denotes the greatest common divisor.
Then the modulated process is stationary if and only if $\{g_t\}$ is zero everywhere or if there
exists two constants $a>0$ and $\gamma\in[-\pi,\pi)$ such that for all $t\in\N$, letting $r=t\bmod\mu$ be the remainder of $t$ divided by $\mu$,
\begin{eqnarray*}
\rho_t &=& a\\
\phi_t &=& \phi_{r} + \gamma \left\lfloor{\frac{t}{\mu}}\right\rfloor \mod 2\pi,
\end{eqnarray*}
where $\left\lfloor{\frac{t}{\mu}}\right\rfloor$
denotes the floor of $\frac{t}{\mu}$ and$\mod 2\pi$ indicates that the equality is true up to an additive multiple of $2\pi$.
In this case the spectral density of the modulated process $\{\widetilde{Z}_t\}$ is
\begin{equation*}
S_{\widetilde{Z}}(\omega) = a^2S_Z\left(\omega-\frac{\gamma}{\mu}\right).
\end{equation*}
\end{proposition}
\begin{proof}
See appendix \ref{proof=stationaryModulatedProcesses}.
\end{proof}
The value of $\mu$ in Proposition \ref{prop=stationaryModulatedProcesses} depends on the location of zeros in the covariance sequence of the latent process.
In particular, if $|c_Z(1;\boldsymbol{\theta})|>0$ then $\mu = 1$ and $\widetilde{Z}_t$ is stationary only if there exists a constant $\gamma\in\R$ such that for all $t\in\N$, $\phi_t = \phi_0 + \gamma t \mod 2\pi$. If $|c_Z(2;\boldsymbol{\theta})|>0$ but $|c_Z(\tau;\boldsymbol{\theta})|=0$ for all $\tau\in\N, \tau\neq0,2$, then $\mu=2$ (this can occur with a second-order moving average process for instance). In that case the modulated process $\widetilde{Z}_t$ is stationary if and only if there exists a constant $\gamma\in[-\pi,\pi)$ such that for all $t\in\N$, $\phi_t = \phi_0 + \gamma \frac{t}{2} \mod 2\pi$ if $t$ is even, or $\phi_t = \phi_1 + \gamma \frac{t-1}{2} \mod 2\pi$ if $t$ is odd.
\subsubsection{A time-varying bivariate autoregressive process}
\label{sec=tvAR}
We now introduce the specific bivariate autoregressive model that will be used in our real-world data application.
We consider the discrete-time complex-valued process $\{\widetilde{Z}_t:t\in\N\}$, defined by
\begin{eqnarray}
\label{eq=tvARCdef}
\widetilde{Z}_t &=& r e^{i\beta_t}\widetilde{Z}_{t-1} + \epsilon_t, \ \ \ \ \ \ \ t\geq1,\ \ 0\leq r < 1, \ \ \beta_t\in\R,\\
\nonumber\widetilde{Z}_0 &\sim& \mathcal{N}_{C}\left(0, \frac{\sigma^2}{1-r^2}\right),\ \ \ \ \sigma>0,\\
\nonumber\epsilon_t &\sim& \mathcal{N}_{C}\left(0, \sigma^2\right),
\end{eqnarray}
where $\mathcal{N}_{C}\left(0, \sigma^2\right)$ denotes the complex-valued normal distribution with mean $0$ and variance $\sigma^2$, and with i.i.d real and imaginary part. Note that the real and imaginary parts of $\epsilon_t$ then each have variance $\sigma^2/2$.
Here $0\leq r < 1$ is commonly known as either the autoregressive or the damping parameter, ensuring the mean-reversion of the process.
By mean-reversion we mean that, beginning at any time $t$, we have $\lim_{\tau\to\infty} \E\left\{\widetilde{Z}_{t+\tau}|\widetilde{Z}_t\right\} = 0$, i.e. irrespective of the size of the perturbation $\epsilon_t$ at time $t$, the process is expected to return to its mean of 0. This is seen from the following inductive relationship,
\begin{equation*}
\tilde{Z}_{t+\tau} = r^\tau e^{i\sum_{j=1}^{\tau}{\beta_{t+j}}}\tilde{Z}_t + \sum_{j=1}^\tau {r^{\tau-j}e^{i\sum_{k=j+1}^\tau\beta_k}\epsilon_{t+j}}, \ \ \tau\geq 0,
\end{equation*}
which leads to
\begin{equation*}
\E\left\{\tilde{Z}_{t+\tau}|\tilde{Z}_t\right\} = r^\tau e^{i\sum_{j=1}^\tau{\beta_{t+j}}}\tilde{Z}_t,
\end{equation*}
which goes to zero exponentially as $\tau$ goes to infinity, since $0\leq r<1$.
A damping parameter $r$ close to 1 will lead to a slowly-decaying autocorrelation sequence.
A value of $r$ close to 0 will lead to a process with very short memory, with the limiting behaviour of a white noise process as $r\rightarrow0$.
The parameter $\beta_t$ is a known, dimensionless time-varying frequency, which we shall take within the interval $[-\pi,\pi)$ without loss of generality.
The process \eqref{eq=tvARCdef} is a nonstationary version of the complex-valued first order autoregressive process \citep{AdamWidelyLinear} introduced by~\citet{le1988note}, and also a discrete-time analogue of the complex-valued Ornstein-Ulhenbeck (OU) process \citep{Arato1962estimation} with time-varying oscillation frequency. We now prove in Proposition \ref{prop=tvarmodulated} that the model defined in (\ref{eq=tvARCdef}) belongs to our class of bivariate modulated processes.
\begin{proposition}[Modulated process representation]
\label{prop=tvarmodulated}
Let $\{\widetilde{Z}_t\}$ be the process defined in (\ref{eq=tvARCdef}).
There exists a unit-magnitude complex-valued modulating sequence $g_t$,
and a stationary complex-valued proper process $\{Z_t\}$ such that $\{\widetilde{Z}_t\}$ is the modulation of $\{Z_t\}$ by the non-random sequence $\{g_t\}$.
More explicitly, we have $\widetilde{Z}_t = g_t Z_t$, for all $t\in\N$, where,
\begin{eqnarray}
\label{eq=tvARcModulatingSeq}
g_t &=& e^{i\sum_{u=1}^{t}{\beta_u}},\\
\nonumber Z_t &=& r Z_{t-1} + \epsilon_t', \ \ t\geq 1,
\end{eqnarray}
and $Z_0\sim\mathcal{N}_{C}(0, \sigma^2/(1-r^2))$. The process $\epsilon_t'$ is a Gaussian white noise process with the same properties
(zero-mean, variance $\sigma^2$ and independence of real and imaginary parts) as those of $\epsilon_t$. Defined as such, the latent complex-valued process $Z_t$ is stationary and proper.
\end{proposition}
\begin{proof}
See Appendix \ref{proof=tvarmodulated}.
\end{proof}
The stationary latent process $Z_t$ defined in~\eqref{eq=tvARcModulatingSeq} is a stationary complex-valued first order autoregressive process, and is Gaussian. Its autocovariance sequence is given by,
\begin{equation*}
c_Z(\tau) = \frac{\sigma^2}{1-r^2}r^\tau, \ \ \tau\in\Z.
\end{equation*}
It is easy to verify that the mapping $\left(r, \sigma\right)\mapsto \left(c_Z(0), c_Z(1)\right)$ is a one-to-one mapping.
In the following proposition, we stipulate a sufficient condition on the frequencies $\beta_t$ so that the process defined in $(\ref{eq=tvARCdef})$ satisfies our assumption of significant correlation contribution, when represented as a modulated process as defined in Proposition \ref{prop=tvarmodulated}.
\begin{proposition}
\label{prop=tvARcCorrelationContribution}
Let $\tilde{Z_t}$ be the process defined by (\ref{eq=tvARCdef}). Assume that there exists $\Xi\in[-\pi,\pi)$ and $0\leq\Delta\leq\frac{\pi}{2}$ such that for all $t\in\N$, $\left|\beta_t-\Xi\right|\leq \Delta$. Then
$\tilde{Z_t}$ is a modulated process with significant correlation contribution.
\end{proposition}
\begin{proof}
See Appendix \ref{proof=tvARcCorrelationContribution}.
\end{proof}
Hence the complex-valued autoregressive process defined by (\ref{eq=tvARCdef}) belongs to the class of processes with a significant correlation contribution, and the expected periodogram is a one-to-one mapping from the parameter set $[0,1)\times[0,\infty)$ to the set of non-negative continuous functions on $[-\pi,\pi]$, according to Proposition~\ref{prop=identifiabilityViaPeriodogram}---a proposition which is readily extended to the complex-valued processes of this section.
\section{Parametric estimation of modulated processes}
\label{sec=estimation}
We have explored a class of univariate and bivariate modulated processes. The next stage is to describe their efficient inference.
In this section we describe how the parameters of the latent model for $\{X_t\}$ can be inferred from observing a single realization of the modulated process $\{\Y_t\}$.
Most authors have focused on the problem of estimating modulated processes under the assumption of asymptotic stationarity as defined in Definition \ref{asympstat} \citep{Parzen1963, Dunsmuir1981c, Dunsmuir1981, Iacobucci2003}. Although non-parametric estimates have been the key concern in most of the relevant literature, there have been instances where parametric estimation has been considered, see for instance \citet{Dunsmuir1981c}. Parametric estimation ensures that the estimated autocovariance sequence $\hat{c}_X(\tau)$ is non-negative definite, as opposed to using non-parametric estimates of the form given in~\eqref{estimate}. Parametric estimation is also preferable when the true model is known, as it uses the observed degrees of freedom more efficiently. Herein we consider the problem of parametric estimation for our class of modulated processes with a significant correlation contribution, which, we recall, is a generalization of asymptotically stationary modulated processes. We propose an adaptation of the Whittle likelihood \citep{whittle1953estimation}, based on the expected periodogram.
We wish to infer the parameter vector $\boldsymbol{\theta}$ of the latent univariate stationary process $\{X_t\}$ within the parameter set $\Theta$, based on the sample $\bold{\Y}=\Y_0,\cdots,\Y_{N-1}$ and the known modulating sequence $\{g_t:t=0,\cdots,N-1\}$.
Because it has been assumed that the latent process is a zero-mean Gaussian process, the same is true for the modulated process. The vector $\mathbf{\Y}$ is multivariate Gaussian with an expected $N\times N$ autocovariance matrix
$C_{\Y}(\boldsymbol{\theta}) = \left\{c_{\Y}(t_1,t_2;{\boldsymbol{\theta}})\right\}$ for $t_1,t_2=0,\cdots,N-1$, where the components of this matrix are given by $c_{\Y}(t_1,t_2;{\boldsymbol{\theta}}) = g_{t_1} g_{t_2} c_X(t_2-t_1;{\boldsymbol{\theta}})$.
However, the parameter vector $\boldsymbol{\theta}$ of the latent process $\{X_t\}$ can be uniquely determined from the modulated process $\{\Y_t\}$ only if $\boldsymbol{\theta}\rightarrow \left\{c_{\Y}(t_1,t_2;{\boldsymbol{\theta}}):t_1,t_2\in\N\right\}$ is injective, i.e. there is no ${\boldsymbol{\theta}}'\in\Theta$ such that $\btheta\neq\btheta'$ and $c_{\Y}(t_1,t_2;{\boldsymbol{\theta}}) = c_{\Y}(t_1,t_2;{\boldsymbol{\theta'}})\ \forall t_1,t_2\in\N$.
This condition is clearly achieved under the assumption of a modulated process with significant correlation contribution.
The negative of the exact time-domain Gaussian log-likelihood is proportional to
\begin{equation}
\label{eq=timedomainLKH}
\ell_{G}(\boldsymbol{\theta}) =
\frac{1}{N}\log{\left|C_{\Y}(\boldsymbol{\theta})\right|}
+ \frac{1}{N}\mathbf{\Y}^T C_{\Y}(\boldsymbol{\theta})^{-1}\mathbf{\Y},
\end{equation}
where $\left|C_{\Y}(\boldsymbol{\theta})\right|$ denotes the determinant of $C_{\Y}(\boldsymbol{\theta})$. Note that one may need to remove from $\bold{\Y}$ points where $g_t$ is zero, to ensure that the determinant of the covariance matrix is non-zero, and since those observations carry no information about $\btheta$. We minimize $\ell_t$ to obtain the time-domain maximum likelihood estimator (MLE), i.e.
\[
\boldsymbol{\hat{\theta}}_{G} =
\arg \min_{\btheta\in\Theta}\ell_t(\boldsymbol{\theta}).
\]
Parameter estimation based on time-domain likelihood has several drawbacks in the context of modulated processes.
For a large sample size $N$, computing the determinant of the covariance matrix is expensive, requiring $\mathcal{O}(N^3)$ elementary operations in general (although in specific cases such as for Markovian processes the likelihood is obtained in only $\mathcal{O}(N)$ computations).
Moreover each computation of the parametric covariance matrix $C_{\Y}(\boldsymbol{\theta})$ within the exact likelihood requires $\mathcal{O}(N^2)$ operations, compared to $\mathcal{O}(N)$ operations in the case of a stationary process.
We propose a computationally efficient estimation method for the parameters of the latent model based on the periodogram of the modulated time series. First recall that for the stationary time series $\{X_t\}$, making use of the Toeplitz property of the autocovariance matrix, one can approximate the log-likelihood using the Whittle likelihood \citep{whittle1953estimation}, which once discretized is evaluated by
\begin{equation}
\label{eq=whittleLKH}
\ell_W(\btheta) = \frac{1}{N}\sum_{\omega\in\Omega_N}\left\{\log S_X(\omega;\btheta) + \frac{\hat{S}_X^{(N)}(\omega)}{S_X(\omega;\btheta)}\right\},
\end{equation}
where again $\Omega_N$ is the set of Fourier frequencies $\frac{2\pi}{N}\cdot\left( -\lceil\frac{N}{2}\rceil+1, \cdots, -1, 0, 1, \cdots, \lfloor\frac{N}{2}\rfloor \right)$. This pseudo-likelihood has the benefit of $\mathcal{O}(N\log N)$ computational complexity using the Discrete Fourier Transform, with the resulting maximum likelihood estimator being asymptotically equivalent to the time-domain log-likelihood.
We adapt this pseudo-likelihood procedure to modulated processes with significant correlation contribution in the following definition.
\begin{definition}[Spectral maximum pseudo-likelihood estimator for univariate modulated processes]
\label{def=ourEstimator}
Let $\{\Y_t\}$ be a modulated process with significant correlation contribution and let $\bf \Y$ be its length-$N$ sample. We define the following pseudo-likelihood:
\begin{equation}
\label{eq=newLKH}
\ell_M(\boldsymbol{\theta}) =
\frac{1}{N}\sum_{\omega\in\Omega_N}\left\{
\log{\overline{S}\sN_{\Y}(\omega;\boldsymbol{\theta})}
+ \frac{\hat{S}_{\Y}\sN(\omega)}{\overline{S}\sN_{\Y}(\omega;\boldsymbol{\theta})}
\right\},
\end{equation}
where $\overline{S}^{(N)}_{\Y}(\omega)$ is defined in Section \ref{sec=samplingproperties} as the expectation of the periodogram of the modulated time series, and is computed using Proposition \ref{prop=expectationPeriodogram2}.
The corresponding estimator of the parameter vector $\boldsymbol{\theta}$ is obtained by a minimization procedure over the parameter set,
\[
\hat{\boldsymbol{\theta}}_M\sN =
\arg\min_{\btheta\in\Theta}\ell_M(\boldsymbol{\theta}).
\]
\end{definition}
The sequence $\{c_g\sN(\tau): \tau=0,\cdots,N-1\}$ defined in (\ref{eq=autocovOfg}) requires $\mathcal{O}(N^2)$ computations in the most general case. This initial step is carried out independently of inferring the parameter of interest $\btheta$. Then any computation of $\{\overline{S}\sN_{\Y}(\omega;\boldsymbol{\theta}):\omega\in\Omega_N\}$ for any value of the parameter vector $\btheta$ will require $\mathcal{O}(N\log N)$ computations, since we can compute $\{\overline{c}_{\Y}\sN(\tau;\btheta):\tau=0, \cdots, N-1\}$ in $\mathcal{O}(N)$ computations using~\eqref{eq=expectedAutocovSequence1} and the precomputed $\{c_g\sN(\tau): \tau=0,\cdots,N-1\}$, and
the quantity $\{\overline{S}\sN_{\Y}(\omega;\boldsymbol{\theta}):\omega\in\Omega_N\}$ is then computed via a fast Fourier transform.
The reason for separating this initial $\mathcal{O}(N^2)$ step from the rest of the computation is that it is carried out independently of the parameter value, and therefore outside any call to a minimization procedure over the parameter set $\Theta$ involving the expected periodogram.
In the trivial case of a modulation sequence equal to 1 everywhere, then the likelihood of Definition~\ref{def=ourEstimator} does not exactly equal the Whittle likelihood of~\eqref{eq=whittleLKH}. This is because the spectral density $S_X(\omega)$ would be replaced by the expected periodogram $\overline{S}^{(N)}_{\Y}(\omega)$, which is the convolution of the true spectral density with the Fej\'er kernel (see~\eqref{eq=DunsmuirSpectrum}). For stationary time series, this type of estimator was investigated in~\cite{WhittleAdam}, and was found to significantly reduce bias and error in parameter estimation as compared with standard Whittle estimation.
For modulated processes that are asymptotically stationary, this signifies the difference between using~\eqref{eq=DunsmuirSpectrum} and the quantity defined by Proposition~\ref{prop=expectationPeriodogram2} to fit the periodogram.
The same estimator to~\eqref{def=ourEstimator} can be used for the complex-valued time series $\widetilde{Z}_t$ considered in Section \ref{sec=bivariate}, i.e. we define our estimator,
\begin{equation}
\label{eq=complexEstimator}
\hat{\boldsymbol{\theta}}_M\sN =
\arg\min_{\btheta\in\Theta}\ell_M(\boldsymbol{\theta}),
\end{equation}
with the objective function given by,
\begin{equation}
\label{eq=newLKHcomplex}
\ell_M(\boldsymbol{\theta}) =
\frac{1}{N}\sum_{\omega\in\Omega_N}
{
\left\{
\log{\overline{S}\sN_{\widetilde{Z}}(\omega;\boldsymbol{\theta})}
+ \frac{\hat{S}_{\widetilde{Z}}\sN(\omega)}{\overline{S}\sN_{\widetilde{Z}}(\omega;\boldsymbol{\theta})}
\right\}
},
\end{equation}
The comments on computational aspects hold for the complex-valued case as well.
In Section~\ref{sec=consistency}, we will prove consistency of the frequency-domain estimator~\eqref{eq=newLKH} and its $\mathcal{O}(N^{-1/2})$ convergence rate.
\section{Applications}
\subsection{Application to Oceanographic Data}
\label{sec=realdata}
In this section we analyse real-world data from the Global Drifter Program (GDP). Specifically we model jointly the latitudinal and longitudinal velocities obtained from instruments known as drifters, which freely drift according to ocean surface flows~\citep{sykulski2016Lagrangian}.
Those velocities are modelled as the aggregation of two independent complex-valued processes, one of which is nonstationary and which we model as the complex-valued AR(1) process described in Section~\ref{sec=tvAR}. We use our estimator~\eqref{eq=complexEstimator} to infer physical quantities that describe the ocean surface currents.
To scrutinize our results and compare with alternative approaches, we also present two simulation studies, the first one being based on a dynamical model of the ocean surface currents and the second one being a simulated version of the model of Section~\ref{sec=tvAR}.
All data and code used in this paper is available for download at \texttt{http://www.ucl.ac.uk/statistics/research/spg/software} and all results in this section and Section~\ref{sec=missingSims} are exactly reproducible.
\subsubsection{The Global Drifter Program}
\label{sec=GDPapp}
The GDP database (www.aoml.noaa.gov/phod/dac) is a collection of measurements obtained from buoys known as surface drifters, which drift freely with ocean currents and regularly communicate measurements to passing satellites at unequally spaced time intervals averaging 1.4hrs. The data is then interpolated onto a regular temporal grid using the approach of~\citet{elipot2016global}. The measurements include position, and often sea surface and temperature. In total, over 11,000 drifters have been deployed, with approximately 70 million position recordings obtained. The analysis of this data is crucial to our understanding of ocean circulation \citep{lumpkin07}, which is known to play a primary role in determining the global climate system, see e.g.~\citet{andrews2012forcing}. Furthermore, GDP data is used to understand the dispersion characteristics of the ocean, which are critical in correctly modelling oil spills \citep{abascal2010analysis} and more generally assist in developing theoretical understanding of ocean fluid dynamics \citep{griffa2007lagrangian}, which is necessary for global climate modelling.
In Fig.~\ref{Equatorial}(a), we display in the left panel the trajectories of 200 drifters which either traverse or are near the equator, interpolated for this application onto a 2 hour grid from raw position fixes available at the GDP web site. We focus on a single drifter trajectory, drifter ID\#43594, in panels (b) and (c), displaying both its latitudinal position and velocity respectively, the latter of which is obtained by differencing the positions. This velocity time series is nonstationary, as it has oscillations which appear to be modulated and change in frequency over time. The oscillations are known as {\em inertial oscillations}---one of the most ubiquitous and readily observable features of the ocean currents accounting for approximately half of the kinetic energy in the upper ocean \citep{ferrari2009arfm}. Inertial oscillations arise due to the deviation of the rotating earth from a purely spherical geometry, together with the appearance of the Coriolis force in the rotating reference frame of an earth-based observer \citep{early2012forces}. The modulation of these oscillations occurs because the drifters are changing latitude---and the {\em Coriolis frequency}, denoted $f$, is equal to twice the rotation rate of the Earth $\Omega$, multiplied by the sine of the latitude $\zeta$, i.e. $f=2\Omega \sin \zeta$ radians per second. The rotation rate of the Earth $\Omega$ is computed as $2\pi/T$ where $T$ is one sidereal day in seconds. Note that the Coriolis frequency $f$ is a signed quantity, implying that oscillations occur in opposite rotational clockwise/anti-clockwise directions from one hemisphere to the other. The Coriolis frequency is positive in the Northern hemisphere whereas the oscillations occur in the mathematically negative sense.
Therefore we define the inertial frequency $\inerf = -f/2\pi K$ as the negative of the Coriolis frequency divided by $2\pi K$, where $K$ is one solar day in seconds, so that $\inerf$ is in cycles per day. The entire drifter dataset is split into segments of 60 inertial periods in length, accounting for the variation of the inertial period along drifter trajectories, and with 50\% overlap between segments. The standard deviation of the inertial frequency along each data segment is taken, and the 200 segments exhibiting the largest ratio of the standard deviation of the inertial frequency, to the magnitude of its mean value along the segment, are identified for use in this study. These exhibit the largest fractional changes in the inertial frequency, and as shown in Fig.~\ref{Equatorial}(a), are located in the vicinity of the equatorial region where inertial frequency vanishes.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{Arthur2V2.eps}
\caption{
Fitted variance of the discrete Fourier transform using either the stationary model (in black) or the nonstationary model (in red) to the periodogram (in blue) for segments of data from drifter IDs
(a) \#79243, (b) \#54656, (c) \#71845 and (d) \#44312.
The solid black vertical line is the average inertial frequency, and the dashed vertical black lines are the minimum and maximum observed inertial frequency over the observed time window. The models are fit in the frequency range of 0 to 0.8 cycles per day in (a)--(c), and from 0 to 1.5 cycles per day in (d) as this drifter is at a higher latitude of $37^\circ$ S where inertial oscillations occur at a frequency of about 1.2 cycles per day. The fitted models are shown in solid lines within the frequency range, and in dashed lines outside the frequency range.}
\label{DrifterAnalysis}
\end{figure}
\subsubsection{Stochastic modelling}
The stochastic modelling of Lagrangian trajectories was investigated in \citet{sykulski2016Lagrangian}, where the term ``Lagrangian" is used because the moving object making the observations (i.e. the drifter) is the frame of reference, as opposed to fixed-point measurements known as {\em Eulerian} observations. In that paper, the Lagrangian velocity time series was modelled as a complex-valued time series, with the following 6-parameter power spectral density:
\begin{eqnarray}
\label{driftermodel}
S(\omega) &=& \frac{A^2}{(\omega-\inerf)^2+\lambda^2}+\frac{B^2}{\left(\omega^2+h^2\right)^\alpha},\\
\nonumber &&A>0, \;\;\; \lambda>0, \;\;\; \inerf\in[-\pi,\pi], \;\;\; h>0, \;\;\; B>0, \;\;\; \alpha>\frac{1}{2},
\end{eqnarray}
where $\omega$ is given in cycles per day. The first component of~\eqref{driftermodel} is the spectral density of a complex Ornstein-Uhlenbeck (OU) process \citep{Arato1962estimation}, and is used to describe the effect of inertial oscillations at frequency $\inerf$.
Denoting $\widetilde{Z}_{\text{OU}}(t)$ the OU component, where $\widetilde{Z}_{\text{OU}}(t)$ is complex-valued, these oscillations are described by the following stochastic differential equation (SDE),
\begin{equation}
\label{eq=OU_SDE_stationary}
d\widetilde{Z}_{\text{OU}}(t) = (-\lambda + i2\pi\inerf)\widetilde{Z}_{\text{OU}}(t)dt + AdW(t),
\end{equation}
where $t$ is expressed in days and $W(t)$ is a complex-valued Brownian process with independent real and imaginary parts. The damping parameter $\lambda>0$ ensures that the OU process is mean-reverting.
The corresponding continuous complex-valued autocovariance is given by
\begin{equation*}
s(\tau) = \frac{A^2}{2\lambda}\exp\left\{-\lambda|\tau|+i2\pi\omega\tau\right\},
\end{equation*}
and the sampled process \citep{ARATO19991} $\widetilde{Z}_{\text{OU},t} = \widetilde{Z}_{\text{OU}}(t\Delta)$, where $\Delta=1/12$ day is the sampling rate corresponding to the 2hr grid, is a complex-valued AR(1),
\begin{equation}
\label{eq=sampled_stationary_OU}
\widetilde{Z}_{\text{OU},t} = re^{i2\pi\inerf\Delta}\widetilde{Z}_{\text{OU}, {t-1}} + \epsilon_t.
\end{equation}
Here $\epsilon_t$ is a Gaussian complex-valued white noise process with independent real and imaginary parts and variance $\sigma^2$. The autocovariance sequence of the stationary sampled process is given by,
\begin{equation}
c_{\widetilde{Z}_{\text{OU}}} = \frac{\sigma^2}{1-r^2}r^{\tau\Delta}.
\end{equation}
The transformation between the parameters of the complex-valued OU and the complex-valued AR(1) are given by,
\begin{equation}
\label{eq=transform_params_OU_AR}
\sigma^2 = \frac{A^2(1-e^{-\lambda\Delta})}{2\lambda\Delta}, \ \ \ \
r = e^{-\lambda\Delta}.
\end{equation}
The second component of~\eqref{driftermodel} is the spectral density of a stationary proper Mat\'ern process \citep{gneiting2010matern}, denoted $Z_{M,t}$, and is used to describe two-dimensional background turbulence, see \citet{FracBrownianMotionJonathan}.
Although the parameter $\inerf$ is varying as the drifter changes latitude, this parameter is fixed to its mean value in each trajectory segment in \citet{sykulski2016Lagrangian}. This leaves five remaining parameters to estimate, $\{A, \lambda, B, h, \alpha \}$, in different regions of the ocean.
The model of~\eqref{driftermodel} is stationary---slowly-varying nonstationarity in the data is accounted for by windowing the data into chunks of approximately 60 inertial periods, and treating the process as locally-stationary within each window. The estimated parameters can then be aggregated spatially to quantify the heterogeneity of ocean dynamics. This method works well on relatively quiescent and stationary regions of the ocean; however this method cannot account for the rapidly-varying nonstationarity evident in Fig.~\ref{Equatorial}, and leads to model misfit and biased parameter estimates, as we shall now investigate in detail.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{Arthur3up.eps}
\caption{\label{Drifter200Analysis}(a) is a scatter plot of the damping timescale $1/\lambda$ as estimated by the stationary and nonstationary models, for each of the 200 trajectories displayed in Fig.~\ref{Equatorial}; (b) is a histogram of the difference between the log-likelihoods of the nonstationary and stationary models for the same 200 trajectories.}
\end{figure}
\subsubsection{Modulated time series modelling and estimation}\label{ss:modulateddrifters}
We now apply the methodological contributions of this paper to improve the model of~\eqref{driftermodel} for highly nonstationary time series, such as those observed in Fig.~\ref{Equatorial}(a). We do this by accounting for changes in the inertial frequency, $\inerf$, within each window of observation. We denote $\inerf(t)$ the continuous time-varying inertial frequency and $\inerf_t = \inerf(t\Delta)$ the inertial frequency value at each observed time step, $t=0,\cdots,N-1$.
The adapted version of the SDE~\eqref{eq=OU_SDE_stationary} is then given by,
\begin{equation}
\label{eq=OU_SDE_nonstationary}
d\widetilde{Z}_{\text{OU}}(t) = \left(-\lambda + i2\pi\inerf(t)\right)\widetilde{Z}_{\text{OU}}(t)dt + AdW(t).
\end{equation}
In analogue to the proof in \ref{proof=tvarmodulated} it is shown that the sampled process $\widetilde{Z}_{\text{OU},t}=\widetilde{Z}_{\text{OU}}(t\Delta)$ satisfies,
\begin{equation*}
\widetilde{Z}_{\text{OU},t} = re^{i2\pi\int_{\Delta(t-1)}^{\Delta t}{\inerf(u)du}}\widetilde{Z}_{\text{OU},t-1} + \epsilon_t.
\end{equation*}
As the inertial frequency is only observed at sampled points, we approximate the term $\int_{\Delta(t-1)}^{\Delta t}{\inerf(u)du}$ by $\Delta\inerf_t$.
Specifically, we use the model of~\eqref{eq=tvARCdef} for complex-valued time series, i.e.
\begin{equation}
\label{eq=sampled_nonstationary_OU}
\widetilde{Z}_{\text{OU}, t} = r e^{i2\pi\Delta\inerf_t}\widetilde{Z}_{\text{OU}, t-1} + \epsilon_t, \ \ t\geq1,
\end{equation}
where $\epsilon_t$ has the same properties as in~\eqref{eq=sampled_stationary_OU} and the transformation between the parameters $\{A, \lambda, \inerf_t\}$ of the nonstationary complex-valued OU process~\eqref{eq=OU_SDE_nonstationary} and the parameters $\{r, \sigma, \inerf_t\}$ of the nonstationary complex-valued AR(1) process~\eqref{eq=sampled_nonstationary_OU} are given by~\eqref{eq=transform_params_OU_AR}.
The required methodology has been developed in Section \ref{sec=bivariate} for bivariate (or complex-valued) time series.
We only perform the modulation on the complex OU component in~\eqref{driftermodel}; the Mat\'ern component for the turbulent background is unchanged and is considered to be stationary in the window, as it is not in general affected by changes in $\inerf$. The two components are however observed in aggregation, and for this reason we cannot simply demodulate the observed nonstationary signal to recover a stationary signal. Instead, to jointly estimate the parameters $\{A,\lambda,B,h,\alpha\}$,
we first compute the modulating sequence, $g_t$, using~\eqref{eq=tvARcModulatingSeq} in Proposition~\ref{prop=tvarmodulated} and accounting for the temporal sample rate $\Delta$:
\begin{equation}
\label{eq=modulationOU}
g_t = e^{i\sum_{u=1}^t{2\pi\Delta\inerf_u}},
\end{equation}
for $t=0,\cdots,N-1$.
Then we obtain the expected periodogram of the OU component, by computing $c_g(\tau)$ according to~\eqref{eq=cgcomplex}, then $\overline{c}_{\widetilde{Z}}\sN(\tau)$, where we use the autocovariance of a stationary OU process,
\begin{equation*}
c_{Z_{\text{OU}}}(\tau;r, \sigma) = \frac{\sigma^2}{1-r^2}r^\tau,
\end{equation*}
and Fourier transforming according to~\eqref{eq=computationOfPeriodogram}.
Next, we compute the expected periodogram of the stationary Mat\'ern as outlined in~\citet{sykulski2016Lagrangian}. Note that this can also be computed from the autocovariance of a Mat\'ern using~\eqref{eq=computationOfPeriodogram}, by setting $g_t=1$ for all $t$. Finally, we additively combine the expected periodograms, i.e.
\begin{equation*}
\overline{S}\sN(\omega;\btheta) = \sum_{\tau=-(N-1)}^{N-1}{\left[c_g(\tau)c_{Z_{\text{OU}}}(\tau) + \left(1-\frac{|\tau|}{N}\right)c_{Z_{\text{M}}}(\tau)\right]e^{-i\omega\tau}},
\end{equation*}
and then minimize the objective function, given in~\eqref{eq=newLKHcomplex}, to obtain parameter estimates for $\{A,\lambda,B,h,\alpha\}$.
Note that the modulation of a complex-valued AR(1) process by~\eqref{eq=modulationOU} will not lead to an asymptotically stationary process, as in general we cannot expect the quantities $c_g\sN(\tau)$ to converge. However, we can see from Fig.~\ref{Equatorial}(a) that the drifters of our dataset have latitudes comprised between $\pm 20$ degrees. Therefore the terms $2\pi\Delta\inerf_t$ in~\eqref{eq=sampled_nonstationary_OU} are comprised between $\pm 0.3591$ radians, so that the conditions of Proposition~\ref{prop=tvARcCorrelationContribution} are verified. Hence the sampled inertial component is a modulated process with a significant correlation contribution, which justifies the use of our estimator~\eqref{eq=complexEstimator}. Note that this results from the latitudes of the drifters and the sampling rate used.
The assumption of Gaussianity is reasonable for modelling the velocity of instruments from the GDP as is discussed in Section 2.4 of~\citet{LaCasce20081} and references therein. To further inspect this, we tested the Gaussianity of the Fourier transform for the four velocity time series in Fig~\ref{DrifterAnalysis}. Specifically, we compared the theoretical ordered statistics of the exponential distribution to the ordered values of the normalized periodogram (normalized by the expected periodogram of the fitted Gaussian model). These results are not included in the paper for space considerations, however the code to perform this analysis can be found in the online code.
\subsubsection{Parameter estimation with equatorial drifters}
We now compare the likelihood estimates and parameter fits for the stationary model~\eqref{driftermodel}, with those for the nonstationary version of this model described in the previous subsection.
In particular, the damping timescale $1/\lambda$ is of primary interest in oceanography~\citep{elipot2010modification}.
In Fig.~\ref{DrifterAnalysis}, we display the Whittle likelihood fits of each model to segments of data from drifters IDs
\#79243, \#54656 and \#71845,
all of which are among the trajectory segments displayed in the left-hand panel of Fig.~\ref{Equatorial}. We also include model fits to a 60-inertial period window of drifter ID\#44312, which is investigated in detail in~\citet{sykulski2016Lagrangian}, as this South Pacific drifter is from a more quiescent region of the ocean, and does not exhibit significant changes in $\inerf$. For the South Pacific drifter in Fig.~\ref{DrifterAnalysis}(d), both fits are almost equivalent (and hence are overlaid), capturing the sharp peak in inertial oscillations at approx 1.2 cycles per day. For the three equatorial drifters, the stationary model~\eqref{driftermodel} has been fit with the inertial frequency set to the average of $\omega_t^{\{f\}}$ across the window. Here in the first three cases the stationary model is a relatively poor fit to the observed time series spectra. The nonstationary modulated model, which incorporates changes in $\inerf$, is a better fit, capturing the spreading of inertial energy between the maximum and minimum values of $\inerf_t$.
In this analysis, we have excluded frequencies higher than 0.8 cycles per day from all the likelihood fits to the equatorial drifters (the Nyquist is 6 cycles per day for this 2-hourly data), to ignore contamination from tidal energy occurring at 1 cycle per day or higher, which is not part of our stochastic model. Furthermore, we also only fit to the side of the spectrum dominated by inertial oscillations, as the model is not always seen to be a good fit on the other side of the spectrum. The modelling and inference approach is therefore semi-parametric \citep{robinson1995gaussian}.
The significance of the misfit of the stationary model is that parameters of the model may be under- or over-estimated as the model attempts to compensate for the misfit. For example, the damping parameter of the inertial oscillations, $\lambda$, will likely be overestimated in the stationary model, as it is used to try to capture the spread of energy around $\inerf$, which is in fact mostly caused by the changing value of $\inerf$, rather than a true high value of $\lambda$.
To investigate this further, we perform the analysis with all 200 drifters shown in Fig.~\ref{Equatorial}. In Fig.~\ref{Drifter200Analysis}(a), we show a scatter plot of the estimates of $1/\lambda$, known as the damping timescale, as estimated by both models. In general, the damping timescales are larger with the nonstationary model (consistent with a smaller $\lambda$), where the median value is 3.42 days, rather than 1.3 days with the stationary model. Previous estimates of the damping timescale in the literature have not included data from the equatorial region, so while direct comparisons are not possible, the former estimates are found to be more consistent with previous estimates at higher latitudes where values of around 3 days are reported in \cite{elipot2010modification}, and values ranging from 2 to 10 days are reported in \cite{watanabe2002global}.
The nonstationary model does not require more parameters to be fitted than the stationary model; both have 5 unknown parameters. Therefore there is no need to penalize the nonstationary model using model choice or likelihood ratio tests.
Even though the models are not nested, comparing the likelihood of the two approaches can be informative.
We can directly compare the likelihood value of each model using~\eqref{eq=whittleLKH} and \eqref{eq=newLKH}, i.e. $\ell_M(\bm{\hat\theta}_M)-\ell_W(\bm{\hat\theta}_W)$. A histogram of the difference between the likelihoods for the 200 drifters is shown in Fig.~\ref{Drifter200Analysis}(b), where positive values indicate that the likelihood of the nonstationary model is higher. Overall, the nonstationary model has a higher likelihood in 146 out of the 200 trajectories and is therefore seen to be the better model in general.
There are other regions of the global oceans, in addition to the equator, where the nonstationary methods of this paper may significantly improve parameter estimates of drifter time series. These include drifters which follow currents that traverse across different latitudes, such as the Gulf Stream or the Kuroshio. Analysis of such data is an important avenue of future investigation.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{Arthur4.eps}
\caption{\label{NumericalDrifters}Trajectories of 9 particles from the dynamical model, with the damping timescale set to 4 days. All particle trajectories are started at 35$^\circ$~N and 40$^\circ$~W with increasing meridional mean flow from $V=0.1$ to $V=0.9$ cm/s going from left to right ($u$ is set to zero for this example). The drifters are offset in longitude by 0.02 degrees for representation.}
\end{figure}
\subsubsection{Testing with Numerical Model Output}
\label{sec=dynamicalModel}
In this section we test the accuracy of the nonstationary modelling and parameter estimation for drifters by analysing output in a controlled setting using a dynamical model for inertial oscillations. The model propagates particles on an ocean surface forced by winds---simulated white noise in our simulations---with a fixed damping parameter, similar to the damped-slab model of~\citet{pollard1970dsr}, but uses the correct spherical dynamics for Earth from \citet{early2012forces}, so that the oscillations occur at the correct Coriolis frequency given the particle's latitude and the model remains valid at the equator. The damping timescale parameter is fixed globally {\em a priori} in the model and the goal is to see if it can be accurately estimated using parametric time series models.
The numerical model is constructed such that the particle can also be given a linear mean flow, $U+iV$. If this mean flow has a significant vertical component $V$, then the particle will cross different latitudes and the frequency of inertial oscillations will significantly change over a single analysis window.
We display particle trajectories from the dynamical model in Fig.~\ref{NumericalDrifters}, with various realistic mean flow values, where the spherical dynamics can clearly be seen for larger latitudinal mean flow values. We observe that the particles subject to small mean flows display stationary oscillation patterns, whereas for the particles with a large latitudinal mean flow, the oscillation frequency appears to diminish as the particle approaches latitude zero.
A more complete description of the numerical model is available in the online code.
To explore the performance of the estimation of damping time-scales, we assess the performance of the parameter estimates of our nonstationary model, by performing a Monte Carlo study based on the dynamical model described in the previous paragraph. We generate 100 trajectories, each of length 60 days and sampled every 2 hours, for a given damping timescale ($1/\lambda$) and latitudinal mean flow ($V$). We estimate the damping parameter using the stationary and nonstationary methods, in exactly the same way as with the real-world drifter data, and average the estimated damping timescales $1/\lambda$ over the 100 time series. We note that as this model has no background turbulence, then we set $B=0$ in~\eqref{driftermodel} such that there is no Mat\'ern component present.
We then repeat this analysis over a range of realistic values for $1/\lambda$ and $V$. The average estimates of $1/\lambda$ are reported in Fig.~\ref{NumericalMatrix}. The stationary method breaks down for large mean flows and long damping timescales, with large overestimates of $\lambda$. The nonstationary method performs well across the entire range of values. We note that long damping timescales are generally harder to estimate, as $\lambda$ becomes close to zero and is estimated over relatively fewer frequencies. We have not reported mean square errors here for space considerations, but we found the parameter biases to be the main contribution to the errors, so it follows that the nonstationary method remains strongly preferable.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{Arthur5.eps}
\caption{\label{NumericalMatrix}Mean estimates of the damping timsescale $1/\lambda$ with (a) the stationary model of~\eqref{driftermodel} and (b) the nonstationary model of Section~\ref{ss:modulateddrifters}, applied to 100 realizations of the dynamical model described in Section~\ref{sec=dynamicalModel}. The experiment is performed over a grid of meridional mean flow values $v$ from 0 to 0.9 cm/s, and over a range of true damping timescales $1/\lambda$ from 1 to 8 days. The estimated damping timescale values, averaged over 100 repeat experiments, is written in each cell and shaded according to the colorbar.}
\end{figure}
\subsubsection{Testing with Stochastic Model Output}
In this section we test with purely stochastic output, which allows us to extensively compare biases, errors and computational times of the stationary and nonstationary methods in a much larger Monte Carlo study. We continue using the bivariate model of~\eqref{eq=tvARCdef} which is suitable for inertial oscillations, except this time we change $\beta_t$ according to a stochastic process. Specifically, we set as our generative mechanism for the frequencies $\beta_t$,
\begin{eqnarray}\label{eq:thetarandom}
\beta_0 &=& \mathcal{D}(\gamma + A\epsilon_t)\\
\beta_t &=& \mathcal{D}(\beta_{t-1} + A\epsilon_t),
\end{eqnarray}
where $\gamma\in[-\pi, \pi)$, $A>0$, $\epsilon_t$ is a standard normal white noise, and $\mathcal{D}(\cdot)$ is the bounding function defined by
\begin{equation}\label{eq:thetarandom2}
\mathcal{D}(x) = \max\{\min(x, \gamma+\Delta), \gamma-\Delta\},
\end{equation}
where $\Delta>0$, and this choice of $\mathcal{D}(x)$ constrains $\beta_t$ in the interval $[\gamma-\Delta,\gamma+\Delta]$.
This way the frequencies $\beta_t$ are generated according to a bounded random walk, i.e. a random walk which is constrained to stay within a fixed bounded interval.
According to Proposition \ref{prop=tvARcCorrelationContribution}, if $\Delta$ is smaller than $\pi/2$, then this ensures that the modulated process belongs to the class of modulated processes with a significant correlation contribution, and our estimator~\eqref{eq=newLKHcomplex} is consistent.
In our simulations we have set $\gamma=\pi/2$, $\Delta=1$, $A=1/20$.
We simulate for a range of sample sizes ranging from $N=128$ to $N=4096$. For each sample size $N$, we independently simulate 2000 time series and estimate $\{r,\sigma\}$ for each series to report ensemble-averaged biases, errors, and computational times. The results are reported in Table~\ref{table1}. The bias and Mean Square Error (MSE) of the estimated parameters with the stationary method are seen to increase with increasing sample size. This is because the random walk of $\beta_t$ increases the range of $\beta_t$ with larger $N$, such that the nonstationarity of the time series is increasing. Conversely, the nonstationary method accounts for these rapidly changing modulating frequencies, and the bias and MSE of parameter estimates rapidly decrease with increasing $N$. The average CPU time is only around 5\% slower using the nonstationary method, as the method is still $\mathcal{O}(N\log N)$ in computational efficiency.
\begin{table}[t]
\centering
\caption{\label{table1}Performance of estimators with the stationary and nonstationary methods for the model of~\eqref{eq=tvARCdef} with $\beta_t$ evolving according to the bounded random walk described by~\eqref{eq:thetarandom}--\eqref{eq:thetarandom2}. The parameters are set as $r=0.8$, $\sigma=1$, $\gamma=\pi/2$, $\Delta=1$, and $A=1/20$. The results are averaged over 2000 independently generated time series for each sample size $N$.
The average CPU times for the optimization are given in seconds, as performed on a 2.40Ghz Intel i7-4700MQ processor (4 cores).
}\vspace{3mm}
\begin{tabular}{lcccccc}
Sample size ($N$) & 128 & 256 & 512 & 1024 & 2048 & 4096\\
\hline \hline
\multicolumn{7}{c}{Stationary frequency domain likelihood}\\
\hline
Bias ($r$) & -2.3481e-02 & -3.2400e-02 & -4.8112e-02 & -6.9807e-02 & -9.3332e-02 & -1.1161e-01\\
Variance ($r$) & 1.8163e-03 & 1.0760e-03 & 1.1422e-03 & 1.5550e-03 & 1.4045e-03 & 8.2890e-04\\
MSE ($r$) & 2.3677e-03 & 2.1258e-03 & 3.4570e-03 & 6.4280e-03 & 1.0115e-02 & 1.3286e-02\\
\hline
Bias ($\sigma$) & 2.5577e-02 & 5.4988e-02 & 8.9480e-02 & 1.3241e-01 & 1.7432e-01 & 2.0651e-01\\
Variance ($\sigma$) & 3.3898e-03 & 2.8178e-03 & 3.3471e-03 & 4.4660e-03 & 3.9885e-03 & 2.1609e-03\\
MSE ($\sigma$) & 4.0440e-03 & 5.8415e-03 & 1.1354e-02 & 2.1999e-02 & 3.4376e-02 & 4.4809e-02\\
\hline
CPU time (sec) & 1.3083e-02 & 1.7776e-02 & 2.5743e-02 & 4.3666e-02 & 5.0948e-02 & 8.6940e-02 \\ \hline\hline
\multicolumn{7}{c}{Nonstationary frequency domain likelihood}\\
\hline
Bias ($r$) & -4.6158e-03 & -2.0129e-03 & -1.4184e-03 & -2.9047e-04 & -2.6959e-04 & 8.8302e-05\\
Variance ($r$) & 1.6508e-03 & 7.5379e-04 & 3.9819e-04 & 2.0710e-04 & 1.0674e-04 & 5.3236e-05\\
MSE ($r$) & 1.6721e-03 & 7.5784e-04 & 4.0020e-04 & 2.0719e-04 & 1.0681e-04 & 5.3244e-05\\
\hline
Bias ($\sigma$) & -1.4999e-02 & -8.8581e-03 & -4.4302e-03 & -2.5292e-03 & -1.4125e-03 & -9.1703e-04\\
Variance ($\sigma$) & 2.2543e-03 & 1.1989e-03 & 6.4245e-04 & 3.4775e-04 & 2.0113e-04 & 1.0759e-04\\
MSE ($\sigma$) & 2.4793e-03 & 1.2774e-03 & 6.6208e-04 & 3.5415e-04 & 2.0312e-04 & 1.0843e-04\\
\hline
CPU time (sec) & 1.6814e-02 & 2.0272e-02 & 3.1397e-02 & 5.5925e-02 & 8.9997e-02 & 2.4147e-01 \\ \hline\hline
\end{tabular}
\end{table}
Finally, we consider the case in which the modulating sequence is only unknown up to a functional form, and we must also estimate its parameters, along with the parameters of the latent process. We consider the following parametric form for $\beta_t$
\begin{equation}
\label{eq=freqParameterModel}
\beta_t = \gamma + \Delta\frac{2t-(N-1)}{2(N-1)},
\end{equation}
with parameters $\gamma\in[-\pi,\pi)$ and $0<\Delta<\pi$. The upper bound for $\Delta$ is chosen so that the resulting modulated process satisfies the assumptions of Proposition~\ref{prop=tvARcCorrelationContribution}. The modulated process then has a significant correlation contribution.
Therefore $\beta_t$ varies linearly from $\gamma-\frac{\Delta}{2}$ to $\gamma+\frac{\Delta}{2}$. We can then show that for all integer value $\tau$,
\begin{equation}
c_g\sN(\tau) = \frac{\sin\left[\frac{\Delta \tau}{2(N-1)}(N-\tau)\right]}{N\sin\left[\frac{\Delta \tau}{2(N-1)}\right]}e^{\left\{i(\gamma \tau+\frac{\Delta \tau}{2(N-1)}\right\}}.
\end{equation}
This allows the kernel in~\eqref{eq=expectedAutocovSequence1} to be precomputed in $\mathcal{O}(N)$ elementary operations for all values of $\tau=0,\cdots,N-1$. This helps to speed up the computation of the expected periodogram in the likelihood for the special case of a linearly varying $\beta_t$. In this problem we have to estimate $\{\gamma,\Delta\}$ from $\beta_t$ as well as $\{r,\sigma\}$ from $Z_t$. We perform a Monte Carlo simulation with a fixed sample size of $N=512$, where we simulate 5,000 independent time series each with parameters set to
$r=0.9$, $\sigma=10$, $\gamma=0.8$, and $\Delta=1$. We report the biases, variances and MSEs with the stationary and nonstationary methods in Table~\ref{table2}. As the stochastic process is Markovian, it is also possible to implement exact maximum likelihood in $\mathcal{O}(N)$ elementary operations for this specific problem, and we report these values in the table also. Our nonstationary inference method performs relatively close to that of exact maximum likelihood, despite the challenge of having to estimate parameters of the modulating sequence, as well as the latent process. The stationary method performs poorly, as with previous examples, as stationary modelling is not appropriate for such rapidly-varying oscillatory structure.
\begin{table}[t]
\centering
\caption{\label{table2}Performance of estimators with the stationary and nonstationary methods for the model of~\eqref{eq=tvARCdef} with $\beta_t$ evolving according to~\eqref{eq=freqParameterModel}. The parameters are set as $r=0.9$, $\sigma=10$, $\gamma=0.8$, and $\Delta=1$. The results are averaged over 5000 independently generated time series for each sample size $N$. N/A stands for Not Applicable.}\vspace{3mm}
\begin{tabular}{lcccc}
Estimated parameter & $r$ & $\sigma$ & $\gamma$ & $\Delta$\\
\hline \hline
\multicolumn{5}{c}{Exact likelihood}\\
\hline
Bias & -1.3244e-03 & 2.1063e-02 & 2.1192e-03 & -3.6145e-03\\
Variance & 1.8668e-04 & 5.2130e-02 & 2.3725e-04 & 2.8323e-03\\
MSE & 1.8844e-04 & 5.2574e-02 & 2.4174e-04 & 2.8454e-03 \\ \hline\hline
\multicolumn{5}{c}{Stationary frequency domain likelihood}\\
\hline
Bias & -1.5392e-01 & 5.2092e+00 & 2.5871e-03 & N/A\\
Variance & 5.5052e-04 & 8.6907e-01 & 9.4628e-03& N/A\\
MSE & 2.4241e-02 & 2.8005e+01 & 9.4695e-03 & N/A \\ \hline\hline\multicolumn{5}{c}{Nonstationary frequency domain likelihood}\\
\hline
Bias & -1.7074e-03 & 6.8215e-03 & 1.1434e-03 & -3.7092e-02\\
Variance & 2.3975e-04 & 1.5285e-01 & 2.0803e-03 & 1.6425e-02\\
MSE & 2.4266e-04 & 1.5290e-01 & 2.0816e-03 & 1.7801e-02 \\ \hline\hline
\end{tabular}
\end{table}
\subsection{Missing data simulation}
\label{sec=missingSims}
In this section we show that the estimator defined in Definition \ref{def=ourEstimator} can be used for the random missing data scheme~\ref{ex=missingDataScheme} of Section \ref{sec=missing}.
Therefore we simulate a real-valued first order autoregressive process with parameters $0\leq a <1$ and $\sigma$ according to
\begin{equation}\label{eq:missing1}
X_t =a X_{t-1} + \epsilon_t, \ \ t\geq 1,
\end{equation}
where $X_0 \sim \mathcal{N}\left[0,\sigma^2/(1-a^2)\right]$, and $\epsilon_t$ is a Gaussian white noise process with mean zero and variance $\sigma^2$. The process $\{X_t\}$ is the latent process of interest. To account for the missing data, we generate a modulated time series $\Y_t=g_t X_t$ and assume we only observe the time series $\{\Y_t\}$, from which we estimate the parameters of the process $\{X_t\}$. The sequence $\{g_t\}$ takes its values in the set $\{0,1\}$ and is generated according to
\begin{equation*}
g_t \sim \mathcal{B}(p_t),
\end{equation*}
where $\mathcal{B}(p)$ represents the Bernoulli distribution with parameter $p$, and where we set
\begin{equation*}\label{eq:missing2}
p_t = \frac{1}{2} + \frac{1}{4}\cos\left(\frac{2\pi}{10}t\right).
\end{equation*}
The observed modulating sequence $\{g_t\}$, made of zeros and ones, is clearly nonstationary as it does not admit a constant expectation. Therefore a spectral representation of the second order structure of the random modulating sequence $\{g_t\}$, as required in \citet{Dunsmuir1981b}, does not exist.
We simulate and estimate such a model for different sample sizes ranging from $N=128$ to $N=16384$. For each value of $N$, we independently simulate 2000 time series and for each time series we estimate $\{a,\sigma\}$. The outcomes of our simulation study are reported in Table~\ref{table3}. The bias, variance and mean square error rapidly decrease with increasing $N$, while the computational time only increases gradually with $N$ such that the methods are still computationally efficient for long time series. Comparing our technique with other methods from the literature is the subject of ongoing work.
\begin{table}[h]\small \caption{\label{table3}Performance of our estimator for the missing data problem defined in~\eqref{eq:missing1}--\eqref{eq:missing2}. The unknown parameters are set as $a=0.8$, and $\sigma=1$. The results are averaged over 2000 independently generated time series for each sample size $N$.
The average CPU times for the optimization are given in seconds, as performed on a 2.40Ghz Intel i7-4700MQ processor (4 cores).
}\vspace{3mm}
\centering
\begin{tabular}{lccccccc}
Sample size & 128 & 512 & 1024 & 2048 & 4096 & 8192 & 16384\\
\hline\hline
\multicolumn{8}{c}{Estimate of parameter $a$}\\
\hline
Bias & -2.0805e-02 & -4.8097e-03 & -3.0920e-04 & -4.1710e-04 & -2.7349e-04 & -4.9356e-04 & -1.5213e-04\\
Variance & 1.0721e-02 & 2.7114e-03 & 1.2795e-03 & 6.3100e-04 & 3.0883e-04 & 1.4380e-04 & 7.1010e-05\\
MSE & 1.1154e-02 & 2.7346e-03 & 1.2796e-03 & 6.3117e-04 & 3.0891e-04 & 1.4404e-04 & 7.1034e-05\\
\hline
\multicolumn{8}{c}{Estimate of parameter $\sigma$}\\
\hline
Bias & -1.7136e-02 & -5.4872e-03 & -6.8949e-03 & -2.7876e-03 & -1.3061e-03 & 2.2880e-04 & -1.7434e-04\\
Variance & 3.3705e-02 & 8.7577e-03 & 4.1674e-03 & 1.9408e-03 & 9.7955e-04 & 4.1524e-04 & 2.2691e-04\\
MSE & 3.3999e-02 & 8.7878e-03 & 4.2150e-03 & 1.9486e-03 & 9.8125e-04 & 4.1529e-04 & 2.2694e-04\\
\hline
\multicolumn{8}{c}{Computational time}\\
\hline
CPU time (s) & 1.7901e-02 & 3.9687e-02 & 7.0515e-02 & 8.2408e-02 & 1.7624e-01 & 4.1774e-01 & 1.2138e+00\\
\hline\hline
\end{tabular}
\end{table}
\section{Consistency}
\label{sec=consistency}
In this section we show in Theorem~\ref{theorem=consistency} that the frequency domain estimator $\hat{\btheta}_M\sN$ (which for simplicity we denote $\hat{\btheta}\sN$ in this section) is consistent in the univariate real-valued case (extension to our class of bivariate processes follows directly). In Theorem~\ref{theorem=convergencerate} we show that this estimator converges with a $\mathcal{O}(N^{-\frac{1}{2}})$ rate.
To guarantee consistency we require the following assumptions to be satisfied:
\begin{enumerate}
\item The parameter set $\Theta\subset\R^d$ is compact with a non-null interior, and the true parameter $\theta$ lies in the interior of $\Theta$.
\item\label{assumption=1} Assume that for all $\btheta\in\Theta$, we have $\sum_{\tau\in\N}{\left |c_X(\tau;\btheta)\right|}<\infty$ (short memory) and that the functions $\theta\rightarrow c_X(\tau;\btheta)$ are continuous with respect to $\btheta$. It follows that the spectral densities are also continuous with respect to $\btheta$. We also assume that for all $\btheta\in\Theta$ and $\omega\in[-\pi,\pi]$, $S_X(\omega;\btheta)>0$. By continuity on a compact set the spectral densities $S_X(\omega;\btheta)$ are therefore bounded below in both variables by a non-zero value. For the same reason they are bounded above.
\item We assume that the spectral densities are continuously differentiable with respect to $\omega$. By continuity on a compact set the derivatives with respect to $\omega$ are bounded above, independently of $\btheta$.
\item The process $\Y_t$ is a modulated process with significant correlation contribution. We recall that this implies the existence of a finite subset $\Gamma\subset\N$ such that the mapping $\btheta\mapsto\left\{c_X(\tau):\tau\in\Gamma\right\}$ is one-to-one. We also assume that the modulating sequence $\{g_t\}$ is bounded in absolute value by some finite constant $g_{\max}>0$.
\end{enumerate}
We start with the following two lemmas which yield uniform bounds of the expected periodogram and its derivative.
\begin{lemma}[Boundedness of the expected periodogram]
\label{lemma=boundexpectedperiodogram}
For all $\btheta\in\Theta$ and $N\in\N$, the expected periodogram $\overline{S}_{\Y}\sN(\omega;\btheta)$ is bounded below (by a positive real number) and above independently of $N$ and $\btheta$. We denote these bounds $\overline{S}_{\Y,\min}$ and $\overline{S}_{\Y,\max}$ respectively.
\end{lemma}
\begin{proof}
See Appendix \ref{proof=boundDerivative}.
\end{proof}
\begin{lemma}[Boundedness of the derivative of the expected periodogram]
\label{lemma=totalvariationnorm}
The derivative of the expected periodogram with respect to $\omega$ exists and is bounded in absolute value independently of $\btheta$ and $N$.
\end{lemma}
\begin{proof}
See appendix \ref{proof=totalvariationnorm}.
\end{proof}
In analogue to \citet{Taniguchi} for stationary processes, we introduce the following quantity,
\begin{equation*}
D\sN\left(\bgamma, f\right) = \frac{1}{N}\sum_{\omega\in\Omega_N}\left\{ \log \overline{S}_{\Y}\sN(\omega;\bgamma) + \frac{f(\omega)}{\overline{S}_{\Y}\sN(\omega;\bgamma)} \right\},
\end{equation*}
for all positive integer $N$, $\bgamma\in\Theta$ and non-negative real-valued function $f$ defined on $\Omega_N$.
We also define
\begin{equation*}
T\sN(f) = \arg\min_{\bgamma\in\Theta}D\sN\left(\bgamma, f\right).
\end{equation*}
This minimum for fixed $f$ is well defined since the set $\Theta$ is compact and since the function $\bgamma\mapsto D\sN\left(\bgamma, f\right)$ is continuous. However in cases where the minimum is reached not uniquely but at multiple parameter values,
$T\sN(f)$ will denote any of these values, chosen arbitrarily. Note that, by the definition of our frequency domain estimator, we have $\hat{\btheta}\sN = T\sN\left(\hat{S}_{\Y}\sN(\cdot)\right)$.
We start with three lemmas that will be required in proving Theorem \ref{theorem=consistency} which establishes consistency.
\begin{lemma}
\label{lemma=uniquenessOfMin}
We have, for $N$ large enough, $T\sN(\overline{S}_{\Y}\sN(\omega;\btheta)) = \btheta$, uniquely.
\end{lemma}
\begin{proof}
See Appendix~\ref{proof=uniquenessOfMin}.
\end{proof}
This shows that for all $N$ large enough, the function $\bgamma\rightarrow D\left(\bgamma, \overline{S}_{\Y}\sN(\cdot;\btheta)\right)$ reaches a global minimum at the true parameter vector $\btheta$. However because $\overline{S}_{\Y}\sN(\cdot;\btheta)$ is changing with $N$ and is not expected to converge to a given function, we need the following stronger result.
\begin{lemma}
\label{lemma=minimalValues}
If $\left\{\bgamma_N\right\}_{N\in\N}\in\Theta^\N$ is a sequence of parameter vectors such that $D\left(\bgamma_N, \overline{S}_{\Y}\sN(\cdot;\btheta)\right)-D\left(\btheta, \overline{S}_{\Y}\sN(\cdot;\btheta)\right)$ converges to zero when $N$ goes to infinity, then $\bgamma_N$ converges to $\btheta$.
\end{lemma}
\begin{proof}
See Appendix~\ref{proof=minimalValues}.
\end{proof}
We now show that the functions $D\left(\bgamma, \overline{S}_{\Y}\sN(\cdot;\btheta)\right)$ and $D\left(\bgamma, \hat{S}_{\Y}\sN(\cdot)\right)$, defined on $\Theta$, behave asymptotically \emph{in the same way}. For this, we first need the following lemma where we bound the asymptotic variance of some linear functionals of the periodogram.
\begin{lemma}
\label{lemma=boundOnVariance}
Let $\left\{a\sN(\omega): \omega\in[-\pi,\pi)\right\}_{N\in\N}$ be a family of real-valued functions, uniformly bounded by a positive real number. We have
\begin{equation*}
\var\left\{\frac{1}{N} \sum_{\omega\in\Omega_N}{a\sN(\omega)\hat{S}_{\Y}\sN(\omega)}\right\} = \mathcal{O}\left(\frac{1}{N}\right).
\end{equation*}
\end{lemma}
\begin{proof}
See Appendix \ref{proof=boundOnVariance}
\end{proof}
Remembering that $\overline{S}_{\Y}\sN(\omega;\btheta) = \E\left\{\hat{S}_{\Y}\sN(\omega);\btheta\right\}$, we thus have
\begin{equation}
\nonumber\sum_{\omega\in\Omega_N}{a\sN(\omega)\hat{S}_{\Y}\sN(\omega)}
= \sum_{\omega\in\Omega_N}{a\sN(\omega)\overline{S}_{\Y}\sN(\omega;\btheta)} + \mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right).
\end{equation}
We are now able to state a consistency theorem for our estimator $\hat{\btheta}\sN$.
\begin{theorem}[Consistency of the frequency domain estimator]
\label{theorem=consistency}
We have $\hat{\btheta}\sN \overset{P}{\longrightarrow} \btheta$ in probability.
\end{theorem}
\begin{proof}
The proof is based on~\citet{Taniguchi}. Denote $\overline{h}\sN(\bgamma;\btheta) = D\left(\bgamma, \overline{S}_{\Y}\sN(\omega;\btheta)\right)$ and
$\hat{h}\sN(\bgamma) = D\left(\bgamma, \hat{S}_{\Y}\sN(\omega)\right)$
defined for any $\bgamma\in\Theta$. We have,
\begin{eqnarray*}
\overline{h}\sN(\bgamma;\btheta) - \hat{h}\sN(\bgamma) &=&
\frac{1}{N}\sum_{\omega\in\Omega_N}\left\{\log{\overline{S}_{\Y}\sN(\omega;\bgamma)}+\frac{\overline{S}_{\Y}\sN(\omega;\btheta)}{\overline{S}_{\Y}\sN(\omega;\bgamma)}-\log{\overline{S}_{\Y}\sN(\omega;\bgamma)}-\frac{\hat{S}_{\Y}\sN(\omega)}{\overline{S}_{\Y}\sN(\omega;\bgamma)}\right\}\\
&=&\frac{1}{N}\sum_{\omega\in\Omega_N}{\frac{\overline{S}_{\Y}\sN(\omega;\btheta)-\hat{S}_{\Y}\sN(\omega)}{\overline{S}_{\Y}\sN(\omega;\bgamma)}}.
\end{eqnarray*}
We have shown in lemma \ref{lemma=boundexpectedperiodogram} that $\overline{S}_{\Y}\sN(\omega;\bgamma)$ is bounded below in both variables $\omega$ and $\gamma$ by a positive real number, independently of $N$. Therefore, making use of lemma~\ref{lemma=boundOnVariance} we have
\begin{equation}
\label{eq=supgoestozero}
\sup_{\gamma\in\Omega}\left|\overline{h}\sN(\bgamma;\btheta) - \hat{h}\sN(\bgamma)\right| \stackrel{P}{\longrightarrow} 0, \ \ (N\rightarrow\infty),
\end{equation}
where the letter \emph{P} indicates that the convergence is in probability, as the difference is of stochastic order $N^{-\frac{1}{2}}$.
In particular~\eqref{eq=supgoestozero} implies that
\[
\left|\min_\gamma \overline{h}\sN(\bgamma;\btheta) - \min_\gamma \hat{h}\sN(\bgamma)\right| \leq \sup_{\gamma\in\Omega}\left|\overline{h}\sN(\bgamma;\btheta) - \hat{h}\sN(\bgamma)\right| \stackrel{P}{\longrightarrow}0
\]
i.e.
\begin{equation}
\label{eq=cvgzeroP1}
\left|\overline{h}\sN\left(T\sN(\overline{S}_{\Y}\sN(\omega;\btheta));\btheta\right) -\hat{h}\sN\left(T\sN(\hat{S}_{\Y}\sN(\omega))\right) \right| \stackrel{P}{\longrightarrow} 0.
\end{equation}
Relation (\ref{eq=supgoestozero}) also implies that
\begin{equation}
\label{eq=cvgzeroP2}
\left| \overline{h}\sN\left(T\sN(\hat{S}_{\Y}\sN(\omega));\btheta\right)-\hat{h}\sN\left(T\sN(\hat{S}_{\Y}\sN(\omega))\right) \right| \stackrel{P}{\longrightarrow} 0 ,
\end{equation}
so that using the triangle inequality,~\eqref{eq=cvgzeroP1} and~\eqref{eq=cvgzeroP2}, we get,
\begin{equation*}
\left| \overline{h}\sN\left(T\sN(\hat{S}_{\Y}\sN(\omega));\btheta\right) - \overline{h}\sN\left(T\sN(\overline{S}_{\Y}\sN(\omega;\btheta));\btheta\right) \right| \stackrel{P}{\longrightarrow} 0.
\end{equation*}
We then obtain the stated theorem making use of Lemma \ref{lemma=minimalValues}.
\end{proof}
We now study the convergence rate of our frequency domain estimator. For this we first need the following two lemmas. Although the Hessian matrix of the likelihood
is not expected to converge for modulated processes with a significant correlation contribution, we can show that its norm is bounded below
by a positive real number.
For this we need to strengthen the assumption of significant correlation contribution. Assuming that the spectral densities of the latent process are twice continuously differentiable with respect to $\btheta$, we assume that the Jacobian determinant of the mapping $\btheta\mapsto \left[c_X(\tau;\btheta):\tau\in\Gamma\right]^T$ taken at the true parameter value $\btheta$, i.e. the determinant of the matrix with elements $\frac{\partial c_X(\tau_i;\btheta)}{\partial\btheta_j}$ (with $\Gamma = \{\tau_1, \tau_2, \cdots, \tau_d\}$ here), is non-zero.
\begin{lemma}
\label{lemma=boundbelowC}
Let $\bold{U}_1, \cdots, \bold{U}_d$ a family of vectors of $\R^d$ with rank $d$. Let $\alpha_1, \cdots, \alpha_d$ be positive real numbers. There exists a positive constant $C>0$ such that for all $\bold{V}\in\R^d$,
\begin{equation}
\sum_{i=1}^d{\alpha_i^2\left(\bold{U}_i^T\bold{V}\right)^2} \geq C \left\|\bold{V}\right\|_2^2,
\end{equation}
where $\|\cdot\|_2$ denotes the Euclidean norm on $\R^N$.
\end{lemma}
\begin{proof}
See Appendix~\ref{proof=boundbelowC}.
\end{proof}
\begin{lemma}
\label{lemma=7}
We have,
\begin{equation}
\frac{\partial l_M\sN}{\partial\theta_i}(\btheta) = \mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right).
\end{equation}
The Hessian matrix of the function $l_M(\btheta)$ satisfies
\begin{equation}
H(\btheta) = \mathcal{I}(\btheta) + \mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right),
\end{equation}
where the matrix norm of $\mathcal{I}(\btheta)$ is bounded below by a positive value, independently of $N$.
\end{lemma}
\begin{proof}
See Appendix~\ref{proof=7}.
\end{proof}
\begin{theorem}[Convergence rate]
\label{theorem=convergencerate}
We have $\hat{\btheta} = \btheta + \mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right)$.
\end{theorem}
\begin{proof}
We have, by Taylor expansion with Lagrange form of the remainder term,
\begin{equation*}
\nabla l_M(\hat{\btheta}) = \bold{0} = \nabla l_M(\btheta) + H(\widetilde{\btheta})(\hat{\btheta}-\btheta),
\end{equation*}
where $\widetilde{\btheta}$ lies between $\hat{\btheta}$ and $\btheta$. Therefore,
\begin{equation}
\hat{\btheta}-\btheta = - H(\widetilde{\btheta})^{-1}\nabla l_M(\btheta).
\end{equation}
We have shown that $\hat{\btheta}$ converges in probability to $\btheta$. By continuity of the Hessian, and using the results of Lemma~\ref{lemma=7}, we obtain
\begin{equation}
\hat{\btheta}-\btheta = - \left[\mathcal{I}(\btheta) + \mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right) + o_P(1)\right]^{-1}\mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right) = \mathcal{O}_P\left(\frac{1}{\sqrt{N}}\right).
\end{equation}
This concludes the proof.
\end{proof}
\section{Conclusion}
The well-established theory for the analysis of stationary time series is often in contradiction with real-world data applications. This is because most real time series are nonstationary. Nonstationary observations have required statisticians to develop new models and more broadly new generating mechanisms. Among the large class of nonstationary models, uniform modulation of time series is an easy way to create nonstationarity, and presents all the advantages of a simple mechanism for the time-varying second order structure of a process.
Modulation has already been used to account for missing data when analysing stationary time series, as well as gentle time variation. In fact, if the modulation is slow, regular theory for locally stationary time series applies. Despite its popularity as a modelling tool, the concept of modulation, when variation can be moderate to rapid, is very poorly understood. We have in this paper shown how modulation of time series can account for much more rapid changes of a time series model.~\citet{JTSA:JTSA12034} already abandoned the assumption of smoothness in time of the time-varying spectral density~\citep{Dahlhaus1997}. However, the class of modulated processes with a significant correlation contribution is one of few instances of nonstationary models where more data in time results in more accurate estimates, and asymptotic consistency under standard assumptions for the latent stationary process.
As we have generalized modulation beyond the assumptions where it is known that models can be estimated, the question naturally arises, as to what types of modulation still permit parameter estimation.
Key to our understanding of modulated processes is the definition of modulated processes with a significant correlation contribution (see Definition \ref{def=univariateSignificantCorrel}), which generalizes the classical concept of asymptotically stationary modulated processes, and corresponds to our main modelling innovation. We require that the sample autocorrelations of the modulating sequences be asymptotically bounded below, so that the information in the autocorrelation of the process does not fade in the observed process. The interpretation of this requirement is that there must be sufficient support in the autocorrelation to retain the information in the modulation.
With this new model class we can implement estimation directly in the Fourier domain, directly after transforming the data from the temporal domain. Estimation is still possible in ${\cal O}\left(N\log(N)\right)$ computational effort, and the further required conditions to ensure consistency were studied.
Most real-world data sets are aggregations of heterogeneous components. To fully show the promise of our newly proposed procedure, we show how estimation is still possible in the setting of unobserved components models, where different types of processes are superimposed.
Real-world data from the Global Drifter Program show its relevance for understanding surface flow measurements at the equator---a challenging region for studying inertial oscillations—where the power of the new method shows that despite rapid modulation we can still uncover the generating mechanism of the process.
There are a number of questions still remaining in our understanding of modulation. We have extended the regimes when estimation is possible, but do not know when an estimable process tips into one from which no information can be recovered.
By introducing a new class of models, many new questions can both be posited and answered, especially as most sources of real-world data show aggregations of components, all obeying different generation mechanisms.
\footnotesize
\setlength{\bibsep}{0pt plus 0.2ex}
\bibliographystyle{apalike}
|
\section{Introduction}
\label{sec:introduction}
In recent decades, quantum mechanical calculations using Kohn-Sham Density Functional Theory (Kohn-Sham DFT) \citep{HK_DFT, KohnSham_DFT} have become the de facto workhorse of computational materials science. The pseudopotential plane-wave method (\emph{Plane-wave DFT}) --- perhaps the most widely used implementation of the Kohn-Sham theory --- is currently available in a number of mature software packages \citep{Kresse_abinitio_iterative, CASTEP_1, Quantum_Espresso_1, Gonze_ABINIT_1}. The Plane-wave DFT approach involves expanding the unknowns (e.g., the Kohn-Sham orbitals and the electron density) into a linear combination of plane-waves\footnote{Plane-waves are functions of the form $\mathrm{e}^{\mathrm{i}{\bf k}\cdot{\bf x}}$. They form eigenfunctions of translational symmetry operators.}, and subsequently carrying out various computations through the use of Fast Fourier Transforms (FFTs) to seamlessly switch between quantities expressed on real space grids and their plane-wave expansion coefficients. Due to the periodic nature of plane-waves, Plane-wave DFT is well suited for the study of systems with translational symmetry. However, the study of structures which are non-periodic may require the use of large computational supercells \citep{Martin_ES}, thereby rendering the method inefficient for such problems.\footnote{Alternatives to Plane-wave DFT include real-space methods based on finite-differences \citep{Octopus_1, PARSEC, Phanish_SPARC_1, ghosh2016sparc2} and finite-elements \cite{Gavini_Kohn_Sham, Pask_FEM_review_1, Pask_FEM_review_2,Gavini_higher_order}. Although these methods allow non-periodic boundary conditions to be imposed more readily, their use in general symmetry-adapted self-consistent first principles calculations has not been considered prior to this work.}
A vast number of materials systems of interest today are non-periodic, but are associated with other physical symmetry groups. This includes for example, nanoclusters and molecules associated with point group symmetries, and helical nanostructures associated with screw transformation symmetries. The importance of such nanostructures associated with alternative (non-periodic) symmetries cannot be overstated. For instance, they are anticipated to exhibit unprecedented materials properties --- particularly, collective properties such as ferromagnetism and ferroelectricity --- in manners that are otherwise unavailable in the bulk phase \citep{James_OS}. Consequently, a substantial body of work has been devoted in recent years to the theoretical framework and mathematical classification of such structures\footnote{In the mechanics of materials literature, such materials systems have been referred to as \emph{Objective Structures} \citep{James_OS}.} \citep{James_OS, DEJ_ObjForm}. Further, judicious use of the symmetries associated with these structures has been made in designing novel computational methods -- both at the level of atomistic simulations \citep{Dumitrica_James_OMD, dayal2010nonequilibrium, aghaei2012symmetry} as well as electronic structure calculations \citep{My_PhD_Thesis, banerjee2015spectral}.
While the study of electronic properties of nanostructures associated with various non-periodic physical symmetries is of much scientific and technological interest in itself, an additional outcome of having access to computational methods specifically designed to exploit the underlying physical symmetries is that these methods allow one to study the associated deformation modes. In the setting of conventional Plane-wave DFT for example, the application of a homogeneous deformation to a periodic system still results in a structure with translational symmetry, and therefore the deformed system can be easily accommodated through a unit cell that has also been deformed in the same manner. On the other hand, it is far more difficult to account for non-homogeneous deformations such as bending and torsion in the periodic setting, whereby their study \citep{wei2012bending, naumov2011gap} is likely to involve various complications, inaccuracies and inefficiencies.
To the best of our knowledge, as far as atomistic systems are concerned, the connections between various physical symmetry groups and specific non-homogeneous deformation modes appears first in \citep{James_OS}. This idea has been subsequently exploited for simulating the effects of bending deformations by means of cyclic symmetry groups \citep{Dumitrica_James_OMD, Dumitrica_Bending_Graphene, ma2015thermal, Pekka_Efficient_Approach, Pekka_CNT_Bending, Pekka_GNR_Bending, Pekka_Revised_Periodic}, and torsional deformations by means of helical symmetry groups \citep{Dumitrica_James_OMD, zhang2008stability, zhang2009electromechanical, Pekka_Efficient_Approach, kit2012twisting, Pekka_Twisted_GNR, Pekka_Revised_Periodic}. A pervasive issue with these aforementioned works however, is that the simulations in question have been carried out within the framework of interatomic potentials or tight binding methods (classical semi-empirical or Density Functional Tight Binding (DFTB)). The failure of these more approximate models in simulating various physical systems is well known \citep{ismail2000ab,hauch1999dynamic, cocco2010gap, koskinen2009density, naumov2011gap} and it has been understood for some time that access to a systematic self-consistent first principles simulation methodology would be highly desirable \citep{Pekka_Revised_Periodic, James_OS, My_PhD_Thesis}. However, such a computational methodology appears to have been out of reach prior to this work.\footnote{Indeed, as remarked in \citep{Pekka_Revised_Periodic}, a central difficulty has been to formulate an analog of plane-waves for non-periodic symmetries. This issue has been subsequently addressed and resolved in \citep{My_PhD_Thesis}.}
In view of the above discussion, we formulate and implement Cyclic Density Functional Theory (Cyclic DFT) --- a self-consistent first principles simulation method for cyclic nanostructures.\footnote{The crystallographic restriction theorem \citep{senechal1996quasicrystals} prevents any periodic code --- including all Plane-wave DFT codes --- from making full use of the symmetry of a general cyclic structure.} Cyclic symmetries are ubiquitous in various clusters and molecular systems \citep{wiki_cyclic_compound, hargittai2009symmetry, willock2009, go1973ring} (see Fig.~\ref{fig:cyclic_structures} for some examples), and therefore the present methodology can be used to study a large variety of nanostructures from first principles in a systematic and efficient manner. Notably, even general point group symmetries associated with complex nanosystems usually contain cyclic groups as proper subgroups \citep{altmann1994point}, which extends the scope of the various materials systems that can be studied with Cyclic DFT.\footnote{For example, the icosahedral group that is associated with many fullerenes contains a cyclic subgroup originating from a 5-fold rotational symmetry.} Furthermore, due to the connections between cyclic symmetries with bending deformations in nanostructures (i.e., the idea that cyclic boundary conditions locally simulate the behavior of a system subjected to uniform bending), Cyclic DFT makes it possible to carry out systematic ab-initio simulations of nanostructures subjected to bending deformations. In particular, this opens up the possibility that electro-mechanical or other multi-physics coupling effects in nanostructures can be faithfully simulated from first principles by means of this novel framework.
Broadly, the present contribution can be viewed as a particular flavor of so called Objective DFT \citep{My_MS_Thesis, My_PhD_Thesis}, i.e., a self-consistent first principles method for an atomistic or molecular system with a non-periodic symmetry (i.e., an \textit{Objective Structure}). While Objective DFT is far more general and even allows systems associated with non-Abelian or non-compact symmetry groups to be simulated, it reduces to a particularly simple and efficient form for the case of cyclic structures, as we show in this work. Furthermore, the aforementioned connection of cyclic symmetries with the efficient simulation of bending deformations in nanostructures provides a compelling case as to why a thorough description of the theoretical and practical aspects of Cyclic DFT is warranted. Finally, Cyclic DFT can easily be extended to study systems of infinite extent (e.g., nanotubes, bending of nanosheets), which is beyond the scope of the spectral scheme originally presented in Objective DFT. This provides the motivation for the current work.
\begin{figure}[ht]
\centering
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{./si_nano_dot.eps}
\caption{Hydrogen passivated silicon quantum nanodot with 4-fold cyclic symmetry. We thank Yunkai Zhou (Southern Methodist University) for providing us with the nanodot structure.}
\label{fig_ex:subfig_a}
\end{subfigure}
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{./C5.eps}
\caption{Pentaphosphaferrocene molecule with a 5-fold symmetry: the atoms between the X axis and the line denoted as e can be rotated in steps of $72^{\circ}$ and replicated to produce the entire structure.}
\label{fig_ex:subfig_b}
\end{subfigure}%
$\quad$
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{./C6.eps}
\caption{Alpha-cyclodextrin molecule with a 6-fold symmetry: the atoms between the X axis and the line denoted as e can be rotated in steps of $60^{\circ}$ and replicated to produce the entire structure.}
\label{fig_ex:subfig_c}
\end{subfigure}
\caption{Examples of cyclic structures from nanotechnology and chemistry.}
\label{fig:cyclic_structures}
\end{figure}
One of the key theoretical steps involved in the formulation of Cyclic DFT is the symmetry cell reduction of the Kohn-Sham equations posed on the entire cyclic structure to the fundamental domain (or cyclic unit cell) associated with the cyclic symmetry group. In order to accomplish this task, we use tools from Group Representation Theory to rigorously formulate and prove a Bloch theorem for cyclic structures (Theorem \ref{thm:Bloch_Theorem}). In addition, we perform a symmetry cell reduction of the electrostatics problem, the system's electronic ground-state free energy, and the Hellmann-Feynman forces on the atoms. Consequently, on every iteration in the Self-Consistent Field (SCF) method \citep{slater1974self}, the reduction due to symmetry is expected to result in computational savings that varies in direct proportion to the cyclic group order of the system under study.\footnote{Asymptotically, the reduction in computational cost is actually quadratic with respect to group order (Footnote \ref{Footnote:QuadraticReduction}).} An additional outcome of the Bloch theorem for cyclic structures is that it allows the Kohn-Sham eigenvalues associated with the cyclic structure to be presented in a manner similar to that of traditional band diagrams for periodic structures. As a result, the onset of instabilities and transitions in such systems can be more easily detected.
We develop a symmetry-adapted finite-difference method to discretize the equations of Cyclic DFT resulting from the symmetry cell reduction. In addition to the systematic convergence properties and relative simplicity of this discretization strategy, it allows bending in nanostructures to be simulated more efficiently by allowing us to place finite-difference nodes only in regions of interest. Subsequently, we develop a completely functional numerical realization of the method using state of the art algorithms. Numerical experiments and comparison with benchmark calculations allow us to establish the accuracy and efficiency of our Cyclic DFT implementation. Further, these experiments reveal that the symmetry cell reduction not only results in the expected computational speed-up on every SCF step, but it also allows overall faster convergence of the SCF fixed-point iteration. As a consequence of these computational savings, highly accurate first principles simulations of cyclic systems containing many hundreds of atoms can be done quite routinely using even a serial implementation of Cyclic DFT.
Finally, as a demonstration of how Cyclic DFT enables the study of materials properties from first principles, we simulate uniform bending deformations in a silicene nanoribbon and obtain its energy-curvature relationship. In particular, our simulations allow us to obtain the bending stiffness of this material in the linear Euler--Bernoulli regime and to compare our results with the bending stiffness of other two-dimensional materials that have been studied in the literature. The usage of first principles approaches in the mechanics of materials brings with it the hope that these techniques will enable engineers to understand the true atomistic dependence of constitutive laws. Our application of Cyclic DFT to directly evaluate the bending stiffness can be broadly viewed in this light and it motivates us to carry out more sophisticated simulations of this nature on other important nanomaterials systems in the near future.\footnote{Exploiting point group symmetries in ab-initio calculations has also been considered in the chemistry literature in the context of Linear Combination of Atomic Orbitals (LCAO) methods (see \citep{LCAO_famous,Roothan_SCF_Symmetry,LCAO_1,LCAO_2,atkins2011molecular}, for example). However, the focus of these works has been towards using symmetry-adapted basis functions for reducing the effort associated with the computation of the multi-center integrals and the entries of the Hamiltonian matrix that appear in LCAO calculations. Thus, they differ in perspective from our own approach which focuses on formulation and solution of symmetry-adapted cell problems instead. Additionally, in contrast to finite-difference methods, it is usually non-trivial to systematically improve the quality of solution obtained via LCAO methods due to basis completeness issues. Finally, the application of cyclic symmetries to the study of bending deformations in nanostructures does not appear to have been considered by chemists before.}
The rest of the paper is organized as follows.\footnote{The atomic unit system with $m_{\text{e}}=1,e=1,\hbar=1,\frac{1}{4\pi\epsilon_0}=1$, is chosen for the rest of the work, unless otherwise mentioned.} In Section \ref{Sec:Formulation}, we present the mathematical underpinnings and symmetry-adapted finite-difference discretization scheme that constitute Cyclic DFT. In Section \ref{Sec:Results}, we verify the accuracy and efficiency of Cyclic DFT, and then use it to study the uniform bending of a silicene nanoribbon. Finally, we provide concluding remarks in Section \ref{Conclusions}.
\section{Formulation} \label{Sec:Formulation}
In this section, we describe the key aspects of Cyclic DFT. We begin by a formal discussion of cyclic groups and cyclic structures in Section \ref{Subsec:CyclicGroups}, and then discuss the Kohn-Sham theory for such structures in Section \ref{Subsec:KohnShamProblem}. Next, we show how cyclic symmetry leads to an appropriate symmetry cell reduction of the Kohn-Sham problem in Section \ref{Subsec:CyclicBlochTheorem}, and pose the resulting equations on the fundamental domain in Section \ref{Subsec:CyclicDFTEq}. Finally, we formulate a symmetry-adapted finite-difference discretization scheme for solving the cell problem in Section \ref{Subsec:FD}.
\subsection{Cyclic groups and cyclic structures} \label{Subsec:CyclicGroups}
Let ${\cal G}$ denote a cyclic group of order $\mathfrak{N}$ generated by the single element $g$, i.e.,
\begin{align}
\label{abstract_cyclic_group}
\displaystyle{\cal G}=\{g^0, g^1,g^2,\ldots,g^{\mathfrak{N}-1}\}\,.
\end{align}
Identifying $g^{\mathfrak{N}} = g^0 = e$ as the group identity element, we see that the inverse element of $g^{\gamma}\in {\cal G}$ is the element $g^{\mathfrak{N}-\gamma} \in {\cal G}$. A physical realization of this abstract group\footnote{Also identifiable as an additive group of integers modulo $\mathfrak{N}$.} can be obtained by considering a discrete group of rotations about a common axis. Let $({\bf e}_1,{\bf e}_2,{\bf e}_3)$ represent the standard basis of $\mathbb{R}^3$. Here, we form a faithful linear representation of ${\cal G}$ on $\mathbb{R}^3$ by using the following rotation matrices about ${\bf e}_3$:
\begin{align}
\label{rotation_cyclic_group}
\displaystyle\{\mathfrak{R}^{\gamma}:\gamma=0,\ldots,\mathfrak{N}-1\} \quad\text{with}\quad
\mathfrak{R}^{\gamma}=\begin{pmatrix}
\cos\frac{2\pi\gamma}{\mathfrak{N}} & -\sin\frac{2\pi\gamma}{\mathfrak{N}} & 0\\
\sin\frac{2\pi\gamma}{\mathfrak{N}} & \;\cos\frac{2\pi\gamma}{\mathfrak{N}} & 0 \\
0 & 0 & 1
\end{pmatrix}\,.
\end{align}
With this representation, the multiplication operation of the group simply becomes multiplication of the matrices $\mathfrak{R}^{\gamma}$. Henceforth, we will identify a cyclic group of order $\mathfrak{N}$ by the collection of matrices shown in Eq.~\ref{rotation_cyclic_group}; we will use $\Gamma_{\mathfrak{N}}$ to denote the collection of integers $\{0,1,\ldots,\mathfrak{N}-1\}$; and for any matrix $\mathfrak{R}^{\gamma} \in {\cal G}$ and set $S \subset \mathbb{R}^3$, we will use $\mathfrak{R}^{\gamma}(S)$ to denote the set $\{\mathfrak{R}^{\gamma}{\bf x} : {\bf x} \in S \}$.
Let ${\cal C} \subseteq \mathbb{R}^3$ be an open set with a regular boundary which is invariant under the action of the group ${\cal G}$, i.e., for every $\gamma \in \Gamma_{\mathfrak{N}}$, we have:
\begin{align}
\label{domain_invariance}
\mathfrak{R}^{\gamma}({\cal C}) = {\cal C}\,.
\end{align}
Then, the \emph{symmetry cell} or \emph{fundamental domain} or \emph{cyclic unit cell} of ${\cal G}$ in ${\cal C}$ is a set ${\cal D}_{{\cal G}} \subset {\cal C}$ such that:
\begin{align}
\bigcup_{\gamma}\mathfrak{R}^{\gamma}({\cal D}_{{\cal G}}) &= {\cal C}\,,
\label{fundamental_domain_1}
\end{align}
and for every $\gamma_1, \gamma_2 \in \Gamma_{\mathfrak{N}}$
\begin{align}
\mathfrak{R}^{\gamma_1}({\cal D}_{{\cal G}}) \bigcap \mathfrak{R}^{\gamma_2}({\cal D}_{{\cal G}}) &= \textrm{a set of Lebesgue measure }0, \textrm{ for } \gamma_1 \neq \gamma_2\,.
\label{fundamental_domain_2}
\end{align}
We will denote the boundary surfaces common to the original domain and the fundamental domain by $\partial{\cal D}_{{\cal G}}^{0} = \partial{\cal D}_{{\cal G}} \cap \partial{\cal C}$, and those that are unique to the fundamental domain by $\partial{\cal D}_{{\cal G}}^{C} = \partial{\cal D}_{{\cal G}} \setminus \partial{\cal D}_{{\cal G}}^{0} $. For example, if ${\cal C}$ is a finite cylinder with axis ${\bf e}_3$, the fundamental domain ${\cal D}_{{\cal G}}$ will be a sector of the cylinder with angle $2\pi / \mathfrak{N}$.
Consider a finite collection\footnote{In a forthcoming contribution, we will consider the case in which ${\cal P}$ consists of an infinite number of points. This can be used for example, to represent a structure such as a bent silicene sheet that is periodic in the ${\bf e}_3$ direction while having cyclic symmetry around the ${\bf e}_3$ axis.} of distinct points\footnote{For the sake of simplicity, we will assume that none of the points are at the origin, although this assumption is not really required as far as theory and methods developed in this work are concerned.} ${\cal P} \subset {\cal D}_{{\cal G}}$, which we will denote by $\big\{ {\bf x}_{0,k} \big\}_{k=1}^{M_{{\cal P}}}$. These points are representative of atomic positions within the fundamental domain and will be referred to as the \emph{simulated points} or \emph{simulated atoms}. A \emph{cyclic structure}\footnote{Alternately, an \emph{objective structure generated by a cyclic group}, using the terminology employed in the work of \citep{My_PhD_Thesis, DEJ_ObjForm}.} ${\cal S}_{\cal_G, {\cal P}}$ is the group orbit of a given cyclic group ${\cal G}$ acting on the simulated atoms in ${\cal P}$:
\begin{align}
\label{cyclic_structure}
\displaystyle{\cal S}_{\cal_G, {\cal P}}=\bigcup_{\gamma}\mathfrak{R}^{\gamma}({\cal P})= \bigcup_{\gamma , k}\mathfrak{R}^{\gamma}{\bf x}_{0,k}\,.
\end{align}
We will represent a general atom of such a structure using the notation ${\bf x}_{\gamma,k} = \mathfrak{R}^{\gamma}{\bf x}_{0,k}$, $\gamma \in \Gamma_{\mathfrak{N}}$. Further, we will denote the nuclear charge\footnote{Following usual convention adopted in ab-initio theories, we will consider nuclear charge to be negative and therefore the electronic charge to be positive. Additionally, we will neglect the core charges and consider only the valence charges, consistent with the use of the pseudopotential approximation.} of the atom at position ${\bf x}_{0,k} \in {\cal P}$ by $Z_k$. It follows from the cyclic symmetry of the structure that the atom at the position ${\bf x}_{\gamma,k} = \mathfrak{R}^{\gamma}{\bf x}_{0,k}$ (for any $\gamma \in \Gamma_{\mathfrak{N}}$) also has nuclear charge $Z_k$.
\subsection{Kohn-Sham problem for cyclic structures} \label{Subsec:KohnShamProblem}
A detailed formulation of the Kohn-Sham problem \citep{KohnSham_DFT} as it applies to the computation of the electronic and structural properties of a finite collection of atoms can be found in numerous references (e.g., see \citep{banerjee2015spectral, Gavini_Kohn_Sham, Phanish_SPARC_1}). In the present work, we are interested specifically in the first principles simulations of cyclic structures ${\cal S}_{\cal_G, {\cal P}}$, as have been defined in the previous subsection. For the sake of simplicity, we assume that the system as a whole is charge neutral, i.e., there are a total of $\mathfrak{N}N_e$ electrons, where $N_{\text{e}} = - \displaystyle \sum_{k=1}^{M_{\cal P}}Z_k$ represents the number of electrons in the fundamental domain.\footnote{The electrons are not physically restricted to be confined to the fundamental domain. This situation is analogous to the band theory of crystalline solids \citep{Reed_Simon4, LeBris_Defranceschi_ReviewBook}, where a fixed number of electrons are prescribed to be in the periodic unit cell, although the electrons can be spatially delocalized over the infinite crystal.} Consequently, we pose the Kohn-Sham equations in terms of $\mathfrak{N} N_{\text{s}}$ orbitals with $N_{\text{s}} > N_{\text{e}} / 2$ and $N_{\text{s}} \approx N_{\text{e}} / 2$.\footnote{In practical calculations, $N_{\text{s}} = N_{\text{e}} / 2 + N_{\text{extra}}$, and $N_{\text{extra}} = 10 - 20$ for typically used electronic temperatures \citep{zhou2014chebyshev}.} Additionally, we employ the local density approximation (LDA) \citep{KohnSham_DFT}, thermalization via the Fermi-Dirac smearing \citep{Mermin_Finite_Temp}, the local pseudopotential approximation \citep{Kohanoff}, and the local reformulation of the electrostatics \citep{suryanarayana2014augmented,Phanish_SPARC_1}.
In the aforementioned setting, the Kohn-Sham equations in ${\cal C}$ can be written as:\footnote{Henceforth, the space of square integrable functions on the domain ${\cal C}$ will be denoted by $\mathsf{L}^2({\cal C})$, while the Sobolev space of functions in $\mathsf{L}^2({\cal C})$ whose $\mathsf{k}^{\textrm{th}}$ order weak derivatives also lie in $\mathsf{L}^2({\cal C})$ will be denoted by $\mathsf{H^k}({\cal C})$. The subspace of $\mathsf{H}^1({\cal C})$ functions that vanish on the boundary $\partial {\cal C}$ in the trace sense will be denoted by $\mathsf{H}^1_0({\cal C})$. The space of $k-$times continuously differentiable functions that are compactly supported on ${\cal C}$ will be denoted by $\textrm{C}^\mathsf{k}_c({\cal C})$. In particular, continuous functions with compact support on ${\cal C}$ will be denoted as $\textrm{C}_c^0({\cal C})$. }
\begin{subequations} \label{Eqns:KS}
\begin{align}
\label{KS_equations}
K(\rho)\,\phi_i&=\lambda_i\,\phi_i\,,\;\innprod{\phi_i}{\phi_j}{\mathsf{L}^2({\cal C})}=\delta_{ij}\;,\\
\label{KS_explain1}
\text{where}\quad
K(\rho)&=-\frac{1}{2} \Delta + V_{\text{es}}(\rho, {\cal S}_{\cal_G, {\cal P}}) + V_{\text{xc}}(\rho)\;,\\
\label{KS_explain2}
\text{with} \quad
\rho({\bf x})&=2 \sum_{i=1}^{\mathfrak{N} N_{\text{s}}} f_i\abs{\phi_i({\bf x})}^2\,,\\
\label{poisson_problem}
\text{and} \quad
-\frac{1}{4\pi}\Delta V_{\text{es}} &= b + \rho\,. \\
\label{KS_explain3}
\text{Here,} \quad f_i &= \left( 1 + \exp\left( \frac{\lambda_i - \lambda_f}{\sigma} \right) \right)^{-1}, \\
\label{KS_explain4}
\text{with $\lambda_f$ s.t.} \quad &2\sum_{i=1}^{\mathfrak{N} N_{\text{s}}} f_i = \mathfrak{N} N_{\text{e}}\,.
\end{align}
\end{subequations}
The corresponding boundary conditions are:
\begin{subequations} \label{Eqns:KS:BC}
\begin{align}
\text{for} \,\,\, {\bf x} \in \partial{\cal C} \,, \quad \phi_i({\bf x}) &= 0 \,, \\
\rho({\bf x}) &= 0 \,, \\
\text{and} \quad V_{\text{es}}(\rho({\bf x}),{\cal S}_{\cal_G, {\cal P}}) &= \int_{{\cal C}}\!\frac{\rho({\bf y}) + b({\bf y}, {\cal S}_{\cal_G, {\cal P}})}{\abs{{\bf x}-{\bf y}}}\; \mathrm{d{\bf y}} \,. \label{Eqn:KS:BC:Ves}
\end{align}
\end{subequations}
Above, $\rho$ is the electron density; $K(\rho)$ is the Kohn-Sham operator with eigenvalues $\lambda_i$ and eigenfunctions $\phi_i \in \mathsf{H}^1({\cal C})$, commonly referred to as the Kohn-Sham orbitals; $V_{\text{es}}$ is the electrostatic potential \citep{Pask2005,Gavini_Kohn_Sham, suryanarayana2014augmented}, written as the solution to the Poisson problem in Eqs.~\ref{poisson_problem} and \ref{Eqn:KS:BC:Ves}; $V_{\text{xc}}$ is the exchange-correlation potential; $\mathfrak{N} N_{\text{s}}$ is the total number of states\footnote{This can be thought of as choosing $N_{\text{s}}$ states corresponding to the fundamental domain.}; $0 \leq f_i \leq 1$ are the thermalized orbital occupations arising from the Fermi-Dirac smearing (Eq.~\ref{KS_explain3}) with parameter $\sigma$; $\lambda_f$ is the Fermi level of the system, determined by the solution of Eq.~\ref{KS_explain4} \footnote{As a result, the electron density satisfies the constraint on the total number of electrons:
\begin{align}
\label{density_normalized}
\int_{{\cal C}}\!\rho({\bf x})\, \mathrm{d{\bf x}}= \mathfrak{N} N_{\text{e}}\,.
\end{align}};
and $b({\bf y}, {\cal S}_{\cal_G, {\cal P}}) = \displaystyle \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} b_k({\bf y}, {\bf x}_{\gamma,k})$ is the total nuclear pseudocharge density, with $b_k$ being the pseudocharge of the $k^{th}$ nucleus.\footnote{Each $b_k(\cdot, {\bf x}_{\gamma,k})$ is a spherically symmetric smooth function that is compactly supported in a small ball centered at ${\bf x}_{\gamma,k}$, with net enclosed charge $Z_k$. This implies that $b_k$ can be expressed using the ansatz $b_k({\bf x}, {\bf x}_{\gamma,k}) = \eta_k(\abs{{\bf x} - {\bf x}_{\gamma,k}})$, where $\eta_k(\cdot)$ is a smooth one-dimensional function that satisfies $\eta_k(r)=0 \, \forall \, r>r_k$ and $\displaystyle \int_{0}^{r_k}\!\eta_k(r) 4\pi r^2 \mathrm{dr} = Z_k$\,. As a result, the total nuclear pseudocharge density satisfies the relation
\begin{align}
\displaystyle \int_{{\cal C}}\! b({\bf x}, {\cal S}_{\cal_G, {\cal P}})\, \mathrm{d{\bf x}} = \mathfrak{N}\sum_{k=1}^{M_{\cal P}}Z_k\,.
\end{align}}
Due to the finite extent of the system under study and its overall charge neutrality, zero-Dirichlet boundary conditions are prescribed on the Kohn-Sham orbitals $\phi_i$ \citep{wavefunc_decay1, wavefunc_decay2, banerjee2015spectral, Gavini_Kohn_Sham}. It therefore follows from Eq.~\ref{KS_explain2} that the electron density $\rho$ also obeys this boundary condition. The Dirichlet boundary conditions prescribed on $V_{\text{es}}$ are obtained by expressing the solution of Eq.~\ref{poisson_problem} in terms of the Green's function of the Laplacian \citep{Phanish_SPARC_1}.\footnote{As explained later, the boundary conditions for $V_{\text{es}}$ can be numerically evaluated by direct quadrature or through the use of cylindrical multipole moments.}
The self-consistent solution of the Kohn-Sham problem described by Eqs. \ref{Eqns:KS} and \ref{Eqns:KS:BC} leads to the electronic ground-state\footnote{Following standard terminology \citep{LeBris_ReviewBook}, we will refer to the electronic ground state to mean the self-consistent solution of the Kohn-Sham equations for a given set of nuclear positions.} for the cyclic structure ${\cal S}_{\cal_G, {\cal P}}$. The nonlinear eigenvalue problem is typically solved using a fixed-point iteration with respect to either the electron density $\rho$ or the total effective potential
\begin{align}
\label{effective_potential}
V_{\textrm{eff}}({\bf x}) = V_{\textrm{es}}(\rho({\bf x}), {\cal S}_{\cal_G, {\cal P}}) + V_{{\textrm{xc}}}(\rho({\bf x})) \,,
\end{align}
and this is commonly referred to as the SCF method \citep{slater1974self}. In each iteration of this method, the Poisson problem in Eqs.~\ref{poisson_problem} and \ref{Eqn:KS:BC:Ves} needs to be solved to calculate $V_{\textrm{eff}}$, and the lowest $\mathfrak{N}N_{\text{s}}$ eigenstates of the linearized Kohn-Sham operator:
\begin{align}
\label{linear_operator}
\mathfrak{H} = -\frac{1}{2} \Delta + V_{\textrm{eff}}({\bf x})
\end{align}
need to be determined to calculate $\rho$.\footnote{It is possible to calculate the electron density directly from $\mathfrak{H}$, without actually calculating the orbitals (e.g., see \cite{suryanarayana2013optimized,suryanarayana2013spectral,pratapa2015spectral}). } The eigenfunctions of $\mathfrak{H}$ are subjected to zero--Dirichlet boundary conditions, consistent with the Kohn-Sham orbitals.
At the electronic ground-state, the system's free energy can be evaluated by means of a Harris-Foulkes \citep{harris1985simplified,foulkes1989tight} type functional \citep{Phanish_SPARC_1}:\footnote{In some cases, it might be necessary to add a term $E_{\text{overlap}}({{\cal S}_{\cal_G, {\cal P}}})$ to the free energy expression in order to account for the possible overlaps of the pseudocharges \citep{suryanarayana2014augmented, Phanish_SPARC_1}.}
\begin{align}
\nonumber
\mathcal{F}({{\cal S}_{\cal_G, {\cal P}}}) =& \;2\sum_{i = 1}^{\mathfrak{N}N_{\text{s}}} f_i \lambda_i + \int_{{\cal C}}\!\varepsilon_{\text{xc}} (\rho({\bf x})) \rho({\bf x}) \, \mathrm{d {\bf x}} - \int_{{\cal C}}\!V_{\text{xc}}(\rho({\bf x}))\rho({\bf x})\,\mathrm{d{\bf x}} \\\nonumber
&+ \frac{1}{2} \int_{{\cal C}}\!\big (b({\bf x}, {\cal S}_{\cal_G, {\cal P}}) - \rho({\bf x})\big) V_{\text{es}}({\bf x}, {\cal S}_{\cal_G, {\cal P}}) \, \mathrm{d{\bf x}} \\\nonumber
&-\frac{1}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal C}}\!b_k({\bf x}, {\bf x}_{\gamma,k})\, V_k({\bf x}, {\bf x}_{\gamma,k})\,\mathrm{d{\bf x}} \nonumber \\
&+ {2\,\sigma} \sum_{i=1}^{\mathfrak{N} N_{s}} \big( f_i \log f_i + (1-f_i) \log (1 - f_i) \big)\,,
\label{Harris_Foulkes_Energy}
\end{align}
where the first term is the so called Kohn-Sham band energy; the second and third terms arise from the exchange-correlation effects; the fourth and fifth terms arise due to electrostatic interactions and pseudocharge self energies, respectively; and the last term represents the free energy associated with the electronic entropy of the system. The corresponding Hellmann-Feynman force \citep{Finnis_book, Parr_Yang, Martin_ES} on the atom ${\bf x}_{\gamma,k}$\footnote{More specifically, its nucleus.} can be written as \citep{suryanarayana2014augmented, Phanish_SPARC_1}:\footnote{If $E_{\text{overlap}}({{\cal S}_{\cal_G, {\cal P}}})$ is included in the free energy expression in Eq. \ref{Harris_Foulkes_Energy}, it becomes necessary to add the term
\begin{align}
{\bf f}_{\text{overlap}}({\bf x}_{\gamma,k}) = - \pd{E_{\text{overlap}}({{\cal S}_{\cal_G, {\cal P}}})}{{\bf x}_{\gamma,k}}
\end{align}
to the atomic force in order to account for possible overlaps of pseudocharges \citep{suryanarayana2014augmented, Phanish_SPARC_1}. \label{footnote:fn_1}}
\begin{align}
\label{force_formula_1}
{\bf f}_{{\bf x}_{\gamma,k}} =- \pd{\mathcal{F}({{\cal S}_{\cal_G, {\cal P}}})}{{\bf x}_{\gamma,k}} = \int_{{\cal C}}\!\nabla b_k({\bf x}, {\bf x}_{\gamma,k})\big(V_{\text{es}}({\bf x}, {\cal S}_{\cal_G, {\cal P}}) - V_k({\bf x}, {\bf x}_{\gamma,k}) \big)\,\mathrm{d{\bf x}}\,.
\end{align}
The problem of determining the equilibrium geometry\footnote{In principle, an equilibrium geometry (i.e., a configuration in which the internal forces on the atoms are all zero) can be associated with a local minimum, maximum or saddle point of the energy landscape. However, most applications in mechanics are usually concerned with energy minimizing structures and we will also restrict our attention to such cases.} of the cyclic structure can be stated as solving the minimization problem:\footnote{In practice, several local structural minima (and not just a single global minimum) may be involved in the energy landscape of a structure.}
\begin{align}
\label{geometry_optimization_original}
\infm{\{{\bf x}_{\gamma,k} \in {\cal S}_{\cal_G, {\cal P}}\}}{\mathcal{F}({{\cal S}_{\cal_G, {\cal P}}})}\,,\,\text{subject to}\;{\bf x}_{\gamma,k} = \mathfrak{R}^{\gamma}{\bf x}_{0,k} \,,\, \text{for} \,\, \gamma \in \Gamma_{\mathfrak{N}}\,.
\end{align}
For every atomic configuration that arises during this geometry optimization procedure, the electronic ground-state needs to be determined by solving the Kohn-Sham equations in Eqs.~\ref{Eqns:KS} and \ref{Eqns:KS:BC}, after which the corresponding free energy and atomic forces can be calculated using Eqs.~\ref{Harris_Foulkes_Energy} and \ref{force_formula_1}, respectively.
\subsection{Symmetry cell reduction of the Kohn-Sham problem for cyclic structures} \label{Subsec:CyclicBlochTheorem}
We now describe how the underlying symmetry of cyclic structures can be utilized to perform a symmetry cell reduction of the Kohn-Sham DFT problem described in the previous subsection to the fundamental domain.
\subsubsection{Reduction of the electron density} \label{Subsubsec:Symm:rho}
A natural starting point is to make the assumption that the electron density $\rho$ inherits the symmetry of the cyclic structure. Thus, for (almost) every ${\bf x} \in {\cal C}$ and for $\gamma \in \Gamma_{\mathfrak{N}}$, we may write:\footnote{As a consequence of Proposition \ref{prop:invariance_to_BC}, Eq.~\ref{rho_symmetry} can be interpreted as cyclic boundary conditions (on the surfaces $\partial{\cal D}_{{\cal G}}^{C}$) applied to the electron density when restricted to the fundamental domain.}
\begin{align}
\label{rho_symmetry}
\rho({\bf x}) = \rho(\mathfrak{R}^{\gamma} {\bf x})\,.
\end{align}
This assumption is very similar in nature to that commonly used in crystal lattice calculations, wherein it is assumed that the electron density inherits the translational symmetry of the lattice \citep{Defranceschi_LeBris, rhodes2010crystallography, rohrer2001structure, giustino2014materials}. Analogous to that case, it is conceivable that period doubling (i.e., cyclic group order halving in our case) charge density waves or Pierels instabilities \citep{peierls1955quantum,gruner2000density, kennedy2004proof} can occur in the system, leading to a structural transition and a breakdown of the assumption in Eq.~\ref{rho_symmetry}. In such a case, a larger cyclic unit cell that corresponds to a lower group order needs to be employed. However, the study of such instabilities, including the detection of their onset as well as their physical consequences is beyond the scope of the present contribution and is a compelling topic for future research.\footnote{It is likely however, that the cyclic band structure and cyclic phonon diagrams obtained from Cyclic DFT computations (Section \ref{subsec:silicene_results}) will be instrumental in the prediction of such instabilities.}
\subsubsection{Reduction of the effective potential} \label{Subsubsec:Symm:Veff}
Let us now consider the consequences of the cyclic group invariance of the electron density (i.e., $\rho({\bf x}) = \rho(\mathfrak{R}^{\gamma} {\bf x})$ for $\gamma \in \Gamma_{\mathfrak{N}}$) on the effective potential. First we observe:
\begin{proposition} \label{prop:PseudochargeDensity}The following conditions hold for the pseudocharges and the corresponding potentials for a cyclic structure:
\begin{enumerate}
\item For any $\gamma \in \Gamma_{\mathfrak{N}}$, the nuclear pseudocharges $b_k(\cdot)$ and the corresponding pseudopotentials $V_k(\cdot)$ obey the conditions:
\begin{align}
\label{bk_condition}
b_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k}) & = b_k({\bf x}, {\bf x}_{\mathfrak{N}-\gamma,k})\,,\\
\label{vk_condition}
V_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k}) &= V_k({\bf x}, {\bf x}_{\mathfrak{N}-\gamma,k})\,.
\end{align}
\item The total nuclear pseudocharge \begin{align} \label{Prop:Invariance_b}
b({\bf x}, {\cal S}_{\cal_G, {\cal P}}) = \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} b_k({\bf x}, {\bf x}_{\gamma,k}) = \sum_{\gamma = 0}^{\mathfrak{N}-1}\sum_{k = 1}^{M_{{\cal P}}} \eta_k(\abs{{\bf x} - {\bf x}_{\gamma,k}})
\end{align}
is invariant under ${\cal G}$.
\end{enumerate}
\end{proposition}
\begin{myproof}\hfill
\begin{enumerate}
\item The radial symmetry of $b_k(\cdot)$ implies that $b_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k}) = b_k({\bf x}, \mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}_{0,k})$, from which Eq.~\ref{bk_condition} follows. Next, using the change of variables $\mathfrak{R}^{\mathfrak{N}-\gamma}{\bf y} = {\bf z}$, we get:
\begin{align}
\nonumber
V_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k}) &= \int_{{\cal C}}\!\frac{b_k({\bf y}, {\bf x}_{0,k})}{\abs{\mathfrak{R}^{\gamma}{\bf x}-{\bf y}}}\; \mathrm{d{\bf y}} = \int_{{\cal C}}\!\frac{b_k({\bf y}, {\bf x}_{0,k})}{\abs{\mathfrak{R}^{\gamma}{\bf x}-\mathfrak{R}^{\gamma}\mathfrak{R}^{\mathfrak{N}-\gamma}{\bf y}}}\; \mathrm{d{\bf y}} \\
&=\int_{{\cal C}}\!\frac{b_k(\mathfrak{R}^{\gamma}{\bf z}, {\bf x}_{0,k})}{\abs{{\bf x}-{\bf z}}}\;\mathrm{d{\bf z}}\, = \int_{{\cal C}}\!\frac{b_k({\bf z}, \mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}_{0,k})}{\abs{{\bf x}-{\bf z}}}\; \mathrm{d{\bf z}}
= V_k({\bf x}, \mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}_{0,k})\,,\label{Vk_calculation}
\end{align}
from which Eq.~\ref{vk_condition} follows.
\item We note that for any $\gamma,\gamma_1 \in \Gamma_{\mathfrak{N}}$, the rotation matrix $\mathfrak{R}^{\mathfrak{N}-\gamma}$ acting on the point ${\bf x}_{\gamma_1,k}= \mathfrak{R}^{\gamma_1}{\bf x}_{0,k}$ produces the point ${\bf x}_{\gamma_2,k} = \mathfrak{R}^{\gamma_2}{\bf x}_{0,k}$, with $\gamma_2 = (\mathfrak{N}-\gamma+\gamma_1)\,\text{mod}\,\mathfrak{N}$. In this situation, if $\gamma$ is held fixed while $\gamma_1$ is allowed to cycle through the elements of $\Gamma_{\mathfrak{N}}$, $\gamma_2$ also cycles through each element of $\Gamma_{\mathfrak{N}}$, thus yielding the group orbit of ${\bf x}_{0,k}$ under ${\cal G}$. Thus, we may write for any $\gamma \in \Gamma_{\mathfrak{N}}$:
\begin{align}
\nonumber
b(\mathfrak{R}^{\gamma}{\bf x}, {\cal S}_{\cal_G, {\cal P}}) &= \sum_{\gamma_1 = 0}^{\mathfrak{N}-1}\sum_{k = 1}^{M_{{\cal P}}} \eta_k(\abs{\mathfrak{R}^{\gamma}{\bf x} - \mathfrak{R}^{\gamma_1}{\bf x}_{0,k}}) \\\nonumber
&= \sum_{\gamma_1 = 0}^{\mathfrak{N}-1}\sum_{k = 1}^{M_{{\cal P}}} \eta_k(\abs{\mathfrak{R}^{\gamma}{\bf x} - \mathfrak{R}^{\gamma}\,\mathfrak{R}^{\mathfrak{N} - \gamma}\,\mathfrak{R}^{\gamma_1}{\bf x}_{0,k}})\\\nonumber
&= \sum_{\gamma_1 = 0}^{\mathfrak{N}-1}\sum_{k = 1}^{M_{{\cal P}}} \eta_k(\abs{{\bf x} - \mathfrak{R}^{\mathfrak{N} - \gamma}\,\mathfrak{R}^{\gamma_1}{\bf x}_{0,k}})\\\nonumber
&= \sum_{\gamma_2 = 0}^{\mathfrak{N}-1}\sum_{k = 1}^{M_{{\cal P}}} \eta_k(\abs{{\bf x} - \mathfrak{R}^{\gamma_2}{\bf x}_{0,k}})\;, \textrm{with}\; \gamma_2 = (\mathfrak{N}-\gamma+\gamma_1)\,\text{mod}\,\mathfrak{N}\\\label{b_invariant}
&= b({\bf x}, {\cal S}_{\cal_G, {\cal P}})\,,
\end{align}
where we have used the spherical symmetry of $\eta_k$ to arrive at the third equality.
\end{enumerate}
\end{myproof}
Next, we use the above result to show:
\begin{proposition}
\label{prop:invariant_potential} The total effective potential
\begin{align}
\nonumber
\displaystyle V_{\emph{\textrm{eff}}}({\bf x}) = V_{\emph{\textrm{es}}}(\rho({\bf x}), {\cal S}_{\cal_G, {\cal P}}) + V_{\emph{\textrm{xc}}}(\rho({\bf x}))
\end{align}
is invariant under ${\cal G}$.
\end{proposition}
\begin{myproof}
It suffices to show that both the terms that constitute $V_{\textrm{eff}}({\bf x})$ are individually group invariant. The exchange-correlation potential $V_{{\textrm{xc}}}(\rho({\bf x}))$ is invariant by virtue of the invariance of $\rho({\bf x})$. To show the invariance of the electrostatic potential, we recall that it can be written as
\begin{align}
V_{\text{es}}(\rho({\bf x}),{\cal S}_{\cal_G, {\cal P}}) &= \int_{{\cal C}}\!\frac{\rho({\bf y}) + b({\bf y}, {\cal S}_{\cal_G, {\cal P}})}{\abs{{\bf x}-{\bf y}}}\; \mathrm{d{\bf y}} \,.
\end{align}
Therefore, for any $\gamma \in \Gamma_{\mathfrak{N}}$, we may write
\begin{align}
V_{\text{es}}(\rho(\mathfrak{R}^{\gamma} {\bf x}),{\cal S}_{\cal_G, {\cal P}}) &= \int_{{\cal C}}\!\frac{\rho({\bf y})+b({\bf y}, {\cal S}_{\cal_G, {\cal P}})}{\abs{\mathfrak{R}^{\gamma}{\bf x}-{\bf y}}}\; \mathrm{d{\bf y}} \nonumber \\
&= \int_{{\cal C}}\!\frac{\rho(\mathfrak{R}^{\gamma} {\bf z}) + b(\mathfrak{R}^{\gamma} {\bf z}, {\cal S}_{\cal_G, {\cal P}})}{\abs{\mathfrak{R}^{\gamma}{\bf x}-\mathfrak{R}^{\gamma} {\bf z}}}\; \mathrm{d{\bf z}} \nonumber \\
&= \int_{{\cal C}}\!\frac{\rho({\bf z})+b({\bf z}, {\cal S}_{\cal_G, {\cal P}})}{\abs{{\bf x}-{\bf z}}}\; \mathrm{d{\bf z}} \nonumber \\
&= V_{\text{es}}(\rho({\bf x}),{\cal S}_{\cal_G, {\cal P}}) \,,
\end{align}
where we have used the change of variables ${\bf y} = \mathfrak{R}^{\gamma}{\bf z}$ and the invariance of the Lebesgue measure under rotations to arrive at the second equality. Further, we have used the invariance of $\rho$ (Eq.~\ref{rho_symmetry}) and $b$ (Eq.~\ref{Prop:Invariance_b}) to arrive at the third equality.
\end{myproof}
As a consequence of the above results --- specifically, the cyclic group invariance of $b({\bf x}, {\cal S}_{\cal_G, {\cal P}})$ and $V_{\text{es}}(\rho({\bf x}),{\cal S}_{\cal_G, {\cal P}})$ --- the Poisson problem in Eqs.~\ref{poisson_problem} and \ref{Eqn:KS:BC:Ves} can be reduced to the fundamental domain, with cyclic boundary conditions on the surfaces $\partial{\cal D}_{{\cal G}}^{C}$.
\subsubsection{Reduction of the linear eigenvalue problem} \label{Subsubsec:EigProblem}
We recall that in the every iteration of the SCF method, it is necessary to compute part of the spectrum associated with the linearized Kohn-Sham operator:
\begin{align}
\mathfrak{H} = -\frac{1}{2} \Delta + V_\textrm{eff}({\bf x}) \,,
\end{align}
where the effective potential $V_\textrm{eff}({\bf x})$ has been evaluated from the electron density calculated during the previous SCF iteration. Assuming that $V_\textrm{eff}({\bf x})$ is a continuous function,\footnote{This is indeed true if $\rho$ is continuous \citep{My_PhD_Thesis}, as is evident from Eq. \ref{effective_potential}.} it can be shown that the operator $\mathfrak{H}$ on $\mathsf{L}^2({\cal C})$, with domain $\textrm{Dom}(\mathfrak{H}) = \mathsf{H}^2({\cal C}) \cap \mathsf{H}^1_0({\cal C})$, is an elliptic self-adjoint operator with a compact resolvent \citep{My_PhD_Thesis, Evans_PDE, Renardy_Rogers, Kato}. Therefore, $\mathfrak{H}$ has an increasing sequence of real eigenvalues, with the associated eigenfunctions\footnote{Although a finite cylinder has edges on its boundary, the eigenfunctions are expected to be at least $\mathsf{H}^2({\cal C})$ regular since the domain is convex \citep{grisvard2011elliptic}.} forming an orthonormal basis of $\mathsf{L}^2({\cal C})$.
Let us assume that the electron density calculated during the previous SCF iteration is group invariant.\footnote{This is indeed the case in Cyclic DFT, as shown in Section \ref{Subsubsec:ConsistencySCF}.} It follows from Proposition \ref{prop:invariant_potential} that the effective potential $V_\textrm{eff}({\bf x})$ in the current SCF iteration is also group invariant. This leads us to the following observation:
\begin{proposition}
\label{prop:invariant_operator}
With \emph{$V_\textrm{eff}({\bf x})$} continuous over ${\cal C}$ and group invariant, the operator $\mathfrak{H}$ commutes with the action of the group ${\cal G}$ on functions in $\mathsf{L}^2({\cal C})$, i.e., for any $\gamma \in \Gamma_{\mathfrak{N}}$, if \emph{$T_\gamma : \mathsf{L}^2({\cal C}) \to \mathsf{L}^2({\cal C})$} is an operator such that \emph{$T_\gamma f({\bf x}) \mapsto f\big((\mathfrak{R}^{\gamma})^{-1}{\bf x}\big)$}, then for any function \emph{$f \in \textrm{Dom}(\mathfrak{H})$}, we have \emph{$\mathfrak{H} T_\gamma f = T_\gamma \mathfrak{H} f$}.
\end{proposition}
\begin{myproof}
Since the function space $\textrm{C}^2_c({\cal C})$ is dense in $\textrm{Dom}(\mathfrak{H})$, it suffices to verify this result for $f \in \textrm{C}^2_c({\cal C})$. Using a change of variables calculation, it can be shown that the Laplacian operator commutes with rotations, and therefore with the operators $T_\gamma$. This combined with the group invariance of $V_\textrm{eff}({\bf x})$ establishes the required result.
\end{myproof}
It can be shown that the collection of operators $\{ T_\gamma: \gamma \in \Gamma_{\mathfrak{N}}\}$ form a faithful unitary representation\footnote{Each operator $T_\gamma$ is unitary and together the collection $\{ T_\gamma: \gamma \in \Gamma_{\mathfrak{N}}\}$ forms an identity preserving homomorphism of ${\cal G}$ on the carrier space $\mathsf{L}^2({\cal C})$. Since the representation is faithful, the map ${\cal G} \ni \mathfrak{R}^{\gamma} \mapsto T_\gamma$ is in fact an isomorphism.} of the group ${\cal G}$ on the space $\mathsf{L}^2({\cal C})$ \citep{My_PhD_Thesis}. Proposition \ref{prop:invariant_operator} shows that the operator $\mathfrak{H}$ is left invariant by these representations, i.e., we may write $T_\gamma \mathfrak{H} (T_\gamma)^{-1} = \mathfrak{H}$. In order to exploit this invariance property for further analysis, it becomes natural to use tools from the representation theory of finite groups \citep{My_PhD_Thesis, Folland_Harmonic, Barut_Reps}. We proceed to do so by first noting the following result (see Theorem 2.3.18 of \citep{My_PhD_Thesis}, as well as \citep{Bossavit_Old, Bossavit_New}) stated here without proof:
\begin{lemma}
For $\nu, \gamma \in \Gamma_{\mathfrak{N}}$, let \emph{$\chi_{\nu}(\gamma) = \expchar{\nu \gamma}{\mathfrak{N}}$} and let
\begin{align}
\label{peter_weyl_projectors}
P^{\nu} = \displaystyle \frac{1}{\mathfrak{N}}\sum_{\gamma = 0}^{\mathfrak{N} - 1}\overline{\chi_{\nu}(\gamma)}\,T_{\gamma} = \displaystyle \frac{1}{\mathfrak{N}}\sum_{\gamma = 0}^{\mathfrak{N} - 1}\expcharconj{\nu \gamma}{\mathfrak{N}}\,T_{\gamma}.
\end{align}
The operators $P^{\nu}$ are projection operators on $\mathsf{L}^2({\cal C})$, and the ranges of these projectors $V^{\nu} = P^{\nu}\big(\mathsf{L}^2({\cal C})\big)$ form closed, mutually orthogonal subspaces of $\mathsf{L}^2({\cal C})$ for different values of $\nu$. The following direct sum decomposition holds:
\begin{align}
\label{direct_sum_subspaces}
\mathsf{L}^2({\cal C}) = \bigoplus_{\nu = 0}^{\mathfrak{N}-1} V^{\nu}\,,
\end{align}
and further $f \in V^{\nu}$ if and only if it obeys the condition:
\begin{align}
\label{Bloch_cyclic}
T_{\gamma}f ({\bf x}) = f\big((\mathfrak{R}^{\gamma})^{-1}{\bf x}\big) = \chi_\nu(\gamma)f({\bf x})\,.
\end{align}
\label{lemma:rep_theory_result}
\end{lemma}
The complex scalars $\displaystyle \chi_{\nu}(\gamma) $ are the so called characters of the group ${\cal G}$, and the projectors $P^{\nu}$ are the associated Peter-Weyl projectors \citep{Folland_Harmonic}. The range $V^{\nu}$ of each projector forms an invariant subspace of $\mathsf{L}^2({\cal C})$ . A useful corollary of the above result follows immediately by looking at the case $\nu = 0$:
\begin{corollary}
For any given scalar function $f$ on ${\cal C}$, the function $\displaystyle\sum_{\gamma = 0}^{\mathfrak{N}-1} f(\mathfrak{R}^{\gamma} {\bf x})$ is group invariant.
\label{coro:group_invariance}
\end{corollary}
\noindent The condition $T_\gamma \mathfrak{H} (T_\gamma)^{-1} = \mathfrak{H}$ implies that the subspaces $V^{\nu}$ are invariant subspaces for the operator $\mathfrak{H}$ as well. This allows us to \emph{block-diagonalize} $\mathfrak{H}$ so that we may write:
\begin{align}
\mathfrak{H} = \bigoplus_{\nu = 0}^{\mathfrak{N}-1} \mathfrak{H}^{\nu}\,,
\label{block_diagonal_hamil}
\end{align}
with each $\mathfrak{H}^{\nu} = \mathfrak{H} P^{\nu}$ denoting the projection of $\mathfrak{H}$ onto the invariant subspace $V^{\nu}$. As a result, the eigenstates of $\mathfrak{H}$ may be computed by solving each of the projected problems associated with $\mathfrak{H}^{\nu}, \nu \in \Gamma_{\mathfrak{N}}$. Consequently, the eigenfunctions of $\mathfrak{H}$ obey the cyclic-Bloch condition shown in Eq.~\ref{Bloch_cyclic}. We may summarize this discussion and rigorously establish the result as follows:
\begin{theorem}[Cyclic Bloch Theorem]
Given a potential \emph{$V_\textrm{eff}({\bf x})$} which is continuous on ${\cal C}$ and invariant under the cyclic group ${\cal G}$, there exists a basis of $\mathsf{L}^2({\cal C})$ such that every eigenstate $v$ of the operator \emph{$\mathfrak{H} = -\frac{1}{2} \Delta + V_\textrm{eff}$} obeys the following condition (for any $\gamma \in \Gamma_{\mathfrak{N}}$):
\begin{align}
\label{Bloch_cyclic_in_theorem}
v\big((\mathfrak{R}^{\gamma})^{-1}{\bf x}\big) = \chi_\nu(\gamma)v({\bf x}) = \expchar{\nu \gamma}{\mathfrak{N}} v({\bf x})\,,
\end{align}
for some $\nu \in \Gamma_{\mathfrak{N}}$. Conversely, for every $\nu \in \Gamma_{\mathfrak{N}}$, there exists at least one eigenvector of $\mathfrak{H}$ obeying the above condition.
\label{thm:Bloch_Theorem}
\end{theorem}
\begin{myproof}
Since $\mathfrak{H}$ commutes with each of the operators in $\{ T_\gamma: \gamma \in \Gamma_{\mathfrak{N}}\}$, and since the projectors $P^{\nu}$ introduced in Eq.~\ref{peter_weyl_projectors} are linear combinations of the operators in $\{ T_\gamma: \gamma \in \Gamma_{\mathfrak{N}}\}$, $\mathfrak{H}$ also commutes with each of the projectors. This means that $\mathfrak{H}$ leaves the range $V^{\nu}$ of each of the projectors invariant. This is because if $f \in V^{\nu}$, then $P^{\nu} f = f$ and further:
\begin{align}
P^{\nu} (\mathfrak{H} f) = \mathfrak{H} (P^{\nu} f) = \mathfrak{H} f\,,
\end{align}
which implies that $\mathfrak{H} f \in V^{\nu}$. Note that this also implies that $\mathfrak{H}$ leaves $(V^{\nu})^{\perp}$ --- the orthogonal complement of $V^{\nu}$ --- invariant as well. This can be easily checked by looking at the projection operator ${\cal I} - P^{\nu}$ (with ${\cal I}$ denoting the identity operator on $\mathsf{L}^2({\cal C})$), which projects onto the orthogonal complement of $V^{\nu}$.
The above conditions imply that $\mathfrak{H}$ must have at least one eigenvector $v \in V^{\nu}$ for every $\nu \in \Gamma_{\mathfrak{N}}$. To establish this (for any $\nu \in \Gamma_{\mathfrak{N}}$), we first recall that the eigenvectors of $\mathfrak{H}$ form a complete orthonormal basis set in $\mathsf{L}^2({\cal C})$. Since $V^{\nu}$ is a closed subspace of $\mathsf{L}^2({\cal C})$, there exists an eigenvector $v$ of $\mathfrak{H}$ which is not completely orthogonal to $V^{\nu}$. We write this eigenvector as $v = v_1 + v_2$ with $v_1 \in V^{\nu}, v_1 \neq 0$ and $v_2 \in (V^{\nu})^{\perp}$ . Subsequently, we have:
\begin{align}
\label{eigenvector_calc}
\mathfrak{H} v = \mathfrak{H} (v_1 + v_2) = \lambda v = \lambda(v_1 + v_2)\,,
\end{align}
which implies that $\mathfrak{H} v_1 - \lambda v_1 = \mathfrak{H} v_2 - \lambda v_2$. Since $\mathfrak{H}$ leaves both $V^{\nu}$ and its orthogonal complement invariant, the left hand side of this expression lies in $V^{\nu}$ while the right hand side lies in $(V^{\nu})^{\perp}$. Since $V^{\nu} \cap (V^{\nu})^{\perp} = \{0\}$, it must be that $\mathfrak{H} v_1 = \lambda v_1$ and $\mathfrak{H} v_2 = \lambda v_2$ and so, (possibly after an appropriate choice of basis of $\mathsf{L}^2({\cal C})$) $v_1 \in V^{\nu}$ and $v_2 \in (V^{\nu})^{\perp}$ are individual eigenvectors of $\mathfrak{H}$, as required.\footnote{We may now repeat this argument for the orthogonal complement of $v_1$ within $V^{\nu}$ --- since this subspace is also left invariant by $\mathfrak{H}$, an eigenvector of $\mathfrak{H}$ can be found in it. Proceeding in this fashion (i.e., considering the orthogonal complement within $V^{\nu}$ of the linear span of the eigenvectors found so far), we can in fact show the existence of infinitely many eigenvectors of $\mathfrak{H}$ in $V^{\nu}$ as long as the subspace $V^{\nu}$ is infinite-dimensional.}
Next, we observe that by suitable choice of an orthonormal basis of $\mathsf{L}^2({\cal C})$, every eigenvector of $\mathfrak{H}$ can be made to lie in one of the subspaces $V^{\nu}$. To see this, we note that since the decomposition of $\mathsf{L}^2({\cal C})$ into the invariant subspaces $V^{\nu}$ is exhaustive (Eq.~\ref{direct_sum_subspaces}), every eigenvector of $\mathfrak{H}$ either lies in one of the subspaces $V^{\nu}$ or it can be written as a linear combination of eigenvectors from these subspaces. If the eigenstates of $\mathfrak{H}$ are non-degenerate (i.e., there is no repeated eigenvalue), the second option is ruled out as follows: writing the eigenvector as $ v = \displaystyle \sum_{\nu = 0}^{\mathfrak{N} - 1} c_{\nu} v_{\nu}$, with $v_{\nu} \in V^{\nu}, c_{\nu} \in \mathbb{C}$, a calculation similar to the one in the previous paragraph shows that each individual component $v_{\nu}$ must be an eigenvector as well (since $V^{\nu} \cap V^{\nu'} = \{0\}$ for $\nu \neq \nu'$). However, this would violate the orthogonality of the eigenvectors of $\mathfrak{H}$ since $\innprod{v}{v_{\nu}}{} \neq 0$ unless $c_{\nu} = 0$. In the presence of degeneracies,\footnote{While it might be tempting to simplify the statement and proof of Theorem \ref{thm:Bloch_Theorem} by assuming lack of any degeneracies, examples appear to indicate that degeneracies in the spectrum of $\mathfrak{H}$ might generically appear (e.g., see \citep{My_Elliott_Symmetry_paper}). Thus, we find it useful to add the caveat that the cyclic-Bloch boundary conditions (Eq.~\ref{Bloch_cyclic_in_theorem}) can be expected to hold for any eigenstate of $\mathfrak{H}$ provided an appropriate choice of basis vectors of $\mathsf{L}^2({\cal C})$ has been made. We are grateful to the anonymous reviewer for pointing out this subtle issue to us.\\} any linear combination of the degenerate eigenstates from different invariant subspaces is still an eigenstate for $\mathfrak{H}$. In this case, we may choose the basis vectors of $\mathsf{L}^2({\cal C})$ appropriately so that the individual eigenstates from the different invariant subspaces (and not their linear combinations) are identified as the eigenvectors of $\mathfrak{H}$.
Finally, since every eigenvector $v$ of $\mathfrak{H}$ lies in one of the invariant subspaces $V^{\nu}$, the condition in Eq.~\ref{Bloch_cyclic_in_theorem} can be simply identified as the characterization of the subspaces presented in Eq.~\ref{Bloch_cyclic}, thus establishing the theorem.\footnote{The proof presented here is slightly different in technical details from that presented in \citep{My_MS_Thesis, My_PhD_Thesis}. In the latter case, a more direct use of Schur's Lemma \citep{Folland_Harmonic}, as applied to the resolvent operator of the Hamiltonian was made.}
\end{myproof}
We may use the Cyclic Bloch theorem to reduce the physical domain of the eigenvalue problem associated with $\mathfrak{H}$ from ${\cal C}$ to the symmetry cell ${\cal D}_{{\cal G}}$. Roughly speaking, since every point in ${\cal C}$ can be obtained as the action of a rotation matrix $\mathfrak{R}^{\gamma}$ (for some $\gamma \in \Gamma_{\mathfrak{N}}$) acting on a point in the fundamental domain ${\cal D}_{{\cal G}}$, it suffices to specify a function $f \in V^{\nu}$ (in particular, an eigenvector $v\in V^{\nu}$ of $\mathfrak{H}$) using only the points within the fundamental domain, as long as boundary conditions consistent with Eq.~\ref{Bloch_cyclic} are prescribed on $\partial{\cal D}_{{\cal G}}^{C}$. We summarize this result as follows:\footnote{It becomes necessary to work with functions in $\mathsf{H}^1_0({\cal C})$ so that boundary conditions on $\partial{\cal D}_{{\cal G}}$ can be interpreted in the trace sense. The eigenvectors of $\mathfrak{H}$ are actually more regular than functions in $\mathsf{H}^1_0({\cal C})$.}
\begin{proposition}
\label{prop:invariance_to_BC}
Let ${\bf x} \in \partial{\cal D}_{{\cal G}}^{C}$ such that $\mathfrak{R}^{\gamma}{\bf x} \in \partial{\cal D}_{{\cal G}}^{C}$ for some $\gamma \in \Gamma_{\mathfrak{N}}$ and \footnote{For the cylindrical geometries that we are considering, $\gamma \in \{0,1,\mathfrak{N}-1\}$. } let \emph{$\text{cl.}({\cal D}_{{\cal G}})$} denote the (topological) closure of the fundamental domain. A function $f \in \big(V^{\nu} \cap \mathsf{H}^1_0({\cal C}) \big)$ if and only if the restriction of the function to \emph{$\text{cl.}({\cal D}_{{\cal G}})$}, denoted by \emph{$\tilde{f} = f|_{\text{cl.}({\cal D}_{{\cal G}})}$}, satisfies
\begin{align}
\label{Bloch_cyclic_restatement}
\tilde{f}(\mathfrak{R}^{\gamma}{\bf x}) = \overline{\chi_{\nu}(\gamma)} \, \tilde{f}({\bf x}) = \expcharconj{\nu \gamma}{\mathfrak{N}}\,\tilde{f}({\bf x})\,\,\,\, \text{for} \,\,\,\,{\bf x} \in \partial{\cal D}_{{\cal G}}^{C}\,.
\end{align}
\end{proposition}
\begin{myproof}
Since the function space $\textrm{C}_c^0({\cal C})$ is dense in $\mathsf{H}^1_0({\cal C})$, it suffices to work with such functions. If $f \in V^{\nu} \cap \textrm{C}_c^0({\cal C})$, then it obeys Eq.~\ref{Bloch_cyclic} for all ${\bf x}\in {\cal C}$ and in particular on the surfaces $\partial{\cal D}_{{\cal G}}^{C} \subset {\cal C}$. On the other hand, if $\tilde{f}$ is a continuous function defined over the fundamental domain such that it obeys Dirichlet boundary conditions on $\partial{\cal D}_{{\cal G}}^{0}$ and the condition in Eq.~\ref{Bloch_cyclic_restatement} on $\partial{\cal D}_{{\cal G}}^{C}$ for some $\gamma, \nu \in \Gamma_{\mathfrak{N}}$, we may define an extension $f$ of $\tilde{f}$ to all of ${\cal C}$ as follows. For any ${\bf y} \in {\cal C}$ (such that ${\bf y}$ does not lie on the axis ${\bf e}_3$) we may identify a point ${\bf x} \in {\cal D}_{{\cal G}}$ and a rotation matrix $\mathfrak{R}^{\gamma_2}\in {\cal G}$ such that ${\bf y} = \mathfrak{R}^{\gamma_2} {\bf x}$. Then, we set
\begin{align}
\label{how_to_define_f}
f({\bf y}) = f(\mathfrak{R}^{\gamma_2} {\bf x}) = \expcharconj{\nu \gamma_2}{\mathfrak{N}}\,\tilde{f}({\bf x})\,,
\end{align}
and $f = 0$ on $\partial {\cal C}$. The function $f$ defined this way is compactly supported on ${\cal C}$ and is continuous across each of the surfaces $\mathfrak{R}^{\gamma_1}(\partial{\cal D}_{{\cal G}})$ for any $\gamma_1 \in \Gamma_{\mathfrak{N}}$, by virtue of Eq.~\ref{Bloch_cyclic_restatement}. By the continuity of $\tilde{f}$ in the interior of ${\cal D}_{{\cal G}}$, $f$ is also continuous in the interior of $\mathfrak{R}^{\gamma_1}({\cal D}_{{\cal G}})$. Thus, $f \in \textrm{C}_c^0({\cal C})$. Further, Eq.~\ref{how_to_define_f} implies that $f$ obeys Eq.~\ref{Bloch_cyclic} and therefore by Lemma \ref{lemma:rep_theory_result}, $f \in V^{\nu}$. This establishes the sought result.
\end{myproof}
As a consequence of the above result, the Cyclic Bloch Theorem may be reinterpreted to mean that the eigenstates of $\mathfrak{H}$ can be obtained by solving $\mathfrak{N}$ independent eigenvalue problems associated with the operators $\mathfrak{H}^{\nu}$ ($\nu \in \Gamma_{\mathfrak{N}}$) over the fundamental domain. Each problem corresponds to a projection of the original problem to one of the subspaces $V^{\nu}$, and therefore it obeys zero-Dirichlet and cyclic-Bloch (Eq.~\ref{Bloch_cyclic_restatement}) boundary conditions on the surfaces $\partial{\cal D}_{{\cal G}}^{0}$ and $\partial {\cal D}_{{\cal G}}^{C}$, respectively. Generically, each of these problems results in an infinite sequence of orthonormal eigenstates.
It is worth noting that Theorem \ref{thm:Bloch_Theorem} represents an extension of the Bloch theorem used in classical solid state physics \citep{Ashcroft_Mermin, Reed_Simon4, Odeh_Keller} to cyclic structures. A further generalization of this result to the case of structures with more complex symmetries (or Objective Structures) can be found in the first author's thesis work \citep{My_PhD_Thesis, My_MS_Thesis}. Theorem \ref{thm:Bloch_Theorem} appears in a somewhat more rudimentary form with certain key details omitted in \cite{Sattlegger_Thesis}. The gist of the result also appears in the physics literature \citep{Pekka_1,Pekka_2} --- however, only in the context of tight binding and not Kohn-Sham DFT --- wherein the derivation follows a line of heuristic reasoning (aimed at guessing the correct extension of the classical Bloch Theorem), and therefore lacks the mathematical rigor adopted here.
\subsubsection{Consistency of the SCF method after symmetry cell reduction} \label{Subsubsec:ConsistencySCF}
While performing the symmetry cell reduction of the linear eigenvalue problem, we have assumed that the electron density calculated during the previous SCF iteration is group invariant. It is therefore important to ensure that the electron density remains group invariant throughout the complete fixed-point iteration. To do so, we note that as a consequence of the reduction achieved by the Cyclic Bloch Theorem, the electron density in any SCF iteration can be calculated using the expression:\footnote{Henceforth, a tilde will be used to denote that the quantity has been defined over the fundamental domain ${\cal D}_{{\cal G}}$. }
\begin{align} \label{Eqn:rho:Bloch-cyclic}
\tilde{\rho}({\bf x})&=2 \sum_{\nu = 0}^{\mathfrak{N} - 1}\sum_{i=1}^{N_{\text{s}}} {f}_i^{\nu}\abs{\tilde{\phi}_i^{\nu}({\bf x})}^2 \,,
\end{align}
where $\tilde{\phi}_i^{\nu}$ are the eigenfunctions of $\mathfrak{H}^{\nu}$, and ${f}_i^{\nu}$ are the corresponding occupations obtained using the Fermi-Dirac smearing.
An inspection of Eq.~\ref{Eqn:rho:Bloch-cyclic} reveals that the electron density calculated in any SCF iteration satisfies (for any $\gamma \in \Gamma_{\mathfrak{N}}$):
\begin{align}
\tilde{\rho}(\mathfrak{R}^{\gamma}{\bf x}) =2 \sum_{\nu = 0}^{\mathfrak{N} - 1}\sum_{i=1}^{N_{\text{s}}} {f}_i^{\nu}\abs{\tilde{\phi}_i^{\nu}(\mathfrak{R}^{\gamma}{\bf x})}^2
=2 \sum_{i=1}^{N_{\text{s}}} {f}_i^{\nu}\abs{\expcharconj{\nu \gamma}{\mathfrak{N}}\,\tilde{\phi}_i^{\nu}({\bf x})}^2 = 2 \sum_{i=1}^{N_{\text{s}}} {f}_i^{\nu}\abs{\tilde{\phi}_i^{\nu}({\bf x})}^2 = \tilde{\rho}({\bf x})\,.
\label{density_invariant_calc}
\end{align}
Thus, the extension of this density to the entire domain ${\cal C}$ is continuous and is group invariant. Consequently, it follows from Proposition \ref{prop:invariant_potential} and Theorem \ref{thm:Bloch_Theorem} that the Kohn-Sham Hamiltonian resulting from this density can once again be reduced to the fundamental domain with the aid of the cyclic-Bloch boundary conditions. This ensures that SCF iterations remain consistent provided one starts performing these iterations with an electron density that is group invariant. In principle, such an electron density may be generated by considering any $\rho' \in \textrm{C}_c^0({\cal C})$ obeying $\displaystyle \int_{{\cal C}}\rho'({\bf x}) \, \mathrm{d{\bf x}}= \mathfrak{N}N_{\text{e}}$ and then constructing
\begin{align}
\rho_{0}({\bf x})= \frac{1}{\mathfrak{N}}\sum_{\gamma = 0}^{\mathfrak{N}-1}\rho'(\mathfrak{R}^{\gamma}{\bf x})\,.
\end{align}
It follows from Corollary \ref{coro:group_invariance} that the guess electron density $\rho_{0}$ is group invariant, and therefore its restriction to the fundamental domain serves as a suitable starting point.
\subsubsection{Reduction of the free energy and the atomic forces}
Having performed a symmetry cell reduction of the self-consistent Kohn-Sham equations, we now express the cyclic structure's free energy at the electronic ground-state, as well as the Hellmann Feynman forces on the atoms, in terms of quantities defined over the fundamental domain.
To simplify the free-energy, we use the cyclic-Bloch reduction to rewrite the first (band energy) and last (electronic entropy energy) terms of Eq. \ref{Harris_Foulkes_Energy} as follows:
\begin{align}
{E}_{\text{band}} &= 2\sum_{\nu =0}^{\mathfrak{N}-1}\sum_{i = 1}^{N_{\text{s}}} f_i^{\nu} \lambda_i^{\nu} \,, \\
{E}_{\text{entropy}} &= 2\,\sigma \sum_{\nu=0}^{\mathfrak{N}-1}\sum_{i=1}^{N_{s}} \big( f_i^{\nu} \log f_i^{\nu} + (1-f_i^{\nu}) \log (1 - f_i^{\nu}) \big) \,,
\end{align}
where $\lambda_i^{\nu}$ are the eigenvalues of $\mathfrak{H}^{\nu}$ at self-consistency. In addition, ${f}_i^{\nu}$ are the corresponding occupations obtained using the Fermi-Dirac smearing. The second, third, and fourth terms of Eq.~\ref{Harris_Foulkes_Energy} are integrals over the complete domain ${\cal C}$ with integrands that are group invariant. Consequently, they can be reduced to the fundamental domain as follows:
\begin{align}
\int_{{\cal C}}\!\varepsilon_{\text{xc}} (\rho({\bf x})) \rho({\bf x}) \, \mathrm{d {\bf x}} &= \mathfrak{N} \int_{{\cal D}_{{\cal G}}}\!\varepsilon_{\text{xc}} (\tilde{\rho}({\bf x})) \tilde{\rho}({\bf x}) \, \mathrm{d {\bf x}} \,, \\
\int_{{\cal C}}\!V_{\text{xc}}(\rho({\bf x}))\rho({\bf x})\,\mathrm{d{\bf x}} &= \mathfrak{N} \int_{{\cal D}_{{\cal G}}}\!V_{\text{xc}}(\tilde{\rho}({\bf x}))\tilde{\rho}({\bf x})\,\mathrm{d{\bf x}} \,, \\
\frac{1}{2} \int_{{\cal C}}\!\big (b({\bf x}, {\cal S}_{\cal_G, {\cal P}})-\rho({\bf x})\big) V_{\text{es}}({\bf x}, {\cal S}_{\cal_G, {\cal P}}) \, \mathrm{d{\bf x}} &= \frac{\mathfrak{N}}{2} \int_{{\cal D}_{{\cal G}}}\!\big (\tilde{b}({\bf x}, {\cal P})-\tilde{\rho}({\bf x})\big) \tilde{V}_{\text{es}}({\bf x}, {\cal P}) \, \mathrm{d{\bf x}} \,.
\end{align}
To see how the fifth term in Eq.~\ref{Harris_Foulkes_Energy}, i.e., pseudocharge self energy term:
\begin{align}
{E}_{\text{self}}({\cal S}_{\cal_G, {\cal P}}) = \frac{1}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal C}}\!b_k({\bf x}, {\bf x}_{\gamma,k}) \, V_k({\bf x}, {\bf x}_{\gamma,k}) \,\mathrm{d{\bf x}}\,,
\end{align}
can be reduced to the fundamental domain, we recall Proposition \ref{prop:PseudochargeDensity} (part 1), to write:
\begin{align}
\nonumber
{E}_{\text{self}}({\cal S}_{\cal_G, {\cal P}}) &= \frac{1}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal C}}\!b_k({\bf x}, \mathfrak{R}^{\gamma}{\bf x}_{0,k})\, V_k({\bf x}, \mathfrak{R}^{\gamma}{\bf x}_{0,k})\, \mathrm{d{\bf x}} \\\nonumber
&= \frac{1}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal C}}\!b_k(\mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}, {\bf x}_{0,k})\,V_k(\mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}, {\bf x}_{0,k})\, \mathrm{d{\bf x}} \\ \nonumber
&= \frac{1}{2} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal C}}\,\sum_{\gamma = 0}^{\mathfrak{N}-1} b_k(\mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}, {\bf x}_{0,k})\, V_k(\mathfrak{R}^{\mathfrak{N}-\gamma}{\bf x}, {\bf x}_{0,k})\, \mathrm{d{\bf x}} \\ \nonumber
&= \frac{\mathfrak{N}}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal D}_{{\cal G}}}\!b_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k})\, V_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k})\, \mathrm{d{\bf x}} \\
&= \frac{\mathfrak{N}}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal D}_{{\cal G}}}\!\tilde{b}_k({\bf x}, {\bf x}_{\gamma,k})\,\widetilde{V}_k({\bf x}, {\bf x}_{\gamma,k})\, \mathrm{d{\bf x}}\,,
\end{align}
where the fourth equality is obtained by observing that each of the integrands in the previous expression are group invariant via Corollary \ref{coro:group_invariance}. Note that, $\tilde{b}_k$ and $\widetilde{V}_k$ are the restriction of $b_k$ and $V_k$ to the fundamental domain.
Next, we consider the reduction of the atomic forces. As a consequence of the frame indifference of the free energy $\mathcal{F}({{\cal S}_{\cal_G, {\cal P}}})$, the Hellmann-Feynman force on the atom located at ${\bf x}_{\gamma,k} = \mathfrak{R}^{\gamma} {\bf x}_{0,k}$ is related to the force experienced by the corresponding atom ${{\bf x}_{0,k}} \in {\cal P}$ within the fundamental domain ${\cal D}_{{\cal G}}$ through the relation:\footnote{This result is not restricted only to the case of Kohn-Sham DFT however -- it holds for more general theories such as the many body Schr\"odinger equation in Born-Oppenheimer quantum mechanics \cite{Dumitrica_James_OMD}.}
\begin{align}
\label{force_equivariance}
{\bf f}_{{\bf x}_{\gamma,k}} = (\mathfrak{R}^{\gamma})^T\,{\bf f}_{{\bf x}_{0,k}} = (\mathfrak{R}^{\mathfrak{N}-\gamma})\,{\bf f}_{{\bf x}_{0,k}} \,,
\end{align}
where
\begin{align}
\label{force_formula_revisited}
{\bf f}_{{\bf x}_{0,k}} = \int_{{\cal C}}\!\nabla b_k({\bf x}, {\bf x}_{0,k})\big(V_{\text{es}}({\bf x}, {\cal S}_{\cal_G, {\cal P}}) - V_k({\bf x}, {\bf x}_{0,k}) \big)\,\mathrm{d{\bf x}}\,,
\end{align}
and $V_{\text{es}}$ is the solution of the Poisson problem in Eqs. \ref{poisson_problem} and \ref{Eqn:KS:BC:Ves} at the electron density associated with the (electronic) ground-state . Therefore, it suffices to calculate the forces for only those atoms that are located within the fundamental domain, the expression for which can be rewritten in terms of quantities expressed over the fundamental domain as follows:
\begin{align}
\label{force_formula_first_term_rhs}
{\bf f}_{{\bf x}_{0,k}} &= \sum_{\gamma = 0}^{\mathfrak{N}-1}\mathfrak{R}^{\mathfrak{N}-\gamma}\int_{{\cal D}_{{\cal G}}}\!\nabla b_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k})\big(V_{\text{es}}(\mathfrak{R}^{\gamma}{\bf x},{\cal S}_{\cal_G, {\cal P}}) - V_k(\mathfrak{R}^{\gamma}{\bf x}, {\bf x}_{0,k}) \big)\, \mathrm{d{\bf x}} \nonumber \\
&= \sum_{\gamma = 0}^{\mathfrak{N}-1}\mathfrak{R}^{\mathfrak{N}-\gamma}\int_{{\cal D}_{{\cal G}}}\!\nabla \tilde{b}_k({\bf x}, \mathfrak{R}^{\gamma}{\bf x}_{0,k})\big(\widetilde{V}_{\text{es}}({\bf x},{\cal P}) - \widetilde{V}_k({\bf x}, \mathfrak{R}^{\gamma}{\bf x}_{0,k}) \big)\,\mathrm{d{\bf x}} \nonumber \\
&= \sum_{\gamma = 0}^{\mathfrak{N}-1}\mathfrak{R}^{\mathfrak{N}-\gamma}\int_{{\cal D}_{{\cal G}}}\!\nabla \tilde{b}_k({\bf x}, {\bf x}_{\gamma,k})\big(\widetilde{V}_{\text{es}}({\bf x},{\cal P}) - \tilde{V}_k({\bf x}, {\bf x}_{\gamma,k}) \big)\,\mathrm{d{\bf x}}\,,
\end{align}
where the second equality is obtained by using cyclic boundary conditions:
\begin{align}
\nonumber
V_{\text{es}}(\mathfrak{R}^{\gamma}{\bf x},{\cal S}_{\cal_G, {\cal P}}) = V_{\text{es}}({\bf x},{\cal S}_{\cal_G, {\cal P}})=\widetilde{V}_{\text{es}}({\bf x},{\cal P})\,,
\end{align}
and Proposition \ref{prop:PseudochargeDensity} (part 1).
\subsection{Cyclic DFT: Kohn-Sham problem on the fundamental domain for cyclic structures} \label{Subsec:CyclicDFTEq}
We now summarize the formulation of Cyclic DFT developed in the previous subsections. After the symmetry cell reduction depicted in Fig.~\ref{fig:cyclic_cell_reduction}, the Kohn-Sham equations on the fundamental domain ${\cal D}_{{\cal G}}$ can be written as (for $\nu \in \Gamma_{\mathfrak{N}}$)
\begin{subequations} \label{Eqns:CyclicDFT:KS}
\begin{align}
\label{Cyclic_DFT_1}
\widetilde{K}(\tilde{\rho})\,\tilde{\phi}_i^{\nu}&={\lambda}_i^{\nu}\,\tilde{\phi}_i^{\nu}\;;\;\innprod{\tilde{\phi}_i^{\nu}}{\tilde{\phi}_j^{\nu}}{\mathsf{L}^2({\cal D}_{{\cal G}})}=\frac{1}{\mathfrak{N}}\,\delta_{ij}\,,\\
\label{Cyclic_DFT_2}
\text{where}\quad
\widetilde{K}(\tilde{\rho})&=-\frac{1}{2} \Delta + \widetilde{V}_{\text{es}} + {V}_{\text{xc}}(\tilde{\rho})\,,\\
\text{with}\quad
\label{Cyclic_DFT_4}
\tilde{\rho}({\bf x})&=2 \sum_{\nu = 0}^{\mathfrak{N} - 1}\sum_{i=1}^{N_{\text{s}}} {f}_i^{\nu}\abs{\tilde{\phi}_i^{\nu}({\bf x})}^2\,,\\
\label{Cyclic_DFT_3}
\text{and} \quad
-\frac{1}{4\pi}\Delta \widetilde{V}_{\text{es}} &= \,(\tilde{\rho} + \tilde{b})\,,\, \tilde{b} = b|_{{\cal D}_{{\cal G}}}\,. \\
\label{Cyclic_DFT_5}
\text{Here,} \quad
{f}_i^{\nu} &= \left( 1 + \exp\left( \frac{{\lambda}_i^{\nu} - \lambda_f}{\sigma} \right) \right)^{-1},\\
\label{Cyclic_DFT_6}
\text{with $\lambda_f$ s.t.} \quad &2 \sum_{\nu = 0}^{\mathfrak{N} - 1}\sum_{i=1}^{N_{\text{s}}} {f}_i^{\nu} = \mathfrak{N} N_{\text{e}}\,.
\end{align}
\end{subequations}
The corresponding boundary conditions on $ \partial{\cal D}_{{\cal G}} = \partial{\cal D}_{{\cal G}}^{C} \cup \partial{\cal D}_{{\cal G}}^{0}$ are
\begin{subequations} \label{Eqns:CyclicDFT:KS:BC}
\begin{align}
\text{for} \,\,\, {\bf x} \in \partial{\cal D}_{{\cal G}}^{C} \,, \quad \tilde{\phi}_i^{\nu}(\mathfrak{R}^{\gamma}{\bf x}) &= \expcharconj{\nu \gamma}{\mathfrak{N}}\,\tilde{\phi}_i^{\nu}({\bf x}) \,, \label{Eq:CyclicDFT:phi_cyclicBlochBC} \\
\tilde{\rho}(\mathfrak{R}^{\gamma}{\bf x}) &= \tilde{\rho}({\bf x}) \,, \\
\text{and} \quad \widetilde{V}_{\text{es}} (\mathfrak{R}^{\gamma}{\bf x} {\cal P}) &= \widetilde{V}_{\text{es}}({\bf x},{\cal P}) \,. \label{Eq:CyclicDFT:Ves:cyclic} \\
\text{For} \,\,\, {\bf x} \in \partial{\cal D}_{{\cal G}}^{0} \,, \quad \tilde{\phi}_i^{\nu}({\bf x}) &= 0 \,, \label{Eq:CyclicDFT:phi_zeroBC} \\
\tilde{\rho}({\bf x}) &= 0 \,, \\
\text{and} \quad \widetilde{V}_{\text{es}}({\bf x},{\cal P}) &= \sum_{\gamma = 0}^{\mathfrak{N}-1} \int_{{\cal D}_{{\cal G}}}\!\frac{\tilde{b}({\bf y}, {\cal P})+\tilde{\rho}({\bf y})}{\abs{{\bf x}-\mathfrak{R}^{\gamma}{\bf y}}}\; \mathrm{d{\bf y}} \,. \label{Eqn:SymmetryReducedBC_Ves}
\end{align}
\end{subequations}
In each iteration of the SCF method in Cyclic DFT, the following linear eigenproblems (there are $\mathfrak{N}$ of them, one for each $\nu$):
\begin{align} \label{Eq:CyclicDFT:linearEigenproblem}
\left( -\frac{1}{2} \Delta + \widetilde{V}_{\text{eff}} \right) \tilde{\phi}^{\nu}_{i} = \lambda_i^{\nu} \tilde{\phi}^{\nu}_{i} \,, \quad \nu=0, 1, \ldots, \mathfrak{N}-1 \,,
\end{align}
where
\begin{align}
\label{effective_potential_reduced}
\widetilde{V}_{\textrm{eff}}({\bf x}) = \widetilde{V}_{\textrm{es}}(\tilde{\rho}({\bf x}), {\cal P}) + V_{{\textrm{xc}}}(\tilde{\rho}({\bf x})) \,,
\end{align}
need to be solved for the lowest $N_s$ eigenvalues and corresponding eigenfunctions. The eigenfunctions are subject to the cyclic-Bloch boundary conditions in Eq.~\ref{Eq:CyclicDFT:phi_cyclicBlochBC} and zero-Dirichlet boundary conditions in Eq.~\ref{Eq:CyclicDFT:phi_zeroBC}. The symmetry cell reduced electrostatic potential $\widetilde{V}_{\textrm{es}}$ is obtained by solving the Poisson equation in Eq.~\ref{Cyclic_DFT_3} subject to the cyclic boundary conditions in Eq.~\ref{Eq:CyclicDFT:Ves:cyclic} and Dirichlet boundary conditions in Eq.~\ref{Eqn:SymmetryReducedBC_Ves}.
At the electronic ground-state, the free energy \emph{per unit cell} of the cyclic structure, i.e., the quantity $\widetilde{\mathcal{F}}({\cal P}) ={\mathcal{F}}({{\cal S}_{\cal_G, {\cal P}}})/\mathfrak{N}$, can be expressed in terms of quantities over the fundamental domain as:\footnote{If the term ${E}_{\text{overlap}}({\cal S}_{\cal_G, {\cal P}})$ is included in the expression in Eq.~\ref{Harris_Foulkes_Energy}, it is also necessary to include a term of the form $\widetilde{E}_{\text{overlap}}({\cal P}) ={E}_{\text{overlap}}({\cal S}_{\cal_G, {\cal P}}) / \mathfrak{N}$ in Eq.~\ref{Harris_Foulkes_Energy_per_cell}. This term can be expressed in terms of integrals over the fundamental domain, analogous to other the terms on the right hand side of Eq.~\ref{Harris_Foulkes_Energy_per_cell}.}
\begin{align}
\nonumber
\widetilde{\mathcal{F}}({\cal P}) =& \;\frac{2}{\mathfrak{N}}\bigg(\sum_{\nu =0}^{\mathfrak{N}-1}\sum_{i = 1}^{N_{\text{s}}} f_i^{\nu} \lambda_i^{\nu}\bigg) + \int_{{\cal D}_{{\cal G}}}\!\varepsilon_{\text{xc}} (\tilde{\rho}({\bf x})) \tilde{\rho}({\bf x}) \, \mathrm{d {\bf x}} - \int_{{\cal D}_{{\cal G}}}\!{V}_{\text{xc}}(\tilde{\rho}({\bf x}))\tilde{\rho}({\bf x})\, \mathrm{d{\bf x}} \\\nonumber
&+\frac{1}{2} \int_{{\cal D}_{{\cal G}}}\!\big (\tilde{b}({\bf x}, {\cal P})-\tilde{\rho}({\bf x})\big) \widetilde{V}_{\text{es}}({\bf x}, {\cal P}) \, \mathrm{d{\bf x}}-\frac{1}{2} \sum_{\gamma = 0}^{\mathfrak{N}-1} \sum_{k = 1}^{M_{{\cal P}}} \int_{{\cal D}_{{\cal G}}}\!\widetilde{V}_k({\bf x}, {\bf x}_{\gamma,k})\,\tilde{b}_k({\bf x}, {\bf x}_{\gamma,k})\,\mathrm{d{\bf x}}\\
&+ \frac{2\,\sigma}{\mathfrak{N}} \bigg(\sum_{\nu=0}^{\mathfrak{N}-1}\sum_{i=1}^{N_{s}} \big( f_i^{\nu} \log f_i^{\nu} + (1-f_i^{\nu}) \log (1 - f_i^{\nu}) \big)\bigg)\,.
\label{Harris_Foulkes_Energy_per_cell}
\end{align}
The corresponding Hellmann-Feynman atomic force takes the form:\footnote{It might be necessary to include the term ${\bf f}_{\text{overlap}}({\bf x}_{0,k})$ --- described earlier in footnote \ref{footnote:fn_1} --- in Eq.~\ref{force_formula_revisited} to account for possible overlaps of pseudocharges. This quantity can be expressed in terms of integrals over the fundamental domain, analogous to the other terms in Eq.~\ref{force_formula_FD}.}
\begin{align}
\label{force_formula_FD}
{\bf f}_{{\bf x}_{0,k}} = \sum_{\gamma = 0}^{\mathfrak{N}-1}\mathfrak{R}^{\mathfrak{N}-\gamma}\int_{{\cal D}_{{\cal G}}}\!\nabla \tilde{b}_k({\bf x}, {\bf x}_{\gamma,k})\big(\widetilde{V}_{\text{es}}({\bf x},{\cal P}) - \widetilde{V}_k({\bf x}, {\bf x}_{\gamma,k}) \big)\,\mathrm{d{\bf x}}\,.
\end{align}
The problem of determining the equilibrium geometry\footnote{Once again, this should be interpreted in the sense of computing local structural minima (i.e. local minima in the ground state electronic free energy as the atomic positions are varied).} of the cyclic structure as described by Eq.~\ref{geometry_optimization_original} can be reformulated as:
\begin{align}
\label{geometry_optimization_reduced}
\infm{\{{\bf x}_{0,k} \in {{\cal P}}\}}{\widetilde{\mathcal{F}}({\cal P})}\,.
\end{align}
Indeed, it follows from Eq.~\ref{force_equivariance} that if the atoms in the fundamental domain are in equilibrium, so is every other atom in the cyclic structure.
\begin{figure}[ht]
\begin{subfigure}{\linewidth}
\centering
\includegraphics[scale=0.4]{./Fig_1.eps}
\caption{The original problem for the cyclic structure posed in the cylinder ${\cal C}$. The orbitals and the electron density are subjected to zero-Dirichlet boundary conditions on the surfaces $\text{S}_1,\text{S}_2,\text{S}_3$. On these surfaces, the electrostatic potential is given by Eq.~\ref{Eqn:KS:BC:Ves}. Note that $\partial {\cal C} = \text{S}_1 \cup \text{S}_2 \cup \text{S}_3$ in the notation of the text.}
\label{fig1:subfig_a}
\end{subfigure}\\[1ex]
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[scale=0.4]{./Fig_2.eps}
\caption{The symmetry cell reduction results in cyclic-Bloch boundary conditions for the orbitals and cyclic boundary conditions for the density on the surfaces $\text{B}_1, \text{B}_2$. Both the electron density and the orbitals are subjected to zero-Dirichlet boundary conditions on the surfaces $\text{D}_1,\text{D}_2,\text{D}_3$.}
\label{fig1:subfig_b}
\end{subfigure}%
$\quad$
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[scale=0.4]{./Fig_3.eps}
\caption{Another view of the cyclic unit cell ${\cal D}$ and the cyclic cell reduction. Bottom surface $\text{D}_2$ has not been shown. The electrostatic potential obeys cyclic boundary conditions on the surfaces $\text{B}_1, \text{B}_2$ and is given by Eq.~\ref{Eqn:SymmetryReducedBC_Ves} on $\text{D}_1,\text{D}_2,\text{D}_3$. Note that $\partial{\cal D}_{{\cal G}}^{0}= D_1 \cup D_2 \cup D_3$, while $\partial{\cal D}_{{\cal G}}^{C}= B_1 \cup B_2$ in the notation of the text. }
\label{fig1:subfig_c}
\end{subfigure}
\caption{Schematic of cyclic unit cell reduction.}
\label{fig:cyclic_cell_reduction}
\end{figure}
\subsection{Symmetry-adapted finite-difference discretization} \label{Subsec:FD}
Having formulated the governing equations of Cyclic DFT, we now describe a discretization strategy. We fix ${\cal C}$ to be an annular cylinder\footnote{The use of an annular region enhances the efficiency of the discretization scheme not only due to reduced degrees of freedom, but also due to the improved conditioning of the various matrices resulting from the discretization. In addition, it avoids the coordinate system singularity along the line $r = 0$, which would otherwise require special treatment (see for e.g. \citep{mohseni2000numerical, lai2001note}).} with axis ${\bf e}_3$, height $H$, inner radius $R_{\text{in}}$, and outer radius $R_{\text{out}}$. Let the cyclic structure have a symmetry group of order $\mathfrak{N}$. It follows that the fundamental domain ${\cal D}_{{\cal G}}$ is a sector of the annular cylinder with angle $2\pi / \mathfrak{N}$. To ensure that the resulting discretization scheme is compatible with the cyclic symmetry of the system, we work in cylindrical coordinates $(r,\vartheta,z)$. In this setting\footnote{In the notation introduced earlier, the points with cylindrical coordinates $\big(r \in [R_{\text{in}}, R_{\text{out}}],\vartheta = 0,z \in [0,H]\big)$ and $\big(r \in [R_{\text{in}}, R_{\text{out}}],\vartheta = \frac{2\pi}{\mathfrak{N}},z \in [0,H]\big)$ form the surfaces $\partial{\cal D}_{{\cal G}}^{C}$, while the points $\big(r \in [R_{\text{in}}, R_{\text{out}}],\vartheta \in [0,\frac{2\pi}{\mathfrak{N}}],z = 0\big)$, $\big(r \in [R_{\text{in}}, R_{\text{out}}],\vartheta \in [0,\frac{2\pi}{\mathfrak{N}}],z = H\big)$, $\big(r = R_{\text{in}},\vartheta \in [0,\frac{2\pi}{\mathfrak{N}}],z \in [0,H]\big)$ and $\big(R_{\text{out}},\vartheta \in [0,\frac{2\pi}{\mathfrak{N}}],z \in [0,H]\big)$ form the surfaces $\partial{\cal D}_{{\cal G}}^{0}$.}, the linear eigenproblems in Eq.~\ref{Eq:CyclicDFT:linearEigenproblem} take the form:
\begin{align}
\nonumber
-\frac{1}{2} \bigg(\hpd{\tilde{\phi}^{\nu}_i(r,\vartheta,z)}{r}{2} &+ \frac{1}{r} \pd{\tilde{\phi}^{\nu}_i(r,\vartheta,z)}{r} + \frac{1}{r^2}\hpd{{\tilde{\phi}}^{\nu}_i(r,\vartheta,z)}{\vartheta}{2}+ \hpd{\tilde{\phi}^{\nu}(r,\vartheta,z)}{z}{2}\bigg )\\
&+\widetilde{V}_{\text{eff}}(r,\vartheta,z)\,\tilde{\phi}^{\nu}_i(r,\vartheta,z) = \lambda_{i}^{\nu}\,\tilde{\phi}^{\nu}_i(r,\vartheta,z)\,,
\label{cylindrical_linear_eigen}
\end{align}
with boundary conditions (Eqs.~\ref{Eq:CyclicDFT:phi_cyclicBlochBC} and \ref{Eq:CyclicDFT:phi_zeroBC}):
\begin{align}
\label{Bloch_BC_in_cylindrical_coordinates}
\tilde{\phi}^{\nu}_i(r, \vartheta = \frac{2\pi}{\mathfrak{N}},z) &= \expcharconj{\nu \gamma}{\mathfrak{N}}\,\tilde{\phi}^{\nu}_i(r, \vartheta = 0,z)\,,\\
\label{Dirichlet_BC_in_cylindrical_coordinates}
\tilde{\phi}^{\nu}_i(r, \vartheta ,z = 0) = \tilde{\phi}^{\nu}_i(r, \vartheta ,z = H) &= \tilde{\phi}^{\nu}_i(r = R_{\text{in}}, \vartheta ,z) = \tilde{\phi}^{\nu}_i(r = R_{\text{out}}, \vartheta, z) = 0\,.
\end{align}
Similarly, the Poisson equation in Eq.~\ref{Cyclic_DFT_3} can be written as:
\begin{align}
\nonumber
-\frac{1}{4 \pi} \bigg( \hpd{\widetilde{V}_{\text{es}}(r,\vartheta,z)}{r}{2} &+ \frac{1}{r} \pd{\widetilde{V}_{\text{es}}(r,\vartheta,z)}{r} + \frac{1}{r^2}\hpd{\widetilde{V}_{\text{es}}(r,\vartheta,z)}{\vartheta}{2}+ \hpd{\widetilde{V}_{\text{es}}(r,\vartheta,z)}{z}{2} \bigg) \\
&=\tilde{\rho}(r,\vartheta,z)+\tilde{b}(r,\vartheta,z) \,,
\label{electrostatics_cylindrical_fd}
\end{align}
with boundary conditions (Eqs.~\ref{Eq:CyclicDFT:Ves:cyclic} and \ref{Eqn:SymmetryReducedBC_Ves}):
\begin{align}
\label{Electrostatic_BC_in_cylindrical_coordinates_1}
\widetilde{V}_{\text{es}}(r,\vartheta = 0 ,z) &= \widetilde{V}_{\text{es}}(r,\vartheta = \frac{2\pi}{\mathfrak{N}},z)\,,\\
\widetilde{V}_{\text{es}}(r = R_{\text{in}},\vartheta ,z) = \mathfrak{v}(r = R_{\text{in}},\vartheta ,z)\,&, \, \widetilde{V}_{\text{es}}(r = R_{\text{out}},\vartheta ,z) = \mathfrak{v}(r = R_{\text{out}},\vartheta ,z)\,,\\\label{Electrostatic_BC_in_cylindrical_coordinates_2}
\widetilde{V}_{\text{es}}(r ,\vartheta ,z = 0) = \mathfrak{v}(r ,\vartheta ,z = 0)\,&, \,\widetilde{V}_{\text{es}}(r,\vartheta ,z = H) = \mathfrak{v}(r ,\vartheta ,z=H)\,.
\end{align}
Above, $\mathfrak{v}(r,\vartheta,z)$ denotes the quantity obtained by evaluation of Eq.~\ref{Eqn:SymmetryReducedBC_Ves} at the point with cylindrical coordinates $(r,\vartheta,z)$.
We discretize the equations of Cyclic DFT using the finite-difference method with a uniform grid spacing of $\Delta r$, $\Delta \vartheta$ and $\Delta z$ in the radial, angular and ${\bf e}_3$ directions, respectively. Specifically, we employ a finite-difference mesh that consists of the points
\begin{align}
{\cal M} = {\cal M}_{R} \times {\cal M}_{\vartheta} \times {\cal M}_{z}\,,
\end{align}
where ${\cal M}_{R}$, ${\cal M}_{\vartheta}$, and ${\cal M}_{z}$ represent the nodes in the radial, angular and ${\bf e}_3$ directions, respectively, i.e.,
\begin{align}
{\cal M}_{R} &= \{R_{\text{in}},R_{\text{in}}+\Delta r,\ldots, R_{\text{in}} + (N_{r}-1)\Delta r = R_{\text{out}}\}\,,\\
{\cal M}_{\vartheta} &= \{0,\Delta \vartheta,\ldots, (N_{\vartheta}-1)\Delta \theta = \frac{2\pi}{\mathfrak{N}}-\Delta \theta\}\,,\\
{\cal M}_{z} &= \{0,\Delta z, (N_{z}-1)\Delta z = H\}\,,
\end{align}
and $N_{r}$, $N_{\vartheta}$ and $N_{z}$ denotes the number of grid points in the corresponding directions. We index the nodes by the triplet of natural numbers $(k_r,k_{\vartheta},k_z)$, with $k_r \in \{1,2,\ldots, N_r\}$, $k_{\vartheta} \in \{1,2,\ldots, N_{\vartheta} \}$ and $k_{z} \in \{1,2,\ldots,N_{z}\}$, so that the grid point with the indices $(k_r,k_{\vartheta},k_z)$ refers to the physical point $(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) \in {\cal M}$. While discretizing the governing equations, we represent the values of the various fields involved by their values at the finite-difference nodes. This casts the infinite dimensional problem to one posed on a linear space of dimension $N_rN_{\vartheta}N_{z}$.
We express the various derivatives in Eqs.~\ref{cylindrical_linear_eigen} and \ref{electrostatics_cylindrical_fd} using high-order finite-differences:
\begin{align}
\label{higher_order_fd_1}
\pd{f}{s}\bigg\vert_{s=s_{k_s}} &\approx \sum_{q=1}^{n_{o}} w_{q,1} \big(f{(s_{k_s+q})} - f{(s_{k_s-q})}\big)\, := D_sf(s_{k_s}),\\
\label{higher_order_fd_2}
\hpd{f}{s}{2}\bigg\vert_{s=s_{k_s}} &\approx \sum_{q=0}^{n_{o}} w_{q,2} \big(f{(s_{k_s+q})} + f{(s_{k_s-q})}\big)\, := D^{2}_sf(s_{k_s})\,,
\end{align}
where the $s$-coordinate represents any one of the $r$, $\vartheta$ or $z$-coordinates. The weights appearing in the above equations can be written as \citep{suryanarayana2014augmented, mazziotti1999spectral}:
\begin{align}
\label{fd_weights_1}
{w}_{q,1} &= \frac{(-1)^{q+1}}{ (\Delta s) q} \frac{(n_{o}!)^2}{(n_{o}-q)! (n_{o}+q)!} \,, \,\, q=1, 2, \ldots, n_{o}\,,\\
\label{fd_weights_2}
w_{0,2} &= - \frac{1}{(\Delta s)^2} \sum_{q=1}^{n_o} \frac{1}{q^2} \,, \\\label{fd_weights_3}
w_{q,2} &= \frac{2 (-1)^{q+1}}{(\Delta s)^2 q^2} \frac{(n_{o}!)^2}{(n_{o}-q)! (n_{o}+q)!} \,, \,\, q=1, 2, \ldots, n_{o}\,,
\end{align}
where $\Delta s$ denotes the mesh spacing along the $s$-coordinate. With these weights, the finite-difference expressions in Eqs.~\ref{higher_order_fd_1} and \ref{higher_order_fd_2} represent $2 n_{o}$-order accurate approximations, i.e., the error is $\mathcal{O}({\Delta s}^{2n_{o}})$.
On approximating the derivatives using finite-differences, the eigenvalue equations in Eq.~\ref{cylindrical_linear_eigen} take the form:
\begin{align}
\nonumber
-\frac{1}{2}\bigg(&D^{2}_r\,\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})+\frac{1}{r_{k_r}}D_{r}\,\tilde{\phi}^{\nu}_i(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})+ \frac{1}{r_{k_r}^2}D^{2}_{\vartheta}\,\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})\\
&+ D^{2}_{z}\,\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})\bigg )+\widetilde{V}_{\text{eff}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})\,\tilde{\phi}^{\nu}_i(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) = \lambda_{i}^{\nu}\,\tilde{\phi}^{\nu}_i(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})\,.
\label{discretized_KS_Hamiltonian}
\end{align}
During the application of the finite-difference stencils, any reference to grid points which do not lie in ${\cal M}$ is resolved by using the discrete representation of the boundary conditions in Eqs.~\ref{Bloch_BC_in_cylindrical_coordinates} and \ref{Dirichlet_BC_in_cylindrical_coordinates}:
\begin{align}
\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &= \expcharconj{\nu \gamma}{\mathfrak{N}}\,\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}+N_{\vartheta}},z_{k_z}) \,\,\,\, \text{if} \,\,\,\, k_{\vartheta} < 1 \,, \\
\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &= \expchar{\nu \gamma}{\mathfrak{N}}\,\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}-N_{\vartheta}},z_{k_z}) \,\,\,\, \text{if} \,\,\,\, k_{\vartheta} > N_{\vartheta} \,, \\
\tilde{\phi}^{\nu}_{i}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &= 0 \,\,\,\, \text{if} \,\,\,\, k_r \notin \{1,2,\ldots, N_r\} \,\, \text{or} \,\, k_z \notin \{1,2,\ldots, N_z\}\,.
\end{align}
Analogously, the Poisson equation in Eq.~\ref{electrostatics_cylindrical_fd} can be written in discrete form as:
\begin{align}
\nonumber
- \frac{1}{4 \pi} \bigg( D^{2}_r\,\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &+ \frac{1}{r_{k_r}} D_r\,\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) + \frac{1}{r_{k_r}^2}D^{2}_{\vartheta}\,\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})\\ &+ D^{2}_z\,\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) \bigg) = \tilde{\rho}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})+\tilde{b}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) \,,
\end{align}
with the discrete representation of the boundary conditions in Eqs.~\ref{Electrostatic_BC_in_cylindrical_coordinates_1} and \ref{Electrostatic_BC_in_cylindrical_coordinates_2} being:\footnote{For $k_r < 1$ or $k_r > N_r$, the point $r_{k_r}$ is used to denote the extension of the mesh in the radial direction, beyond the points $R_{\text{in}}$ and $R_{\text{out}}$, respectively. Thus, $k_r = 0$ denotes points with radial coordinate $R_{\text{in}}-\Delta r$, while $k_r = N_r + 1$ denotes points with radial coordinate $R_{\text{out}}+\Delta r$. Similar considerations hold in the $z$-direction.}
\begin{align}
\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &= \widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}+N_{\vartheta}},z_{k_z}) \,\,\,\, \text{if} \,\,\,\, k_{\vartheta} < 1 \,, \\
\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &= \widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}-N_{\vartheta}},z_{k_z}) \,\,\,\, \text{if} \,\,\,\, k_{\vartheta} > N_{\vartheta} \,, \\
\widetilde{V}_{\text{es}}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) &= \mathfrak{v}(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z}) \,\,\,\, \text{if} \,\,\,\, k_r \notin \{1,2,\ldots, N_r\} \,\, \text{or} \,\, k_z \notin \{1,2,\ldots, N_z\} \,.
\end{align}
The values of $\mathfrak{v}(r,\vartheta,z)$ are computed by means of direct numerical quadrature or using a multipole expansion in cylindrical coordinates \citep{cohl1999compact, wiki_cylindrical_multipole}.
The evaluation of the electronic ground-state free energy via Eq.~\ref{Harris_Foulkes_Energy_per_cell} and the atomic forces via Eq.~\ref{force_formula_FD} requires a recipe for evaluating integrals over the fundamental domain. To this end, we employ the quadrature rule:
\begin{align}
\nonumber
\int_{{\cal D}_{{\cal G}}}\!f({\bf x})\,\mathrm{d{\bf x}} &= \int_{r=R_{\text{in}}}^{R_{\text{out}}}\!\int_{\vartheta=0}^{\frac{2\pi}{\mathfrak{N}}}\!\int_{z=0}^{H}f(r,\vartheta,z)\,r\,\mathrm{dr\,d\vartheta\,dz} \\ &\approx \sum_{k_r=1}^{N_r}\sum_{k_{\vartheta}=1}^{N_{\vartheta}}\sum_{k_z=1}^{N_z}f(r_{k_r},\vartheta_{k_{\vartheta}},z_{k_z})\,r_{k_r} \Delta r\, \Delta \vartheta \Delta z\,.
\label{integration_discrete}
\end{align}
The numerical evaluation of the gradient operator for computation of the atomic forces merits further consideration. Since the derivation of Eq.~\ref{force_formula_1} (and consequently, Eq.~\ref{force_formula_FD}) implicitly assumes that the gradient operator is expressed in Cartesian coordinates (for e.g., see \citep{Phanish_SPARC_1}), we need to express the Cartesian gradient in cylindrical coordinates so that Eq.~\ref{force_formula_FD} can be evaluated by means of the finite-difference operators $D_r$, $D_{\vartheta}$ and $D_{z}$. For example, the computation of the ${\bf e}_1$ component of the force requires application of the operator $\pd{}{x}$, or equivalently, $\cos(\vartheta) \pd{}{r} - \sin(\vartheta) \pd{}{\vartheta} \approx \cos(\vartheta) D_r - \sin(\vartheta) D_{\vartheta}$. Such an approach \citep{my_analog_of_planewaves} results in the atomic forces being evaluated in Cartesian coordinates directly, which is more convenient for atomic relaxation and molecular dynamics calculations.
We note that the Laplacian and Hamiltonian matrices resulting from the above discretization are non-Hermitian, even though the infinite-dimensional operators from which they arise are Hermitian (i.e., self-adjoint). Having said this, each matrix does approach a Hermitian matrix as the discretization is refined and/or the finite difference order $n_{o}$ is increased. This issue is well known in the literature (for example, in the context of adaptive coordinates \citep{gygi1995real, Octopus_1}) and has been shown to not interfere with the quality of the solution obtained in practical electronic structure calculations. In particular, the eigenvalues of the Hamiltonian matrix (Eq.~\ref{discretized_KS_Hamiltonian}), which are required for computation of the band energy (Eq.~\ref{Harris_Foulkes_Energy_per_cell}), turn out to be real valued (or have vanishingly small imaginary parts that can be ignored) as required.
An alternative discretization scheme to the one proposed above is the spectral method introduced in \citep{My_PhD_Thesis, banerjee2015spectral} for cluster systems. This scheme --- identical to the plane-wave method in many respects --- is capable of leveraging arbitrary point group symmetries. It is therefore capable of solving the equations of Cyclic DFT on the fundamental domain \citep{My_PhD_Thesis}. However, the basis functions are global in nature (like plane-waves) and therefore the approach is not well suited for simulating large bent structures in which the atoms may be located far from the origin.\footnote{This issue can be handled by making suitable modifications in the radial basis functions.} In this respect, the localized nature of the proposed finite-difference approach (each node interacts with only a small set of neighboring points) enables the degrees of freedom (i.e., grid points) to be judiciously expended close to regions of interest while avoiding empty regions (i.e., grid points can be placed in regions close to a bent structure), thus improving the overall efficiency of the discretization. With regards to accuracy, it has been shown that the use of high-order finite-difference stencils within the pseudocharge formulation allows the systematic and accurate computation of total energies and forces with minimal interference from numerical issues such as the egg-box effect \citep{Phanish_SPARC_1,ghosh2016sparc2}.
It is worth mentioning that an alternative real-space discretization to the one adopted here is that based on finite-elements. In particular, the finite-element method allows various non-conventional domains, boundary conditions, and geometries to be easily handled. Moreover, high-order finite-elements have been shown to be an efficient choice for performing DFT calculations \citep{Pask_FEM_review_1,Gavini_higher_order, motamarri2014subquadratic}. However, the relatively simple and easy-to-implement high-order finite-difference discretization employed in this work is accurate and highly computationally efficient\footnote{This can be mainly attributed to the particularly compact representation of the Laplacian compared to other real-space alternatives for achieving the accuracies desired in electronic structure calculations.}, which enables reliable mechanistic simulations (Section \ref{subsec:silicene_results}) to be performed. Indeed, as suggested by the anonymous reviewer, investigating the efficacy of alternate real-space discretization strategies (such as finite-elements and wavelets \citep{BigDFT}) in the present context is a worthy subject of future work.
We mention in passing that even though the finite-difference discretization has been considered by several workers for performing electronic structure calculations \citep{Chelikowsky_Saad_1, Octopus_1, Phanish_SPARC_1}, to the best our knowledge, the present work is the first to do so within the cylindrical coordinate system. This choice of coordinate system presents its own challenges for the finite-difference method. For example, even in the Cartersian coordinate system where the mesh is uniformly spaced, calculation of accurate atomic forces is challenging \cite{ono1999timesaving,bobbitt2015high}. This and a number of such issues have been overcome in the formulation and implementation of Cyclic DFT so as to enable the accurate and efficient evaluation of energies and forces, as verified in Section \ref{Subsec:Verification}.
\section{Results and discussion} \label{Sec:Results}
\subsection{Implementation of Cyclic DFT} \label{subsec:implementation_details}
We implement the strategies and algorithms developed in the previous section using the MATLAB software package. In all the simulations, we utilize the pseudopotentials introduced in \citep{Carter_bulk_pseudo_1, Carter_bulk_pseudo_2}, a smearing of $\sigma = 0.0862$ eV (i.e., an electronic temperature of $1000$ K), and the local density approximation (LDA) \citep{KohnSham_DFT} with the Perdew-Wang parametrization \citep{perdew1992accurate} for the correlation functional. In addition, we choose $R_{\text{in}}$, $R_{\text{out}}$, and $H$ such that all atoms are at least $10-12$ Bohr away from the boundary $\partial{\cal D}_{{\cal G}}^{0}$ (i.e., the boundary on which Dirichlet boundary conditions are applied), so as to allow sufficient decay of the orbitals and the electron density.\footnote{In \citep{Phanish_SPARC_1}, it has been demonstrated that a vacuum of $10$ Bohr is sufficient to achieve accuracy of $10^{-5}$ Ha/atom in the energy and $10^{-5}$ Ha/Bohr in the forces for even polar systems like CO and H$_2$O.} Also, we employ $12^{th}$ order accurate finite-differences (i.e., $n_o = 6$) for all our calculations. The local nature of the finite-difference scheme results in the Laplacian and Hamiltonian matrices being sparse, and these are stored as such. In what follows, rather than use the three discretization parameters $\Delta r$, $\Delta \vartheta$ and $\Delta z$ (or equivalently $N_{r}$, $N_{\vartheta}$ and $N_{z}$), we utilize a single parameter to characterize the discretization: $h = \displaystyle \max \left\{ \Delta r, \frac{(R_{\text{in}}+R_{\text{out}})}{2}\Delta\vartheta, \Delta z \right\}$.
We generate the guess electron density at the start of the SCF iteration by summing the electron densities corresponding to the isolated-atom Kohn-Sham solutions. Random numbers are used to generate the initial guess for the orbitals, which are then orthonormalized for the purposes of stability. We use the Chebyshev polynomial filtered subspace iteration (CheFSI) method \citep{Serial_Chebyshev, Parallel_Chebyshev, zhou2014chebyshev} (with filter orders of $80-120$) for computing approximations to the Kohn-Sham orbitals in each SCF iteration.\footnote{Due to the nature of the cylindrical coordinate system, all the finite-difference mesh points are not uniformly spaced. Consequently, the discretized operators in this case are not as well conditioned as the operators arising from uniform grids in the Cartesian coordinate system. This makes it necessary to employ relatively high Chebyshev polynomial filter orders here.} As an alternative, we retain the option of using the Generalized Preconditioned Locally Harmonic Residual (GPLHR)\footnote{This method can be thought of as an analog of the LOBPCG algorithm \cite{LOBPCG_1} for the case of non-Hermitian matrices. Due to the ability of this method to use preconditioners (based on incomplete LU factorization, for example) it becomes computationally advantageous to use this method whenever the Hamiltonian is poorly conditioned. This can happen, for example, when studying a severely bent nanostructure since $R_{\text{in}}$ is likely to be close to $0$ in this case. This will lead to excessive clustering of finite difference nodes near $R_{\text{in}}$, thus resulting in the poor conditioning. The use of CheFSI in this case would involve very high filter orders, resulting in the repeated computation of the product of the Hamiltonian matrix with a block of vectors which would make the calculation expensive. In contrast, the use of GPLHR, particularly in the early SCF iterations allows the overall computational cost to be kept manageable. We would like to thank Eugene Vecharynski (Lawrence Berkeley National Lab) for his help with the use of the GPLHR method in Cyclic DFT.} algorithm \citep{vecharynski2015generalized}. We solve the Poisson equation using the Generalized Minimal Residual (GMRES) method \citep{saad1986gmres} with an incomplete-LU factorization based preconditioner \citep{saad2003iterative}, while retaining the option of using the AAJ \citep{pratapa2016anderson} and rPJ \citep{pratapa2015restarted} linear solvers.\footnote{These alternative solvers tend to produce better performance while dealing with poorly-conditioned problems \citep{pratapa2016anderson,pratapa2015restarted,suryanarayana2016alternating}.} We use a relative convergence tolerance of $10^{-5}-10^{-6}$ on the effective potential for convergence of the SCF method, and accelerate it using the Periodic Pulay mixing scheme \citep{banerjee2016periodic}.
We perform all computations using a single node of the Mesabi cluster of the Minnesota Supercomputing Institute. Each node of Mesabi has 24 Intel Haswell E5-2680v3 processors operating at $2.50$ GHz and sharing $64$-GB of RAM.
\subsection{Verification studies for the accuracy and efficiency of Cyclic DFT} \label{Subsec:Verification}
We now verify the accuracy and efficiency of Cyclic DFT. As the representative example, we choose a cyclic aluminum nanostructure with $M_{{\cal P}}=3$ atoms in the fundamental domain and symmetry group order of $\mathfrak{N}=12$, as shown in Fig.~\ref{fig:aluminum_ring}. The atoms are positioned randomly within the fundamental domain such that the radial coordinates are between $11-13$ Bohr and $z$-coordinates are between $10-13$ Bohr.
First, we confirm the convergence of the computed energy and atomic forces to plane-wave results determined using ABINIT \citep{Gonze_ABINIT_1} --- a well-established and optimized code for performing DFT calculations. In ABINIT, we use a plane-wave energy cutoff of $20$ Ha and a cubic supercell with edge of $40$ Bohr, which results in energy and forces that are converged to within $2.5 \times 10^{-7}$ Ha/atom and $10^{-5}$ Ha/Bohr, respectively. In Cyclic DFT, we choose the dimensions of the annular cylinder ${\cal C}$ to be $H=23$ Bohr, $R_{\text{in}} = 1$ Bohr, and $R_{\text{out}}=23$ Bohr. We present in Table \ref{Table:Convergence} the error in computed energy and atomic forces as the spatial discretization is refined. It is clear that there is systematic convergence in energy and forces to the reference plane-wave result, with accuracies of even $10^{-4}$ Ha/atom and $10^{-4}$ Ha/Bohr being readily achieved. This verifies the accuracy of Cyclic DFT in determining the electronic ground-state energy and atomic forces of cyclic structures.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\linewidth]{./Al_Ring_3.eps}
\caption{Cyclic aluminum nanostructure with $M_{{\cal P}}=3$ and group order $\mathfrak{N}=12$. The atoms within the fundamental domain are colored red.}
\label{fig:aluminum_ring}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|}
\hline
$h$ & Error in energy & Error in atomic force \\
(Bohr) & ($\mu$Ha/atom) & ($\mu$Ha/Bohr) \\
\hline
$1.00$ & $6980$ & $130537$ \\
\hline
$0.71$ & $740$ & $47000$ \\
\hline
$0.48$ & $27$ & $4300$ \\
\hline
$0.39$ & $12$ & $658$ \\
\hline
$0.30$ & $4$ & $93$ \\
\hline
\end{tabular}
\caption{Convergence of the energy and atomic forces computed by Cyclic DFT to the reference plane-wave result. The system under consideration is a cyclic aluminum nanostructure with $M_{{\cal P}}=3$ and group order $\mathfrak{N}=12$. The error in the energy denotes the absolute value of the difference and the error in the atomic force denotes the maximum magnitude of the difference among all the atoms. }
\label{Table:Convergence}
\end{table}
Next, we investigate the efficiency of the symmetry cell reduction in Cyclic DFT. To do so, for the aforementioned cyclic aluminum nanostructure, we consider unit cells where the sectors of the annular cylinder have angles that are $12$, $6$, $4$, $3$, $2$, and $1$ times the angle of the fundamental domain (i.e., $\pi/6$).\footnote{The $12$-fold cyclic symmetry of the system allows us to choose $1$, $2$, $3$, $4$, $6$, and $12$-fold cyclic symmetry subgroups.} This translates to exploiting the cyclic symmetry reduction from groups of order $\mathfrak{N}=1$, $2$, $3$, $4$, $6$, and $12$, respectively. For these cases, we present in Table \ref{Table:BreakdownTimeCyclic} the computational time taken by the various components of Cyclic DFT for $h=0.43$ Bohr. We observe that there is significant reduction in the computational time as the value of $\mathfrak{N}$ is increased.\footnote{During the SCF iterations using CheFSI, the most computationally intensive step tends to be computation of the action of the discretized Hamiltonian matrix on a block of vectors. It is easy to see that this step would stand to gain a speedup by a factor of $\mathfrak{N}$ when a cyclic group of order $\mathfrak{N}$ is used. This is because the Hamiltonian is represented on a smaller physical domain resulting in a fewer number of grid points (i.e., a factor of $\mathfrak{N}$) while the total number of electronic states involved (i.e., the number of states for every value of $\nu$, times the number of values of $\nu$) effectively remains the same. This trend is quite apparent in column $4$ of Table \ref{Table:BreakdownTimeCyclic}. However, for large enough systems, subspace diagonalization becomes the dominant cost within CheFSI, whereby the reduction in cost due to symmetry is expected to scale quadratically with respect to group order. This is also the case for other eigensolvers such as GPLHR, all of which asymptotically scale cubically with system size. Since subspace diagonalization can consume a significant portion of the computational time during large scale electronic structure computations \citep{My_DG_Cheby_paper}, we anticipate that the quadratic speed up obtained by the cyclic group reduction can result in a significant computational saving in large scale calculations. \label{Footnote:QuadraticReduction}} The worse than linear reduction of some of the components within one SCF iteration and the better than linear reduction of the atomic forces can be attributed to the use of inbuilt MATLAB optimizations, as well as the scope for further improvement in the implementation. We note that the energy and atomic forces obtained for the various group orders chosen above are identical to each other within $10^{-8}$ Ha/atom and $10^{-8}$ Ha/Bohr, respectively, further verifying the accuracy of Cyclic DFT.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$\mathfrak{N}$ & Poisson & Electron density & CheFSI & Energy & Total time & Atomic forces \\
\hline
$1$ & $26.4$ & $1.05$ & $353$ & $0.057$ & $371$ & $700$ \\
\hline
$2$ & $12.3$ & $0.51$ & $194$ & $0.013$ & $204$ & $316$ \\
\hline
$3$ & $8.7$ & $0.42$ & $118$ & $0.020$ & $129$ & $216$ \\
\hline
$4$ & $6.3$ & $0.25$ & $92$ & $0.007$ & $102$ & $125$ \\
\hline
$6$ & $4.0$ & $0.17$ & $57$ & $0.005$ & $66$ & $58$ \\
\hline
$12$ & $2.6$ & $0.08$ & $33$ & $0.006$ & $42$ & $29$ \\
\hline
\end{tabular}
\caption{Computational time in seconds for various components in Cyclic DFT as a function of the group order $\mathfrak{N}$ chosen for symmetry reduction. The system under consideration is a aluminum nanostructure with $12$-fold cyclic symmetry. The Poisson time is for one solution with an initial guess that is identically zero. The electron density and CheFSI times are for a representative iteration within the SCF method.}
\label{Table:BreakdownTimeCyclic}
\end{table}
As demonstrated in Table \ref{Table:SCF:Cyclic}, the symmetry cell reduction also reduces the number of iterations required to achieve convergence within the SCF method. A possible explanation for this phenomenon is as follows. In the CheFSI method applied to the full problem (or a problem with reduced group order), the eigenvectors (and eigenvalues) of the linearized Hamiltonian arising in every SCF iteration are calculated only approximately. Therefore, the eigenvectors(of the full problem, or a problem with reduced group order) do not exactly satisfy cyclic-Bloch boundary conditions on the fundamental domain. Consequently, the electron density and the effective potential do not precisely satisfy cyclic boundary conditions on the fundamental domain. This introduces low frequency error components into the SCF fixed-point iteration, which negatively impacts its convergence. Indeed, at the electronic ground-state, the eigenfunctions, electron density, and effective potential all satisfy the desired boundary conditions on the fundamental domain, as verified by the nearly identical results obtained for the various choices of group order. In contrast, directly enforcing the cyclic-Bloch boundary conditions on the fundamental domain using the appropriate group order bypasses the above described issue and results in better SCF convergence.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$\mathfrak{N}$ & $1$ & $2$ & $3$ & $4$ & $6$ & $12$ \\
\hline
SCF iterations & $127$ & $85$ & $44$ & $36$ & $30$ & $24$ \\
\hline
\end{tabular}
\caption{Number of SCF iterations as a function of the group order $\mathfrak{N}$. The system under consideration is a aluminum nanostructure with $12$-fold cyclic symmetry. }
\label{Table:SCF:Cyclic}
\end{table}
\subsection{Application of Cyclic DFT to the bending of a silicene nanoribbon}
\label{subsec:silicene_results}
Silicene\footnote{The silicon analogue of graphene.} is a two-dimensional allotrope of silicon consisting of atoms arranged in a buckled honeycomb lattice \cite{vogt2012silicene}. Silicene has fascinating electronic properties\footnote{Just like graphene, silicene has a linear dispersion relation and is semi-metallic.} that include high charge mobility \cite{shao2013first} and the quantum spin Hall effect \cite{liu2011quantum}. Though there have been some electronic structure studies to understand the effect of strains on the mechanical and electronic properties of silicene \cite{qin2012first,peng2013mechanical}, they have been restricted to uniform in-plane deformations.\footnote{This is because the commonly used plane-wave approaches are restricted to periodic boundary conditions, i.e., translational symmetry.} Using silicene as a representative example, we now show how Cyclic DFT can be used for the first principles study of uniform bending in nanostructures. To the best of our knowledge, this is the first self-consistent ab-initio simulation of such a nature.
The uniform bending simulation of a silicene nanoribbon in Cyclic DFT --- shown schematically in Fig.~\ref{Fig:SiliceneBending} --- proceeds as follows. First, we determine the equilibrium configuration of an infinite silicene sheet, using which we generate a nanoribbon of the desired width $W$. Next, depending on the desired bending curvature $\kappa$, we generate a cyclic structure with radius $\kappa^{-1}$ by `rolling' the silicene nanoribbon. For this cyclic system, we determine the fundamental domain, the atoms which belong to it, and the resulting group order. Finally, we compute the cyclic structure's electronic ground-state and energy using the formulation described in Section~\ref{Sec:Formulation}. We note that the use of a cyclic structure is expected to provide an accurate representation of pure bending,\footnote{For small enough curvatures.} since it replicates the desired bending curvature locally, and furthermore, the electronic interactions are short ranged (i.e., matter is near-sighted \cite{prodan2005nearsightedness}).
\begin{figure}[ht]
\begin{subfigure}{\linewidth}
\centering
\includegraphics[scale=0.35]{./Si_patch_Flat.eps}
\caption{Silicene nanoribbon}
\label{Fig:SiliceneSheet}
\end{subfigure}\\
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[scale=0.28]{./Silicene_Nu_15_top.eps}
\caption{Top view of the cyclic structure generated using the silicene nanoribbon.}
\label{Fig:SiliceneBent:Top}
\end{subfigure}%
$\quad$
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[scale=0.28]{./Silicene_Nu_15_side.eps}
\caption{Side view of the cyclic structure generated using the silicene nanoribbon.}
\label{Fig:SiliceneBent:Side}
\end{subfigure}
\caption{Schematic for the silicene nanoribbon bending simulation using Cyclic DFT. The radius of the cyclic structure corresponds to the bending radius of curvature $\kappa^{-1}$. The atoms within the fundamental domain are colored red.}
\label{Fig:SiliceneBending}
\end{figure}
We first use the above procedure to determine the electronic ground-state of a silicene nanoribbon of width $W=20.79$ Bohr with bending radius of curvature $\kappa^{-1}=17.49$ Bohr. The corresponding cyclic structure has $M_{{\cal P}}=8$ atoms in the fundamental domain and group order $\mathfrak{N}=15$. We present the results obtained by Cyclic DFT in Fig.~\ref{Fig:BendingSimulationResults}. Specifically, in Fig.~\ref{Fig:ContourPlot}, we present the contours of the electron density on the $z \approx 25$ Bohr plane. In Fig.~\ref{Fig:BandPlot}, we present the cyclic band structure plot\footnote{Analogue of the traditional band structure plot for systems with translational symmetry.} for this system, i.e., the eigenvalues $\lambda_i^{\nu}$ for $i=1, 2, \ldots, 20$ and $\nu=0, 1, \ldots \mathfrak{N}-1$. From this figure\footnote{Disregarding the $\nu = 0$ point, the symmetry of the curves about a vertical line passing through the point $\nu = 7.5$ can be explained in terms of the theoretical framework presented in \citep{My_Elliott_Symmetry_paper}.}, we can see that there is a negligible bandgap for the bent system. In addition, the system appears to be electronically stable. An analogous plot of the phonon eigenvalues can be used to study the structural stability of such systems, and can therefore be used to predict the onset and modes of instabilities\footnote{See \citep{aghaei2012symmetry} for examples of phonon-band structure diagrams in nano systems with helical symmetries, computed using empirical inter-atomic potentials.}. This is a worthy topic of research and is currently being pursued by the authors.
\begin{figure}[ht]
\begin{subfigure}{0.45\linewidth}
\centering
\includegraphics[scale=0.48]{./ContourPlot.eps}
\caption{Electron density contours on $z \approx 25$ Bohr plane.}
\label{Fig:ContourPlot}
\end{subfigure}%
$\quad\quad$
\begin{subfigure}{0.48\linewidth}
\centering
\includegraphics[scale=0.48]{./CyclicBandPlot.eps}
\caption{Cyclic band structure plot.}
\label{Fig:BandPlot}
\end{subfigure}
\caption{Results for bending of a silicene nanoribbon of width $W=20.79$ Bohr and radius of curvature $\kappa^{-1} = 17.49$ Bohr. The cyclic structure has $M_{{\cal P}}=8$ atoms in the fundamental domain and group order $\mathfrak{N}=15$.}
\label{Fig:BendingSimulationResults}
\end{figure}
Next, we use Cyclic DFT to study the variation of the silicene nanoribbon's bending energy $\mathcal{E}_b$ with its curvature $\kappa$. We define the strain energy due to bending at a given $\kappa$ to be the difference between the free energy of two configurations --- flat silicene nanoribbon\footnote{The free energy is computed using ABINIT.} (i.e. $\kappa^{-1} \rightarrow 0$) and the cyclic silicene nanostructure with radius of curvature $\kappa^{-1}$. For this study, we choose a silicene nanoribbon of width $W=20.79$ Bohr and bending radii of $\kappa^{-1}=14.05, 17.49, 27.87, 34.80, 52.14, 69.50, \, \text{and} \, 83.39$ Bohr. The corresponding cyclic structures have $M_{{\cal P}}=8$ atoms in the fundamental domain and cyclic group orders of $\mathfrak{N}=12, 15, 30, 45, 60, \, \text{and} \, 72$, respectively. In order to be able to accurately calculate the bending energy, we choose all the parameters within Cyclic DFT so as to achieve an overall accuracy of $10^{-5}$ Ha/atom in the energy. Anticipating Euler-Bernoulli type bending behavior, we plot the bending energy $\mathcal{E}_b$ so calculated as a function of $\kappa^{-2}$ in Fig.~\ref{fig:bendingEnergyCurvature}. We observe that $\mathcal{E}_b$ is proportional to $\kappa^{-2}$ --- consistent with Euler-Bernoulli bending\footnote{For a majority of the data points in Fig.~\ref{fig:bendingEnergyCurvature}, the straight line fit replicates the simulation data to accuracies of over $98 \%$.} --- with the predicted bending modulus being $32.6$ eV/atom $\text{\AA}^2$. It is worth noting that at the electronic level, we find that there is no noticeable bandgap opening in the silicene nanoribbon for the aforementioned bending curvatures.
On normalizing the calculated bending stiffness with the surface area \citep{nikiforov2014tight, jiang2013elastic}, we obtain a value of $D = 6.16$ eV, which lies between the values of $1.45$ eV and $9.61$ eV that have been obtained in the literature for graphene \citep{nikiforov2014tight, kudin2001c} and molybdenum disulphide \citep{jiang2013elastic}, respectively. To get an intuitive understanding of this observation, we recall that according to continuum theories of bending, there is usually a strong influence of the effective cross section on the bending stiffness of a material specimen. Due to the presence of out of plane atoms in a single layer relaxed silicene structure, it is likely that this material has an effective thickness\footnote{If one assumes that continuum shell theory is applicable in the present setting, we have the following relation for the effective thickness \citep{shenderova2002carbon}:
\begin{align}
t = \sqrt{\frac{12 D (1-\nu^2)}{Y}} \,,
\end{align}
where $Y$ is the in-plane Young's modulus, and $\nu$ is the Poisson's ratio. On substituting the values of $D$, $\nu$, and $Y$ for silicene into the above expression, we obtain $t \sim 0.4$ nm, which is in good agreement with the values that have been typically employed in literature \citep{pei2014effects,peng2013mechanical}. We note that the effective thickess is sometimes also referred to as the intrinsic finite thickness \citep{zhang2015elastic}.} that is intermediate when compared to the planar single atomic layer graphene structure and the three atomic layer molybdenum disulphide structure, thus leading to a similar trend in the bending stiffness in these three materials. On similar lines, we comment that due to the overall similarities in structure between the buckled silicene geometry studied here and the puckered phosphorene structures studied in the literature \citep{zhang2015elastic, yang2015temperature}, the bending stiffness of these two materials turns out to be quite similar (i.e., our value of $6.16$ eV for silicene vs. the values of $4.88 - 7.99$ eV obtained for phosphorene in the literature \citep{zhang2015elastic}).
We should mention in passing that our simulations do not involve actual relaxation of the atoms in the bent geometry. This is representative of practical scenarios where the movement of the atoms is constrained, e.g. when substrates are utilized to impart the desired strain state \citep{kerszberg2015ab,ding2010stretchable}. Additionally, relaxation effects appear to be unimportant in systems subjected to low curvatures \citep{Dumitrica_Bending_Graphene, nikiforov2014tight}. Edge effects and/or out of plane atomic displacements are likely to influence the mechanical behavior when atomic relaxations are allowed in systems subjected to large curvatures --- particularly so, when the arrangement of atoms within the fundamental domain is more complicated than the relatively simple nanoribbon geometry considered here. We hope to investigate these effects in future work.
Finally, it is worth pointing out that the silicene nanoribbon bending simulations as described above would be extremely challenging to perform --- if not practically impossible --- with existing conventional first principles techniques (especially when large values of the radius of curvature are involved), even if high performance computing resources are used. In contrast, our serial MATLAB implementation of Cyclic DFT enables these studies to be carried out conveniently within a few hours of simulation wall time, thus demonstrating the utility of the approach.
\begin{figure}[ht]
\centering
\includegraphics[width=0.6\textwidth]{./BendingEnergyCurvature.eps}
\caption{Bending energy as a function of the curvature for a silicene nanoribbon of width $W=20.79$ Bohr.}
\label{fig:bendingEnergyCurvature}
\end{figure}
\section{Conclusions and future directions} \label{Conclusions}
In this work, we have presented the theoretical foundations as well as the numerical formulation and implementation of Cyclic DFT --- a self-consistent first principles simulation method for nanostructures with cyclic symmetries. The Cyclic DFT methodology allows us to exploit the cyclic symmetry in a systematic and efficient manner. Additionally, it enables us to probe the behavior of various nanosystems under uniform bending deformations and therefore to directly obtain the associated materials properties from first principles. We have demonstrated these capabilities of Cyclic DFT through bending simulations of silicene nanoribbons.
With the foundational work on Cyclic DFT complete, we now briefly touch on future extensions and applications:
\begin{itemize}
\item \textbf{\small{Extending the range of applicability of Cyclic DFT:}} We are currently implementing accurate norm conserving pseudopotentials \citep{hamann1979norm, troullier1991efficient} and more elaborate exchange-correlation functionals \cite{perdew1996generalized,perdew1996rationale} into Cyclic DFT. We are also developing a well-optimized, large-scale parallel implementation in C/C++. Among other attractive parallelization features of Cyclic DFT, we are making use of the attribute that the $\mathfrak{N}$ eigenvalue problems associated with the different values of $\nu \in \Gamma_{\mathfrak{N}}$ can be solved independent of one another. This lends itself to an embarrassingly parallel implementation of Cyclic DFT. Together, these developments are going to allow us to study a wide variety of materials systems using Cyclic DFT and to reach large system sizes by making effective use of high performance computing resources.
\item \textbf{\small{Mechanistic simulation studies of low-dimensional nanomaterials under bending deformations:}} Single layers of low dimensional nanomaterials such as graphene, silicene, germanene, phosporene and hexagonal boron nitride (both ribbons and sheets), as well as their multi-layered counterparts have risen to scientific prominence in recent years due to their unique material properties \citep{xu2013graphene,butler2013progress}. Cyclic DFT provides a natural means of studying the effect of uniform bending deformations on such materials and understanding their mechanical behavior through ab-initio atomic relaxation and molecular dynamics simulations \citep{Hutter_abinitio_MD}.
An extension of Cyclic DFT to study non-uniform bending in low-dimensional nanomaterials can be effectively accomplished by appealing to coarse-gaining ideas developed previously for the study of crystal defects using DFT \citep{suryanarayana2013coarse, ponga2016sublinear}. The general idea is to treat non-uniform bending as a defect which breaks the cyclic symmetry of the system. Therefore, coarse-graining may be achieved by considering the non-uniformly bent structure to be under uniform bending locally. Such an approach has already been developed for atomistic systems \citep{hakobyan2012objective} using quasicontinuum type \citep{tadmor1996quasicontinuum} ideas. We anticipate that it may be also carried out at the level of electronic structure calculations.
\item \textbf{\small{Investigation of multi-physics coupling in nanosystems:}} Since Cyclic DFT is a true first principles simulation methodology, it allows for the possibility of investigating the effect of bending deformations on electronic, magnetic, transport and optical materials properties in nanosystems of interest. The study of electronic properties of two-dimensional nanomaterials systems under deformations has received much scientific attention in recent years \cite{naumov2011gap,kerszberg2015ab,johari2012tuning}. In line with this, the investigation of (for example) whether an electronic band gap can be introduced in existing two dimensional materials or nanotubes by subjecting them to bending deformations is a topic worthy of further research. Importantly, there is also the tantalizing possibility that some novel nanomaterial might develop a significant magnetization or polarization in the cyclic unit cell when subjected to a bending deformation. Such materials, if discovered, are likely to have a profound impact on the design of future sensors or energy conversion technologies. We anticipate that such studies can be conveniently accomplished using Cyclic DFT.
On these lines, we also find it worthwhile to mention the possible use of Cyclic DFT in studying the nanoscale flexoelectric effect \citep{deng2014flexoelectricity, ahmadpoor2015flexoelectricity, kalinin2008electronic, nguyen2013nanoscale, dumitricua2002curvature}. Conventional first principles calculations of flexoelectric coefficients \citep{hong2013first, ponomareva2012finite, hong2010flexoelectricity} have usually relied on Plane-wave DFT. The computation of the polarization in the periodic unit cell of an infinite crystal involves both theoretical and computational complications \citep{king1993theory, resta2007theory, spaldin2012beginner, resta1994macroscopic}. Additionally, since simulation of the flexoelectric effect, by definition, requires inhomogeneous strains to be imposed on the system under study, the setup of the simulation is further complicated \citep{xu2013direct, chandratre2012coaxing}. In contrast, estimation of the flexoelectric coefficient (for example, of a nanoribbon) using Cyclic DFT naturally addresses these two issues since the system being simulated is of finite extent (which is likely to help in bypassing the difficulties of computing polarization) and inhomogenous strains can be introduced via uniform bending of the nanoribbon.
\end{itemize}
\begin{center}
---
\end{center}
\section*{Acknowledgement}
Support for this work, in part, was provided through Scientific Discovery through Advanced Computing (SciDAC) program funded by U.S. Department of Energy, Office of Science, Advanced Scientific Computing Research and Basic Energy Sciences. PS acknowledges the support of the National Science Foundation under Grant Number 1553212. This work was partially carried out while ASB was at the University of Minnesota, Minneapolis. ASB acknowledges support from the following grants while at Minnesota: AFOSR FA9550-15-1-0207, NSF-PIRE OISE-0967140, ONR N00014-14-1-0714 and the MURI project FA9550-12-1-0458 (administered by AFOSR). The authors would like to acknowledge informative discussions with Richard James (Univ. of Minnesota), Ryan Elliott (Univ. of Minnesota), Kaushik Bhattacharya (Caltech), Lin Lin (Univ. of California, Berkeley) and Chao Yang (Lawrence Berkeley National Lab). The authors express their gratitude to the anonymous reviewers and the Editor for their comments and suggestions on the manuscript. The authors would also like to thank the Minnesota Supercomputing Institute for making the computing resources used in this work available.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\label{sec:intro}
\subsection{Composition theorems}
A basic structural question that can be asked in any model of computation is whether there can be resource savings when computing the same function on several independent inputs.
We say a direct sum theorem holds in a model of computation if solving a problem on $n$ independent inputs requires roughly $n$ times the resources needed to solve one instance of the problem.
Direct sum theorems hold for deterministic and randomized query complexity \cite{JKS10}, fail for circuit size \cite{Pan12}, and remain open for communication complexity \cite{KRW95,BBCR13,FKNN95}.
More generally, instead of merely outputting the $n$ answers, we could compute another function of these $n$ answers.
If $f$ is an $n$-bit Boolean function and $g$ is an $m$-bit Boolean function, we define the composed function $f\circ g$ to be an $nm$-bit Boolean function such that $f \circ g (x_1, \ldots, x_n) = f(g(x_1),\ldots, g(x_n))$, where each $x_i$ is an $m$-bit string.
The composition question now asks if there can be significant savings in computing $f \circ g$ compared to simply running the best algorithm for $f$ and using the best algorithm for $g$ to evaluate the input bits needed to compute $f$.
If we let $f$ be the identity function on $n$ bits that just outputs all its inputs, we recover the direct sum problem.
Composition theorems are harder to prove and are known for only a handful of models, such as deterministic~{\cite{Tal13,Mon14}} and quantum query complexity~{\cite{Rei11,LMR+11,Kim12}}.
More precisely, let $\D(f)$, $\R(f)$, and $\Q(f)$ denote the deterministic, randomized, and quantum query complexities of $f$. Then for all (possibly partial) Boolean%
\footnote{Composition theorems usually fail for trivial reasons for non-Boolean functions. Hence we restrict our attention to Boolean functions, which have domain $\{0,1\}^n$ (or a subset of $\{0,1\}^n$) and range $\{0,1\}$.}
functions $f$ and $g$, we have
\begin{equation}
\D(f\circ g) = \D(f)\D(g) \qquad \mathrm{and}
\qquad \Q(f \circ g) = \Theta(\Q(f)\Q(g)).
\end{equation}
In contrast, in the randomized setting we only have the upper bound direction, $\R(f \circ g) = O(\R(f)\R(g)\log \R(f))$. Proving a composition theorem for randomized query complexity remains a major open problem.
\begin{open}
Does it hold that $\R(f \circ g) = \Omega(\R(f)\R(g))$ for all Boolean functions $f$ and $g$?
\end{open}
In this paper we prove something close to a composition theorem for randomized query complexity.
While we cannot rule out the possibility of synergistic savings in computing $f \circ g$, we show that a composition theorem does hold if we insert a small gadget in between $f$ and $g$ to obfuscate the output of $g$.
Our gadget is ``small'' in the sense that its randomized (and even deterministic) query complexity is $\Theta(\log \R(g))$.
Specifically we choose the index function, which on an input of size $k+2^k$ interprets the first $k$ bits as an address into the next $2^k$ bits and outputs the bit stored at that address.
The index function's query complexity is $k+1$ and we choose $k=\Theta(\log \R(g))$.
\begin{theorem}
\label{thm:comp}
Let $f$ and $g$ be (partial) Boolean functions and let $\textsc{Ind}$ be the index function with $\R(\textsc{Ind}) = \Theta(\log \R(g))$. Then $\R(f\circ \textsc{Ind} \circ g) = \Omega(\R(f)\R(\textsc{Ind})\R(g)) = \Omega(\R(f)\R(g)\log \R(g))$.
\end{theorem}
\thm{comp} can be used instead of a true composition theorem in many applications.
For example, recently a composition theorem for randomized query complexity was needed in the special case when $g$ is the $\textsc{And}$ function \cite{ABK15}.
Our composition theorem would suffice for this application, since the separation shown there
only changes by a logarithmic factor
if an index gadget is inserted between $f$ and $g$.
We prove \thm{comp} by introducing a new lower bound technique for randomized query complexity. This is not surprising since the composition theorems for deterministic and quantum query complexities are also proved using powerful lower bound techniques for these models, namely the adversary argument and the negative-weights adversary bound \cite{HLS07} respectively.
\subsection{Sabotage complexity}
\label{sec:intro_sab}
To describe the new lower bound technique, consider the problem of computing a Boolean function $f$ on an input $x\in\{0,1\}^n$ in the query model.
In this model we have access to an oracle, which when queried with an index $i\in [n]$ responds with $x_i\in \{0,1\}$.
Imagine that a hypothetical saboteur damages the oracle and makes some of the input bits unreadable. For these input bits the oracle simply responds with a $*$.
We can now view the oracle as storing a string $p\in\{0,1,*\}^n$ as opposed to a string $x\in\{0,1\}^n$.
Although it is not possible to determine the true input $x$ from the oracle string $p$, it may still be possible to compute $f(x)$ if all input strings consistent with $p$ evaluate to the same $f$ value.
On the other hand, it is not possible to compute $f(x)$ if $p$ is consistent with a $0$-input and a $1$-input to $f$. We call such a string $p\in\{0,1,*\}^n$ a \emph{sabotaged input}.
For example, let $f$ be the $\textsc{Or}$ function that computes the logical $\textsc{Or}$ of its bits.
Then $p=00\!*\!0$ is a sabotaged input since it is consistent with the $0$-input $0000$ and the $1$-input $0010$.
However, $p=01\!*\!0$ is not a sabotaged input since it is only consistent with $1$-inputs to $f$.
Now consider a new problem in which the input is promised to be sabotaged (with respect to a function $f$) and our job is to find the location of a $*$.
Intuitively, any algorithm that solves the original problem $f$ when run on a sabotaged input must discover at least one $*$, since otherwise it would answer the same on $0$- and $1$-inputs consistent with the sabotaged input.
Thus the problem of finding a $*$ in a sabotaged input is no harder than the problem of computing $f$, and hence naturally yields a lower bound on the complexity of computing $f$. As we show later, this intuition can be formalized in several models of computation.
As it stands the problem of finding a $*$ in a sabotaged input has multiple valid outputs, as the location of any star in the input is a valid output. For convenience we define a decision version of the problem as follows: Imagine there are two saboteurs and one of them has sabotaged our input.
The first saboteur, Asterix, replaces input bits with an asterisk ($*$) and the second, Obelix, uses an obelisk ($\dagger$).
Promised that the input has been sabotaged exclusively by one of Asterix or Obelix, our job is to identify the saboteur.
This is now a decision problem since there are only two valid outputs.
We call this decision problem $f_\mathrm{sab}$, the \emph{sabotage problem} associated with $f$.
We now define lower bound measures for various models using $f_\mathrm{sab}$. For example, we can define the \emph{deterministic sabotage complexity of $f$} as $\DS(f)\coloneqq \D(f_\mathrm{sab})$ and in fact, it turns out that for all $f$, $\DS(f)$ equals $\D(f)$ (\thm{DS}).
We could define the \emph{randomized sabotage complexity of $f$} as $\R(f_\mathrm{sab})$, but instead we define it as $\RS(f)\coloneqq \R_0(f_\mathrm{sab})$, where $\R_0$ denotes zero-error randomized query complexity, since $\R(f_\mathrm{sab})$ and $\R_0(f_\mathrm{sab})$ are equal up to constant factors (\thm{RSR0S}).
$\RS(f)$ has the following desirable properties.
\begin{enumerate}[itemsep=0.3ex]
\item \makebox[10em][l]{(Lower bound for $\R$)} For all $f$, $\R(f) = \Omega(\RS(f))$ \hfill (\thm{RgeqRS})
\item \makebox[10em][l]{(Perfect composition)} For all $f$ and $g$, $\RS(f\circ g) \geq \RS(f) \RS(g)$ \hfill (\thm{RS_compose})
\item \makebox[10em][l]{(Composition with $\R$)} For all $f$ and $g$, ${\R}(f\circ g) = \Omega(\R(f) \RS(g))$ \hfill (\thm{R_compose})
\item \makebox[10em][l]{(Superior to $\prt(f)$)} There exists a total $f$ with $\RS(f) \geq \prt(f)^{2-o(1)}$ \hfill (\thm{comparison})
\item \makebox[10em][l]{(Superior to $\Q(f)$)} There exists a total $f$ with $\RS(f) = \widetilde{\Omega}(\Q(f)^{2.5})$ \hfill (\thm{comparison})
\item \makebox[10em][l]{(Quadratically tight)} For all total $f$, $\R(f) = O(\RS(f)^2 \log \RS(f))$ \hfill (\thm{rootR0})
\end{enumerate}
Here $\prt(f)$ denotes the partition bound \cite{JK10,JLV14}, which subsumes most other lower bound techniques such as approximate polynomial degree, randomized certificate complexity, block sensitivity, etc. The only general lower bound technique not subsumed by $\prt(f)$ is quantum query complexity, $\Q(f)$, which can also be considerably smaller than $\RS(f)$ for some functions.
In fact, we are unaware of any total function $f$ for which $\RS(f) = o(\R(f))$, leaving open the intriguing possibility that this lower bound technique captures randomized query complexity for total functions.
\subsection{Lifting theorems}
Using randomized sabotage complexity we are also able to show a relationship between lifting theorems in communication complexity.
A lifting theorem relates the query complexity of a function $f$ with the communication complexity of a related function created from $f$.
Recently, G\"o\"os, Pitassi, and Watson \cite{GPW15} showed that there is a communication problem $G_\textsc{Ind}$, also known as the two-party index gadget, with communication complexity $\Theta(\log n)$ such that for any function $f$ on $n$ bits, $\D^{\mathrm{cc}}(f \circ G_\textsc{Ind}) = \Omega(\D(f)\log n)$, where $\D^\mathrm{cc}(F)$ denotes the deterministic communication complexity of a communication problem $F$.
Analogous lifting theorems are known for some complexity measures, but no such theorem is known for either zero-error randomized or bounded-error randomized query complexity.
Our second result shows that a lifting theorem for zero-error randomized query complexity implies one for bounded-error randomized query.
We use $\R_0^\mathrm{cc}(F)$ and $\R^\mathrm{cc}(F)$ to denote the zero-error and bounded-error communication complexities of $F$ respectively.
\begin{theorem}
\label{thm:lifting}
Let $G:\mathcal{X} \times \mathcal{Y} \to \{0,1\}$ be a communication problem with $\min\{|\mathcal{X}|,|\mathcal{Y}|\}= O(\log n)$.
If it holds that for all $n$-bit partial functions $f$,
\begin{equation}
\R_0^{\mathrm{cc}}(f \circ G) = \Omega(\R_0(f)/\polylog n),
\end{equation}
then for all $n$-bit partial functions $f$,
\begin{equation}
\R^{\mathrm{cc}}(f \circ G_\textsc{Ind}) = \Omega(\R(f)/\polylog n),
\end{equation}
where $G_\textsc{Ind}:\{0,1\}^b \times \{0,1\}^{2^b} \to \{0,1\}$ is the index gadget (\defn{commindex}) with $b = \Theta(\log n)$.
\end{theorem}
Proving a lifting theorem for bounded-error randomized query complexity remains an important open problem in communication complexity.
Such a theorem would allow the recent separations in communication complexity shown by Anshu et al.~\cite{ABB+16} to be proved
simply by establishing their query complexity analogues, which
was done in \cite{ABK15} and \cite{AKK15}.
Our result shows that it is sufficient to prove a lifting theorem for zero-error randomized protocols instead.
\subsection{Open problems}
The main open problem is to determine whether $\R(f) = \widetilde{\Theta}(\RS(f))$ for all total functions $f$.
This is known to be false for partial functions, however.
Any partial function where all inputs in $\Dom(f)$ are far apart in Hamming distance necessarily has low sabotage complexity.
For example, any sabotaged input to the collision problem\footnote{In the collision problem, we are given an input $x\in[m]^n$, and we have to decide if $x$ viewed as a function from $[n]$ to $[m]$ is 1-to-1 or 2-to-1 promised that one of these holds.} has at least half the bits sabotaged making $\RS(f) = O(1)$, but $\R(f) = \Omega(\sqrt{n})$.
It would also be interesting to extend the sabotage idea to other models of computation and see if it yields useful lower bound measures.
For example, we can define quantum sabotage complexity as $\QS(f) \coloneqq \Q(f_\mathrm{sab})$, but we were unable to show that it lower bounds $\Q(f)$.
\subsection{Paper organization}
In \sec{prelim}, we present some preliminaries and useful properties of randomized algorithms (whose proofs appear in \app{properties} for completeness). We then formally define sabotage complexity in \sec{sabotage} and prove some basic properties of sabotage complexity.
In \sec{sum_and_composition} we establish the composition properties of randomized sabotage complexity described above (\thm{RS_compose} and \thm{R_compose}).
Using these results, we establish the main result (\thm{comp}) in \sec{composition}.
We then prove the connection between lifting theorems (\thm{lifting}) in \sec{lifting}.
In \sec{comparison} we compare randomized sabotage complexity with other lower bound measures.
We end with a discussion of deterministic sabotage complexity in \sec{other}.
\section{Preliminaries}
\label{sec:prelim}
In this section we define some basic notions in query complexity that will be used throughout the paper.
Note that all the functions in this paper have Boolean input and output, except sabotaged functions whose input alphabet is $\{0,1,*,\dagger\}$.
For any positive integer $n$, we define $[n]\coloneqq\{1,2,\ldots,n\}$.
In the model of query complexity, we wish to compute an $n$-bit Boolean function $f$
on an input $x$ given query access to the bits of $x$.
The function $f$ may be total, i.e., $f:\{0,1\}^n\to\{0,1\}$, or partial,
which means it is defined only on a subset of $\{0,1\}^n$, which we denote by $\Dom(f)$.
The goal is to output $f(x)$ using as few queries to the bits of $x$ as possible.
The number of queries used by the best possible deterministic
algorithm (over worst-case choice of $x$) is denoted $\D(f)$.
A randomized algorithm is a probability distribution over deterministic algorithms.
The worst-case cost of a randomized algorithm is the worst-case (over all the deterministic algorithms in its support) number of queries made by the algorithm on any input $x$.
The expected cost of the algorithm is the expected number of queries
made by the algorithm (over the probability distribution) on an input $x$ maximized over all inputs $x$.
A randomized algorithm has error at most $\epsilon$ if it outputs
$f(x)$ on every $x$ with probability at least $1-\epsilon$.
We use $\R_\epsilon(f)$ to denote the worst-case cost of the best randomized algorithm that computes $f$ with error $\epsilon$. Similarly, we use $\bar{\R}_\epsilon(f)$ to denote the expected cost of the best randomized algorithm that computes $f$ with error $\epsilon$.
When $\epsilon$ is unspecified it is taken to be $\epsilon=1/3$.
Thus $\R(f)$ denotes the bounded-error randomized query complexity of $f$.
Finally, we also define zero-error expected randomized query complexity, $\bar{\R}_0(f)$,
which we also denote by $\R_0(f)$ to be consistent with the literature.
For precise definitions of these measures as well as the definition of quantum query complexity
$\Q(f)$, see the survey by Buhrman and de Wolf~\cite{BdW02}.
\subsection{Properties of randomized algorithms}
We will assume familiarity with the following basic properties of randomized algorithms.
For completeness, we prove these properties in \app{properties}.
First, we have Markov's inequality, which allows us to convert an algorithm with a guarantee on the expected number of queries into
an algorithm with a guarantee on the maximum number of queries with a constant factor loss in the query bound and a constant factor increase in the error. This can be used, for example, to convert zero-error randomized algorithms into bounded-error randomized algorithms.
\begin{restatable}[Markov's Inequality]{lemma}{markov}
\label{lem:markov}
Let $A$ be a randomized algorithm that makes $T$ queries in expectation (over its internal randomness). Then for any $\delta\in (0,1)$, the
algorithm $A$ terminates within $\lfloor T/\delta\rfloor$ queries with probability at least $1-\delta$.
\end{restatable}
The next property allows us to amplify the success probability of an $\epsilon$-error randomized algorithm.
\begin{restatable}[Amplification]{lemma}{amplification}
\label{lem:amplification}
If $f$ is a function with Boolean output and $A$ is a randomized
algorithm for $f$ with error $\epsilon<1/2$, repeating $A$ several times and taking the majority vote of the outcomes decreases the error. To reach error $\epsilon^\prime>0$, it suffices to repeat the algorithm
$\frac{2\ln(1/\epsilon^\prime)}{(1-2\epsilon)^2}$ times.
\end{restatable}
Recall that we defined $\bar{\R}_\epsilon(f)$ to be the minimum expected number of queries made by a randomized algorithm
that computes $f$ with error probability at most $\epsilon$.
Clearly, we have $\bar{\R}_\epsilon(f)\leq \R_\epsilon(f)$, since the expected number of queries made by an algorithm is at most the maximum number of queries made by the algorithm.
Using \lem{markov}, we can now relate them in the other direction.
\begin{restatable}{lemma}{Rexpprime}\label{lem:Rexp_prime}
Let $f$ be a partial function, $\delta>0$, and $\epsilon \in [0,1/2)$.
Then we have
$\R_{\epsilon+\delta}(f)\leq \frac{1-2\epsilon}{2\delta}\bar{\R}_{\epsilon}(f) \leq \frac{1}{2\delta}\bar{\R}_{\epsilon}(f)$.
\end{restatable}
The next lemma shows how to relate these measures with the same error $\epsilon$ on both sides of the inequality.
This also shows that $\bar{\R}_\epsilon(f)$ is only a constant factor away from $\R_\epsilon(f)$ for constant $\epsilon$.
\begin{restatable}{lemma}{Rexp}
\label{lem:Rexp}
If $f$ is a partial function,
then for all $\epsilon\in(0,\frac{1}{2})$, we have
$\R_\epsilon(f)\leq
14\frac{\ln(1/\epsilon)}
{(1-2\epsilon)^2}
\bar{\R}_\epsilon(f)$.
When $\epsilon=\frac{1}{3}$, we can improve this to
$\R(f)\leq 10\bar{\R}(f)$.
\end{restatable}
Although these measures are closely related for constant error, $\bar{\R}_\epsilon(f)$ is more convenient than $\R_\epsilon(f)$ for discussing composition and direct sum theorems.
We can also convert randomized algorithms that
find certificates with bounded error into zero-error
randomized algorithms.
\begin{restatable}{lemma}{repeat}
\label{lem:repeat}
Let $A$ be a randomized algorithm that uses $T$ queries in expectation and finds a certificate with probability $1-\epsilon$. Then repeating $A$ when it fails to find a certificate turns it into an algorithm that always finds a certificate (i.e., a zero-error algorithm) that uses at most $T/(1-\epsilon)$ queries in expectation.
\end{restatable}
Finally, the following lemma is useful for proving lower bounds on randomized algorithms.
\begin{restatable}{lemma}{block}
\label{lem:block}
Let $f$ be a partial function. Let $A$ be a randomized algorithm
that solves $f$ using at most $T$ expected queries and
with error at most $\epsilon$.
For $x,y\in\Dom(f)$ if $f(x)\neq f(y)$ then
when $A$ is run on $x$,
it must query an entry on which $x$ differs from $y$
with probability at least $1-2\epsilon$.
\end{restatable}
\section{Sabotage complexity}
\label{sec:sabotage}
We now formally define sabotage complexity. Given a (partial or total) $n$-bit Boolean function $f$,
let $P_f\subseteq\{0,1,*\}^n$
be the set of all partial assignments of $f$ that
are consistent with both a $0$-input and a $1$-input. That is,
for each $p\in P_f$, there exist $x,y\in\Dom(f)$
such that $f(x)\neq f(y)$ and $x_i=y_i=p_i$ whenever
$p_i\neq*$. Let $\smash{P^\dagger_f}\subseteq\{0,1,\dagger\}^n$
be the same as $P_f$,
except using the symbol $\dagger$ instead of $*$.
Observe that $P_f$ and $\smash{P_f^\dagger}$ are disjoint.
Let $Q_f=P_f \cup \smash{P^\dagger_f}\subseteq\{0,1,*,\dagger\}^n$.
We then define $f_\mathrm{sab}$ as follows.
\begin{definition}\label{defn:sabotage}
Let $f$ be an $n$-bit partial function. We define
$f_\mathrm{sab}:Q_f\to\{0,1\}$ as $f_\mathrm{sab}(q)=0$ if $q\in P_f$
and $f_\mathrm{sab}(q)=1$ if $q\in \smash{P^\dagger_f}$.
\end{definition}
Note that even when $f$ is a total function, $f_\mathrm{sab}$ is always a partial function.
See \sec{intro_sab} for more discussion and motivation
for this definition. Now that we have defined $f_\mathrm{sab}$,
we can define deterministic and randomized sabotage complexity.
\begin{definition}\label{def:RS}
Let $f$ be a partial function.
Then $\DS(f)\coloneqq \D(f_{\mathrm{sab}})$ and $\RS(f)\coloneqq\R_0(f_{\mathrm{sab}})$.
\end{definition}
We will primarily focus on $\RS(f)$ in this work and only discuss $\DS(f)$ in \sec{other}.
To justify defining $\RS(f)$ as $\R_0(f_{\mathrm{sab}})$
instead of $\R(f_{\mathrm{sab}})$ (or $\bar{\R}(f_\mathrm{sab})$), we now show these definitions
are equivalent up to constant factors.
\begin{theorem}\label{thm:RSR0S}
Let $f$ be a partial function.
Then $\R_0(f_\mathrm{sab}) \geq \bar{\R}_\epsilon(f_{\mathrm{sab}})\geq
(1-2\epsilon)\R_0(f_{\mathrm{sab}})$.
\end{theorem}
\begin{proof}
The first inequality follows trivially. For the second,
let $x\in Q_f$ be any valid input to $f_{\mathrm{sab}}$.
Let $x'$ be the input $x$ with asterisks replaced with obelisks and vice versa.
Then since $f_\mathrm{sab}(x)\neq f_\mathrm{sab}(x')$, by \lem{block} any $\epsilon$-error
randomized algorithm that solves $f_\mathrm{sab}$ must find a position on which
$x$ and $x'$ differ with probability at least $1-2\epsilon$.
The positions at which they differ are either asterisks or obelisks.
Since $x$ was an arbitrary input, the algorithm must always
find an asterisk or obelisk with probability at least $1-2\epsilon$.
Since finding an asterisk or obelisk is a certificate for $f_\mathrm{sab}$,
by \lem{repeat}, we get a zero-error algorithm for $f_\mathrm{sab}$ that
uses $\bar{\R}_\epsilon(f_{\mathrm{sab}})/(1-2\epsilon)$ expected queries. Thus
$\R_0(f_{\mathrm{sab}})\leq \bar{\R}_\epsilon(f_{\mathrm{sab}})/(1-2\epsilon)$,
as desired.
\end{proof}
Finally, we prove that $\RS(f)$ is indeed a lower bound on $\R(f)$, i.e., $\R(f) = \Omega(\RS(f))$.
\begin{theorem}\label{thm:RgeqRS}
Let $f$ be a partial function. Then
$\R_\epsilon(f)\geq\bar{\R}_\epsilon(f)\geq(1-2\epsilon)\RS(f)$.
\end{theorem}
\begin{proof}
Let $A$ be a randomized algorithm for $f$
that uses $\bar{\R}_\epsilon(f)$
randomized queries and outputs the correct answer on every input
in $\Dom(f)$ with probability at least $1-\epsilon$.
Now fix a sabotaged input $x$, and let $p$ be the probability
that $A$ finds a $*$ or $\dag$ when run on $x$.
Let $q$ be the probability that $A$ outputs $0$ if it
doesn't find a $*$ or $\dag$ when run on $x$.
Let $x_0$ and $x_1$ be inputs consistent with $x$ such that
$f(x_0)=0$ and $f(x_1)=1$.
Then $A$ outputs $0$ on $x_1$ with probability at least
$q(1-p)$, and $A$ outputs $1$ on $x_0$ with probability
at least $(1-q)(1-p)$. These are both errors, so we have
$q(1-p)\leq\epsilon$ and $(1-q)(1-p)\leq\epsilon$.
Summing them gives
$1-p\leq 2\epsilon$, or $p\geq 1-2\epsilon$.
This means $A$ finds a $*$ entry within
$\bar{\R}_\epsilon(f)$ expected queries with probability
at least $1-2\epsilon$. By \lem{repeat}, we get
$\frac{1}{1-2\epsilon}\bar{\R}_\epsilon(f)\geq \RS(f)$,
or $\bar{\R}_\epsilon(f)\geq(1-2\epsilon)\RS(f)$.
\end{proof}
We also define a variant of $\RS$ where the number
of asterisks (or obelisks) is limited to one. Specifically,
let $U\subseteq\{0,1,*,\dagger\}^n$ be the set of all
partial assignments with exactly one $*$ or $\dagger$.
Formally, $U \coloneqq \{x\in\{0,1,*,\dagger\}^n: |\{i\in[n]:x_i \notin \{0,1\}\}| = 1\}$.
\begin{definition}\label{def:RS1}
Let $f$ be an $n$-bit partial function. We define
$f_\mathrm{usab}$ as the restriction of $f_\mathrm{sab}$ to $U$,
the set of strings with only one asterisk or obelisk.
That is, $f_\mathrm{usab}$ has domain $Q_f\cap U$, but is equal
to $f_\mathrm{sab}$ on its domain. We then define
$\RS_{\mathrm{u}}(f)\coloneqq \R_0(f_\mathrm{usab})$. If $Q_f\cap U$ is empty,
we define $\RS_{\mathrm{u}}(f)\coloneqq 0$.
\end{definition}
The measure $\RS_{\mathrm{u}}$ will play a key role in our lifting result in \sec{lifting}.
Since $f_\mathrm{usab}$ is a restriction of $f_\mathrm{sab}$ to a promise,
it is clear that its zero-error randomized query complexity
cannot increase, and so $\RS_{\mathrm{u}}(f)\leq\RS(f)$.
Interestingly, when $f$ is total, $\RS_{\mathrm{u}}(f)$ equals $\RS(f)$.
In other words, when $f$ is total, we may assume without
loss of generality that its sabotaged version has only one
asterisk or obelisk.
\begin{theorem}\label{thm:RS1}
If $f$ is a total function, then $\RS_{\mathrm{u}}(f)=\RS(f)$.
\end{theorem}
\begin{proof}
We already argued that $\RS(f)\geq\RS_{\mathrm{u}}(f)$. To show
$\RS_{\mathrm{u}}(f)\geq\RS(f)$, we argue that any zero-error algorithm
$A$ for $f_\mathrm{usab}$ also solves $f_\mathrm{sab}$.
The main observation we need is that any input to $f_\mathrm{sab}$
can be completed to an input to $f_\mathrm{usab}$ by replacing
some asterisks or obelisks with $0$s and $1$s.
To see this, let $x$ be an input to $f_\mathrm{sab}$. Without
loss of generality, $x\in P_f$. Then there are two strings
$y,z\in\Dom(f)$ that are consistent with $x$,
satisfying $f(y)=0$ and $f(z)=1$.
The strings $y$ and $z$
disagree on some set of bits $B$, and $x$ has a $*$ or $\dag$
on all of $B$. Consider starting with $y$
and flipping the bits of $B$ one by one, until we reach
the string $z$. At the beginning, we have $f(y)=0$,
and at the end, we reach $f(z)=1$. This means that
at some point in the middle, we must have flipped a bit
that flipped the string from a $0$-input to a $1$-input.
Let $w_0$ and $w_1$ be the inputs where this happens.
They differ in only one bit. If we replace that bit with $*$ or $\dag$,
we get a partial assignment $w$ consistent with both,
so $w\in P_f$. Moreover, $w$ is consistent with $x$.
This means we have completed an arbitrary input to
$f_\mathrm{sab}$ to an input to $f_\mathrm{usab}$, as claimed.
The algorithm $A$, which correctly solves $f_\mathrm{usab}$, when run on $w$ (a valid input to $f_\mathrm{usab}$) must find an asterisk or obelisk in $w$.
Now consider running $A$ on the input $x$ to $f_\mathrm{sab}$ and compare its execution to when it is run on $w$. If $A$ ever queries a position that is different in $x$ and $w$, then it has found an asterisk or obelisk and the algorithm can now halt. If not, then it must find the single asterisk or obelisk present in $w$, which is also present in $x$.
This shows that the slightly modified version of $A$ that halts if it queries an asterisk or obelisk solves $f_\mathrm{sab}$ and hence
$\RS(f)=\R_0(f_\mathrm{sab})\leq\R_0(f_\mathrm{usab})=\RS_{\mathrm{u}}(f)$.
\end{proof}
\section{Direct sum and composition theorems}
\label{sec:sum_and_composition}
In this section, we establish the main composition theorems
for randomized sabotage complexity.
To do so, we first need to establish direct sum theorems
for the problem $f_{\mathrm{sab}}$. In fact,
our direct sum theorems hold more generally for
zero-error randomized query complexity of partial functions
(and even relations).
To prove this, we will require Yao's minimax theorem \cite{Yao77}.
\begin{theorem}[Minimax]\label{thm:yao}
Let $f$ be an $n$-bit partial function.
There is a distribution $\mu$ over inputs in $\Dom(f)$
such that all zero-error algorithms for $f$
use at least $\R_0(f)$ expected queries on $\mu$.
\end{theorem}
We call any distribution $\mu$ that satisfies the assertion in Yao's theorem a \emph{hard distribution} for $f$.
\subsection{Direct sum theorems}
\label{sec:sum}
We start by defining the $m$-fold direct sum of a function $f$, which is simply the function that accepts $m$ inputs to $f$ and outputs $f$ evaluated on all of them.
\begin{definition}\label{def:sum}
Let $f:\Dom(f)\to \mathcal{Z}$, where $\Dom(f)\subseteq \mathcal{X}^n$, be a partial function
with input and output alphabets $\mathcal{X}$ and $\mathcal{Z}$.
The $m$-fold direct sum of $f$ is the partial function $f^{\oplus m}:\Dom(f)^m\to\mathcal{Z}^m$
such that for any $(x_1,x_2,\ldots,x_m) \in \Dom(f)^m$, we have
\begin{equation}
f^{\oplus m}(x_1,x_2,\ldots,x_m)=(f(x_1),f(x_2),\ldots,f(x_m)).
\end{equation}
\end{definition}
We can now prove a direct sum theorem for zero-error randomized and more generally $\epsilon$-error expected randomized
query complexity, although we only require the result about zero-error algorithms.
We prove these results for partial functions, but they also hold for arbitrary relations.
\begin{theorem}[Direct sum]\label{thm:direct_sum}
For any $n$-bit partial function $f$ and any positive
integer $m$, we have
$\R_0(f^{\oplus m})=m\R_0(f)$. Moreover,
if $\mu$ is a hard distribution for $f$
given by \thm{yao}, then $\mu^{\otimes m}$ is
a hard distribution for $f^{\oplus m}$.
Similarly, for $\epsilon$-error randomized algorithms we get $\bar{\R}_\epsilon(f^{\oplus m}) \geq m\bar{\R}_\epsilon(f)$.
\end{theorem}
\begin{proof}
The upper bound follows from running the $\R_0(f)$
algorithm on each of the $m$ inputs to $f$. By linearity
of expectation, this algorithm solves all $m$ inputs
after $m\R_0(f)$ expected queries.
We now prove the lower bound. Let $A$ be a zero-error
randomized algorithm for $f^{\oplus m}$ that uses $T$ expected
queries when run on inputs from $\mu^{\otimes m}$.
We convert $A$ into an algorithm $B$ for $f$
that uses $T/m$ expected queries when run on inputs from $\mu$.
Given an input $x\sim\mu$, the algorithm $B$ generates
$m-1$ additional ``fake'' inputs from $\mu$.
$B$ then shuffles these together with $x$, and runs $A$
on the result. The input to $A$ is then distributed
according to $\mu^{\otimes m}$, so $A$ uses $T$ queries
(in expectation) to solve all $m$ inputs. $B$ then reads the
solution to the true input $x$.
Note that most of the queries $A$ makes are to fake inputs,
so they don't count as real queries. The only real queries $B$
has to make happen when $A$ queries $x$. But since $x$ is
shuffled with the other (indistinguishable) inputs,
the expected number of queries $A$ makes to $x$ is the
same as the expected number of queries $A$ makes to each
fake input; this must equal $T/m$. Thus $B$ makes $T/m$
queries to $x$ (in expectation) before solving it.
Since $B$ is a zero-error randomized algorithm for $f$
that uses $T/m$ expected queries on inputs from $\mu$,
we must have $T/m\geq\R_0(f)$ by \thm{yao}.
Thus $T\geq m\R_0(f)$, as desired.
The same lower bound proof carries through for $\epsilon$-error expected query complexity, $\bar{\R}_\epsilon(f)$, as long as we use a version of Yao's theorem for this model. For completeness, we prove this version of Yao's theorem in \app{yao}.
\end{proof}
\thm{direct_sum} is essentially \cite[Theorem 2]{JKS10}, but our theorem statement looks different since we deal with expected query complexity instead of worst-case query complexity. From \thm{direct_sum}, we can also prove a direct sum theorem for worst-case randomized query complexity since for $\epsilon\in(0,1/2)$,
\begin{equation}
\R_\epsilon(f^{\oplus m}) \geq \bar{\R}_\epsilon(f^{\oplus m}) \geq m \bar{\R}_\epsilon(f) \geq 2\delta m \R_{\epsilon+\delta}(f),
\end{equation}
for any $\delta>0$, where the last inequality used \lem{Rexp_prime}.
For our applications, however,
we will need a strengthened version of this theorem,
which we call a threshold direct sum theorem.
\begin{theorem}[Threshold direct sum]\label{thm:thresh_sum}
Given an input to $f^{\oplus m}$ sampled from $\mu^{\otimes m}$,
we consider solving only some of the $m$ inputs to $f$.
We say an input $x$ to $f$ is solved if a $z$-certificate was queried that proves
$f(x)=z$. Then any randomized algorithm
that takes an expected $T$ queries and solves an expected
$k$ of the $m$ inputs when run on inputs from $\mu^{\otimes m}$
must satisfy $T\geq k\R_0(f)$.
\end{theorem}
\begin{proof}
We prove this by a reduction to \thm{direct_sum}.
Let $A$ be a randomized algorithm that, when run on
an input from $\mu^{\otimes m}$, solves an expected $k$
of the $m$ instances, and halts after
an expected $T$ queries. We note that these expectations
average over both the distribution $\mu^{\otimes m}$
and the internal randomness of $A$.
We now define a randomized algorithm $B$ that solves the $m$-fold direct sum $f^{\oplus m}$
with zero error. $B$ works as
follows: given an input to $f^{\oplus m}$, $B$ first runs
$A$ on that input. Then $B$ checks which of the $m$ instances of $f$
were solved by $A$ (by seeing if a certificate proving
the value of $f$ was found for a given instance of $f$). $B$ then runs the
optimal zero-error algorithm for $f$, which makes $\R_0(f)$ expected queries,
on the instances of $f$ that were not solved by $A$.
Let us examine the expected number of queries used by $B$
on an input from $\mu^{\otimes m}$. Recall that a randomized
algorithm is a probability distribution over deterministic
algorithms; we can therefore think of $A$ as a distribution.
For a deterministic algorithm $D\sim A$ and an input
$x$ to $f^{\oplus m}$, we use $D(x)$ to denote the number
of queries used by $D$ on $x$, and $S(D,x)\subseteq [m]$
to denote the set of inputs to $f$ the algorithm $D$ solves
when run on $x$. Then by assumption
\begin{equation}
T=\mathop{\mathbb{E}}_{x\sim\mu^{\otimes m}}
\mathop{\mathbb{E}}_{D\sim A} D(x) \qquad \textrm{and} \qquad
k=\mathop{\mathbb{E}}_{x\sim\mu^{\otimes m}}
\mathop{\mathbb{E}}_{D\sim A} |S(D,x)|.
\end{equation}
Next, let $R$ be the randomized algorithm that uses $\R_0(f)$
expected queries and solves $f$ on any input.
For an input $x$ to $f^{\oplus m}$, we write
$x=x_1x_2\ldots x_m$ with $x_i\in\Dom(f)$. Then the expected number
of queries used by $B$ on input from $\mu^{\otimes m}$
can be written as
\begin{align}
\phantom{=}& \mathop{\mathbb{E}}_{x\sim\mu^{\otimes m}}
\mathop{\mathbb{E}}_{D\sim A}\left( D(x) +
\mathop{\mathbb{E}}_{D_1\sim R}
\mathop{\mathbb{E}}_{D_2\sim R}
\cdots
\mathop{\mathbb{E}}_{D_m\sim R}
\sum_{i\in [m]\setminus S(D,x)} D_i(x_i)\right)\\
=& \mathop{\mathbb{E}}_{x\sim\mu^{\otimes m}}
\mathop{\mathbb{E}}_{D\sim A}\left( D(x) +
\sum_{i\in[m]\setminus S(D,x)}
\mathop{\mathbb{E}}_{D_i\sim R} D_i(x_i)\right)\\
\leq & \mathop{\mathbb{E}}_{x\sim\mu^{\otimes m}}
\mathop{\mathbb{E}}_{D\sim A}\left( D(x) +
\sum_{i\in [m]\setminus S(D,x)} \R_0(f)\right)\\
=& \mathop{\mathbb{E}}_{x\sim\mu^{\otimes m}}
\mathop{\mathbb{E}}_{D\sim A}\left( D(x) +
(m-|S(D,x)|) \R_0(f)\right)\\
=& \quad T+(m-k)\R_0(f).
\end{align}
Since $B$ solves the direct sum problem on $\mu^{\otimes m}$, the expected number
of queries it uses is at least $m\R_0(f)$ by \thm{direct_sum}.
Hence $T+(m-k)\R_0(f)\geq m\R_0(f)$, so $T\geq k\R_0(f)$.
\end{proof}
\subsection{Composition theorems}
\label{sec:basic_composition}
Using the direct sum and threshold direct sum theorems we have established, we can now prove composition theorems for randomized sabotage complexity.
We start with the behavior of $\RS$ itself under composition.
\begin{theorem}\label{thm:RS_compose}
Let $f$ and $g$ be partial functions. Then
$\RS(f\circ g)\geq\RS(f)\RS(g)$.
\end{theorem}
\begin{proof}
Let $A$ be any zero-error algorithm for $(f\circ g)_\mathrm{sab}$, and let $T$
be the expected query complexity of $A$ (maximized over all inputs).
We turn $A$ into a zero-error algorithm
$B$ for $f_\mathrm{sab}$.
$B$ takes a sabotaged input $x$ for $f$.
It then runs $A$ on a sabotaged input to $f\circ g$
constructed as follows.
Each $0$ bit of $x$ is replaced
with a $0$-input to $g$, each $1$ bit of $x$ is replaced
with a $1$-input to $g$, and each $*$ or $\dag$ of $x$ is replaced
with a sabotaged input to $g$. The sabotaged inputs are
generated from $\mu$, the hard distribution
for $g_\mathrm{sab}$ obtained from \thm{yao}.
The $0$-inputs are generated by
first generating a sabotaged input, and then
selecting a $0$-input consistent with that sabotaged input.
The $1$-inputs are generated analogously.
This is implemented in the following way.
On input $x$, the algorithm $B$ generates $n$ sabotaged inputs
from $\mu$ (the hard distribution for $g_{\mathrm{sab}}$),
where $n$ is the length of the string $x$.
Call these inputs $y_1,y_2,\dots,y_n$.
$B$ then runs the algorithm $A$ on this collection
of $n$ strings, pretending that it is an input to
$f\circ g$, with the following caveat: whenever
$A$ tries to query a $*$ or $\dag$ in an input $y_i$, $B$ instead
queries $x_i$. If $x_i$ is $0$, $B$ selects an input from
$f^{-1}(0)$ consistent with $y_i$, and replaces $y_i$ with
this input. It then returns to $A$ an answer consistent
with the new $y_i$.
If $x_i$ is $1$, $B$ selects a consistent input
from $f^{-1}(1)$ instead. If $x_i$ is a $*$ or $\dag$, $B$ returns a $*$ or $\dag$ respectively.
Now $B$ only makes queries to $x$ when it finds a $*$ or $\dagger$ in an input to $g_\mathrm{sab}$.
But this solves that instance of $g_\mathrm{sab}$, which was drawn from the hard distribution for $g_\mathrm{sab}$.
Thus the query complexity of $B$ is upper bounded by the number of instances of $g_\mathrm{sab}$ that can be solved by a $T$-query algorithm with access to $n$ instances of $g_\mathrm{sab}$.
We know from \thm{thresh_sum} that if $A$ makes $T$ expected queries,
the expected number of $*$ or $\dag$ entries it finds among $y_1,y_2,\dots,y_n$ is at most $T/\RS(g)$.
Hence the expected number of queries $B$ makes to $x$ is at most $T/\RS(g)$.
Thus we have $\RS(f)\leq T/\RS(g)$, which gives
$T\geq \RS(f)\RS(g)$.
\end{proof}
Using this we can lower bound the randomized query complexity of composed functions.
In the following, $f^n$ denotes the function $f$ composed with itself $n$ times, i.e., $f^1 = f$ and $f^{i+1} = f \circ f^{i}$.
\begin{corollary}
Let $f$ be a partial function. Then
$\R(f^n)\geq\RS(f)^n/3$.
\end{corollary}
This follows straightforwardly from observing that $\R(f^n)=\R_{1/3}(f^n)\geq(1-2/3)\RS(f^n)$ (using \thm{RgeqRS}) and $\RS(f^n) \geq \RS(f)^n$ (using \thm{RS_compose}).
We can also prove a composition theorem for zero-error and bounded-error randomized query complexity in terms of randomized sabotage complexity. In particular this yields a composition theorem for $\R(f\circ g)$ when $\R(g)=\Theta(\RS(g))$.
\begin{theorem}\label{thm:R_compose}
Let $f$ and $g$ be partial functions. Then
$\bar{\R}_\epsilon(f\circ g)\geq\bar{\R}_\epsilon(f)\RS(g)$.
\end{theorem}
\begin{proof}
The proof follows a similar argument to the proof of
\thm{RS_compose}. Let $A$ be a randomized
algorithm for $f\circ g$ that uses $T$ expected queries
and makes error $\epsilon$.
We turn $A$ into an algorithm $B$ for $f$
by having $B$ generate inputs from $\mu$, the hard distribution for $g_\mathrm{sab}$,
and feeding them to $A$, as before. The only difference is that this
time, the input $x$ to $B$ is not a sabotaged input.
This means it has no $*$ or $\dag$ entries, so all
the sabotaged inputs that $B$ generates turn into
$0$- or $1$-inputs if $A$ tries to query a $*$ or $\dag$
in them.
Since $A$ uses $T$ queries, by \thm{thresh_sum},
it finds at most $T/\RS(g)$ asterisks or obelisks (in expectation).
Therefore, $B$ makes at most $T/\RS(g)$ expected
queries to $x$. Since $B$ is correct whenever $A$ is correct,
its error probability is at most $\epsilon$.
Thus $\bar{\R}_\epsilon(f)\leq T/\RS(g)$, and thus $T\geq \bar{\R}_\epsilon(f)\RS(g)$.
\end{proof}
Setting $\epsilon$ to $0$ yields the following corollary.
\begin{corollary}
Let $f$ and $g$ be partial functions. Then
$\R_0(f\circ g)\geq\R_0(f)\RS(g)$.
\end{corollary}
For the more commonly used $\R(f\circ g)$, we obtain the following composition result.
\begin{corollary}\label{cor:R_compose}
Let $f$ and $g$ be partial functions. Then
$\R(f\circ g)\geq\R(f)\RS(g)/10$.
\end{corollary}
This follows from \lem{Rexp}, which gives $\bar{\R}_{1/3}(f)\geq\R(f)/10$, and \thm{R_compose}, since
$\R(f\circ g)\geq\bar{\R}_{1/3}(f\circ g) \geq \bar{\R}_{1/3}(f)\RS(g)
\geq \R(f)\RS(g)/10$.
Finally, we can also show an upper bound composition result for randomized sabotage complexity.
\begin{restatable}{theorem}{uppercompose}
\label{thm:upper_compose}
Let $f$ and $g$ be partial functions. Then
$\RS(f\circ g)\leq \RS(f)\R_0(g)$. We also have
$\RS(f\circ g)=O(\RS(f)\R(g)\log\RS(f))$.
\end{restatable}
\begin{proof}
We describe a simple algorithm for finding a $*$ or $\dag$
in an input to $f\circ g$. Start by running
the optimal algorithm for the sabotage problem of $f$.
This algorithm uses $\RS(f)$ expected queries.
Then whenever this algorithm tries to query a bit,
run the optimal zero-error algorithm for $g$ in the
corresponding input to $g$.
Now, since the input to $f\circ g$ that we are given
is a sabotaged input, it must be consistent with both
a $0$-input and a $1$-input of $f\circ g$. It follows
that some of the $g$ inputs are sabotaged, and moreover,
if we represent a sabotaged $g$-input by $*$ or $\dag$,
a $0$-input to $g$ by $0$, and a $1$-input to $g$ by $1$,
we get a sabotaged input to $f$. In other words, from the inputs to $g$ we can derive a sabotaged input for $f$.
This means that the outer algorithm runs uses an expected $\RS(f)$
calls to the inner algorithm, and ends up calling the
inner algorithm on a sabotaged input to $g$.
Meanwhile, each call to the inner algorithm
uses an expected $\R_0(g)$ queries,
and will necessarily find a $*$ or $\dag$ if the input it is
run on is sabotaged. Therefore, the described algorithm
will always find a $*$ or $\dag$, and its expected running time
is $\RS(f)\R_0(g)$ by linearity of expectation
and by the independence of the internal randomness
of the two algorithms.
Instead of using a zero-error randomized algorithm for $g$, we can
use a bounded-error randomized algorithm for $g$ as long as its error probability is small.
Since we make $O(\RS(f))$ calls to the inner algorithm, if we boost the bounded-error algorithm's
success probability to make the error much smaller than $1/\RS(f)$
(costing an additional $\log \RS(f)$ factor), we will get a bounded-error
algorithm for $(f\circ g)_\mathrm{sab}$. Since $\R((f\circ g)_\mathrm{sab})$ is the same as $\RS(f\circ g)$ up to a constant factor (\thm{RSR0S}),
\begin{equation}
\RS(f\circ g)=O(\RS(f)\R(g)\log\RS(f)),
\end{equation}
as desired.
\end{proof}
\section{Composition with the index function}
\label{sec:composition}
We now prove our main result (\thm{comp}) restated more precisely as follows.
\begin{namedtheorem}{Theorem \ref*{thm:comp}}{Precise version}
Let $f$ and $g$ be (partial) functions, and
let $m=\Omega(\R(g)^{1.1})$. Then
$\R(f\circ\textsc{Ind}_m\circ g)=\Omega(\R(f)\R(g)\log m)
=\Omega(\R(f)\R(\textsc{Ind}_m)\R(g))$.
\end{namedtheorem}
Before proving this, we formally define the index function.
\begin{definition}[Index function]
The index function on $m$ bits, denoted $\textsc{Ind}_m:\{0,1\}^m \to \{0,1\}$, is defined as follows. Let $c$ be the largest integer such that $c+2^c\leq m$. For any input $x\in \{0,1\}^m$, let $y$ be the first $c$ bits of $x$ and let $z=z_0z_1\cdots z_{2^c-1}$ be the next $2^c$ bits of $x$. If we interpret $y$ as the binary representation of an integer between $0$ and $2^c-1$, then the output of $\textsc{Ind}_m(x)$ equals $z_y$.
\end{definition}
To prove \thm{comp}, we also require the strong direct product theorem for randomized query complexity that was established by Drucker \cite{Dru12}.
\begin{theorem}[Strong direct product]\label{thm:dir_prod}
Let $f$ be a partial Boolean function,
and let $k$ be a positive integer.
Then any randomized algorithm for $f^{\oplus k}$ that uses
at most $\gamma^3k\R(f)/11$ queries has success probability
at most $(1/2+\gamma)^k$, for any $\gamma\in(0,1/4)$.
\end{theorem}
The first step to proving $\R(f\circ \textsc{Ind} \circ g) = \Omega(\R(f)\R(\textsc{Ind})\R(g))$ is to establish that $\R(\textsc{Ind}\circ g)$ is essentially the same as $\RS(\textsc{Ind} \circ g)$ if the index gadget is large enough.
\begin{lemma}\label{lem:index}
Let $f$ be a partial Boolean function and let
$m=\Omega(\R(f)^{1.1})$.
Then
\begin{equation}
\RS(\textsc{Ind}_m\circ f)=\Omega(\R(f)\log m)
=\Omega(\R(\textsc{Ind}_m)\R(f)).
\end{equation}
Moreover, if $f^{\oplus c}_{\mathrm{ind}}$
is the defined as the index function on $c+2^c$ bits
composed with $f$ in only the first $c$ bits,
we have $\RS(f^{\oplus c}_{\mathrm{ind}}) \geq \RS_{\mathrm{u}}(f^{\oplus c}_{\mathrm{ind}})=\Omega(c\R(f))$
when $c\geq 1.1\log \R(f)$.
\end{lemma}
Before proving \lem{index}, let us complete the proof of \thm{comp} assuming \lem{index}.
\begin{proof}[Proof of \protect{\thm{comp}}]
By \cor{R_compose}, we have
$\R(f\circ\textsc{Ind}_m\circ g)\geq \R(f)\RS(\textsc{Ind}_m\circ g)/10$.
Combining this with \lem{index} gives
$\R(f\circ\textsc{Ind}_m\circ g)=\Omega(\R(f)\R(g)\log m)$,
as desired.
\end{proof}
We can now complete the argument by proving \lem{index}.
\begin{proof}[Proof of \protect{\lem{index}}]
To understand what the inputs to $(\textsc{Ind}_m\circ f)_{\mathrm{sab}}$
look like, let us first analyze the function $\textsc{Ind}_m$.
We can split an input to $\textsc{Ind}_m$ into a small index
section and a large array section. To sabotage
an input to $\textsc{Ind}_m$, it suffices to sabotage
the array element that the index points to (using
only a single $*$ or $\dag$). It follows that to sabotage
an input to $\textsc{Ind}_m\circ f$, it suffices to sabotage
the input to $f$ at the array element that the index
points to. In other words, we consider sabotaged inputs where
the only stars in the input are in one array cell whose index is the output
of the first $c$ copies of $f$, where $c$ is the largest integer such that $c+2^c\leq m$. Note that
$c=\log m-\Theta(1)$.
We now convert any $\RS(\textsc{Ind}_m\circ f)$ algorithm
into a randomized algorithm for $f^{\oplus c}$.
First, using \lem{markov}, we get a
$2\RS(\textsc{Ind}_m\circ f)$ query randomized
algorithm that finds a $*$ or $\dag$ with probability $1/2$
if the input is sabotaged. Next, consider running
this algorithm on a non-sabotaged input. It makes
$2\RS(\textsc{Ind}_m\circ f)$ queries. With probability
$1/2$, one of these queries will be in the array cell
whose index is the true answer to $f^{\oplus c}$ evaluated
on the first $cn$ bits. We can then consider
a new algorithm $A$ that runs the
above algorithm for $2\RS(\textsc{Ind}_m\circ f)$ queries,
then picks one of the $2\RS(\textsc{Ind}_m\circ f)$
queries at random, and if that query is in an array
cell, it outputs the index of that cell.
Then $A$ uses $2\RS(\textsc{Ind}_m\circ f)$ queries
and evaluates $f^{\oplus c}$ with probability
at least $\RS(\textsc{Ind}_m\circ f)^{-1}/4$.
Next, \thm{dir_prod} implies that for any $\gamma\in(0,1/4)$,
either $A$'s success
probability is smaller than $(1/2+\gamma)^c$,
or else $A$ uses at least
$\gamma^3c\R(f)/11$ queries. This means either
\begin{equation}
\label{eq:dpt}
\RS(\textsc{Ind}_m\circ f)^{-1}/4 \leq (1/2+\gamma)^{c}
\qquad \textrm{or} \qquad
2\RS(\textsc{Ind}_m\circ f)\geq \gamma^3c\R(f)/11.
\end{equation}
Now if we choose $\gamma=0.01$, it is clear that the second inequality in \eq{dpt} yields
$\RS(\textsc{Ind}_m\circ f)=\Omega(c \R(f))=\Omega(\R(f)\log m)$ no matter what $m$ (and hence $c$) is chosen to be.
To complete the argument, we show that the first inequality in \eq{dpt} also yields the same.
Observe that the first inequality is equivalent to
\[
\RS(\textsc{Ind}_m\circ f)
=\Omega\left(\Bigl(\frac{2}{1+2\gamma}\Bigr)^c\right)
=\Omega\left(\Bigl(\frac{2}{1+2\gamma}\Bigr)^{\log m-\Theta(1)}\right)
=\Omega(m^{\log_2(2/1.02)})
=\Omega(m^{0.97}).
\]
We now have $m^{0.97}=\Omega(m^{0.96}\log m)
=\Omega(\R(f)^{1.1\times 0.96}\log m)
=\Omega(\R(f)\log m)$, as desired.
The lower bound on $\RS_{\mathrm{u}}(f^{\oplus c}_{\mathrm{ind}})$
follows similarly once we makes two observations. First, this argument works
equally well for $f^{\oplus c}_{\mathrm{ind}}$ instead of $\textsc{Ind}_m \circ f$.
Second, sabotaging the array cell indexed by the outputs
to the $c$ copies of $f$ in $f^{\oplus c}_{\mathrm{ind}}$ introduces only
one asterisk or obelisk, so the argument above
lower bounds $\RS_{\mathrm{u}}(f^{\oplus c}_{\mathrm{ind}})$ and not only $\RS(f^{\oplus c}_{\mathrm{ind}})$.
\end{proof}
\section{Relating lifting theorems}
\label{sec:lifting}
In this section we establish \thm{lifting}, which proves that a lifting theorem for zero-error randomized communication complexity implies one for bounded-error randomized communication complexity.
To begin, we introduce the two-party index gadget (also used in \cite{GPW15}).
\begin{definition}[Two-party index gadget]\label{def:commindex}
For any integer $b>0$, and finite set $\mathcal{Y}$, we define the index function $G_\textsc{Ind}:\{0,1\}^b \times \mathcal{Y}^{2^b} \to \mathcal{Y}$ as follows. Let $(x,y)\in\{0,1\}^b \times \mathcal{Y}^{2^b}$ be an input to $G_\textsc{Ind}$. Then if we interpret $x$ as the binary representation of an integer between $0$ and $2^b-1$, the function $G_\textsc{Ind}(x,y)$ evaluates to $y_x$, the $x^\textrm{th}$ letter of y. We also let $G_b$ be the index function with $\mathcal{Y} = \{0,1\}$ and let $G'_b$ be the index function with $\mathcal{Y} = \{0,1,*,\dagger\}$.
\end{definition}
The index gadget is particularly useful in communication complexity because it is ``complete'' for functions with a given value of $\min\{|\mathcal{X}|,|\mathcal{Y}|\}$.
More precisely, any problem $F:\mathcal{X} \times \mathcal{Y} \to \{0,1\}$ can be reduced to $G_b$ for $b = \lceil \log \min\{|\mathcal{X}|,|\mathcal{Y}|\}\rceil$.
To see this, say $|\mathcal{X}|\leq |\mathcal{Y}|$ and let $|\mathcal{X}|=2^b$.
We now map every input $(x,y)\in\mathcal{X} \times \mathcal{Y}$ to an input $(x',y')$ for $G_b$.
Since $\mathcal{X}$ has size $2^b$, we can view $x$ as a string in $\{0,1\}^b$ and set $x' = x$.
The string $y'=y'_0 y'_1 \cdots y'_{2^b-1} \in \{0,1\}^{2^b}$ is defined as $y'_x = F(x,y)$.
Hence we can assume without loss of generality that a supposed lifting theorem for zero-error protocols is proved using the two-party index gadget of some size.
Our first step is to lower bound the bounded-error randomized communication complexity
of a function in terms of the zero-error randomized communication complexity of a related function.
\begin{lemma}\label{lem:communication_sab}
Let $f$ be an $n$-bit (partial) Boolean function
and let $G_b:\{0,1\}^b\times\{0,1\}^{2^b}\to\{0,1\}$ be the index gadget with $b=O(\log n)$. Then
\begin{equation}\R^\mathrm{cc}(f\circ G_b)=\Omega\left(
\frac{\R^\mathrm{cc}_0(f_\mathrm{usab}\circ G^\prime_b)}
{\log n\log\log n}\right),\end{equation}
where $G^\prime_b$ is the index gadget mapping
$\{0,1\}^b\times\{0,1,*,\dagger\}^{2^b}$ to $\{0,1,*,\dagger\}$.
\end{lemma}
\begin{proof}
We will use a randomized protocol $A$ for $f\circ G_b$
to construct a zero-error protocol $B$ for
$f_\mathrm{usab}\circ G^\prime_b$. Note the given input to
$f_\mathrm{usab}\circ G^\prime_b$ must have a unique
copy of $G^\prime_b$ that evaluates to $*$ or $\dagger$,
with all other copies evaluating to $0$ or $1$.
The goal of $B$ is to find this copy and determine if
it evaluates to $*$ or $\dagger$. This will evaluate
$f_\mathrm{usab}\circ G^\prime_b$ with zero error.
Note that if we replace all $*$ and $\dagger$ symbols
in Bob's input with $0$ or $1$, we would get a valid input to
to $f\circ G_b$, which we can evaluate using $A$.
Moreover, there is a single special
$*$ or $\dagger$ in Bob's input
that governs the value of this input to $f\circ G_b$
no matter how we fix the rest of the $*$ and $\dagger$ symbols.
Without loss of generality, we assume that if the special
symbol is replaced by $0$, the function $f\circ G_b$
evaluates to $0$, and if it is replaced by $1$,
it evaluates to $1$.
We can now binary search to find this special symbol.
There are at most $n 2^b$ asterisks and obelisks in Bob's input.
We can set the left half to $0$ and the right half to $1$,
and evaluate the resulting input using $A$. If the answer
is $0$, the special symbol is on the left half; otherwise,
it is on the right half. We can proceed to binary search
in this way, until we have zoomed in on one gadget that must
contain the special symbol. This requires narrowing
down the search space from $n$ possible gadgets to $1$,
which requires $\log n$ rounds. Each round requires a call
to $A$, times a $O(\log\log n)$ factor for error reduction.
We can therefore find the right gadget with bounded
error, using $O(\R^\mathrm{cc}(f\circ G_b)\log n\log\log n)$
bits of communication.
Once we have found the right gadget, we can certify its validity
by having Alice send the right index to Bob, using $b$ bits
of communication, and Bob can check that it points to an asterisk or obelisk.
Since we found a certificate with constant
probability, we can use \lem{repeat} to turn this into
a zero-error algorithm. Thus
\begin{equation}\R^\mathrm{cc}_0(f_\mathrm{usab}\circ G^\prime_b)=
O(b+\R^\mathrm{cc}(f\circ G_b)\log n\log\log n).\end{equation}
Since $b=O(\log n)$, we obtain the desired result
$\R^\mathrm{cc}_0(f_\mathrm{usab}\circ G^\prime_b)=
O(\R^\mathrm{cc}(f\circ G_b)\log n\log\log n)$.
\end{proof}
Equipped with this lemma we can prove the connection between lifting theorems (\thm{lifting}), stated more precisely as follows.
\begin{namedtheorem}{Theorem \ref*{thm:lifting}}{Precise version}
Suppose that for all partial Boolean functions $f$
on $n$ bits, we have
\begin{equation}\R_0^\mathrm{cc}(f\circ G_b)=\Omega(\R_0(f)/\polylog n)\end{equation}
with $b=O(\log n)$.
Then for all partial functions Boolean functions, we also have
\begin{equation}\R^\mathrm{cc}(f\circ G_{2b})=\Omega(\R(f)/\polylog n).\end{equation}
The $\polylog n$ loss in the $\R^\mathrm{cc}$ result
is only $\log n\log\log^2 n$ worse than the loss in the
$\R_0^\mathrm{cc}$ hypothesis.
\end{namedtheorem}
\begin{proof}
First we show that
for any function $f$ and positive integer $c$,
\begin{equation}\label{eq:find}
\R^\mathrm{cc}(f\circ G_{2b})=\Omega\left(
\frac{\R^\mathrm{cc}(f^{\oplus c}_{\mathrm{ind}}\circ G_{2b})}
{c\log c}\right).
\end{equation}
To see this, note that
we can solve $f^{\oplus c}_{\mathrm{ind}}\circ G_{2b}$
by solving the $c$ copies of $f\circ G_{2b}$
and then examining the appropriate cell of the array.
This uses $c\R^\mathrm{cc}(f\circ G_{2b})$ bits of communication,
times $O(\log c)$ since we must amplify the randomized
protocol to an error of $O(1/c)$.
Next, using \eq{find} and \lem{communication_sab} on
$\R^\mathrm{cc}(f^{\oplus c}_{\mathrm{ind}}\circ G_{2b})$, we get
\begin{equation}\label{eq:Rcc}
\R^\mathrm{cc}(f\circ G_{2b})=\Omega\left(
\frac{\R^\mathrm{cc}(f^{\oplus c}_{\mathrm{ind}}\circ G_{2b})}
{c\log c}\right)
=\Omega\left(
\frac{\R_0^\mathrm{cc}((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}
\circ G^\prime_{2b})}
{c\log c\log n\log\log n}\right).
\end{equation}
From here we want to use the assumed lifting theorem
for $\R_0$. However, there is a technicality: the gadget
$G^\prime_{2b}$ is not the standard index gadget, and
the function $(f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}$
does not have Boolean
alphabet. To remedy this, we use two bits to represent
each of the symbols $\{0,1,*,\dagger\}$.
Using this representation, we define a new function
$(f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}^{\mathrm{bin}}$ on twice as many bits.
We now compare
$(f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}^{\mathrm{bin}}\circ G_b$
to $(f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}\circ G^\prime_{2b}$.
Note that the former uses two pointers of size $b$ to index
two bits, while the latter uses one pointer of size $2b$
to index one symbol in $\{0,1,*,\dagger\}$
(which is equivalent to two bits).
It's not hard to see that the former function
is equivalent to the latter function restricted to a promise.
This means the communication complexity of the former
is smaller, and hence
\begin{equation}\label{eq:usab1}
\R_0^\mathrm{cc}((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}
\circ G^\prime_{2b})
=\Omega({\R_0^\mathrm{cc}((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}^{\mathrm{bin}}
\circ G_b)}).
\end{equation}
We are now ready to use the assumed lifting theorem for $\R_0$.
To be more precise, let's suppose a lifting result
that states $\R_0^\mathrm{cc}(f\circ G_b)=\Omega(\R_0(f)/\ell(n))$
for some function $\ell(n)$. Thus
\begin{equation}\label{eq:usab2}
{\R_0^\mathrm{cc}((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}^{\mathrm{bin}}\circ G_b)}
=\Omega({\R_0((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}^{\mathrm{bin}})}/
{\ell(n)}).
\end{equation}
We note that
\begin{equation}\label{eq:usab3}
\R_0((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}^{\mathrm{bin}})
=\Omega(\R_0((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}))
=\Omega(\RS_{\mathrm{u}}(f^{\oplus c}_{\mathrm{ind}})).\end{equation}
Setting $c=1.1\log\R(f)$, we have
$\RS_{\mathrm{u}}(f^{\oplus c}_{\mathrm{ind}})=\Omega(c\R(f))$ by \lem{index}. Combining this with \eq{usab1}, \eq{usab2}, and \eq{usab3}, we get
\begin{equation}
\R_0^\mathrm{cc}((f^{\oplus c}_{\mathrm{ind}})_\mathrm{usab}
\circ G^\prime_{2b})
=\Omega(c\R(f)/\ell(n)).
\end{equation}
Combining this with \eq{Rcc} yields
\begin{equation}
\R^\mathrm{cc}(f\circ G_{2b})=\Omega\left(
\frac{c\R(f)}{\ell(n)c\log c\log n\log\log n}\right)
=\Omega\left(
\frac{\R(f)}{\ell(n)\log n\log\log^2 n}\right).
\end{equation}
This gives the desired lifting theorem for bounded-error randomized communication with
$\polylog n$ loss that is at most $\log n\log\log^2 n$ worse than the loss in the
assumed $\R_0^\mathrm{cc}$ lifting theorem.
\end{proof}
\section{Comparison with other lower bound methods}
\label{sec:comparison}
\setlength{\intextsep}{0pt}%
\setlength{\columnsep}{18pt}%
\begin{wrapfigure}{r}{0.24\textwidth}
\vspace{-0.25em}
\centering
\begin{tikzpicture}[x=0.7cm,y=1.25cm]
\node (R) at(2,2){$\R$};
\node (RS) at(0,1){$\RS$};
\node (prt) at(2,1){$\prt$};
\node (RC) at(1,0){$\RC$};
\node (Q) at(4,1){$\Q$};
\node (adeg) at(3,0){$\adeg$};
\path[-] (RS) edge (R);
\path[-] (prt) edge (R);
\path[-] (Q) edge (R);
\path[-] (RC) edge (RS);
\path[-] (RC) edge (prt);
\path[-] (adeg) edge (Q);
\path[-] (adeg) edge (prt);
\end{tikzpicture}
\vspace{-.5em}
\caption{Lower bounds on $\R(f)$.\label{fig:lower}}
\end{wrapfigure}
In this section we compare $\RS(f)$ with other lower bound techniques for bounded-error randomized query complexity.
\fig{lower} shows the two most powerful lower bound techniques for $\R(f)$, the partition bound ($\prt(f)$) and quantum query complexity ($\Q(f)$), which subsume all other general lower bound techniques.
The partition bound and quantum query complexity are incomparable, since there are functions for which the partition bound is larger, e.g., the $\textsc{Or}$ function, and functions for which quantum query complexity is larger \cite{AKK15}.
Another common lower bound measure, approximate polynomial degree ($\adeg$) is smaller than both.
Randomized sabotage complexity ($\RS$) can be much larger than the partition bound and quantum query complexity as we now show.
We also show that randomized sabotage complexity is always as large as randomized certificate complexity ($\RC$), which itself is larger than block sensitivity, another common lower bound technique.
Lastly, we also show that $\R_0(f) = O(\RS(f)^2 \log \RS(f))$, showing that $\RS$ is a quadratically tight lower bound, even for zero-error randomized query complexity.
\subsection{Partition bound and quantum query complexity}
We start by showing the superiority of randomized sabotage complexity against the two best lower bounds for $\R(f)$.
Informally, what we show is that any separation between $\R(f)$ and a lower bound measure like $\Q(f)$, $\prt(f)$, or $\adeg(f)$ readily gives a similar separation between $\RS(f)$ and the same measure.
\begin{theorem}
\label{thm:comparison}
There exist total functions $f$ and $g$ such that $\RS(f) \geq \prt(f)^{2-o(1)}$ and $\RS(g) = \widetilde{\Omega}(\Q(g)^{2.5})$. There also exists a total function $h$ with $\RS(h) \geq \adeg(h)^{4-o(1)}$.
\end{theorem}
\begin{proof}
These separations were shown with $\R(f)$ in place of
$\RS(f)$ in \cite{ABK15} and \cite{AKK15}.
To get a lower bound on $\RS$, we can simply compose
$\textsc{Ind}$ with these functions and apply \lem{index}.
This increases $\RS$ to be the same as $\R$ (up to logarithmic
factors), but it does not increase $\prt$, $\adeg$, or $\Q$
more than logarithmically, so the desired separations follow.
\end{proof}
As it turns out, we didn't even need to compose $\textsc{Ind}$ with
these functions. It suffices to observe that they all use
the cheat sheet construction, and that an argument
similar to the proof of \lem{index}
implies that $\RS(f_{\mathrm{CS}})=\widetilde{\Omega}(\R(f))$ for all $f$
(where $f_{\mathrm{CS}}$ denotes the cheat sheet version of $f$,
as defined in \cite{ABK15}). In particular,
cheat sheets can never be used to separate $\RS$ from $\R$
(by more than logarithmic factors).
\subsection{Randomized certificate complexity}
Finally, we also show that randomized sabotage complexity upper bounds randomized certificate complexity. To show this, we first define randomized certificate complexity.
Given a string $x$, a block is a set of bits of $x$
(that is, a subset of $\{1,2,\ldots,n\}$).
If $B$ is a block and $x$ is a string, we denote by $x^B$
the string given by flipping the bits specified by $B$
in the string $x$. If $x$ and $x^B$ are both in the domain
of a (possibly partial) function $f:\{0,1\}^n\to\{0,1\}$
and $f(x) \neq f(x^B)$, we say that $B$
is a sensitive block for $x$ with respect to $f$.
For a string $x$ in the domain $f$, the maximum number
of disjoint sensitive blocks of $x$ is called the
block sensitivity of $x$, denoted by $\bs_x(f)$.
The maximum of $\bs_x(f)$ over all $x$ in the domain
of $f$ is the block sensitivity of $f$, denoted by $\bs(f)$.
A fractionally disjoint set of sensitive blocks of $x$
is an assignment of non-negative weights to the sensitive blocks
of $x$ such that for all $i\in\{1,2,\ldots,n\}$,
the sum of the weights of blocks containing
$i$ is at most $1$. The maximum total weight
of any fractionally disjoint set of sensitive blocks
is called the fractional block sensitivity of $x$.
This is also sometimes called the randomized certificate
complexity of $x$, and is denoted by $\RC_x(f)$ \cite{Aar08,Tal13,GSS16}.
The maximum of this over all $x$ in the domain of $f$
is $\RC(f)$ the randomized certificate complexity of $f$.
Aaronson \cite{Aar08} observed that $\bs_x(f)\leq\RC_x(f)\leq\C_x(f)$.
We therefore have
\begin{equation}
\bs(f)\leq\RC(f)\leq\C(f)\leq\R_0(f)\leq\D(f).
\end{equation}
The measure $\RC(f)$ is also a lower bound for $\R(f)$;
indeed, from arguments in \cite{Aar08} it follows that
$\R_\epsilon(f)\geq\RC(f)/(1-2\epsilon)$,
so $\R(f)\geq\RC(f)/3$.
\begin{theorem}\label{thm:RC}
Let $f:\{0,1\}^n\to\{0,1\}$ be a partial function. Then
$\RS(f)\geq\RC(f)/4$.
\end{theorem}
\begin{proof}
Let $x$ be the input that maximizes $\RC_x(f)$.
Let $B_1,B_2,\dots B_m$ be all the (not necessarily disjoint)
sensitive blocks of $x$. For each $i \in \{1,2,\dots,m\}$,
let $y_i$ be the sabotaged input formed by replacing
block $B_i$ in $x$ with $*$ entries.
Finding a $*$ in an input chosen from
$Y=\{y_1,y_2,\dots,y_m\}$ is a special case of the sabotage
problem for $f$, so it can be done in $\RS(f)$ expected
queries.
We now use reasoning from \cite{Aar08} to turn this
into a non-adaptive algorithm.
By \lem{markov}, after $\lfloor 2\RS(f)\rfloor$ queries,
we find a $*$ with probability at least $1/2$.
For each $t$ between $1$ and $T=\lfloor2\RS(f)\rfloor$,
let $p_t$ be the probability that the
adaptive algorithm finds a $*$ on query $t$,
conditioned on the previous queries not finding a $*$.
Then we have
\begin{equation}p_1+p_2+\dots+p_T\geq \frac{1}{2}.\end{equation}
If we pick $t\in\{1,2,\dots,T\}$ uniformly and simulate
query $t$ of the adaptive algorithm (which
is possible since we know $x$ and are assuming
the previous $t-1$ queries did not find a $*$),
we must find
a $*$ with probability at least
$1/(2T) \geq 1/(4\RS(f))$.
This is a non-adaptive algorithm for finding a $*$,
so it is also a non-adaptive algorithm for finding
a difference from $x$.
Let the probability distribution over inputs bits obtained from this non-adaptive
algorithm be $(q_1,q_2,\dots,q_n)$, so that the algorithm
queries bit $i$ with probability $q_i$.
We have $\sum_{i=1}^n q_i=1$ and for each sensitive block $B_j$, we have
$\sum_{i\in B_j} q_i\geq 1/(4\RS(f))$.
For each sensitive block $B_j$, let $w_j$ be the weight of $B_j$ under the maximum fractional set of disjoint
blocks. Then $\sum_{j=1}^m w_j=\RC(f)$ and
for each bit $i$, we have $\sum_{j:i\in B_j} w_j\leq 1$.
We then have
\[\frac{\RC(f)}{4\RS(f)}
=\sum_{j=1}^m w_j\cdot\frac{1}{4\RS(f)}
\le \sum_{j=1}^m w_j\sum_{i\in B_j} q_i
=\sum_{i=1}^n q_i\sum_{j:i\in B_j} w_j
\leq \sum_{i=1}^n q_i\cdot 1
\leq 1.\]
Hence $\RS(f)\geq\RC(f)/4$.
\end{proof}
\subsection{Zero-error randomized query complexity}
\begin{theorem}\label{thm:rootR0}
Let $f:\{0,1\}^n\to\{0,1\}$ be a total function. Then $\R_0(f) = O(\RS(f)^2 \log \RS(f))$ or alternately,
$\RS(f)=\Omega(\sqrt{\R_0(f)/\log\R_0(f)})$.
\end{theorem}
\begin{proof}
Let $A$ be the $\RS(f)$ algorithm. The idea
will be to run $A$ on an input to $x$ for long enough
that we can ensure it queries a bit in every sensitive
block of $x$; this will mean $A$ found a certificate
for $x$. That will allow us to turn the algorithm
into a zero-error algorithm for $f$.
Let $x$ be any input, and let $b$ be a sensitive block of $x$.
If we replace the bits of $x$ specified by $b$ with stars,
then we can find a $*$ with probability $1/2$
by running $A$ for $2\RS(f)$ queries
by \lem{markov}. This means that if we run $A$ on $x$
for $2\RS(f)$ queries,
it has at least $1/2$ probability of querying a bit
in any given sensitive block of $x$. If we repeat this $k$ times,
we get a $2k\RS(f)$ query algorithm that queries a bit
in any given sensitive block of $x$ with probability at least
$1-2^{-k}$.
Now, by \cite{KT13}, the number of minimal sensitive blocks in $x$
is at most $\RC(f)^{\bs(f)}$ for a total function $f$.
Our probability of querying a bit in all of these sensitive blocks
is at least $1-2^{-k}\RC(f)^{\bs(f)}$ by the union bound.
When $k\geq 1+\bs(f)\log_2\RC(f)$, this is at least $1/2$.
Since a bit from every sensitive block is a certificate,
by \lem{repeat}, we can turn this into a zero-error
randomized algorithm with expected query complexity
at most $4(1+\bs(f)\log_2\RC(f))\RS(f)$, which gives
$\R_0(f) = O(\RS(f)\bs(f)\log \RC(f))$.
Since $\bs(f)\leq\RC(f)=O(\RS(f))$ by \thm{RC}, we have
$\R_0(f)=O(\RS(f)^2\log \RS(f))$,
or
$\RS(f)=\Omega(\sqrt{\R_0(f)/\log\R_0(f)})$.
\end{proof}
\section{Deterministic sabotage complexity}
\label{sec:other}
Finally we look at the deterministic analogue of randomized sabotage complexity.
It turns out that deterministic sabotage complexity (as defined in \defn{RS}) is exactly the same as deterministic query complexity for all (partial) functions.
Since we already know perfect composition and direct sum results for deterministic query complexity, it is unclear if deterministic sabotage complexity has any applications.
\begin{theorem}\label{thm:DS}
Let $f:\{0,1\}^n\to\{0,1\}$ be a partial function. Then $\DS(f)=\D(f)$.
\end{theorem}
\begin{proof}
For any function $\DS(f) \leq D(f)$ since a deterministic algorithm that correctly computes $f$ must find a $*$ or $\dag$ when run on a sabotaged input, otherwise its output is independent of how the sabotaged bits are filled in.
To show the other direction, let $\D(f) = k$. This means for every $k-1$ query algorithm, there are two inputs $x$ and $y$ with $f(x) \neq f(y)$, such that they have the same answers to the queries made by the algorithm. If this is not the case then this algorithm computes $f(x)$, contradicting the fact that $\D(f)=k$. Thus if there is a deterministic algorithm for $f_\mathrm{sab}$ that makes $k-1$ queries, there exist two inputs $x$ and $y$ with $f(x) \neq f(y)$ that have the same answers to the queries made by the algorithm. If we fill in the rest of the inputs bits with either asterisks or obelisks, it is clear that this is a sabotaged input (since it can be completed to either $x$ or $y$), but the purported algorithm for $f_\mathrm{sab}$ cannot distinguish them. Hence $\D(f_\mathrm{sab}) \geq k$, which means $\DS(f) \geq \D(f)$.
\end{proof}
\section*{Acknowledgements}
We thank Mika G\"o\"os for finding an error in an earlier proof of \thm{thresh_sum}. We also thank the anonymous referees of Theory of Computing for their comments.
This work is partially supported by ARO grant number W911NF-12-1-0486. This preprint is MIT-CTP \#4806.
\clearpage
|
\section{Introduction}
On 2015 September 14, both detectors from the Laser Interferometer Gravitational-Wave Observatory (LIGO) made the first direct detection of gravitational waves \citep{LIGO:2016_main}. The gravitational waves were emitted by two merging black holes of $M_1=36_{-4}^{+5}\msun$ and $M_2=29_{-4}^{+4}\msun$ located at redshift $z=0.09_{-0.04}^{+0.03}$. While the detection of black holes much heavier than any mass measured in X-ray binaries revives the study of the evolution of massive stellar binaries \citep{Mandel:2016,Marchant:2016}, the determination of host galaxies has been mostly ignored. So far, binary population syntheses (BPS) models have argued that GW150914 can only have formed in a low metallicity environment, below $0.25\Zsun$, most probably around $0.1\Zsun$ \citep{LIGO:2016_implications,Belczynski:2016}.
This strong limit on the progenitor metallicity allows one to determine in what type of galaxy and at what time the progenitors of massive ($M_1+M_2 \geqslant 40 \msun$, $M_1,M_2\geqslant 15\msun$) binary black holes (BBH) are born. While previous work has determined merger rates as a function of redshift \citep{Dominik:2013,Dvorkin:2016}, this work presents the first determination of the formation conditions for the massive BBH mergers we currently observe. As the delay time between progenitor formation and BBH merger often exceeds several Gyr, one has to consider star formation through cosmic history to correctly model the progenitor population.
Low metallicity gas is typically found in high redshift galaxies or in local dwarf galaxies. Using a two component model for the star formation and metallicity as a function of redshift, \citet{OShaughnessy:2010} showed that elliptical galaxies dominate BBH mergers hosts. Based on the redshift evolution of a mass-independent metallicity distribution with significant scatter, \citet{Belczynski:2016} suggests two roughly equally probable formation times for GW150914 around $z\simeq 3$ and below $z\simeq 0.2$. In this work we use a complete, redshift dependent mass-metallicity relation (MZR) consistent with recent high-redshift observations \citep{Erb:2006,Mannucci:2009}. Additionally, we explicitly account for galaxy mergers that bring low metallicity stars/black holes formed in low mass galaxies to higher mass galaxies, where the BBH mergers take place.
In order to determine the environment in which GW150914 formed, we assume the progenitors have the metallicity of the gas in which they form. First, we determine the amount of low metallicity star formation through cosmic history (\S\ref{sec:lowZ}). Using a binary population synthesis (BPS) model, we then determine the delay time distribution for various progenitor metallicities (\S\ref{sec:mergers}). We finally combine both computations to determine where GW150914 most likely formed (\S\ref{sec:combine}) and discuss the implications for future detections (\S\ref{sec:discussion}). In this paper, we assume a $\Lambda$CDM cosmology with $h=0.7$, $\Omega_{\Lambda}=0.7$ and $\Omega_m=0.3$ and use $\Zsun=0.02$.
\section{Forming low metallicity stars}\label{sec:lowZ}
The total merger rate $\mathcal{R}$ at merger time $\tm$ is given by
\begin{eqnarray}\label{eqn:NofMgal}
&&\frac{d\mathcal{R}}{d\log \mgal d\tform d\Zc}= \frac{dN}{d\log \mgal d\Zc} \times \\
&&\int_{t_{\infty}}^{\tm} d \tform \frac{dN_{\mathrm{m}}(t_m-\tform,\Zc)}{dt_f} SFR\left(\tform,\mgal \right) \Psi \left(\tform,\mgal,\Zc\right), \nonumber
\end{eqnarray}
where $\mgal$ is the galaxy stellar mass, $\Zc$ the progenitor metallicity. We perform the integral over the formation time of the progenitors $\tform$. SFR is the star formation rate (in $M_{\odot}$ yr$^{-1}$) and $dN/d\log \mgal$ the stellar mass function (SMF) at the time of the merger (per unit comoving volume, taken from \citealt{Tomczak:2014}). $\Psi$ is the fraction of stellar mass forming at metallicity $\Zc$ with respect to the total stellar mass formed. $dN_{\mathrm{m}}/dt$ is the delay time distribution of massive black hole mergers per unit solar mass. In this section we will determine the distribution of low metallicity stellar mass as a function of host mass and formation time while in \S\ref{sec:mergers} we will determine the number of black hole mergers per unit solar mass $N_m$.
To determine the amount of low metallicity stellar mass in each galaxy, we include stars formed within that galaxy as well as stars that were brought into the galaxy through mergers. While this ``ex-situ'' star formation is around $30\%$ in present day galaxies \citep{Lackner:2012}, the accreted stars are typically formed in lower mass galaxies, which have a lower metallicity. As such, the ex-situ component cannot be neglected for our work. The study of galaxy merger histories can be done with hydrodynamic cosmological simulations \citep[e.g][]{Maller:2006} or semi-analytic models of galaxy formation within dark matter halos \citep[e.g][]{Cole:2000,Guo:2008,Fakhouri:2008}. We use the global fit function from \citet{Cole:2008}, based on the Extended Press-Schechter formalism \citep{Lacey:1993}, which provides the redshift dependent mass distribution of progenitor halos that will merge within the main halo by $z=0$. We neglect the evolution of the main halo and galaxy mass between $\zm$ and $z=0$, but for the observed $\zm$ this is extremely small.
For each galaxy mass we consider, we determine the mass of the corresponding dark matter halo using abundance matching by \citet{Behroozi:2013}. Using 10 Myr timesteps, we build up the amount of low metallicity stars within that halo between $z_{\infty}=8$ and $\zm$ according to the merger tree. We neglect stars formed before $z=8$ due to the lack of observational constraints. While the SFR density at such early times is at least two orders of magnitude below its value at the peak of star formation the low metallicity environment will increase their respective contribution to the total merger rate.
For each timestep, we determine the progenitor halo mass function. For each of the progenitors, we determine the corresponding galaxy mass and SFR at that redshift, again using data from \citet{Behroozi:2013}. We then determine the fraction of stars forming at $\Zc$ using the redshift dependent MZR. Finally, we add up the low metallicity contributions of all the progenitors to get the total amount of low metallicity stars formed at the considered redshift that will be in $\mgal$ at $\zm$.
We model 11 metallicity bins between $\Zc=0.01\Zsun$ and $\Zc=\Zsun$, each bin being 0.2 dex wide. We specifically examine BBH progenitors formed at $\Zc /\Zsun =$ 0.3, 0.1, and 0.01 ; $\Zc = 0.1 \Zsun$ is broadly the most likely value \citep{Belczynski:2016}, but progenitors form for $\Zc \lesssim 0.5\Zsun$ . The observational determination of gas-phase metallicities, which is needed to tell us where low metallicity stars form, unfortunately, has systematic uncertainties of $\simeq 0.5$~dex owing to different nebular calibrations \citep{Kewley:2008,Steidel:2014}.
We therefore determine the mean metallicity of the star forming gas using the mass-metallicity relation from \citet{Ma:2016}
\begin{equation}
\label{eq:MZR}
12+\log(\mathrm{O/H})=0.35(\log(\mgal)-10)+0.93 \exp^{-0.43z}+7.95.
\end{equation}
This MZR is based on high-resolution cosmological zoom-in simulations suite FIRE \citep{Hopkins:2014_FIRE}, which reproduce the observed stellar mass-halo mass relation, Kennicutt-Schmidt law, star forming main sequence and star formation histories. More importantly, the simulated MZR agrees with both gas phase and stellar metallicity measurements observed at low redshifts for $10^4\leqslant~\mgal\leqslant~10^{11}\msun$ \citep{Tremonti:2004,Lee:2006} as well as the data at higher redshifts \citep{Erb:2006,Mannucci:2009}. This MZR agrees well with the \citet{Pettini:2004} calibration, removing some of the systematic uncertainties. If, however, we systematically increase all metallicities by switching to the \citet{Kobulnicky:2004} calibration, we obtain the same relative merger rates but lower the total rate by a factor of five.
To determine the actual amount of low metallicity star forming gas within a galaxy, we need to assess the scatter with respect to the mean metallicity, as increased scatter will increase the number of BBH progenitors. \citet{Tremonti:2004} indicate a scatter with $\sigma\simeq .1$ dex between different galaxies independent of redshift. This is significantly lower than the scatter derived from Damped Ly$\alpha$ systems (DLA) \citep{Rafelski:2012}. In the latter, galaxy masses are not measured, and their scatter likely accounts for most of the scatter in metallicity \citep{Dvorkin:2015}. A significant scatter may also be present within a given galaxy. In spiral galaxies, the metallicity decreases by about $0.03-0.06$ dex $\mathrm{kpc}^{-1}$ with galactocentric radius \citep{Henry:2012,Berg:2013}. At a given radius, scatter is typically $\simeq 0.1$ dex. Assuming both radial and non-radial variations of $\sigma\simeq 0.2$, we have a total standard deviation of $\sigma=0.3$. Using a normal distribution for $[O/H]$, we then determine $\Psi$, the fraction of gas at $\Zc$.
Fig.~\ref{fig:stars} shows the low metallicity stellar mass density as a function of lookback time (and redshift) to its formation for various galaxy masses (taken at $\zm=0$). Stars with $\Zc = 0.01 \Zsun$ (dotted lines) form before $z\simeq 2$ and can be found in dwarf galaxies. Stars with $\Zc=0.3\Zsun$ (dashed lines) formed more recently (1$\leqslant \zform \leqslant$ 2) in Milky Way type galaxies. Stars with $\Zc=0.1\Zsun$ show a combination of both trends. When we neglect galaxy mergers, low metallicity star formation is reduced by at least an order of magnitude and limited to $\zform \geqslant 2$, with little dependence on $\mgal$. We find that most of the metal poor stars formed at low redshifts were brought in through mergers and were formed in galaxies smaller that their final host. In the next section we will determine the typical time between progenitor star formation and BBH merger in order to determine from which of these environments GW150914 most likely originated.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{lowZ_star_formation_evol_density_lookbackPP04_dt1e7}
\caption{Cosmic mass density (in $\msun $ Mpc$^{-3}$ Gyr$^{-1}$ $\msun^{-1}$) of stars at different metallicities (linestyles) in present-day galaxies with a total galaxy stellar mass $\mgal=10^{7-11} \msun$ (color as labeled), as a function of lookback time (redshift) to when the stars actually formed. \label{fig:stars}}
\end{figure}
\section{Time delays for massive BBH mergers}\label{sec:mergers}
To link the SFR, progenitor metallicity, and host mass evolution discussed above with BBH mergers that are detectable by LIGO, we compute a set of BPS models. Many phases in the evolution of binary stars remain poorly understood and previous BPS studies have shown that this results in large uncertainties in the BBH merger rate \citep[e.g.][]{Lipunov:1997,Sipior:2002,Dominik:2013}. Since this work focuses on host galaxies, and not binary evolution, we consider a simple, single set of standard assumptions consistent with observational constraints. We note that our models do not include the recently proposed massive overcontact binary BBH formation channel \citep{Marchant:2016,Mandel:2016}. We focus on field binaries and neglect BBHs that are dynamically formed in globular clusters \citep[e.g.][]{Downing:2011,Rodriguez:2015}, which would typically form at high redshifts and preferentially reside in more massive galaxies. Lacking observational constraints, we also neglect BBH stemming from Pop III stars \citep{Kinugawa:2014}, which are not likely candidates for GW 150914 \citep{Hartwig:2016} and which contribution to the gravitational wave background is still uncertain \citep{Dvorkin:2016}.
The BPS models are computed with the binary star evolution code \texttt{BSE} described in \citet{Hurley:2002}, which we have updated to improve the treatment of massive binaries. We use the weaker, metallicity dependent wind mass loss prescriptions from \citet{Belczynski:2010}. Updated remnant mass prescriptions are taken from \citet{Belczynski:2008}. BH birth kicks are modeled following \citet{Dominik:2013}. This results in the production of BBHs with component masses $\ga 25 \msun$ that are not disrupted by powerful natal kicks. The kicks are drawn from a Maxwellian distribution of width 265 km s$^{-1}$, reduced according to the amount of material that falls back after core collapse.
We have also updated the treatment of some mass transfer scenarios in \texttt{BSE}. We force systems that experience a common envelope phase while the mass donor is in the Hertzsprung gap to merge{\footnote{Stars in the Hertzsprung gap lack a steep density gradient between the core and envelope so there is no clear boundary to halt the inspiral of the companion and prevent a stellar merger \citep{Ivanova:2004,Belczynski:2007}.}. For stars that have evolved beyond the Hertzsprung gap, we take the common envelope efficiency to be unity, and compute the envelope binding energies with the \texttt{BSE}-default, evolutionary-state-dependent formulae. Furthermore, we allow stable Roche lobe overflow mass transfer to be non-conservative and assume that only half of the mass lost by the donor is accreted by the companion \citep{Dominik:2013}. With this updated version of \texttt{BSE} we are able to produce a reasonable estimate for the BBH merger delay time distribution given an initial population of binary stars.
We construct the delay time distribution from a Monte Carlo ensemble of $2.5\times10^6$ binaries. Primary masses range from $25-150 \msun$ and are drawn from the initial mass function (IMF) given by \citet{Kroupa:2001}. This allows for a wider mass distribution than the GW150914 event, which will be representative for future massive black hole binary detections. When we select a narrow mass range, set by the uncertainties on the GW150914 detection ($M_1=36_{-4}^{+5}\msun$ and $M_2=29_{-4}^{+4}\msun$), we find qualitatively very similar trends. The initial mass ratios and orbital periods are drawn from the distributions measured by \citet{Sana:2012}. Initial eccentricities are drawn from a thermal distribution $f(e) \propto 2e$. We evolve the same population of binaries for the 11 metallicity bins we consider.
Fig.~\ref{fig:delaytime} shows the number of BBH mergers per solar mass of stars formed that occur a time $t_\mathrm{delay}$ after the stellar binary forms. We only considered BBH mergers with total mass larger than $40 \msun$. Due to the metallicity dependence of the wind mass loss rates, binaries formed at $\Zc = 0.01 \Zsun$ produce the most massive BHs. Accordingly, these extremely low metallicity stars have the largest number of massive BBH mergers per unit stellar mass. However, at very late times higher metallicity stars account for a comparable number of mergers.
If we include BBH mergers of all masses (not shown here), $dN_\mathrm{m}/dt$ at each metallicity considered here approaches the standard $t^{-1}$ dependence \citep[e.g.][]{Dominik:2013,Belczynski:2016}. This agreement with previous work is encouraging because, for our purposes, it is most important to properly capture the {\em shape} of the delay time distributions. When we restrict our study to BBH mergers with total mass larger than $40 \msun$, only the $\Zc = 0.01 \Zsun$ delay time distribution $dN_\mathrm{m}/dt$ follows the $t^{-1}$ dependence, as is shown by the flat line for $N_{\mathrm{m}}(t)$. At higher metallicity, short mergers are absent because of larger stellar radii, which make many systems merge as stellar binaries before producing a BBH. On top of that, some binaries contract less during the common envelope phase, because of the lower envelope binding energy, resulting in BBHs that merge at later times. Except for the very low metallicity progenitors, we do not expect mergers from recently formed stars.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{delaytime_hist_loglog}
\caption{Number of massive BBH mergers per solar mass of star formation $N_{\mathrm{m}}$ as a function of time since formation for a stellar population with a Kroupa IMF and BBH mass $> 40\;M_\odot$. The upper limit in $t_\mathrm{delay}=t_{\mathrm{m}}-\tform$ is the Hubble time. For massive BBH mergers, only the $0.01 \Zsun$ population follows the standard $dN_\mathrm{m}/dt \propto t^{-1}$ evolution, shown with a red line. } \label{fig:delaytime}
\end{figure}
\section{Formation of BBH merger candidates }\label{sec:combine}
We now combine the number of low metallicity stars formed in different galaxies at different epochs with the number of mergers after a certain delay time for different progenitor metallicities (see Eq.~\ref{eqn:NofMgal}). We assume a binary fraction of $0.7$ \citep{Sana:2012}. Fig.~\ref{fig:mergers} shows the merger rates as a function of host galaxy mass, progenitor formation time and metallicity. The distribution is bimodal with early formation of $\Zc\gtrsim 0.1 \Zsun$ progenitors now present in massive galaxies and lower metallicity progenitors forming later in dwarf galaxies. The latter have limited star formation but are numerous and have a low metallicity. The contribution of dwarf galaxies is sensitive to the extrapolation of the low-mass galaxy SMF below observational completeness but the relatively flat galaxy mass distribution is robust to those uncertainties.
Integrated over all galaxy masses, the formation time of the progenitors is a rather flat distribution over the last 8 Gyr. We do not recover the strongly bimodal birth time distribution from \citet{Belczynski:2016} because of our more accurate treatment of the star forming gas metallicity and star formation. Most of the progenitors form around $\Zc\simeq 0.1\Zsun$. Many stars form at higher $\Zc$, but the number of mergers per unit solar mass is drastically reduced. At lower progenitor metallicity, more systems merge, but the amount of stars formed is low. If we were to include recently proposed fast merger channels \citep{Mandel:2016,Marchant:2016}, the distribution of host galaxies and formation times would be very similar to the distribution of low metallicity stars, with a possible contribution from low redshift galaxies.
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{Final_lMgal_Zcut_dt1e7_2dplot.png}
\includegraphics[width=\columnwidth]{Final_lMgal_zform_dt1e7_2dplot.png}
\caption{Comoving merger rate (Gpc$^{-3}$ yr$^{-1}$) as a function of present day galaxy mass and metallicity (left) and lookback time to the formation of the progenitor (right). The distribution has been integrated over formation time (left) and progenitor metallicity (right). Side panels show the integrated rates for our total sample (solid lines) as well as the restricted GW150614-like sample (dotted lines).}
\label{fig:mergers}
\end{figure*}
The total merger rate we find is $\mathcal{R}=850$ Gpc$^{-3}$ yr$^{-1}$ for our total mass sample and $\mathcal{R}=150$ Gpc$^{-3}$ yr$^{-1}$ when we restrict ourselves to the exact masses observed in GW150914. After the first observing run, the LIGO estimate of the merger rate of GW150914-like black holes is 2-53 Gpc$^{-3}$ yr$^{-1}$ \citep{LIGO:2016_rate}. While our model is based on standard assumptions for galaxy evolution and massive binary evolution, the total predicted merger rate overestimates the observed rate only by a factor 3. Choosing a metallicity calibration that predicts a lower MZR, a lower binary fraction and/or higher common envelope binding energy will naturally decrease these numbers. As we focus on formation conditions rather than absolute rates, we choose not to fine tune our model.
\section{Discussion and Conclusion}\label{sec:discussion}
In this paper, we compute when and where GW150914 most likely formed. Using only the strong constraint on the progenitor's metallicity and combining a state of the art binary population synthesis model with a complete cosmological description for the evolution of low metallicity gas, we find that GW150914 likely formed in a massive galaxy at $1\leqslant\zform\leqslant2$, but later formation in a dwarf galaxy is also possible. In fact the distribution of BBH merger progenitor formation times is remarkably flat for $\tform \simeq 1-10$ Gyr, and differs from the strongly bimodal distribution from \citet{Belczynski:2016}. Their computation is based on the metallicity evolution of DLAs, which ignores the crucial mass dependence of the metallicity. Our model also includes galactic mergers, which allow BBH progenitors formed in dwarf galaxies to end up in massive systems at $\tm$ and strongly increase the amount of mergers in the latter. Still, we find a large contribution of mergers in dwarf galaxies, whic is radically different from the distribution of present day stars, supernovae and BH which are strongly concentrated around $\mgal\simeq 10^{11}\msun$. This work presents the first determination of the formation conditions for the massive BBH mergers we currently observe. Without fine tuning, the total merger rate we predict is compatible with the LIGO detection rate.
Our work assumes that the only environmental impact on stellar evolution is progenitor metallicity, allowing us to decouple galactic evolution and stellar evolution, including multiplicity and the initial mass function. As such, the large uncertainties in massive stellar evolution only affect our absolute merger rate, but not its dependence on galaxy mass and formation time. Unless the metallicity dependence of stellar evolution were to be drastically revised, our model can be easily rescaled for different models of massive stellar evolution.
Uncertainties also affect our model for galaxy evolution, especially in small galaxies at high redshifts where star formation rates and particularly metallicity are very hard to determine observationally. We have assumed dwarf galaxies form the same amount of massive binaries per unit solar mass than larger galaxies, neglecting the fact that they may not host large enough molecular clouds to do so. As our understanding of high redshift star formation and stellar evolution improves with data from the James Webb Space Telescope (JWST) and the Wide Field Infrared Survey Telescope (WFIRST), our method will become a valuable tool to understand BBH mergers.
As Advanced LIGO and Advanced $Virgo$ reach their design sensitivity \citep{LIGO:2016_detector}, they will detect hundreds of BBH mergers, up to $\zm\lesssim 1$. BBHs merging at these redshifts formed during the peak of cosmic star formation, with a rather flat distribution of galaxy mass. In this context, this will provide strong tests of our models and the otherwise elusive nature of high redshift star formation and/or the metallicities of high-redshift or faint galaxies. Our method can further be combined with galaxy catalogs to predict typical distance distributions and sky localizations for future detections.
\section*{Acknowledgments}
Support for AL and PFH was provided by an Alfred P. Sloan Research Fellowship, NASA ATP Grant NNX14AH35G, and NSF Collaborative Research Grant \#1411920 and CAREER grant \#1455342. Support for SGK was provided by NASA through Einstein Postdoctoral Fellowship grant \#PF5-160136 awarded by the Chandra X-ray Center, which is operated by the Smithsonian Astrophysical Observatory for NASA under contract NAS8-03060. DC was supported through the Walter Burke Institute for Theoretical Physics and the Sherman Fairchild Foundation and the Caltech Department of Astronomy. The authors thank Chris Pankow, Evan Kirby, Xiangcheng Ma, Fangzhou Jiang, and Peter Behroozi for very helpful and stimulating discussions. We thank the referee for a constructive report that improved and clarified the manuscript.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{sec:intro}
The High Luminosity Large Hadron Collider (HL-LHC) machine will induce higher background radiations compared to the current operating conditions. It is important to study the performance and stability of the currently installed and future detectors in a high radiation environment. Focused on these requirements, the CERN Engineering- (EN) and Physics- Department (PH) made a joint project, the Gamma Irradiation Facility (GIF++) \cite{gif}. GIF++ is the new CERN irradiation facility located in the North Area of the CERN Super Proton Synchrotron (SPS). It is a unique place where high energy ($\sim$100 GeV) charged particles (mainly muons) are combined with a high flux of gamma radiation (662 keV) produced by 13.9 TBq $^{137}$Cs source \cite{atlas-gif}. An attenuator system is installed to vary the gamma flux on the two sides of the source independently. A schematic overview of the GIF++ is shown in figure \ref{gif}.\\
The Compact Muon solenoid (CMS) is one of the two general-purpose detectors at the LHC \cite{cms-lhc}, which uses Resistive Plate Chambers (RPCs) along with other detectors \cite{muon-sys} for the muon detection. In April 2015, the RPC detectors were installed at the GIF++ to study the performance at a radiation dose equivalent at 3000 fb$^{-1}$ of the CMS. A dedicated control system has been built to control these detectors and archive the relevant parameters using the WinCC-OA (PVSS) Supervisory Control And Data Acquisition (SCADA) system \cite{wincc-oa}. The system controls high voltage and low voltage supplies and monitors temperature, pressure and humidity of both the RPC gas and the environment. The source status and attenuator values are accessed through the Data Interchange Protocol (DIP), published centrally by the Engineering Department. The RPC gas supply is controlled and monitored by an external WinCC-OA project, that shares relevant parameters with this project. All relevant parameters are archived in a Structured Query Language (SQL) database (DB) for offline analysis.\\
One of the features of the GIF++ RPC DCS system is accessing the source status and attenuator values. Based on this information, RPC performance parameters (efficiency, working point, cluster size and resistivity) are measured. To retrieve the data from the database, a specific algorithm has been developed to synchronize the detector parameters (current and voltage) with the external parameters (temperature, pressure and humidity). It enables to monitor precisely the effect of external parameters on the detector.
\begin{figure}[H]
\centering
\hspace{-0.5cm}
\includegraphics[scale=0.35]{GIF.png}
\caption{An overview of the GIF++.}
\label{gif}
\end{figure}
\section{Motivation}
At HL-LHC, RPCs and associated electronics will operate at much higher luminosity ($\sim$ 5 $\times$ 10$^{34}$ cm$^{-2}$ s$^{-1}$) while accumulating large radiation dose expected from more than 3000 fb$^{-1}$. A precise understanding of the ageing of detector's material and currently installed electronics is needed in intensively high radiation.
The GIF++ gives the same conditions which will be present at CMS during HL-LHC. The RPCs are irradiated with photons, produced by a $^{137}$Cs radiation source of 13.9 TBq. At the same time, a high-momentum-particles beam extracted from the SPS, mainly consist of muons, is used to study the performance of the detectors by measuring and comparing their efficiency. The two main gas components of the RPC detector are C$_{2}$H$_{2}$F$_{4}$ and SF$_{6}$ which are going to be gradually banned due to greenhouse effects. An R\&D of different gases mixture is under study at GIF++ to replace the two gases by eco-friendly gases. When a good candidate for a new RPC gas mixture is identified, the study will be further carried out at GIF++ to include long-term assessment of the radiation tolerance of the chambers operated with the new gas mixture. Further details can be found in \cite{phase2}.
\section{WinCC-OA}
Large experiments at CERN use WinCC-OA as a tool to develop control systems. It allows for the description of a device in terms of a data point, with data point elements representing its parameters. These can then be addressed directly to write to and read from the corresponding device. Parameters of interest can then be stored by WinCC-OA in an internal database for offline analyses. WinCC-OA provides the facility to build a graphical user interface (GUI).
\section{The CMS RPC DCS project at GIF++}
The CMS RPC DCS at GIF++ has been developed by using WinCC-OA 3.11 and extended with the standard Joint Control Project (JCOP) framework \cite{jcop}. The JCOP framework provides extra functions such as standardized Finite State Machine (FSM), the additional Graphical User Interface (GUI), the alarm handlers and the ORACLE database interface \cite{g-polese}. The high voltage and low voltage power supplies used in the RPC GIF++ setup consist of CAEN SY1527 mainframe modules as well as CAEN EASY modules, with additional ADC modules used to read out gas and environmental sensors (pressure, temperature and humidity). The project has access to the hardware registered through an Object Linking and Embedding (OLE) for Process Control (OPC) server provided by CAEN using the OPC protocol \cite{opc}. The project controls the high voltage and low voltage system through the OPC protocol. The environmental and gas sensors ( for pressure, temperature and humidity) are also readout via the OPC protocol. The source status and attenuator values are available centrally via the Data Interchange Protocol (DIP). The project has been designed as a distributed one in order to be able to communicate with other projects and to read valuable information. Communication has been established with the central GIF++ DCS, such that the information from the gas system, like flow rates are readable.\\
The Finite State machine (FSM) hierarchy of the project is based on the naming convention of the trolley, where the detectors are installed. Each trolley has six sections and each section accommodates one detector. Currently three CMS RPCs trolleys are installed in the GIF++. Trolley 1 (RPC Consolidation) is equipped with spare RPCs, trolley 2 with small glass RPCs and trolley 3 with prototypes of improved RPCs. Detailed information of the trolleys are given here \cite{salvador}.
A schematic overview of the DCS project is shown in figure \ref{DCS_sys}.
\bigbreak \bigbreak \bigbreak \bigbreak
\begin{figure}[H]
\centering
\hspace{-0.5cm}
\includegraphics[scale=0.4,trim=60 30 60 30,clip]{DCS_sys.png}\\
\caption{DCS project overview.}
\label{DCS_sys}
\end{figure}
\subsection{High and Low Voltage System}
The high and low voltage system is controlled and monitored through the CAEN OPC server. Each gap of a chamber is independently connected to a single high voltage channel which improves the granularity of control. The RPC front-end electronics requires digital and analog power supplies \cite{feb}. Each low voltage line has been shared between two front-end-boards (FEBs) for digital as well as for analog. The low voltage boards are installed in the CAEN main frame and controlled by DCS.
\subsection{Environmental and Gas Parameters}
The performance of RPCs strongly depends on the temperature and pressure of the environment. Hence, it is important to measure the environmental parameters (temperature, pressure and humidity) at different locations. The applied voltage is corrected for the environmental temperature and pressure in order to include its effects. This procedure is described in detail in \cite{env-rpc}.
Figure \ref{scan_temp}a gives a plot for environmental temperature, pressure and humidity. The environmental and gas sensors (temperature, pressure and humidity) are readable through ADC (analog-to-digital converter) board which is installed in the EASY crate. The JCOP framework gives the opportunity to convert online the ADC counts into physical values. The trending feature provides a comparison among different sensors located at different positions.
\subsection{High Voltage Scan and Stability Test}
The project has been designed for R\&D of detectors, hence, should be able to perform high voltage scanning or stability tests. For high voltage scanning, a separate branch has been incorporated in the Finite State Machine (FSM) tree, where the user operates each detector independently. The stability test runs for long time. Based on the requirements, a dedicated manager is used to apply the stability script and restart it automatically. A high voltage scanning plot for one of the CMS RPC chambers (RE3) at GIF++ is shown in figure \ref{scan_temp}b.
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\hspace{-0.5cm}
\includegraphics[scale=0.32,trim=50 70 20 90,clip]{tprh.png}
& \hspace{-0.60cm} \includegraphics[scale=0.334,trim=50 100 45 80,clip]{HV_Scan2.png}\\
($\mathbf{a}$)\qquad&($\mathbf{b}$)\qquad\\
\end{tabular}
\caption{(a): Environmental temperature ($^o$C), pressure (mb) and humidity (\%). Time is on x-axis while red, blue and green lines show the values of pressure, temperature and humidity respectively on y-axis. Pressure and temperature values used for operating voltage correction of RPCs. (b): A high voltage scan for CMS RE3 chamber. The x-axis shows time while the y-axis shows the voltage (V) and current ($\mu$A) values. The red, blue and green lines are voltages while cyan, brown and orange lines are the corresponding current values for bottom, top narrow and top wide gaps respectively.}\label{scan_temp}
\end{figure}
\subsection{FSM and GUI}
The JCOP framework provides FSM toolkits in WinCC-OA based on State Machine Interface (SMI++). It offers an easy, robust and safe way to control the full detector through the definition of a finite number of states, transitions and actions (ON, OFF, STANDBY, Ramping Up, Ramping Down). All the DCS hierarchy nodes are implemented through the FSM mechanism, shown in figure \ref{gui}.\\
WinCC-OA provides a user friendly Graphic User Interface (GUI) panel- an intuitive tool to control, monitor and operate the detectors in safe mode. It provides flexibility to combine text, graphical objects and synoptic diagrams. GUI panels can be used to see the online behavior of the detector in the form of plots, tables and histograms.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.61,trim=0 175 10 80,clip]{GUI2.png}\\
\caption{FSM main tree and high voltage scan panel using GUI.}
\label{gui}
\end{figure}
\subsection{DataBase}
To study the behavior of the detector over time and to make offline data analysis, it is necessary to store all the important parameters in a database. WinCC-OA's internal SQL database is used in this project. The stored data can be extracted for the offline analysis using a GUI.
\section{Detector performance study}
\subsection{RPCs setup at GIF++ and Method for efficiency study}\label{eff_calc}
Four RPC chambers (two RE2 and two RE4 \cite{feb}) are placed parallel to each other in vertical position. Several high voltage scans were performed using different radiation levels (starting from the absence of radiation source), to define the optimal operating voltage of each chamber, called working point (WP). The efficiency (E) for different background radiation levels is calculated using the formula:\\
\centerline {E = N$_{RPC}$/N$_{TRACK}$.}
The muon track is reconstructed using three RPC planes and extrapolated to the RPC under test, looking to the closest cluster (a strip or set of continuous strips). N$_{TRACK}$ corresponds to the number of muons passing through the reference three RPC detectors at the same time. N$_{RPC}$ corresponds to the number of fired clusters in the chamber under test. The dependency of the efficiency E with respect to the effective high voltage HV$_{eff}$ \cite{hv-eff} can be fitted using the sigmoidal curve described by the subsequent formula:\\
\centerline {$\langle$E$\rangle$ = E$_{max}$ / (1 + exp (-$\lambda$ (HV$_{eff}$ - HV$_{50\%}$))}
where E$_{max}$ is the maximum efficiency reached by chambers at HV$\rightarrow \infty$, $\lambda$ is proportional to the slope of sigmoid at flex point and HV$_{50\%}$ is the high voltage at which a chamber reaches 50\% of its maximum efficiency. Working point of a chamber is defined by the formula HV$_{wp}$ = HV$_{knee}$ + 150V, where HV$_{knee}$ is the voltage at which efficiency is 95\% of the maximal one.
\subsection{Efficiency Results}
In figure \ref{eff}a, the efficiency as a function of HV$_{eff}$ for different radiation levels is shown. The maximum efficiency decreases with the amount of radiation received by the detector. In figure \ref{eff}b, the maximum efficiency as a function of the gamma hits rate is presented for four RPC chambers. The RPCs were placed parallel to each other, RPC-1 being the closest to the source and RPC-4 being the most distant. The radiation doze depends on the distance between the RPCs and the gamma source, RPCs 3 and 4 received a smaller dose as compared to RPCs 1 and 2.
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\hspace{-0.5cm}
\includegraphics[scale=0.30,trim=0 13 0 0,clip]{eff.png}
& \hspace{-0.60cm} \includegraphics[scale=0.329,trim=0 0 0 0,clip]{rate_eff.png}\\
($\mathbf{a}$)\qquad&($\mathbf{b}$)\qquad\\
\end{tabular}
\caption{(a): Efficiency vs HV$_{eff}$ for different gamma attenuator factors. (b): Maximum efficiency vs gamma rate at HV$_{wp}$ for four RPCs.}\label{eff}
\end{figure}
\section{Conclusion}
A DCS project for CMS RPCs has successfully been implemented and tested in the CERN GIF++. Since June, 2015 the project is running in a stable state, operating the detectors and archiving the data. The hardware integrated in the project, fully controls high voltage scanning and stability tests. The environmental and gas sensors are included and used for Temperature/Pressure corrections. Gas flow-meters are read through central DCS at GIF++ and the data are used to study the behavior of different gases. All useful parameters are archived in the internal database for offline analysis. As the project is designed for detector R\&D studies, any new hardware can be added easily and safely.\\
The performance of the CMS RPC chambers at GIF++ has been studied and compared at different radiation levels. At a rate of 600 Hz-cm$^{-2}$, the Eff$_{max}$ of the chamber was 95\%.
\acknowledgments We wish to congratulate our colleagues in the CERN Engineering- (EN) and Physics- Department (PH) for successful operation of the GIF++. We thank the technical and administrative staff at CERN, other CMS institutes and RPC group. Many thanks to ATLAS colleague Marino Romano for his technical support to develop the project.
|
\section{Introduction}
It is well known that the simply connected 3-dimensional homogeneous Riemannian manifolds have the isometry groups of dimension $3$, $4$ and $6$. When the dimension of the isometry group is $6$, then we have a space form. When the dimension of the isometry group is $3$, then the manifold has the geometry of the Lie group $Sol_3$.
We denote $E(\kappa,\tau)$, $\kappa\neq 4\tau^2$, as the homogeneous $3$-manifolds whose isometry groups are of dimension $4$, which are fibrations over $2$-dimensional simply connected space forms $\Bbb{M}^2(\kappa)$ of constant curvature $\kappa$. In other words, there exists a Riemannian submersion $\Pi:E(\kappa,\tau)\rightarrow\Bbb{M}^2(\kappa)$, where the constant number $\tau$ is the bundle curvature. The fibers are geodesics and there exists a one-parameter family of translations along the fibers, generated by a unit Killing vector field $\xi$.
When $\tau=0$, we get a product manifold $\Bbb{M}^2(\kappa)\times \Bbb{R}$. When $\tau\neq0$, the manifolds are of three types: the Berger sphere $\Bbb{S}_{\kappa,\tau}^3$ ($\kappa>0$), the Heisenberg Group $Nil_3$ ($\kappa=0$) and $P\widetilde{SL_2}$
($\kappa<0$). The so called {\em Hopf cylinder} is defined as the preimage $\Pi^{-1}(\gamma)$ of a regular closed curve $\gamma$ in $\Bbb{M}^2(\kappa)$. When considering $\Bbb{S}_{\kappa,\tau}^3$ and $P\widetilde{SL_2}$ (the fibers are circles), the corresponding Hopf cylinder is also called the {\em Hopf torus}. For more details on $E(\kappa,\tau)$, we refer to \cite{Da}.
The surface with constant mean curvature $H$ is called the cmc $H$-surface. There are a lot of researches on the geometry of $E(\kappa,\tau)$, for instance (\cite{AbRosenb}, \cite{AbRosenb2}, \cite{BaFe}, \cite{brs}, [9-11], \cite{Hu}, [17-21]), among others. A study on the cmc $H$-surfaces in more general spaces can be found in \cite{MePe}.
In this paper we focus on $E(\kappa,\tau)$, $\tau\neq 0$, and we get a Simons' type formula for cmc $H$-surfaces which can be stated as follows. Notice that the method used in this paper is elementary.
\begin{thm}\label{mainthm}
Let $f:M\rightarrow E(\kappa,\tau)$ ($\tau\neq 0$) be an immersion of a compact cmc $H$-surface. Denote $\Phi:=A-HI$ and the angle function by $C:={\langle} \xi,N{\rangle}$, where $N$ is a unit normal vector field on $M$ and $A$ is the second fundamental form. Then it holds the Simons' type formula
\begin{equation} \int_{M}\Big\{|\Phi|^4-[2(H^2+\tau^2)+(\kappa-4\tau^2)(5C^2-1)]|\Phi|^2+2(\kappa-4\tau^2)(H^2+\tau^2)(3C^2-1)\Big\}d\sigma \geq0, \label{Simon}\end{equation}
where the equality holds if and only of $f$ is of parallel second fundamental form, and in this case $M$ is a Hopf cylinder with $|A|^2=2(2H^2+\tau^2)$.
\end{thm}
\begin{rem} It is well known that the compact minimal immersed surface $M$ in the unit sphere $\Bbb{S}^3$ satisfies the Simons' integral inequality \[\int_M|A|^2(|A|^2-2)d\sigma\geq 0,\]
where the equality holds if and only of $M$ is the great $2$-sphere or the Clifford torus (\cite{CCK}).
\end{rem}
\begin{rem} For the minimal surface in $E(\kappa,\tau)$ ($\tau\neq 0$), $H=0$, the corresponding Simons' type formula was recently given in \cite{Hu}. \end{rem}
\section{Preliminaries}
For an immersed surface $M$ in $E(\kappa,\tau)$, define the {\it contact angle} $\beta(p)$ at $p\in M$ by the angle between the vector $\xi(p)$ and the tangent space $T_pM$. By definition, the contact function $C=\sin\beta$. Since the distribution ${\langle}\xi{\rangle}^\perp$ in $E(\kappa,\tau)$ is not integrable (see \cite{Aebischer}), we consider the set $\mathcal{W}=\{p\in M: \beta(p)\neq \pm\frac{\pi}{2}\}\subset M$, whose complementary $\mathcal{W}^{C}=\{p\in M: \beta(p)= \pm\frac{\pi}{2}\}$ has empty interior by Frobenius Theorem. In this paper we will work by making use of the method of moving frames on $\mathcal{W}$ and then extend the formulas to the whole surface by the continuity.
\subsection{Adapted frame}
In this section we introduce the method of choosing an appropriate orthonomal frame $\{e_1,e_2,e_3\}$ on $\mathcal{W}\subset M$ to study its geometry. It is well known that $E(\kappa,\tau)$ admits an orthonormal frame $\{f_1,f_2,f_3\}$ with $f_3=\xi$, satisfying (see \cite{Da}) \[[f_1,f_2]=-2\tau f_3,\ \ \ [f_2,f_3]=-\kappa/(2\tau) f_1\ \ \ \mbox{and}\ \ \ [f_3,f_1]=-\kappa/(2\tau) f_2.\label{framelee}\]
It is straightforward to verify that the coframe $\{w^1,w^2,w^3\}$ of $\{f_1,f_2,f_3\}$ satisfies
\[ dw^1=\kappa/(2\tau) w^2{\wedge} w^3,\ \ \ dw^2=\kappa/(2\tau) w^3{\wedge} w^1\ \ \ \mbox{and}\ \ \ dw^3=2\tau w^1{\wedge} w^2,\label{BBBB}\] and the Levi-Civita connection $1$-forms are given by
\[ w^1_2=(\kappa/\tau-\tau)w^3,\ \ \ w^1_3=-\tau w^2\ \ \ \mbox{and}\ \ \ w^2_3=\tau w^1.\label{connection}\]
An interesting observation (see Lemma 1 in \cite{CuiGomes}) is that, if one rotates $f_1$ and $f_2$ but leaves $f_3$ fixed, the relation of $w^1_3$ and $w^2_3$ will not change ($w^1_2$ may change). Thus, the connection 1-forms can be summarized as
\begin{equation}\label{w_i^j}
w^1_2,\ \ \ w^1_3=-\tau w^2\ \ \ \mbox{and}\ \ \ w^2_3=\tau w^1.
\end{equation}
If one studies a surface $M$ isometrically immersed in $E(\kappa,\tau)$ one can firstly choose $\{f_1,f_2,f_3\}$ by rotating ${f_1,f_2}$ and leaving $f_3$ fixed at $p\in \mathcal{W}$ such that $f_1\in T_pM$, and then choose the local orthonormal frame $\{e_1,e_2,e_3\}$, called the {\em adapted frame}, where
\begin{equation} \label{framesurface}
e_1=f_1, \ \ \ e_2=\sin\beta f_2+\cos\beta f_3 \ \ \ \mbox{and}\ \ \ e_3=-\cos\beta f_2+\sin\beta f_3,
\end{equation} so that $e_1$ and $e_2$ are tangent to the surface. This frame was firstly introduced for $\Bbb{S}^3$ in \cite{Gomes} and for $E(\kappa,\tau)$ in \cite{CuiGomes}.
\subsection{Structure equations} In this section we study the structure equations under the adapted frame in the open set $\mathcal{W}\subset M$ in $E(\kappa,\tau)$. The corresponding coframe $\{\theta^1,\theta^2,\theta^3\}$ is given by
\begin{equation} \label{coframe}
\theta^1=w^1, \ \ \ \theta^2=\sin\beta w^2+\cos\beta w^3\ \ \ \mbox{and}\ \ \ \theta^3=-\cos\beta w^2+\sin\beta w^3.
\end{equation}
The structure equations of $E(\kappa,\tau)$ are given by
\begin{equation*}
\bar{\Omega}^A_B=d\theta^A_B+\theta^A_C{\wedge}\theta^C_B\ \ \ \mbox{and}\ \ \ d\theta^A =-\theta^A_B{\wedge}\theta^B, \ \ \mbox{with}\ \ \theta^A_B+\theta^B_A=0.
\end{equation*}
Here and from now on, we assume $1\leq A,B,\ldots\leq 3$, $1\leq i,j,\ldots\leq 2$ and we will use the Einstein summation convention.
Since, when restricted to $M$, $\theta^3=0$ we get that $\theta^3_i{\wedge}\theta^i=0$. The Cartan's lemma thus implies $\theta^3_i=h_{ij}\theta^j$ with smooth functions $h_{ij}=h_{ji}$.
We now immediately deduce the structure equations of $M$
\begin{equation}\label{gauss11}
d\theta^i = - \theta^i_j{\wedge}\theta^j, \ \ \mbox{with}\ \ \theta^i_j+\theta^j_i=0\ \ \mbox{and}\ \
\Omega^i_j := d\theta^i_j+ \theta^i_k{\wedge}\theta^k_j=\bar{\Omega}^i_j-\theta^i_3{\wedge}\theta^3_j,
\end{equation}
where the last equation is called the {\it Gauss equation}. For $A=3$ and $B=i$ one gets the {\it Codazzi equations}:
\begin{equation} \label{codazzi11}
\bar{\Omega}^3_i=d\theta^3_i+ \theta^3_k{\wedge}\theta^k_i=d(h_{ij}\theta^j)+ h_{kj}\theta^j{\wedge}\theta^k_i=h_{ij|k}\theta^k{\wedge}\theta^j,
\end{equation}
where
\begin{equation} \label{KKKK}
h_{ij|k}\theta^k:=dh_{ij}- h_{kj}\theta^k_i- h_{ik}\theta^k_j.
\end{equation}
By using the classical notation $\bar{\Omega}^A_B={1\over2}\bar{R}^A_{BCD}\theta^C{\wedge}\theta^D\,$ and $\,\Omega^i_j={1\over2}R^i_{jkl}\theta^k{\wedge}\theta^l$
we get that the Gauss equation \eqref{gauss11} and Codazzi equation \eqref{codazzi11} are respectively given by
\begin{equation} \label{Gauss12}
R^i_{jkl}=\bar{R}^i_{jkl}+h_{ik}h_{jl}-h_{ij}h_{kl},\nonumber
\end{equation}
\begin{equation} \label{Codazzi12}
h_{ij|k}-h_{ik|j}=-\bar{R}^3_{ijk}.\nonumber
\end{equation}
Throughout this paper, for simplicity of notation we denote $\beta_i=e_i(\beta)$, $\beta_{ij}=e_je_i(\beta)$ and $\beta_{ijk}=e_ke_je_i(\beta)$. The following lemma was actually proved in \cite{CuiGomes}, that gives us the Levi-Civita connection on the surface $M$ and the second fundamental form under the adapted frame. For the readers' convenience we give a proof here.
\begin{lem} (\cite{CuiGomes})\label{second} Let $\nabla$ be the Levi-Civita connection on $M$. Then \[\nabla_{e_1}e_1=-(2H+\beta_2)\tan\beta e_2,\ \ \ \ \ \ \ \ \nabla_{e_2}e_2=-(2\tau+\beta_1)\tan\beta e_1,\]
\begin{equation} \nabla_{e_2}e_1=(\beta_1+2\tau)\tan\beta e_2\ \ \ \ \mbox{and}\ \ \ \ \nabla_{e_1}e_2=(2H+\beta_2)\tan\beta e_1,\label{NB}\end{equation}
where $H:=(h_{11}+h_{22})/2$ is the mean curvature.
Moreover, the coefficients of the second fundamental form of $M$ are given by
\begin{equation} \label{sec}
h_{11}=2H+\beta_2,\ \ \ h_{12}=h_{21}=-\tau-\beta_1 \ \ \ \mbox{and}\ \ \ h_{22}=-\beta_2.
\end{equation}
Also, we have the relation
\begin{equation} \label{mean}
\cos\beta w^1_2(e_1)=2H+\beta_2.
\end{equation}
\end{lem}
\begin{proof}
Since $\theta^3=0$ on $M$ we have
\[\cos\beta w^2=\sin\beta w^3,\]
which together with the expression \eqref{coframe} for $\theta^2$ gives us
\begin{equation} \label{cof2}
\theta^1=w^1,\ \ \ \cos\beta\theta^2= w^3\ \ \ \mbox{and}\ \ \ \sin\beta\theta^2=w^2.
\end{equation}
So, from equations \eqref{w_i^j} and \eqref{coframe} we have
\begin{eqnarray*}
d\theta^1 &=& dw^1=-w^1_2{\wedge} w^2-w^1_3{\wedge} w^3=-\sin\beta w^1_2{\wedge} \theta^2,\\
d\theta^2 &=& \sin\beta dw^2+\cos\beta dw^3 = \sin\beta(w_2^1-\tau\cos\beta\theta^2){\wedge} \theta^1,\\
d\theta^3 &=& d\beta{\wedge} \theta^2-\cos\beta dw^2+\sin\beta dw^3 = d\beta{\wedge} \theta^2+\cos\beta w_1^2{\wedge} \theta^1+\tau(1+\sin^2\beta)\theta^1{\wedge} \theta^2.
\end{eqnarray*}
We thus obtain
\begin{equation} \theta^1_2=\sin\beta(w_2^1-\tau\cos\beta\theta^2),\ \ \ \ \cos\beta w^1_2(e_2)=-\beta_1 -\tau(1 + \sin^2\beta), \label{theta12} \end{equation}
and thus
\[\nabla_{e_1}e_1=\theta^2_1(e_1)e_2=-\sin\beta w^1_2(e_1)e_2,\ \ \ \ \ \ \ \ \ \ \nabla_{e_2}e_2=\theta^1_2(e_2)e_1=-(2\tau+\beta_1)\tan\beta e_1,\]
\begin{equation} \nabla_{e_2}e_1=\theta^2_1(e_2)e_2=(2\tau+\beta_1)\tan\beta e_2\ \ \ \ \mbox{and}\ \ \ \ \nabla_{e_1}e_2=\theta^1_2(e_1)e_1=\sin\beta w^1_2(e_1) e_1.\label{NNNB}\end{equation}
Now, using equations \eqref{w_i^j}, \eqref{framesurface} and \eqref{cof2} we compute
\begin{eqnarray*}
\bar{D}e_3 &=& \bar{D}(-\cos\beta f_2+\sin\beta f_3)\\
&=&\sin\beta d\beta f_2-\cos\beta(w_2^1f_1+w_2^3f_3)+\cos\beta d\beta f_3+\sin\beta(w_3^1f_1+w_3^2f_2)\\
&=&(-\cos\beta w_2^1-\tau\sin^2\beta\theta^2)e_1 + (d\beta+\tau\theta^1)e_2,
\end{eqnarray*}
where $\bar{D}$ is the Levi-Civita connection on $E(\kappa,\tau)$. On the other hand, $\bar{D}e_3=\theta_3^1e_1+\theta^2_3e_2$, whence
\[ \label{connectoin1}
\theta^1_3=-\cos\beta w^1_2-\tau \sin^2\beta \theta^2\ \ \ \mbox{and}\ \ \ \theta^2_3=d\beta+\tau\theta^1.
\]
Hence, from the second equation of \eqref{theta12} and $\theta^3_i=h_{ij}\theta^j$, we get
\begin{equation} \label{secd}
h_{11}=\cos\beta w^1_2(e_1),\ \ \ h_{12}=\cos\beta w^1_2(e_2)+\tau\sin^2\beta = -\tau-\beta_1\ \ \mbox{and}\ \ h_{22} =-\beta_2.
\end{equation}
Since by the definition $2H=h_{11}+h_{22}$, we get (\ref{mean}) and then (\ref{NB}) and (\ref{sec}) follow from (\ref{mean}), (\ref{NNNB}) and (\ref{secd}) immediately.
\end{proof}
By using the formula of the Riemannian curvature tensor in \cite{Da}, one can obtain the Codazzi equations under the adapted frame. See the proof in \cite{CuiGomes}.
\begin{lem} (\cite{CuiGomes}) The Codazzi equations under the adapted frame are given by
\begin{equation}\beta_1w^1_2(e_1)\cos\beta=\beta_2(2\tau+\beta_1)\label{Codazzi1}\end{equation} and
\begin{eqnarray}\label{Codazzi2}
\nonumber0&=&\cos\beta e_2(w^1_2(e_1))+\sin\beta \cos\beta (w^1_2(e_1))^2+\beta_{11}+2(\tau+\beta_1)(2\tau+\beta_1)\tan\beta\\
&&+(\kappa-4\tau^2)\sin\beta \cos\beta.
\end{eqnarray}
\end{lem}
At the end of this section, we shall derive some identities which will be useful in the next section.
From (\ref{mean}) and (\ref{Codazzi1}), we have $\beta_1(2H+\beta_2)=\beta_2(2\tau+\beta_1)$, that is\begin{equation} \tau\beta_2=H\beta_1.\label{beta12}\end{equation}
According to Lemma \ref{second}, we observe that \begin{eqnarray} \beta_{12}&=&e_2e_1\beta=[e_2,e_1]\beta+\beta_{21}=\nabla_{e_2}e_1\beta-\nabla_{e_1}e_2\beta+\beta_{21}\nonumber\\&=&(2\tau+\beta_1)\tan\beta\beta_2-(2H+\beta_2)\tan\beta\beta_1+\beta_{21}=\beta_{21}.\nonumber\label{beta111A}\end{eqnarray}
By (\ref{beta12}), if $H$ is a constant, we have $\tau\beta_{21}=H\beta_{11}$ and $\beta_1\beta_{21}=\beta_2\beta_{11}$. Then a similar way shows that \[ \beta_{112}=e_2e_1(\beta_1)=[e_2,e_1]\beta_1+\beta_{121}=\nabla_{e_2}e_1\beta_1-\nabla_{e_1}e_2\beta_1+\beta_{121}=\beta_{121}.\label{beta112}\]
Thus by (\ref{beta12}) again, for the later use, we deduce that \begin{equation} \beta_{12}=\beta_{21}=(H/ \tau)\beta_{11},\ \ \beta_{22}=(H/ \tau)\beta_{12}=(H/ \tau)\beta_{21}=(H/ \tau)^2\beta_{11}\label{BBBd}\end{equation}
and \begin{equation} \beta_{122}=(H/ \tau)\beta_{112}=(H/ \tau)\beta_{121}=(H/ \tau)^2\beta_{111}.\label{beta113}\end{equation}
\section{Simons' type formula for cmc surafces}
In this section we consider the tensor $\Phi:=A-HI$ with constant $H$. By Lemma \ref{second} and (\ref{beta12}), we have \begin{eqnarray} |\Phi|^2&=&|A|^2-2H^2=(2H+\beta_2)^2+2(\tau+\beta_1)^2+\beta_2^2-2H^2\nonumber\\
&=&2(\tau+\beta_1)^2+2(H+\beta_2)^2=2[1+({H/\tau})^2](\tau+\beta_1)^2.\label{Phi2}\end{eqnarray}
We now start to compute the Laplacian of $|\Phi|^2$. By using (\ref{beta12}), (\ref{BBBd}) and (\ref{beta113}), we compute that
\begin{eqnarray} \Delta |\Phi|^2&=&e_1e_1(|\Phi|^2)+e_2e_2(|\Phi|^2)-\nabla_{e_1}e_1 |\Phi|^2-\nabla_{e_2}e_2 |\Phi|^2\nonumber\\
&=&4[1+({H/\tau})^2][\beta_{11}^2+(\tau+\beta_1)\beta_{111}+\beta_{12}^2+(\tau+\beta_1)\beta_{122}\nonumber\\
&&+(2H+\beta_2)(\tau+\beta_1)\tan\beta\beta_{12}+(2\tau+\beta_1)(\tau+\beta_1)\tan\beta\beta_{11}]\nonumber\\
&=&4[1+({H/\tau})^2]^2[\beta_{11}^2+(\tau+\beta_1)\beta_{111}+(2\tau+\beta_1)(\tau+\beta_1)\tan\beta\beta_{11}].\label{DeltaPhi2}
\end{eqnarray}
Next, we shall express $\beta_{11}$ and $\beta_{111}$ to the expressions involving in $\beta_1$.
First, by Lemma \ref{second} we have $2H+\beta_2=\cos\beta w^1_2(e_1)$ and thus we get
\begin{equation*}\beta_{22}=-\tan\beta\beta_2(2H+\beta_2)+\cos\beta e_2(w^1_2(e_1)).
\end{equation*}
So, the Codazzi equation \eqref{Codazzi2} becomes
\begin{eqnarray}
\beta_{11}+\beta_{22} &=& -\tan\beta[(2H+\beta_2)\beta_2+(2H+\beta_2)^2+2(\tau+\beta_1)(2\tau+\beta_1)]\nonumber\\
&&-(\kappa-4\tau^2)\sin\beta \cos\beta.\label{c2-H}
\end{eqnarray}
Again, by using (\ref{beta12}) and (\ref{beta113}), the equation (\ref{c2-H}) becomes \begin{equation} [1+({H/\tau})^2] \beta_{11}=-2[1+({H/\tau})^2](\tau+\beta_1)(2\tau+\beta_1)\tan\beta-(\kappa-4\tau^2)\sin\beta\cos\beta,\label{beta11}\end{equation}
and taking the derivative by $e_1$ gives us \begin{eqnarray} [1+({H/\tau})^2] \beta_{111}&=&-2[1+({H/\tau})^2][(3\tau+2\beta_1)\beta_{11}\tan\beta+(\tau+\beta_1)(2\tau+\beta_1)\sec^2\beta\beta_1]\nonumber\\
&&-(\kappa-4\tau^2)\cos(2\beta)\beta_1\nonumber\\
&=&4[1+({H/\tau})^2](3\tau+2\beta_1)(\tau+\beta_1)(2\tau+\beta_1)\tan^2\beta+2(\kappa-4\tau^2)(3\tau+2\beta_1)\sin^2\beta\nonumber\\
&&-2[1+({H/\tau})^2](\tau+\beta_1)(2\tau+\beta_1)\sec^2\beta\beta_1-(\kappa-4\tau^2)\cos(2\beta)\beta_1.\label{beta111}\end{eqnarray}
Substituting (\ref{beta11}) and (\ref{beta111}) into (\ref{DeltaPhi2}), by a straightforward computation we have
\begin{eqnarray} \Delta |\Phi|^2&=&4[1+({H/\tau})^2]^2\Big\{\beta_{11}^2+2(\tau+\beta_1)^2(2\tau+\beta_1)[2(2\tau+\beta_1)\tan^2\beta-\beta_1]\Big\}\nonumber\\
&&+4[1+({H/\tau})^2](\kappa-4\tau^2)(\tau+\beta_1)[4(\tau+\beta_1)\sin^2\beta-\beta_1\cos^2\beta]. \label{DeltaPhi2F}\end{eqnarray}
\begin{lem}\label{AAA} \begin{equation} |\nabla A|^2=2[1+({H/\tau})^2]^2[\beta_{11}^2+4(\tau+\beta_1)^2(2\tau+\beta_1)^2\tan^2\beta].\label{Asquare}\end{equation}
\end{lem}
\begin{proof} By (\ref{KKKK}), (\ref{BBBd}) and Lemma \ref{second}, we compute that
\begin{eqnarray} h_{11|1}&=&e_1(2H+\beta_2)-h_{21}\theta^2_1(e_1)-h_{12}\theta^2_1(e_1)\nonumber\\
&=&\beta_{21}-2(\tau+\beta_1)(2H+\beta_2)\tan\beta\nonumber\\
&=&(H/\tau)[\beta_{11}-2(\tau+\beta_1)(2\tau+\beta_1)\tan\beta].\nonumber
\end{eqnarray}
A similar computation yields
\[h_{11|2}=(H/\tau)^2\beta_{11}+2(\tau+\beta_1)(2\tau+\beta_1)\tan\beta,\ \ \ \ h_{12|1}=-\beta_{11}-2(H/\tau)^2(\tau+\beta_1)(2\tau+\beta_1)\tan\beta\]
and the following identities hold true: \[h_{12|2}=-h_{11|1}=-h_{22|1}\ \ \ \mbox{and}\ \ \ h_{12|1}=-h_{22|2}.\]
Thus by definition, \[|\nabla A|^2=h_{11|1}^2+h_{11|2}^2+h_{22|1}^2+h_{22|2}^2+2(h_{12|1}^2+h_{12|2}^2)=2(2h_{11|1}^2+h_{11|2}^2+h_{12|1}^2),\]
and Eq. (\ref{Asquare}) follows by a straightforward computation.
\end{proof}
By (\ref{Phi2}), (\ref{DeltaPhi2F}), (\ref{Asquare}) and $|\nabla\Phi|^2=|\nabla A|^2$, we have
\begin{eqnarray} {1\over 2}\Delta |\Phi|^2
&=&|\nabla\Phi|^2-|\Phi|^2[|\Phi|^2-2(H^2+\tau^2)]\nonumber\\
&&+(\kappa-4\tau^2)\Big\{4|\Phi|^2\sin^2\beta-2[1+({H/\tau})^2](2\tau+\beta_1)\beta_1\cos^2\beta\Big\}\nonumber\\
&&+2(\kappa-4\tau^2)[1+({H/\tau})^2]\tau\beta_1\cos^2\beta\nonumber\\
&=&|\nabla\Phi|^2-|\Phi|^2[|\Phi|^2-2(H^2+\tau^2)]\nonumber\\
&&+(\kappa-4\tau^2)[4|\Phi|^2\sin^2\beta-(|\Phi|^2-2(H^2+\tau^2))\cos^2\beta]\nonumber\\
&&+2(\kappa-4\tau^2)[1+({H/\tau})^2]\tau\beta_1\cos^2\beta. \label{DeltaPhi2FF}
\end{eqnarray}
We now deal with the last term in (\ref{DeltaPhi2FF}). Inspired by the computations in \cite{Hu} and \cite{TU}, we prove the following lemma.
\begin{lem}\label{LLL} Let $T:=f_3-{\langle} f_3,N{\rangle} N$ be the projection of the vertical vector field $f_3=\xi$ on $\mathcal{W}\subset M$. Then
\begin{equation} {1\over2}\Delta|T|^2-div(\nabla_TT)=2\tau(\beta_1\cos^2\beta+2\tau\sin^2\beta). \label{CVC}\end{equation}\end{lem}
\begin{proof} By (\ref{framesurface}), we can express $T=f_3-\sin\beta e_3=\cos\beta e_2$. Then we compute
\[\nabla_TT=\cos\beta\nabla_{e_2}(\cos\beta e_2)=-\sin\beta\cos\beta [\beta_2e_2+(2\tau+\beta_1)e_1]=-(1/2)\sin(2\beta)(\nabla\beta+2\tau e_1).\]
Noticing that \[ {1\over2}\Delta|T|^2={1\over2}\Delta(\cos^2\beta)=-\cos(2\beta)|\nabla\beta|^2-\sin\beta\cos\beta\Delta\beta,\] we get
\begin{eqnarray} div(\nabla_TT)&=&{\langle} \nabla_{e_i}\nabla_TT,e_i{\rangle}=-\cos(2\beta)(|\nabla\beta|^2+2\tau \beta_1)-\sin\beta\cos\beta(\Delta\beta+2\tau{\langle} \nabla_{e_i}e_1,e_i{\rangle})\nonumber\\
&=&{1\over2}\Delta|T|^2-2\tau(\beta_1\cos^2\beta+2\tau\sin^2\beta),\nonumber\end{eqnarray}
which proves (\ref{CVC}).
\end{proof}
By (\ref{DeltaPhi2FF}), (\ref{CVC}) and noticing that the contact function is $C=\sin\beta$, we have
\begin{eqnarray} &&{1\over 2}\Delta |\Phi|^2-(\kappa-4\tau^2)[1+({H/\tau})^2]\Big({1\over2}\Delta|T|^2-div(\nabla_TT)\Big)\nonumber\\
&=&|\nabla\Phi|^2-|\Phi|^2[|\Phi|^2-2(H^2+\tau^2)]\nonumber\\
&&+(\kappa-4\tau^2)[4|\Phi|^2\sin^2\beta-(|\Phi|^2-2(H^2+\tau^2))\cos^2\beta]\nonumber\\
&&-4(\kappa-4\tau^2)(H^2+\tau^2)\sin^2\beta\nonumber\\
&=&|\nabla\Phi|^2-|\Phi|^2[|\Phi|^2-2(H^2+\tau^2)]\nonumber\\
&&+(\kappa-4\tau^2)[|\Phi|^2(5C^2-1)-2(H^2+\tau^2)(3C^2-1)].\label{DP}\end{eqnarray}
Although the above formula is derived on $\mathcal{W}\subset M$, it can be extended to the whole $M$ since the complimentary $\mathcal{W}^C$ has empty interior.
\begin{rem} The Berger sphere is the unit sphere $\Bbb{S}^3$ endowed with the metric
\[ {\langle} X,Y{\rangle}=\frac{4}{\kappa}\Big[{\langle} X,Y{\rangle}_{\Bbb{S}^3}+\Big(\frac{4\tau^2}{\kappa}-1\Big){\langle} X,\xi {\rangle}_{\Bbb{S}^3} {\langle} Y,\xi {\rangle}_{\Bbb{S}^3}\Big],\label{metric1}\]
where $\langle\, , \,\rangle_{\Bbb{S}^3}$ stands for the standard metric on $\Bbb{S}^{3}$, $\kappa>0$ and $\tau\neq0$. Although in this paper we assume $\kappa\neq4\tau^2$, the formula (\ref{DP}) still holds for $\kappa=4\tau^2$. By taking $\kappa=4\tau^2=4c>0$, the sphere is the Riemannian space form $\Bbb{S}^3(c)$ with constant sectional curvature $c$. In this case, (\ref{DP}) reduces to
\[{1\over 2}\Delta |\Phi|^2=|\nabla\Phi|^2-|\Phi|^2[|\Phi|^2-2(H^2+c)],\]
which is the well known formula in \cite{NoS} by observing that $tr\Phi^3=0$ for surfaces.
\end{rem}
{\bf Proof of Theorem \ref{mainthm}}:
The Simons' type formula (\ref{Simon}) is a direct consequence by integrating (\ref{DP}) on $M$, no matter whether $M$ is orientable or not. The equality holds if and only if $|\nabla A|=|\nabla \Phi|=0$, and this implies that $\sin\beta=0$ from (\ref{beta11}) and (\ref{Asquare}). Therefore, the vertical Killing vector field of the fibration $\Pi:E(\kappa,\tau)\rightarrow\Bbb{M}^2(\kappa)$ is tangent to $M$, so $\Pi(M)$ is a closed curve $\gamma$ in $S^2$ and then $M=\Pi^{-1}(\gamma)$ is the Hopf cylinder in $E(\kappa,\tau)$. In this case, $|A|^2=2(2H^2+\tau^2)$ follows from (\ref{Phi2}) immediately. \qed
\begin{rem} The authors in \cite{BeDi} proved that the Hopf cylinder is the only surface in $E(\kappa,\tau)$ ($\tau\neq 0$) with parallel second fundamental form. This fact can be used to prove the case of the equality in Theorem \ref{mainthm} instead of using Eqs. (\ref{beta11}) and (\ref{Asquare}).
\end{rem}
As an application, we shall use Theorem \ref{mainthm} to give a pinching result of the second fundamental form. Observe that in the case of $\kappa>4\tau^2$, the equation
\[x^2-[2(H^2+\tau^2)+(\kappa-4\tau^2)(5C^2-1)]x+2(\kappa-4\tau^2)(H^2+\tau^2)(3C^2-1)=0\]
has two distinct real solutions:
\[x_1=a(\kappa,\tau,H,C):=[2(H^2+\tau^2)+(\kappa-4\tau^2)(5C^2-1)-\sqrt{\rho}]/2\]
and \[x_2=b(\kappa,\tau,H,C):=[2(H^2+\tau^2)+(\kappa-4\tau^2)(5C^2-1)+\sqrt{\rho}]/2,\]
where \[\rho=4(H^2+\tau^2)^2+4(H^2+\tau^2)(\kappa-4\tau^2)(1-C^2)+(\kappa-4\tau^2)^2(5C^2-1)^2.\]
Obviously, it holds the relation $a(\kappa,\tau,H,C)\leq H^2+\tau^2<2(H^2+\tau^2)\leq b(\kappa,\tau,H,C) $. By Theorem \ref{mainthm} and the identity $|\Phi|^2=|A|^2-2H^2$, we get immediately
\begin{cor}\label{maincor}
Let $f:M\rightarrow E(\kappa,\tau)$ ($\kappa>4\tau^2$ and $\tau\neq 0$) be an immersion of a compact cmc $H$-surface. If $a(\kappa,\tau,H,C)+2H^2\leq |A|^2\leq b(\kappa,\tau,H,C)+2H^2$, then $M$ is a Hopf cylinder with $|A|^2=2(2H^2+\tau^2)$. In particular, if $3H^2+\tau^2\leq |A|^2\leq 2(2H^2+\tau^2)$, then $M$ is a Hopf cylinder with $|A|^2=2(2H^2+\tau^2)$.
\end{cor}
\begin{rem} When $H=0$, in the case of $\kappa>4\tau^2$, the authors in \cite{Hu} obtained that $M$ is a Clifford torus in the Berger sphere under the assumption that $a(\kappa,\tau,0,C)\leq |A|^2\leq b(\kappa,\tau,0,C)$, by observing the fact that the Clifford torus is the only minimal Hopf torus in the Berger sphere (see \cite{BaFe} or \cite{Pink}). We also mention that there exist cmc (even minimal) tori which are not Hopf tori in the Berger sphere (see \cite{To}).
\end{rem}
\begin{rem} Similar to the Remark 4.8 in \cite{Hu}, the case of $\kappa<4\tau^2$ for Corollary \ref{maincor} is still open.
\end{rem}
|
\section{Introduction}
Given a group $G$ and a set of involutions $S:=\{\rho_0, \ldots, \rho_{r-1}\}$ which generate $G$, such that
\[ \forall i,j \mathrm{\;with\;} |i-j|>1, \rho_i \mathrm{\;and\;} \rho_j \mathrm{\;commute\;(the\;\emph{string\;property}}),
\]
we call the pair $(G,S)$ a {\em string group generated by involutions} (or {\em sggi} for short).
We denote by $\Gamma_I$ the group generated by $\{\rho_i:i\in I\}$ for $I\subseteq\{0,\ldots,r-1\}$.
The pair $(G,S)$ satisfies the {\em intersection property} if for every $I,J \subseteq\{0,\ldots,r-1\}$,
$\Gamma_I\cap\Gamma_J=\Gamma_{I\cap J}$.
A sggi $\Gamma:= (G,S)$ that satisfies the intersection property is called a {\em string C-group} of {\em rank $|S|$}.
If $\Gamma:=(G,S)$ is a string C-group, we sometimes will abuse language and talk about the group $\Gamma$ and denote the {\em rank} of $G$ as the largest size of a set $S$ of involutions such that $\Gamma:=(G,S)$ is a string C-group.
For $i \in \{0, \ldots, r-1\}$, we denote by $\Gamma_i$ the group generated by all the elements of $S$ except $\rho_i$.
It is known that string C-groups are automorphism
groups of abstract regular polytopes and that, given an abstract regular polytope and a base flag of the polytope, one can construct a string C-group whose group $G$ is the automorphism group of the polytope~\cite[Section 2E]{ARP}. Hence the study of string C-groups has interests not only in group theory, but also in geometry.
Classifications of string C-groups from almost simple groups started with experimental work of Leemans and Vauthier~\cite{LVatlas} (see also~\cite{Halg, HHalg, LMalg, CLM2012} for more experimental results) and quickly led to the determination of the rank of a string C-group of Suzuki type~\cite{Leemans:2006}. A series of results then followed for the almost simple groups with socle $\mathrm{PSL}(2,q)$~\cite{ls07,ls09,DiJuTho}, groups $\mathrm{PSL}(3,q)$ and $\mathrm{PGL}(3,q)$~\cite{Brooksbank:2010}, groups $\mathrm{PSL}(4,q)$~\cite{Brooksbank:2015}, small Ree groups~\cite{Leemans:2015} and finally, symmetric groups~\cite{fl,Corr} and alternating groups~\cite{flm,flm2}. In particular, only the last two families gave rise to string C-groups of arbitrary large rank. It was proved in \cite{transitive} that the maximal rank of a string C-group for transitive subgroups of $S_n$ that are neither $\Alt_n$ nor $\Sym_n$ is $\frac{n}{2}+1$.
A symmetric group $\Sym_n$ is known to have rank $n-1$~\cite{fl} and an alternating group $\Alt_n$ with $n\geq 12$ is known to have rank at least $\lfloor\frac{n-1}{2}\rfloor$ when $n\geq 12$~\cite{flm2}. It is conjectured in~\cite{flm2} that this is the highest possible rank for a string C-group of alternating type.
In this paper, we prove this conjecture. Our main result is as follows.
\begin{thm}\label{maintheorem}
The rank of $\Alt_n$ is $0$ if $n=3, 4, 6, 7, 8$; $3$ if $n=5$; $4$ if $n=9$; $5$ if $n=10$; $6$ if $n=11$ and $\lfloor\frac{n-1}{2}\rfloor$ if $n\geq 12$.
\end{thm}
The cases where $n\leq 11$ had already been dealt with the use of {\sc Magma}~\cite{flm}.
In this paper, we show that if $\Gamma:=(\Alt_n,S)$ is a string C-group and $n\geq 12$, then $|S| \leq \lfloor\frac{n-1}{2}\rfloor$.
In some parts of this proof we use induction over $n$.
The proof is divided into three parts. In Sections~\ref{s:prim} and
\ref{s:imprim}, we deal with the case where some subgroup $\Gamma_i$ is
primitive or transitive imprimitive, respectively, and our main tool here is permutation group theory. In the remainder of the
paper we have to deal with the case where all $\Gamma_i$'s are intransitive.
Our main tool for this case is the use of \emph{fracture graphs}; but these
are also used elsewhere, so we give a brief introduction here.
Let $\Gamma=\langle\rho_0,\,\ldots,\,\rho_{r-1}\rangle$ be an sggi acting as a permutation group on a set $\{1,\,\ldots,\,n\}$.
We define the {\it permutation representation graph} $\mathcal{G}$ as the $r$-edge-labeled multigraph with $n$ vertices and with a single $i$-edge $\{a,\,b\}$ whenever $a\rho_i=b$ with $a\neq b$.
Suppose we have a sggi $\Gamma$ which is a transitive subgroup of
the symmetric group $S_n$, such that every subgroup $\Gamma_i$ is intransitive.
Then, for each $i$, the permutation $\rho_i$ has a cycle whose points lie in
different $\Gamma_i$-orbits. Choosing one such cycle for each $i$, and
regarding them as the edges of a graph on the vertex set $\{1,\ldots,n\}$,
we obtain a \emph{fracture graph} for $\Gamma$. The fracture graph is of
course not unique, and indeed much of our proof involves showing how to
replace a fracture graph by a more convenient one.
If $\Gamma$ is contained in the group of even permutations (as
in our main theorem), then each permutation $\rho_i$ has at least two cycles.
If it happens that for each $i$ we can find two pairs of points in
different $\Gamma_i$-orbits, then taking an $i$-edge between each of these pairs of points we obtain a
\emph{$2$-fracture graph}. Section~\ref{s:2frac} handles the case where a
$2$-fracture graph exists. Section~\ref{s:no2frac} the case where it
does not, and we then use fracture graphs instead.
\section{$\Gamma_i$ is primitive for some $i$}
\label{s:prim}
Now we embark on the proof of the main theorem. In this section we prove the
theorem in the case where some $\Gamma_i$ is primitive.
Given a string C-group $\Phi:=(G,T)$ with $T:= \{\rho_0, \ldots, \rho_{d-1}\}$, the {\em diagram} of $\Phi$ is a graph with $d$ vertices and an edge between vertices $i$ and $j$ whenever $\rho_i\rho_j$ is not an involution. Moreover, the edge $\{i,j\}$ is then labelled with the order of $\rho_i\rho_j$.
Observe that, by the string property, the diagram of $\Phi$ is a union of paths.
We say that a set $S\subseteq T$ is {\em connected} provided the labels of the generators of $S$ form an interval.
Let us first state a Theorem due to Mar\'oti that will be useful in the proof of the next proposition and also later on.
\begin{thm}[Mar\'oti~\cite{Maroti}]\label{marotiThm}
Let $G$ be a primitive group of degree~$n$ which is not $S_n$ or $A_n$. Then
one of the following possibilities occurs:
\begin{enumerate}
\item For some integers $m,k,l$, we have $n={m\choose k}^l$, and $G$ is a
subgroup of $S_m\wr S_l$, where $S_m$ is acting on $k$-subsets of
$\{1,\ldots,m\}$;
\item $G$ is $M_{11}$, $M_{12}$, $M_{23}$ or $M_{24}$ in its natural
$4$-transitive action;
\item $\displaystyle{|G|\le n\cdot\prod_{i=0}^{\lfloor \log_2n\rfloor-1}
(n-2^i)}$.
\end{enumerate}
\end{thm}
\begin{prop}\label{prim}
Let $n\geq 12$. If $\Phi:=(G,T)$ is a string C-group of rank $d$ with $G<A_n$ and $G$ primitive, then $d\leq (n-3)/2$.
\end{prop}
Indeed, in this case $d$ is asymptotically much smaller than $n/2$.
\begin{proof}
We use the methods of~\cite{transitive}.
Suppose first that the diagram of $\Phi$ is not connected. Then the
primitive group $G$ is the direct product of two proper subgroups,
each of which is necessarily simple and acts regularly; so $|G|=n^2$, and $n\ge60$. But
clearly $|G|\ge2^{d}$; so $d\le 2\log_2n<(n-3)/2$ for $n\ge60$.
So we may suppose that the diagram of $\Phi$ is connected. Now we combine Conder's
lower bound $2^{2d-1}$ for the order of a string C-group of rank $d$
\cite{Conder} with well-known upper bounds for the order of primitive groups,
such as Mar\'oti's (see Theorem~\ref{marotiThm}). We deal with the three cases of Mar\'oti's Theorem.
Case (b) is handled by computer. In case (a), since we are only interested in
an upper bound for $|G|$, we can assume that $G$ is maximal in $A_n$, so that
either $G$ is $S_m$ acting on $k$-sets, or $G=S_m\wr S_l$ with $l>1$. In the first
subcase, $d\leq m-1$, while $n={m\choose k}\ge m(m-1)/2$, hence $d\leq \frac{n-3}{2}$ for $n\geq 12$. In the second subcase, we can use the main result of \cite{transitive} to conclude that, $d\leq \frac{ml}{2}+1$ while $n={m\choose k}^l$ and $n\geq 12$; again this gives $d\leq \frac{n-3}{2}$.
In these cases, $G$ is embeddable in a smaller symmetric group, of degree $m$
in the first case, or $ml$ in the second.
Finally, in case (c), we have
\[2^{2d-1}\leq |G|\le n^{1+\log_2n},\]
If we assume $d\geq \frac{n-2}{2}$, we get
\[2^{n-3}\leq 2^{2d-1}\leq |G|\le n^{1+\log_2n},\]
thus
\[n\leq (\log_2n)(\log_2n+1)+3\]
which
gives a contradiction for $n\geq 34$.
For $n\leq 33$, we give in Table~\ref{primBound} the list of primitive groups of degree $n$ such that their order is $\geq 2^{2\lfloor\frac{n}{2}\rfloor-3}$, following numbering of Sims's list~\cite{BL96}.
When {\sc Magma} is mentioned in the references column, it means we computed all string C-groups representations of the corresponding group using {\sc Magma} and the bound is sharp.
\begin{table}
\begin{tabular}{|c|c|c|c|c|}
\hline
Degree&Number&$G$&Max rank&Reference\\
\hline
12&1&$PSL(2,11)$&4&\cite{LVatlas,ls07}\\
&2&$PGL(2,11)$&3&\cite{LVatlas,ls09}\\
&3&$M_{11}$&0&\cite{LVatlas}\\
&4&$M_{12}$&4&\cite{LVatlas}\\
\hline
13&7&$PSL(3,3)$&0&\cite{LVatlas,Brooksbank:2010}\\
\hline
14&2&$PGL(2,13)$&3&\cite{LVatlas,ls09}\\
\hline
15&3&$A_7$&0&\cite{LVatlas}\\
&4&$PSL(4,2)$&0&{\sc Magma}\\
\hline
16&18&$2^4:S_6$&5&{\sc Magma}\\
&19&$2^4:A_7$&0&{\sc Magma}\\
&20&$2^4:PSL(4,2)$&0&{\sc Magma}\\
\hline
17&8&$P\Gamma L(2,16)$&0&\cite{LVatlas}\\
\hline
22&2&$M_{22}:2$&4&\cite{LVatlas}\\
\hline
23&5&$M_{23}$&0&\cite{HHalg}\\
\hline
24&3&$M_{24}$&5&\cite{HHalg}\\
\hline
\end{tabular}
\caption{Primitive groups $G$ of degree $\leq 33$ with $|G| \geq 2^{2\lfloor\frac{n}{2}\rfloor-3}$.}\label{primBound}
\end{table}
So $d\le(n-3)/2$ in all cases.
\end{proof}
We remark that Mar\'oti's bound uses the Classification of Finite Simple Groups.
The use of such heavy machinery could in principle be avoided by using the
slightly weaker bounds proved by `elementary' means by Babai and
Pyber~\cite{Babai,Pyber}; however, this would require examining of many more
`small' cases, some of which are too large for practical computation.
The previous proposition gives the following corollary that finishes a case for our main theorem, when some $\Gamma_i$ is primitive.
\begin{coro}
Let $n\geq 12$ and let $\Gamma:=(A_n,S)$ be a string C-group of rank $r$. If $\Gamma_i$ is primitive for some $i$ then $r\leq \frac{n-1}{2}$.
\end{coro}
\begin{proof}
If $\Gamma_i$ is primitive for some $i$, then $\Gamma_i < A_n$ and satisfies the hypotheses of Proposition~\ref{prim}. Hence the rank $r-1$ of $\Gamma_i$ is bounded by $\frac{n-3}{2}$.
\end{proof}
\section{$\Gamma_i$ is transitive imprimitive for some $i$}
\label{s:imprim}
In this section, we prove the main theorem in the case where $\Gamma_i$ is
transitive but imprimitive for some $i$.
Let $\Gamma\cong A_n$ with $n\geq 12$ and $\Gamma_i$ be transitive imprimitive for some $i\in\{0,\ldots, r-1\}$.
Let $k$ and $m$ be such that $\Gamma_i$ is embedded into $S_k\wr S_m$. We assume that the blocks of imprimitivity are maximal (so $\Gamma_i$ acts primitively
on the set of blocks), but do not require that $k$ is as big as possible.
Consider the following sets of generators of $\Gamma_i$:
\begin{itemize}\itemsep0pt
\item $L$ an independent generating set for the block action;
\item $C$ the set of generators that commute with all elements of $L$;
\item $R$ the remaining generators.
\end{itemize}
Let us first recall some important results found in \cite{transitive}. We have that $|L|\leq m-1$ and $|C|\leq k-1$. The group $\langle L\rangle$ is primitive on the set of blocks, and $L$ has at most two connected components. When $L$ has two components, $r-1\leq2\log_2m+(k-1)+4$ and $m\geq 60$, thus in that case we have $r\leq \frac{n}{2}-1$. So, in what follows, we assume that $L$ is connected and generates a primitive group on the set of blocks.
\begin{prop}\label{R<=1}
If $m\neq 2$ then $\{\rho_i\}\cup L$ must be connected and $|R|\leq 1$.
\end{prop}
\begin{proof}
As $\Gamma$ is primitive, $\rho_i$ must break the imprimitivity of $\Gamma_i$, thus it must swap at least one pair of points in different blocks.
On the other hand $\langle L\rangle$ is primitive, thus $\rho_i$ cannot commute with every element of $L$. Hence $\{\rho_i\}\cup L$ must be connected and $|R|\leq 1$.
\end{proof}
\subsection{The case $k,m>2$}
\begin{prop}
If $k>2$ and $m>2$, then $r\le\lfloor(n-1)/2\rfloor$.
\end{prop}
\begin{proof}
As observed at the beginning of the section, $|L|\leq m-1$, $|C|\leq k-1$. By Proposition~\ref{R<=1}, $|R|\leq 1$, hence $r-1\leq (m-1)+(k-1)+1$. When $n=12$ the bound that we get for the rank is $7$, but using {\sc Magma}~\cite{BCP97} we found out that there are no polytopes of ranks $6$ or $7$ for $A_{12}$. So we may assume that $n>12$.
If $r>\lfloor(n-1)/2\rfloor$, then
\[\frac{n-1}{2}=\frac{km-1}{2}\leq r\leq k+m,\]
so $(k-2)(m-2)\le5$. The solutions with $km>12$ are $(k,m)=(3,5)$, $(3,6)$,
$(3,7)$, $(4,4)$, $(5,3)$, $(6,3)$, $(7,3)$.
Now we consider these cases. If $(k,m)=(3,7)$ or $(7,3)$, then $r\le10=(21-1)/2$, as required. If $(k,m)=(3,5)$, $(3,6)$, $(4,4)$, $(5,3)$ or
$(6,3)$, then we have $r\le\lfloor(n-1)/2\rfloor$ unless $|L|=m-1$, $|C|=k-1$,
and $|R|=1$. So $\langle C\rangle\cong S_k$, and since $\langle C \rangle $ commutes with a group
acting primitively on the blocks, it acts in the same way on each block.
We also see that $\langle L\rangle$ acts as $S_m$ on the set of blocks, and
since it commutes with $S_k$ fixing the blocks, we have
$\langle L,C\rangle\cong S_k\times S_m$. Transpositions in $S_k$ (resp.~$S_m$)
act as products of $m$ (resp.~$k$) transpositions on the point set. So if either
$m$ or $k$ is odd, then $\Gamma$ contains an odd permutation, a contradiction.
Now suppose that $(k,m)=(4,4)$ and $r=8$. We know that $\rho_i$ commutes with a subgroup
$S_3\times S_4$ with orbits of sizes $4$ and $12$.
We know from the previous paragraph that $S_4\times S_4$
is acting on the product of two sets of size $4$. So when we descend to
$S_3\times S_4$, the orbit of size $4$ has $S_4$ acting in the usual action (and its
centraliser is trivial), while the orbit of size $12$ is the product of sets
of sizes $3$ and $4$. A permutation which commutes with it must fix the two
systems of imprimitivity, so its projection onto each factor commutes with
the corresponding symmetric group, and so is trivial.
\end{proof}
\subsection{The case $k=2$}
The estimate above gives $r\le m+2=\frac{n}{2}+2$. We have to knock three off
this bound. The group induced on the blocks is primitive. It follows that
the centraliser of $\langle L\rangle$ in the symmetric group is generated by
the involution $z$ which interchanges the points of each block. Now if $m$ is
odd, then $z$ is an odd permutation, and so $C=\emptyset$. If $m$ is even, then
$|C|\le1$, and if $z\in\langle L\rangle$ then the intersection property forces
$C=\emptyset$.
We separate the argument into three cases, according as the group $H$ induced on blocks is
$S_m$, $A_m$, or neither of these. The cases $H=S_m$ and $H=A_m$ use similar
arguments, but differ in detail, so we have kept them separate.
\paragraph{\underline{Case $H=S_m$}} We assume that $m\ge7$ for this proof.
Note first that $|R\cup C|\le 1$. Indeed, if both $R$ and $C$ are nonempty, $m$ is even,
and $z$ is contained in $\langle L\cup R\rangle\cap \langle C\rangle$, a contradiction. [This is because the kernel
of the action of $\Gamma_i$ on blocks is an $S_m$-submodule of $(C_2)^m$;
the only such submodules are the trivial ones, the module $M_1$ generated
by $z$, and the module $M_2$ consisting of elements interchanging an even
number of blocks; if $m$ is even then $M_1\le M_2$, and there cannot exist
two independent submodules.]
So $r\le|L|+2$, and if either $|L|\le m-3$ or
$R\cup C=\emptyset$ and $|L|\le m-2$ then we have the required result.
Up to duality, there are the three possibilities, either
\begin{itemize}
\item[(A)] $\rho_i=\rho_0, \, L=\{\rho_1,\ldots, \rho_{r-2}\}, \, R=\{\rho_{r-1}\}\mbox{ and }C=\emptyset,$
\item[(B)] $\rho_i=\rho_{r-2},\, L=\{\rho_0,\ldots, \rho_{r-3}\}, \, C=\{\rho_{r-1}\}\mbox{ and }R=\emptyset,$ or
\item[(C)] $\rho_i=\rho_0,\, L=\{\rho_0,\ldots, \rho_{r-1}\}, \mbox{ and }R=C=\emptyset.$
\end{itemize}
Let $G=\langle L\rangle$. Now $G$ induces the symmetric group $S_m$ on the
set of blocks, and $L$ is an independent set of generators for $G$ as a
string group (not necessarily a string C-group!). We have $|L|\le m-1$.
Assume that $|L|=m-1$ and $m\ge7$. The elements of $L$ induce the Coxeter
generators on the set of blocks: for a certain numbering of the blocks, $\rho_j$ swaps blocks $j$ and $j+1$, for $j\in\{1,\ldots,m-1\}$. (This
is an easy deduction from the result of \cite{CC}.)
In (A) and (C), $\rho_i$ commutes with $\rho_2,\ldots,\rho_{m-1}$, and these elements
generate a group acting as $S_{m-1}$ on blocks, fixing the first block.
Since $m\ge7$, we see that $\rho_i$ must fix all the blocks numbered from
$2$ to $m$, and clearly also block $1$; so it preserves the block system.
In (B) $\rho_i$ commutes with $\rho_0,\ldots,\rho_{m-3}$, which
also acts as $S_{m-1}$ on blocks, and the same applies. So $\Gamma$ preserves
the block system, and is imprimitive, a contradiction to the assumption that
$\Gamma$ is the alternating group.
So we can assume that $|L|=m-2$ and $R\cup C\neq\emptyset.$
Let $K$ be the kernel of the action of $\Gamma_i$ on
the blocks, and let $K_1=K\cap G$. Then $K$ and $K_1$ are $S_m$-submodules of
the permutation module $F^m$, where $F$ is the field with two elements. The
only submodules have dimensions $0$, $1$ (spanned by the all-$1$ vector),
$m-1$ (the vectors of even weight), and $m$. Now since $\Gamma_i$ consists
of even permutations, we cannot have $K=F^m$.
We show that $K=K_1$ is impossible. If $K=K_1$, then $G=\Gamma_i$, and so
$R\cup C=\emptyset$, contrary to our assumption.
Next we show that $K_1=1$ is impossible. In this case, $L$ generates $S_m$ as
string C-group. By the main result of ~\cite{fl,Corr}, there is a unique
possibility, up to duality. In (A) the group generated by
$\rho_2,\ldots,\rho_{r-2}$ is $S_{m-1}$, and $\rho_0$ commutes with this
group, so $\rho_0$ preserves the block system, a contradiction. In (B) the group generated by $\rho_0,\ldots,\rho_{r-4}$ is $S_{m-1}$, and
$\rho_i$ commutes with this group.
We then get the same contradiction as before.
So we are left with the case $K=[(C_2)^m]^+$ and $K_1=\langle z\rangle$.
In this case $G$ is an extension of $C_2$ by $S_m$ and $C=\emptyset$ (as in case (A)).
The involutions $z$ and $\rho_0$ both commute with $\Gamma_{0,1}$, since
$z$ is in the centre of $\Gamma_0$. So the dihedral group $D$ they generate also
commutes with $\Gamma_{0,1}$. Moreover, $\rho_0$ and $z$ do not commute with
each other; if they did, then $\Gamma=\langle\Gamma_0,\rho_0\rangle$ would be
contained in the centralizer of $z$, contradicting the fact that $\Gamma$ is
the alternating group. So $D$ has order $2d$ with $d\geq 3$. We now separate in two
cases.
In the case where $\Gamma_{0,1}$ is transitive, the group $D$ is semiregular;
thus $\Gamma_{0,1}$ has $m/d$ blocks of imprimitivity
each of size $2d$, and is contained in $D\wr S_{m/d}$. Now since $d\ge3$, we
can replace the action of $\Gamma_{0,1}$ by one where each orbit of $D$ has
size $d$, rather than $2d$; this action is still faithful (since $D$ acts
faithfully on $d$ points if $d\ge3$). So $\Gamma_{0,1}$ is isomorphic to a
transitive imprimitive group of degree $m$. By the main result of
\cite{transitive}, we have $r-2\le\frac{m}{2}+1$ (whence $m\le 6$,
which is not so).
So suppose that $\Gamma_{0,1}$ is intransitive on blocks; then also
$\Gamma_{0,1,r-1}$ is intransitive. Let $H_1$ be the group it induces on the
blocks. Now the images of $\rho_2,\ldots,\rho_{r-2}$ form a set of $r-3$
generators for $H_1$ as an sggi (not necessarily a string C-group). If $r\ge m$,
then we conclude that
\begin{itemize}\itemsep0pt
\item $H_1$ has at most three orbits;
\item if it has three orbits, then it acts on each as the symmetric group;
\item if it has two orbits, then it acts on one of them as the symmetric
group.
\end{itemize}
Suppose there are three orbits. Then $H_1$ commutes with the group induced by $D$
(which has at least one orbit of size $d\ge3$), the three orbits must be
isomorphic and a $D$-orbit meets each in one point. But then $H_1\le S_{m/3}$
and the number of independent generators for $H_1$ is at most $m/3-1$. So we
have $m/3-1\ge m-3$, which is impossible.
Suppose that $H_1$ has two orbits $O_1$ and $O_2$ with the action of $H_1$ on $O_1$ being that of the symmetric group $S_{|O_1|}$. We have a dihedral group $D$ commuting with the symmetric group such that each $D$-orbit meets $O_1$ in one point, as these intersections form a system of imprimitivity for $D$ on $O_1$.
Suppose that
the action on the other orbit is not faithful. Then there is a non-trivial
subgroup fixing all points in this orbit (and hence fixing all $D$-orbits)
but non-trivially on $O_1$, and so moving the intersections of $D$-orbits
with $O_1$ (since these have size $1$, only the trivial group fixes them all).
As to the size, each $D$-orbit has one point in $O_1$ and $d-1$ in $O_2$,
so $|O_2|$ is $(d-1)/2$ times the degree, that is, $(d-1)m/d$.
Thus $H_1$ has at most $(d-1)m/d-1$ independent generators. If our
inequality holds, then $(d-1)m/d-1\ge m-3$, from which we get $m\le 2d$. But
then the dihedral group has at most two orbits, and
$\Gamma_{0,1}\le D\wr C_2$.
A group of order $2m^2$ has
largest independent set of size at most $2\log_2m+1$.
This number cannot be $m-2$ or greater for
$m>8$; the remaining cases are resolved by a computer check.
\paragraph{\underline{Case $H=A_m$}} As before, let $L$ be an independent set of generators
for the action of $\Gamma_i$ on blocks, and let $G=\langle L\rangle$. If
$G$ is intransitive, then its orbits form a transversal for the blocks, and
so $G\cong A_m$. By the induction hypothesis, if $m\ge12$, then
$|L|\le\lfloor\frac{m-1}{2}\rfloor$, and so
$r\le\lfloor\frac{m+3}{2}\rfloor\le m-1$, since $m\ge7$. For $m<12$, if
the bound fails, we have either $m=10$ and $|L|=5$, or $m=11$ and $|L|=6$; the
required bound is satisfied in either case.
So we may assume that $G$ is transitive.
Let $K$ be the kernel of the action of $\Gamma_i$ on blocks, and $K'=K\cap G$.
Since $\Gamma_i/K$ and $G/K'$ are both isomorphic to $A_m$, we see that
$K\ne K'$. Moreover, $A_m$ cannot act transitively on $2m$ points, and so
$K'\ne 1$. Since $K$ and $K'$ are submodules of the $A_m$-module $(C_2)^m$,
and neither is the whole of $(C_2)^m$ (which contains odd permutations), we
must have $|K|=2^{m-1}$, $|K'|=2$. The generator of $K'$ is the involution $z$
which interchanges the points in each block. Since this is an even permutation,
$m$ must be even. Moreover $C=\emptyset$. Thus up to duality we may assume that $i=0$.
If $R=\emptyset$ then $r\leq m-1$, hence we now assume that $R=\{\rho_{r-1}\}$.
Now $z$ is in the centre of $\Gamma_0$, and so commutes with $\Gamma_{0,1}$.
The involution $\rho_0$ also commutes with $\Gamma_{0,1}$, and by the
intersection property $\rho_0\ne z$. Let $D=\langle z,\rho_0\rangle$, a
dihedral group of order $2d$, say. Now $\rho_0$ and $z$ do not commute: for,
if they did, then $\langle\rho_0,\Gamma_0\rangle$ would be contained in the
centraliser of $z$, whereas in fact this group is $A_{2m}$. In particular,
$d\ge3$.
Suppose first that $\Gamma_{0,1}$ is transitive. Then $D$, which commutes
with a transitive group, is semiregular; and $\Gamma_{0,1}$, which commutes
with $D$, is isomorphic to a subgroup of $D\wr S_{m/d}$. Since $D$ acts
faithfully on $d$ points (as $d\ge3$), $\Gamma_{0,1}$ is isomorphic to a transitive
imprimitive group on $m$ points. By the main theorem of \cite{transitive},
we get $r-2\le\frac{m}{2}+1$, so $m\le6$.
Now suppose that $\Gamma_{0,1}$ is intransitive.
We know that $\langle\rho_1,\ldots,\rho_{r-2}\rangle\cong C_2\times A_m$,
and $\{\rho_1,\ldots,\rho_{r-2}\}$ are string C-group generators. We claim that their images, $\bar{\rho}_1,\ldots,\bar{\rho}_{r-2}$,
in $\Gamma_{0,r-1}/\langle z\rangle\cong A_m$ are independent.
Suppose not. Note that they generate $A_m$ as an sggi. If they fail to be
independent, one of them can be expressed in terms of the others. Suppose that
it is $\bar{\rho}_h$. We cannot have $1<h<r-2$, since then these elements would
generate a commuting product of two subgroups. We cannot have $h=1$, since
we are assuming that $\Gamma_{0,1}$ is intransitive. And finally, we cannot
have $h=r-2$. For if so, then
$C_{A_{2m}}(A_m)=\langle z\rangle$; but $\rho_{r-1}$ centralises
$\langle\rho_1,\ldots,\rho_{r-3}\rangle=A_m$, so $\rho_{r-1}=z$, contradicting
the intersection property, since $z\in\langle\rho_1,\ldots,\rho_{r-2}\rangle$.
Now the images mod $z$ of $\rho_2,\ldots,\rho_{r-2}$ are independent, and
generate an intransitive subgroup of $A_m$. So $r-3\le m-3$. If equality
holds, then this group has just two orbits; it acts on each orbit as the
symmetric group.
But this contradicts the fact that these elements belong to
$\Gamma_{0,1}$, which centralises the dihedral group $D=\langle\rho_0,z\rangle$
having at least one orbit of size greater than $2$ on the set of blocks.
\paragraph{\underline{Case $H\ne S_m,A_m$}}
In this case we prove the following result on independent sets.
\begin{prop} \label{IGSprim}
Let $G$ be a primitive group of degree $n\ge8$, not isomorphic to $A_n$ or
$S_n$. Then the maximum size of an independent generating set of $G$ is at
most $n-4$.
\end{prop}
\begin{proof}
Let $M(G)$ be the maximum size of an independent generating set of $G$.
We consider separately the three possibilities given by Theorem~\ref{marotiThm}.
In case (a) when $l\geq 2$ we have
$n-4 = (^m_k)^l-4\geq m^l-4\geq ml-2\geq M(G)$. When $l=1$ we have $k\geq 2$, $m\leq n/2$ and the group is a subgroup of $S_m$ or $A_m$, so $M(G) \leq m-1$, much smaller than $n-4$.
In case (b) we have to consider the groups $M_{11}, M_{12}, M_{23}$ or $M_{24}$. The maximal length of a chain of subgroups of $M_{11}$, $M_{23}$ or $M_{24}$ is 7, 11 and 14 resp. (see \cite{Whis00}).
If $G$ is isomorphic to $M_{12}$ then $M(G) \leq 9$ by~\cite{Whis00}.
Suppose that $M(G) = 9$. Then one of the following subgroups $H$ of $M_{12}$, namely $M_{11}$ or $P\Gamma L(2,9)$, has to have $M(H) = 8$.
As $M(M_{11}) \leq 7$ (see~\cite{Whis00}), we must have $M(P\Gamma L(2,9))=8$. A quick look at the subgroup lattice of $M_{12}$ shows that this is impossible as two subgroups of order 1440 never intersect in a subgroup of order 720.
Hence $M(M_{12}) \leq 8$.
In case (c) the chain length is bounded by $\log_2 \left[n.\prod_{i=0}^{\lfloor\log _2 n-1\rfloor}(n-2^i)\right]$ that is at most $n-4$ for $n\geq 26$.
We also know that if $|G| = p_1^{e_1}\ldots p_k^{e_k}$ then the chain length (and hence $M(G)$) is bounded by $e_1+ \ldots + e_k$.
Combining these two bounds, and using {\sc Magma}, we conclude the result holds for $n>9$ and for $n=8$ we are left with $P\Gamma L_1(8)$, $PSL_2(7)$, $PGL_2(7)$ and $ASL_3(2)$. But for those, looking at the subgroup lattice we get $M(G)\leq 4$.
\end{proof}
\subsection{The case $m=2$}
Suppose that $m=2$, so that $\Gamma_i\le S_k\wr S_2$.
An involution interchanging the blocks is fixed-point-free so $k$ is even and $n = 0 \mod 4$.
We separate the argument into two cases, according as there is or is not a
value of $j\ne i$ such that $\Gamma_{i,j}$ is transitive. First, suppose
that such a $j$ exists.
\begin{prop}
If $\Gamma_{i,j}$ is transitive for some $j\neq i$ and $\Gamma_i$ is transitive imprimitive embedded into $S_{n/2}\wr S_2$, then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
Suppose $r>\frac{n-1}{2}$ and $\Gamma_{i,j}$ is transitive for $j\neq i$.
Then the two
groups $\Gamma_i$ and $\Gamma_j$ are transitive, so each has two
blocks of size $n/2$. The stabilisers of the blocks for the two subgroups
each have index $2$, so their intersection is a normal subgroup of index $4$
in $\Gamma_{i,j}$.
The two block systems cannot be the same, since then they
would be preserved by $\langle\Gamma_i,\Gamma_j\rangle=\Gamma$. These
intersections are blocks for $\Gamma_{i,j}$, of size $n/4=l$, say, and the action
on the blocks (which is not primitive in this case) is isomorphic to the Klein
group $C_2\times C_2$ generated, modulo the normal subgroup, by a set $L$ of
size $2$.
Now we play the usual game: let $C$ be the set of generators commuting with
$L$ (so $C$ acts in the same way on each block, and has rank at most $l-1$),
and $R$ the remaining generators of $\Gamma_{i,j}$, so $|R|\le4$.
Thus, $r-2=|R|+|C|+|L|\le l+5$. As $r>\frac{n-1}{2}$, $l+7\ge2l$, so $l\le7$, and $n\le 28$. Since $n$ is a multiple of $4$,
we only need to consider the cases $n=16$, $20$, $24$ and $28$.
To deal with the exceptions, we first subdivide into two cases, according as
the permutations of $L$ commute. Let $L=\{\rho_s, \rho_r\}$.
Suppose first that $\rho_s$ and $\rho_r$ do not commute. Then they are
adjacent in the diagram; so we can improve our estimate to $|R|\le2$. Also,
there is a vertex $v$ such that, if we follow a path with labels $r,s,r,s$,
we arrive at a different point $w$ in the same block. Then the stabilizer of
$v$ in $\langle C\rangle$ also fixes $w$. So $\langle C\rangle$ is not the
symmetric group, and we have $|C|\le l-2$. Then we have $r-2\le l+2$, so
$n\le16$.
Now suppose that $\rho_s$ and $\rho_r$ commute. If $|R|=4$, or if
$|R|=3$ and the diagram of $C$ is connected, then at
least one element of $R$, say $\rho_h$, also commutes with $C$. If
$\langle C\rangle$ acts primitively on a block, then the centralizer of $\langle C\rangle$
is generated by $\rho_r$ and $\rho_s$, and so
$\rho_h\in\langle\rho_r,\rho_s\rangle$, a contradiction. So we conclude that
either $C$ is disconnected (giving $|C|\le l-2$), or $\langle C\rangle$ is imprimitive (giving $|C|\le l/2+1$)
or $|R|\le 2$. Putting these into our estimates shows that $n\le24$.
If $n=24$, in the worst case scenario, we have $|R| = 4$, $|C| = 4$ and $|L| = 2$ possibly giving $r=12$.
If there are two permutations $\rho_a$ and $\rho_b$ such that $\Gamma_{i,j,a,b}$ is transitive, then it may be assumed that $\Gamma_a$ and $\Gamma_b$ are both embedded into $S_{n/2}\wr S_2$.
This forces $n$ to be divisible by 16, a contradiction. If $|R|\geq 2$ then $\langle C\rangle$ is intransitive within the blocks, otherwise another two generators can be removed from $\Gamma_{i,j}$ and the group remains transitive. Let $\langle C\rangle$ be intransitive within the blocks. If $|C|=4$ then one element of $R$ fuses the orbits of $\langle C\rangle$ inside a block, hence $|R|\leq 2$, by the same reason. Thus the case $|C|=|R|=4$ leads to a contradiction.
We eliminate the case $n=20$ as follows. For $n=20$, we must have either
$|C|=|R|=3$, or $|C|=4$, $|R|=2$. In the first case, the diagram of $C$
is disconnected, so we must have $C\cong S_3\times S_2$, with orbit lengths
$3$ and $2$. Take an element of $R$ which commutes with one component of
the diagram of $C$, and modify it using $\rho_i$ and $\rho_j$ so that it
fixes all the blocks; this element commutes with $S_2\times S_3$, but this
group contains its centraliser in $S_5$, a contradiction. In the second
case, $\langle C\rangle=S_5$. Then $\langle C,\rho_j\rangle\cong S_5\times S_2$
fixes the blocks of imprimitivity for $\Gamma_i$. But this group is maximal
in the stabiliser of this block system. There can be at most one more generator
in $\Gamma_i$, giving $|R|\le1$, a contradiction.
For $n=16$, we used the computer to show the nonexistence of such a group.
\end{proof}
In what follows we consider that $\Gamma_{i,j}$ is intransitive for every $j\neq i$. We prove that $r\leq \frac{n-1}{2}$ using a certain fracture graph and we use the following results that are immediate consequences of the definition of a fracture graph of a given sggi $G$ with permutation representation graph $\mathcal{G}$. By an alternating square, in $\mathcal{G}$, we mean a square having opposed edges with the same label.
\begin{itemize}
\item (\cite{extension}, Proposition 3.2 ) if two edges of $\mathcal{G}$, $e$ and $e'$, have the same label and $e$ is an edge of a fracture graph, then at least one vertex of $e'$ is in a different component of that fracture graph;
\item (\cite{extension}, Proposition 3.6) when two edges of an alternating square of $\mathcal{G}$, belong to a fracture graph of $G$, the vertex of the square not in these edges is in another component of that fracture graph.
\end{itemize}
When representing $\mathcal{G}$, dashed edges are used to represent edges that are in $\mathcal{G}$ but not in the chosen fracture graph.
\begin{prop}
If $\Gamma_{i,j}$ is intransitive for every $j\neq i$ and $\Gamma_i$ is transitive imprimitive embedded into $S_{n/2}\wr S_2$, then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
Let $L=\{\rho_l\}$.
Suppose that $R=\emptyset$. In this case $\Gamma_i\leq 2\times S_{n/2}$ and therefore we could see $\Gamma_I$ as an imprimitive group with blocks of size two which we dealt with before. Hence $r\leq \frac{n-1}{2}$.
We may now assume that $R$ is nonempty.
First observe that if $\rho_j\in R$ then $\rho_j$ connects at least two pairs of $\langle \rho_l\rangle$-orbits.
Indeed as $\rho_j$ is an even permutation and $\rho_j$ does not commute with $\rho_l$, there must be an even number (different from zero) of pairs of $\langle\rho_l\rangle$-orbits joined by a single $\rho_j$-edge.
Suppose $\rho_{l-1}\in R$.
Consider the graph $\mathcal{L}$ whose vertices are the $\langle \rho_l\rangle$-orbits and with a $j$-edge for each element of $C\cup R$ connecting
$\langle \rho_l\rangle$-orbits in different $\Gamma_{i,j}$-orbits.
Note that $\mathcal{L}$ is a fracture graph for the group action on the $\langle \rho_l\rangle$-orbits of $\Gamma_i$.
Since $\Gamma_{i,j}$ is intransitive for every $j\neq i$, we have that $\mathcal{L}$ has no cycles.
Then $\rho_{l-1}$ must connect at least two pairs of $\langle \rho_l\rangle$-orbits, as observed before.
Let us denote two of them by $(L_1,L'_1)$, and $(L_2,L'_2)$. We have that $|\{L_1,L'_1,L_2,L'_2\}|\geq 3$.
Suppose that $\{L_1,L'_1\}$ is the $(l-1)$-edge of $\mathcal{L}$.
Then either $L_2$ or $L'_2$ is another connected component of $\mathcal{L}$, different from the component of $L_1$.
Let $L_1$ and $L_2$ be in different connected components of $\mathcal{L}$.
Therefore $|C\cup R|\leq n/2-2$.
Suppose we have the equality.
Then $\mathcal{L}$ has exactly two connected components, and
either $|C|=n/2-3$ and $|R|=1$, or $|C|=n/2-4$ and $R=2$.
Let $|C|=n/2-3$ and $R=\{\rho_{l-1}\}$.
We claim that the incident edges of $\mathcal{L}$ have consecutive labels.
Suppose the contrary.
Let $g$ and $h$ be non-consecutive labels of incident edges of $\mathcal{L}$. Consider first $g,h\neq l-1$.
Then there exists an alternating square, with labels $g$ and $h$, whose vertices are $\langle\rho_l\rangle$-orbits.
Three vertices of the square belong to one connected component of $\mathcal{L}$ and the fourth belongs to another component of $\mathcal{L}$.
Let us denote this fourth vertex by $L$.
\[ \[email protected]{ *++[][F]{} \ar@{-}[d]_g\ar@{-}[r]^h & *++[][F]{}\ar@{--}[d]^g \\
*++[][F]{} \ar@{--}[r]_h& *+[][F]{L}}\]
Hence it may be assumed that $L$ and $L_2$ are in the same connected component of $\mathcal{L}$.
\[ \[email protected]{ *+[][F]{L'_2} \ar@{--}[d]_{l-1} \ar@{~}[rr] && *++[][F]{} \ar@{-}[d]_g \ar@{-}[r]^h & *++[][F]{}\ar@{--}[d]^g \\
*+[][F]{L_2} \ar@{~}[rr]&& *++[][F]{} \ar@{--}[r]_h& *+[][F]{L}}\]
There are paths in $\mathcal{L}$ from the two ends of the $g$-edge to $L_2$ and $L_2'$.
But then these two vertices are in the same $\Gamma_g$-orbit, a contradiction.
Now suppose there is an $h$-edge with $h\neq l-2$ incident to an $(l-1)$-edge. These two edges are on an alternating square.
Let $h$ be minimal with this property.
As $k>4$, there is a vertex incident to that square.
Now the label of the edge connecting that vertex to the square must be $h+1$.
Hence there is also an alternating square adjacent to the first one, with labels $h+1$ and $l-1$.
These two squares give three different connected components of $\mathcal{L}$, a contradiction.
Thus two incident edges in $\mathcal{L}$ must be consecutive.
This gives a unique possibility for the permutation representation graph of $\Gamma_i$.
$$\[email protected]{ *+[o][F]{} \ar@{-}[r]^l & *+[o][F]{} \ar@{-}[d]^{l+1}\\
*+[o][F]{} \ar@{-}[d]^{l+1} \ar@{-}[r]^l & *+[o][F]{} \\
*+[o][F]{} \ar@{-}[d]^{l+2} \ar@{-}[r]^l & *+[o][F]{} \ar@{-}[d]^{l+2} \\
*+[o][F]{} \ar@{-}[d]^{l+3} \ar@{-}[r]^l & *+[o][F]{} \ar@{-}[d]^{l+3} \\
*+[o][F]{} \ar@{-}[d]^{l+4} \ar@{-}[r]^l & *+[o][F]{} \ar@{-}[d]^{l+4} \\
*+[o][F]{} \ar@{.}[d] \ar@{-}[r]^l & *+[o][F]{} \ar@{.}[d] \\
*+[o][F]{} & *+[o][F]{} } $$
Now there are two possibilities for $i$, namely either $i=0$ or $i=r-1$.
But as $\rho_i$ must break the block system, the only possibility is $i=r-1$ and $\rho_i$ a single transposition connecting the two bottom vertices of the picture below, a contradiction.
Let $|C|=n/2-4$ and $R=\{\rho_{l-1},\rho_{l+1}\}$. As observed at the beginning of the proof of this proposition, $\rho_{l+1}$ connects at least two pairs of $\langle\rho_l\rangle$-orbits. If it does not connect $L_2$ to $L'_2$, then $\mathcal{L}$ has three connected components, a contradiction.
Hence assume both $\rho_{l-1}$ and $\rho_{l+1}$ connect $L_2$ to $L'_2$.
As $k\geq 3$ there exists another $\langle \rho_l\rangle$-orbit $L$ that is adjacent to either $L_2$ or $L'_2$. Let $\rho_h\in R\cup C$ be a permutation connecting $L$ and $L_2$. If $\rho_h\in C$ then, as it commutes with at least one of $\rho_{l-1}$ and $\rho_{l+1}$,
there is an alternating square in the permutation representation graph of $\Gamma$ with labels $h$ and one of $l-1$ and $l+1$.
This implies that $\rho_h$ also connects two pairs of $\langle \rho_l\rangle$-orbits different from $(L_2,L_2')$.
Hence $|C\cup R|\leq n/2-3$, a contradiction.
Thus $\rho_h\in R$.
We may assume that $h=l-1$.
If a $(l+1)$-edge is incident to a $(l-1)$-edge then there exists an alternating square and we get a contradiction as before.
Hence there exists a double edge with labels $l+1$ and $l-1$ and no other incidence between edges with these labels.
Also if incident edges have labels in $C$, then the labels must be consecutive.
Moreover $\Gamma_i$ has the following permutation representation graph.
$$\[email protected]{ *+[o][F]{} \ar@{=}[d]_{l-1}^{l+1} \ar@{-}[r]^l & *+[o][F]{} \\
*+[o][F]{} \ar@{-}[r]_l & *+[o][F]{} \ar@{-}[d]^{l-1} \\
*+[o][F]{} \ar@{-}[d]_{l+1} \ar@{-}[r]^l & *+[o][F]{} \\
*+[o][F]{} \ar@{-}[d]_{l+2} \ar@{-}[r]^l & *+[o][F]{} \ar@{-}[d]^{l+2} \\
*+[o][F]{} \ar@{-}[d]_{l+3} \ar@{-}[r]^l & *+[o][F]{} \ar@{-}[d]^{l+3} \\
*+[o][F]{} \ar@{.}[d] \ar@{-}[r]^l & *+[o][F]{} \ar@{.}[d] \\
*+[o][F]{} & *+[o][F]{} } $$
Now there are two possibilities for $i$, namely either $i=0$ or $i=r-1$. In each case either $\rho_i$ is an odd permutation or $\rho_i$ fixes the blocks, a contradiction.
\end{proof}
\section{All $\Gamma_i$'s intransitive: $2$-fracture graphs exist}
\label{s:2frac}
In this section and the next, we handle the case where all subgroups
$\Gamma_i$ are intransitive. As explained in the introduction, we use the
techniques of fracture graphs. In this section we deal with the case where
$2$-fracture graphs exist. In fact, for later use, our results are more
general: we do not assume that $\Gamma\cong A_n$ until the end of the section.
Let $\mathcal{G}$ be a permutation representation graph on $n$ vertices for $\Gamma$, which is connected (meaning that this permutation representation of $\Gamma$ is transitive). From now on, we assume that $\Gamma$ is indeed the permutation group defined by $\mathcal{G}$.
Suppose that $\Gamma$ has no transitive maximal parabolic subgroup $\Gamma_i$.
We make the further assumption that each $\rho_i$ interchanges at least two pairs of points which lie in different $\Gamma_i$-orbits for all $i$.
Then $\mathcal{G}$ has a subgraph with $n$ vertices and $2r$ edges corresponding to two pairs of vertices in different $\Gamma_i$-orbits. We call this graph a \emph{2-fracture graph}.
For ease of notation, we denote by $q_{i,j}$ an alternating square with labels $i$ and $j$, and call a cycle with more than four vertices a \emph{big cycle}.
We give some basic results that follow immediately from the definition of a 2-fracture graph.
\begin{prop}\label{paths}
If $e=\{v,w\}$ is an $i$-edge of a 2-fracture graph $Q$ of $\Gamma$, then any path from $v$ to $w$ in $\mathcal{G}$ which does not contain $e$ must contain another $i$-edge.
\end{prop}
\begin{prop}\label{isoncy}
A cycle in a 2-fracture graph has either zero or two $i$-edges. In particular, a 2-fracture graph has no multiple edges.
\end{prop}
\begin{prop}\label{common}
If one $i$-edge is in the intersection of a pair of cycles of a 2-fracture graph, then both $i$-edges are in the intersection of those cycles. In particular, an edge of a square of a 2-fracture graph cannot be common to any other cycle.
\end{prop}
\begin{proof}
If only one $i$-edge is common to two cycles of a 2-fracture graph, then there is a cycle with only one $i$-edge, which is not possible by Proposition~\ref{isoncy}.
\end{proof}
\begin{prop}\label{cons}
A big cycle of a 2-fracture graph has adjacent edges with nonconsecutive labels.
\end{prop}
\begin{proof}
Let $i$ be the smallest edge label in a big cycle. By Proposition~\ref{isoncy} there are two $i$-edges in that cycle. These $i$-edges must be adjacent to at least three further edges. The only edges in the cycle with labels consecutive with $i$ have label $(i+1)$ (by minimality of $i$), and there are at most two of these; so the cycle has adjacent edges with nonconsecutive labels.
\end{proof}
In the following two propositions we show two useful ways of getting a 2-fracture graph from another. In the first one an $i$-edge of a 2-fracture graph that is in a cycle is replaced by another $i$-edge in the same cycle. In the second proposition, one $i$-edge of a cycle is in a 2-fracture graph and
the other $i$-edge in that cycle is replaced by the $i$-edge not
in the cycle in the 2-fracture graph.
\begin{prop}\label{swap}
If there is a cycle $C$ in $\mathcal{G}$ containing exactly two $i$-edges $e_1, e_2$, such that $e_1$ is in a 2-fracture graph $\mathcal{Q}$ and $e_2$ is not, then there is another 2-fracture graph $\mathcal{Q}'$ obtained by removing $e_1$ and adding $e_2$.
\end{prop}
\begin{proof}
The graph $\mathcal{Q}'$ is a 2-fracture graph, because the edge $e_2$ is between vertices in different $\Gamma_i$-orbits.
\end{proof}
\begin{prop}\label{putincycle}
If there is a cycle $C$ in $\mathcal{G}$ containing exactly two $i$-edges $e_1, e_2$, such that $e_1$ is in a 2-fracture graph $\mathcal{Q}$ and $e_2$ is not, then there is another 2-fracture graph $\mathcal{Q}'$ obtained by removing the $i$-edge of $\mathcal{Q}$ which is not $e_1$ and adding $e_2$.
\end{prop}
\begin{proof}
The proof is the same as that of the previous Proposition.
\end{proof}
Both of the above Propositions will be of particular use when the cycle $C$ is an alternating square.
\begin{prop}\label{movingsquares}
Let $q_{i,j}$ be an alternating square with a vertex $v$ $l$-adjacent to a vertex $w$ in a 2-fracture graph $\mathcal{Q}$. If $l$ is not consecutive with $i$, then the square can be moved to include the edge $\{v,w\}$. That is, there is another 2-fracture graph $\mathcal{Q}'$ obtained from $\mathcal{Q}$ by changing exactly two edges, as pictured below. Furthermore $\mathcal{Q}'$ does not have more alternating squares than $\mathcal{Q}$.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^j \ar@{-}[dd]^i && *+[o][F]{v} \ar@{-}[rr]^l \ar@{-}[dd]^i&&*+[o][F]{w}\ar@{--}[dd]^i\\
&&&&\\
*+[o][F]{}\ar@{-}[rr]_j &&*+[o][F]{}\ar@{--}[rr]_l&& *+[o][F]{}}\qquad\longrightarrow\qquad\[email protected]{*+[o][F]{} \ar@{-}[rr]^j \ar@{--}[dd]^i && *+[o][F]{v} \ar@{-}[rr]^l \ar@{-}[dd]^i&&*+[o][F]{w}\ar@{-}[dd]^i\\
&&&&\\
*+[o][F]{}\ar@{-}[rr]_j &&*+[o][F]{}\ar@{-}[rr]_l&& *+[o][F]{}}$$
\end{prop}
\begin{proof}
Let $l$ be not consecutive with $i$. There is an alternating square in $\mathcal{G}$, sharing an $i$-edge with $q_{i,j}$. We apply Proposition~\ref{swap} to the other $i$-edge of $q_{i,j}$, and Proposition~\ref{putincycle} for the $l$-edges.
\end{proof}
\begin{prop}\label{changelabels}
If a 2-fracture graph $\mathcal{Q}$ contains the subgraph on the left of the following figure, then there is a 2-fracture graph $\mathcal{Q}'$ containing the subgraph on the right, such that $\mathcal{Q}$ and $\mathcal{Q}'$ differ only in four edges.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&*+[o][F]{} &&\\
&&&&\\
*+[o][F]{}\ar@{-}[rr]_{i-1}&&*+[o][F]{}\ar@{-}[rr]_i\ar@{--}[uurr]^l && *+[o][F]{}\ar@{-}[rr]_l&& *+[o][F]{}\ar@{--}[uull]_i}\qquad\longrightarrow\qquad\[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&*+[o][F]{} &&\\
&&&&\\
*+[o][F]{}\ar@{-}[rr]_{i-1}&&*+[o][F]{}\ar@{--}[rr]_i\ar@{-}[uurr]^l && *+[o][F]{}\ar@{--}[rr]_l&& *+[o][F]{}\ar@{-}[uull]_i} $$
\end{prop}
\begin{proof}
This is a consequence of applying Proposition~\ref{swap} twice.
\end{proof}
\begin{prop}\label{nobigcicles}
If $\Gamma$ has a 2-fracture graph, then it has one with no big cycle.
\end{prop}
\begin{proof}
Consider a 2-fracture graph $\mathcal{Q}$ with $q$ big cycles. Suppose that $C$ is a big cycle of $\mathcal{Q}$. By Proposition~\ref{cons} there is at least one pair of adjacent edges with nonconsecutive labels $i$ and $j$. Hence these edges belong to an alternating square $q_{i,j}$ of $\mathcal{G}$. The other two edges of this square are not edges of $\mathcal{Q}$. Indeed the remaining $i$ and $j$ edges of $\mathcal{Q}$ must belong to $C$ by Proposition~\ref{isoncy}.
Consider the 2-fracture graph $\mathcal{Q}'$ that is obtained from $\mathcal{Q}$ by applying Proposition~\ref{swap} twice, replacing the $i$-edge and the $j$-edge of $C$ which are not in $q_{i,j}$ by those that are not in $C$ but are in $q_{i,j}$.
By Proposition~\ref{common} the edges of the square cannot belong to another cycle, thus $\mathcal{Q}'$ has $q-1$ big cycles.
Continuing this process we obtain a 2-fracture graph without big cycles.
\end{proof}
\begin{prop}\label{max1squareforcomp}
If $\mathcal{G}$ has a 2-fracture graph, then it has a 2-fracture graph such that each connected component has at most one cycle, and any cycle is an alternating square.
\end{prop}
\begin{proof}
By Proposition~\ref{nobigcicles}, $\mathcal{G}$ contains a 2-fracture graph $\mathcal{Q}_0$ having no big cycle.
If $\mathcal{Q}_0$ has a cycle, then it is an alternating square by Proposition~\ref{isoncy}.
Suppose that a connected component of $\mathcal{Q}_0$ has $p$ alternating squares with $p>1$.
We will prove that there exists a 2-fracture graph with $p-1$ squares and without any big cycle.
The proof is a double induction. Suppose that $q_{i,j}$ and $q_{l,k}$ are squares of $\mathcal{Q}_0$ that have distance $s$ which is the smallest distance between two squares in $\mathcal{Q}_0$. We produce another 2-fracture graph with
either one fewer square, or the minimal distance $s$ reduced by $1$.
Either the squares $q_{i,j}$ and $q_{l,k}$ share a vertex, or there is a path in $\mathcal{Q}_0$ connecting them.
Suppose that they share a vertex. Then there exists another alternating square in $\mathcal{G}$. We may assume that it is $q_{i,k}$, which shares an edge with both $q_{i,j}$ and $q_{l,k}$.
$$\[email protected]{*+[o][F]{}\ar@{-}[rr]^j &&*+[o][F]{} \ar@{--}[rr]^k &&*+[o][F]{}\ar@{--}[dd]^i\\
&&&&\\
*+[o][F]{}\ar@{-}[uu]^i \ar@{-}[rr]^j&& *+[o][F]{} \ar@{-}[uu]^i \ar@{-}[rr]^k \ar@{-}[dd]_l&& *+[o][F]{} \ar@{-}[dd]_l\\
&&&&\\
&& *+[o][F]{} \ar@{-}[rr]_k &&*+[o][F]{} }
$$
By Proposition~\ref{swap}, we can replace one of the $i$-edges of $\mathcal{Q}_0$ with the $i$-edge of $q_{i,k}$ not in $\mathcal{Q}_0$ to obtain a 2-fracture graph $\mathcal{Q}_1$. Note that this new $i$-edge of $\mathcal{Q}_1$ does not belong to any cycle by Proposition~\ref{common}. Thus $\mathcal{Q}_1$ has $p-1$ squares.
Now assume that the shortest path between $q_{i,j}$ and $q_{k,l}$ has $s \geq 1$ edges. If the first or the last edge of this path has a label not consecutive with the square that it meets, then we can use Proposition~\ref{movingsquares}, to create a new 2-fracture graph which does not have more squares, and with a smaller minimal distance between squares.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^j \ar@{-}[dd]^i && *+[o][F]{} \ar@{--}[rr]^x \ar@{-}[dd]^i&&*+[o][F]{}\ar@{--}[dd]^i&& *+[o][F]{} \ar@{-}[rr]^k \ar@{-}[dd]^l&& *+[o][F]{} \ar@{-}[dd]^l\\
&&&&&&&&\\
*+[o][F]{}\ar@{-}[rr]_j &&*+[o][F]{}\ar@{-}[rr]_x&& *+[o][F]{}\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_k&&*+[o][F]{}}\qquad\longrightarrow\qquad\[email protected]{*+[o][F]{} \ar@{--}[rr]^j \ar@{--}[dd]^i && *+[o][F]{} \ar@{-}[rr]^x \ar@{-}[dd]^i&&*+[o][F]{}\ar@{-}[dd]^i&& *+[o][F]{} \ar@{-}[rr]^k \ar@{-}[dd]^l&& *+[o][F]{} \ar@{-}[dd]^l\\
&&&&&&&&\\
*+[o][F]{}\ar@{-}[rr]_j &&*+[o][F]{}\ar@{-}[rr]_x&& *+[o][F]{}\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_k&&*+[o][F]{}}$$
Otherwise, we can use Proposition~\ref{changelabels} in order to create a new 2-fracture graph with the first or the last edge of a new path having a label not consecutive with the square that it meets, and then use Proposition~\ref{movingsquares} as before.
Continuing this process, we get a 2-fracture graph where the minimal distance between squares is zero, and then applying the argument above, we get a 2-fracture graph with fewer squares, and without big cycles. This construction terminates with a 2-fracture graph with no big cycles and with at most one square in each connected component.
\end{proof}
\begin{prop}\label{3cases}
Suppose that $\Gamma$ has a disconnected 2-fracture graph $\mathcal{Q}$ with no big cycles such that every connected component has exactly one square.
Then $\Gamma$ has a 2-fracture graph $\mathcal{Q}'$ with the same characteristics as $\mathcal{Q}$ and such that the minimal distance between two squares in $\mathcal{G}$ is either one, two or three accordantly to one of the following three cases.
$$(1)\[email protected]{*+[o][F]{} \ar@{-}[rr]^i \ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&\\
*+[o][F]{} \ar@{-}[rr]^i && *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]^l&& *+[o][F]{} }
\quad (2)\[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} }$$
$$(3) \[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&&&& *+[o][F]{} \ar@{-}[rr]^{j-1} \ar@{-}[dd]_{j+1}&& *+[o][F]{}\ar@{-}[dd]_{j+1} \\
&&&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_j && *+[o][F]{}\ar@{-}[rr]_{j-1} && *+[o][F]{} }$$
\end{prop}
\begin{proof}
As $\mathcal{Q}$ contains all vertices of $\mathcal{G}$, any component of $\mathcal{Q}$ is at distance one from another component of $\mathcal{Q}$. Consider two components in $\mathcal{Q}$ at distance one. Repeatedly using Propositions~\ref{movingsquares} and \ref{changelabels}, we can obtain a 2-fracture graph that contains one of the above subgraphs.
\end{proof}
\subsection{Disconnected 2-fracture graphs}
We now separate the argument into two cases, dealing first with the case of
a disconnected 2-fracture graph.
\begin{prop}\label{dislc}
If $\mathcal{G}$ has no connected 2-fracture graph, then it has a 2-fracture graph that has at least one component which is a tree, all the others having only one cycle (which is an alternating square).
\end{prop}
\begin{proof}
Suppose that $\mathcal{Q}$ is a disconnected 2-fracture graph of $\Gamma$.
By Proposition~\ref{max1squareforcomp} there exists a 2-fracture graph $\mathcal{Q}$ of $\Gamma$ having at most one cycle, that is an alternating square, in each connected component. We proceed by induction on the number of connected components of $\mathcal{Q}$.
Suppose that $\mathcal{Q}$ has $p$ connected components, each with an alternating square. By Proposition~\ref{3cases}, we can choose $\mathcal{Q}$ so that it has a subgraph as shown in one of the three cases of Proposition~\ref{3cases}.
In what follows we deal with each of these cases separately.
Let us consider that $\mathcal{G}$ has the subgraph (1) of Proposition~\ref{3cases}. Observe here that $x$ is not consecutive with at least one of the labels of the squares. In addition since the squares of $\mathcal{Q}$ are in different connected components of $\mathcal{Q}$, the $x$-edge in not in $\mathcal{Q}$. Suppose $x$ and $l$ are nonconsecutive. Then we have an alternating square as in the following picture on the left. Using Proposition~\ref{swap} we obtain $\mathcal{Q}'$ on the right.
$$\[email protected]{*+[o][F]{} \ar@{-}[rr]^i \ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&\\
*+[o][F]{} \ar@{-}[rr]^i && *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]^l&& *+[o][F]{}\\
&&&&&&\\
&&&&*+[o][F]{}\ar@{--}[uurr]_x\ar@{--}[uull]^l && }\qquad\longrightarrow\qquad \[email protected]{*+[o][F]{} \ar@{-}[rr]^i \ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[dd]_j&& *+[o][F]{} \ar@{--}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&\\
*+[o][F]{} \ar@{-}[rr]^i && *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]^l&& *+[o][F]{}\\
&&&&&&\\
&&&&*+[o][F]{}\ar@{--}[uurr]_x\ar@{-}[uull]^l && }$$
If the other $x$-edge is not in $\mathcal{Q}$, then we created a connected component which is a tree. On the other hand, if the other $x$-edge is in $\mathcal{Q}$, then the new 2-fracture graph has $p-1$ connected components.
Let us consider the second case of Proposition~\ref{3cases}.
First suppose that $x$ is not consecutive either with $l$ or $k$.
Without loss of generality we assume that it is not consecutive with $l$; then we use Proposition~\ref{swap} as shown in the following diagram.
$$\[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} \\
&&&&&&&&\\
&&&&&&*+[o][F]{}\ar@{--}[uurr]_x\ar@{--}[uull]^l &&}\qquad\longrightarrow\qquad \[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&& *+[o][F]{} \ar@{--}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} \\
&&&&&&&&\\
&&&&&&*+[o][F]{}\ar@{--}[uurr]_x\ar@{-}[uull]^l &&}$$
The argument for why this exchange does not create any new cycles is the same as the previous case.
Similar to above, if neither of the $x$-edges is in $\mathcal{Q}$, then we created a connected component which is a tree. On the other hand, if one of the $x$-edges is in $\mathcal{Q}$, then the new 2-fracture graph has $p-1$ connected components.
Next, consider that $x$ is consecutive with both $l$ and $k$, then we have the following diagram.
$$\[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} \\
&&&&&&&&\\
&&&&*+[o][F]{}\ar@{--}[uurr]_i\ar@{--}[uull]^x&& &&\\
&&*+[o][F]{}\ar@{--}[urr]_{i-1}\ar@{--}[uuull]^x&&&&&&}\qquad\longrightarrow\qquad\[email protected]{*+[o][F]{} \ar@{--}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_k&& *+[o][F]{}\ar@{-}[dd]_k \\
&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} \\
&&&&&&&&\\
&&&&*+[o][F]{}\ar@{--}[uurr]_i\ar@{--}[uull]^x&& &&\\
&&*+[o][F]{}\ar@{-}[urr]_{i-1}\ar@{--}[uuull]^x&&&&&&}$$
Note that $x$ is not consecutive with $i$, and not consecutive with either $i-1$ or $i+1$. Without loss we assume it is not consecutive with $i-1$. Suppose that both $x$-edges of the square $q_{x,i-1}$
are in $\mathcal{Q}$. Then using Proposition~\ref{swap}, we create a square $q_{x,i-1}$ in $\mathcal{Q}'$.
Now if both $i$-edges are in $\mathcal{Q}$ we have reduced the number of connected components; otherwise, we go back to case (1).
Similarly, if an $x$-edge of $q_{x,i-1}$ is not in $\mathcal{Q}$, then we either created a 2-fracture graph with $p-1$ components or with a component that is a tree.
Finally, we consider the diagram below for case (3). Here, $x$ is not consecutive with either $i$ or $j$. We may assume it is not consecutive with $i$. Furthermore, $x$ is not consecutive either with $i-1$ or $i+1$. Let us assume it is not consecutive with $i-1$. A similar argument shows that the new graph either has $p-1$ components or a component that is a tree.
$$\[email protected]{*+[o][F]{} \ar@{-}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&&&& *+[o][F]{} \ar@{-}[rr]^{j-1} \ar@{-}[dd]_{j+1}&& *+[o][F]{}\ar@{-}[dd]_{j+1} \\
&&&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_j && *+[o][F]{}\ar@{-}[rr]_{j-1} && *+[o][F]{} \\
&&&&&&&&&&\\
&&&& *+[o][F]{}\ar@{--}[uurr]_i\ar@{--}[uull]^x &&&&&&\\
&&*+[o][F]{}\ar@{--}[urr]_{i-1}\ar@{--}[uuull]^x&&&&&&&&}\quad\longrightarrow\quad\[email protected]{*+[o][F]{} \ar@{--}[rr]^{i-1} \ar@{-}[dd]_{i+1}&& *+[o][F]{} \ar@{-}[dd]_{i+1}&&&&&& *+[o][F]{} \ar@{-}[rr]^{j-1} \ar@{-}[dd]_{j+1}&& *+[o][F]{}\ar@{-}[dd]_{j+1} \\
&&&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_{i-1} && *+[o][F]{} \ar@{-}[rr]_i&& *+[o][F]{} \ar@{--}[rr]^x&& *+[o][F]{} \ar@{-}[rr]_j && *+[o][F]{}\ar@{-}[rr]_{j-1} && *+[o][F]{} \\
&&&&&&&&&&\\
&&&& *+[o][F]{}\ar@{--}[uurr]_i\ar@{--}[uull]^x &&&&&&\\
&&*+[o][F]{}\ar@{-}[urr]_{i-1}\ar@{--}[uuull]^x&&&&&&&&}$$
After any of the exchanges seen above, we have a 2-fracture graph with no big cycle, and either $p-1$ connected components or a component which is a tree. Therefore, by induction, there exists a 2-fracture graph that is either connected or disconnected with at least one component being a tree.
\end{proof}
\subsection{Connected 2-fracture graphs}
In this subsection we will assume that $\Gamma$ has a connected 2-fracture graph.
\begin{prop}\label{squareinconeclcomp0}
Let $q_{i,j}$ be an alternating square of a connected 2-fracture graph $\mathcal{Q}$ of $\Gamma$. Let $q_{i,l}$ be an alternating square in $\mathcal{G}$ sharing an $i$-edge with $q_{i,j}$. Then neither or both the $l$-edges of $q_{i,l}$ are in $\mathcal{Q}$.
\end{prop}
\begin{proof}
Suppose that exactly one $l$-edge of $q_{i,l}$ is in $\mathcal{Q}$. Let $u$ be the vertex of $q_{i,l}$ which is adjacent to $q_{i,j}$ in $\mathcal{G}$ by the $l$ edge of $q_{i,l}$ which is not in $\mathcal{Q}$.
$$\[email protected]{*+[o][F]{} \ar@{-}[rr]^j \ar@{-}[dd]_i&& *+[o][F]{} \ar@{-}[rr]^l \ar@{-}[dd]_i&& *+[o][F]{} \ar@{--}[dd]^i\\
&&&&\\
*+[o][F]{} \ar@{-}[rr]_j && *+[o][F]{} \ar@{--}[rr]_l && *+[o][F]{u} }$$
Any possibility to connect $u$ to $q_{i,j}$ gives a contradiction to Proposition~\ref{paths}.
\end{proof}
\begin{prop}\label{squareinconeclcomp2}
Let $q_{i,j}$ be an alternating square of a connected 2-fracture graph $\mathcal{Q}$.
All edges of $\mathcal{G}$ meeting the vertices of $q_{i,j}$ belong to $\mathcal{Q}$ and have labels consecutive either with $i$ or $j$.
\end{prop}
\begin{proof}
Let $w$ be a vertex of $q_{i,j}$ $k$-adjacent (in $\mathcal{G}$) to a vertex $v$.
Suppose that $\{v,w\}$ is not in $\mathcal{Q}$.
$$\[email protected]{&&*+[o][F]{v}\ar@/^/@{.}[ddrr] &&\\
&&&&\\
*+[o][F]{} \ar@{-}[rr]^j \ar@{-}[dd]_i&& *+[o][F]{w} \ar@{--}[uu]^k \ar@{-}[rr]^l \ar@{-}[dd]_i&& *+[o][F]{} \ar@{--}[dd]^i\\
&&&&\\
*+[o][F]{} \ar@{-}[rr]_j && *+[o][F]{} \ar@{--}[rr]_l && *+[o][F]{u} }$$
There exists a path in $\mathcal{Q}$ from $v$ to $w$.
Let $l$ be the label of the edge of this path meeting $q_{i,j}$.
Suppose that there is a square $q_{i,l}$ or $q_{j,l}$ sharing an edge with $q_{i,j}$.
Suppose first that $q_{i,l}$ contains $w$.
Let $u$ be the vertex of $q_{l,i}$ diagonally opposed to $w$. One of the $l$-edges of $q_{l,i}$ is not in $\mathcal{Q}$, because both are in the path from $w$ to $v$. By Proposition \ref{squareinconeclcomp0}, we have a contradiction.
Similar arguments rule out the case where $q_{i,l}$ shares the other $i$-edge with $q_{i,j}$.
Thus $|i-j|=2$ and $l$ is consecutive with both $i$ and $j$. But then $k$ is not consecutive either with $i$ or with $j$.
Suppose without loss of generality it is not consecutive with $j$, thus we have a square $q_{j,k}$ sharing an edge with $q_{i,j}$.
Let $\{v',w'\}$ be the other $k$-edge of $q_{i,k}$, with $v'$ not in $q_{i,j}$.
$$\[email protected]{*+[o][F]{v'}\ar@{--}[dd]_k\ar@{--}[rr]^j&&*+[o][F]{v}\ar@/^/@{.}[ddrr] &&\\
&&&&\\
*+[o][F]{w'} \ar@{-}[rr]^j \ar@{-}[dd]_i&& *+[o][F]{w} \ar@{--}[uu]^k \ar@{-}[rr]^l \ar@{-}[dd]_i&& *+[o][F]{} \\
&&&&\\
*+[o][F]{} \ar@{-}[rr]_j && *+[o][F]{} && }$$
By Proposition~\ref{squareinconeclcomp0}, $\{v',w'\}$ is not in $\mathcal{Q}$.
Thus there is a path in $\mathcal{Q}$, connecting $w'$ to $v'$, and this path does not have any $l$-edges. Let $l'$ be the label of the edge of this path meeting $w'$. As $l'$ is not consecutive with both $i$ and $j$ we have a contradiction as before. Thus $\{v,w\}$ is in $\mathcal{Q}$.
Furthermore, as all edges of $\mathcal{G}$ adjacent to $q_{i,j}$ are in $\mathcal{Q}$, $k$ must be consecutive to either $i$ or $j$. Otherwise, we would have at least four edges with the same label in $\mathcal{Q}$.
\end{proof}
\begin{prop}\label{squareedge}
If $q_{i,j}$ is an alternating square of a connected 2-fracture graph $\mathcal{Q}$ with $n \geq 9$ vertices, then each vertex of $q_{i,j}$ has degree at most three in $\mathcal{Q}$.
\end{prop}
\begin{proof}
Suppose that $q_{i,j}$ has a vertex adjacent, in $\mathcal{Q}$, to two other edges, and let $k$ and $l$ be the labels of those edges.
By Proposition~\ref{squareinconeclcomp2} $k$ and $l$ must be consecutive with one of the labels of $q_{i,j}$.
We may assume that $i<j$ and $l<k$. Then there are three possibilities either $i<l<k<j$, $l<i<j<k$ or $i<l<j<k$ corresponding to the following graphs.
$$\[email protected]{*+[o][F]{}\ar@{--}[rr]^i &&*+[o][F]{} &&\\
&&&&\\
*+[o][F]{}\ar@{-}[uu]^k \ar@{-}[rr]^i\ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[uu]^k \ar@{-}[rr]^l \ar@{-}[dd]_j&& *+[o][F]{} \ar@{--}[dd]^j\\
&&&&\\
*+[o][F]{} \ar@{-}[rr]_i && *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} }\quad
\[email protected]{*+[o][F]{}\ar@{--}[rr]^i &&*+[o][F]{w} \ar@{--}[rr]^l &&*+[o][F]{v} \ar@{--}[dd]^k\\
&&&&\\
*+[o][F]{}\ar@{-}[uu]^k \ar@{-}[rr]^i\ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[uu]^k \ar@{-}[rr]^l \ar@{-}[dd]_j&& *+[o][F]{} \ar@{--}[dd]^j\\
&&&&\\
*+[o][F]{} \ar@{-}[rr]_i && *+[o][F]{} \ar@{-}[rr]_l && *+[o][F]{} }\quad
\[email protected]{*+[o][F]{}\ar@{--}[rr]^i &&*+[o][F]{} \ar@{--}[rr]^l &&*+[o][F]{}\ar@{--}[dd]^k\\
&&&&\\
*+[o][F]{}\ar@{-}[uu]^k \ar@{-}[rr]^i\ar@{-}[dd]_j&& *+[o][F]{} \ar@{-}[uu]^k \ar@{-}[rr]^l \ar@{-}[dd]_j&& *+[o][F]{} \\
&&&&\\
*+[o][F]{} \ar@{-}[rr]_i && *+[o][F]{} && }
$$
Consider the second case. As $\mathcal{Q}$ is connected, there is a path in $\mathcal{Q}$ between $v$ and $w$. Then by Proposition~\ref{swap}, we can create a cycle in a 2-fracture graph having only one edge with some label, contradicting Proposition~\ref{isoncy}.
The third case becomes the same as the first case by using Proposition~\ref{swap} on the $i$-edges.
Now we consider the first case. If the labels are not all consecutive, then another pair of labels give an alternating square and we get the same contradiction as in the second case.
Suppose the labels are consecutive, then we use the fact that there is another vertex of $\mathcal{Q}$ connected to one of the vertices of one of the graphs above. Again using Propositions~\ref{swap} and \ref{squareinconeclcomp2}, independent of how this other vertex is connected, we can obtain the same contradiction as in the second case.
\end{proof}
\begin{prop}\label{minimalsquare}
Let $n\geq 9$. If $\Gamma$ has a 2-fracture graph with a square, then it has a 2-fracture graph having a square $q_{i,j}$ such that $|i-j|=2$.
\end{prop}
\begin{proof}
Choose $\mathcal{Q}$, $i$, and $j$ so that the alternating square $q_{i,j}$ has the property that $|i- j|$ is minimal.
First suppose that $|i - j| \geq 3$. By Proposition~\ref{squareinconeclcomp2} there is a vertex of the square $q_{i,j}$ which is $k_1$-adjacent to another vertex with $k_1\in \{ i-1, i+1, j-1, j+1\}$. Without loss of generality, we can assume $k_1 = i \pm 1$, and thus there is an alternating square $q_{k_1,j}$ sharing an edge with $q_{i,j}$. By Proposition~\ref{swap}, there is another connected 2-fracture graph $\mathcal{Q}'$ with $q_{k_1,j}$ as its alternating square. Since we assumed that $| i- j |$ is minimal we know that $ | k_1- j| > | i - j |$.
Consider the vertices $v_i, \, i\in\{0,\ldots,6\}$ as in the following figure.
$$\[email protected]{
*+[o][F]{v_1}\ar@{-}[rr]^i\ar@{-}[dd]_j &&*+[o][F]{v_3} \ar@{-}[dd]^j\ar@{-}[rr]^{k_1}&& *+[o][F]{v_5}\ar@{--}[dd]^j \\
&&&&\\
*+[o][F]{v_2}\ar@{-}[rr]_i&&*+[o][F]{v_4} \ar@{-}[rr]_{k_1}&& *+[o][F]{v_6} }$$
By Proposition~\ref{squareedge} the degree of the vertices $v_3$ and $v_4$ is three. Suppose that $v_1$ has degree three, then the label $k$ of the edge incident to $v_1$, not in $q_{i,j}$, would satisfy $ | k- j | < | i- j|$. As above, using Proposition~\ref{swap}, we would obtain a contradiction. Thus the degree of $v_1$ is two, and similarly the degree of $v_2$ is two.
As $n \geq 9$ the degree of either $v_5$ or $v_6$ is three. Assume without loss of generality that $v_5$ is $k_2$-adjacent to a vertex $v_7$. We have that $k_2$ must be consecutive with $k_1$; furthermore, there is an alternating square $q_{k_2,j}$ with vertices $v_5,v_6,v_7, v_8$, and thus $ | k_2 - j| > | k_1 - j | > | i- j| $. Continuing this process we get a sequence of labels $k_3,\ldots, k_s$ with $ | k_s - j| >\ldots> | k_1 - j | > | i- j| $. Thus $\mathcal{Q}$ does not have edges with labels between $i$ and $j$. We can then conclude that $\Gamma$ is not connected, a contradiction. Hence $|i - j| \leq 2$.
Suppose that $|i-j|=1$, say $j=i+1$.
Then by Proposition~\ref{squareinconeclcomp2} there is a vertex either $i-1$ or $i+2$ adjacent to this square. This situation guarantees another alternating square either $q_{i-1,i+1}$ or $q_{i+2,i}$. By Proposition~\ref{swap}, there is another connected 2-fracture graph having a square $q_{i,j}$ with $j=i+2$, as wanted.
\end{proof}
\begin{prop}\label{i-j=2}
Let $n\geq 9$. If $q_{i-1,i+1}$ is an alternating square of a connected 2-fracture graph $\mathcal{Q}$, then both $i$-edges are incident to the square as shown in the following figure.
$$\[email protected]{
*+[o][F]{}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^i&& *+[o][F]{v} \\
&&&&\\
*+[o][F]{}\ar@{-}[rr]_{i+1}&&*+[o][F]{} \ar@{-}[rr]_i && *+[o][F]{w} }$$
Moreover the degree, in $\mathcal{Q}$, of vertices $v$ and $w$ is one.
\end{prop}
\begin{proof}
Suppose that there is no $i$-edge incident to a vertex of the square $q_{i+1,i-1}$.
Then, by Proposition~\ref{squareinconeclcomp2}, an edge incident to $q_{i+1,i-1}$ has labels $i+2$ or $i-2$.
By duality we may assume there is an edge incident to the square having label $i+2$. Then there is a square
sharing an edge with $q_{i+1,i-1}$, as shown in the following figure.
$$\[email protected]{
*+[o][F]{v_3}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{v_1} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^{i+2}&& *+[o][F]{v} \ar@{--}[dd]^{i-1} \\
&&&&\\
*+[o][F]{v_4}\ar@{-}[rr]_{i+1}&&*+[o][F]{v_2} \ar@{-}[rr]_{i+2}&& *+[o][F]{w} }$$
By Proposition~\ref{squareedge} there is no $i$-edge incident to the vertices of the figure above. Furthermore, the vertices $v_1$ and $v_2$ have degree three and the vertices $v_3$ and $v_4$ have degree two. Thus one of the vertices on the right must have degree three. Suppose $v$ has degree three. Then it is $(i+3)$-incident to another vertex, and we get a square $q_{i+3,i-1}$ as pictured below.
$$\[email protected]{
*+[o][F]{}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^{i+2}&& *+[o][F]{} \ar@{-}[rr]^{i+3} \ar@{--}[dd]^{i-1} && *+[o][F]{} \ar@{--}[dd]^{i-1}\\
&&&&&&\\
*+[o][F]{}\ar@{-}[rr]_{i+1}&&*+[o][F]{} \ar@{-}[rr]_{i+2}&& *+[o][F]{} \ar@{-}[rr]_{i+3} && *+[o][F]{} }$$
Again, by Proposition~\ref{squareedge}, there is no $i$-edges incident to the vertices of this figure. Continuing this process we get a graph that doesn't have $i$-edges, hence $\Gamma$ is disconnected, a contradiction.
Hence there is a vertex of the square $i$-adjacent to another vertex $v$. Consider $v_1$, $v_2$, $v_3$ and $v_4$ as in the following figure.
$$\[email protected]{
*+[o][F]{v_3}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{v_1} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^i&& *+[o][F]{v} \\
&&&&\\
*+[o][F]{v_4}\ar@{-}[rr]_{i+1}&&*+[o][F]{v_2} && && }$$
If the degree of $v$, in $\mathcal{Q}$, is greater than one, then there exist an edge with label $j$ not consecutive with $i$ and incident to $v$.
Then there is an alternating square $q_{i,j}$ containing $v$ and $v_1$.
Hence $v_1$ has degree at least four in $\mathcal{G}$, which is not possible by Proposition ~\ref{squareedge}.
Thus the degree of $v$ is one in $\mathcal{Q}$.
Suppose that $v_4$ is $i$-adjacent, in $\mathcal{Q}$, to a vertex $w$. Then the degree of $w$, analogously to $v$, is one. Thus either $v_2$ or $v_3$ is $k$-adjacent to another vertex. As $k$ is not consecutive either with $i+1$ or $i-1$ there is a square $q_{k, i\pm 1}$ containing $v_1$ or $v_4$. Thus $v_1$ or $v_4$ is of degree four and we get a contradiction.
Now suppose that neither $v_2$ nor $v_3$ is $i$-adjacent to any vertex.
Either $v_2$, $v_3$, or $v_4$ has degree three. Using duality we may assume that either $v_3$ or $v_4$ has degree three. Suppose at first that $v_3$ has degree three.
By the same arguments as above $v_3$ is $(i+2)$-adjacent to another vertex $v_5$. Then there exists an alternating square $q_{i+2,i-1}$ containing $v_3$, $v_4$, $v_5$, and $v_6$ as seen in the following figure. Thus, both $v_3$ and $v_4$ have degree three.
$$\[email protected]{
*+[o][F]{v_5}\ar@{-}[rr]^{i+2}\ar@{--}[dd]_{i-1}&&*+[o][F]{v_3}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{v_1} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^i&& *+[o][F]{v} \\
&&&&&&\\
*+[o][F]{v_6}\ar@{-}[rr]_{i+2}&&*+[o][F]{v_4}\ar@{-}[rr]_{i+1}&&*+[o][F]{v_2} && }$$
On the other hand, if we assume that $v_4$ has degree three, then using duality, we may assume it is incident to an edge of label $i+2$. Thus there is an alternating square $q_{i-1,i+2}$ containing $v_3$; thus both situations give the same result.
Now if the other $i$-edge of $\mathcal{Q}$ is incident to either $v_5$ or $v_6$, this creates an alternating square $q_{i,i+2}$ containing either $v_3$ or $v_4$, which is not possible because they have degree three. Continuing this process we never get a way to connect the other $i$-edge, giving a contradiction.
Hence either $v_2$ or $v_3$ is $i$-adjacent, in $\mathcal{Q}$, to some vertex $w$. Moreover the same argument used to prove that $v$ has degree one holds for $w$.
\end{proof}
\begin{prop}\label{notconwithsquare}
Let $n\geq 9$. If $\mathcal{G}$ has a connected 2-fracture graph, then either the 2-fracture graph is a tree or $\Gamma$ is embedded in $C_2\wr S_{n/2}$.
\end{prop}
\begin{proof}
Suppose $\mathcal{G}$ has a connected 2-fracture graph that is not a tree and let $\mathcal{Q}$ be a 2-fracture graph with a square $q_{i+1,i-1}$ accordantly with Proposition~\ref{minimalsquare}.
By Proposition~\ref{i-j=2} there are two $i$-edges incident to the square. Let $v_1$, $v_2$, $v_3$ and $v_4$ be the vertices of $q_{i+1,i-1}$, as in the following figure.
$$\[email protected]{
*+[o][F]{v_3}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{v_1} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^i&& *+[o][F]{v} \\
&&&&\\
*+[o][F]{v_4}\ar@{-}[rr]_{i+1}&&*+[o][F]{v_2} \ar@{-}[rr]_i && *+[o][F]{w} }$$
The degree of $v$ and $w$ are both one. Hence, by Proposition~\ref{i-j=2}, either $v_3$ or $v_4$ has degree three.
Suppose without loss of generality that $v_3$ has degree three.
Now the label of the edge incident to $v_3$, not in the square $q_{i-1,i+1}$, must have label $i+2$.
This creates an alternating square $q_{i-1,i+2}$ in $\mathcal{G}$.
In particular the degree of $v_4$ is also three.
$$\[email protected]{
*+[o][F]{v_5}\ar@{-}[rr]^{i+2}\ar@{--}[dd]_{i-1}&&*+[o][F]{v_3}\ar@{-}[rr]^{i+1} \ar@{-}[dd]_{i-1} &&*+[o][F]{v_1} \ar@{-}[dd]^{i-1}\ar@{-}[rr]^i&& *+[o][F]{v} \\
&&&&&&\\
*+[o][F]{v_6}\ar@{-}[rr]_{i+2}&&*+[o][F]{v_4}\ar@{-}[rr]_{i+1}&&*+[o][F]{v_2} \ar@{-}[rr]_i&& *+[o][F]{w} }$$
As $n\geq 9$ either $v_5$ or $v_6$ has degree three. Assume that $v_5$ is incident to another vertex $v_7$. Then the label of the edge $\{v_5,v_7\}$ must be $(i+3)$ and the degree of $v_5$ must be three. Then there exists an alternating square $q_{i-1,i+3}$ in $\mathcal{G}$, and therefore the degree of $v_6$ must be three. If $n=10$ we are done. Otherwise we continue this process. The result is the graph below where $i=1$.
$$\[email protected]{
*+[o][F]{}\ar@{--}[dd]_0\ar@{-}[rr]^{r-1}&&*+[o][F]{}\ar@{--}[dd]_0\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]^5\ar@{--}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^4\ar@{--}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^3\ar@{--}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^2 \ar@{-}[dd]_0 &&*+[o][F]{} \ar@{-}[dd]^0\ar@{-}[rr]^1&& *+[o][F]{} \\
&&&&&&&&&&&&&&\\
*+[o][F]{}\ar@{-}[rr]^{r-1}*+[o][F]{}&&*+[o][F]{}\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_1&& *+[o][F]{} }$$
Now the permutation graph of $\Gamma$ is the graph above or has another $0$-edge connecting the vertices on the right. Indeed these are all the edges of $\mathcal{G}$ for otherwise there is a cycle containing an edge of $\mathcal{Q}$ and no other edge with the same label, a contradiction.
In any case we get $\Gamma$ embedded into $C_2\wr S_{\frac{n}{2}}$ (since all the permutations $\rho_i$ preserve the partition whose parts are the $0$-edges and the pair of vertices on the right).
\end{proof}
\begin{coro}
With the assumptions of Proposition~\ref{notconwithsquare}, if the 2-fracture graph is not a tree, there are two possibilities for $\Gamma$, namely
\begin{enumerate}
\item The permutation representation graph is the following and $\Gamma\cong C_2\times S_{n/2}$.
$$\[email protected]{
*+[o][F]{}\ar@{-}[dd]_0\ar@{-}[rr]^{r-1}&&*+[o][F]{}\ar@{-}[dd]_0\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^4\ar@{-}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^3\ar@{-}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^2 \ar@{-}[dd]_0 &&*+[o][F]{} \ar@{-}[dd]^0\ar@{-}[rr]^1&& *+[o][F]{} \ar@{-}[dd]^0\\
&&&&&&&&&&&&&&\\
*+[o][F]{}\ar@{-}[rr]_{r-1}*+[o][F]{}&&*+[o][F]{}\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_1&& *+[o][F]{} }$$
\item
The permutation representation graph is the following and $\Gamma$ depends on parity of $n$.
If $n/2$ is even, then $\Gamma\cong C_2\wr S_{n/2}$ and if $n/2$ is odd then $\Gamma\cong C_2^{n/2-1}: S_{n/2}$ which is a subgroup of index $2$ in $C_2\wr S_{n/2}$.
$$\[email protected]{
*+[o][F]{}\ar@{-}[dd]_0\ar@{-}[rr]^{r-1}&&*+[o][F]{}\ar@{-}[dd]_0\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^4\ar@{-}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^3\ar@{-}[dd]_0&&*+[o][F]{}\ar@{-}[rr]^2 \ar@{-}[dd]_0 &&*+[o][F]{} \ar@{-}[dd]^0\ar@{-}[rr]^1&& *+[o][F]{} \\
&&&&&&&&&&&&&&\\
*+[o][F]{}\ar@{-}[rr]^{r-1}*+[o][F]{}&&*+[o][F]{}\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_1&& *+[o][F]{} }$$
\end{enumerate}
\label{2fracfinal}
\end{coro}
Finally in this section, we prove the main theorem in the case where all
$\Gamma_i$ are intransitive and $2$-fracture graphs exist. So our assumptions
now are:
\begin{itemize}\itemsep0pt
\item $\Gamma\cong A_n$;
\item all $\Gamma_i$'s are intransitive, and each $\rho_i$ interchanges at least
two pairs of points in different $\Gamma_i$-orbits.
\end{itemize}
The conclusions we reached from the second assumption are given in Propositions
\ref{dislc} and \ref{notconwithsquare}: either
\begin{itemize}\itemsep0pt
\item there is a $2$-fracture graph of which one component is a tree and the others are unicyclic; or
\item the permutations are as given in Corollary~\ref{2fracfinal}.
\end{itemize}
In the first case, such a graph has $n-1$
edges. So $2r\le n-1$, as required.
In the second, the group $\Gamma$ is not $A_n$.
\section{All $\Gamma_i$'s intransitive: no $2$-fracture graphs}
\label{s:no2frac}
In this section we continue to handle the case where all subgroups $\Gamma_i$ are intransitive. In particular, we deal with the case where $\Gamma$ does not have any possible 2-fracture graph. Although some of our results are more general, throughout this section we will make the assumption that $\Gamma$ is isomorphic to $A_n$.
Suppose that all maximal parabolic subgroups of $\Gamma$ are intransitive but there exists $i\in\{0,\ldots, r-1\}$ such that $\rho_i$ permutes only one pair $\{a,b\}$ of vertices in different $\Gamma_i$-orbits.
Consequently, the only generators that can act non-trivially on $a$ and $b$ are $\rho_{i-1}$ and $\rho_{i+1}$.
We will say that the orbit of $a$ is the first $\Gamma_i$-orbit and the orbit of $b$ is the second $\Gamma_i$-orbit.
Let $n_1$ and $n_2$ be the sizes of the first and the second $\Gamma_i$-orbit, respectively; and let $A$ and $B$ be the correspondent groups, determined by the action of $\Gamma_i$ on each orbit. Both $A$ and $B$ are string groups generating by involutions (or sggi's). Indeed let $\rho_j=\alpha_j\beta_j$ with $\alpha_j$ and $\beta_j$ being the permutations in each $\Gamma_i$-orbit.
Then $A := \langle \alpha_i\, |\, i\in \{0,\ldots, r-1\}\rangle$ and $B := \langle \beta_i\, |\, i\in \{0,\ldots, r-1\}\rangle$.
\begin{prop}\label{CCD}
If $A$ is primitive, then the set $J_A:= \{ i \,|\, i \in \{0,\ldots, r-1\}$ and $\alpha_i \neq 1_A\}$ is an interval. The same result holds for $B$.
\end{prop}
\begin{proof}
Suppose that $J_A$ is not an interval.
Then $A = H \times K$ for $H = \langle \alpha_j \mid j \in J_1 \rangle$ and $K = \langle \alpha_j \mid j \in J_2 \rangle$ for some disjoint index sets $J_1$ and $J_2$ such that $J_A = J_1 \cup J_2$. As both $H$ and $K$ are transitive on the $n_1$ points, the cardinality of $J_1$ and $J_2$ is at least two.
Moreover each generator $\alpha_j$ commutes with all generators of a transitive group on $n_1$ points, either $H$ or $K$, which implies that it has full support on the first $\Gamma_i$-orbit. Therefore, it has a nontrivial action on $a$. However, we have seen that the only generators that can act nontrivially on $a$ are $\rho_{i-1}$ and $\rho_{i+1}$. This gives a contradiction, so $J_A$ is an interval.
The proof also works for $B$.
\end{proof}
Thanks to Proposition~\ref{CCD} we consider (up to duality) separately the following cases : Case (1) $A$ and $B$ are both imprimitive; Case (2): $J_A$ and $J_B$ are intervals and $i\notin\{ 0,r-1\}$; Case (3): we deal with the remaining cases, particularly we assume that $J_B= \emptyset$ or an interval.
\subsection{Case (1): $A$ and $B$ are both imprimitive}
Let $A$ be embedded into $S_{k_1}\wr S_{m_1}$ and $B$ be embedded into $S_{k_2}\wr S_{m_2}$ with $n_1=m_1k_1$ and $n_2=m_2k_2$.
Consider a minimal subset $M$ of the set of generators of $\Gamma_i$ generating the group induced on the two block systems. Let $R$ be the set containing the remaining generators of $\Gamma_i$.
Consider the permutation representation graph $\mathcal{X}$ for the block action, that is, a graph having $m_1+m_2$ vertices, corresponding to the blocks, and with a $j$-edge between two blocks whenever $\rho_j$ swaps them.
As $\Gamma_i$ has exactly two orbits, the graph $\mathcal{X}$ has two connected components.
Also, consider the subgraph $\mathcal{\bar{X}}$ of $\mathcal{X}$ with the same vertices and with a $j$-edge for each element of $M$, between blocks in different $\Gamma_j$-orbits. This is a fracture subgraph of $\mathcal{X}$, particularly $\mathcal{\bar{X}}$ has no cycles. Hence $|M| \leq m_1+m_2- 2 $.
Similarly, consider the graph $\mathcal{Y}$ with $k$ vertices corresponding to the $\langle M\rangle$-orbits, with a $j$-edge between a pair of $\langle M\rangle$-orbits $L$ and $L'$ whenever there is $x\in L$ such that $x\rho_j\in L'$ with $\rho_j\in R$. Let $\mathcal{\bar{Y}}$ be a fracture subgraph of $\mathcal{Y}$ having only one $j$-edge for each element $\rho_j\in R$ between $\langle M\rangle$-orbits in different $\Gamma_j$-orbits. As before, $\mathcal{\bar{Y}}$ has no cycles and has at least two components. Hence $|R| \leq k - 2$. Note that $k\leq k_1+k_2$, hence $|R| \leq k_1+k_2- 2$.
\begin{prop}\label{M=m-2}
If $|M|= m_1+m_2-2$ then $\mathcal{X}$ has two connected components and consecutive labels. Up to a duality, $\mathcal{X}$ is the following graph.
$$\[email protected]{ *+[F]{} \ar@{-}[rr]^{i-m_1+1} && *+[F]{ } \ar@{.}[rr] && *+[F]{ } \ar@{-}[rr]^{i-2} &&*+[F]{} \ar@{-}[rr]^{i-1} && *+[F]{a} && *+[F]{b} \ar@{-}[rr]^{i+1} && *+[F]{ } \ar@{-}[rr]^{i+2} && *+[F]{ } \ar@{.}[rr]&&*+[F]{ } \ar@{-}[rr] ^{i+m_2-1}&&*+[F]{ }} $$
\end{prop}
\begin{proof}
Since $|M|= m_1+m_2-2$ and $\mathcal{\bar{X}}$ has two components, $\mathcal{\bar{X}} = \mathcal{X}$.
As $\rho_i$ only swaps the points $a$ and $b$ in different $\Gamma_i$-orbits, a $j$-edge of $\mathcal{X}$ incident to the blocks containing $a$ and $b$, must be consecutive with $i$.
The graph $\mathcal{X}$ does not have any alternating squares, and thus only edges with consecutive labels are incident. Up to duality, we have determined $\mathcal{X}$.
\end{proof}
\begin{prop}\label{M=m-3(1)}
If $|M|= m_1+m_2-3$ then, up to duality, either $m_1=2$ or $m_1\geq 4$, accordantly to one of the following graphs.
$$(1) \,\[email protected]{ *+[F]{} \ar@{-}[rr]^{i-1} &&*+[F]{a} &&*+[F]{b} \ar@{-}[rr]^{i-1} && *+[F]{ } \ar@{-}[rr]^{i-2}& &*+[F]{} \ar@{.}[rr] &&*+[F]{} \ar@{-}[rr] ^{i-m_2-1}&&*+[F]{ } } $$
$$(2) \,\[email protected]{ *+[F]{} \ar@{-}[rrr]^{i+m_1-3} &&&*+[F]{} \ar@{-}[rrr]^{i+m_1-2} &&&*+[F]{} \ar@{-}[rrr]^{i+m_1-3} &&&*+[F]{} \ar@{.}[rr] &&*+[F]{} \ar@{-}[rr] ^{i+1}&&*+[F]{a } & &*+[F]{b} \ar@{-}[rr]^{i-1} &&*+[F]{} \ar@{.}[rr]&&*+[F]{} \ar@{-}[rr] ^{i-m_2-1}&&*+[F]{ }} $$
\end{prop}
\begin{proof}
As $\mathcal{\bar{X}}$ has $m_1+m_2-3$ edges, it has 3 connected components. Since $\mathcal{X}$ has two components, there exists a $j$-edge $e_j$ of $\mathcal{X}$ such that $\mathcal{\bar{X}} \cup e_j$ has two components and no cycles. Hence, as in Proposition~\ref{M=m-2} incident edges of $\mathcal{\bar{X}} \cup e_j$ must have consecutive labels.
First suppose that the $j$-edges of $\mathcal{\bar{X}} \cup e_j$ are different $\Gamma_i$-orbits. We claim that $j = i \pm 1$. Assume the contrary. Then there is a path in the first $\Gamma_i$-orbit connecting the block containing $a$ to the block moved by $\rho_j$; the same happens in the second $\Gamma_i$-orbit: there is a path connecting the block containing $b$ to the block moved by $\rho_j$.
$$\[email protected]{ *+[F]{} \ar@{-}[rr]^j &&*+[F]{} \ar@{.}[rr] && *+[F]{} \ar@{-}[rr]^{i-1} &&*+[F]{a} &&*+[F]{b} \ar@{-}[rr]^{i+1} && *+[F]{ } \ar@{.}[rr]& &*+[F]{} \ar@{-}[rr]^j &&*+[F]{} } $$
Assume that $ j > i +1$. Then both of these paths have to contain the label $i+1$, a contradiction. Thus $j=i\pm 1$. Up to duality we may consider $j=i-1$ corresponding to the first possibility for $\mathcal{X}$.
Now consider the $j$-edges of $\mathcal{\bar{X}} \cup e_j$ in the same $\Gamma_i$-orbit; assume it is the first orbit. There is a path in $\mathcal{\bar{X}} \cup e_j$ joining the four blocks swapped by $\rho_j$ which has consecutive labels and no repeating labels other than $j$. Thus this path is a single edge with label $j \pm 1$. Assume that this edge has label $j+1$. At least one of the four vertices of $\mathcal{\bar{X}} \cup e_j$ that are incident to the $j$-edges has degree one. Otherwise, there would be another repeated label. This gives the second possibility for the graph $\mathcal{X}$ where $j=i-m_1-3$ with $m_1\geq 4$.
\end{proof}
\begin{prop}\label{R<=k-3}
If an element of $R$ has a nontrivial action between more than one pair of $\langle M\rangle$-orbits, then $|R|\leq k -3$.
\end{prop}
\begin{proof}
In this case $\mathcal{\bar{Y}}$ has at least three connected components, and still has no cycles. Thus $|R|\leq k-3$.
\end{proof}
In what follows let $L_a$ be the $\langle M\rangle$-orbit containing $a$ and $L_b$ the $\langle M\rangle$-orbit containing $b$.
\begin{prop}\label{M=m-3(2)}
If $|M|=m_1+m_2-3$ then $|R|\leq k_1+k_2-3$.
\end{prop}
\begin{proof}
As $\mathcal{\bar{Y}}$ is a forest and has at least two components, $|R|\leq k-2$.
If $k< k_1+k_2$ then $|R|\leq k_1+k_2-3$. Assume that $k=k_1+k_2$ and that we have
the equality $|R|= k_1+k_2-2$. Then $\mathcal{\bar{Y}}$ has exactly two components and there are at least two $\langle M\rangle$-orbits in each $\Gamma_i$-orbit.
As $|M|=m_1+m_2-3$ we have, up to duality, one of the two possibilities for $\mathcal{X}$ given in Proposition~\ref{M=m-3(1)}.
First suppose that $m_1+m_2> 4$. There are, up to duality, the two possibilities for $\mathcal{X}$ given in Proposition~\ref{M=m-3(1)}. In graph (1) there is only one possibility to connect $L_a$ to another $\langle M\rangle$-orbit, that is using a pair of $(i+1)$-edges. Then in both cases, (1) and (2), there is only one possibility to connect $L_b$ to another $\langle M\rangle$-orbit, that is, with a single edge with label $l=i-m_2-2$ (between vertices of the last block of the second $\Gamma_i$-orbit).
Furthermore, $\rho_l$ swaps exactly one pair of vertices of these $\langle M\rangle$-orbits. Then as $\Gamma$ is even $\rho_l$ must swap another pair of $\langle M\rangle$-orbits, hence we have a contradiction with Proposition~\ref{R<=k-3}.
Now let $m_1+m_2=4$. In this case $\mathcal{X}$ is as in figure (1) of Proposition~\ref{M=m-3(1)} and $M=\{\rho_{i-1}\}$. To connect $L_a$ to another $\langle M\rangle$-orbit, or $L_b$ to another $\langle M\rangle$-orbit, there are only two possibilities for the labels, either $l=i-2$ or $l=i+1$. By Proposition~\ref{R<=k-3} we may assume that either $L_a$ or $L_b$ is $(i-2)$-adjacent to another $\langle M\rangle$-orbit. Then we use the fact that $\Gamma$ is even and Proposition~\ref{R<=k-3} to get a contradiction.
\end{proof}
\begin{prop}\label{MM=m-2}
If $|M| = m_1+m_2-2$ then $|R| \leq k_1+k_2-4$.
\end{prop}
\begin{proof}
Up to duality we may consider $\mathcal{X}$ as in Proposition~\ref{M=m-2}. First the elements of $R$ must fix all the blocks.
Otherwise there is $\rho_j\in M$ such that $\Gamma_j$ is transitive, a contradiction. Let $C$ be the set of generators in $R$ that commute with all elements of $M$. We have that $|R \setminus C| \leq 2$.
There is at most one $\langle M\rangle$-orbit adjacent to $L_a$ in $\mathcal{\bar{Y}}$, and the label of the edge which might connect them is $\rho_f := \rho_{i - m_1}$. Similarly, there is at most one $\langle M\rangle$-orbit adjacent to $L_b$ in $\mathcal{\bar{Y}}$, and the label of the edge which might connect them is $\rho_l := \rho_{i + m_2}$. We denote the $\langle M\rangle$-orbits adjacent to $L_a$ and $L_b$, $L'_a$ and $L'_b$ resp. if they exist.
Furthermore, both $\rho_f$ and $\rho_l$, if they exist, swap a single pair of points in these $\langle M\rangle$-orbits.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{f-1}\ar@{-}[dd]^{f} &&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr] ^{i-1}&& *+[o][F]{a}\ar@{-}[rr] ^i &&*+[o][F]{b} \ar@{-}[rr]^{i+1} && *+[o][F]{ } \ar@{.}[rr]&&*+[o][F]{} \ar@{-}[rr]^{l-1} &&*+[o][F]{} \ar@{-}[dd]^l\\
&&&&&&&&&&&&&&\\
*+[o][F]{} \ar@{-}[rr]^{f-1} &&*+[o][F]{} \ar@{.}[rr]&& *+[o][F]{}\ar@{-}[rr]^{i-1}&& *+[o][F]{} &&*+[o][F]{} \ar@{-}[rr]^{i+1} && *+[o][F]{ } \ar@{.}[rr]&&*+[o][F]{} \ar@{-}[rr]^{l-1} &&*+[o][F]{}
} $$
Since $\Gamma$ is even, $\rho_f$ and $\rho_l$ must both have nontrivial action on a point in another $\langle M\rangle$-orbit. We now consider separately the cases: $|R \setminus C| = 0$, $|R \setminus C| = 1$ and $|R \setminus C| = 2$.
If $R \setminus C = \emptyset$, then $L_a$ and $L_b$ are the unique $\langle M\rangle$-orbits, $R = \emptyset$ and $k=2$. Thus $|R|=0\leq k_1+k_2-4$.
Suppose that $R \setminus C=\{\rho_f\}$. In this case $L_b$ coincide with the second $\Gamma_i$-orbit. As $\mathcal{Y}$ has at least two $f$-edges, by Proposition~\ref{R<=k-3}, $|R| \leq k-3$. In addition, $k\leq k_1+k_2-1$, thus $|R| \leq k-3\leq k_1+k_2-4$.
Now let $ R \setminus C=\{ \rho_f, \rho_l \}$. We may assume $\mathcal{\bar{Y}}$ with an $f$-edge between $L_a$ and $L'_a$, and with an $l$-edge between $L_b$ and $L'_b$. Since $\Gamma$ is even, both $\rho_f$ and $ \rho_l$ act non-trivially on a point in a $\langle M\rangle$-orbit other than $L_a, L_b, L'_a$ and $L'_b$. By Proposition~\ref{R<=k-3},
$\mathcal{\bar{Y}}$ has at least three components. Furthermore, to have three components there must exist a pair of $\langle M\rangle$-orbits $\{L,L'\}$ such that both $ \rho_l$ and $\rho_f$ swap a point in $L$ with a point in $L'$. Indeed, $\rho_l$ acts trivially on the points not in $L, L', L_b$ or $L_b'$ and $\rho_f$ acts trivially on the points not in $L, L', L_a$ or $L_a'$. Let us assume that $L$ and $L'$ are both in the first $\Gamma_i$-orbit. Then $\rho_l$ swaps $L$ and $L'$ entirely.
First suppose $L = L'_a$. Then, both $\rho_f$ and $\rho_l$, act nontrivially on a point in $L'_a$. Thus there is an alternating square, with labels $f$ and $l$, containing this point. Hence $\rho_l$ acts nontrivially on a point in $L_a$, which we have shown is impossible. Consequently $L \neq L'_a$, as well as $L'\neq L'_a$.
Since $\Gamma_i$ has only two orbits, there is a generator $\rho_j\in R$ that sends a point in either $L$ or $L'$ to a point not in these two $\langle M\rangle$-orbits; without loss of generality we may assume that $\rho_j$ sends a point in $L'$ to a point in a $\langle M\rangle$-orbit which we denote $L''$. Then the unique possibility is $j = l + 1$. However, then $\rho_j$ commutes with all the elements of $M$ which act between blocks in the first $\Gamma_i$-orbit. Thus $\rho_j$ swaps $L'$ and $L''$. This guarantees an alternating square with labels $j$ and $f$, with $\rho_f$ acting nontrivially on $L''$. This implies that $L''=L'_a$. Moreover this forces $\rho_j$ to have a nontrivial action on $L_a$, a contradiction.
\end{proof}
\begin{prop}\label{bothimp}
If $A$ and $B$ are both imprimitive then $r \leq \frac{n -1}{2}$.
\end{prop}
\begin{proof}
By Propositions~\ref{M=m-3(2)} and~\ref{M=m-2}, we have $r = |R| + | M |+1\leq k_1+k_2+ m_1+m_2 - 5$. As $k_1+ m_1 + k_2 + m_2 -5 \leq \frac{k_1m_1 +k_2m_2 - 1}{2}$, we conclude that $r \leq \frac{n -1}{2}$.
\end{proof}
\subsection{Case 2: $J_A$ and $J_B$ are intervals and $i\notin\{ 0,r-1\}$}
We first recall two propositions on sggi that can be found in \cite{flm2} that we use to deal with this case.
\begin{prop}\cite[Proposition 3.3]{flm2}\label{genAn}
Let $\Phi = \langle \alpha_0,\,\ldots,\,\alpha_{d-1} \rangle$ be a transitive permutation group acting on the set of points $\{ 1,\ldots, n \}$ with $n \geq 5$, and let $\Phi^* = \langle \alpha_0,\,\ldots,\,\alpha_{d-1},\,\alpha_d,\, \alpha_{d+1} \rangle$, where
\[\begin{array}{rl}
\alpha_r&=(i,n+1)(n+2,n+3) \mbox{ for some } i \in \{1,\ldots,n\}\\
\alpha_{r+1}&=(n+1,n+2)(n+3,n+4).
\end{array}\]
Then $\Phi^*$ is isomorphic to $S_{n+4}$ if it contains an odd permutation, and to $A_{n+4}$ otherwise .
\end{prop}
The term sesqui-extension was first introduced in \cite{flm}. Let us recall its meaning.
Let $\Phi = \langle \alpha_0,\ldots,\alpha_{d-1} \rangle$ be a sggi, and let $\tau$ be an involution in a supergroup of $\Phi$ such that $\tau \not \in \Phi$ and $\tau$ commutes with all of $\Phi $. For fixed $k$, we define the group $\Phi^*= \langle \alpha_i \tau^{\eta_i}\,|\, i\in \{0,\,\ldots,\,d-1\} \rangle$ where $\eta_i = 1$ if $i=k$ and 0 otherwise, the {\it sesqui-extension} of $\Phi $ with respect to $\alpha_k$ and $\tau$.
\begin{prop} \label{sesqui}\cite[Proposition 5.4]{flm2}
If $\Phi =\left<\alpha_i\,|\, i=0,\,\ldots,\,d-1\right>$ and $\Phi^*=\langle \alpha_i \tau^{\eta_i}\,|\, i\in \{0,\,\ldots,\,d-1\} \rangle$ is a sesqui-extension of $\Phi$ with respect to $\alpha_k$, then:
\begin{enumerate}
\item $ \Phi^* \cong \Phi$ or $\Phi^* \cong \Phi\times \langle \tau\rangle\cong\Phi\times 2$.
\item whenever $\tau \notin \Phi^*$, $\Phi$ is a string C-group if and only if $\Phi^*$ is a string C-group.
\end{enumerate}
\end{prop}
In this case we may assume that $A=\Gamma_{<i}$ and $B=\Gamma_{>i}$.
As $A$ or $B$ can have small degree $\leq 11$, in what follows we list all primitive even string C-groups of small degree, with intransitive maximal parabolic subgroups, having rank $r\geq\frac{n-1}{2}$.
\begin{prop}\label{smalln}
Let $\Phi$ be string C-group with a connected Coxeter diagram isomorphic to an even primitive group of degree $n\leq 11$. If $\Phi_j$ is intransitive for every $j\in\{0,\ldots,r-1\}$ then either $r \leq \frac{n-2}{2}$ or $\Phi$ has one of the permutation representation graphs given in Table~\ref{small}.
\begin{small}
\begin{table}
\begin{tabular}{|ccc|}
\hline
(1)&$D_{10}$& $ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} }$\\
(2)&$L_2(5)$ & $ \xymatrix@-1pc{
*+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} }$\\
(3)&$L_2(5)$& $ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^2_0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} }$\\
(4)&$A_9$ &$ \xymatrix@-1pc{ *+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2& *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2& *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} } $ \\
(5)&$A_9$ &$ \xymatrix@-1pc{ *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^1& *+[o][F]{} \ar@{-}[r]^0 \ar@{-}[d]_2 & *+[o][F]{} \ar@{=}[d]^1_2\\
&&&&& *+[o][F]{} \ar@{-}[r]_0 & *+[o][F]{} } $ \\
(6)&$A_9$ &$ \xymatrix@-1pc{ *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^1 \ar@{-}[d]_3& *+[o][F]{} \ar@{-}[r]^0 \ar@{-}[d]_3& *+[o][F]{}\ar@{-}[d]_3\\
&*+[o][F]{} \ar@{-}[r]_2 & *+[o][F]{} \ar@{-}[r]_1 & *+[o][F]{} \ar@{-}[r]_0 & *+[o][F]{} } $ \\
(7)&$A_9$& $ \xymatrix@-1pc{ *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^1 \ar@{-}[d]_3& *+[o][F]{} \ar@{-}[r]^0 \ar@{=}[d]_3^2& *+[o][F]{}\ar@{=}[d]_3^2\\
&*+[o][F]{} \ar@{-}[r]_2 & *+[o][F]{} \ar@{-}[r]_1 & *+[o][F]{} \ar@{-}[r]_0 & *+[o][F]{} } $ \\
(8)&$A_9$& $ \xymatrix@-1pc{ *+[o][F]{} \ar@{-}[r]^1 &*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 \ar@{-}[d]_0& *+[o][F]{} \ar@{-}[r]^3 \ar@{-}[d]_0& *+[o][F]{}\ar@{-}[d]_0\\
&&& *+[o][F]{} \ar@{-}[r]_2 & *+[o][F]{} \ar@{-}[r]_3 & *+[o][F]{} } $\\
(9)&$A_9$ & $ \xymatrix@-1pc{ *+[o][F]{} \ar@{=}[r]^0_2&*+[o][F]{} \ar@{-}[r]^1 &*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 \ar@{-}[d]_1& *+[o][F]{}\ar@{-}[d]^1\\
&&&& & *+[o][F]{} \ar@{-}[r]_3 & *+[o][F]{} } $\\
(10)&$A_{10}$ & $ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^0& *+[o][F]{} \ar@{-}[r]^1& *+[o][F]{} \ar@{=}[r]^2_0 & *+[o][F]{} }$\\
(11)&$A_{10}$& $ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4& *+[o][F]{} \ar@{-}[r]^3 &*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} }$\\
(12)&$A_{10}$& $ \xymatrix@-1pc{ *+[o][F]{} \ar@{-}[r]^4&*+[o][F]{} \ar@{-}[r]^3 &*+[o][F]{} \ar@{=}[r]^4_2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^1& *+[o][F]{} \ar@{-}[r]^0 \ar@{-}[d]_2& *+[o][F]{}\ar@{=}[d]^1_2\\
&&&&&& *+[o][F]{} \ar@{-}[r]_0 & *+[o][F]{} } $\\
(13)&$A_{10}$&$ \xymatrix@-1pc{ *+[o][F]{} \ar@{=}[r]^0_2&*+[o][F]{} \ar@{-}[r]^1 &*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3& *+[o][F]{} \ar@{-}[r]^4 \ar@{-}[d]_2& *+[o][F]{}\ar@{=}[d]^3_2\\
&&&&&& *+[o][F]{} \ar@{-}[r]_4 & *+[o][F]{} } $\\
(14)&$A_{11}$ & $ \xymatrix@-1pc{
*+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4& *+[o][F]{} \ar@{=}[r]^3_5 &*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} }$\\
(15)&$A_{11}$& $ \xymatrix@-1pc{
*+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4& *+[o][F]{} \ar@{-}[r]^5 &*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{=}[r]^5_3 & *+[o][F]{} }$\\
(16)&$A_{11}$ & $ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4& *+[o][F]{} \ar@{=}[r]^3_5 &*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} }$\\
\hline
\end{tabular}
\caption{Even transitive string C-groups of degree $m$ with connected Coxeter diagram having intransitive maximal parabolic subgroups and rank $\geq \frac{m-1}{2}$ .}\label{small}
\end{table}
\end{small}
\end{prop}
\begin{proof}
We used \textsc{Magma} to get this result.
\end{proof}
\begin{prop}\label{small123}
Let $j\in \{0,\ldots,\, r-1\}$. Suppose that $\Gamma_{<j}$ is transitive on $m$ points and fixes the remaining $m-n$ points.
If $\Gamma_{<j}$ is a primitive group of degree $m<12$ and $j\geq \frac{m-1}{2}$ then it must have permutation representation graph (1), (2) or (3) of Table~\ref{small}.
\end{prop}
The dual version of this proposition is also true.
\begin{proof}
By Proposition~\ref{smalln}, $\Gamma_{<j}$ has one of the 16 permutation representation graphs given in Table~\ref{small}.
Let $X:=\{1,\ldots,n\}\setminus \mathrm{Fix}(\Gamma_{<j})$ with $|X|=m$.
Then there exists a generator $\rho_k$ of $\Gamma$, with $k\geq j$, such that $c\rho_k=d$ with $c\in X$ and $d\not \in X$.
As $\Gamma_{<j}$ is primitive, $k$ must be consecutive with $j-1$.
Moreover, $c$ has degree two with the edges incident to $c$ having labels $j-1$ and $j$.
In graphs (8), (9), (13) and (15) there is no such vertex $c$. Let us consider the remaining graphs.
Now let (4) be the permutation representation graph of $\Gamma_{<j}$. Then $\Gamma_{<j}=\Gamma_{<4}\cong A_9$ and $\Gamma_{>1}\cong A_{n-4}$ by Proposition~\ref{genAn} (note that $n-4\geq 5$), thus $\Gamma_{>1} \cap \Gamma_{<4}\cong A_5$. However, $\langle \rho_2, \rho_3 \rangle$ is a dihedral group. Consequently $\Gamma$ does not satisfy the intersection condition, a contradiction.
Only in cases (11) and (16) there are two possibilities for the vertex $c$, but as in case (4), case (16) is self-dual.
Using similar arguments, summarised below, we conclude that $\Gamma_{<j}$ cannot be any of the graphs of Table~\ref{small} except graphs (1), (2) and (3). Let (11)* denote the dual of (11) with its labels interchanged by $k \leftrightarrow 4-k$.
\begin{center}
\begin{tabular}{cl}
(5), (6), (7): &$\Gamma_{>1}\cong 2\times A_{n-4}$ or $\Gamma_{>1}\cong A_{n-4}$ (by Proposition~\ref{sesqui}); \\[5pt]
&$\,\Gamma_{<4}\cong A_9$; $\,\Gamma_{>1} \cap \Gamma_{<4}\cong A_5\not \cong \langle \rho_2, \rho_3 \rangle$.\\[5pt]
(10), (11), (11)*, (12): &$\Gamma_{>2}\cong A_{n-5}$; $\,\Gamma_{<5}\cong A_{10}$; $\,\Gamma_{>2} \cap \Gamma_{<5}\cong A_5\not \cong \langle \rho_3, \rho_4 \rangle$.\\[5pt]
(14), (16):& $\Gamma_{>3}\cong A_{n-6}$; $\,\Gamma_{<6}\cong A_{11}$; $\,\Gamma_{>3} \cap \Gamma_{<6}\cong A_5\not \cong \langle \rho_4, \rho_5 \rangle$.
\end{tabular}
\end{center}
\end{proof}
\begin{prop}\label{RI}
Let $\Phi=\langle \alpha_0,\ldots,\alpha_{d-1}\rangle$ be a transitive sggi embedded into $S_a\wr S_b$ with $X$ being the set of generators of $\Phi$ generating the block action independently.
If $\Phi_j$ is intransitive for all $j\in\{0,\ldots,d-1\}$, then
\begin{enumerate}
\item $d\leq |X|+a-1$;
\item $d\leq \frac{ab-2}{2}$ for $a,b\neq 2$ and $ab\neq 9$;
\item if $(a,d)=(2,b)$ or $(b,d)=(2,a)$ or $(a,b)=(3,3)$ then $\Phi$
either contains an odd
permutation or has disconnected diagram.
\end{enumerate}
\end{prop}
\begin{proof}
Consider a partition $P$ of $\{1,\ldots,ab\}$ determined by the orbits of $\langle X\rangle$.
Now let $\mathcal{X}$ be a graph with $|P|$ vertices corresponding the partitions of $P$ and with exactly one $j$-edge for each generator of $\Phi$ that is not in $X$, connecting two partitions in different $\Phi_j$-orbits. This graph has no cycles and $|P|\leq a$, thus
$d-|X|\leq a-1$. Hence, $d\leq |X|+a-1$. Moreover as $X$ generate the bock action independenty $|X|\leq b-1$. Hence $d\leq \leq a+b-2 \frac{ab-2}{2}$ for $a,b\neq 2$ and $ab\neq 9$.
Now suppose that $(a,d)=(2,b)$ or $(b,d)=(2,a)$ and $\Phi$ is even, up to duality, $\Phi$ has the following permutation representation graph, and
$$
\[email protected]{*+[o][F]{} \ar@{-}[rr]^1 \ar@{-}[dd]_0 &&
*+[o][F]{}\ar@{-}[rr]^2\ar@{-}[dd]_0 &&
*+[o][F]{} \ar@{-}[rr]^3\ar@{-}[dd]_0 &&
*+[o][F]{} \ar@{.}[rr]\ar@{-}[dd]_0 &&
*+[o][F]{}\ar@{-}[rr]^{d-1}\ar@{-}[dd]_0 &&
*+[o][F]{}\ar@{-}[dd]_0 \\
&& && && && &&\\
*+[o][F]{} \ar@{-}[rr]_1 &&
*+[o][F]{}\ar@{-}[rr]_2&&
*+[o][F]{} \ar@{-}[rr]_3&&
*+[o][F]{} \ar@{.}[rr]&&
*+[o][F]{}\ar@{-}[rr]_{d-1}&&
*+[o][F]{}\\ }
$$
Thus $\Phi$ has a disconnected diagram.
\end{proof}
Using induction over $n$ and Proposition~\ref{RI} we get the following result.
\begin{prop}\label{induction}
Suppose that $\Phi$ is string C-group generated by involutions of rank $d$, with connected diagram, having all maximal parabolic subgroups intransitive. If $\Phi$ is a transitive even group of degree $m$ with $12\leq m< n$ then, $d\leq \frac{m-1}{2}$. Moreover if $d= \frac{m-1}{2}$ then $\Phi$ is the alternating group $A_m$.
\end{prop}
\begin{proof}
If $\Phi$ is primitive and not the alternating group then, by Proposition~\ref{prim}, $d\leq \frac{m-2}{2}$. If $\Phi$ is imprimitive, then it is embedded into a group $S_a\wr S_b$ with $ab=m$ and with the block action being generated by at most $b-1$ elements.
As the maximal parabolic subgroups of $\Phi$ are intransitive we may use Proposition~\ref{RI} to get
\[d\leq \frac{m-2}{2}.\]
for $a,b\neq 2$. If either $a=2$ or $b=2$ then, as $\Phi$ is even and has a connected diagram, $d< \frac{m}{2}$ with $m$ even, hence $d\leq \frac{m}{2}$.
Finally if $\Phi\cong A_m$ with $12\leq m< n$, we then get the result by induction on $n$.
\end{proof
\begin{prop}\label{small(1)}
Suppose that $\Gamma_{<3}$ is not one of the string $C$-groups (2) or (3) Table~\ref{small}.
If $A$ is the string $C$-group (1) of Table~\ref{small}, then $r \leq \frac{n-1}{2}$. The same result holds for $B$.
\end{prop}
\begin{proof}
Consider first $n-5 < 12$. If $r-3\geq \frac{(n-5)-1}{2}$ then, by Proposition~\ref{small123}, $B\,(=\Gamma_{>2})$ is one of the small examples (1), (2) or (3) of Table~\ref{small}, thus $n<12$, a contradiction. Hence $r-3<\frac{(n-5)-1}{2}$. Let $n-5 \geq 12$.
Suppose first that $\rho_2$ has a trivial action on the first $\Gamma_i$- orbit. In this case $\Gamma_{>1}$ is an even transitive group on $n-4$ points.
By Proposition~\ref{sesqui} $\Gamma_1 \cong \langle \rho_0\rangle \times \Gamma_{>1}$. Now, by Proposition~\ref{induction}, we get that $r-2 \leq \frac{(n-4)-1}{2}$ and hence $r\leq \frac{n-1}{2}$.
Now suppose that $\rho_2$ has a nontrivial action on the first block. As $\Gamma_{<3}$ cannot be one of the string C-groups (2) or (3) of Table~\ref{small}, it is a sesqui extension of it (with respecto to $\rho_2$).
If $B$ is not the alternating group, then by Proposition~\ref{induction}, we get that $r-3 \leq \frac{(n-5)-2}{2}$. We need only to consider the case $B=\Gamma_{>2}\cong A_{n_2}$.
Let $\Phi$ be either the string C-group (2) or (3) of Table~\ref{small} and $\tau$ be the action of $\rho_2$ in the second $\Gamma_2$-orbit. As either $\tau=(\rho_1\rho_2)^3$ or $\tau=(\alpha_1\alpha_2)^5$ (according to each case (1) or (2)), $\Gamma_{<3}$ is isomorphic to $\langle \tau\rangle\times \Phi$. But then, as $\Gamma_{>2}\cong A_{n_2}$, $\tau\in \Gamma_{>2}\cap \Gamma_{<3}$, a contradiction.
Hence $r \leq \frac{n-1}{2}$.
\end{proof}
\begin{prop}\label{inot0}
Suppose neither $\Gamma_{<3}$ nor $\Gamma_{>r-4}$ is one of the string $C$-groups (2) or (3) of Table~\ref{small}, or their duals.
If $A$ and $B$ are not both the alternating groups, then $r \leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
By Proposition~\ref{small(1)} it may be assumed that neither $A$ nor $B$ is the string $C$-group (1) of Table~\ref{small}.
If $A$ and $B$ are not alternating groups then by Propositions~\ref{induction} and \ref{smalln}, $i \leq \frac{n_1-2}{2}$ and $r-1-i \leq \frac{n_2-2}{2}$. If $A$ is the alternating group, then either $n\leq 12$ and $i\leq \frac{n_1-2}{2}$ (by Proposition~\ref{smalln}) or, $n_1\geq 12$ and then, by Proposition~\ref{induction} (induction) we conclude that $i\leq \frac{n_1-1}{2}$. Analogously if $B$ is the alternating group then $r-1-i\leq \frac{n_2-1}{2}$ . In any case if $A$ and $B$ are not both the alternating groups, $r \leq 1 + \frac{n_1+n_2-3}{2}= \frac{n-1}{2}$.
\end{proof}
\begin{prop}\label{bothAn}
Suppose neither $\Gamma_{<3}$ nor $\Gamma_{>r-4}$ is one of the string $C$-groups (2) or (3) of Table~\ref{small}, or their duals.
If $A$ and $B$ are both alternating groups then $r \leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
In this case $\Gamma_{<i+1}$ is a sesqui extension of a sggi $\Phi$ with respect to $\rho_i$, where $\Phi$ is a group of degree $n_1+1$.
By Proposition~\ref{sesqui} $\Gamma_{<i+1}$ is isomorphic either to $2\times \Phi$ or $\Phi$.
Suppose that $\Gamma_{<i+1}\cong 2\times \Phi$. In that case there is an even permutation $\tau$ on the second $\Gamma_i$-orbit that belongs to $\Gamma_{<i+1}$.
As $\Gamma_{>i}\cong A_{n_2}$, $\tau\in \Gamma_{>i}$ and therefore $\tau\in \Gamma_{<i+1}\cap \Gamma_{>i}$, a contradiction.
Hence $\Gamma_{<i+1}\cong \Phi$ and $\Phi$ is itself a string $C$-group.
Using the same argument $\Gamma_{>i-1}$ is also isomorphic to a transitive group $\Psi$ of degree $n_2+1$.
Moreover either $\Phi$ or $\Psi$ is a even group. Suppose $\Phi$ is even.
Then, as $A$ is not one of the string C-groups (1), (2) or (3), either $n_1<12$ and $i\leq \frac{n_1-2}{2}$, or $n_1+1\geq 12$.
In latest case, by Proposition~\ref{induction} $i+1\leq \frac{(n_1+1)-1}{2}$.
In addition, $r-1-i\leq \frac{n_2-1}{2}$, hence $r \leq \frac{n-1}{2}$.
\end{proof}
\begin{prop}\label{Case2}
Suppose neither $\Gamma_{<3}$ nor $\Gamma_{>r-4}$ is one of the string $C$-groups (2) or (3) of Table~\ref{small}, or their duals.
Let $i\notin\{0,r-1\}$. If $A\cong \Gamma_{<i}$ and $B=\Gamma_{>i}$ then $r \leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
This is a consequence of Propositions~\ref{inot0} and \ref{bothAn}.
\end{proof}
To complete this case we still need to deal with $\Gamma_{<3}$, $\Gamma_{>r-4}$, or both, being one of the string $C$-groups (2) or (3) of Table~\ref{small}.
This is included in Case 3, at the end.
\subsection{Case 3: The remaining cases}
Assume in this case that $J_B$ is either empty or an interval. As before let $\mathcal{G}$ be the permutation representation graph of $\Gamma$.
\begin{prop}\label{path}
If $e$ is an $f$-edge of $\mathcal{G}$ not in an alternating square, then any path (not containing another $f$-edge) from $e$ to an edge with label $l$, with $l<f$ (resp. $l>f$), contains all labels between $l$ and $f$.
Moreover, there exists a path from $e$ to an $l$-edge, that is fixed by $\Gamma_{>l}$ (resp. $\Gamma_{<l}$).
\end{prop}
\begin{proof}
Consider a path starting in $e$ and containing an $l$-edge. Let $f<l$.
Suppose that none of the edges of the path has label $k$, for some $f<k<l$.
Then in this path there is an edge with label $<k$ meeting an edge with label $u>k$.
Suppose that this is the first time in the path that this happens.
Then there is an alternating square, containing $e$ and a $u$-edge, a contradiction, as shown in the following figure.
$$ \[email protected]{*+[o][F]{}\ar@{.}[rr] &&*+[o][F]{}&&&&&& *+[o][F]{} \ar@{.}[ddll]_u\ar@{.}[ddrr]&&&&&&&&\\
&&&& \ldots&&&&&&\\
*+[o][F]{} \ar@{-}[rr]_f \ar@{.}[uu]^u&&*+[o][F]{} \ar@{.}[uu]^u\ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{.}[rr]&&*+[o][F]{} \ar@{-}[rr]^u && *+[o][F]{} \ar@{~}[rrrr] &&&&*+[o][F]{}\ar@{-}[rr]^l&&*+[o][F]{}
}
$$
\end{proof}
\begin{prop}\label{i=0andh=r-1}
Let $i=0$.
If $\rho_{r-1}$ acts non-trivially on both $\Gamma_i$-orbits, then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
By Proposition~\ref{path} there are two paths, one in the first and the other in the second $\Gamma_i$-orbit, each containing all labels from $1$ to $r-1$.
Thus $2(r-1)+1\leq n-1$. Hence $r\leq\frac{n}{2}$. Suppose we have the equality. In that case the paths have consecutive labels, as in the following figure and $n$ is precisely the number of vertices of the two paths.
$$\[email protected]{*+[o][F]{}\ar@{-}[rr]^{r-1} && *+[o][F]{} \ar@{-}[rr]^{r-2}&&*+[o][F]{} \ar@{.}[rr] && *+[o][F]{}\ar@{-}[rr]^1 && *+[o][F]{}\ar@{-}[rr]^0 &&*+[o][F]{} \ar@{-}[rr]^1&& *+[o][F]{}\ar@{.}[rr]&& *+[o][F]{} \ar@{-}[rr]^{r-2}&& *+[o][F]{} \ar@{-}[rr]^{r-1}&&*+[o][F]{} }
$$
But then there is no place for extra $i$-edges unless $r=3$ (which gives the bound trivially as $n\geq 12$).
\end{proof}
\begin{prop}\label{i=0andh<>r-1}
Let $i=0$. Let $h\neq r-1$ be the maximal label such that $\rho_h$ acts non-trivially on both $\Gamma_i$-orbits.
There exists a set of vertices $X$, contained in the $\Gamma_i$-orbit fixed by $\rho_{r-1}$, such that $h\leq \frac{n-|X|-1}{2}$ and $\Gamma_{>h}$ fixes $\{1,\ldots, n\}\setminus X$. Moreover if $h= \frac{n-|X|-1}{2}$ then $\Gamma_{<h}$ has the following permutation representation graph, where the black dots represent the vertices of $\{1,\ldots, n\}\setminus X$.
$$\[email protected]{\bullet\ar@{-}[rr]^h && \bullet \ar@{-}[rr]^{h-1}&&\bullet \ar@{.}[rr] && \bullet\ar@{-}[rr]^1 && \bullet\ar@{-}[rr]^0 &&\bullet \ar@{-}[rr]^1&& \bullet \ar@{.}[rr]&& \bullet \ar@{-}[rr]^{h-1}&& \bullet \ar@{-}[rr]^{h}&&*+[o][F]{} }
$$
\end{prop}
\begin{proof}
By Proposition~\ref{path} there exist a path $\mathcal{P}_1$ from the $h$-edge in the first $\Gamma_i$-orbit to the vertex $a$, and a path $\mathcal{P}_2$ from the $h$-edge in the second $\Gamma_i$-orbit to the vertex $b$, each of them containing all labels from $1$ to $h-1$, and fixed by $\Gamma_{>h}$.
In that case, the two paths give us $2h+1 \leq |\mathrm{Fix}(\Gamma_{>h})|=n-|X|$ with $X := \{1, \ldots, n\}\setminus \mathrm{Fix}(\Gamma_{>h})$ as required. When equality holds, the permutation representation graph is the one given in this proposition.
\end{proof}
\begin{prop}\label{<jon>jtran}
Let $X:= \{1,\ldots,n\} \setminus \mathrm{Fix}(\Gamma_{>j})$.
If $\Gamma_{>j}$ is transitive on $X$ and there exists a permutation $\rho_l$ with $l<j$ acting non trivially on $X$, then $r-1-j\leq \frac{|X|}{2}-1$. Moreover if $r-1-j= \frac{|X|}{2}-1$ then, $\Gamma_{>j}$ has one of the following permutation representation graphs for some $k\in\{j+1,\ldots, r-1\}$.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{j+1} \ar@{.}[dd]_l &&
*+[o][F]{}\ar@{.}[rr] \ar@{.}[dd]_l&&
*+[o][F]{} \ar@{-}[rr]^k\ar@{.}[dd]_l&&
*+[o][F]{} \ar@{-}[rr]^{k+1}\ar@{.}[dd]_l&&
*+[o][F]{}\ar@{-}[rr]^{k+2}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k \ar@{.}[rr] &&
*+[o][F]{} \ar@{-}[rr]^{r-1} \ar@{.}[dd]_l \ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k \\
&& && && && && && &&\\
*+[o][F]{} \ar@{-}[rr]_{j+1} && *+[o][F]{}\ar@{.}[rr] &&*+[o][F]{} \ar@{-}[rr]_k && *+[o][F]{}\ar@{-}[rr]_{k+1}&& *+[o][F]{}\ar@{-}[rr]_{k+2} &&*+[o][F]{} \ar@{.}[rr] &&*+[o][F]{}\ar@{-}[rr]_{r-1} && *+[o][F]{}}
$$
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{j+1} \ar@{.}[dd]_l \ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[rr] \ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{} \ar@{-}[rr]^{k-2}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{} \ar@{-}[rr]^{k-1}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k&&
*+[o][F]{}\ar@{-}[rr]^{k}\ar@{.}[dd]_l &&
*+[o][F]{}\ar@{.}[dd]_l \ar@{.}[rr] &&
*+[o][F]{} \ar@{-}[rr]^{r-1} \ar@{.}[dd]_l &&
*+[o][F]{}\ar@{.}[dd]_l\\
&& && && && && && &&\\
*+[o][F]{} \ar@{-}[rr]_{j+1} && *+[o][F]{}\ar@{.}[rr] &&*+[o][F]{} \ar@{-}[rr]_{k-2} && *+[o][F]{}\ar@{-}[rr]_{k-1}&& *+[o][F]{}\ar@{-}[rr]_k &&*+[o][F]{} \ar@{.}[rr] &&*+[o][F]{}\ar@{-}[rr]_{r-1} && *+[o][F]{}}
$$
\end{prop}
\begin{proof}
If $\rho_l$ acts non trivially on $X$ then $\Gamma_{>j}$ must be imprimitive with blocks of size two, with $\rho_l$ swapping all pairs of vertices inside the blocks.
Let $M$ an independent generating set for the block action of $\Gamma_{>j}$ and suppose that the group generated by $M$ is intransitive on $X$. Then there must be two orbits on $X$ under the action of $\langle M \rangle$ and some $k>j$ such that $\rho_k$ swaps points of distinct orbits. But then, as $\rho_l \in \Gamma_k$, $\Gamma_k$ is transitive on $X$.
Moreover, $\rho_k$ only moves points in $X$ which is the union of two $\Gamma_k$-orbits. These two orbits are already fused by $\rho_l\in\Gamma_k$. Therefore, $\Gamma_k$ has to be transitive on $\{1,\ldots, n\}$, a contradiction.
So the action of $\langle M\rangle$ must be transitive on $X$.
If $|M|=\frac{|X|}{2}-1$ there are only the given possibilities for $\Gamma_{>j}$.
\end{proof}
\begin{comment}
We also have the dual version of the previous result.
\begin{prop}\label{dual<jon>jtran}
Let $X:= \{1,\ldots,n\} \setminus \mathrm{Fix}(\Gamma_{<j})$.
If $\Gamma_{<j}$ is transitive on $X$ and there exists a permutation $\rho_l$ with $l>j$ acting non trivially on $X$, then $j\leq \frac{|X|}{2}-1$. Moreover if $j=\frac{|X|}{2}-1$ then $\Gamma_{<j}$ has one of the following permutation representation graph for some $k\in\{0,\ldots, j-1\}$.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{j-1} \ar@{.}[dd]_l &&
*+[o][F]{}\ar@{.}[rr] \ar@{.}[dd]_l&&
*+[o][F]{} \ar@{-}[rr]^k\ar@{.}[dd]_l&&
*+[o][F]{} \ar@{-}[rr]^{k-1}\ar@{.}[dd]_l&&
*+[o][F]{}\ar@{-}[rr]^{k-2}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k \ar@{.}[rr] &&
*+[o][F]{} \ar@{-}[rr]^0 \ar@{.}[dd]_l \ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k \\
&& && && && && && &&\\
*+[o][F]{} \ar@{-}[rr]_{j-1} && *+[o][F]{}\ar@{.}[rr] &&*+[o][F]{} \ar@{-}[rr]_k && *+[o][F]{}\ar@{-}[rr]_{k-1}&& *+[o][F]{}\ar@{-}[rr]_{k-2} &&*+[o][F]{} \ar@{.}[rr] &&*+[o][F]{}\ar@{-}[rr]_0 && *+[o][F]{}}
$$
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{j+1} \ar@{.}[dd]_l \ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[rr] \ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{} \ar@{-}[rr]^{k+2}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{} \ar@{-}[rr]^{k+1}\ar@{.}[dd]_l\ar@<.3ex>@{-}[dd]^k&&
*+[o][F]{}\ar@{-}[rr]^{k}\ar@{.}[dd]_l &&
*+[o][F]{}\ar@{.}[dd]_l \ar@{.}[rr] &&
*+[o][F]{} \ar@{-}[rr]^0\ar@{.}[dd]_l &&
*+[o][F]{}\ar@{.}[dd]_l\\
&& && && && && && &&\\
*+[o][F]{} \ar@{-}[rr]_{j+1} && *+[o][F]{}\ar@{.}[rr] &&*+[o][F]{} \ar@{-}[rr]_{k+2} && *+[o][F]{}\ar@{-}[rr]_{k+1}&& *+[o][F]{}\ar@{-}[rr]_{k} &&*+[o][F]{} \ar@{.}[rr] &&*+[o][F]{}\ar@{-}[rr]_0 && *+[o][F]{}}
$$
\end{prop}
\begin{proof}
This is just the dual of Proposition~\ref{<jon>jtran}.
\end{proof}
\end{comment}
In what follows we suppose that $i > 0$ and that all generators acting on the second $\Gamma_i$-orbit have labels $> i$. As $J_A$ is not an interval, $\Gamma_{<i}$ can either transitive or intransitive on the first $\Gamma_i$-orbit. We consider these cases separately.
\begin{prop}\label{<itran}
Let $r>\frac{n-1}{2}$. Let $h>i$ be the maximal label of a permutation acting non-trivially on both $\Gamma_i$-orbits.
If $\Gamma_{<i}$ is transitive on the first $\Gamma_i$-orbit then $h=i+1$, $h<r-1$ and there exists a set of vertices $X$, contained in the second $\Gamma_i$-orbit, such that $h\leq \frac{n-|X|-1}{2}$ and $\Gamma_{>h}$ fixes $\{1,\ldots, n\}\setminus X$.
Moreover if $h= \frac{n-|X|-1}{2}$ then $\Gamma_{<h}$ (with $h=i+1$) has the following permutation representation graph for some $k\in\{2,\ldots,i-1\}$ where the black dots represent the vertices of $\{1,\ldots, n\}\setminus X$.
$$
\[email protected]{\bullet\ar@{-}[rr]^0 \ar@{=}[d]_{i+1}^k && \bullet \ar@{-}[rr]^1\ar@{=}[d]_{i+1}^k&&\bullet\ar@{=}[d]_{i+1}^k\ar@{.}[rr] && \bullet\ar@{-}[rr]^{i-1}\ar@{-}[d]_{i+1} && \bullet\ar@{-}[rr]^i \ar@{-}[d]_{i+1} &&\bullet \ar@{-}[rr]^{i+1}&&*+[o][F]{} \\
\bullet\ar@{-}[rr]_0 && \bullet \ar@{-}[rr]_1&&\bullet \ar@{.}[rr] && \bullet\ar@{-}[rr]_{i-1}&& \bullet&&&&
}
$$
\end{prop}
\begin{proof}
In this case, $\Gamma_{<i}$ is transitive on the first $\Gamma_i$-orbit $O_1$, which has size $n_1$. Moreover there exists $h>i$ such that $\rho_h$ acts non-trivially on $O_1$.
This action is fixed-point-free and hence $\rho_h$ moves $a$ and $n_1$ is even. But since the $i$-edge $\{a,b\}$ is not in a square by hypothesis, $h=i+1$ and $\Gamma_{>h}$ acts trivially on $O_1$.
Moreover by the dual of Proposition~\ref{<jon>jtran} we have that $i\leq \frac{n_1}{2}-1$.
Thus $h=i+1\leq \frac{n_1}{2}$.
As $J_B$ is an interval, $\rho_{i+1}$ is the unique permutation acting nontrivially on $b$. If $h\neq r-1$, $\Gamma_{>h}$ fixes $O_1$ and a vertex in the second $\Gamma_i$-orbit, so at least $n_1+1$ vertices altogether.
Then $h\leq \frac{|\mathrm{Fix}(\Gamma_{>h})|-1}{2}\leq \frac{n-|X|-1}{2}$. The equality cannot occur as it would imply that $n-|X| = n_1+1$ is even, a contradiction. So $h\leq \frac{n-|X|-1}{2}$ as wanted.
Moreover, when the equality holds, by the dual of Proposition~\ref{<jon>jtran}, the only possibility is the permutation representation graph given in the statement of this proposition.
Now suppose that $h=r-1$. Then $r-1\leq \frac{n_1}{2}$. As $n\geq n_1+2$, $r\leq \frac{n}{2}$. As by hypothesis $r> \frac{n-1}{2}$, we have the equality $r= \frac{n}{2}$. Then $\Gamma$ has the permutation representation graph given in the statement of this proposition with $h=i+1=r-1$, with some additional $i$-edges. As all $\Gamma_j$'s must be intransitive, the extra $i$-edges must be vertical edges. Hence one of $\rho_i$ and $\rho_{i+1}$ has to be an odd permutation, a contradiction.
\end{proof}
\begin{prop}\label{<iint}
Let $r>\frac{n-1}{2}$. Let $h>i$ be the maximal label of a permutation acting non-trivially on both $\Gamma_i$-orbits.
If $\Gamma_{<i}$ is intransitive in the first $\Gamma_i$-orbit, then $h<r-1$ and there exists a set of vertices $X$, contained in the second $\Gamma_i$-orbit, such that $h\leq \frac{n-|X|-1}{2}$ and $\Gamma_{>h}$ fixes $\{1,\ldots, n\}\setminus X$.
Moreover if $h= \frac{n-|X|-1}{2}$ then $h=i+1$ and $\Gamma_{<h+1}$ has the following permutation representation graph, where the black dots represent the vertices of $X$.
$$\[email protected]{\bullet\ar@{-}[rr]^0 \ar@{-}[d]_{i+1} && \bullet \ar@{-}[rr]^1\ar@{-}[d]_{i+1}&&\bullet\ar@{-}[d]_{i+1} \ar@{.}[rr] && \bullet\ar@{-}[rr]^{i-1}\ar@{-}[d]_{i+1} && \bullet\ar@{-}[rr]^i \ar@{-}[d]_{i+1} &&\bullet \ar@{-}[rr]^{i+1}&& *+[o][F]{} \\
\bullet\ar@{-}[rr]_0 && \bullet \ar@{-}[rr]_1&&\bullet \ar@{.}[rr] && \bullet\ar@{-}[rr]_{i-1}&& \bullet&&\\
}
$$
\end{prop}
\begin{proof}
In this case $J_A$ is not an interval thus, by Proposition~\ref{CCD}, $A$ is imprimitive embedded into $S_k\wr S_m$ and $\Gamma_{<i}$ is fixing all the blocks (with $k,m\geq 2$).
By Proposition~\ref{path} there exist a path $\mathcal{P}_1$ from the $h$-edge in the first $\Gamma_i$-orbit to the vertex $a$, and a path $\mathcal{P}_2$ from the $h$-edge in the second $\Gamma_i$-orbit to the vertex $b$, each of them containing all labels from $i+1$ to $h-1$, and fixed by $\Gamma_{>h}$.
In addition there is a path $\mathcal{P}_3$ in the block $\beta$ containing the vertex $a$, and edges with all labels from $0$ to $i-1$.
Moreover, there is also a path $\mathcal{P}_4$ in the block $\beta\rho_{i+1}$ also containing edges with all labels from $0$ to $i-1$.
Let us first assume that there is an edge with label $l>i$ inside one of the two blocks $\beta$ or $\beta\rho_{i+1}$.
Then the generator $\rho_l$ is fixed-point-free on the block containing that edge.
Suppose that block is $\beta$. Then $l = i+1$, a contradiction. Suppose then that the edge is in $\beta\rho_{i+1}$. Then $l=i+2 = h$ and there are only two blocks in the block system.
Moreover, $\rho_l$ acts inside $\beta\rho_{i+1}$ and provides an embedding of $\Gamma_{<i}$ into $S_2\wr S_{k/2}$.
Thus, since $\Gamma_j$ is intransitive for every $j$, a similar argument to the one used in the proof of Proposition~\ref{<jon>jtran} shows that $i \leq k/2-1$.
Hence $h-2 \leq k/2-1$.
Now if $h=r-1$, as $n\geq 2k+2$ (the extra 2 coming from the $h$-edge in the second $\Gamma_i$-orbit),
we get $r-1\leq \frac{n+2}{4} < \frac{n-2}{2}$ (as $n$ must be at least 7 in this case)
giving a contradiction with the hypotheses of the proposition.
Thus $h< r-1$ and $\Gamma_{>h}$ fixes a set $P$ of size $2k+2$.
Therefore, $h\leq \frac{|P|+2}{4} \leq \frac{|P|-1}{2}$ for every $|P| \geq 4$ which is obviously the case here as $k\geq 2$.
Let us now assume that all edges with label $l>i$ are not contained in $\beta$ nor in $\beta\rho_{i+1}$.
Then the paths $\mathcal{P}_1$, $\mathcal{P}_2$, $\mathcal{P}_3$ and $\mathcal{P}_4$ have no edge in common, as shown in the following figure.
$$\[email protected]{ *+[o][F]{} \ar@{-}[rr]^h&&*+[.][F]{} \ar@{~}[rrrr]^{\mathcal{P}_1} &&&&*+[.][F]{}\ar@{-}[d]_{i-1}\ar@{-}[rr]^{i+1} && *+[o][F]{a}\ar@{-}[d]^{i-1} \ar@{-}[rr]^i&& *+[o][F]{b}\ar@{~}[rrrr]^{\mathcal{P}_2}&&&& *+[.][F]{}\ar@{-}[rr]^h&& *+[o][F]{c}\\
&&&&&&*+[.][F]{}\ar@{~}[dd]_{\mathcal{P}_4}\ar@{.}[rr]_{i+1}&&*+[.][F]{}\ar@{~}[dd]^{\mathcal{P}_3}&&&&&&\\
&&&&&&&&&&&&&&\\
&&&&&&*+[o][F]{}&& *+[o][F]{} &&&&&&}
$$
Suppose that $h\neq r-1$ and that $\rho_{h+1}$ acts trivially on the first $\Gamma_i$-orbit.
Let $P$ be the set of vertices of $\mathcal{P}_1\cup\mathcal{P}_2\cup\mathcal{P}_3\cup \mathcal{P}_4$ excluding the vertex $c$ on the right hand side of the diagram above, and $X=\{1,\ldots,n\}\setminus P$.
The set $P$ is fixed by $\Gamma_{>h}$ and $2h+1\leq |P|=n-|X|$.
Moreover when $2h+1= |P|$ the permutation representation of $\Gamma_{<h}$ is the one given in this proposition with $h=i+1$.
Now if $h=r-1$, we have $2(r-1)+1\leq n-1$. Hence $r\leq \frac{n}{2}$, and by hypotheses we must have the equality.
Therefore $\Gamma$ has precisely the permutation representation graph given in the statement of this proposition with some additional $i$-edges.
As all $\Gamma_j$'s must be intransitive, the extra $i$-edges must be vertical. Hence one of $\rho_i$ and $\rho_{i+1}$ has to be an odd permutation, a contradiction.
It remains to prove that $\Gamma_{>h}$ fixes the first $\Gamma_i$-orbit. Suppose the contrary. Then $\rho_{r-1}$ acts nontrivially on the first $\Gamma_i$-orbit.
In this case consider the path $\mathcal{P}_1$ as before, with the label of its first edge being $r-1$ instead of $h$, and consider $\mathcal{P}_3$ and $\mathcal{P}_4$ has before.
Now let $\mathcal{P}_2$ be a copy of $\mathcal{P}_1$ whose last vertex is $a\rho_{i-1}$, as in the following figure.
$$\[email protected]{ *+[o][F]{} \ar@{-}[rr]^{r-1}&&*+[.][F]{} \ar@{~}[rrrr]^{\mathcal{P}_1} &&&&*+[.][F]{}\ar@{-}[d]_{i-1}\ar@{-}[rr]^{i+1} && *+[o][F]{a}\ar@{-}[d]^{i-1} \ar@{-}[rr]^i&& *+[o][F]{b}\ar@{~}[rrrr]&&&& *+[.][F]{}\ar@{-}[rr]^h&& *+[o][F]{}\\
*+[o][F]{} \ar@{-}[rr]_{r-1}&&*+[.][F]{} \ar@{~}[rrrr]_{\mathcal{P}_2} &&&&*+[.][F]{}\ar@{~}[dd]_{\mathcal{P}_4}\ar@{.}[rr]_{i+1}&&*+[.][F]{}\ar@{~}[dd]^{\mathcal{P}_3}&&&&&&\\
&&&&&&&&&&&&&&\\
&&&&&&*+[o][F]{}&& *+[o][F]{} &&&&&&}
$$
In this case $2(r-1)+1\leq (n-1)-1$, a contradiction.
\end{proof}
\begin{prop}\label{h}
Let $r>\frac{n-1}{2}$. Let $i\neq 0$ and $J_B$ be an interval with labels $>i$.
If $h>i$ is the maximal label of a permutation acting non-trivially on both $\Gamma_i$-orbits, then $h< r-1$ and there exists a set of vertices $X$, contained in the second $\Gamma_i$-orbit, such that $h\leq \frac{n-|X|-1}{2}$ and $\Gamma_{>h}$ acts trivially on $\{1,\ldots, n\}\setminus X$.
\end{prop}
\begin{proof}
This is consequence of Propositions~\ref{<itran} and \ref{<iint}.
\end{proof}
\begin{prop}\label{ineq0Btrivial}
Let $n\geq 8$.
If $i\neq 0$ and $J_B=\emptyset$ then $r\leq\frac{n-1}{2}$.
\end{prop}
\begin{proof}
Similarly to the proof of Proposition~\ref{<iint}, $A$ is embedded into a wreath product $S_k\wr S_m$ with $n=km+1$ and $\Gamma_{<i}$ fixing the blocks.
If there is a label $l>i$ such that $\rho_l$ acts nontrivially inside a block then $m=2$, $r-1=l=i+2$ and $\Gamma_{<i}$ is embedded into $S_2\wr S_{k/2}$, giving the inequality $i=r-3\leq k/2-1\leq \frac{n-1}{4}-1$. Hence for $n\geq 8$, $r\leq\frac{n-1}{2}$.
Suppose then that
every generator with label $>i$ acts trivially on the blocks it fixes.
Thus there are four paths $\mathcal{P}_1$, $\mathcal{P}_2$, $\mathcal{P}_3$ and $\mathcal{P}_4$, as in the following graph, containing all but one label twice and one cycle that is an alternating square.
$$\[email protected]{ *+[o][F]{} \ar@{-}[rr]^{r-1}&&*+[.][F]{} \ar@{~}[rrrr]^{\mathcal{P}_1} &&&&*+[.][F]{}\ar@{-}[d]_{i-1}\ar@{-}[rr]^{i+1} && *+[o][F]{a}\ar@{-}[d]^{i-1} \ar@{-}[rr]^i&& *+[o][F]{b}\\
*+[o][F]{} \ar@{-}[rr]_{r-1}&&*+[.][F]{} \ar@{~}[rrrr]_{\mathcal{P}_2} &&&&*+[.][F]{}\ar@{~}[dd]_{\mathcal{P}_4}\ar@{-}[rr]_{i+1}&&*+[.][F]{}\ar@{~}[dd]^{\mathcal{P}_3}&&\\
&&&&&&&&&&\\
&&&&&&*+[o][F]{}&& *+[o][F]{} &&}
$$
Hence $2(r-1)+1\leq n$.
If equality holds, the paths contain exactly one edge for each label, but then as all $\Gamma_j$'s are intransitive, $\rho_i$ acts trivially on $\{1,\ldots,n\}\setminus \{a,b\}$, and thus is a transposition, a contradiction. Therefore $r\leq \frac{n}{2}$. If $n$ is odd then $r\leq \frac{n-1}{2}$. If $n$ is even then $n_1$ is odd, and neither $k=2$ nor $m=2$. Hence there at least two more vertices not in the paths, and then $2(r-1)+1\leq n-2$. Therefore $r\leq\frac{n-1}{2}$.
\end{proof}
\begin{prop}\label{>htran}
Let $n\geq 12$. Suppose that $h>i$ is the maximal label of a permutation acting non-trivially on both $\Gamma_i$-orbits.
If $h\neq r-1$ and $\Gamma_{>h}$ is transitive on $X:= \{1,\ldots,n\} \setminus \mathrm{Fix}(\Gamma_{>h})$ then, either $r\leq \frac{n-1}{2}$, or one of the groups $\Gamma_{<3}$ or $\Gamma_{>r-4}$ is one the string C-groups (2) and (3) given in Table~\ref{small} or their duals.
\end{prop}
\begin{proof}
Suppose that $\Gamma_{>h}$ is neither the dual of the string C-group (2) nor the string C-group (3) given in Table~\ref{small}. Then, by Propositions~\ref{small123}, \ref{induction} and \ref{small(1)}, $r-1-h\leq \frac{|X|-1}{2}$. Moreover $r-1-h= \frac{|X|-1}{2}$ if $\Gamma_{>h}$ is the alternating group $A_{|X|}$.
By Proposition~\ref{h}, $h\leq \frac{n-|X|-1}{2}$. Suppose that $\Gamma_{>h}\cong A_{|X|}$ and $h= \frac{n-|X|-1}{2}$.
If $\rho_i$ has a nontrivial action on $X$ then, by Proposition~\ref{<jon>jtran}, $\Gamma_{>h}$ must be imprimitive with blocks of size two, a contradiction.
Thus we may assume that $\rho_i$ fixes $X$ pointwise and swaps a pair of vertices in $ \{1,\ldots,n\} \setminus X$.
If $i=0$ then $\Gamma_{<h}$ has the permutation representation graph given in Proposition~\ref{i=0andh<>r-1}. The string condition implies that $h=2$ and $\Gamma_{<3}$ must be the string C-group (2) given in Table~\ref{small}, a contradiction.
When $i\neq 0$, either $r\leq \frac{n-1}{2}$ or $\Gamma_{<h}$ has either the permutation representation graph given in Proposition~\ref{<itran} or the one given in Proposition~\ref{<iint}.
As all $\Gamma_j$'s must be intransitive, the extra $i$-edges must be vertical edges. Hence one of $\rho_i$ and $\rho_{i+1}$ has to be an odd permutation, a contradiction.
Hence either $h<\frac{n-|X|-1}{2}$ or $r-1-h<\frac{|X|-1}{2}$, which implies in both cases that $r\leq \frac{n-1}{2}$.
\end{proof}
In the previous proposition we considered the case where $\rho_h$ acts nontrivially in both $\Gamma_i$-orbits and $\Gamma_{>h}$ is transitive on the points it does not fix.
Let us now deal with the case where $\Gamma_{>h}$ is intransitive.
In what follows we use the results of the previous section on 2-fracture graphs.
Observe that in the proofs of Propositions~\ref{max1squareforcomp} and \ref{dislc} none of the following three conditions on $\Gamma$ were needed: intersection condition, connectedness of the diagram and the group being even.
Indeed all we need is a string group $G$ generated by a set $\{\delta_j \,|\, j \in \{0,\ldots ,d-1\}\}$ of involutions
where $\delta_j$ swaps two pairs of vertices in different $G_j$-orbits for each $j \in \{0,\ldots ,d-1\}$.
\begin{prop}\label{t}
Let $t\in\{0,\ldots,r-2\}$ and $U:=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{>t})$. If $t$ is such that
\begin{itemize}
\item $t\leq \frac{n-|U|-1}{2}$,
\item $\Gamma_{>t}$ has a 2-fracture graph,
\item $\Gamma_{>t}$ acts intransitively on $U$,
\end{itemize}
then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
Assume first that $t=r-2$. In this case, $|U| \geq 4$ for $\rho_{r-2}$ to be an even permutation. Hence $r\leq \frac{n-1}{2}$.
Let $t<r-2$ be the maximal label satisfying the conditions of this proposition.
Suppose that $\Gamma_{>t}$ has $c$ nontrivial orbits $U_1,\ldots, U_c$.
For each set $U_s$ with $s\in\{1,\ldots, c\}$ denote by $I_s$ the set of labels
$l\,(>t)$ of edges in $U_s$.
Consider the graph $\mathcal{C}$
whose vertices are the orbits $U_1, \ldots, U_c$, and two orbits $U_s$ and $U_q$ are joined by an $l$-edge if there exist a point in $U_s$ and a point in $U_q$ that are swapped by $\rho_l$.
\begin{comment}
Suppose that $\mathcal{C}_t$ is connected. Then, for all $s\in\{1,\ldots,k\}$, we have $|I_s|=r-1-t$ and $|U_s| = \frac{|U|}{k}$.
By Proposition~\ref{path}, there is a path from the $i$-edge $\{a,b\}$ to an $(r-1)$-edge containing all labels from $i+1$ to $r-1$. Hence, in this case, each component $U_s$ contains a path with all labels from $t+1$ to $r-1$. Thus $r-1-t\leq \frac{|U|}{k}-1\leq \frac{|U|}{2}-1$ and therefore, using the first assumption of the proposition, $r\leq \frac{n-1}{2}$.
Suppose now that $\mathcal{C}_t$ is disconnected with $c$ components $\{C_1, \ldots, C_c\}$.
In each connected component $C_j$ of $\mathcal{C}$, we pick one orbit $U_{x_j}$ that is connected to some other orbit (which could be a unique point) in another component using a $t$-edge.
Let $k_j := \frac{|C_j|}{|U_{x_j}|} \in \mathbb{N}$.
We have that $|U| = k_1|U_{x_1}|+\ldots+ k_c|U_{x_c}|$.
\end{comment}
Consider a (simple) fracture graph $\mathcal{F}$ of $\Phi:=\Gamma_{>t}$, that is, a graph with $|U|$ vertices and $r-1-t$ edges. Such a graph exists by the second hypothesis of the proposition.
Each $l$-edge of $\mathcal{F}$ connects vertices in different $\Phi_l$-orbits. Let $s\in\{1,\ldots,c\}$ and $F_s$ be the set of labels of edges of $\mathcal{F}$ within $U_s$.
Clearly $F_s\subseteq I_s$.
Choose $\mathcal{F}$ such that it satisfies the following property:
\begin{enumerate}
\item[P1] if $l\in F_s$ is the label of the unique $l$-edge in one component swapping vertices in different $\Gamma_l$-orbits, then no other component has more than one pair of vertices in different $\Gamma_l$-orbits.
\end{enumerate}
Let $G_s$ be the group action of $\Gamma_{>t}$ in $U_s$. We have that $G_s$ is generated by a set of involutions (not necessarily independent) with labels in $I_s$. The subset of involutions with labels in $F_s$ is independent since $F_s$ is the subset of labels of edges of $\mathcal{F}$.
We denote by $(G_s)_j$ the group generated by all involutions of the generating set of $G_s$ except the one with label $j$.
If $G_s$ does not admit a 2-fracture graph with set of labels $F_s$, then there exists an edge $e$ with label $l\in F_s$ with $\rho_l$ swapping only one pair of vertices of $U_s$, in different $\Gamma_l$-orbits. Let $m$ be the minimal label and $x$ be the maximal label of an edge inside $U_s$.
By Proposition~\ref{path} there exists a path $\mathcal{P}_1$ containing all labels from $l-1$ to $m$ and another path $\mathcal{P}_2$ containing all labels from $l+1$ $x$ in $U_s$. Let $\mathcal{P}=\mathcal{P}_1\cup
\{e\} \cup\mathcal{P}_2$. Now we deal separately with the cases $m>t+1$ and $m=t+1$ and we conclude, in both cases, that $x \leq \frac{n-|X|-2}{2}$ where $X := \{1,\ldots,n\}\setminus \mathrm{Fix}(\Gamma_{>x})$.\\
$\underline{m>t+1}$: If $m>t+1$ then a component $U_q$ adjacent to $U_s$ also contains all labels from $m$ to $x$.
Let $U_z$ be a component containing an edge with label $m-1$.
We can reach it from $U_s$ with a shortest path in $\mathcal{C}$.
The last component before $U_z$ in this path, say $U_w$, contains a copy $\mathcal{P}'$ of $\mathcal{P}$.
Thus there is a $t$-edge from $U_w$ to $U_z$.
$$\[email protected]{ &&*+[o][F]{} \ar@{-}[rr]^m\ar@{-}[d]^t&&*+[o][F]{} \ar@{-}[rr]^{m+1}\ar@{-}[d]^t&&*+[o][F]{} \ar@{~}[rrrr]\ar@{-}[d]^t&&&&*+[o][F]{}\ar@{-}[rr]^x\ar@{-}[d]^t&&*+[o][F]{}\ar@{-}[d]^t& U_w\\
*+[o][F]{v}\ar@{-}[rr]_{m-1}&&*+[o][F]{} \ar@{-}[rr]_m&&*+[o][F]{} \ar@{-}[rr]_{m+1}&&*+[o][F]{} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]_x&&*+[o][F]{}&U_z}
$$
As $m$ is the minimal label in $U_{w}$, $m-1\notin U_{w}$, thus $m-1=t+1$. Let $P$ be the set of vertices of $\mathcal{P}\cup\mathcal{P}'$.
We have $2(x-(t+1))\leq |P|-2$, hence $x\leq t+\frac{|P|}{2}\leq \frac{n-|U|+|P|-1}{2}$.
As $\Gamma_{>t}$ has a 2-fracture graph by hypothesis, there is at least one more edge with label $t+1$, so $|U|>|P|+1$ and $x\leq \frac{n-3}{2}$.
If $x=r-1$ then $r\leq\frac{n-1}{2}$.
Let $x\neq r-1$. The vertex $v$ is fixed by $\Gamma_{>x}$ as well as all vertices of $P$. Let $X := \{1,\ldots,n\}\setminus \mathrm{Fix}(\Gamma_{>x})$.
Then we have $x \leq \frac{n-|X|-2}{2}$.\\
$\underline{m=t+1}$: Suppose that $m=t+1$. Furthermore assume that any component containing a unique $l$-edge between vertices in different $\Gamma_l$-orbits, has minimal label $t+1$.
Let $x$ be the maximal label in $U_s$ and let $\mathcal{P}$ be as before.
We now use the fact that $\Gamma_{>t}$ has a 2-fracture graph, and thus there exists another component $U_q$ having one $l$-edge $e'$ between vertices in different components. Moreover by Property (P1) this $l$-edge cannot be in an alternating square inside $U_q$. By assumption the minimal label in $U_q$ is also $t+1$.
Let $y$ be the maximal label in $U_q$.
Assume that $y\geq x$. Then, there exists a path $\mathcal{P}'_1$ containing all labels from $l-1$ to $t+1$ and another path $\mathcal{P}'_2$ containing all labels from $l+1$ to $x$, both in $U_q$. Consider $\mathcal{P}'=\mathcal{P}'_1\cup\{e'\} \cup\mathcal{P}'_2$.
There is at most one vertex of $\mathcal{P}'$ that is not fixed by $\Gamma_{>x}$.
Let $Q$ be the set of vertices of $\mathcal{P}\cup\mathcal{P}'$.
$$\[email protected]{ *+[o][F]{} \ar@{-}[rr]^{t+1}&&*+[o][F]{} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]^x&&*+[o][F]{}&U_s&&&&\\
*+[o][F]{} \ar@{-}[rr]^{t+1}&&*+[o][F]{} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]^x&&*+[o][F]{c} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]^y&&*+[o][F]{}&U_q}$$
In this case $\Gamma_{>x}$ fixes the set $Q\setminus\{c\}$.
Hence $2(x-t)\leq |Q|-2$ and we get
$$x\leq \frac{n-|X|-2}{2}$$
where $X := \{1,\ldots,n\}\setminus \mathrm{Fix}(\Gamma_{>x})$.\\
In both cases, thanks to the maximality of $t$ we conclude that either $x=r-1$ or $\Gamma_{>x}$ is transitive on $X$. If $x=r-1$, $r\leq \frac{n-2}{2}$.
Suppose $\Gamma_{>x}$ is transitive on $X$. In this case $X$ is a subset of $U_s$ for some $s\in\{1,\ldots,c\}$.
If $\Gamma_{>x}$ is fix-point-free on $U_s$, then $\rho_t$ centralizes $\Gamma_{>x}$, thus, by Proposition~\ref{<jon>jtran}, $r-1-x\leq \frac{|X|}{2}-1$.
Hence $r\leq \frac{n-1}{2}$. If $\Gamma_{>x}$ has fixed points in $U_s$, then $x\neq \frac{n-|X|-2}{2}$ and, as by Proposition~\ref{induction} $r-1-x\leq \frac{X}{2}$, we get $r\leq \frac{n-1}{2}$.
This concludes the case where $G_s$ has no 2-fracture graph.
Therefore we may assume that $G_s$ admits a 2-fracture graph with set of labels $F_s$ and therefore $F_s\leq \frac{|U_s|}{2}$. We can use the results of Section~\ref{s:2frac}.
Suppose that $\Gamma_{>t+1}$ is transitive on $U_s$.
Take the closest orbit $U_q$ to $U_s$ in $\mathcal{C}$ such that $\Gamma_{> t+1}$ is intransitive on $U_q$ and let $P$ be a shortest path from $U_s$ to $U_q$.
We can use $\rho_t$ modify $\mathcal{F}$ along this path, to concentrate all edges of the fracture graph in $U_q$. We obtain another fracture graph also satisfying P1.
Hence, we may assume that $F_x$ is empty for every orbit $U_x$ of $P$ except $U_q$.
Therefore $|F_x| = 0 \leq \frac{|U_x|-2}{2}$.
Thus in every component with $F_s\neq \emptyset$ there exists an edge with label $t+1$ between vertices in different $\Gamma_{t+1}$-orbits.
Consequently there is at most one component having a connected 2-fracture graph, and $r-1-t\leq \frac{|U|-(c-1)}{2}$
where $c$ is the number of $\Gamma_{>t}$-orbits of size at least 2.
Whenever $c>2$ or if $|F_x| = 0$ for some $x$, we get $r \leq \frac{n-1}{2}$.
Let us now assume $c=2$. In this case, if $r = \frac{n}{2}$, the 2-fracture graph on $U_s$ is connected and has an alternating square (by Proposition~\ref{max1squareforcomp}), and the 2-fracture graph on $U_q$ is disconnected, and has two components, one having an alternating square and the other being a tree. These two components in $U_q$ must moreover be connected by a $t$-edge.
These components satisfy the following equalities: $|F_s|=\frac{|U_s|}{2}$, $|F_q|=\frac{|U_q|-1}{2}$, $I_s=F_s$ and $I_q=F_q\cup\{t+1\}$.
Let $G_s$ be the group action in $U_s$ and let $G_q$ be the group action in $U_q$.
Recall that $t < r-2$. If $(G_q)_{t+2}$ is transitive on $U_q$ then $\rho_{t+1}$ centralizes $G_q$, thus the two components of the 2-fracture graph of $G_q$ have the same shape, so they must both be trees, a contradiction. Thus $(G_q)_{t+2}$ is intransitive on $U_q$. Moreover $t+2\in F_q$ as $I_q=F_q\cup\{t+1\}$. Therefore $t+2\notin F_s=I_s$, hence $\rho_{t+1}$ centralizes $G_s$. But then there exists an edge with label $l\neq t+1$ in $U_s$ incident to the $t$-edge that connects $U_s$ to $U_q$, thus $l\in I_s\cap I_q$, a contradiction.
\end{proof}
\begin{prop}\label{case3}
Let $n\geq 12$. Suppose that $i$ is the maximal label such that
\begin{itemize}
\item $\Gamma_i$ has two orbits with a single $i$-edge joining them;
\item there exists $h>i$ such that $\rho_h$ acts non-trivially on both $\Gamma_i$-orbits.
\end{itemize}
If $r> \frac{n-1}{2}$, then one of the following occurs.
\begin{enumerate}
\item $\Gamma_{>r-4}$ is the dual of the string C-group (2), or the string C-group (3) given in Table~\ref{small};
\item $\Gamma_{r-1}$ has exactly two orbits, one being trivial.
\end{enumerate}
\end{prop}
\begin{proof}
Let $h$ be maximal.
Let us choose the vertices $a$ and $b$, and consequently the groups $A$ and $B$, such that $J_B$ is an interval with all labels in $J_B$ being $>i$. If $J_A$ is also an interval, then, as $h\in J_A\cap J_B$, $i=0$; we assume without loss of generality that $r-1\in J_A\setminus J_B$ (Recall that the case $h=r-1$ has been dealt with in Proposition~\ref{i=0andh=r-1}).
Denote by $O_1$ and $O_2$ the orbits of $A$ and $B$ respectively.
By Proposition~\ref{i=0andh<>r-1} when $i=0$, and Proposition~\ref{h} when $i\neq 0$, we have that $h< r-1$ and $h\leq \frac{n-|X|-1}{2}$ with $X:=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{>h})\subseteq O_2$. If $\Gamma_{>h}$ is transitive, by Proposition~\ref{>htran}, one of the groups $\Gamma_{<3}$ or $\Gamma_{>r-4}$ is one of the string C-groups (2) or (3) given in Table~\ref{small} or their duals. Thus we may assume that $\Gamma_{>h}$ is intransitive.
Now if $\Gamma_{>h}$ has a 2-fracture graph, by Proposition~\ref{t}, $r\leq \frac{n-1}{2}$, a contradiction.
Hence $\Gamma_{>h}$ does not admit a 2-fracture graph.
Then there exists $j>h$ such that $\rho_j$ only swaps one pair of vertices in different $\Gamma_j$-orbits.
Choose $j$ minimal with this property. Then $\Gamma_{\{h+1,\ldots j-1\}}$ has a 2-fracture graph.
Let $C$ and $D$ be the group actions of $\Gamma_j$ in each $\Gamma_j$-orbit with $C$ acting on the orbit $L_1$ containing the $i$-edge $\{a,b\}$. Let $L_2$ denote the other $\Gamma_j$-orbit.
If one of the groups is trivial then either we get case (b) of the statement of this proposition or $r\leq \frac{n-1}{2}$ by the dual of Proposition~\ref{ineq0Btrivial}, a contradiction.
Thus assume both $C$ and $D$ are nontrivial sggi's. Indeed let $\rho_j=\gamma_j\delta_j$ with $\gamma_j$ and $\delta_j$ being the permutations in each $\Gamma_j$-orbit.
Then $C = \langle \gamma_i\, |\, i\in \{0,\ldots, r-1\}\rangle$ and $D := \langle \delta_i \,| \,i\in \{0,\ldots, r-1\}\rangle$.
Let $J_C:= \{ i \in \{0,\ldots, r-1\}\,|\,\gamma_i \neq 1_C\}$ and $J_D:= \{ i \in \{0,\ldots, r-1\}\,|\,\delta_i \neq 1_D\}$.
By Propositions~\ref{CCD} and \ref{bothimp} either $J_C$ or $J_D$ is an interval.
If $C=\Gamma_{<j}$ and $D=\Gamma_{>j}$ then by Proposition~\ref{Case2}, $r\leq \frac{n-1}{2}$ or $\Gamma_{>r-4}$ is, up to duality, one of the string C-groups (2) or (3) given in Table~\ref{small}. This gives case (a) of the statement.
It remains to consider the case where there exists a permutation $\rho_g$ acting non-trivially on both $\Gamma_j$-orbits.
Choose $g$ minimal.
Thanks to the maximality of $i$, $g<j$.
As $g\in J_D$ we have that $g>i$. We now consider four cases:
(1) If $j=r-1$ and $g=0$, then by the dual of Proposition~\ref{i=0andh=r-1}, we get a contradiction.
(2) If $j=r-1$ and $g\neq 0$, then both $J_C$ and $J_D$ are intervals and $g$ is the minimal label of a permutation acting nontrivially on $L_2$.
Then we can use the dual of Proposition~\ref{i=0andh<>r-1} to conclude that
$r-g\leq \frac{n-|Y|-1}{2}$ with $Y:= \{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{<g})\subseteq L_1$.
If $g\leq h$ then, by Proposition~\ref{path}, there is a path in $X$ containing all labels from $h$ to $r-2$ twice. Let $V$ be the set of vertices of this path, then $r-h\leq\frac{|V|}{2}$. Hence $r\leq \frac{n-|X|-|V|-1}{2}$, giving a contradiction.
Hence $g> h$ and $X\cap Y\neq \emptyset$.
(3) Let $j\neq r-1$ and $J_D$ be an interval. In this case both $\rho_0$ and $\rho_{r-1}$ act nontrivially on the first $\Gamma_j$-orbit. Hence we can use the dual argument to the one used in the last paragraph of the proof of Proposition~\ref{<iint}, to conclude that $r\leq \frac{n-1}{2}$, a contradiction.
(4) Let $j\neq r-1$ and $J_C$ be an interval. By the dual of Proposition~\ref{h}, $g<0$ and $r-g\leq \frac{n-|Y|-1}{2}$ with $Y:= \{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{<g})\subseteq L_1$. By the same reasoning as in (2), $h>g$.
Only cases (2) and (4) are possible and both give the inequality $r-g\leq \frac{n-|Y|-1}{2}$ with $Y:= \{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{<g})\subseteq L_1$ and $h>g$.
By choice of $j$, there is no other label $l$ between $h$ and $g$ having only one pair of vertices in different $\Gamma_l$-orbits.
Then $\Gamma_{\{h+1,\ldots, g-1\}}$ has a 2-fracture graph. Moreover $\Gamma_{>h} \cap \Gamma_{<g}$
is intransitive on $X\cap Y$
for if it were transitive, then $\Gamma_{>h}$ would be transitive on $X$, contradicting the assumptions of the proposition.
Hence the 2-fracture graph for $\Gamma_{\{h+1,\ldots, g-1\}}$ is disconnected and $g-h-1\leq \frac{|X\cap Y|-1}{2}$.
We have $h\leq \frac{n-|X|-1}{2}$, $g-h\leq \frac{|X\cap Y|-1}{2}+1$, $r-g \leq \frac{n-|Y|-1}{2}$ and $n= |X| + |Y| - |X \cap Y|$, thus $r\leq \frac{n-1}{2}$, a contradiction.
\end{proof}
We now consider that $A$ is trivial or has the permutation representation graph (2) or (3) of Table~\ref{small}.
\begin{prop}\label{0trivialtran}
Let $i=0$. If $A$ is trivial and $\Gamma_{>1}$ is transitive on $\{1,\ldots,n\}\setminus\{a,b\}$, then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
As $\rho_0$ acts nontrivially on $\{1,\ldots,n\}\setminus\{a,b\}$ and centralizes $\Gamma_{>1}$, by Proposition~\ref{<jon>jtran},
$r-2\leq \frac{n-2}{2}-1$ and therefore $r\leq n/2$. Suppose that we have the equality. Then $\Gamma_{>1}$ is as given in Proposition~\ref{<jon>jtran}, and either $\rho_0$ or $\rho_1$ is odd, a contradiction. Hence $r\leq \frac{n-1}{2}$.
\end{proof}
\begin{prop}\label{12Tran}
Let $A$ have the permutation representation $F_x=\emptyset$ or (3) of Table~\ref{small} (this implies that $n\geq 7$).
If $\Gamma_{>4}$ is transitive on $n-7$ vertices then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
In this case $i=3$ and $\rho_i$ has full support on the $n-7$ vertices that are not fixed by $\Gamma_{>4}$.
So $n-7$ is even. By Proposition~\ref{<jon>jtran}, $r-5\leq \frac{n-7}{2}-1$.
Moreover if we have the equality then either $\rho_3$ or $\rho_4$ is odd, a contradiction.
Hence $r-5< \frac{n-7}{2}-1$ and $r\leq \frac{n-1}{2}$.
\end{proof}
\begin{prop}\label{specialbubbles}
Suppose that $A$ is trivial or has permutation representation graph (2) or (3) of Table~\ref{small} and $B$ is the alternating group.
If $\Gamma_{>i+1}$ is intransitive on the second $\Gamma_i$-orbit and $\Gamma_{>i}$ has a 2-fracture graph, then $r\leq \frac{n-1}{2}$.
Moreover if $r> \frac{n-3}{2}$ then one of the following possibilities must occur.
\begin{enumerate}
\item there exists $x\in\{0,\ldots, r-1\}$ such that $x\leq\frac{n-|X|-1}{2}$ with $X=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{>x})$.
\item $\Gamma$ has the following permutation representation graph.
$$\[email protected]{ &&*+[o][F]{} \ar@{-}[rr]^1&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$$
\end{enumerate}
\end{prop}
\begin{proof}
Consider the graph $\mathcal{C}$ as in the proof of Proposition~\ref{t}, with $t=i+1$.
Let $U$, $U_s$, $G_s$, $F_s$ and $I_s$ be as in Proposition~\ref{t}.
Suppose there is a component $U_s$ that does not have a 2-fracture graph.
Let $m$ and $x$ be the minimal and the maximal label of that component respectively.
We proved in Proposition~\ref{t} that $m\in\{t+1, t+2\}$ and accordantly to these possibilities for $m$ the permutation representation graph of $\Gamma$ contains one of the following graphs.
\begin{tabular}{cc}
$m=t+2:$&$\[email protected]{ &&*+[o][F]{} \ar@{-}[rr]^{t+2}\ar@{-}[d]^t&&*+[o][F]{} \ar@{-}[rr]^{t+3}\ar@{-}[d]^t&&*+[o][F]{} \ar@{~}[rrrr]\ar@{-}[d]^t&&&&*+[o][F]{}\ar@{-}[rr]^x\ar@{-}[d]^t&&*+[o][F]{}\ar@{-}[d]^t\\
*+[o][F]{}\ar@{-}[rr]_{t+1}&&*+[o][F]{} \ar@{-}[rr]_{t+2}&&*+[o][F]{} \ar@{-}[rr]_{t+3}&&*+[o][F]{} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]_x&&*+[o][F]{}}$\\
$m=t+1$:& $\[email protected]{ *+[o][F]{} \ar@{-}[rr]^{t+1}&&*+[o][F]{} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]^x&&*+[o][F]{}&&&&\\
*+[o][F]{} \ar@{-}[rr]^{t+1}&&*+[o][F]{} \ar@{~}[rrrr]&&&&*+[o][F]{}\ar@{-}[rr]^x&&*+[o][F]{c} }$
\end{tabular}
Let $P$ be the set of vertices of the first graph and $Q$ be the set of vertices of the second graph.
If $x\neq r-1$, in the first case $\Gamma_{>x}$ fixes $P$, and in the second case $\Gamma_{>x}$ fixes $Q\setminus\{c\}$. If $i=0$ and $t=1$ we have that $x\leq \frac{|P|+1}{2}\mbox{ or }x\leq \frac{|Q|}{2}.$
If $i=3$ (and $t=4$), $x\leq \frac{|P|+7}{2}\mbox{ or }x\leq \frac{|Q|+6}{2}.$
Suppose that $x=r-1$ and $r\geq \frac{n-2}{2}$.
Consider first that $A$ is trivial and $m=t+2$. In this case $n\geq |P|+2$ and $\frac{|P|+1}{2}\geq \frac{n-4}{2}$, thus $|P|\in\{n-5,n-4,n-3,n-2\}$. Then it is possible to determined the permutation representation graph of $\Gamma$ according to the value of $|P|$. If $|P|=n-2$ then $\Gamma$ is the graph (1) of Table~\ref{GammaWhenATrivial1} which is not an even group. It is also not possible to get an even group when $|P|=n-3$, as there must exist a $\Gamma_{>1}$-orbit with a 2-edge. For $|P|=n-4$ it is possible to create such a component and we get the graph (2) of Table~\ref{GammaWhenATrivial1}. But as $\Gamma_{>1}$ does not have a 2-fracture graph we get a contradiction.
If $|P|=n-5$, it is not possible to create a third $\Gamma_{>1}$-orbit, thus $\Gamma_{>1}$ has exactly two components, one containing $P$
and the other having an even number of vertices swapped pairwise by $\rho_0$, a contradiction.
\begin{table}
\begin{small}
\[\begin{tabular}{|c|c|c|}
\hline
\#&$ |P|$& \textbf{Possibilities for $\Gamma$ when $A$ is trivial }\\
\hline
\textbf{1}&$n-2$&$\[email protected]{ &&&&&&*+[o][F]{} \ar@{-}[rr]^3\ar@{-}[d]^1&&*+[o][F]{}\ar@{-}[d]^1 \ar@{.}[rr]\ar@{-}[d]^1&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{-}[d]^1&&*+[o][F]{}\ar@{-}[d]^1\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$
\\
\hline
\textbf{2}&$n-4$&$\[email protected]{&&&& &&&&&&*+[o][F]{} \ar@{-}[rr]^3\ar@{-}[d]^1&&*+[o][F]{}\ar@{-}[d]^1 \ar@{.}[rr]\ar@{-}[d]^1&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{-}[d]^1&&*+[o][F]{}\ar@{-}[d]^1\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{=}[rr]_2^0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$
\\
\hline
\end{tabular}\]
\end{small}
\caption{Possibilities depending on $|P|$ for $\Gamma$ when $A$ is trivial}\label{GammaWhenATrivial1}
\end{table}
Now let $A$ be trivial and $m=t+1$. In this case $|Q|\in\{n-4,n-3,n-2\}$. In Table~\ref{GammaWhenATrivial2} we list all possibilities for the permutation representation graph of $\Gamma$ for each value of $|Q|$. If the permutation representation graph of $\Gamma$ is one of the graphs (3), (5a), (5b) or (5c), then $\Gamma$ is odd, a contradiction. Thus the only possibility is the permutation representation graph (4), giving the graph of the statement of this proposition.
\begin{table}
\begin{small}
\[\begin{tabular}{|c|c|c|}
\hline
\#&$|Q|$ & \textbf{Possibilities for $\Gamma$ when $A$ is trivial }\\
\hline
\textbf{3}&$n-2$&$\[email protected]{ &&&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$
\\
\hline
\textbf{4}&$n-3$&$\[email protected]{ &&*+[o][F]{} \ar@{-}[rr]^1&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$
\\
\hline
\textbf{5a}&$n-4$&$\[email protected]{ *+[o][F]{} \ar@{-}[rr]^2&&*+[o][F]{} \ar@{-}[rr]^1&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$
\\
&& or \\
&&$\[email protected]{ *+[o][F]{} \ar@{=}[rr]^2_0&&*+[o][F]{} \ar@{-}[rr]^1&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{5b} &$n-4$ &$\[email protected]{ &&&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\ar@{-}[rr]^{r-2}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}}$
\\
\hline
\textbf{5c} &$n-4$ &$\[email protected]{ &&&&&&&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
&& or \\
&&$\[email protected]{ &&&&&&&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{=}[rr]_2^0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\end{tabular}\]
\end{small}
\caption{Possibilities depending on $|Q|$ for $\Gamma$ when $A$ is trivial}\label{GammaWhenATrivial2}
\end{table}
Now let $A$ be the permutation representation graph (2) of Table~\ref{small} and $m=t+2$. If $r\geq \frac{n-2}{2}$ and $x=r-1$ then $|P|\in\{n-11,\, n-10,\,n-9,\,n-8,\,n-7\}$.
In Table~\ref{GammaWhenA(1)} we list all possibilities for $\Gamma$ according to $|P|$. For $|P|=n-7$ we get the permutation representation graph (1) and $\Gamma$ is odd, a contradiction. Then $|P|<n-8$, as there must exist a $\Gamma_{>4}$-orbit containing a $5$-edge and a $3$-edge. Thus for $|P|=n-9$ we get the permutation representation graph (2), but $\Gamma_{>4}$ doesn't have a 2-fracture graph, a contradiction. Now suppose $|P|\leq n-10$.
Either there are two $\Gamma_{>4}$-orbits, one having a set of vertices $P$
and another having $\rho_3$ swapping all its vertices pairwise, or
there exist a third $\Gamma_{>4}$-orbit containing a $5$-edge. For this to happen at least two additional vertices are needed, thus $|P|= n-11$. This gives the permutation representation graph (3), which again corresponds to an odd group, a contradiction.
\begin{table}
\begin{small}
\[\begin{tabular}{|c|c|c|}
\hline
\# &$|P|$& \begin{tabular}{c} \textbf{Possibilities for $\Gamma$ when $A$}\\\textbf{ has the permutation representation graph (2)} \end{tabular}\\
\hline
\textbf{1} & $n-7$ &$\[email protected]{&& && && && && && && &&*+[o][F]{}\ar@{-}[rr]^6\ar@{-}[d]^4&& *+[o][F]{}\ar@{.}[rr]\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[d]^4\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{2} & $n-9$ &$\[email protected]{&& && && && && && && && && &&*+[o][F]{}\ar@{-}[rr]^6\ar@{-}[d]^4&& *+[o][F]{}\ar@{.}[rr]\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[d]^4\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{=}[rr]_5^3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{3} & $n-11$ &$\[email protected]{&& && && && && && && *+[o][F]{}\ar@{=}[rr]^3_4&&*+[o][F]{} && &&*+[o][F]{}\ar@{-}[rr]^6\ar@{-}[d]^4&& *+[o][F]{}\ar@{.}[rr]\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[d]^4\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[u]^5\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[u]_5\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\end{tabular}\]
\end{small}
\caption{Possibilities depending on $|P|$ for $\Gamma$ when $A$ has permutation representation graph (2)}\label{GammaWhenA(1)}
\end{table}
In Table~\ref{GammaWhenA(1)2}, we list all possibilities for $\Gamma$ when $A$ has the permutation representation graph (2) of Table~\ref{small} and $m=t+1$. As before $r\geq \frac{n-2}{2}$ and $x=r-1$, hence $|Q|\in\{n-10,\,n-9,\,n-8,\,n-7\}$. In (4), (6a), (6b), (6c), (6d), (7b), (7c), (7d), (7e) and (7f) $\Gamma$ is odd. In the remaining case the intersection condition fails.
\begin{table}
\begin{small}
\[\begin{tabular}{|c|c|c|}
\hline
\# &$|Q|$& \begin{tabular}{c} \textbf{Possibilities for $\Gamma$ when $A$}\\\textbf{ has the permutation representation graph (2)} \end{tabular}\\
\hline
\textbf{4} & $n-7$ &$\[email protected]{&& && && && && && &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{5} & $n-8$ &$\[email protected]{&& && && && && &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{6a} & $n-9$ &$\[email protected]{&& && && && &&*+[o][F]{}\ar@{-}[rr]^5 &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{6b} & $n-9$ &$\[email protected]{&&&& && && && && && && &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&
*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{6c} & $n-9$ & $\[email protected]{&& && && && && && &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}}$\\
\hline
\textbf{6d} & $n-9$ &$\[email protected]{&& && && && &&*+[o][F]{}\ar@{=}[rr]^5_3 &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{7a} & $n-10$ & $\[email protected]{&& && && && && &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}}$\\
\hline
\textbf{7b} & $n-10$ &$\[email protected]{&& && && && *+[o][F]{}\ar@{-}[rr]^4&&*+[o][F]{}\ar@{-}[rr]^5 &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{7c} & $n-10$ &$\[email protected]{&&&& && && && && && && *+[o][F]{}\ar@{-}[rr]^4&&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&
*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{7d} & $n-10$ & $\[email protected]{&& && && && && && *+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-2}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}\ar@{-}[rr]_{r-2}&&*+[o][F]{}}$\\
\hline
\textbf{7e} & $n-10$ &$\[email protected]{&& && && && *+[o][F]{}\ar@{-}[rr]^6&&*+[o][F]{}\ar@{-}[rr]^5 &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\textbf{7f} & $n-10$ &$\[email protected]{&& && && && *+[o][F]{}\ar@{-}[rr]^4&&*+[o][F]{}\ar@{=}[rr]^5_3 &&*+[o][F]{}\ar@{-}[rr]^4 &&*+[o][F]{}\ar@{-}[rr]^5\ar@{-}[d]^3 &&*+[o][F]{}\ar@{-}[rr]^6\ar@{=}[d]^4_3&& *+[o][F]{}\ar@{.}[rr]\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{-}[rr]^{r-1}\ar@{=}[d]^4_3&&*+[o][F]{}\ar@{=}[d]^4_3\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r-1}&&*+[o][F]{}}$\\
\hline
\end{tabular}\]
\end{small}
\caption{Possibilities depending on $|Q|$ for $\Gamma$ when $A$ has permutation representation graph (2)}\label{GammaWhenA(1)2}
\end{table}
If $A$ has the permutation representation graph (3) of Table~\ref{small} we get the same contradictions as in Tables~\ref{GammaWhenA(1)} and~\ref{GammaWhenA(1)2}.
Now suppose that $x\neq r-1$. The group $\Gamma_{>x}$ fixes the first $\Gamma_i$-orbit and the vertex $b$.
Let $X:=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{>x})$.
When $i=0$ and $t=1$ we have $|X|\leq n-(|P|+2)$ and $x\leq \frac{|P|+1}{2}$, or $|X|\leq n-(|Q\setminus\{c\}|+2)$ and $x\leq\frac{|Q|}{2} $, giving in any case $x\leq\frac{n-|X|-1}{2}$ .
When $i=3$ and $t=4$, we have $|X|\leq n-(|P|+7)$ and $x\leq\frac{|P|+7}{2} $, or $|X|\leq n-(|Q\setminus\{c\}|+7)$ and $x\leq\frac{|Q|+6}{2}$. Hence $x\leq\frac{n-|X|}{2}$. Suppose we have the equality. Then $\Gamma$ contains one of following two graphs, having all vertices fixed by $\Gamma_{>x}$, except one of vertices $c$ or $d$ of graph (2).
$$(1)\; \[email protected]{&& && && && && && && &&*+[o][F]{}\ar@{-}[rr]^6\ar@{-}[d]^4&& *+[o][F]{}\ar@{.}[rr]\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[rr]^{x}\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[d]^4\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{x}&&*+[o][F]{}}$$
or
$$(2)\; \[email protected]{&& && && && && && &&*+[o][F]{}\ar@{-}[rr]^5&&*+[o][F]{}\ar@{-}[rr]^6&& *+[o][F]{}\ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]^x&&*+[o][F]{c}\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_x&&*+[o][F]{d}}$$
Then as $x\neq r-1$ there is another vertex $v$, not fixed by $\Gamma_{>x}$, incident to one of the vertices of one of the two graphs above.
If $\Gamma$ contains the graph (1), $v$ must be attached to it by a $4$-edge, which is not possible. This rules out graph (1).
In the second graph, the two components of the figure are not adjacent, since otherwise $c$ and $d$ are both fixed by $\Gamma_{>x}$.
Suppose first that $c$ is the vertex not fixed by $\Gamma_{>x}$.
Then there is another component $U_l$ adjacent to the $\Gamma_{>4}$-orbit containing the vertex $d$, as in the following figure.
$$\[email protected]{&& && && && && && && &&*+[o][F]{}\ar@{-}[rr]^6\ar@{-}[d]^4&& *+[o][F]{}\ar@{.}[rr]\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[rr]^{x}\ar@{-}[d]^4&&*+[o][F]{}\ar@{-}[d]^4&U_l\\
*+[o][F]{}\ar@{=}[rr]_0^2&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{}\ar@{-}[rr]_3&&*+[o][F]{}\ar@{-}[rr]_4&&*+[o][F]{}\ar@{-}[rr]_5&&*+[o][F]{}\ar@{-}[rr]_6&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{x}&&*+[o][F]{d}&}$$
But $\Gamma_{>x}$ is fix-point-free in $U_l$, a contradiction. This shows that $x\leq\frac{n-|X|-1}{2}$.
If $d$ is the vertex not fixed by $\Gamma_{>x}$, then the component adjacent to the component having the vertex $c$ must be fixed by $\Gamma_{>x}$, giving a contradiction as before.
This finishes the case where some component does not have a 2-fracture graph.
We now consider that each group $G_s$ has a 2-fracture graph for $F_s$. Particularly each component $U_s$ with $F_s \ne \emptyset$ has at least four vertices.
In addition let $\mathcal{F}$ be a fracture graph satisfying the property (P1) and such that $E:=\{F_s\,|\, F_s=\emptyset, s\in\{1,\ldots,c\}\}$ has maximal size.
Denote by $S$ and $\delta_s$ the following numbers.
$$\delta_s= |F_s|-\frac{|U_s|}{2}\mbox{ and } S:=\displaystyle\sum_{s=1}^c\delta_s.$$
As $G_s$ has a 2-fracture graph for $F_s$, $\delta_s\leq 0$ for all $s\in\{1,\ldots, c\}$.
Let $U_p$ be the component such that $t+1\in F_p$.
In Proposition~\ref{t} we proved that $(G_s)_{t+1}$ cannot be transitive in $U_s$. Therefore a fracture graph for $G_s$ with $s\neq p$
is disconnected. Hence $\delta_s\leq -0.5$ for $s\neq p$.
If $S\leq -3$, then $r-1-t\leq \frac{|U|}{2}-3$. For $(i,t)=(0,1)$, $|U|\leq n-2$ hence $r-2\leq \frac{n-2}{2}-3$. For $(i,t)=(3,4)$, $|U|\leq n-7$ hence $r-5\leq \frac{n-7}{2}-3$. In any case $r\leq \frac{n-3}{2}$. In what follows we prove that either $S\leq -3$ or we have (a) of the statement of this proposition.
Note that if $F_s\neq \emptyset$, then as $G_s$ has a 2-fracture graph for $F_s$, $|U_s|$ has at least four vertices.
As $\rho_i$ is an even permutation and fixes the first $\Gamma_i$-orbit except the vertex $a$, it must act nontrivially as an odd permutation in $U$.
In what follows we consider separately the following cases: (1) $\rho_i$ swaps an odd number of pairs of vertices $(v,w )$ with $v\in U_s$ and $w\in U_x$ with $s\neq x$;
(2) $\rho_i$ acts as an odd permutation inside a component $U_s$ with $F_s\neq \emptyset$ and $s\neq p$; (3) $\rho_i$ acts as an odd permutation inside $U_p$; (4) $\rho_i$ acts as an odd permutation inside a component $U_s$ with $F_s= \emptyset$.
(1) If $\rho_i$ swaps an odd number of pairs of vertices $(v,w )$ with $v\in U_s$ and $w\in U_x$, then $|U_s|=|U_x|$ is odd. If $|U_x|=|U_s|=3$ then $F_x=F_s=\emptyset$, hence $S\leq -3$. Consider $|U_x|=|U_s|\geq 5$.
As $|E|$ is maximal, either $F_x=\emptyset$ or $F_s=\emptyset$. We may assume that $|F_x|=0$. Then $|F_s|< \frac{5}{2}$.
Thus $\delta_x\leq -2,5$ and $\delta_s\leq -0,5$. Hence $S\leq -3$.
(2) Consider now that $\rho_i$ acts as an odd permutation inside a $\Gamma_{>t}$-orbit $U_s$ with $F_s\neq \emptyset$ and $s\neq p$.
In this case $\rho_i$ centralizes $G_s$, therefore $|U_s|$ is even and $|U_s|\geq 6$.
Moreover there is a 2-fracture graph for $G_s$ with labels in $S:=F_s\cup\{t+1\}$ being disconnected and having no cycles.
Hence $|F_s|+1\leq \frac{|U_s|-2}{2}$, thus $\delta_s\leq -2$.
Suppose $\delta_s=-2$. In that case the permutation representation graph of $G_s$ is as follows, where $x$ is the maximal label in $U_s$ and $k\in\{t+1,\ldots, x\}$.
$$ \[email protected]{*+[o][F]{} \ar@{-}[rr]^{t+1} \ar@{.}[dd]_i &&
*+[o][F]{}\ar@{.}[rr] \ar@{.}[dd]_i&&
*+[o][F]{} \ar@{-}[rr]^k\ar@{.}[dd]_i&&
*+[o][F]{} \ar@{-}[rr]^{k+1}\ar@{.}[dd]_i&&
*+[o][F]{}\ar@{-}[rr]^{k+2}\ar@{.}[dd]_i\ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[dd]_i\ar@<.3ex>@{-}[dd]^k \ar@{.}[rr] &&
*+[o][F]{} \ar@{-}[rr]^x \ar@{.}[dd]_i \ar@<.3ex>@{-}[dd]^k &&
*+[o][F]{}\ar@{.}[dd]_i\ar@<.3ex>@{-}[dd]^k \\
&& && && && && && &&\\
*+[o][F]{} \ar@{-}[rr]_{t+1} && *+[o][F]{}\ar@{.}[rr] &&*+[o][F]{} \ar@{-}[rr]_k && *+[o][F]{}\ar@{-}[rr]_{k+1}&& *+[o][F]{}\ar@{-}[rr]_{k+2} &&*+[o][F]{} \ar@{.}[rr] &&*+[o][F]{}\ar@{-}[rr]_x && *+[o][F]{}}
$$
Suppose that $x\neq r-1$. Then $\Gamma_{>x}$ fixes $U_s$, the vertex $b$ and the first $\Gamma_i$-orbit. Hence there exists $x\in\{0,\ldots, r-1\}$ as in statement (a) of this proposition. Thus we may consider that $\delta_s< -2$. As in this case $|U_s|$ is even, $\delta_s$ is an integer thus $\delta_s\leq -3$ and $S\leq -3$.
If $x=r-1$, we have $r-1-t\leq \frac{|U_s|}{2}-1$. For $(i,t)=(0,1)$, $|U_s|\leq (n-2)-4$ where $4$ is the minimal size of $U_p$. For $(i,t)=(3,4)$, $|U_s|\leq (n-7)-4$.
In any case we get $r\leq \frac{n-3}{2}$.
(3) If $\rho_i$ acts as an odd permutation inside $U_p$, then a 2-fracture graph of $G_p$ is disconnected without cycles.
Hence $|F_p|\leq \frac{|U_p|-2}{2}$, that is $\delta_p\leq -1$. Suppose we have the equality $|F_p|=\frac{|U_p|-2}{2}$. Then the permutation representation graph of $G_s$ is as the permutation representation graph given in case (2).
The only difference is that in this case $t+1\in F_p$. In this case we get exactly the same result as before, that is statement (a) of this proposition.
Assume now that $\delta_p<-1$. As in case (2), $|U_p|$ is even, thus $\delta_p$ is an integer. Then we may assume that $\delta_p\leq -2$.
Suppose that $t+2\notin F_p$. If $t+2\notin I_p$ or $(G_p)_{t+2}$ is transitive in $U_p$, then $\rho_{t+1}$ also centralizes $G_p$. But then there exists $l\in I_p\cap I_x$ with $I_x$ being the set of labels of a component $U_x$ adjacent to $U_p$. Then a 2-fracture graph of $G_x$ has at least three components, hence $\delta_x\leq -1$. Thus $\delta_p+\delta_s\leq -3$ and $|S|\leq -3$. If $t+2\in I_p$ and $(G_p)_{t+2}$ is intransitive in $U_p$, then $G_p$ has a disconnected 2-fracture graph without cycles for $F_s\cup \{\rho_{t+2}\}$. Hence $2(|F_p|+1)\leq |U_p|-2$. Moreover if we have the equality then $G_s$ is as the permutation representation graph given in case (2). Hence $2(|F_p|+1)< |U_p|-2$ and $\delta_p\leq -3$.
We now consider that $t+2\in F_p$.
Suppose $|S|=-2.5$. Then $c=2$ and the second component $U_s$ is such that $\delta_s= -0.5$. In particular, $F_s\neq \emptyset$, for otherwise $\delta_s\leq -1$.
First suppose that $(G_s)_{t+2}$ is transitive in $U_s$. Then $\rho_{t+1}$ centralizes $G_s$. Hence $2|F_s|\leq |U_s|-2$ and $\delta_s\leq -1$, a contradiction.
Thus $G_{t+2}$ is intransitive in $U_s$. As $t+2\notin F_s$, a 2-fracture graph for $G_s$, with labels in $F_s$, has at least 3 components with at most one cycle, hence $2|F_s|\leq |U_s|-2$, as before, a contradiction.
(4) Suppose that $\rho_i$ acts as an odd permutation inside a component $U_s$ with $F_s=\emptyset$. Either $|U_s|=2$ or $|U_s|\geq 6$. But in the second case clearly $\delta_s\leq -3$, thus we consider $U_s=\{u,v\}$. Let $l$ be the number of components of size two, whose vertices are swapped by $\rho_i$. We need to consider the case $l$ odd. If $l\geq 3$ then $S\leq -3$, hence we assume $l=1$. Moreover let $\rho_i$ be a 2-transposition, $\rho_i=(a\,b)(u\,v)$. Then $I_s=\{t+1\}$.
$$ \[email protected]{
*+[o][F]{} \ar@{-}[r]^t & *+[o][F]{u} \ar@{=}[r]^{t+1}_{i=t-1} & *+[o][F]{v} \ar@{-}[r]^t & *+[o][F]{} }$$
If $t+1=r-1$ then $\Gamma$ has one of the following permutation representation graphs.
$(i,t)=(0,1):$
$$ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2& *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{.}[r] &*+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2&*+[o][F]{} } $$
$(i,t)=(3,4):$
$$ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4& *+[o][F]{} \ar@{=}[r]^3_5 &*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} \ar@{-}[r]^4&*+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} \ar@{.}[r] &*+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} }$$
As by hypotheses $\Gamma_{>i}$ has a 2-fracture graph, for $(i,t)=(0,1)$, we have $n\geq 9 $, and for $(i,t)=(3,4)$, we have $n\geq 15$. In both cases $r\leq \frac{n-3}{2}$.
Assume $t+1\neq r-1$.
In this case $\Gamma_{>t+1}$ fixes the first $\Gamma_i$-orbit and $\{b,u,v,u\rho_t, v\rho_t\}$.
If $|\{b,u,v,u\rho_t, v\rho_t\}|=5$, consider $x=t+1$. Then $x$ satisfies (a) of this proposition.
Otherwise, $|\{b,u,v,u\rho_t, v\rho_t\}|=4$. Then $\Gamma$ contains one of the following graphs.
\vspace{5pt}
For $(i,t)=(0,1)$:
$$ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{.}[r] & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{}\ar@{.}[r]& *+[.][F]{} }$$
For $(i,t)=(3,4)$:
$$ \xymatrix@-1pc{
*+[o][F]{} \ar@{-}[r]^0 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{=}[r]^0_2 & *+[o][F]{} \ar@{-}[r]^1 & *+[o][F]{} \ar@{-}[r]^2 & *+[o][F]{} \ar@{-}[r]^3 & *+[o][F]{} \ar@{-}[r]^4& *+[o][F]{} \ar@{=}[r]^3_5 &*+[o][F]{} \ar@{-}[r]^4 & *+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{} \ar@{.}[r] & *+[o][F]{} \ar@{-}[r]^5 & *+[o][F]{}\ar@{.}[r]& *+[.][F]{}}$$
In the first case, $\Gamma_{>2}$ has at least five fixed points. Therefore $x=2$ satisfies the statement (a) of this proposition.
Consider the second case. As $B=\Gamma_{>3}\cong A_{n-6}$, $\Gamma_{>2}\cong A_{n-5}$. In addition $\Gamma_{<5}\cong A_{10}$ or $\Gamma_{<5}\cong A_{10}\times \langle \tau\rangle $ where $\tau$ is an even involution. Thus $\Gamma_{>2}\cap\Gamma_{<5}$ is either $A_5$ or $A_5\times \langle \tau\rangle$. But $\langle \rho_3,\rho_4\rangle\cong D_5\times \langle \tau\rangle$, contradicting the intersection condition.
\end{proof}
\begin{prop}\label{specials}
Let $n\geq 12$. If $A$ is trivial or $A$ has the permutation representation graph (2) or (3) of Table~\ref{small} and $B$ is an alternating group, then $r\leq \frac{n-1}{2}$.
\end{prop}
\begin{proof}
Assume for contradiction that $r > \frac{n-1}{2}$.
By the dual of Proposition~\ref{ineq0Btrivial} we may consider, up to duality, that when $A$ is trivial, $i=0$.
By Propositions~\ref{0trivialtran}, \ref{12Tran} and \ref{specialbubbles} we may assume that $\Gamma_{>i+1}$ is intransitive on $\{1,\ldots,n\}\setminus {\rm Fix}(\Gamma_{>i+1})$, but $\Gamma_{>i}$ does not have a 2-fracture graph. In addition, by Propositions~\ref{Case2} and \ref{case3} we can consider that if $\Gamma_j$ has exactly two orbits and a single $j$-edge connecting them, then the group orbits are either $C=\Gamma_{r-1}\cong A_{n-1}$ and $D$ trivial, or $C=\Gamma_{<j}\cong A_{n-6}$ and $D=\Gamma_{>j}\cong A_5$, $D$ having the permutation representation graph dual of (2) or the graph (3).
Consider the group $C$. Let $r_C$ be the rank of $C$ and $n_C$ the degree of $C$. Thanks to the intersection condition $\Gamma_{i+1,\ldots,j-1}$ is the alternating group.
In addition $C_{>i+1}$ is intransitive on the second $C_i$-orbit (the orbit containing the vertex $b$), for otherwise $\Gamma_{>i+1}$ is transitive on $\{1,\ldots,n\}\setminus {\rm Fix}(\Gamma_{>i+1})$. Thus $C$ satisfies the condition of Proposition~\ref{specialbubbles}.
Accordantly with that proposition there are three possibilities.
The first one is $r_C\leq \frac{n_C-3}{2}$ which implies $r\leq \frac{n-1}{2}$, a contradiction.
The second one gives the following permutation representation graph for $C$.
$$\[email protected]{ &&*+[o][F]{} \ar@{-}[rr]^1&&*+[o][F]{} \ar@{-}[rr]^2\ar@{-}[d]^0&&*+[o][F]{} \ar@{-}[rr]^3\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0 \ar@{.}[rr]\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{-}[rr]^{r_C-1}\ar@{=}[d]^1_0&&*+[o][F]{}\ar@{=}[d]^1_0\\
*+[o][F]{}\ar@{-}[rr]_0&&*+[o][F]{}\ar@{-}[rr]_1&&*+[o][F]{}\ar@{-}[rr]_2&&*+[o][F]{} \ar@{-}[rr]_3&&*+[o][F]{} \ar@{.}[rr]&&*+[o][F]{}\ar@{-}[rr]_{r_C-1}&&*+[o][F]{}}$$
But then, it is not possible to attach the permutation representation graph of $D$ by a single $j$-edge, a contradiction.
The last establishes that
there exists $x>i$ such that $x\leq \frac{n-|X|-1}{2}$ with $X:=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{>x})$.
On the other hand the dual of Proposition~\ref{specialbubbles} gives the same three possibilities for $B$ and, as before, one of them gives $r\leq \frac{n-1}{2}$, and the second gives a contradiction. Thus it may be assumed that there exists $y>i$ such that $r-y\leq \frac{n-|Y|-1}{2}$ with $X:=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{<y})$. If $\Gamma_{\{x+1,\ldots,y-1\}}$ is intransitive on $X\cap Y=\{1,\ldots, n\}\setminus {\rm Fix}(\Gamma_{\{x+1,\ldots,y-1\}})$ then it has a disconnected 2-fracture graph. Hence $y-1-x\leq \frac{|X\cap Y|-1}{2}$. Otherwise, as $\Gamma_{\{x+1,\ldots,y-1\}}$ has a 2-fracture graph it cannot be one of the graphs (2) or (3) of Table~\ref{small}. Thus by Propositions~\ref{small123} and \ref{induction}, we also have $y-1-x\leq \frac{|X\cap Y|-1}{2}$. As $n= |X| + |Y| - |X \cap Y|$, we get $r\leq \frac{n-1}{2}$, a contradiction.
\end{proof}
The cases we have covered, that some $\Gamma_i$ is primitive, or transitive
imprimitive, or all $\Gamma_i$ are intransitive and $2$-fracture graphs do or do
not exist, exhaust all possibilities; so Theorem~\ref{maintheorem} is proved.
\section{Acknowledgements}
This research was supported by a Marsden grant (UOA1218) of the Royal Society of New Zealand, and by the Portuguese Foundation for Science and Technology (FCT-Funda\c{c}\~{a}o para a Ci\^{e}ncia e a Tecnologia), through CIDMA - Center for Research and Development in Mathematics and Applications, within project UID/MAT/04106/2013.
\bibliographystyle{amsplain}
|
\section{Introduction}
\subsection{Faraday Rotation in Radio Sources}
There are two ways of probing the geometry of magnetic fields in
astrophysical plasmas: (i) the intrinsic orientation of the polarization
angle of synchrotron radiation provides information about the projection
of the synchotron magnetic field onto the plane of the sky and (ii)
Faraday rotation (FR) --- the wavelength-dependent rotation of the intrinsic
polarization angle of linearly polarized radiation as it traverses a
magnetized plasma --- provides information about the line-of-sight
component of the magnetic field in the region of Faraday rotation.
Because Faraday rotation provides information only about the
line-of-sight magnetic field, the full three-dimensional (3D)
structure of the field cannot be deciphered uniquely solely from
direct or synthesized FR observations.
This drawback can, in
principle, be overcome by modeling the polarized emission regions
and the embedded magnetic field and projecting model Faraday
rotation-measure (RM)
maps onto the observer's sky to enable direct comparisons (Murgia
et al. 2004; Laing et al. 2006, 2008; Guidetti et al. 2010; Govoni
et al. 2010; Bonafede et al. 2010). Model RM maps constructed in
this way use a minimal set of assumptions to avoid imposing
preconceived notions about the plasma and the magnetic field and
their interactions with the surrounding intergalactic environment.
This type of modeling for galaxy clusters implies mean inter-cluster
magnetic-field
strengths of a few to several tens of $\mu$G and the presence of a
random component on scales of $\sim$1~kpc in radio galaxies that
are members of large clusters and small groups (Laing et al. 2006,
2008; Guidetti et al. 2010; Bonafede et al. 2010).
The polarization structure observed in the jets of extragalactic
radio sources can be used to construct
models for the structure of the underlying magnetic field in the jets.
Such models of the magnetic field and the FR of the
radiation emitted by the jets and the lobes of
extragalactic radio sources have been broadened in scope by including
physical assumptions about the 3D structure of the field, the
kinematics and emissivity of the jet outflow, and the presence of
cavities around the jets (Laing et al. 2006, 2008).
\subsection{Magnetic fields in jets}
It is now well understood that an active galactic
nucleus (AGN) consists of two components: a
rotating supermassive ($\sim 10^8 M_\odot$) black hole with its
event horizon extending out to $\sim$1~AU, and a surrounding
rotating accretion disk extending out to $\sim$1~pc.
It is widely believed that the physical mechanism that drives these
systems is magnetohydrodynamical: the rotating accretion disk is
threaded by a magnetic field that
is wound up by the differential rotation of
the disk plasma (i.e., the field becomes helical), driving a collimated
outflow (jet) above and below the symmetry plane of the disk. The
asymptotic velocity attained by the jet material is of the same
order as the rotational velocity at the base of the jet. The jets
are also expected to include two components:
an inner, relativistic, axial jet and an outer, nonrelativistic,
extended disk wind (Pelletier et al. 1988; Ferreira et al. 2006).
The theoretical details of the jet launching mechanism have been
well-studied in the literature (Blandford \& Payne 1982; Contopoulos
\& Lovelace 1994; Contopoulos 1995; see also the review by
Spruit~2010 and references therein). It is expected that
the magnetic field is efficiently wound up only beyond the
Alfv\'en distance, which is $\sim$10 times the
radial extent of the outflow at its base. Thus, these theoretical
models suggest that the magnetic fields
of the inner jet and the outer extended wind will develop
significant toroidal components on distances beyond $\sim$10~AU and
$\sim$10~pc, respectively, from the base of the jet (see Fig.~1).
\begin{figure}[t]
\centering
\includegraphics[width=0.35\textwidth,angle=90]{CB-newfig-alfven.eps}
\caption{Schematic of the key transition points along the jet axis (not to
scale): the
Alfven point for the inner field, which lies roughly 10~AU from the jet base,
and the Alfven point for the outer field, which lies roughly 10~pc from the
jet base. The jet base is located further to the left.
Ellipses representing the azimuthal component of the inner field
are shown in black, and ellipses representing the azimuthal component of
the outer field in blue. In each case, the orientation of the field is
shown by corresponding arrows. The associated currents are shown by the
dashed lines. CW RM gradients and inward currents dominate on parsec scales,
while CCW RM gradients and outward currents dominate on decaparsec--kiloparsec
scales.} \label{Figure-Alfven}
\end{figure}
The old idea that loops of weak
field can be generated randomly by the Biermann (1950) battery in
the accretion disks, where they are amplified by turbulent dynamos,
so that their tangled fields are embedded in the outflows, has not
been successful in explaining the observations. Recent radio observations
indicate the presence of magnetic fields organized on scales of at least
$\sim$10--30 kpc in AGN (e.g., Carilli \& Taylor 2002; Widrow 2002; Eilek
2003; Kronberg 2005, 2010). Furthermore, the inefficacy of this
mechanism has been recognized in various theoretical studies (e.g.,
Vainshtein \& Rosner 1991; Subramanian 2008; Kulsrud \& Zweibel
2008). Despite these problems, the Biermann battery continues to
be adopted as the physical basis in investigations of magnetic-field generation in AGN, because of the perceived lack of an alternative
mechanism.
\subsection{The Cosmic Battery}
\begin{figure}[t]
\centering
\includegraphics[width=0.35\textwidth]{CBplot3.eps}
\caption{Schematic of the helical-field structure of an AGN for
arbitrary observer orientation as predicted by the Cosmic Battery.
The central black hole is represented by the gray sphere, the surrounding accretion disk by the flattened
yellow disk, and the axis of the system by the gray line. The white arrow indicates the direction of the
disk/black hole rotation. Inner field lines threading
the black hole are yellow and green. Outer field lines threading the
disk are orange and red. The Cosmic Battery predicts that the axial
direction of the inner/outer field is along/opposite to
$\boldsymbol\omega$, respectively (eq.~[\ref{muB}]). This corresponds
to an axial electric current flowing inward along the jet axis and
outward farther out where the outer
field lines are drawn. Field lines are wound by the black hole/disk
rotation. An observer will see the transverse RMs
increasing on the sky CW/CCW relative to the jet base for the inner/outer azimuthal field.
} \label{Figure2}
\end{figure}
In the past few years, a promising alternative to the Biermann
battery has been found, which seems capable of generating strong,
ordered, large-scale magnetic fields in AGN accretion disks and of
providing the physical background needed to understand
some of the most prominent features observed in magnetized jets.
This alternative mechanism relies on the Poynting--Robertson drag
on plasma electrons to generate large--scale azimuthal currents in
the inner disks of AGN (Contopoulos \& Kazanas 1998; Contopoulos
et al. 2006; Christodoulou et al. 2008;
Lynden-Bell~2013; Koutsantoniou \& Contopoulos~2014; Contopoulos
et al.~2015). Its most important element is that it supports and
maintains a large-scale poloidal magnetic field whose direction is
inextricably tied to the rotation of the disk (see also
Lynden-Bell~2013). In the vicinity of a $10^8 M_\odot$ black hole,
this mechanism can generate a field corresponding to equipartition
between the magnetic and relativistic-particle energies on timescales of
the order of one billion years (Contopoulos \& Kazanas 1998; Contopoulos
et al.~2015). We call this mechanism the Cosmic Battery (hereafter CB).
According to the CB, the Poynting-Robertson drag force on the
electrons at the inner edge of the accretion disk around an AGN's
supermassive black hole\footnote{In fact, around an accreting black
hole of any size; see e.g., Contopoulos \& Kazanas (1998), Kylafis
et al. (2012).} generates a toroidal electric current which gives
rise to poloidal magnetic field loops around the inner edge of the
disk with a magnetic moment $\boldsymbol{\mu_B}$. The outer footpoint of
each loop resides well inside the accretion disk and diffuses
outward on the disk's local diffusion timescale, provided that this
is at least about a factor of two shorter than the local accretion
timescale. The inner footpoint is dragged inward by the accretion
flow and eventually ends up near the black hole. Under these
conditions, the magnetic flux that accumulates in the vicinity of
the black hole due to the continuous (secular) growth of the inner
field always points in the same direction as the angular
velocity vector $\boldsymbol\omega$ of the accretion disk:
\begin{equation} \boldsymbol{\mu_B} \parallel
\boldsymbol\omega\ . \label{muB}
\end{equation}
The footpoints of each poloidal loop also participate in
the rotation of the disk, thus they are both dragged along
the direction of rotation. The resulting
wound-up field configuration corresponds to a large-scale axial
electric current ${\cal I}_{\rm inner}$ that flows toward the disk
along the symmetry axis of the inner (core) jet, i.e.,
\begin{equation}
{\cal I}_{\rm inner} \ \mbox{flows opposite to the jet direction},
\end{equation}
and a large-scale axial electric current ${\cal I}_{\rm outer}$
that outflows from the disk and into the outer wind, i.e.,
\begin{equation} {\cal I}_{\rm outer} \
\mbox{flows along the jet direction}
\end{equation}
(see Fig.~1). The central jet current closes along the interface between
the core
jet and the outer wind, whereas the outer wind electric current
closes farther out (not along the inner core jet).
This universal configuration of the axial electric current
circuit in extragalactic jets, predicted by Contopoulos et
al.~(2009) (hereafter CCKG), is unique to the CB mechanism and it was
confirmed numerically by the simulations of Christodoulou et al.
(2008) and Contopoulos et al.~(2015).
\subsection{Observational signatures of the Cosmic Battery}
The most obvious diagnostic tool of the above magnetic-field
configuration is the observation of characteristically oriented
transverse Faraday RM gradients across the jets
and the lobes of at least some radio sources that are not too
disturbed by interactions with their environment or other
internal factors. As depicted in Figs.~1 and 2, when the observed RMs
increase in the clockwise (CW) direction on the sky relative to
the base of the jet outflow, the electric
current flows inward opposite to the jet direction, whereas if the
RMs increase in the counterclockwise (CCW) direction on
the sky relative to the jet base, the electric current flows
outward along the jet direction. In summary:
\begin{equation} \mbox{RM increases CW} \ \Longrightarrow \ {\cal I}_{\rm inner} \
\mbox{opposite to the jet direction},
\end{equation}
\begin{equation} \mbox{RM increases CCW} \ \Longrightarrow \ {\cal I}_{\rm outer} \
\mbox{along the jet direction}.
\end{equation}
These results are independent of observer location
(see Fig.~2 and CCKG for more details). Now, as was pointed out above
(see Fig.~1),
the inner field should be efficiently wound up only outside the
corresponding Alfv\'en point, meaning at distances greater than about
10~AU from the jet base; in terms of observed transverse Faraday RM
gradients due to the systematic change in the line-of-sight component
of the azimuthal field, we expect the inner azimuthal field and
${\cal I}_{\rm inner}$ to give rise to significant transverse Faraday
RM gradients only at distances greater than about 10~AU from the jet
base. Analogously, the outer field should be efficiently wound up only
at distances greater than about 10~pc from the jet base; accordingly,
we expect the outer azimuthal field and
${\cal I}_{\rm outer}$ to begin making significant contributions to
the Faraday RM gradients only at distances greater than about 10~pc
from the jet base. This suggests that the net Faraday rotation due
to both the inner and outer azimuthal fields (i.e., the net transverse
RM gradient), should be dominated by the inner azimuthal field at
distances out to about 10~pc from the jet base, then by the outer
azimuthal field starting at distances of about 10~pc or more (see Fig.~1).
There could also be a region at distances of a few tens of pc where no
clear transverse RM gradients are observed, because the contributions
from the inner and outer azimuthal fields are comparable.
CCKG looked for precisely this effect using published RM maps
for roughly 30 AGNs whose parsec-scale jets mapped with very long
baseline interferometry (VLBI) seemed to show reasonably clear transverse
RM gradients. CCKG reported evidence for a predominance of
CW transverse RM gradients within $\simeq 20$~pc of the center,
whereas CCW transverse RM gradients were found
in some of these sources at larger distances. These results were
disputed by Taylor \& Zavala (2010), who claimed that, in the vast
majority of cases, the VLBI jets were too poorly resolved to make
the observed transverse RM gradients reliable. However, the Monte
Carlo simulations of Hovatta et al. (2012) and Mahmud et al. (2013)
subsequently demonstrated that transverse RM gradients could be detected
even when the intrinsic jet width was much narrower than the beam
width for the VLBI array used, removing the doubts cast by Taylor
\& Zavala (2010). Previously firm results, recently reported new
results and reanalyses of a number of previously published RM
images applying the improved error estimation approach developed
by Hovatta et al. (2012) have now brought the list of reliable
(monotonic, with significances $> 3\sigma$) parsec-scale transverse
RM gradients
to 27, of which 20 are CW and 7 are CCW on the sky, relative to
their jet bases (Gomez et al.
2008; Kharb et al. 2009; Hovatta et al. 2012; Gabuzda et al. 2014a,
2014b, 2015a); a simple binomial probability distribution analysis
indicates that the probability of at least 20 out of 27 of the
observed transverse RM gradients having the same orientation (CW)
by chance is about $0.95\%$,
supporting the reports of CCKG of a predominance of CW transverse
RM gradients on parsec scales. In the sizeable minority of parsec-scale
jets displaying CCW transverse RM gradients (7 out of 27), the gradients
may be present at distances from the cores that are greater
than the transition distance of about 20~pc; a more
detailed analysis of this question will be considered by Gabuzda
et al. (in preparation). In
addition, it is always possible that the battery mechanism
is not efficient in some jet--disk systems.
Work is ongoing to try to add to the
list of AGN jets with reliable transverse RM gradients on parsec
scales (Gabuzda et al., in preparation).
A precise demarkation line between the distances at which the
inner and outer azimuthal magnetic fields dominate the observed
transverse RM gradients cannot be determined with certainty from
the available
VLBI data, but the outer (return) field with its CCW RM gradients
appears to dominate at distances larger than about 20~pc (Fig.~2
in CCKG), at least in those objects in which the RM gradients are
resolved and are not obscured by substantial random RM components.
\subsection{Outline of the paper}
With the above theoretical considerations in mind, we set out to
examine the available information about transverse RM gradients on
larger scales extending out to kiloparsecs, exceeding the expected
distance of the Alfv\'en point for the outer azimuthal magnetic
field from the jet base. The purpose of
this study is essentially twofold: (i) to establish whether or not
transverse RM gradients reasonably interpreted as reflecting the
presence of helical jet magnetic fields are present on scales of
tens to thousands of parsec and, if present, (ii) to establish
whether such RM
gradients show any evidence for a preferred direction.
In Section~2, we describe observations for three sources we
have analyzed for this paper.
In Section~3, we consider three previous firm reports of transverse
RM gradients on scales exceeding about 20~parsec, for which quantative
analyses have already been carried out in previous publications,
to which we add
our new detection of a $3.0\sigma$ transverse RM gradient in A2142A. We
also consider
four reliable cases of reversals in the direction of the observed RM
gradients with distance from the jet base. Finally, we consider an
additional seven transverse RM gradients that are visible in previously published
RM maps, and present the results of quantitative analyses of the RM results
for three of these sources. All seven of these transverse RM gradients
must be considered tentative, due to insufficiently high significance
of the gradients ($< 3\sigma$) , lack of information about the uncertainties
in the observed RM values, and/or uncertainty in the orientation of the
gradients relative to the local jet direction.
In Section~4, we discuss the implications of these collected
results both in general and in the context of the CB mechanism.
Finally, we conclude in Section~5 with a
summary of our results and some remarks about related ongoing
research of the large-scale magnetic fields in AGNs and in our own
Galaxy.
\section{Observations and reduction}
\subsection{VLA Data}
A. Bonafede and F. Govoni kindly provided the calibrated visibility data
that had been used to make the published RM maps for A2142A (Govoni et al.
2010) and 5C4.152 (Bonafede et al. 2010). Data are available for 5C4.152
at 4.635, 4.835 and 8.275~GHz, and for A2142A at 4.535, 4.835, 8.085, and
8.465~GHz.
The observations and data calibration and reduction methods used
in the initial analyses carried out for these objects are given by
Bonafede et al. (2010) and Govoni et al. (2010).
We could not use the RM maps published by Bonafede et al. (2010) and
Govoni et al. (2010) directly, because the associated error maps did not
take into account the finding of Hovatta et al (2012) that the uncertainties
in the Stokes $Q$ and $U$ fluxes in individual pixels on-source are
somewhat higher than the off-source rms fluctuations, potentially
increasing the resulting RM uncertainties.
To address this, we imported the final, fully self-calibrated
visibility data into the {\sc AIPS} package, then
used these data to make naturally weighted $I$, $Q$ and $U$ maps
at all wavelengths, with matching image sizes, cell sizes and beam
parameters specified by hand in the {\sc AIPS} task {\sc IMAGR}.
These images were all convolved with a circular Gaussian beam having a
full-width at half-maximum of 3$^{\prime\prime}$. We obtained
maps of the polarization angle, $\chi = \frac{1}{2}\arctan(U/Q)$, and
used these to construct corresponding RM maps in the {\sc AIPS} and
{\sc CASA} packages. The uncertainties in the polarization
angles used to obtain the RM fits were calculated from the uncertainties in $Q$
and $U$, which were estimated using the approach of Hovatta et al. (2012).
In all cases, satisfactory RM fits were obtained without applying
$n\pi$ rotations of the observed polarization angles.
\subsection{VLBA data}
We also analyzed the RM distribution of 3C120, constructed using data
obtained with the Very Long Baseline Array (VLBA) at 1.358, 1.430, 1.493 and
1.665~GHz.
This RM map was originally published by Coughlan et al. (2010); again, we
were not able to use the previously published RM map directly because
the associated error maps did not take into account the improved method
of Hovatta et al. (2012) for estimating the on-source $Q$ and $U$
uncertainties in individual pixels.
We addressed this by using the $Q$ and $U$ maps used by Coughlan et al.
(2010) to produce new RM maps in both {\sc AIPS} and {\sc CASA}, assigning
$Q$ and $U$ uncertainties in accordance with the approach of Hovatta et
al. (2012).
In all cases, when differences in RM values across the jet were obtained, the
uncertainties of the RM values did not include the effect of uncertainty
in the polarization angle calibration, since this cannot introduce spurious
RM gradients (Mahmud et al. 2009, Hovatta et al. 2012). The
uncertainty of the difference between the RM values at the two ends of
a slice was estimated by adding the uncertainties for the two RM values
in quadrature.
\begin{figure}
\begin{center}
\hspace*{-1.0cm}
\includegraphics[width=.45\textwidth,angle=0]{A2142_RM_sigma_figure.eps}
\end{center}
\hspace*{0.3cm}
\caption{Intensity map at 4.535~GHz of A2142A with
the RM distribution
superposed (upper panel). The lowest contour is 1\% of the peak intensity of 8.8~mJy/beam,
the contours increase in increments of a factor of two, and the beam size
is shown in the lower left corner of the image. The black arrow across
the jet highlights the direction of the transverse RM gradients. The
gray box shows the region for which the significances of series of
parallel, monotonic transverse RM gradients are plotted in the lower
panel; the gray arrow outside the box pointing outward along the jet
shows the direction
of increasing pixel number in the lower panel, and pixel~0 corresponds
to the inner edge of the gray box. Pixel size is 0.25~arcsec.
The $2\sigma$ level is shown by the dashed blue horizontal line, and the
$3\sigma$ level by the dashed red horizontal line.}
\end{figure}
\section{Transverse RM gradients on scales exceeding $\simeq$~20~pc}
\subsection{Previously published firm gradients}
Virtually all the work done on transverse RM gradients has been
carried out for high-resolution data obtained on the VLBA at wavelengths between 15 and 5~GHz; most of the
RM gradients detected cross their jets at projected distances from the
jet base of no more than a few milliarcseconds (less than about
20--30~pc).
Very few studies probing larger scales have been carried out, and
we will now consider the few results available here. In all cases,
the statistical significances of the transverse RM gradients have
been reliably estimated, and found to be at least $3\sigma$.
When we refer to a transverse RM gradient being
CW or CCW, we mean that its orientation is clockwise or
counter-clockwise on the sky, relative to its own jet base.
\begin{enumerate}
\item {\em 1652+398 (Mrk~501)}.---The RM map of this AGN based on VLBA data at
frequencies between 1.6 and 8.4~GHz published by
Croke et al. (2010) shows a very clear transverse RM gradient
throughout the extended jet before it turns sharply about 30~mas
(about 20~pc projected distance) from the core, oriented CCW. Croke et al.
(2010) present a quantative analysis of this gradient.
Even taking into account the fact that the RM uncertainties indicated
by Croke et al. (2010) were not obtained using the improved method
of Hovatta et al. (2012), and therefore could be up to about a factor
of two too small, the significance of this gradient is far in excess
of $3\sigma$.
\smallskip
\item {\em 3C380}.---The RM map of this AGN based on VLBA data at wavelengths
between 1.4 and 5.0~GHz published by Gabuzda et al. (2014a) shows a
transverse RM gradient oriented CCW; the quantative analysis carried
out by Gabuzda et al. (2014a) shows that this gradient has a significance
of about $4\sigma$.
\smallskip
\item {\em 5C4.114}.---An arcsecond-scale VLA RM map for the kiloparsec-scale
jets of 5C4.114 recently analyzed by Gabuzda et al. (2015b) demonstrates
transverse RM gradients across both the northern and
southern jets with significances of about $4\sigma$ and $3\sigma$,
respectively, both oriented CCW.
\end{enumerate}
\subsection{New detection of a firm transverse RM gradient in A2142A}
The radio source A2142A is located in a cluster, and we therefore
expect some contribution to the RM distribution from the magnetized
intercluster gas; this should most likely not show any large-scale order,
and be predominantly ``patchy''. In addition to some patchiness, the
RM distribution of A2142A shows a
tendency for the RM values along the northern side of the jet to be
less negative than those along the southern side (Fig.~13 of
Govoni et al. 2010), corresponding to a possible gradient in the RM
values across the jet. We note that the core is at the eastern end of
the observed radio structure, so that this implied transverse RM
gradient is oriented CCW.
Fig. 3 presents our 4.535-GHz intensity image of A2142A, with the RM image
superposed in color; this essentially reproduces the images in Fig.~13
of Govoni et al. (2010). The output pixels in the RM map were blanked when the
RM uncertainty resulting from the $\chi$ vs. $\lambda^2$ fits exceeded
80~rad/m$^2$. Our analysis of the entire RM distribution showed the
presence of monotonic RM gradients across the jet in the region surrounded
by the gray box. At the
redshift of A2142A, $z = 0.091$, these correspond to projected distances of
roughly 10~kpc from the jet base. The points in the lower panel
of Fig.~3 correspond to monotonic transverse RM gradients obtained
for a series of parallel RM slices across the jet, inside the gray box in
the upper panel.
As can be seen, comparisons
of the RM values at the two ends of the RM slices considered indicate that
the transverse RM gradients about 4~arcsec from the start of the boxed region
have significances reaching about $3.5\sigma$, with another region of
gradients reaching nearly $2\sigma$ slightly further out from the core.
Thus, we consider this a firm case of a transverse RM gradient on
kiloparsec scales.
\subsection{Previously published firm RM-gradient reversals}
To the four sources in Sections~3.1 and 3.2, we add information from reports of
reversals in the directions of the observed transverse RM gradients
in four more AGNs: 0716+714 (Mahmud et al. 2013, reversal at a projected
distance of a few pc from the jet base), 0923+392 (Gabuzda et al. 2014b,
reversal at a projected distance of about 15~pc from the jet base),
1749+701 (Mahmud et al 2013, reversal at a projected distance of about
35~pc from the jet base), and 2037+511 (Gabuzda et al. 2014b, reversal about
35~pc from the jet base). Although the observed reversal in the
direction
of the observed transverse RM gradient in 0716+714 was relatively close
to the core, the 1.4--1.7~GHz data analyzed by Healy (2013) show that the
orientation of the outer gradient was maintained to projected distances
of about 35~pc from the core. In all four cases, quantitative analyses
carried out in the papers cited above indicate that both transverse RM
gradients detected had significances of at least $3\sigma$, and
the inner gradient was oriented CW and the outer gradient CCW relative
to the jet base.
\subsection{Tentative gradients from an initial inspection of RM maps
in the literature}
The transverse RM gradients in 5C4.114 and A2142A listed in Sections~3.1
and 3.2 were initially identified via an initial inspection of
published Faraday RM maps
of extragalactic radio sources on kiloparsec scales, carried out for
85 objects
(Athreya et al. 1998; Best et al. 1998, 1999; Feretti et al.
1999; Venturi \& Taylor 1999; Taylor et al. 2001; Eilek \& Owen
2002; Goodlet et al. 2004; Govoni et al. 2006; Laing et al. 2006,
2008; Kharb et al. 2008; Guidetti et al. 2008, 2010; Feain et al.
2009; Kronberg 2009; Bonafede et al. 2010; Govoni et al. 2010;
Kronberg et al. 2011; Algaba et al. 2013).
In all cases, the authors of these
studies took care to ensure that the RM images were constructed
using only data with sufficiently high signal-to-noise ratios, and
that the resulting $\lambda^2$ fits for the RM values were
sufficiently good. Unfortunately, full information about the
uncertainties in the RM values at individual points in the RM
maps is not available; therefore we were able to identify
candidate transverse RM gradients, but have not been able to carry out
quantitative analyses to test their significances.
We were unable to identify any obvious large-scale monotonic
transverse RM gradients in 77 of these 85 published
maps, in which (a)~the jets appeared to be strongly influenced by
their surroundings (they were strongly bent or appeared to be
partially disrupted); (b)~the RM distribution was patchy; (c)~limited to
just a few pixels across or along the images (insufficient resolution);
(d)~the RM images were based on only two
frequencies; or (e)~the results were shown in gray scale that did
not make the structures in the RM images sufficiently clear.
We sought to obtain higher-resolution or
color images from some authors, and we appreciate their efforts to
provide us with the maps we requested.
In the end, we identified eight potential objects whose RM maps showed
extended, monotonic, nearly transverse RM gradients. Of these,
two (5C4.114 and A2142A) have now been shown to be
statistically significant (Gabuzda et al. 2015b; this paper), leaving
six additional tentative gradients. To these we add results for 3C120
on slightly smaller scales of 30--100~pc presented by Coughlan
et al. (2010). We have not attempted to measure or analyse the
magnitudes of these tentative RM gradients, which would be
extremely difficult to interpret unambiguously due to convolution with the
observing beam. Instead, we focus on the presence of transverse gradients
visible in the RM distributions and their directions relative to
the bases of the jets, quantatively estimating the significances
of these gradients when possible.
We briefly describe the RM distributions of these seven sources, and
the basis for our suggestion that these may contain transverse RM
gradients. The patchiness of many of the 85 published RM distributions we considered is consistent with a picture in which random distributions
of the magnetic field and electron density in the general vicinity of the
radio source (e.g., in the cluster or inter-cluster medium in which the
source is located) generally dominate on these large scales; these seven
sources appear to be those in which an ordered RM pattern is dominant
in at least some regions.
As above, when we refer to a transverse RM gradient being
CW or CCW, we mean that its orientation is clockwise or
counter-clockwise on the sky, relative to its own jet base.
\begin{enumerate}
\item {\it 0156$-$252}.---There appears to be an RM gradient
across the eastern jet (Fig.~3 of Athreya et al.~1998), oriented in the CCW
direction. Unfortunately, we do not have access to these data and were
not able to quantatively analyze the significance of this gradient.
Therefore, this remains a tentative transverse RM gradient.
\smallskip
\item {\it 3C120}.---In contrast to the other objects considered
in this section, the Faraday-rotation map of Coughlan et al.
(2010) is based on VLBA data at four frequencies between 1.4 and
1.7~GHz, and thus probes slightly smaller angular scales. Transverse RM
gradients are present at several locations on scales of approximately
30--100~pc from the core, all oriented CCW.
\begin{figure}[H]
\begin{center}
\includegraphics[width=.45\textwidth,angle=0]{3c120_fig.eps}
\end{center}
\hspace*{0.3cm}
\caption{Top panel: Intensity map at 1.358~GHz of 3C120 convolved
with the ``intrinsic'' beam and the RM distribution superposed. The lowest
contour is 0.125\% of the peak of 1.37~Jy/beam. The middle panel
shows the same intensity and RM maps convolved with a circular beam of equal
area; the lowest contour is 0.125\% of the peak of 1.43~Jy/beam. In both cases,
the contour step is a factor of two, the pixel size is 0.25~arcsec, and
the beam size is shown in the lower left corner of the image.
The bottom panel shows a plot of the significances of a series of transverse
RM gradients calculated inside the gray boxes in the middle panel; the
$2\sigma$ level is shown by the dashed blue horizontal line. See text for
more detail.}
\end{figure}
\smallskip
The top panel of Fig. 4 presents our 1.358~GHz intensity and RM images of
3C120 (beam size $19.6~\textrm{mas}
\times 7.6~\textrm{mas}$ in position angle $27^{\circ}$); these essentially
reproduce the images of Coughlan et al. (2010). The middle panel gives
a version of the same RM map convolved with a circular beam (beam
radius 12.2~mas).
In both cases, the output pixels in the RM maps were blanked when the
RM uncertainty resulting from the $\chi$ vs. $\lambda^2$ fits exceeded
20~rad/m$^2$. Both maps show very similar RM structures, and we
present our further analysis for the map convolved with the circular beam.
\smallskip
Our analysis of the entire RM distribution showed the presence of
monotonic RM gradients across the jet in the regions surrounded
by the gray boxes.
The black arrows in the middle panel highlight the main regions of
transverse RM gradients.
The points in the bottom
panel of Fig.~4 correspond to monotonic transverse RM gradients obtained
for a series of parallel RM slices across the jet, inside the gray boxes in
the middle panel.
The gray arrows pointing outward along the jet in the middle panel
show the direction of increasing pixel number in the lower panel: pixel~0
corresponds to the inner edge of the left-hand gray box, and pixel~50 to the
inner edge of the right-hand gray box.
We can see that these slices reach significances $\simeq 2\sigma$ at
several locations along the jet, but none reach $3\sigma$. Thus,
these transverse RM gradients cannot yet be considered firm. However, 3C120
remains on the list of sources displaying tentative transverse RM
gradients; deeper VLBA observations with a wider frequency range could
help determine whether these gradients are statistically significant.
\smallskip
\item {\it M87}.---Algaba et al.~(2013) suggest that there are clear
RM gradients across HST-1 and knots A and C. The gradients are all
directed CCW. Until a more detailed quantitative analysis is
available, this remains a tentative transverse RM gradient.
\smallskip
\item {\it 5C4.152}.---This source has an intriguing RM
distribution, with RM gradients present across both lobes
(Fig.~15 of Bonafede et al.~2010). The gradient across the
southern lobe seems to be fairly orthogonal to the jet direction,
while the gradient across the northern lobe appears to be offset
from orthogonality as the jet bends toward the east just before
the hot spot. If the gradient across the southern lobe can be taken
to be transverse to the jet, its direction is CCW.
\smallskip
We reconstructed the previously published RM map in order to estimate
the significance of these RM gradients. Our quantative analysis
of a series of RM
slices taken across the southern lobe indicates that the strongest
transverse RM gradients have significances of about $1.5\sigma$.
This demonstrates that these gradients cannot be considered
statistically significant. However, we suggest that it is appropriate to
retain 5C4.152 on the list of sources displaying tentative transverse RM
gradients, in particular, because the 8-GHz data of Bonafede et al. (2010)
had an atypically high noise level, and the uncertainties in the RM
values were relatively high as a result of the limited range and number of frequencies
used (8.2~GHz, 4.6~GHz, 4.8~GHz). Analysis of
kiloparsec-scale observations with lower noise levels and/or a wider
frequency range could help determine more conclusively whether these
gradients are significant or not.
\smallskip
\item {\it Cen A}.---After subtracting the overall mean
RM (which is presumably foreground), the residual RM map shows a
tendency for positive residuals on the eastern side and
negative residuals on the western side of the entire northern
lobe (Feain et al.~2009). This seems to imply an RM gradient that
extends for some 235~kpc across the northern lobe (Fig.~7 of Feain
et al.~2009), oriented CCW. The southern lobe is strongly
bent and no systematic trends are visible.
\smallskip
Although there appears to be a difference in the dominant signs
of the residual RM values on either side of the northern jet axis, the
RM values themselves do not form a clear gradient across the jet.
In addition, RM measurements are available only at the positions
of background sources, so that we do not have measurements of a continuous
RM distribution across the source. Nevertheless, we were able to test
whether the difference in the dominant RM signs was statistically significant
by dividing the northern jet/lobe into three sections across the jet,
encompassing the declination range from $-38^{\circ}$ to $-41^{\circ}$
and the right-ascension ranges 13$^h$\,15$^m$0$^s$--13$^h$\,21$^m$59$^s$
(western side), 13$^h$\,22$^m$0$^s$--13$^h$\,29$^m$59$^s$ (central region),
and 13$^h$\,30$^m$0$^s$--13$^h$\,36$^m$59$^s$ (eastern side). We then
determined the number of positive and negative RM values in the eastern
and western sections; this yielded 15/21 negative RM values in the
eastern section and 15/21 positive RM values in the western section.
\smallskip
A simple binomial probability analysis indicates that the probability of
obtaining 15 or more of 21 values of a particular single sign is
3.92\%; if we are interested in the probability of having this fraction
of either sign, this must be multiplied by two, and so increases to 7.84\%.
However, the probability of obtaining 15 or more of 21 values of one sign
on one side of the jet, and simultaneously 15 or more of 21 values of
the opposite sign on the other side of the jet, is the product of these
the two separate probabilities, or $(0.0784)(0.0392)(100) = 0.31\%$,
which corresponds to $3\sigma$.
\smallskip
Thus, our quantative analysis demonstrates that the asymmetry in
the signs of the RM values across the northern jet of Cen~A is statistically
significant at the $3\sigma$ level. However, because it is not possible to
test whether this transverse trend in the observed RM values is monotonic
and systematic due to the availability of RM measurements only at the
positions of background sources, we
retain Cen~A on the list of sources whose jets display tentative
transverse RM gradients.
\smallskip
\item {\it 3C303}.---Kronberg et al.~(2011) report the presence of a
transverse RM gradient at the location of knot~E3, and use information
about this gradient together with certain assumptions to estimate the
associated current in the jet. The RM gradient is clearly visible
in Fig.~3 of Kronberg et al. (2011), although it is quite narrow
and is not well resolved. The direction of the gradient is CCW. Although
this RM map has been published in the refereed literature, we
include 3C303 in our list of tentative gradients because the original
paper does not provide information about the uncertainties of the
single-pixel RM values that can be used to estimate the significance
of the RM gradient. We do not have access to these data,
and so were not able to carry out our own quantitative analysis
for this source.
\smallskip
\item {\it 3C465}.---The southern jet provides a clear
example of a systematic, extended transverse RM gradient that
extends along most of this jet (Fig.~6 of Eilek \& Owen~2002).
The color scale chosen for the RM distribution is such
that the RM map presented by Eilek \& Owen (2002) is essentially
an RM sign map. The orientation of the RM gradient across the
southern jet is CCW. The northern jet shows signs that it
intrinsically had an overall pattern, but the jet and the RM
pattern have become distorted. We do not have access to these data,
and so were not able to carry out our own quantitative analysis
for this source.
\end{enumerate}
Table~1 summarizes information about all the firm (upper rows)
and tentative (lower rows) transverse RM gradients considered in
this Section. The references in bold in Table~1 indicate the
papers in which quantitative analyses are carried out to determine
the statistical significances of these gradients.
Treating the detections of the transverse RM gradients across the northern
and southern jets of 5C4.114 as independent measurements, we have
in total nine firm cases of transverse RM gradients.
All these transverse RM gradients are oriented CCW
relative to their jet bases on comparatively
large scales, greater than about 20~pc.
\section{Discussion}
\subsection{Transverse RM gradients on scales out
to kiloparsecs from the central AGN}
The most fundamental implication of the results discussed above
is that transverse Faraday rotation gradients --- predicted to
exist if the jets carry toroidal or helical magnetic fields and
detected earlier in some 27 AGNs on parsec scales --- are
also present on considerably larger scales, extending out to
thousands of parsecs (e.g., 5C4.114; Gabuzda et al. 2015b). This
is an important result,
because if AGN jets carry helical magnetic fields, these
should be present on essentially all scales where the jets propagate,
provided that the intrinsic field structure of the jet is not disrupted
by interactions with the surrounding environment. We note that this
conclusion follows from the firm results listed in Sections~3.1--3.3,
and does not depend on the tentative identifications of
transverse RM gradients considered in Section~3.4.
At the same time, our literature search for transverse RM
gradients on kiloparsec scales yielded only eight tentative cases out of 85 objects considered, indicating that
it is comparatively difficult to detect transverse RM gradients
due to the presence of helical magnetic
fields in kiloparsec-scale jets. In fact, it is easy to understand
why this should be the case: there is an
appreciable turbulent, inhomogeneous component to the thermal ambient
media surrounding the jets on these scales, which superposes a more or
less random pattern over the systematic pattern due to the helical fields.
This random component in the RM distribution apparently
dominates in the majority of
cases. This makes it perfectly natural that most of the observed RM distributions
appear random and patchy, but the overall pattern due to the helical fields
sometimes comes through. This suggests that, on average, it may be easier
to detect the systematic RM component due to helical jet magnetic fields on
parsec scales, where the ordered inner field is more dominant, a result that
seems to be borne out from the observations (see also Section~4 below).
The tentative transverse RM gradients that we consider in
Section~3.4 were not, in most cases,
noticed or appreciated by the authors of the original papers; this is
primarily due to the fact
that their main interest was in the magnetic fields of the clusters in which the radio
sources are located, rather than in the individual radio sources themselves.
Eilek \& Owen (2002) did note the striking RM distribution in 3C465, but they
did not consider the possibility that it is associated with an embedded magnetic field
because the observed Faraday rotation is external to the main radiation source.
Bonafede et al. (2010) similarly suggested that the observed RM distributions are
due primarily to Faraday rotating material that is not associated with the radio
sources themselves, based in part on the observation that the observed Faraday
rotation is external; however, their arguments are based on general
considerations and statistical relations, rather than on an individual examination
of particular objects. Govoni et al. (2010) essentially assumed that the observed
RMs are associated only with the intracluster medium, and used the RM observations
to deduce the properties of this medium.
In our interpretation, the large-scale Faraday rotation for
the vast majority of extragalactic radio sources is indeed
dominated by turbulence and fluctuations in the cluster media; but
the eight sources we have identified as displaying evidence for
monotonic, extended, transverse RM gradients on scales exceeding about
20~parsec (including 5C4.114;
Gabuzda et al. 2014b) essentially represent
a handful of objects in which the dominant contribution to the
Faraday rotation in some regions is due to material associated
with the radio source itself, rather than the patchy cluster
media. In these cases, the observed Faraday rotation can still be
external (not occurring throughout the radiating volume of the
source), but it is nevertheless associated with regions carrying
the imprint of a helical magnetic field in the immediate vicinity
of each AGN jet. Direct evidence that the observed
Faraday rotation is external to the main jet volume is indicated by
the fact that the RM fits that were obtained when constructing the RM
maps do not show significant deviations from $\lambda^2$ behavior,
within the uncertainties (see, e.g., the plots presented by Bonafede et
al (2010), Croke et al. (2010), Govoni et al. (2010), Mahmud et al.
(2013), and Gabuzda et al. (2014a)).
In some cases, the three-dimensional structure of
the emitting regions may not be entirely clear; for example, it may
not be obvious whether the observed regions of the Faraday-rotation
gradients are associated with the outflowing jet or a possibly
back-flowing lobe structure. However, in either case, these
regions could carry the imprint of the helical magnetic field
carried by the jet, so that this uncertainty does not affect the
basic interpretation of the Faraday-rotation gradients we offer
here.
\subsection{Predominance of CCW RM gradients on large scales}
A striking feature of the results considered in Section~3.1--3.3
is that there is an obvious predominance of CCW RM gradients:
all nine firm gradients on scales exceeding about 20~pc, for all of
which quantative analyses have been carried out, are CCW
(upper part of Table~1). Based on a simple unweighted binomial probability
function, the probability for all nine of these gradients to be
CCW (in agreement with the CB mechanism) by chance is about 0.2\%,
which corresponds to about $3.1\sigma$.
In addition, all seven tentative gradients (Section~3.4, lower part
of Table~1) are CCW. Although the statistical significances of
most of these gradients are not known as we could not carry out
quantitative analyses of these data, this again appears not to be random,
in the same sense.
The probability that all seven of these tentative gradients
would be CCW by chance is about 0.8\%, corresponding to about
$2.65\sigma$. This suggests that at least some of these
currently tentative RM gradients may well eventually be proven to be
statistically significant, further strengthening the evidence for
the presence of a helical or
toroidal field component in these AGN jets on kiloparsec scales. We note
that one factor limiting the uncertainties in the RM measurements is
the requirement that we work with the RM values in individual pixels.
In principle, it should be possible to decrease the RM uncertainties by
averaging over some number of neighboring pixels, on a scale that
remains significantly smaller than the beam size; however, this is
not possible in practice as we do not fully understand
the correlations between the uncertainties in neighboring pixels. Some
initial progress on this problem has been made (Coughlan 2014), but
more work remains before it will be possible to carry out such averaging
while also obtaining accurate estimates of the uncertainty on the average.
Thus, the results for the nine firm transverse RM gradients are
statistically significant at the $3\sigma$ level, and this significance
will increase if any of the tentative transverse RM gradients
we have identified are subsequently shown to be firm detections.
In all four cases when reversals of the RM gradients
are observed along the jet of a single AGN (sources in Table~1 marked
with an asterisk), the gradients are CW on small scales and CCW on larger
scales, consistent with the prediction of the CB mechanism.
The obviously non-random distribution of
the observed transverse RM gradients' orientations relative to
their jet bases leads us to take seriously the idea that there
is a preferred orientation of transverse
RM gradients associated with helical jet magnetic fields.
Although this idea may seem
strange at first, it has a very natural interpretation, as we have already discussed in Section~1.4. The direction
of the RM gradient implies a direction for the azimuthal magnetic-field
component giving rise to the RM gradient. This azimuthal field
component, in turn, implies a certain direction for the dominant current
flowing in the jet --- either inward or outward. Thus, the evidence
we find for a preference for CCW transverse RM gradients on scales
greater than about 20~parsec implies a
preference for the dominant currents in the large-scale jets of AGN to
flow outwards from the AGN centers. In fact, this is predicted by
the Cosmic Battery mechanism (Eqs.~(3) and~(5) above), since the regions of
the transverse Faraday-rotation gradients in most of the
sources considered in Sections~3.1--3.3 lie beyond the expected
Alfv\'en points for the inner region corresponding to the magnetized jet
outflows.
The preponderance of CCW transverse RM gradients and the presence
of helical magnetic fields on large scales can be explained
physically in the framework of the CB outlined in Section~1 above: the
return helical field dominates the total observed Faraday rotation
on relatively large angular scales on the sky, where the detected
radio emission extends to fairly large distances from the jet
axis. It is much less likely for such observations to be dominated
by Faraday rotation due to the inner helical field whose field
lines are clinging close to the jet axis. In contrast, as was shown by
the analysis of CCKG, the inner helical field is more likely
to dominate the RM measurements on parsec scales.
In the above picture, one would still not expect to
find only CW transverse RM gradients on parsec scales, i.e., inside the
Alfv\'en point, and only CCW transverse RM gradients on
kiloparsec scales, beyond the Alfv\'en point, since
various physical and observational factors perturbing the jet and
its magnetic field are bound to play a role (e.g., Broderick \&
McKinney~2010). This may explain why a minority of the sources studied
on parsec scales have displayed CCW transverse RM gradients. The exact
distance from the jet base where the
transition from CW to CCW transverse RM gradients occurs may also
vary from source to source, so that a clear-cut demarkation line may
not exist.
Finally, CCW RM gradients are also expected when they are observed
right on top of termination shocks, provided that the shock fronts
have not been bent too much away from orthogonality to the
direction of jet propagation. Such localized CCW RM gradients may
be present in the hot spots of the FR II sources 5C4.74 and
5C4.152 (Bonafede et al. 2010), although the statistical significance
of these gradients remains to be tested.
\subsection{A theoretical alternative}
K\"onigl (2010) has argued that a predominance of CW RM gradients
on parsec scales could be the result of ordered, large-scale,
magnetic fields that were produced in the outer weakly ionized
accretion disks by Hall currents, and that were launched from the
disks in centrifugally driven wind outflows. This model may be
able to give rise to a predominance of CCW RM gradients on larger
scales, but only if the overall observed Faraday rotation is
dominated by the contribution of a ``return field'' whose origin
is not clear. Furthermore, Hall currents are usually believed to be
unimportant in AGN physics, because, unlike protostellar disks
(Krasnopolsky et al. 2011), the accretion disks of AGN (especially
those associated with the VLBA sources) are thought to be highly
ionized, in particular near the compact object where all the
prominent magnetohydrodynamical phenomena associated with the
parsec-scale emission are thought to originate (e.g., Gaskell 2009,
2010). This view is supported by the highly ionized oxygen ions observed
in certain broad-line radio-galaxies, i.e., radio loud AGN viewed at
fairly small angles to the jet direction (e.g., K\"onigl et al.
1995). On the other hand, if the magnetic field is brought in to
the outer, cold, weakly ionized part of the disk from farther out
(K\"onigl 2010), then different polarities are likely to undergo
fast reconnection long before a strong toroidal field component
can develop because of the much slower differential rotation of the
outer disk. Therefore, we consider this model physically less
plausible than the Cosmic Battery for AGNs.
\section{Summary and concluding remarks}
We have considered nine firm (having significances of at least $3\sigma$)
and seven tentative (visible in the RM maps but whose significances are
either below $3\sigma$ or unknown)
cases of monotonic transverse RM gradients detected across the jet
outflows of AGNs and radio galaxies on scales exceeding about 20~pc,
listed in Table~1. These observations provide
direct evidence that these jets carry helical magnetic fields, whose
toroidal component sometimes survives to distances of hundreds or even
thousands of parsec from the central AGN.
On both the relatively large scales considered here and on parsec
scales, transverse RM gradients have been reported for only some
fraction of the observed sources.
The simple reason why transverse RM
gradients should in fact not always be observed, even if all
AGN jets carry helical fields, is that the systematic pattern in the
RM distribution due to the helical magnetic field can be disrupted
by interactions with and entrainment of the ambient medium
through which the jet is propagating, patchiness of this surrounding
medium, and turbulence in the outer layers of the jet. The higher
detection rate of transverse RM gradient on parsec scales may then
indicate that these effects are less important and less disruptive
on smaller scales.
The data presented in Table~1 show a preponderance of large-scale transverse
RM gradients in which the RMs increase CCW on the sky relative to the jet bases,
corresponding to a dominant outward current along the associated jet structures.
The significance of this result is currently just above $3\sigma$, and
this tendency is clear enough to warrant further study, both observational and
theoretical.
Together with the results for smaller (parsec, VLBA) scales cited in
Section~1.4, the collected results considered in this paper are consistent with
the prediction of the CB model that CCW transverse RM gradients (corresponding
to outward currents) should be visible across AGN jets on relatively large
scales, while CW transverse RM gradients (corresponding to inward currents) should
be found closer to the centers of activity. The analysis of CCKG, whose validity
has now largely been confirmed by subsequent studies (e.g., Gabuzda 2014b,
2015a), suggested that the division between these regimes occurred at projected
distances of about 20~pc from the central AGN, and the results of our analysis
here bear this out.
This prediction will be tested more extensively by
future higher-resolution, long-wavelength radio observations that
will be made possible by the latest advances in radio telescopes
(ATA--256, EVLA, LOFAR, ASKAP, MeerKAT, SKA, VSOP--2; Gaensler et
al. 2004; Gaensler 2009; Perley et al. 2009; Hagiwara et al. 2009;
Law et al. 2011b). Further results from multi-wavelength
VLBA polarization observations of AGNs probing scales of
tens to hundreds of parsec may also shed more light on this question (e.g.,
Coughlan et al. 2010, Gabuzda et al. 2014a).
Finally, in a striking development, this picture has found
additional support from recent radio observations of the Galactic
center: Law, Brentjens, \& Novak (2011a) have reported a
powerful transverse CCW RM gradient ($\Delta$(RM)$\simeq
1100$~rad~m$^{-2}$ extending over 150~pc). Combining previous
observations of the area with their own observations, Law et al.
(2011a) find a "return" poloidal magnetic field in the central
$2^{\circ}$ of the Galactic center, with a dipolar configuration above and
below the Galactic plane. The measured RM values then flatten and
become positive in the central 30~pc, where the inner magnetic
field is expected to be superposed along the line of sight.
Additional observations by Pshirkov et al. (2011) have shown that
this CCW RM gradient extends to all Galactic latitudes above
and below the Galactic plane. This extended CCW RM gradient, the
flattening of the RM values towards the Galactic center, and the mapped
dipolar magnetic field are all consistent with the expectations of the CB
mechanism. This supposes that the Galaxy had jets in the past,
which are now not detectable, but they have left a relic magnetic
field. In this case, the results of Pshirkov et al. (2011) amount to
two more independent detections of CCW transverse RM gradients on large
scales (above and below the Galactic plane).
It is also interesting to note that this magnetic field structure is coincident
with the so called ``Fermi Bubbles" (Su \& Finkbeiner 2012; Ackermann et al. 2014),
large-scale ($\sim$50$^o$) microwave and $\gamma$-ray structures above and below
the Galactic plane centered at the Galactic center.
In closing, it is clear that the data analyzed thus far show
intriguing trends that could be of cardinal importance to our
understanding of electromagnetic processes occurring in the jet--disk
systems of AGN. In particular, they support the global magnetic field
topology predicted by the Cosmic Battery model (Eqs.~(1)--(5)). If this
mechanism is indeed the dominant source of the initial magnetic fields in
AGN jets, this would indicate that AGNs could be an important source
of magnetic flux in the Universe. The results
that we have presented here are undeniably based on small-number
statistics and require further confirmation from FR measurements
of additional radio sources. Our aim in reporting these results is to
bring these trends to the attention of the AGN community, in order
to spur further studies in this area, both observational and
theoretical.
\begin{acknowledgements}
We acknowledge insights and assistance with data provided by Drs.
Philip Best, Annalisa Bonafede, George Contopoulos, Federica Govoni,
Christian Kaiser, and Preeti Kharb. We especially thank Annalisa Bonafede
and Federica Govoni for presenting us with the calibrated data for
A2142A and 5C4.152. We also acknowledge the assistance of
Antonios Nathanail in the preparation of Fig.~2. This work was
supported by the General Secretariat for Research and Technology
of Greece, the Irish Research Council (IRC) and the European Social
Fund in the framework of Action
Excellence.
We thank the referee whose comments have led to an expansion of the
paper that helped improve the clarity and significance of our results.
\end{acknowledgements}
|
\section{Introduction}\label{sec:intro}
We consider the Hermitian-definite generalized eigenvalue problem
\begin{equation}\label{eq:ghep}
Hu=\lambda Su,
\end{equation}
where $H$ and $S$ are $n$-by-$n$ Hermitian matrices and $S$ is
positive-definite. The scalar $\lambda$ and nonzero vector $u$
satisfying \eqref{eq:ghep} are called
{\em eigenvalue} and {\em eigenvector}, respectively.
The pair $(\lambda,u)$ is called an eigenpair.
All eigenvalues of \eqref{eq:ghep} are known to be real.
Our task is to compute few smallest eigenvalues and the
corresponding eigenvectors. We are particularly interested in solving
the eigenvalue problem~\eqref{eq:ghep},
where the matrices $H$ and $S$ are large and sparse, and
there is no obvious gap between the eigenvalues of interest
and the rest. Furthermore, $S$ is nearly singular
and $H$ and $S$ share a near-nullspace.
It is called an ill-conditioned generalized eigenvalue problem
in \cite{fix1972algorithm}, a term we will adopt in this paper.
The ill-conditioned generalized eigenvalue problem is considered to
be an extremely challenging problem.\footnote{W. Kahan,
Refining the general symmetric definite eigenproblem, poster presentation at
Householder Symposium XVIII 2011,
available http://www.cs.berkeley.edu/$\sim$wkahan/HHXVIII.pdf}
\medskip
Beside examples such as those cited in \cite{fix1972algorithm},
the ill-conditioned eigenvalue problem~\eqref{eq:ghep} arises from the
discretization of enriched Galerkin methods. The partition-of-unity
finite element (PUFE) method~\cite{melenk1996partition},
which falls within the class of enriched
Galerkin methods, is a promising approach
in quantum-mechanical materials calculations,
see \cite{cai2013hybrid} and references therein.
In the PUFE method, physics-based basis functions are added to the classical
finite element (polynomial basis) approximation, which affords the
method improved accuracy at reduced costs versus
existing techniques. However, due to near linear-dependence
between the polynomial and enriched basis functions,
the system matrices that stem from such methods are ill-conditioned, and
share a large common near-nullspace. Furthermore, there is in
general no clear gap between the eigenvalues that are sought and the rest.
Another example of the ill-conditioned eigenvalue problem~\eqref{eq:ghep} arises
from modeling protein dynamics using
normal-mode analysis \cite{levitt1985protein,nishikawa1987normal,brooks1995harmonic,levi:15}.
\medskip
In this paper, we focus on a preconditioned
steepest descent with implicit deflation method, {PSD-{id}} method in short,
to solve the eigenvalue problems~\eqref{eq:ghep}.
The basic idea of the {PSD-{id}} method is simple.
Denote all the eigenpairs of \eqref{eq:ghep} by $(\lambda_1, u_1)$,
$(\lambda_2,u_2)$, \dots, $(\lambda_n,u_n)$, and the eigenvalue and
eigenvector matrices by
$\Lambda=\diag(\lambda_1,\lambda_2,\dots,\lambda_n)$
and $U=[u_1 \; u_2\; \cdots\; u_n]$, respectively.
Assume that the eigenvalues $\{\lambda_i\}$ are in an ascending order
$\lambda_1 \le\lambda_2\le\cdots\le \lambda_n$.
The following variational principles are well-known,
see \cite[p.99]{wilkinsonalgebraic} for example:
\begin{align}\label{eq:evew-i}
\lambda_i=\min_{U_{i-1}^HS z=0}\rho(z)
\quad \mbox{and} \quad u_i=\argmin_{U_{i-1}^HS z=0}\rho(z),
\end{align}
where $U_{i-1}=[u_1\; u_2\; \cdots \; u_{i-1}]$ and
$\rho(z)$ is the Rayleigh quotient
\begin{equation}\label{eq:rq}
\rho(z)=\frac{z^HHz}{z^H Sz}.
\end{equation}
On assuming that $U_{i-1}$ is known,
one can find the $i$th eigenpair by minimizing the
the Rayleigh quotient $\rho(z)$
with $z$ being $S$-orthogonal against $U_{i-1}$ under the algorithmic framework
of the preconditioned steepest descent minimization.
\medskip
The idea of computing the algebraically largest eigenvalue and its corresponding
eigenvector of \eqref{eq:ghep} (with $B = I$) using the steepest descent (SD)
method dates back to early 1950s \cite{hestenes1951method} and
\cite[Chap.7]{fadeev1963computational}.
In \cite{longsine1980simultaneous}, block steepest descent (BSD)
methods are proposed to compute several eigenpairs simultaneously.
The preconditioned steepest descent (PSD) method was
introduced around late 1950s~\cite{samokish1958steepest,petryshyn1968eigenvalue}.
The block PSD (BPSD) methods have appeared in
the literature, see \cite{bramble1996subspace,neymeyr2014iterative}
and references therein.
Like the PSD method, the {PSD-{id}} method studied in this paper computes
one eigenpair at a time. To compute the $i$th eigenpair, the search
subspace of {PSD-{id}}
is {\em implicitly} orthogonalized against the previously computed
$i-1$ eigenvectors. Furthermore, the preconditioner at each iteration is
flexible (i.e, could change at every iteration) and can be indefinite, instead
of being fixed and positive definite as
in~\cite{samokish1958steepest,bramble1996subspace,neymeyr2014iterative}.
\medskip
Over the past six decades, there has been significant work on
the convergence analysis of the SD, PSD and BPSD methods.
The convergence of the SD method to compute a single eigenpair
is presented in \cite[Chap.7]{fadeev1963computational}.
For the BSD method, the convergence of the first eigenpair
is presented in \cite{longsine1980simultaneous} and
``ordered convergence'' for multiple eigenpairs is declared.
The (nonasymptotic) rate of convergence of the PSD method is first studied
in \cite{samokish1958steepest}, which later is proven to
be sharp \cite{ovtchinnikov2006sharp}.
A comprehensive review of the convergence estimates of the PSD method,
is presented in \cite{bramble1996subspace}.
The theoretical proofs of the convergence of the BPSD method
have still {largely eluded us}, we refer the readers
to \cite{bramble1996subspace,ovtchinnikov2006cluster}
and two recent papers \cite{neymeyr2014block,neymeyr2014iterative}.
In this paper, we present two main results
(Theorems~\ref{thm:psd2} and~\ref{thm:Samokish2})
on the convergence and nonasymptotic rate of convergence
of the {PSD-{id}} method. These results extend the classical
ones due to Faddeev and Faddeeva \cite[sec.74]{fadeev1963computational}
and Samokish \cite{samokish1958steepest} for the SD and PSD methods.
We show that with the proper choice of the shift,
the well-known {indefinite} shift-and-invert preconditioner
is a flexible and locally accelerated preconditioner,
and is asymptotically optimal that
leads to superlinearly converge of the {PSD-{id}} method.
Numerical examples shows the superlinear convergence of the
{PSD-{id}} method with locally accelerated preconditioners for solving
ill-conditioned generalized eigenvalue problems~\eqref{eq:ghep} arising from
full self-consistent electronic structure calculations.
\medskip
We would like to note that the main objective of this paper is to provide
a rigorous convergence analysis of the {PSD-{id}} method with flexible and
locally accelerated preconditioners than to advocate the usage of the
{PSD-{id}} method in practice.
The BPSD methods~\cite{bramble1996subspace, neymeyr2014iterative} and
a recent proposed locally accelerated BPSD (LABPSD) presented in our
previous work~\cite{cai2013hybrid} have demonstrated their efficiency
for finding several eigenpairs simultaneously.
While a rigorous and full-scale convergence proof of the
the BPSD methods still largely eludes us,
we believe the analysis of the {PSD-{id}} method presented in this paper
can shed light on an improved understanding of the
convergence behavior of the BPSD methods
such as the LAPBSD method~\cite{cai2013hybrid} for
solving the ill-conditioned generalized eigenvalue problem~\eqref{eq:ghep}
arising from the PUFE simulation of electronic
structure calculations.
\medskip
The rest of this paper is organized as follows.
In section~\ref{sec:dpsd}, we present the {PSD-{id}} method
and discuss its basic properties.
In section~\ref{sec:analysis}, we provide a
convergence proof and a nonasymptotic estimate of
the convergence rate of the {PSD-{id}} method.
An asymptotically optimal preconditioner is discussed
in section~\ref{sec:opt}. Numerical examples
to illustrate the theoretical results are presented in section~\ref{sec:numer}.
We close with some final remarks in section~\ref{sec:conclusion}.
\medskip
In the spirit of reproducible research,
Matlab scripts of an implementation of the {PSD-{id}} method, and
the data that used to generate numerical results
presented in this paper can be obtained from the URL
http://dsec.pku.edu.cn/$\sim$yfcai/psdid.html.
\section{Algorithm} \label{sec:dpsd}
Assuming that $U_{i-1}$ is already known, by \eqref{eq:evew-i},
one can find the $i$th eigenpair by minimizing the Rayleigh quotient $\rho(z)$
with $z$ being $S$-orthogonal against $U_{i-1}$.
Specifically, let us denote by $(\lambda_{i;j},u_{i;j})$
the $j$th approximation of $(\lambda_i,u_i)$
and assume that
\begin{equation} \label{eq:jassume}
U^H_{i-1}S u_{i;j} = 0, \quad
\|u_{i;j}\|_S=1 \quad \mbox{and} \quad
\lambda_{i;j}=\rho(u_{i;j}).
\end{equation}
To compute the $(j+1)$st approximate eigenpair
$(\lambda_{i;j+1}, u_{i;j+1})$, by the steepest descent approach,
the steepest decreasing direction
of $\rho(z)$ is opposite to the gradient of $\rho(z)$ at $z=u_{i;j}$:
\[
\nabla\rho(u_{i;j})= 2(H-\lambda_{i;j} S)u_{i;j}=2r_{i;j}.
\]
Furthermore, to accelerate the convergence, we use the following
preconditioned search vector
\begin{equation}\label{eq:pij}
p_{i;j}=-K_{i;j}r_{i;j},
\end{equation}
where $K_{i;j}$ is a preconditioner.
By a Rayleigh-Ritz projection based implementation,
the $(j+1)$st approximate eigenpair
$(\lambda_{i;j+1}, u_{i;j+1})$
computed by the preconditioned steepest descent method
is given by
\begin{align}\label{eq:liju}
(\lambda_{i;j+1}, u_{i;j+1})
= (\gamma_i, Z_j w_i),
\end{align}
where $(\gamma_i,w_i)$ is the $i$th eigenpair of
the projected matrix pair $(H_R, S_R) = (Z^H_j H Z_j, Z^H_jSZ_j)$,
$\|w_i\|_{S_R}=1$, and $Z_j=[U_{i-1}\; u_{i;j} \; p_{i;j}]$ is the basis
matrix of the projection subspace.
Here we assume that $Z_j$ is of full column rank.
\bigskip
Algorithm~\ref{alg:psdid} is a summary of the aforementioned procedure.
Since the first eigenvectors $U_{i-1}$ are implicitly deflated in the
Rayleigh-Ritz procedure,
we call Algorithm~\ref{alg:psdid} a preconditioned steepest descent
with implicit deflation, {{PSD-{id}} } in short.
We note that the preconditioner $K_{i;j}$ is flexible.
It can be changed at each iteration.
If the preconditioner is fixed as a uniform positive definite matrix,
i.e., $K_{i;j} = K > 0$, then Algorithm~\ref{alg:psdid}
is the SIRQIT-G2 algorithm in \cite{longsine1980simultaneous} with $K=I$
and initial vectors $X^{(0)}=[U_{i-1}\, u_{i;0}]$, and is the BPSD
method \cite{knyazev2003efficient} with initial vectors $[U_{i-1}\, u_{i;0}]$.
\begin{algorithm}
\caption{{PSD-{id}} } \label{alg:psdid}
\begin{algorithmic}[1]
\REQUIRE $U_{i-1}$
and initial vector $u_{i;0}$
\ENSURE Approximate eigenpair $(\lambda_i, u_i)$ of $(\lambda_i, u_i)$
\STATE $\lambda_{i;0}= \rho(u_{i;0})$
\FOR {$j=0,1,\ldots$, until convergence}
\STATE compute $r_{i;j}=Hu_{i;j} - \lambda_{i;j}Su_{i;j}$
\STATE precondition $p_{i;j} = -K_{i;j}r_{i;j}$
\STATE compute the $i$th eigenpair $(\gamma_i,w_i)$ of
$(H_R,S_R)=(Z^H_j H Z_j, Z^H_jS Z_j)$, $Z_j=[U_{i-1}\; u_{i;j} \; p_{i;j}]$
\STATE update $\lambda_{i;j+1} = \gamma_i$ and $u_{i; j+1} = Z_j w_i$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\medskip
If Algorithm~\ref{alg:psdid} does not breakdown,
i.e., the matrices $Z_j$ on line 5 are full column rank for all $j$,
then a sequence of approximate eigenpairs
$\{(\lambda_{i;j},u_{i;j})\}_j$ are produced.
The following proposition gives basic properties of the sequence.
In particular, if the initial vector $u_{i;0}$ does not satisfy
the assumption~\eqref{eq:jassume}, the first approximate vector
$u_{i;1}$ computed by Algorithm~\ref{alg:psdid} will suffice.
\begin{proposition} \label{prop:basicproperties}
If $Z_j$ is of full column rank, then
\begin{itemize}
\item[(a)] $U^H_{i-1}S u_{i;j+1} = 0$.
\item[(b)] $\|u_{i;j+1}\|_S=1$.
\item[(c)] $\lambda_{i;j+1} \geq \lambda_i$.
\item[(d)] $\lambda_{i;j+1} \le \lambda_{i;j}$.
\end{itemize}
\end{proposition}
\begin{proof}
Results (a) and (b) are verified by straightforward calculation.
The result (c) follows from the inequality
\begin{equation*
\lambda_{i;j+1}=\rho({u}_{i;j+1})\ge \min_{U_{i-1}^HSz=0}\rho(z)=\lambda_i.
\end{equation*}
Finally, the result (d) follows from the facts that
\[
\lambda_{i;j+1}=\lambda_i(H_R, S_R)
=\min_{U_{i-1}^H S Z_j w=0 }\rho(Z_j w)
\le \rho(Z_j w)|_{w=e_i}=\rho(u_{i;j})=\lambda_{i;j},
\]
where $e_i$ is the $i$th column vector of identity matrix of order $i+1$.
\end{proof}
The following proposition shows that with the proper choice of
the preconditioner $K_{i;j}$, the basis matrix
$Z_j=[U_{i-1}\; u_{i;j} \; p_{i;j}]$
is of full column rank,
which implies that Algorithm~\ref{alg:psdid} does not breakdown.
\begin{proposition}\label{lem:fr}
If $r_{i;j}\ne 0$ and $K_{i;j}$ is chosen such that
\begin{equation}\label{k-con}
K^{c}_{i;j} :=(U_{i-1}^{c})^H S K_{i;j}SU_{i-1}^c>0,\end{equation}
then the basis matrix $Z_j$ is of full column rank. Here
$U_{i-1}^c$ is complementary eigenvector matrix of
$U_{i-1}$, i.e., $U_{i-1}^c= [u_{i}\; \cdots \; u_n]$.
\end{proposition}
\begin{proof}
We prove that $Z_j$ is of full column rank by showing that
\[
\det(H_{R}-\lambda_{i;j}S_{R}) =
\det( Z^H_j( H -\lambda_{i;j}S)Z_j ) \neq 0.
\]
First, it can be verified that the projected matrix pair
$(H_R, S_R)$ can be factorized as follows:
\begin{equation} \label{eq:hr}
(H_R, S_R) = L^{-1}\left( \begin{bmatrix}
\Lambda_{i-1} & 0\\
0& H_{\bot}\end{bmatrix},
\begin{bmatrix}
I_{i-1} & 0\\ 0 & S_{\bot}\end{bmatrix}
\right) L^{-H},
\end{equation}
where
\[
L=\begin{bmatrix}
I_{i-1} & 0 & 0\\
0 & 1 & 0\\
-p_{i;j}^H SU_{i-1} & 0 & 1
\end{bmatrix}, \quad
H_{\bot} = Z_{\bot}^H HZ_{\bot}, \quad
S_{\bot} = Z_{\bot}^H SZ_{\bot},
\]
and
$Z_{\bot}=[u_{i;j} \; p_{\bot}]$ and
$p_{\bot}=U_{i-1}^c(U_{i-1}^c)^HSp_{i;j}$.
Consequently, we have
\begin{align}\label{eq:dethrsr}
\det(H_R-\lambda_{i;j} S_R)=\det(\Lambda_{i-1}-\lambda_{i;j} I)
\det(H_{\bot}-\lambda_{i;j}S_{\bot}).
\end{align}
By Proposition~\ref{prop:basicproperties}(c),
we have $\lambda_{i;j}\geq \lambda_i$.
Since $r_{i;j}\ne 0$, $\lambda_{i;j}>\lambda_i$.
Hence, we conclude that
\begin{equation} \label{eq:det1}
\det(\Lambda_{i-1}-\lambda_{i;j} I)\ne 0.
\end{equation}
Next we show that $\det(H_{\bot}-\lambda_{i;j}S_{\bot}) \neq 0$.
We first note that since $U_{i-1}^HSu_{i;j}=0$,
there exists a vector $a$ such that $u_{i;j}=U_{i-1}^c a$.
Then it follows that
\begin{align}\label{eq:rij}
r_{i;j
=(H-\lambda_{i;j}S)U_{i-1}^ca
=SU_{i-1}^c(\Lambda_{i-1}^c-\lambda_{i;j}I)a
=SU_{i-1}^c(U_{i-1}^c)^Hr_{i;j},
\end{align}
where $\Lambda_{i-1}^c = \diag(\lambda_{i},\ldots, \lambda_n)$.
Note that $(U_{i-1}^c)^Hr_{i;j}\ne 0$ since $r_{i;j}\ne 0$.
Furthermore, using \eqref{eq:rij} and \eqref{k-con}, we have
\begin{align}
\det(H_{\bot}-\lambda_{i;j}S_{\bot})
& =\det\begin{bmatrix} 0 & r_{i;j}^H p_{\bot}\\ p_{\bot}^H r_{i;j}& p_{\bot}^H (H-\lambda_{i;j}S) p_{\bot}\end{bmatrix}\notag\\
& = -|p_{\bot}^H r_{i;j}|^2\notag\\
& = -|r_{i;j}^H U_{i-1}^c (U_{i-1}^c)^H S K_{i;j} SU_{i-1}^c (U_{i-1}^c)^H r_{i;j}|^2\notag\\
& = -|r_{i;j}^H U_{i-1}^c {K}^c_{i;j} (U_{i-1}^c)^H r_{i;j}| < 0. \label{eq:det2}
\end{align}
By \eqref{eq:dethrsr}, \eqref{eq:det1} and \eqref{eq:det2},
we conclude that $H_{R}-\lambda_{i;j}S_{R}$ is nonsingular,
which implies that $Z_j$ is of full column rank.
\end{proof}
\begin{definition} \label{def:epdp}
A preconditioner $K_{i;j}$ satisfying the condition \eqref{k-con} is
called an {\rm effectively positive definite preconditioner}.
\end{definition}
We note that an effectively positive definite preconditioner $K_{i;j}$ with $i>1$
is not necessarily to be symmetric positive definite.
For example, for any $\lambda_1<\sigma <\lambda_i$ and
$\sigma$ is not an eigenvalue of $(H,S)$, $K_{i;j}=(H-\sigma S)^{-1}$ is
effectively positive definite, although $K_{i;j}$ is indefinite.
\medskip
If the preconditioner $K_{i;j}$ is chosen such that
the search vector $p_{i;j}=-K_{i;j}r_{i;j}$ satisfies
\begin{equation}\label{ideal-p}
U^HS({u}_{i;j}+p_{i;j})=\xi=(\xi_1,\xi_2,\dots,\xi_n)^H\quad
\mbox{with $\xi_i\ne 0$ and $\xi_j=0$ for $j>i$},
\end{equation}
then
\begin{align}
\lambda_{i;j+1}
& = \min_{U_{i-1}^H S Z_j w=0}\rho(Z_j w) \notag \\
& =\min_w \rho(U_{i-1}^c (U_{i-1}^c)^H SZ_jw)\notag \\
& =\min_v\rho(U_{i-1}^c (U_{i-1}^c)^H S[u_{i;j} \ p_{i;j}]v) \notag \\
& \le \rho(U_{i-1}^c (U_{i-1}^c)^H S[u_{i;j} \ p_{i;j}]v)|_{v=[1 \; 1]^\top} \notag \\
& =\rho(U_{i-1}^c (U_{i-1}^c)^H S(u_{i;j}+ p_{i;j}))\notag\\
& =\rho(U_{i-1}^c (U_{i-1}^c)^H SU\xi) \notag \\
& =\rho(\xi_i u_i ) =\lambda_i. \label{eq:lambdaless}
\end{align}
Therefore, combining the inequality \eqref{eq:lambdaless}
and Proposition~\ref{prop:basicproperties}(c),
we have $\lambda_{i;j+1}=\lambda_i$.
In this case, we refer to $p_{i;j}$ satisfying the equation \eqref{ideal-p}
as an {\em ideal search direction}.
The notion of an ideal search direction not only helps assessing the quality
of a preconditioned search direction, but also tells the desired property
for the solution of the preconditioning equation $p_{i;j}=-K_{i;j} r_{i;j}$.
\section{Convergence analysis}\label{sec:analysis}
In this section, we prove the convergence of the {PSD-{id}} method and
derive a nonasymptotic estimate of the convergence rate.
For brevity, we assume that for the desired
$i$th eigenvalue $\lambda_i$, it satisfies
$\lambda_{i-1} < \lambda_{i} < \lambda_{i+1}$.
Otherwise by replacing $\lambda_{i+1}$ by the smallest eigenvalue
of $(H,S)$ which is larger than $\lambda_i$, all results in this section
still hold, the proofs are similar.
\subsection{Convergence results}\label{sec:convergence}
Assume that the preconditioner $K_{i;j}$ is effectively positive definite,
then by Proposition~\ref{prop:basicproperties}(d) and \eqref{eq:det2},
we have
that $\lambda_{i;j+1}$ is strictly less than $\lambda_{i;j}$,
\begin{equation}\label{lamstrictless}
\lambda_{i;j+1}<\lambda_{i;j}.
\end{equation}
Furthermore, by Proposition~\ref{prop:basicproperties}(c)
and \eqref{lamstrictless}, the approximate eigenvalue sequence
$\{\lambda_{i;j}\}_j$ is a monotonically
decreasing and is bounded below by $\lambda_i$, i.e.,
\begin{equation} \label{eq:lambdadecrease}
\lambda_{i;0} > \lambda_{i;1} > \cdots > \lambda_{i;j}>\lambda_{i;j+1}
> \dots \ge \lambda_i.
\end{equation}
Therefore, the sequence $\{\lambda_{i;j}\}_j$ must converge.
Does it converge to the $i$th eigenvalue $\lambda_i$ of $(H,S)$?
How about the corresponding $\{u_{i;j}\}_j$?
We will answer these questions in this subsection.
First, we give the following lemma to quantify the
difference between two consecutive approximates
$\lambda_{i;j}$ and $\lambda_{i;j+1}$ of $\lambda_i$.
\begin{lemma}\label{thm:lam}
If $r_{i;j}\ne 0$ and the preconditioner $K_{i;j}$ is
effectively positive definite, then
\begin{equation}\label{eq:lamdif}
\lambda_{i;j}-\lambda_{i;j+1} \ge \sqrt{g^2+\phi^2} - g,
\end{equation}
where
$g=(\lambda_n-\lambda_i)/2$ and
$\phi ={\|r_{i;j}\|_{S^{-1}} }/{\kappa(K^c_{i;j})}$,
$\kappa(K^c_{i;j})$ is the condition number of ${K}^c_{i;j}$
defined in \eqref{k-con}.
\end{lemma}
\begin{proof}
Let $H_{\bot}$ and $S_{\bot}$ be the matrices defined in \eqref{eq:hr}.
By $S$-orthogonalizing $p_{i;j}$ against $[U_{i-1} \; u_{i;j}]$,
the resulting vector $\hat{p}=(I-u_{i;j}u_{i;j}^HS)P_{i-1}p_{i;j}$
must be nonzero since $Z_j=[U_{i-1} \; u_{i;j}\; p_{i;j}]$ is of full
column rank (Lemma~\ref{lem:fr}).
Therefore, it holds that $\hat{p}^HS\hat{p}> 0$.
By straightforward calculations, we have
\begin{align}
\det(H_{\bot}-\mu S_{\bot})
& = \det\left(
\begin{bmatrix}\lambda_{i;j}& u_{i;j}^HH\hat{p}\\ \hat{p}^HHu_{i;j} & \hat{p}^HH\hat{p}\end{bmatrix} -
\mu\begin{bmatrix} 1 & 0\\ 0 & \hat{p}^HS\hat{p}\end{bmatrix}\right) \notag \\
& = \hat{p}^HS\hat{p} (\lambda_{i;j}-\mu)
(\rho(\hat{p})-\mu)-|u_{i;j}^HH\hat{p}|^2 \notag \\
& = \hat{p}^HS\hat{p} \left[ (\lambda_{i;j}-\mu)^2 +
(\rho(\hat{p})-\lambda_{i;j})(\lambda_{i;j}-\mu)
-\frac{ |u_{i;j}^HH\hat{p}|^2}{\hat{p}^HS\hat{p}}\right]. \label{eq:detHS}
\end{align}
By the definition of $\lambda_{i;j+1}$ in~\eqref{eq:liju},
the identity~\eqref{eq:dethrsr},
we know that $\lambda_{i;j+1}$ is the smaller
root of the quadratic polynomial \eqref{eq:detHS} of $\mu$.
In addition, by \eqref{eq:lambdadecrease},
we know that $\lambda_{i;j}-\lambda_{i;j+1}$ is positive.
Therefore $\lambda_{i;j}-\lambda_{i;j+1}$ is
the positive root of the following quadratic equation in $t$:
\begin{equation*}
t^2 + (\rho(\hat{p})-\lambda_{i;j}) t
- \frac{|u_{i;j}^HH\hat{p}|^2}{ \hat{p}^HS\hat{p}}=0.
\end{equation*}
Then it follows that
\begin{align}\label{eq:tplus}
\lambda_{i;j}-\lambda_{i;j+1}
=-\frac{\rho(\hat{p})-\lambda_{i;j}}{2}
+\sqrt{\left(\frac{\rho(\hat{p})-\lambda_{i;j}}{2}\right)^2
+\frac{|u_{i;j}^HH\hat{p}|^2}{ \hat{p}^HS\hat{p}}}.
\end{align}
In what follows, we give the estimates of
the quantities $|\rho(\hat{p})-\lambda_{i;j}|$,
$|u_{i;j}^HH\hat{p}|^2$ and $\hat{p}^HS\hat{p}$, respectively.
For the quantity $|\rho(\hat{p})-\lambda_{i;j}|$,
using the fact that for any nonzero $z$ satisfying $U_{i-1}^HSz=0$,
it holds $\lambda_i\le\rho(z)\le\lambda_n$, then
using $U_{i-1}^HS\hat{p}=0$ and $U_{i-1}^H S u_{i;j}=0$, we have
\begin{equation} \label{eq:rhoest}
0\le |\rho(\hat{p})-\lambda_{i;j}|\le \lambda_n-\lambda_i= 2g.
\end{equation}
For the quantity $|u_{i;j}^HH\hat{p}|^2$, we have
\begin{subequations}
\begin{align}
|u_{i;j}^HH\hat{p}|
& =|u_{i;j}^HH(I-u_{i;j}u_{i;j}^HS)U_{i-1}^c(U_{i-1}^c)^HSK_{i;j}r_{i;j}|
\label{uhp-1}\\
&=|\big[u_{i;j}^HH-\lambda_{i;j}u_{i;j}^HS\big] U_{i-1}^c \big[(U_{i-1}^c)^HSK_{i;j} SU_{i-1}^c \big] (U_{i-1}^c)^H r_{i;j}|\label{uhp-2}
\\
&=|r_{i;j}^HU_{i-1}^c K^c_{i;j} (U_{i-1}^c)^Hr_{i;j}|\label{uhp-3}\\
&\ge \lambda_{\min}(K^c_{i;j})\|(U_{i-1}^c)^Hr_{i;j}\|^2\label{uhp-4}\\
& = \lambda_{\min}(K^c_{i;j})\|r_{i;j}\|_{S^{-1}}^2\label{uhp-5},
\end{align}
\end{subequations}
where
\eqref{uhp-1} uses the definition of $\hat{p}$ and \eqref{eq:pij},
\eqref{uhp-2} uses the fact that
$r_{i;j}=SU_{i-1}^c (U_{i-1}^c)^H r_{i;j}$,
\eqref{uhp-3} and \eqref{uhp-4} use
the definition of $K^c_{i;j}$ in \eqref{k-con} and the assumption
that $K^c_{i;j}$ is symmetric positive definite, and
\eqref{uhp-5} is based on the following calculations:
\begin{align*}
\|(U_{i-1}^c)^Hr_{i;j}\|^2
&=r_{i;j}^H U_{i-1}^c(U_{i-1}^c)^Hr_{i;j} &\\
&=r_{i;j}^H U_{i-1}^c(U_{i-1}^c)^Hr_{i;j}+r_{i;j}^H U_{i-1}U_{i-1}^Hr_{i;j} & (U_{i-1}^Hr_{i;j}=0)\\
&=r_{i;j}^H U U^H r_{i;j}&\\
&=r_{i;j}^H S^{-1}r_{i;j} & (UU^H=S^{-1})\\
&= \|r_{i;j}\|_{S^{-1}}^2. &
\end{align*}
For the quantity $\hat{p}^HS\hat{p}$, we have
\begin{align}
\hat{p}^HS\hat{p}
& = (r_{i;j})^HK_{i;j} SU_{i-1}^c \big[ (U_{i-1}^c)^H(I-Su_{i;j}u_{i;j}^H)
S (I-u_{i;j}u_{i;j}^HS) U_{i-1}^c \big] (U_{i-1}^c)^H S K_{i;j} r_{i;j}\notag\\
& \le\| (U_{i-1}^c)^H(I-Su_{i;j}u_{i;j}^H)
S (I-u_{i;j}u_{i;j}^HS) U_{i-1}^c \| \|(U_{i-1}^c)^H S K_{i;j} r_{i;j} \|^2\notag\\
& \le \|(U_{i-1}^c)^H S K_{i;j} r_{i;j} \|^2 \notag \\
& = \|K^c_{i;j} (U_i^{c})^Hr_{i;j}\|^2 \notag\\
& \le \lambda_{\max}(K^c_{i;j})^2\|(U_{i-1}^c)^Hr_{i;j}\|^2\notag\\
& = \lambda_{\max}(K^c_{i;j})^2\|r_{i;j}\|_{S^{-1}}^2, \label{php-7}
\end{align}
where
the second inequality use the fact that
\begin{align*}
\| &(U_{i-1}^c)^H(I-Su_{i;j}u_{i;j}^H) S (I-u_{i;j}u_{i;j}^HS) U_{i-1}^c \|\\
= \| &(U_{i-1}^c)^HS^{{\frac{1}{2}}}(I-S^{\frac{1}{2}} u_{i;j}u_{i;j}^H S^{\frac{1}{2}})
(I-S^{\frac{1}{2}} u_{i;j}u_{i;j}^HS^{\frac{1}{2}}) S^{\frac{1}{2}} U_{i-1}^c \| \\
\le \| &(I-S^{\frac{1}{2}} u_{i;j}u_{i;j}^H S^{\frac{1}{2}})\|^2 \|S^{\frac{1}{2}} U_{i-1}^c \|^2 \le 1.
\end{align*}
Finally, by \eqref{eq:rhoest}, \eqref{uhp-5} and
\eqref{php-7}, it follows from \eqref{eq:tplus} that
\begin{align*}
\lambda_{i;j}-\lambda_{i;j+1}
\ge& -\frac{|\rho(\hat{p})-\lambda_{i;j}|}{2}+\sqrt{\bigg(\frac{\rho(\hat{p})-\lambda_{i;j}}{2}\bigg)^2 + \frac{|u_{i;j}^HH\hat{p}|^2}{ \hat{p}^HS\hat{p}}}\\
\ge & -\frac{\lambda_n-\lambda_i}{2} + \sqrt{\bigg(\frac{\lambda_n-\lambda_i}{2}\bigg)^2+\frac{\|r_{i;j}\|^2_{S^{-1}}}{\kappa^2(K^c_{i;j})}}\\
= &-g+\sqrt{g^2+\phi^2}.
\end{align*}
This completes the proof.
\end{proof}
We note that in \cite[Chap.7]{fadeev1963computational}, for the
steepest descent method to compute the largest eigenvalue
$\lambda_n$ of a Hermitian matrix, it shows that
\begin{equation*
\lambda_{n;j+1}-\lambda_{n;j}\ge
\frac{\|r_{n;j}\|^2}{\lambda_n-\lambda_1}.
\end{equation*}
Then it is established that $\lambda_{n;j}$ converges to $\lambda_n$, and
$u_{n;j}$ converges to $u_n$ directionally.
Lemma~\ref{thm:lam} and the following theorem
are generalizations that are not limited to the largest eigenpair,
and include the usage of flexible preconditioners.
\begin{theorem}\label{thm:psd2}
If the initial estimate eigenvalue
$\lambda_{i;0}$ satisfying $\lambda_i < \lambda_{i;0}<\lambda_{i+1}$, and
the flexible preconditioners $K_{i;j}$ are an effectively positive definite
for all $j$ and $\mbox{sup}_j \kappa(K^c_{i;j})=q<\infty$,
then the sequence $\{(\lambda_{i;j}, u_{i;j})\}_j$ generated by the
{PSD-{id}} method converges to the desired pair $(\lambda_{i}, u_{i})$, i.e.,
\begin{itemize}
\item[(a)] $\lim_{j\rightarrow \infty}\lambda_{i;j}=\lambda_i$.
\item[(b)] $\lim_{j\rightarrow \infty}\|r_{i;j}\|_{S^{-1}}=0$, namely $u_{i;j}$
converges to $u_i$ directionally.
\end{itemize}
\end{theorem}
\begin{proof} To prove (a), we first notice
that $\{\lambda_{i;j}\}_j$ is a monotonic decreasing sequence,
and is bounded by $\lambda_i$ from below.
So there exists a real number $\tilde{\lambda}_i$ such that
$\lambda_{i;j} \rightarrow \tilde{\lambda}_i$ as
$j\rightarrow \infty$. Now we show by contradiction
that $\tilde{\lambda}_i=\lambda_i$.
For any $u_{i;j}$ ($\|u_{i;j}\|_S=1$), we have
\begin{align*}
\|r_{i;j}\|_{S^{-1}}
& =\|(H-\lambda_{i;j}S)u_{i;j}\|_{S^{-1}} \\
& \ge \|(H-\tilde{\lambda}_iS)u_{i;j}\|_{S^{-1}}-(\lambda_{i;j}-\tilde{\lambda}_i)\|Su_{i;j}\|_{S^{-1}}\\
& \ge \min_k{|\lambda_k-\tilde{\lambda}_i|}-(\lambda_{i;j}-\tilde{\lambda}_i).
\end{align*}
As $\lim_{j\rightarrow\infty}\lambda_{i;j}=\tilde{\lambda}_i$,
there exists a $j_0$ such that for any $j\ge j_0$,
\[
\|r_{i;j}\|_{S^{-1}}> {\frac{1}{2}}\min_k|\lambda_k-\tilde{\lambda}_i|.
\]
By defining $d(r,\kappa):=-g+\sqrt{g^2+(r/\kappa)^2}$,
it follows from Lemma~\ref{thm:lam} that
for any $j\ge j_0$, it holds that
\begin{align*}
\lambda_{i;j}-\lambda_{i;j+1}
\ge d(\|r_{i;j}\|_{S^{-1}},\kappa_{i;j})
> d(\min_k|\lambda_k-\tilde{\lambda}_i|/2, \kappa_{i;j})
\ge d(\min_k|\lambda_k-\tilde{\lambda}_i|/2, q),
\end{align*}
which in the limit becomes
\[
0> d(\min_k|\lambda_k-\tilde{\lambda}_i|/2, q).
\]
This is a contradiction to the fact that
$d(\min_k|\lambda_k-\tilde{\lambda}_i|/2, q)$ is a positive constant.
To prove (b), by using $\lim_{j\rightarrow \infty}\lambda_{i;j}
=\lambda_i$ and Lemma~\ref{thm:lam}, we have
\begin{align*}
\lim_{j\rightarrow\infty} d( \|r_{i;j}\|_{S^{-1}},q)
\le \lim_{j\rightarrow\infty} d( \|r_{i;j}\|_{S^{-1}},\kappa_{i;j})
\le \lim_{j\rightarrow\infty}(\lambda_{i;j}-\lambda_{i;j+1})=0.
\end{align*}
Consequently, $\lim_{j\rightarrow\infty}d(\|r_{i;j}\|_{S^{-1}},q)=0$,
which leads to $\lim_{j\rightarrow\infty}\|r_{i;j}\|_{S^{-1}}=0$
since $1\le q<\infty$.
\end{proof}
We note that
in Theorem~\ref{thm:psd2}, without assuming $\lambda_{i;0}<\lambda_{i+1}$, by similar argument, we can conclude
$\{\lambda_{i;j}\}_j$ converges to an eigenvalue $\lambda_k$ for some $k\ge i$,
and $\{u_{i;j}\}_j$ directionally converges to the corresponding eigenvector $u_k$.
\subsection{Rate of convergence}\label{subsec:asymptotic}
Theorem~\ref{thm:psd2} concludes the convergence of the
sequence $\{\lambda_{i;j}\}_j$,
what follows we derive a nonasymptotic estimate of
the convergence rate of $\{\lambda_{i;j}\}_j$ based on
the work of Samokish in 1958~\cite{samokish1958steepest}.
We begin by recalling the following equalities
for the projection matrix
$P_{i-1} = I - U_{i-1}U^H_{i-1}S = {U}_{i-1}^c({U}_{i-1}^c)^HS$:
\begin{subequations}
\begin{align}
P_{i-1} u_{i;j} & =u_{i;j}, \label{eq-1}\\
P_{i-1}^2 & =P_{i-1}, \label{eq-2}\\
P_{i-1}^H(H-\lambda_i S)& =(H-\lambda_i S)P_{i-1}, \label{eq-3}\\
P_{i-1}^HS & =SP_{i-1}, \label{eq-4}
\end{align}
\end{subequations}
First, we have the following lemma.
\begin{lemma}\label{lem:eigen-distribution}
Define
\begin{equation}
M=P_{i-1}^H (H-\lambda_i S)P_{i-1}
\end{equation}
and assume that $K_{i;j}$ is effectively positive definite.
\begin{itemize}
\item[(a)] $M$ is positive semi-definite and
$M = G G^H$, where $G=S{U}_i^c({\Lambda}_i^c-\lambda_i I)^{{\frac{1}{2}}}$
is of full column rank.
\item[(b)] All eigenvalues of $G^HK_{i;j}G$ are positive.
\item[(c)] The eigenvalues of $K_{i;j}M$ are given by
\begin{equation}
\lambda(K_{i;j}M)=\{0_{[i]}\}\cup\lambda\big(G^HK_{i;j}G\big),
\end{equation}
where $0_{[i]}$ stands for the multiplicity $i$ of the number 0.
\end{itemize}
\end{lemma}
\begin{proof}
(a) By the definitions of $M$ and $P_{i-1}$, it easy to see that
\begin{align*}
M& =S{U}_{i-1}^c({U}_{i-1}^c)^H(H-\lambda_i S){U}_{i-1}^c({U}_{i-1}^c)^HS\notag\\
& =S{U}_{i}^c({\Lambda}_{i}^c-\lambda_i I)({U}_{i}^c)^HS
= GG^H \ge 0,
\end{align*}
where $G=S{U}_i^c({\Lambda}_i^c-\lambda_i I)^{{\frac{1}{2}}}$.
(b) Direct calculation leads to
\begin{align*}
G^HK_{i;j} G
=({\Lambda}_i^c-\lambda_i I)^{{\frac{1}{2}}} \widetilde{K}_{22}({\Lambda}_i^c-\lambda_i I)^{{\frac{1}{2}}},
\end{align*}
where $\widetilde{K}_{22}$ is the trailing $(n-i)$-by-$(n-i)$ principal
submatrix of $K^c_{i;j}$ by deleting its first $i-1$ rows and first $i-1$ columns.
Since $K_{i;j}$ is effectively positive definite, we know that
$K^c_{i;j}>0$ and hence $\widetilde{K}_{22}>0$.
Thus all eigenvalues of $G^HK_{i;j} G$ are positive.
(c) It follows that
\begin{align*}
\lambda(K_{i;j}M)=\lambda(K_{i;j} GG^H)=\{0_{[i]}\}\cup\lambda(G^HK_{i;j} G),
\end{align*}
where we use the well-known identity
$\lambda(AB)=\{0_{[m-n]}\}\cup\lambda(BA)$ for
$A\in\mathbb{C}^{m\times n}$ and $B\in\mathbb{C}^{n\times m}$ and $m\ge n$.
\end{proof}
We now give a nonasymptotic estimate of the
convergence rate of {PSD-{id}} (Algorithm~\ref{alg:psdid}).
\begin{theorem}\label{thm:Samokish2}
Let $\epsilon_{i;j}= \lambda_{i;j}-\lambda_i$ and
$\lambda_{i;j}$ be {localized}, namely
\begin{equation}\label{asym-con}
\tau (\sqrt{\theta_{i;j}\epsilon_{i;j}}+\delta_{i;j}\epsilon_{i;j}) < 1,
\end{equation}
then
\begin{equation}\label{ineq:Samokish2}
\epsilon_{i;j+1}
\le \Bigg[\frac{\Delta + \tau\sqrt{\theta_{i;j}\epsilon_{i;j}}}
{1-\tau (\sqrt{\theta_{i;j}\epsilon_{i;j}}+
\delta_{i;j}\epsilon_{i;j})} \Bigg]^2 \epsilon_{i;j},
\end{equation}
where
$\theta_{i;j}= \|S^{{\frac{1}{2}}}K_{i;j}MK_{i;j}S^{{\frac{1}{2}}}\|$,
$\delta_{i;j}= \|S^{{\frac{1}{2}}}K_{i;j}S^{{\frac{1}{2}}}\|$,
$\Delta={(\Gamma - \gamma)}/{(\Gamma + \gamma)}$,
$\Gamma$ and $\gamma$ are
the largest and smallest positive eigenvalues of $K_{i;j}M$,
respectively, and $\tau={2}/{(\Gamma+\gamma)}$.
\end{theorem}
\begin{proof}
Recall $Z_{\bot}=[\, u_{i;j}\; P_{i-1}p_{i;j}\,]$ defined in~\eqref{eq:hr}.
It is easy to see that by using $Z_{\bot}$, the $(j+1)$th approximate eigenpair
$(\lambda_{i;j+1}, u_{i;j})$ can be written as
\[
\lambda_{i;j+1} = \min_{v}\rho(Z_{\bot}v).
\]
Considering a choice of the vector $v$ for the line search, we have
\[
\lambda_{i;j+1} = \min_{v}\rho(Z_{\bot}v)
\leq \rho(Z_{\bot}v)|_{v=[1 \; \tau]^{\top}} = \rho(z),
\]
where $z=Z_{\bot}[1\; \tau]^{\top}= u_{i;j}+\tau P_{i-1} p_{i;j}$.
Consequently, we have
\begin{align} \label{ineq:samo}
\epsilon_{i;j+1} = \lambda_{i;j+1}-\lambda_i
\le \rho(z)-\lambda_i=
\frac{z^H (H - \lambda_i S) z} {z^H S z}.
\end{align}
In the following, we provide estimates for the
numerator and denominator of the upper bound \eqref{ineq:samo}.
For the {numerator} of the upper bound in \eqref{ineq:samo}, it follows that
\begin{subequations}\label{eq-zmz}
\begin{align}
z^H(H-\lambda_i S)z
& = (u_{i;j}+\tau p_{i;j})^HP_{i-1}^H(H-\lambda_i S)P_{i-1}(u_{i;j}
+\tau p_{i;j})\notag \\
& =\|u_{i;j}+\tau p_{i;j}\|^2_{M}\notag\\
&=\|u_{i;j}-\tau P_{i-1}K_{i;j}(H-\lambda_{i;j}S)u_{i;j}\|_M^2\notag\\
&=\|u_{i;j}-\tau P_{i-1}K_{i;j}[(H-\lambda_iS)-\epsilon_{i;j}S]u_{i;j}\|_M^2\notag\\
&=\|[I-\tau P_{i-1}K_{i;j}(H-\lambda_iS)]u_{i;j}
+\tau \epsilon_{i;j} P_{i-1}K_{i;j}Su_{i;j}\|_M^2\notag\\
&=\|[I-\tau P_{i-1}K_{i;j}P_{i-1}^H(H-\lambda_iS)P_{i-1}]u_{i;j}
+\tau \epsilon_{i;j} P_{i-1}K_{i;j}S u_{i;j}\|_M^2\label{eq:zmz1}\\
& \le \big(\|[I-\tau P_{i-1}K_{i;j} M]u_{i;j}\|_{M}
+ \tau \epsilon_{i;j}\|P_{i-1}K_{i;j} Su_{i;j}\|_{M}\big)^2\label{eq:zmz4},
\end{align}
\end{subequations}
where the equality \eqref{eq:zmz1} uses the
identities \eqref{eq-1} and \eqref{eq-3}.
The inequality \eqref{eq:zmz4} uses the triangular inequality of the
vector norm induced by the semi-positive definite matrix $M$.
For the first term in \eqref{eq:zmz4}, using $M=GG^H$ and $G^HP_{i-1}=G^H$,
where $G$ is defined in Lemma~\ref{lem:eigen-distribution},
we have
\begin{align}
\|[I-\tau P_{i-1}K_{i;j} M]u_{i;j}\|_{M}
&= \| G^H[I-\tau P_{i-1}K_{i;j} GG^H ]u_{i;j}\|\notag\\
&= \|(I -\tau G^HK_{i;j} G)(G^H u_{i;j})\|\notag\\
&\le \|(I -\tau G^H K_{i;j} G)\| \|u_{i;j}\|_M.\label{eq:samo-11}
\end{align}
Note that by Lemma~\ref{lem:eigen-distribution}, it yields that
\begin{align}
\|(I -\tau G^H K_{i;j} G)\| & =\max_k|1-\tau\lambda_k(G^H K_{i;j} G)| \notag \\
& =\max\{|1-\tau\gamma|,|1-\tau\Gamma|\} \notag \\
& =\frac{\Gamma-\gamma}{\Gamma+\gamma}=\Delta. \label{eq:samo-12}
\end{align}
Consequently, we can rewrite \eqref{eq:samo-11} as
\begin{align}\label{numerator1}
\|[I-\tau P_{i-1}K_{i;j}M]u_{i;j}\|_{M}
\le \Delta\,\|u_{i;j}\|_M = \Delta \sqrt{\epsilon_{i;j}}.
\end{align}
For the second term in \eqref{eq:zmz4}:
\begin{subequations}
\label{numerator2}
\begin{align}
\|P_{i-1}K_{i;j} Su_{i;j}\|^2_M
& = {u_{i;j}^H SK_{i;j} P_{i-1}^H M P_{i-1} K_{i;j} S u_{i;j}} \notag\\
&\le \|S^{{\frac{1}{2}}}K_{i;j}MK_{i;j}S^{{\frac{1}{2}}}\| \|S^{{\frac{1}{2}}}u_{i;j}\|^2\label{eq22}\\
&=\|S^{{\frac{1}{2}}}K_{i;j}MK_{i;j}S^{{\frac{1}{2}}}\|\label{eq23}\\
&=\theta_{i;j},\notag
\end{align}
\end{subequations}
where \eqref{eq22} uses \eqref{eq-2},
\eqref{eq23} uses the fact $\|u_{i;j}\|_S=1$.
Combining \eqref{numerator1} and \eqref{numerator2},
an estimate of the {numerator} of the upper bound in \eqref{ineq:samo} is given by
\begin{align}
z^H(H-\lambda_i S)z &\leq
(\Delta+\tau\sqrt{\theta_{i;j}\epsilon_{i;j}})^2 \epsilon_{i;j}.\label{numerator}
\end{align}
For the {denominator} of the upper bound \eqref{ineq:samo}, we first note that
\begin{align}
z^HSz & =\|u_{i;j}+\tau P_{i-1} p_{i;j}\|_S^2 \notag \\
& \ge (\|u_{i;j}\|_S - \tau\|P_{i-1} p_{i;j}\|_S)^2 \notag \\
& = (1- \tau\|P_{i-1} p_{i;j}\|_S)^2. \label{ineq:zsz}
\end{align}
By calculations, we have the following upper bound for $\|P_{i-1} p\|_S$:
\begin{subequations}
\label{denominator}
\begin{align}
\|P_{i-1} p_{i;j}\|_S
&= \| P_{i-1} K_{i;j} (H-\lambda_{i;j} S)u_{i;j} \|_S \notag\\%\label{eq31}\\
&=\|P_{i-1} K_{i;j}(H-\lambda_i S)u_{i;j} -\epsilon_{i;j}P_{i-1} K_{i;j}Su_{i;j}\|_S\notag\\%\label{eq32}\\
&\le \|P_{i-1} K_{i;j}(H-\lambda_i S)u_{i;j}\|_S +\epsilon_{i;j} \|P_{i-1} K_{i;j}Su_{i;j}\|_S\notag\\
& = \|P_{i-1} K_{i;j}Mu_{i;j}\|_S +\epsilon_{i;j} \|P_{i-1} K_{i;j}Su_{i;j}\|_S\label{eq33}\\
& \le \|S^{{\frac{1}{2}}}P_{i-1}S^{-{\frac{1}{2}}}\|\|S^{{\frac{1}{2}}} K_{i;j} M^{{\frac{1}{2}}}\| \|M^{{\frac{1}{2}}}u_{i;j}\|
+ \epsilon_{i;j} \|S^{{\frac{1}{2}}}P_{i-1} S^{-{\frac{1}{2}}}\| \|S^{{\frac{1}{2}}}K_{i;j}S^{{\frac{1}{2}}}\| \|S^{\frac{1}{2}} u_{i;j}\|\notag\\
&\le\sqrt{\theta_{i;j}\epsilon_{i;j}} +\delta_{i;j}\epsilon_{i;j}, \label{eq34}
\end{align}
\end{subequations}
where the equality \eqref{eq33} uses \eqref{eq-1} and \eqref{eq-3},
the inequality \eqref{eq34} uses the fact $\|S^{\frac{1}{2}} P_{i-1} S^{-{\frac{1}{2}}}\|\le 1$.
By \eqref{ineq:zsz} and \eqref{eq34}, if
\[
\tau (\sqrt{\theta_{i;j}\epsilon_{i;j}}+\delta_{i;j}\epsilon_{i;j}) < 1,
\]
then the {denominator} of the upper bound \eqref{ineq:samo} satisfies
\begin{equation}\label{ineq:zsz2}
z^HSz \geq (1- \tau (\sqrt{\theta_{i;j}\epsilon_{i;j}}+\delta_{i;j}\epsilon_{i;j}))^2.
\end{equation}
By combining \eqref{ineq:samo}, \eqref{numerator}
and \eqref{ineq:zsz2}, we derive the
estimate \eqref{ineq:Samokish2}. This concludes the proof.
\end{proof}
Theorem~\ref{thm:Samokish2} indicates that
if $\lambda_{i;j}$ is localized (i.e., \eqref{asym-con} is satisfied),
and $\Delta +\tau\sqrt{\theta_{i;j}\epsilon_{i;j}}\rightarrow 0$
as $j\rightarrow \infty$, then the {PSD-{id}} method converges
{\em superlinearly}. In this case, we may call that the preconditioner $K_{i;j}$
is {\em asymptotically optimal}. In next section, we will consider
such a preconditioner.
\medskip
To end this section
we note that for the smallest eigenvalue $\lambda_1$, if
the preconditioner $K_{i;j}$ is chosen to be fixed and positive definite,
i.e., $K_{i;j} = K > 0$,
one can verify that $\theta_{i;j}\le\Gamma\delta_{i;j}$.
Theorem~\ref{thm:Samokish2} becomes the classical
Samokish's theorem~\cite{samokish1958steepest,ovtchinnikov2006sharp},
which remains asymptotically most accurate estimate of the convergence
rate of the PSD method and is proven to be {\em sharp} \cite{ovtchinnikov2006sharp}.
The proof of Theorem~\ref{thm:Samokish2} relies
on the triangular inequality~\eqref{eq-zmz},
which is inspired by the proof of Samokish's theorem
presented in \cite{ovtchinnikov2006sharp}. However,
the treatment of each term in \eqref{ineq:samo} needs to be handled
diligently in order to accommodates the effects of the projection
matrix $P_{i-1}$ and the flexible preconditioner $K_{i;j}$.
\section{An asymptotically optimal preconditioner}\label{sec:opt}
In this section, we consider the shift-and-invert preconditioner:
\begin{equation}\label{eq:kbeta}
\widehat{K}_{i;j}=\big(H-\beta_{i;j} S\big)^{-1},
\end{equation}
where $\beta_{i;j}$ is the shift.
The following theorem shows that with a proper choice of $\beta_{i;j}$,
$\widehat{K}_{i;j}$ is asymptotically optimal and consequently,
the {PSD-{id}} method converges superlinearly.
\begin{theorem}\label{lem:requirement}
Consider the shift
\begin{equation}\label{eq:beta}
\beta_{i;j} = \lambda_{i;j} - c \|r_{i;j}\|_{S^{-1}},
\end{equation}
where the constant
$c= \inf_k {\sqrt{(\lambda_{i;k}-\lambda_i)(\lambda_{i+1}-\lambda_{i;k})}}/{\|r_{i;k}\|_{S^{-1}}}$.
If
\begin{subequations}
\begin{equation} \label{requirement-c}
c >3\sqrt{\Delta_{i;j}}
\end{equation}
and
\begin{equation} \label{requirement2-1}
0 <\Delta_{i;j}<\min \left\{\frac{\Delta_i^2}{4}, \frac{1}{10}\right\},
\end{equation}
\end{subequations}
where $\Delta_i={(\lambda_i-\lambda_{i-1})}/{(\lambda_{i+1}-\lambda_i)}$ and
$\Delta_{i;j}={(\lambda_{i;j}-\lambda_i)}/{(\lambda_{i+1}-\lambda_{i;j})}$.
Then
\begin{itemize}
\item[(a)]
$\beta_{i;j}<\lambda_i$ and
$\widehat{K}_{i;j}$ is effectively positive definite.
\item[(b)]
$\beta_{i;j} \rightarrow \lambda_i$ as $j\rightarrow\infty$.
\item[(c)]
The condition \eqref{asym-con} of Theorem~\ref{thm:Samokish2} is satisfied,
namely, $\lambda_{i;j}$ is {\em localized}.
\item[(d)]
$\Delta+\tau\sqrt{\theta_{i;j}\epsilon_{i;j}} \rightarrow 0$
as $j\rightarrow\infty$.
\end{itemize}
By (c) and (d), the preconditioner $\widehat{K}_{i;j}$
is asymptotically optimal.
\end{theorem}
\begin{proof} ~
(a) By the condition \eqref{requirement2-1},
the relation $0<\Delta_{i;j}<0.1$ implies that
$\lambda_i$ is the closest eigenvalue to $\lambda_{i;j}$.
Let $u_{i;j}=U_{i-1}^c a$ for some $a$,
then $(\lambda_{i;j}, a)$ is an approximated eigenpair of $\Lambda_{i-1}^c$.
Using the Kato-Temple inequality \cite{kato1950upper}, we have
\begin{equation} \label{eq:ktineq}
(\lambda_{i;j}-\lambda_i)(\lambda_{i+1}-\lambda_{i;j})\le \| (\Lambda_{i-1}^c-\lambda_{i;j} I) a\|_2
=\|r_{i;j}\|_{S^{-1}}^2.
\end{equation}
Therefore, the result $\beta_{i;j} < \lambda_i$ is verified as follows:
\begin{align*}
\beta_{i;j}-\lambda_i
& =\lambda_{i;j} - c \|r_{i;j}\|_{S^{-1}} - \lambda_i \\
& \le \lambda_{i;j} -\lambda_i -c\sqrt{(\lambda_{i;j}-\lambda_i)(\lambda_{i+1}-\lambda_{i;j})} \\
& =(\lambda_{i+1}-\lambda_{i;j})(\Delta_{i;j}-c\sqrt{\Delta_{i;j}})<0,
\end{align*}
where for the last inequality we used the condition~\eqref{requirement-c}.
The preconditioner $\widehat{K}_{i;j}$ is effectively positive definite since
\begin{equation}\label{eq:kijc}
\widehat{K}^c_{i;j}=(U_{i-1}^c)^HS\widehat{K}_{i;j}SU_{i-1}^c
=\diag\Big(\frac{1}{\lambda_i-\beta_{i;j}},\frac{1}{\lambda_{i+1}-\beta_{i;j}},
\dots,\frac{1}{\lambda_{n}-\beta_{i;j}}\Big)
\end{equation}
and $\beta_{i;j} <\lambda_i$.
(b)
By Theorem~\ref{thm:psd2}, we have
\begin{equation} \label{eq:betatolambda}
\beta_{i;j} \rightarrow \lambda_i \quad \mbox{as $j\rightarrow\infty$}.
\end{equation}
(c) With the choice of $\beta_{i;j}$ in \eqref{eq:beta}, for $\theta_{i;j}$, we have
\begin{equation} \label{eq:theta}
\theta_{i;j}
= \|S^{{\frac{1}{2}}} \widehat{K}_{i;j}M \widehat{K}_{i;j} S^{{\frac{1}{2}}}\|
= \max_{i+1 \le k \le n} \frac{\lambda_k-\lambda_i}{(\lambda_k-\beta_{i;j})^2}
= \frac{\lambda_{i+1}-\lambda_i}{(\lambda_{i+1}-\beta_{i;j})^2},\\
\end{equation}
where for the last equality,
we only need to show that $f'(x)<0$ for $x\ge \lambda_{i+1}$,
where $f(x)=\frac{x-\lambda_i}{(x-\beta_{i;j})^2}$.
By calculations, we have
\[
f'(x)=\frac{2\lambda_i -x-\beta_{i;j}}{(x-\beta_{i;j})^3}<0
\]
since $x-\beta_{i;j}>0$ and
\begin{align*}
2\lambda_i -x - \beta_{i;j}
&\le 2\lambda_i -\lambda_{i+1} -\lambda_{i;j}+c\|r_{i;j}\|_{S^{-1}}
< \lambda_{i;j} -\lambda_{i+1} + \sqrt{(\lambda_{i;j}-\lambda_i)(\lambda_{i+1}-\lambda_{i;j})}\\
&= (\lambda_{i+1}-\lambda_{i;j})(-1+\sqrt{\Delta_{i;j}})<0.
\end{align*}
For $\delta_{i;j}$, we have
\begin{equation} \label{eq:delta}
\delta_{i;j}
=\|S^{{\frac{1}{2}}} \widehat{K}_{i;j} S^{{\frac{1}{2}}}\|
=\frac{1}{\min_{1\le k \le n}|\lambda_k-\beta_{i;j}|}
=\frac{1}{\lambda_i-\beta_{i;j}},\\
\end{equation}
where for the last equality, we only need to show that $\beta_{i;j}-\lambda_{i-1}>\lambda_i-\beta_{i;j}$,
which is equivalent to
\begin{align*}
2c\|r_{i;j}\|_{S^{-1}} < 2\lambda_{ij}-\lambda_i-\lambda_{i-1}.
\end{align*}
Notice that the right hand side of the above inequality is no less than $\lambda_i-\lambda_{i-1}$,
thus, we only need to show
\[
2c\|r_{i;j}\|_{S^{-1}}<\lambda_i-\lambda_{i-1}.
\]
By calculations, we have
\begin{align*}
\frac{\lambda_i-\lambda_{i-1}}{2c\|r_{i;j}\|_{S^{-1}}}
\ge \frac{\lambda_i-\lambda_{i-1}}{ 2 \sqrt{(\lambda_{i;j}-\lambda_i)(\lambda_{i+1}-\lambda_{i;j})}}
>\frac{ \Delta_i}{2\sqrt{\Delta_{i;j}}} \ge 1.
\end{align*}
In addition, using Lemma~\ref{lem:eigen-distribution}(c) and \eqref{eq:kijc},
it is easy to see that
\begin{align} \label{eq:temp}
\Gamma
= \frac{\lambda_n-\lambda_i}{\lambda_{n}-\beta_{i;j}}
\quad \mbox{and} \quad
\gamma
= \frac{\lambda_{i+1}-\lambda_i}{\lambda_{i+1}-\beta_{i;j}}.
\end{align}
Then it follows that
\begin{align*}
\tau & = 2/(\Gamma + \gamma) \leq 1/\gamma, \\
\tau\sqrt{\theta_{i;j}\epsilon_{i;j}} &
\le\frac{1}{\gamma}\sqrt{\theta_{i;j}\epsilon_{i;j}}
= \frac{\lambda_{i+1}-\beta_{i;j}}{\lambda_{i+1}-\lambda_i}
\sqrt{\frac{\lambda_{i+1}-\lambda_i}{(\lambda_{i+1}-\beta_{i;j})^2}(\lambda_{i;j}-\lambda_{i})}
= \sqrt{\Delta_{i;j}},\\
\frac{1}{\gamma}
= \frac{\lambda_{i+1} - \lambda_{i;j} +{c\|r_{i;j}\|_{S^{-1}}} }{\lambda_{i+1}-\lambda_i}
< \frac{\lambda_{i+1}-\lambda_{i;j}+\sqrt{(\lambda_{i;j}-\lambda_i)(\lambda_{i+1}-\lambda_{i;j})}}{\lambda_{i+1}-\lambda_{i;j}}
=1+\sqrt{\Delta_{i;j}},\\
\delta_{i;j}\epsilon_{i;j}&=\frac{\lambda_{i;j}-\lambda_i}{\lambda_i-\beta_{i;j} }
=\frac{\lambda_{i;j}-\lambda_i}{\lambda_i- \lambda_{i;j} +{c\|r_{i;j}\|_{S^{-1}}} }
<\frac{\lambda_{i;j}-\lambda_i}{\lambda_i- \lambda_{i;j} +{3\sqrt{\Delta_{i;j}}\|r_{i;j}\|_{S^{-1}}} }\\
&\le\frac{\lambda_{i;j}-\lambda_i}{\lambda_i- \lambda_{i;j} +{3\sqrt{\Delta_{i;j}} \sqrt{(\lambda_{i;j}-\lambda_i)(\lambda_{i+1}-\lambda_{i;j})} } }=\frac{1}{-1+3}=\frac{1}{2}.
\end{align*}
Therefore,
\begin{align*}
\tau (\sqrt{\theta_{i;j}\epsilon_{i;j}}+\delta_{i;j}\epsilon_{i;j})
\le \sqrt{\Delta_{i;j}} +\frac{1+\sqrt{ \Delta_{i;j}}}{2} < 1 .
\end{align*}
(d) By the expressions \eqref{eq:temp} of $\Gamma$ and $\gamma$, we have
\begin{align*}
\Delta=\frac{\Gamma-\gamma}{\Gamma+\gamma}
&=\frac{(\lambda_n-\lambda_i)(\lambda_{i+1}-\beta_{i;j}) -
(\lambda_{i+1}-\lambda_i)( \lambda_{n}-\beta_{i;j}) }
{(\lambda_n-\lambda_i)(\lambda_{i+1}-\beta_{i;j}) +
(\lambda_{i+1}-\lambda_i)( \lambda_{n}-\beta_{i;j})}\notag\\
&=\frac{(\lambda_n-\lambda_{i+1})(\lambda_{i}-\beta_{i;j}) }
{(\lambda_n-\lambda_i)(\lambda_{i+1}-\beta_{i;j}) +
(\lambda_{i+1}-\lambda_i)( \lambda_{n}-\beta_{i;j})}\notag\\
&< \frac{(\lambda_n-\lambda_{i+1})(\lambda_{i}-\beta_{i;j}) }{2(\lambda_n-\lambda_i)(\lambda_{i+1}-\lambda_i) }
< \frac{\lambda_{i}-\beta_{i;j}}{2(\lambda_{i+1}-\lambda_i)}, \label{ineq:beta}
\end{align*}
Combining the above estimates of $\Delta$,
$\theta_{i;j}\epsilon_{i;j}$ and $\tau$,
we have
\begin{equation} \label{eq:deltatau}
\Delta+\tau\sqrt{\theta_{i;j}\epsilon_{i;j}}
< \frac{\lambda_{i}-\beta_{i;j}}{2(\lambda_{i+1}-\lambda_i)} +
\sqrt{\Delta_{i;j}}.
\end{equation}
By Theorem~\ref{thm:psd2}(a) and
the result (a) of this theorem,
the upper bound of \eqref{eq:deltatau} converges to zero as $j\rightarrow \infty$.
\end{proof}
Four remarks are in order.
\begin{remark}{\rm
By the definition of the constant $c$ in \eqref{eq:beta}, we have
\[
c\|r_{i;j}\|_{S^{-1}} \le
\sqrt{(\lambda_{i+1}-\lambda_i)(\lambda_{i;j}-\lambda_i)}
\]
and
\begin{equation}\label{eq:betaapprox}
\beta_{i;j}=\lambda_{i;j}+\mathcal{O}((\lambda_{i;j}-\lambda_i)^{\frac{1}{2}}).
\end{equation}
Therefore, in practice,
we can replace the shift $\beta_{i;j}$ by $\lambda_{i;j}$, and use
the preconditioner
\begin{equation}\label{eq:ki}
{\widetilde{K}}_{i;j} = (H-\lambda_{i;j}S)^{-1}.
\end{equation}
We call the preconditioner
${\widetilde{K}}_{i;j}$ a {\em locally accelerated} preconditioner.
} \end{remark}
\begin{remark}{\rm
With the locally accelerated preconditioner $\widetilde{K}_{i;j}$,
the corresponding search vector
$\widetilde{p}_{i;j}=-\widetilde{K}_{i;j}r_{i;j}$.
A direct calculation gives rise to
\begin{align*}
U^HS(u_{i;j}+\widetilde{p}_{i;j})
&= \begin{bmatrix}
0 & \dots & 0 &
\frac{\lambda_{i;j}-\beta_{i;j}}{\lambda_{i}-\beta_{i;j}}a_{i} &\dots &
\frac{\lambda_{i;j}-\beta_{i;j}}{\lambda_{n}-\beta_{i;j}}a_{n} \end{bmatrix}^{\top},
\end{align*}
where we use the fact $U^HSu_{i;j}=a=[0,\dots,0,a_i,\dots,a_n]^{\top}$.
Then by Theorem~\ref{lem:requirement}(b),
we have $U^HS(u_{i;j}+\widetilde{p}_{i;j}) \rightarrow e_i^{\top}$
as $j\rightarrow \infty$.
In the notion of an ideal search vector introduced at the end of
section~\ref{sec:dpsd}, the search vector $\widetilde{p}_{i;j}$
is an asymptotically ideal search vector.
} \end{remark}
\begin{remark}{\rm
Before $\lambda_{i;j}$ is localized,
we can use a fixed preconditioner $K_{i;j} = K$ for all $j$.
An obvious choice is to set $K_{i;j}\equiv K_{\sigma}=(H-\sigma S)^{-1}$
for some $\sigma<\lambda_1$. $K_{\sigma}$ is symmetric positive definite and
can be regarded as a global preconditioner for the initial few iterations.
By the convergence of {PSD-{id}} (Theorem~\ref{thm:psd2}),
it is guaranteed that the sequence $\{\lambda_{i;j}\}_j$ is strictly
monotonically decreasing, albeit the convergence may be slow before the locally
accelerated preconditioner $\widetilde{K}_{i;j}$ is applied, see
the numerical illustration in section~\ref{sec:numer}.
}\end{remark}
\begin{remark}{\rm
As we discussed in section~\ref{sec:intro},
we are particularly interested in solving
ill-conditioned Hermitian-definite generalized eigenvalue problem~\eqref{eq:ghep}
where $H$ and $S$ sharing a common near-nullspace $\mathcal{V}$,
whose dimension can be large.
If we set the preconditioner $K_{i;j}\equiv I$, then
$K_{i;j}M=M=S{U}_{i}^c({\Lambda}_{i}^c-\lambda_i I)({U}_{i}^c)^HS$,
which has a near-nullspace $\mathcal{V}$, and a nullspace $\subspan(U_i)$.
As $\dim(\mathcal{V})>\dim(\subspan(U_i))$, $K_{i;j}M$ has
very small positive eigenvalues.
Therefore, ${\Gamma}/{\gamma}\gg 1$, and $\Delta\approx 1$.
By Theorem~\ref{thm:Samokish2}, we know that the {PSD-{id}} method
would converge linearly.
By a similar arguments, we can declare that for any well-conditioned preconditioner
$K_{i;j}$, the {PSD-{id}} method would also converge linearly.
Therefore, in order to achieve the fast convergence, one has to
apply an ill-conditioned preconditioner such as
the locally accelerated preconditioner ${\widetilde{K}}_{i;j}$.
} \end{remark}
\section{Numerical examples}\label{sec:numer}
In this section, we use a MATLAB implementation
for the {PSD-{id}} method (Algorithm~\ref{alg:psdid}) with
locally accelerated preconditioners ${\widetilde{K}}_{i;j}$
defined in \eqref{eq:ki} to generate
two numerical examples to verify the convergence and the rate of
convergence of the method.
To illustrate the efficiency of the method, we focus on
two ill-conditioned generalized eigenvalue problems~\eqref{eq:ghep} arising
from the PUFE approach to solve differential eigenvalue equations arsing
in quantum mechanics.
Matlab scripts of the implementation of the {PSD-{id}} method and
the data that used to generate numerical results
presented in this section can be obtained from the URL
http://dsec.pku.edu.cn/$\sim$yfcai/psdid.html.
\medskip
To apply ${\widetilde{K}}_{i;j}$, we need to test the
localization conditions \eqref{requirement-c} and \eqref{requirement2-1}
of the $j$th approximate eigenvalue $\lambda_{i;j}$.
For the condition \eqref{requirement-c}, note that $c$ is a constant and
$\Delta_{i;j}$ in limit is zero.
Therefore, when the residual $r_{i;j}$ is sufficiently small,
$\lambda_{i;j}$ is close enough to $\lambda_i$,
then the condition \eqref{requirement-c} will be satisfied.
Therefore, the test of the condition \eqref{requirement-c} can be
replaced by the following residual test:
\begin{equation} \label{con-2}
\mbox{Res}[\lambda_{i;j},u_{i;j}]=
\frac{\|Hu_{i;j}-\lambda_{i;j}Su_{i;j}\|}
{\|Hu_{i;j}\|+|\lambda_{i;j}| \|Su_{i;j}\|} \leq \tau_1,
\end{equation}
where $\tau_1$ is some prescribed threshold, say $\tau_1 = 0.1$.
For the condition \eqref{requirement2-1},
we need the estimates of eigenvalues $\lambda_i$ and $\lambda_{i+1}$
to approximate the quantities
$\Delta_i = {(\lambda_{i}-{\lambda}_{i-1})}/{({\lambda}_{i+1} - \lambda_{i})}$
and
$\Delta_{i;j} = {(\lambda_{i;j}-{\lambda}_{i})}/{({\lambda}_{i+1} - \lambda_{i;j})}$.
For $\Delta_i$,
it is natural to take
the $j$th approximates
${\lambda}_{i;j}$ and ${\lambda}_{i+1;j}$
of ${\lambda}_{i}$ and ${\lambda}_{i+1}$ respectively and yields the following estimate of $\Delta_i$
\[
\Delta_i \approx
\widehat\Delta_i =
\frac{\lambda_{i;j}-{\lambda}_{i-1}}{{\lambda}_{i+1;j} - \lambda_{i;j}}.
\]
For $\Delta_{i;j}$,
if we simply use $\lambda_{i;j}$ to estimate $\lambda_i$,
then it leads to $\Delta_{i;j}=0$.
This violates the condition \eqref{requirement2-1}.
A better estimate of $\lambda_i$ is
to use the {linear extrapolation}
$\widehat{\lambda}_i=2\lambda_{i;j}-\lambda_{i;j-1}$
of $\lambda_{i;j-1}$ and $\lambda_{i;j}$
for $j > 1$. Note that when $j=1$, all approximated eigenvalues
are assumed to be not localized.
Then it yields the following estimate of $\Delta_{i;j}$:
\[
\Delta_{i;j} \approx
\widehat\Delta_{i;j} = \frac{\lambda_{i;j}-\widehat{\lambda}_{i}}{{\lambda}_{i+1;j}-\lambda_{i;j}}
= \frac{\lambda_{i;j-1}-\lambda_{i;j}}{\lambda_{i+1;j} - \lambda_{i;j}}.
\]
In order to estimate ${\lambda}_{i+1}$, the Rayleigh-Ritz projection subspace
in {PSD-{id}} is spanned by the columns of
$Z=[U_{i-1}\; u_{i;j}\; \dots\; u_{i+\ell;j}\; p_{i;j}]$ for some $\ell > 1$.
In this case, the {PSD-{id}} method will also compute $\lambda_{i+1;j}$,
which can be used to approximate $\lambda_{i+1}$.
By the estimates $\widehat\Delta_{i}$ and $\widehat\Delta_{i;j}$,
the localization condition \eqref{requirement2-1} of the
$j$th approximate eigenvalue $\lambda_{i;j}$ of $\lambda_i$
can be verified by the following condition
\begin{equation} \label{con-1}
\widehat\Delta_{i;j} <
\min\left\{\frac{1}{4} \widehat\Delta^2_i, 0.1 \right\} \equiv \tau_2.
\end{equation}
Note that for computing the smallest eigenvalue $\lambda_1$,
we let the initial approximate $\lambda_{0,j} = \sigma$
for some $\sigma < \lambda_1$. Here $\sigma$ is a user given parameter or
a lower bound of $\lambda_1$,
say $\lambda_{1;j}-\|r_{1;j}\|_{S^{-1}}\approx \lambda_{1;j}-\|r_{1;j}\|$.
\medskip
We use the preconditioned MINRES \cite{paige1975solution} to compute
the preconditioned search vector
\begin{equation}\label{eq:pre}
p_{i;j}=- {\widetilde{K}}_{i;j} r_{i;j}
= -(H-\lambda_{i;j} S)^{-1} r_{i;j}.
\end{equation}
In practice, the vector $p_{i;j}$ is just needed to be computed
approximately such that
\begin{equation}\label{pij-con}
\|(H-\lambda_{i;j})p_{i;j}+r_{i;j}\| \le \eta_{i;j}\|r_{i;j}\|,
\end{equation}
where $\eta_{i;j}<1$ is a parameter. In our numerical experiments,
the preconditioner of the MINRES is $S^{-1}$,
$\eta_{i;j}=\mbox{Res}[\lambda_{i;j},u_{i;j}]$,
and the maximum number of MINRES iterations is set to be 200.
\medskip
All numerical experiments are
performed on a quad-core $\text{Intel}^{\tiny\textregistered} \text{ Xeon}^{\tiny\textregistered}$ Processor E5-2643 running at 3.30GHz
with 31.3GB RAM, machine epsilon $\varepsilon\approx 2.2\times 10^{-16}$.
\begin{example} \label{eg:1dho} {\rm
Consider the following Schr\"{o}dinger equation for
a one-dimensional harmonic oscillator:
\begin{align}\label{eq:oscillator}
-{\frac{1}{2}} \psi''(x)+{\frac{1}{2}} x^2 \psi(x) =
E \psi(x), \qquad -L\le x\le L,\qquad \psi(-L)=\psi(L)=0,
\end{align}
where $E$ is the energy, $\psi(x)$ is the wavefunction.
This is a well-known model system
in quantum mechanics \cite{liboff2003introductory,griffiths2004introduction}.
If $L=\infty$, the eigenvalues
of the equation \eqref{eq:oscillator} are $\lambda_i=i-0.5$ and
the corresponding eigenfunctions are $\psi_i(x)=H_i(x)e^{-0.5x^2}$,
where $H_i(x)$ is the $i$th order Hermite polynomial \cite[Chap. 18]{olver2010nist}.
\medskip
For numerical experiments,
we set $L=10$ since $\psi_i(x)$ is numerically zero for $|x|>10$.
We discretize the equation \eqref{eq:oscillator} by linear finite element (FE),
cubic FE and partition of unit FE (PUFE) \cite{melenk1996partition}, respectively.
In all three cases, the eigenfunction $\psi(x)$ is approximated by
\begin{equation} \label{eq:pufebasis}
\psi^h(x) =\sum_{i} c_i \phi_i (x)+
\sum_{\alpha}\sum_{j} c_{j\alpha}\phi_j^{PU}(x)\tilde{\psi}_{\alpha}(x)
\equiv \sum_{k=1}^n u_k \Phi_k(x),
\end{equation}
where $\phi_i(x)$ are the FE basis functions,
$\phi_j^{PU}$ are the FE basis function to form
enriched basis functions, $\tilde{\psi}_{\alpha}(x)$ are enrichment functions,
and $c_i$ and $c_{j\alpha}$ are coefficients.
The enrichment term vanishes in the linear and cubic FE cases.
In our numerical experiments, the interval $[-10,10]$ is divided uniformly,
and for PUFE,
$\phi_i(x)$, $\phi_j^{PU}(x)$ are chosen to be cubic and linear, respectively,
and $\tilde{\psi}_{\alpha}(x)=e^{-0.4x^2}$ for $x\in[-5,5]$ and zero elsewhere.
\medskip
Converting \eqref{eq:oscillator} into its weak form, and
using $\Phi_i$ as the test functions,
we obtain an algebraic generalized eigenvalue problem~\eqref{eq:ghep},
where $u=[u_1\, u_2\, \dots\, u_n]$ and $(i,j)$ elements
$h_{ij}$ and $s_{ij}$ of $H$ and $S$ are given by
\[
h_{ij}=\int_{-10}^{10}(\Phi'(x)\Phi'_j(x)+\frac12 x^2\Phi_i(x)\Phi_j(x))dx
\quad \mbox{and} \quad
s_{ij}=\int_{-10}^{10} \Phi_i(x)\Phi_j(x)dx,
\]
respectively.
The left plot of Figure~\ref{fig:1dh}
shows the errors of the sums of the four
smallest eigenvalues of $(H,S)$ with respect to the number of FEs of
three different finite element discretizations.
The matrix sizes of linear FE are $7, 15, 31, 63, 127, 255$ and $511$.
The matrix sizes for the cubic FE are $23, 47, 95$ and $191$.
The matrix sizes for the PUFE are $28, 56$ $112$.
By the plot, we can see that to achieve the same accuracy,
the matrix sizes of the PUFE are much smaller.
However, the condition numbers of PUFE matrices $H$, $S$
are large; $(\kappa_2(H), \kappa_2(S)) =
(3.0\times 10^6, 5.0\times 10^6),
(6.5\times 10^8, 2.7\times 10^9),
(8.8\times 10^{10}, 8.1\times 10^{11})$, respectively.
\medskip
For demonstrating the convergence behavior of {PSD-{id}} ,
let us compute $m=4$ smallest eigenvalues of
the PUFE matrices $H$ and $S$ of order $n=112$, which corresponds
to the mesh size $h=2L/32$. The matrices $H$ and $S$ are ill-conditioned,
$(\kappa_2(H), \kappa_2(S)) = (8.8\times 10^{10}, 8.1\times 10^{11})$.
Furthermore, $H$ and $S$ share a common near-nullspace, namely
there exists a subspace $\subspan(V)$ of dimension 17 such that
$\|H V\| = \|SV\| = O(10^{-5})$.
To compute $4$ smallest eigenpairs,
we run the {PSD-{id}} algorithm for $i=1,2,3,4$ with
$\ell=4$. The accuracy threshold of computed eigenvalues
is $\tau_{\rm eig}=10^{-9}$.
$\tau_1=0.1$ is used for the residual test \eqref{con-2}.
\medskip
The right plot of Figure~\ref{fig:1dh} shows
the convergence history in the relative residuals
$\mbox{Res}[\lambda_{i;j},u_{i;j}]$ of the {PSD-{id}} method
for computing four smallest eigenvalues.
The localization (i.e., the conditions \eqref{con-2} and \eqref{con-1} are
satisfied) of the $j$ approximate eigenpair $(\lambda_{i;j},u_{i;j})$ for
computing the $i$th eigenvalue $\lambda_i$ is marked by ``+'' sign.
The locally accelerated preconditioner
${\widetilde{K}}_{i;j} = (H-\lambda_{i;j} S)^{-1}$
is used once $\lambda_{i;j}$ is localized.
As Theorem~\ref{lem:requirement} predicts,
the locally accelerated preconditioner ${\widetilde{K}}_{i;j}$
is asymptotically optimal and leads to superlinear convergence
of the {PSD-{id}} algorithm.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{./eigerr.eps}
\includegraphics[width=0.45\textwidth]{./PSDid_Harmonic1d.eps}
\end{center}
\caption{Left: error of the sum of four smallest eigenvalues of
$(H,S)$ with respect to the number of FEs of three different FE discretizatoins
in Example~\ref{eg:1dho}.
Right: convergence of the {PSD-{id}} method for computing four smallest eigenvalues.
} \label{fig:1dh}
\end{figure}
} \end{example}
\begin{example}\label{eg:ceal} {\rm
The Hermitian-definite generalized eigenvalue problem \eqref{eq:ghep} is a
computational kernel in quantum mechanical methods employing a
nonorthogonal basis for {\em ab initio} three-dimensional electronic
structure calculations, see~\cite{cai2013hybrid} and references therein.
In this example, we select a sequence of eigenproblems produced by the PUFE
method for a self-consistent pseudopotential density functional
calculation for metallic, triclinic CeAl~\cite{sukumar2009classical,pask2012linear,pask2011partition}. The Brillouin zone
is sampled at two ${\bf k}$-points:
${\bf k} = (0.00, \, 0.00,\, 0.00)$ and
${\bf k} = (0.12, \, -0.24,\, 0.37)$. The PUFE approximation for the
wavefunction is of the form given in the equation~\eqref{eq:pufebasis}
and we apply a standard Galerkin procedure to set up
the discrete system matrices. The unit cell is a triclinic box,
with atoms displaced from ideal positions. The primitive lattice
vectors and the position of the atomic centers are
\[
{\bf a}_1 = a(1.00 \quad 0.02 \quad -0.04), \quad
{\bf a}_2 = a(0.01 \quad 0.98 \quad 0.03), \quad
{\bf a}_3 = a(0.03 \quad -0.06 \quad 1.09)
\]
and
\[
{\bf \tau}_{\rm{Ce}} = a(0.01 \quad 0.02 \quad 0.03), \quad
{\bf \tau}_{\rm{Al}} = a(0.51 \quad 0.47 \quad 0.55),
\]
with lattice parameter $a = 5.75$ bohr.
Since Ce has a full complement of $s$, $p$, $d$, and $f$ states in
valence, it requires 17 enrichment functions to span the occupied
space. The near-dependencies between the finite element basis
functions and the enriched basis functions lead to an
ill-conditioned generalized eigenvalue problem~\eqref{eq:ghep}.
In this numerical example, the matrix size of $H$ and $S$
is $n=7\times 8^3 + 1752=5336$.
Both $H$ and $S$ are ill conditioned and their condition numbers are
$(\kappa_2(H), \kappa_2(S)) = (1.1641\times 10^{10}, 2.5731\times 10^{11})$.
Furthermore, $H$ and $S$ share a common near-nullspace $\subspan(V)$ of
dimension 1000 such that $\|H V\| = \|SV\| = O(10^{-4})$,
where $V$ is orthonormal. This is an extremely ill-conditioned eigenvalue problem.
Figure~\ref{fig:CeAl} shows
the convergence history of the {PSD-{id}} method for computing
four smallest eigenvalues. As in Figure~\ref{fig:1dh},
the localization of the $j$ approximate eigenpair $(\lambda_{i;j},u_{i;j})$
is marked by ``+'' sign.
Once $\lambda_{i;j}$ is localized,
the locally accelerated preconditioner
${\widetilde{K}}_{i;j} = (H-\lambda_{i;j} S)^{-1}$
is used.
Again, as Theorem~\ref{lem:requirement} predicts,
the locally accelerated preconditioner ${\widetilde{K}}_{i;j}$
leads to superlinear convergence of the {PSD-{id}} algorithm.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{./PSDid_CeAl_PUFE.eps}
\end{center}
\caption{
Convergence of the {PSD-{id}} method for computing four smallest eigenvalues
of the CeAl matrix pair described in Example~\ref{eg:ceal}.
} \label{fig:CeAl}
\end{figure}
} \end{example}
\section{Conclusion}\label{sec:conclusion}
In this paper, we proved the convergence of
the {PSD-{id}} method, and derived a nonasymptotic estimate of
the rate of convergence of the method.
We show that with the proper choice of the shift,
the indefinite shift-and-invert preconditioner
is a locally accelerated preconditioner and
leads to superlinear convergence.
Two numerical examples are presented to verify the theoretical results
on the convergence behavior of the {PSD-{id}} method for solving
ill-conditioned Hermitian-definite generalized eigenvalue problems.
\bibliographystyle{abbrv}
|
\section{Introduction}
Strongly correlated materials form one of the most challenging problems in condensed matter physics. Determining the electronic structure of a realistic material usually involves a multi-step process. First, the original problem is downfolded to a model system that describes the strongly correlated sector. Subsequently, an approximate solution for the model system is sought.
The model system should be simple enough that it is computationally tractable.
The Hubbard model of itinerant electrons with a local interaction is a popular choice~\cite{Hubbard63,Gutzwiller63,Kanamori63,Hubbard64,Gutzwiller64}.
The original electrons of the model, however, have a long-range Coulomb interaction, so in general one would prefer to use an extended Hubbard model that also includes the non-local interactions, if this was computationally feasible.
The Peierls-Feynman-Bogoliubov variational principle~\cite{Peierls38,Bogoliubov58,Feynman72} can be used to map extended Hubbard models with non-local interactions to effective models with only local interactions~\cite{schuler_optimal_2013}. Previously, this method has been used to estimate the effective local interaction in materials with a hexagonal lattice, such as graphene~\cite{schuler_optimal_2013}. Our focus here is on the square lattice Hubbard model with nearest-neighbor interaction. This system has been studied extensively as a testbed for theories developed over the last few years, using, e.g., Quantum Monte Carlo~\cite{zhang_extended_1989}, EDMFT+GW~\cite{Sun02,Ayral13,Huang14}, Dual Boson~\cite{van_loon_beyond_2014} and a slave boson approach \cite{Lhoutellier15,Fresard15}.
Of particular interest is how the renormalized effective interaction depends on the parameters of the Hubbard model. We show that the screening that leads to a renormalization of the local interaction depends strongly on interaction strength and filling.
We use four different computational methods, namely Determinant Quantum Monte Carlo~\cite{blankenbecler_monte_1981} (DQMC), Dual Boson~\cite{rubtsov_dual_2012,van_loon_beyond_2014,stepanov_self-consistent_2016} (DB), Dynamical Mean-Field Theory~\cite{metzner89,georges_dynamical_1996} (DMFT) and the Random Phase Approximation (RPA)~\cite{mahan00}. Two of these, DQMC and DMFT, are restricted to systems with local interaction. This is one of the main motivations for the effective Hubbard model approach: the variational principle allows predictions about the extended Hubbard model using computational methods that are restricted to the Hubbard model.
Since DQMC is numerically exact, it provides the perfect benchmark for the other methods.
DMFT, on the other hand, is an approximation that can be extended to realistic, multi-orbital systems. Coupled with DFT, it forms the workhorse of the strongly correlated materials community.
With the variational principle, non-local interactions can be incorporated into DMFT calculations with relative ease.
RPA and DB allow us to do calculations in the extended Hubbard model. The RPA has only a limited range of validity, since it is a theory for weakly interacting systems. At the same time, computationally it is the simplest of all the theories considered here. DB, on the other hand, is an extension of DMFT that incorporates strong correlation effects. We use it to determine observables in the presence of non-local interaction effects, and to study which observables follow the predictions of the variational principle.
The remainder of this work is structured as follows:
In Sec.~\ref{sec:methods} we give a short overview of the variational principle used to determine the effective local interaction and of the methods used to obtain numerical results. In Sec.~\ref{sec:corrFuncs}, we determine the effective interaction strength in the half-filled Hubbard model. In Sec.~\ref{sec:benchmarkingobservables}, we use DQMC as benchmark for several observables calculated in DB, and in Sec.~\ref{sec:observables} we calculate these observables in the corresponding extended Hubbard model using DB. In Sec.~\ref{sec:alpha:doped} and \ref{sec:observables:doped}, we perform the same analysis for a strongly doped Hubbard model.
\section{Methods}
\label{sec:methods}
The extended Hubbard model on the square lattice with nearest-neighbor hopping reads
\begin{align}
H = -t\sum_{\langle i,j\rangle,\sigma} c_{i\sigma}^\dagger c^{\phantom{\dagger}}_{j\sigma} + U\sum_i n_{i\ensuremath{\uparrow}}n_{i\ensuremath{\downarrow}} + \frac{1}{2} \sum_{\stackrel{i\neq j}{\sigma,\sigma'}}V_{ij} n_{i\sigma}n_{j\sigma'}, \label{eq:exHub}
\end{align}
where $t$ is the nearest-neighbor hopping-matrix element and $\langle i,j\rangle$ is a sum over nearest neighbors. $U$ and $V_{ij}$ are the local and non-local Coulomb matrix elements, respectively. The extended Hubbard model is a particular case of the so-called polar model that has been studied since the 1930's~\cite{Schubin34,Vonsovsky79_1,Vonsovsky79_2}.
We briefly review the main results of the variational method to map extended Hubbard models to \textit{effective} Hubbard models with only local interactions\cite{schuler_optimal_2013}. The effective Hubbard model, reading
\begin{align}
\tilde H = -t\sum_{\langle i,j\rangle,\sigma} c_{i\sigma}^\dagger c_{j\sigma} + \tilde U\sum_i n_{i\ensuremath{\uparrow}}n_{i\ensuremath{\downarrow}}, \label{eq:effHub}
\end{align}
is varied with respect to the effective local interaction $\tilde U$ in order to minimize a free energy functional. This is equivalent to solving
\begin{align}
\tilde U =U-\sum_{j\neq 0}V_{0j} \frac{\partial_{\tilde U}\langle{n_{0}n_{j}}\rangle_{\tilde H}}{\partial_{\tilde U} \langle n_{0}n_{0}\rangle_{\tilde H}} \label{eq:ustar}
\end{align}
for $\tilde U$. For only nearest-neighbor interaction $V$, Eq. \ref{eq:ustar} simplifies to
\begin{align}
\tilde U =U-V \alpha(\tilde U) \label{eq:ustarAlpha},
\end{align}
where we have introduced the nearest-neighbor renormalization strength
\begin{align}
\alpha(\tilde U) = \sum_{\langle 0,j \rangle} \frac{\partial_{\tilde U}\langle{n_{0}n_{j}}\rangle_{\tilde H}}{\partial_{\tilde U} \langle n_{0}n_{0}\rangle_{\tilde H}}. \label{eq:alpha}
\end{align}
Physically, this $\alpha$ describes the effective screening of the local interaction by the non-local interaction effects. $\alpha$ is a function of $\tilde U$, so it can be determined with knowledge of the effective local model alone. There is no assumption on the magnitude of any of the parameters in the original model. The only limit of applicability is that not all physical effects present in the extended Hubbard model can be captured in an effective local interaction, as discussed below.
The variational principle involves the calculation of charge correlation functions of the effective Hubbard model: $\langle n_i n_j \rangle_{\tilde H}$, where $n_i = n_{i\ensuremath{\uparrow}} + n_{i\ensuremath{\downarrow}}$.
In this work, we use several methods to calculate the charge correlation functions: The DQMC~\cite{blankenbecler_monte_1981}, RPA~\cite{mahan00}, DMFT~\cite{metzner89,georges_dynamical_1996} and DB ~\cite{rubtsov_dual_2012,van_loon_beyond_2014} methods. In the remainder of this section, we give a short summary of these methods and associated numerical details. This part can safely be skipped on the first read.
The DQMC method is numerically exact and has been used for the variational principle before~\cite{schuler_optimal_2013}. We use the implementation of the \textsc{quest} code\footnote{``QUantum Electron Simulation Toolbox'' \textsc{quest} 1.3.0 A. Tomas, C-C. Chang, Z-J. Bai, and R. Scalettar, (\url{http://quest.ucdavis.edu/})}. We obtain susceptibilities on Matsubara frequencies by Fourier transforming imaginary-time data. Disadvantages of DQMC are that it cannot be applied to the extended Hubbard model with non-local interaction straightforwardly \cite{zhang_extended_1989,PhysRevLett.68.353,golor_nonlocal_2015} and that it suffers from a sign problem away from half-filled systems. We perform the DQMC calculations for finite $12 \times 12$ (for half-filling) and $8\times 8$ (away from half-filling) systems with Trotter discretisation $\Delta \tau = 0.025$.
The RPA approach~\cite{mahan00} is geared towards weakly-interacting systems.
In a noninteracting system, the susceptibility is given by a ``bubble'' diagram of two Green's functions,
\begin{align}
\chi_0(i\omega_n,\mathbf{q}) =& -\frac{1}{\beta} \sum_{\nu_m} [G^0 G^0]_{\nu\omega \mathbf{q}},\notag \\
[G^0 G^0]_{\nu\omega \mathbf{q}} =& \frac{1}{N} \sum_{\mathbf{k}} G^0(i\nu_m+i\omega_n,\mathbf{k}+\mathbf{q}) G^0(\nu_m,\mathbf{k}),
\end{align}
where $G^0(\nu_m,\mathbf{k})$ is the noninteracting Green's function, $N$ is the number of $\mathbf{k}$ points, and $\nu_m,\omega_n$ are fermionic and bosonic Matsubara frequencies, respectively.
Then, the susceptibility of the (weakly) interacting system is
\begin{align}
\chi^{-1}_\text{RPA}(i\omega_n,\mathbf{q}) = \chi^{-1}_0(i \omega_n,\mathbf{q}) + U+V(\mathbf{q}), \label{eq:RPA}
\end{align}
where $V(\mathbf{q})$ is the Fourier transform of $V_{ij}$ in Eq.~\eqref{eq:exHub}.
Finally, the (equal-time) charge correlation functions are obtained by a Fourier transform to real space and by summing over the Matsubara frequency $\omega_n$:
\begin{align}
\langle n_{0}n_{i}\rangle - \langle n_0\rangle \langle n_i \rangle = \frac{2}{\beta N}\sum_n \sum_\mathbf{q} e^{i\mathbf{q} \mathbf{r}_i}\chi_\text{RPA}(i\omega_n,\mathbf{q}).
\end{align}
Dynamical mean-field theory~\cite{georges_dynamical_1996} is an approximate method that includes \emph{local} correlation effects. The method applies to systems with local interactions only, i.e., $V=0$. The approach is based on a self-consistently determined auxiliary single-site problem.
As in RPA, we obtain the correlation functions from the susceptibility in momentum and frequency space, which is given by
\begin{align}
\chi_{\text{DMFT}}^{-1}(i\omega_n,\mathbf{q}) = -\hat{[\mathcal{G}\mathcal{G}]}(i\omega_n,\mathbf{q})^{-1} - \hat{\Gamma}_{\omega_n}.
\end{align}
Here $\mathcal{G}$ is the DMFT Green's function, which has a local self-energy, $\Gamma$ is the particle-hole irreducible two-particle vertex of the auxiliary single-site problem. $[\mathcal{G}\mathcal{G}]$ is shorthand for the product of two Green's functions. The equation has a matrix structure in fermionic frequencies~\cite{georges_dynamical_1996,PavariniJulich}, as indicated by the hats, which we have suppressed for notational convenience.
The DB method~\cite{rubtsov_dual_2012} is a diagrammatic extension of DMFT that allows for the treatment of non-local interactions directly, via an effective frequency dependent interaction $U(i\omega_n)$. We apply self-consistent DB~\cite{stepanov_self-consistent_2016} in the charge and magnetic ($S^z$) channel to obtain consistent correlation functions~\cite{van_loon_double_2016}. In DB, the expression for the susceptibility is
\begin{align}
\chi_{\text{DB}}^{-1}(i\omega_n,\mathbf{q}) = \chi_{\text{DMFT}}^{-1}(i\omega_n,\mathbf{q}) + U+V(\mathbf{q})- U(i\omega_n), \label{eq:susc:db}
\end{align}
with the important caveat that $\chi_{\text{DMFT}}$ is determined using the DB auxiliary single-site problem, so that the $\mathcal{G}$ and $\Gamma$ that enter this equation are different from the ones in DMFT.
DMFT is recovered when $V(\mathbf{q})=0$ and $U=U(i\omega_n)$.
The DB calculations are performed on a $64\times 64$ lattice. The DB implementation and the CT-HYB~\cite{Werner06} impurity solver are based on the ALPS libraries~\cite{ALPS2,Hafermann13}. The impurity solver takes into account retarded interactions~\cite{Ayral13} and uses improved estimators for the two-particle quantities~\cite{Hafermann14}.
We use converged extended Dynamical-Mean Field Theory (EDMFT, c.f. Appendix~\ref{app:edmft}) calculations as the starting point for the DB self-consistency.
In Appendix~\ref{app:uw}, we show the converged dynamic interaction $U(i\omega_n)$ in both EDMFT and DB.
The number of iterations needed to achieve the self-consistent hybridization and dynamic interaction increases from less than ten to approximately forty between $U/t=7$ and $U/t=8$. In this region, the vertex corrections to the susceptibility are very strong~\cite{van_loon_double_2016}, and $U(i\omega_n)$ converges slowly. This issue makes calculations at higher $U$ very expensive computationally, and for that reason most of the DB calculations are at $U/t<8$.
\section{Effective local interaction at half-filling}
\label{sec:corrFuncs}
Results from all methods presented in this work are obtained at the temperature $\beta t=2$. We discuss the temperature dependence of our results at the end of this section. We use $t=1$ as the unit of energy.
\begin{figure}[b]
\begin{center}
\includegraphics{fig_1_new-crop}
\end{center}
\caption{(Color online) Local (red) and nearest-neighbor (blue) charge correlation functions of half filled nearest-neighbor hopping Hubbard model on a square lattice obtained from DQMC (full line), DB (diamonds), DMFT (crosses), and RPA (dashed line).}
\label{fig:corrFuncHalf}
\end{figure}
To begin with, we discuss the half filled, $\langle n_0\rangle=1.0$, extended Hubbard model with only nearest-neighbor interaction terms. Therefore, the evaluation of Eq. \ref{eq:ustar} involves the calculation of the local and nearest-neighbor charge correlation functions. The results from DQMC, DB, DMFT and RPA are depicted in Fig. \ref{fig:corrFuncHalf} for interaction strengths up to $U/t=10$. As expected, all methods reproduce the non-interacting case exactly. Wick's theorem theorem applies to the non-interacting system, so $\langle n_0 n_0 \rangle=\langle n_0 \rangle + 2\langle n_{0\uparrow} \rangle \langle n_{0\downarrow} \rangle $ and $\langle n_0 n_1 \rangle=\langle n_0 \rangle\langle n_1 \rangle - \sum_{\sigma}\langle c^\dagger_{0\sigma}c_{1\sigma} \rangle \langle c^\dagger_{1\sigma}c_{0\sigma} \rangle $. The DQMC results approach the strong coupling result for $U\gg t$, where every site has one exactly electron and $\langle n_0 n_0 \rangle=\langle n_0 \rangle=1$ and $\langle n_0 n_1 \rangle =\langle n_0 \rangle \langle n_1 \rangle=1$. For intermediate interaction strengths ($U/t \lesssim 5$), results from DQMC, DMFT and DB are virtually indistinguishable on this scale. For larger interactions, differences between the exact DQMC and both the DMFT and the DB approximations are visible. Clearly, the DB method improves the DMFT results. The RPA results are considerably off the DQMC results, especially at larger interaction strengths. We note that the agreement between the methods is considerably better for the nearest-neighbor correlation function than for the local correlation function.
Next, we consider the nearest-neighbor renormalization strength $\alpha(\tilde U)$, depicted in Fig. \ref{fig:alphaHalf} (a), calculated from Eq. \ref{eq:alpha} with the different approximations. The simplest approximation to Eq. \ref{eq:ustar} discussed in Ref. \onlinecite{schuler_optimal_2013} leads to $\tilde U = U - V$, i.e., a constant $\alpha(\tilde U) = 1$.
This approximation can be derived by assuming that the correlation between sites that are not nearest neighbors is zero, i.e., that the system is very strongly localized.
The DQMC result indeed shows that the approximation $\alpha(\tilde U) = 1$ is only valid for $\tilde U\gg t$. In this limit, sites that are more than one lattice spacing apart are uncorrelated, which is sufficient to prove $\alpha=1$~\cite{schuler_optimal_2013}.
In fact, $\alpha(\tilde U)$ has a minimum at intermediate $\tilde U$ before increasing towards 1. DB agrees quite well with the exact DQMC results, with the largest deviations occurring around the minimum of $\alpha$. For $\tilde U/t>5$, DMFT starts to deviate from DQMC. The DMFT results do show a minimum, however, that minimum is located at slightly larger $\tilde U$. For small interactions ($\tilde U/t \lesssim 5$) the RPA follows the behavior found in DQMC. However, RPA does not reproduce the minimum found around $U/t\sim 5$ and is off by a factor of almost 2 at large interactions.
\begin{figure}[htb]
\begin{center}
\includegraphics{fig_2_new_with_temp-crop}
\end{center}
\caption{(Color online) Nearest-neighbor renormalization strength $\alpha(\tilde U)$ of the half-filled nearest-neighbor hopping Hubbard model on a square lattice. (a) $\alpha(\tilde U)$ at $\beta t = 2$, as obtained from DQMC (green dots), DB (red diamonds), DMFT (blue crosses) and RPA (dashed line). (b) At $\beta t=2$ (green dots) and $\beta t = 5$ (magenta circles) using DQMC.}
\label{fig:alphaHalf}
\end{figure}
To study the role of temperature, we have also done DQMC calculations at $\beta t = 5$ instead of $\beta t = 2$. There, $\alpha(\tilde{U})$ is qualitatively similar, as visible in Fig. \ref{fig:alphaHalf}(b). At $\tilde{U}=0$, the nearest-neighbor renormalization strength $\alpha$ is larger, as $\tilde{U}$ increases it goes to a slightly deeper minimum that occurs at smaller $\tilde{U}$, and finally for large $\tilde{U}$ the renormalization strength goes towards 1.
To give a sense of scale, we remind the reader that $\beta t =5$ and $\tilde{U}/t=4$ is in a region where antiferromagnetic fluctuations are strong. This point is close to the metal-insulator transition according to a combination of diagrammatic and Monte Carlo techniques~\cite{Schafer15} and also to the DMFT N\'eel temperature. The (single-site) DMFT Mott metal-insulator transition, on the other hand, occurs at higher interaction strength $\tilde{U}/t=10$ -- $12$.
\section{Benchmarking DB observables}
\label{sec:benchmarkingobservables}
The variational principle only deals with the free energy.
In practical calculations, the main interest often lies with other observables, such as the Green's function and the double occupancy. The question is how well the optimal $\tilde U$ Hubbard model reproduces the observables of the original extended Hubbard model with parameters $U$ and $V$. We want to use DB to calculate the observables of the extended Hubbard model. Before we do that, we study the accuracy of DB at $V=0$, where we can use DQMC as a benchmark.
\begin{figure}
\includegraphics{observables-crop}
\caption{
Observables in the Hubbard model ($V=0$) obtained using DQMC (lines) and the DB method (diamonds).
}
\label{fig:observables2}
\end{figure}
In the previous section, we have seen that the DB results for the local and nearest-neighbor correlation function are accurate. In Fig.~\ref{fig:observables2}, we extend this conclusion to other observables, by comparing them to the DQMC values.
In Fig.~\ref{fig:observables2}(a), we show the imaginary part of the local Green's function at the lowest Matsubara frequency and the double occupancy. The latter is equal to $\av{n_0 n_0}/2 - 1/2$, cf. Fig.~\ref{fig:corrFuncHalf}.
In Fig.~\ref{fig:observables2}(b), we show some zero (Matsubara) frequency spin susceptibilities. These provide insight into the response of the system to static external fields.
They are natural observables for DB, since we calculate the entire momentum and frequency dependent susceptibility according to Eq.~\eqref{eq:susc:db}.
In DQMC, the susceptibility is determined as a function of imaginary time and we obtain the static component by a Fourier transform.
In the figure, we show the $q=(\pi,\pi)$ (checkerboard), $q=(0,0)$ (uniform) and $q$-averaged (local) spin susceptibility.
In Fig.~\ref{fig:observables2}(c), we do the same for the charge susceptibility.
For all the observables shown, we find a good qualitative and quantitative agreement between DQMC and DB. Deviations start to set in at interaction strengths $\tilde{U}/t\gtrsim 6$.
Fig.~\ref{fig:observables2} shows that the local repulsion $\tilde{U}$ suppresses the double occupancy and charge excitations in general, whereas spin excitations are enhanced. The checkerboard spin susceptibility, corresponding to antiferromagnetism, increases the most. We should note that the temperature studied here, $\beta t = 2$, is above the Mott transition temperature, and all observables depend smoothly on $\tilde{U}$.
\section{Observables at finite $V$}
\label{sec:observables}
\begin{figure*}
\includegraphics[]{fig_3}
\caption{
Observables in the extended Hubbard model obtained using the DB method. Every colored square represents a DB calculation, they are separated by $\Delta U=0.1$ and $\Delta V=0.1$. The solid lines show isolines where the observable is constant, the dashed lines show lines of constant $\tilde U$ according to DQMC.
The constant values of the isolines correspond to the tick labels in the colorbars.
}
\label{fig:observables}
\end{figure*}
Now that we have confidence in the predictions of DB, we can use them as a benchmark for the variational principle at finite values of $V$.
To study this, we have done DB calculations for $1.0 \leq U/t \leq 2.2$ and $-0.3 \leq V/t \leq 0.3$, the results of which are shown in Fig.~\ref{fig:observables}. The colored plots give the value of the observable according to DB, with each colored square corresponding to a specific value of $U$ and $V$. The solid lines are approximate isolines in the $(U,V)$-plane along which the observable is constant. The dashed lines, on the other hand, indicate constant $\tilde U$. If the variational principle were exact, these would be the isolines of the observables.
Fig.~\ref{fig:observables}(a) contains the value of the Green's function on the first Matsubara frequency $\pi/\beta$, $-\Im g_{\nu=\pi/\beta}$. The DB observables does show a roughly linear dependence on $V$, albeit with a different slope than the variational principle predicts.
Next, in Fig.~\ref{fig:observables}(b), for the double occupancy, there is a very good match between the DB calculations and the variational scheme. This correspondence is not accidental, the variational principle gives the exact double occupancy to first order in $V$, as is shown in Appendix \ref{app:smallV}. The double occupancy is a special operator in this context, since it is directly connected to the variational parameter $\tilde{U}$.
This explains the matching tangents at $V=0$ in Fig. 4(b). For larger $\left|V\right|$, the curvature in the DB results shows a divergence from the simple $\tilde U = U-\alpha V$ prescription. As $V$ becomes large compared to the Hubbard parameters $\tilde U$ and $t$, it is no longer reasonable to expect the effective Hubbard model to do a good job in describing the relevant physics.
In Fig.~\ref{fig:observables}(c), (e) and (g), we move to the zero-frequency spin susceptibilities, namely the checkerboard [$q=(\pi,\pi)$], uniform [$q=(0,0)$] and local part [$q$-average] of the spin susceptibility, respectively. All three show a reasonable, though not perfect, match between the variational principle prediction and the DB results.
The corresponding correlation functions in the charge sector, in Fig.~\ref{fig:observables}(d), (f) and (h), show a very different dependence on $V$. The checkerboard correlation function does have a linear dependence on $V$, with very small slope. In the uniform charge susceptibility, the sign of the $V$-dependence has changed, and for the local susceptibility the dependence is even quadratic instead of linear.
This poor match is not a surprise. The non-local interaction $V$ directly and explicitly enters the charge dynamics, as in Eq.~\eqref{eq:susc:db}, and the effective local interaction can only give a poor description of that dependence.
As an alternative to DB, it is also possible to use EDMFT to do calculations at finite $V$. In Appendix~\ref{app:edmft}, we show the resulting observables. EDMFT has the advantage of being simpler, however the approximate treatment of the momentum structure of correlation functions leads to quadratic instead of linear scaling in $V$, and a poor description of the spin susceptibility, as shown in the Appendix.
\begin{figure}[htb]
\begin{center}
\includegraphics{fig_3_new-crop}
\end{center}
\caption{(Color online) Local (red) and nearest-neighbor (blue) charge correlation functions of the hole doped ($\langle n_0\rangle=0.18$) nearest-neighbor hopping Hubbard model on a square lattice obtained from DQMC (full line), DB (diamonds), DMFT (crosses), and RPA (dashed line). }
\label{fig:corrFuncDope}
\end{figure}
\section{Effective local interaction away from half-filling}
\label{sec:alpha:doped}
In order to study the performance of the different approximations as well as the renormalizations in dependence of the filling, we study the same Hubbard model as above with a filling of $\langle n_0\rangle=0.18$, well below the optimal filling for high-$T_c$ superconductors, of $\langle n_0 \rangle \approx 0.8$. At this small filling, the DQMC sign problem is not very severe and computations are feasible. In addition, we restrict ourselves to intermediate interaction strengths.
We follow the same approach as before and start by determining the charge correlation functions.
These are shown in Fig. \ref{fig:corrFuncDope}. We come to very similar conclusions as in the case of half filling. DQMC, DB, and DMFT results agree closely in the investigated regime. RPA agrees well for the nearest-neighbor correlation function, and poorly for the local correlator.
DQMC, DMFT and DB all operate in the grand-canonical ensemble, at fixed chemical potential $\mu$. The results at fixed density are obtained by interpolating between simulations at fixed chemical potential. This interpolation step introduces additional uncertainty into the determination of the correlators. This is especially visible when taking the numerical derivative of the correlation functions to obtain $\alpha$, since the difference quotient is very susceptible to noise. To estimate $\alpha$ for DB, we used a linear fit through all the data points in Fig.~\ref{fig:corrFuncDope}, and then we used the linear coefficients of these fits to determine $\alpha$. The resulting $\alpha$ is shown Fig.~\ref{fig:alphaDope}. To obtain an error estimate, a quadratic fit of the data in Fig.~\ref{fig:corrFuncDope} was done, this results in $\tilde{U}$-dependent derivatives and the spread in the derivates was used to determine the error bar. DMFT (not shown) gives a result within the (rather large) error bars of DB.
The first thing that is clear is that $\alpha \neq 1$, so the simple formula $\tilde{U}=U-V$ does not hold.
Interestingly, for this hole doped case, the nearest-neighbor renormalization strength is even negative, i.e., the nonlocal interaction increases the effective local interaction, $\alpha<0$. This is in line with findings in the context of doped benzene models in Ref. \onlinecite{schuler_optimal_2013} and can be understood in terms of Wigner crystallization~\cite{wigner34}.
In a very empty system, the local interaction $\tilde{U}$ suppresses not only the probability to find a second electron on the same site, but also in the vicinity of the first electron.
The effect of the non-local interaction $V$ is also to keep electrons away from each other, so a positive $V$ leads to a larger effective $\tilde{U}$.
The DQMC results show that this renormalization increases for larger interaction. While RPA gives the correct sign, it underestimates $|\alpha|$ at all finite interaction strengths and it predicts a decreasing renormalization for growing interaction, which is the wrong trend.
\begin{figure}[htb]
\begin{center}
\includegraphics{fig_4_new-crop}
\end{center}
\caption{(Color online) Nearest-neighbor interaction strength $\alpha(\tilde U)$ of the hole doped ($\langle n_0\rangle=0.18$) nearest-neighbor hopping Hubbard model on a square lattice obtained from DQMC (green dots), DB (red error bar), and RPA (dashed line). }
\label{fig:alphaDope}
\end{figure}
\section{Observables in doped systems}
\label{sec:observables:doped}
Finally, we also study the observables of the doped system.
As before, we start by comparing DQMC and DB observables at $V=0$, this is shown in Fig.~\ref{fig:doped:observables2}. We again find a good match between the DQMC and DB results. Secondly, we show the observables at finite $V$ in Fig.~\ref{fig:observables_doped}. As before, the isolines predicted by DQMC are shown as dashed lines. Here, however, they do not match at all with the observables from DB.
\begin{figure}
\includegraphics{observables-doped-crop}
\caption{
Observables in the Hubbard model away from half-filling ($\av{n_0}$ and $V=0$) obtained using DQMC (lines) and the DB method (diamonds), cf. Fig.~\ref{fig:observables2} for the half-filled system.
}
\label{fig:doped:observables2}
\end{figure}
\begin{figure*}
\includegraphics[]{fig_3-doped}
\caption{
Observables in the extended Hubbard model, away from half-filling at $\av{n_0}=0.18$, obtained using the DB method. See Fig.~\ref{fig:observables_doped} for the half-filled system.
}
\label{fig:observables_doped}
\end{figure*}
Physically, the strongly doped $\av{n_0}=0.18$ system is very different from the half-filled Hubbard model.
The local interaction $U$ only affects electron pairs that occupy the same site, and as a result, an observable like $g_{\nu=\pi/\beta}$ depends only very weakly on $U$, as seen in Fig.~\ref{fig:doped:observables2}. This weak dependence holds even at higher values of $\tilde{U}/t$ (not shown).
As mentioned above, the local and non-local interaction creates precursors to Wigner crystallization~\cite{wigner34}.
Even in the non-interacting system, the probability to find two electrons at the same site is $\av{n_0}^2/4 < 1\%$.
As such, it is very difficult to encapsulate the effect of $V$ into an effective $\tilde{U}$. Since the physics of the strongly-doped extended Hubbard model is not Hubbard-like, the effective mapping is not able to provide relevant results.
The susceptibility provides another clear difference between Figs.~\ref{fig:observables} and \ref{fig:observables_doped}. In the half-filled model, perfect nesting of the Fermi surface with the checkerboard wavevector $q=(\pi,\pi)$ creates a tendency towards checkerboard ordering (antiferromagnetic ordering in the spin channel) at lower temperatures. This is visible in the large values of the checkerboard susceptibilities in Fig.~\ref{fig:observables}. At the much lower density $\av{n_0}=0.18$, checkerboard ordering is not favored, and the checkerboard susceptibility is smaller than the uniform and local susceptibilities, and it depends only weakly on the interaction strength.
DMFT and DB both use a single-site auxiliary problem as the starting point of their approach. In this way, they are able to incorporate strong local correlation effects. Local correlation requires two particles at the same site, so this is expected to be less important at strong doping.
Based on this, we can expect the non-local correlations that DB includes on top of DMFT to become more important at strong doping~
\cite{van_loon_thermodynamic_2015,van_loon_double_2016}, and we can also expect that the local correlation effects included in DB do not significantly improve on simpler theories like the $GW$-method.
This almost empty system clearly requires an explicit treatment of the nonlocal interactions. The extended Hubbard model is physically very different from the purely local Hubbard model, so any attempt to use the variational principle to relate the two is ill-fated.
\section{Conclusions and Discussion}
We have studied the mapping of the extended Hubbard model onto effective local Hubbard models using a variational principle. In the half-filled Hubbard model, the simple prescription $\tilde{U} = U - V$ is only applicable at very high values of $U$. At intermediate $U$, the effective renormalization of the local interaction $\alpha$, with $\tilde{U} = U - \alpha V$, is reduced by as much as a factor of 2.
To determine the effective interaction, the local and nearest-neighbor correlation function of the Hubbard model are needed.
We find that the self-consistent DB approximation accurately reproduces the numerically exact DQMC results for the correlators, so that even the numerical derivatives come out similarly.
DMFT performs qualitatively correct but quantitatively slightly worse at larger interaction strengths $\tilde{U}/t>5$, and misses the location of the minimum of $\alpha$.
RPA obtains the correct nearest-neighbor renormalization strength at small interaction, however it does not have a minimum and fails in the limit of large interaction.
For the nearly empty system (i.e. heavy hole doping, $\av{n_0}=0.18$), we find that non-local repulsion actually predicts a larger effective local interaction. Numerical calculations are more difficult in this parameter regime, which makes it difficult to assess exactly how well DMFT and DB match with the DQMC results.
We have also studied how observables of the extended Hubbard model behave as a function of $V$, again using the DB approach. The effective Hubbard model predictions work well for most observables at half-filling. For charge correlation functions, though, the effective Hubbard model does not match with the finite $V$ results. This was to be expected, since the charge correlations depend explicitly on $V$.
Away from half-filling, the match between observables in DB and in the effective Hubbard model is much worse. This parameter regime is dominated by Wigner crystallization physics, which is difficult to capture using the variational principle.
The contrasting behavior in these two scenarios teaches us that the mapping to an optimal local Hubbard model only has a chance to succeed when the physics of the system is essentially Hubbard-like. In the very empty system, where doubly occupied sites are rare, only changing the effective Hubbard parameter is insufficient to recover the Wigner localization physics. Similarly, the charge susceptibility in the half-filled system, which is directly driven by the non-local interaction, is not captured in the effective model.
On the other hand, the extended Hubbard model at half-filling is sufficiently similar to the local Hubbard model that a renormalization of the interaction strength suffices to explain the Green's function, double occupancy and spin susceptibility.
\acknowledgments
E.G.C.P. v. L. and M.I.K. acknowledge support from ERC Advanced Grant 338957 FEMTO/NANO. M.S. and T.O.W. acknowledge support from the Deutsche Forschungsgemeinschaft through the research unit 1346 and the University of Bremen through the Zentrale Forschungsf{\"o}rderung.
|
\section{INTRODUCTION}
$\delta$ Sct stars are main sequence and subgiant stars with spectral types from about A2 to F2 (Rodr\'iguez \& Breger 2001).
They pulsate in low-order radial/non-radial pressure ($p$) modes with short periods of 0.02$-$0.2 d (Breger 2000). The pulsations are
driven by the $\kappa$ mechanism acting in the He II partial ionization region. Recently, Balona (2014) showed that nearly all
$\delta$ Sct stars pulsate in the frequency range of 0$-$5 d$^{-1}$ characteristic of $\gamma$ Dor stars and suggested that
the low frequencies are directly associated with the presence of $\delta$ Sct pulsations. The $\delta$ Sct and $\gamma$ Dor variables
share a similar parameter space in the Hertzsprung-Russell (HR) diagram which is partly overlapped, but the latter stars pulsate
in high-order non-radial gravity ($g$) modes driven by a mechanism known as convective blocking (Guzik et al. 2000) with relatively
longer periods of 0.4$-$3 d (Handler \& Shobbrook 2002; Henry, Fekel \& Henry 2005). The hybrid stars that contain the $p$ and
$g$ modes simultaneously are of particular interest, since they probe both the envelope and the deep interior near the core region
of the pulsators (Kurtz et al. 2015).
Over 90 pulsating stars have been known as the $\delta$ Sct, $\gamma$ Dor, and hybrid stars in eclipsing binaries (EBs).
Approximately 73 of them are the so-called oEA (oscillating eclipsing Algol) stars; the mass-accreting components of classical
semi-detached Algols that lie inside the instability strip and show $\delta$ Sct-like oscillations (Mkrtichian et al. 2004).
The $\delta$ Sct stars in EBs exhibit almost the same pulsating characteristics as single $\delta$ Sct pulsators, but
their evolutionary processes are entirely different from each other by tidal interaction and mass transfer between components.
For the pulsating EBs, a possible relation between the binary orbital periods ($P_{\rm orb}$) and the dominant pulsation periods
($P_{\rm pul}$) was firstly given by Soydugan et al. (2006) and updated by Liakos et al. (2012):
$P_{\rm pul}$ = 0.020$P_{\rm orb}$ $-$ 0.005 for 20 EBs and $\log P_{\rm pul}$ = 0.58$P_{\rm orb}$ $-$ 1.53 for 70 EBs in
the same order. Later, the $P_{\rm pul}-P_{\rm orb}$ relation was theoretically established by Zhang, Luo \& Fu (2013), in which
the pulsation period could be described as a function of the orbital period, the pulsation constant, the mass ratio, and
the filling factor. According to those studies, the eclipsing $\delta$ Sct stars pulsate with shorter pulsational periods than
single $\delta$ Sct pulsators. The pulsational period seems to depend on the gravitational force exerted by a companion onto
the pulsating star. This indicates that the companion star may influence the pulsation frequencies and modes of the pulsating component.
The pulsating stars in EBs offer a unique opportunity to study the effects of tidal forces and companions on the pulsations
(Hambleton et al. 2013), as well as to measure directly the masses and radii of the pulsators from binary modelling.
The EBs with pulsating stars are the Rosetta stone for the study of stellar structure and evolution through asteroseismology
and binary properties. Such examples are KIC 10661783 (Southworth et al. 2011; Lehmann et al. 2013),
KIC 4544587 (Hambleton et al. 2013), and KIC 3858884 (Maceroni et al. 2014). Because space missions such as {\it Kepler} and
{\it CoRot} provide ultra-precise photometric data, they allow the detection of many pulsation frequencies with amplitudes down
to the micromagnitude level. In order to look for pulsating components in EBs and to understand their physical properties,
we choose the {\it Kepler} target KIC 6220497 (R.A.$_{2000}$ = 19$^{\rm h}$44$^{\rm m}$39$\fs547$; decl.$_{2000}$ = +41$^{\circ}$33${\rm '}$21$\farcs$17;
$K_{\rm p}$ = $+$14.749; $g$=$+$14.935; $g-r$=$+$0.247), which was announced to be probably a pulsating EB with an orbital period
of 1.323 d by Gaulme \& Guzik (2014). This paper is the third contribution in a series assessing the detections and properties of
pulsating stars in the {\it Kepler} EBs (Lee et al. 2014, 2016). Here, we present the binary system as an oEA star showing
multiperiodic $\delta$ Sct pulsations, based on light-curve synthesis and frequency analysis for the light residuals from
the binary model.
\section{{\it KEPLER} PHOTOMETRY AND LIGHT-CURVE SYNTHESIS}
KIC 6220497 was observed during Quarters 14 and 15 in the long cadence mode, which has a sampling time of 29.4 min. We used
the data in the {\it Kepler} EB catalogue\footnote{http://keplerebs.villanova.edu/} detrended and normalised from the raw SAP
(Simple Aperture Photometry) time series (Pr\v sa et al. 2011; Slawson et al. 2011). The contamination level of the measurements
is estimated to be 0.046. This value suggests that the {\it Kepler} target suffers minimally from third light, if any. Figure 1
depicts the {\it Kepler} light curve for KIC 6220497, where there are no significant trends in the eclipse depths and
the light maxima (Max I and Max II) present equal light levels. As shown in the phase-folded light curve, the binary star
displays a considerable variation outside eclipses due to tidal distortions. The depth difference between the primary and
secondary eclipses indicates a large temperature difference between the component stars.
For the light-curve synthesis of KIC 6220497, we used the 2007 version of the Wilson-Devinney synthesis code
(Wilson \& Devinney 1971, van Hamme \& Wilson 2007; hereafter W-D). The {\it Kepler} data of this system were analysed in
a manner almost identical to that for the pulsating EBs V404 Lyr (KIC 3228863; Lee et al. 2014) and KIC 4739791
(Lee et al. 2016). The effective temperature ($T_1$) of the brighter, and presumably more massive, star was initialised to be
7,254 K from the {\it Kepler} Input Catalogue (KIC; Kepler Mission Team 2009). The logarithmic bolometric ($X_{1,2}$)
and monochromatic ($x_{1,2}$) limb-darkening coefficients were interpolated from the values of van Hamme (1993).
The gravity-darkening exponents were fixed at standard values of $g_1$=1.0 and $g_2$=0.32, while the bolometric albedos at
$A_1$=1.0 and $A_2$=0.5, as surmised from the components' temperatures. Adjustable parameters were the orbital ephemeris
($T_0$ and $P$), the mass ratio ($q$), the orbital inclination ($i$), the effective temperatures ($T_{1,2}$) and
the dimensionless surface potentials ($\Omega_{1,2}$) of the components, and the monochromatic luminosity ($L_{1}$).
This synthesis was repeated until the correction of each adjustable parameter became smaller than its standard deviation using
the differential correction (DC) programme of the W-D code. In this paper, the subscripts 1 and 2 refer to the primary and
secondary components being eclipsed at Min I and Min II, respectively.
The mass ratio ($q$ = $M_2$/$M_1$) is a very important parameter in the light-curve synthesis. However, there exist neither
photometric solutions nor spectroscopic orbits for KIC 6220497. Because the binary star is a faint object with a short orbital
period, 8$-$10 m class telescopes are needed to measure its radial velocities. However, it is very difficult to obtain
such observation times. Thus, we conducted an extensive $q$-search procedure, meaning that we calculated a series of
models with the mass ratios in step of 0.01 between 0.1 and 1.0. For each assumed mass ratio, the DC programme was applied
for various modes but showed acceptable photometric solutions only for semi-detached mode 5 in which the secondary component
fills its inner Roche lobe. As displayed in Figure 2, the $q$ searches indicate that the minimum value of the weighted sum of
the squared residuals ($\sum{W(O-C)^2}$; hereafter $\sum$) is around $q$ = 0.24. In the subsequent calculations, this $q$ value
was treated as an adjustable parameter. The result is listed in the second and third columns of Table 1 and appears as
a solid curve in the right panel of Figure 1. The light residuals from Model 1 are plotted as the grey circles in the lower panel
of Figure 3, wherein it can be seen that the model light curves describe the {\it Kepler} data satisfactorily.
The photometric solutions for Model 1 could be affected by the multiperiodic pulsations of the primary component, which will be
discussed in the following section. We removed the pulsation signatures from the original {\it Kepler} data, leaving
behind the light variations due to binarity effects. Then, the new light curve was solved by using Model 1 as initial values.
The final result is illustrated in Figure 3, and is given as Model 2 in the fourth and fifth columns of Table 1. As listed in
the table, the binary parameters for Model 2 are in good agreement with those for Model 1. We can see that the photometric
solutions of KIC 6220497 are immune from the light variation due to the pulsations. Our light-curve synthesis represents
KIC 6220497 as a semi-detached eclipsing system in which the primary component fills its limiting lobe by
$\Omega_1$/$\Omega_{\rm in}$ = 87\%, where $\Omega_{\rm in}$ is the potential for the inner critical surface. The Roche-geometry
configuration of the system permits some mass transfer from the lobe-filling secondary to the detached primary component.
In all the procedures that have been described, we included the orbital eccentricity ($e$) as a free parameter but found that
the parameter remained indistinguishable from zero within its error. This indicates that KIC 6220497 is in a circular orbit,
as expected for semi-detached classical Algols.
It is known that the stand errors of the adjustable parameters taken from the W-D code are unrealistically small because
of the strong correlations between relatively many parameters and partly the non-normal distribution of measurement errors
(Maceroni \& Rucinski 1997). In order to obtain more reliable errors for the Model 1 and Model 2 parameters, we followed
the procedure described by Koo et al. (2014). First of all, we divided the observed (or prewhitened) {\it Kepler} data of
KIC 6220497 into 138 segments at the interval of an orbital period and separately analysed them with the W-D code. Then,
we computed the standard deviations of each parameter from the 138 different values. The parameter errors presented in Table 1
are the 1$\sigma$-values adopted from this procedure.
The temperatures of the primary component in Table 1 correspond to a normal main-sequence star with a spectral type of A9. Because
the temperature errors are certainly underestimated, it was assumed that the temperature of each component had an error of 200 K.
Using the correlations between spectral type and stellar mass (Harmanec 1988), we estimated the primary's mass to be
$M_1$=1.60$\pm$0.08$M_\odot$. The absolute dimensions for KIC 6220497 can be computed from our photometric solutions and $M_1$.
These are given in the bottom of Table 1, where the radii are the mean volume radii calculated from the tables of Mochnacki (1984).
The luminosities and the bolometric magnitudes were computed by adopting $T_{\rm eff}$$_\odot$=5,780 K and $M_{\rm bol}$$_\odot$=+4.73
for solar values.
\section{LIGHT RESIDUALS AND PULSATIONAL CHARACTERISTICS}
In order to obtain more reliable frequencies, the observed {\it Kepler} data were split into 138 subsets as before and
modeled individually with the W-D code through adjusting only the ephemeris epoch ($T_0$) in the Model 1 of Table 1. The light
residuals from the analyses are displayed in Figure 4 as magnitudes versus BJDs, wherein the lower panel presents a short section
of the residuals. Light variations with a total peak-to-peak amplitude of $\sim$10 mmag are clearly seen in the residuals.
In order to examine the oscillating features, we applied a multiple frequency analysis to the whole data sets at once.
The PERIOD04 programme by Lenz \& Breger (2005) was performed on the frequency range from 0 to the Nyquist limit of 24.47 d$^{-1}$.
Because the primary component lies within the $\delta$ Sct instability strip of the HR diagram, it would be a candidate for
such pulsations. During primary eclipses, the secondary component partially blocks the lights of the pulsating primary star, which
can have an effect on the observed frequencies and amplitudes. In contrast, the light contribution from the secondary is only 2.2 \%
to the total luminosity of the binary system, so the pulsations are almost unaffected during the secondary eclipses. Thus, we made
use of only the light residuals having orbital phases between 0.12 and 0.88 after eliminating the data of the primary eclipses.
The amplitude spectra for KIC 6220497 are shown in the top panel of Figure 5. After the successive prewhitening of each frequency
peak, we detected 33 frequencies with the signal to noise amplitude (S/N) ratios larger than 4.0 (Breger et al. 1993). At each step
of this procedure, a multiperiodic least-squares fit to the light residuals was carried out using the fitting formula of
$Z$ = $Z_0$ + $\Sigma _{i}$ $A_i \sin$(2$\pi f_i t + \phi _i$). Here, $Z$ and $Z_0$ denote the calculated magnitude and zero point,
respectively, $A_i$ and $\phi _i$ are the amplitude and phase of the $i$th frequency, respectively, and $t$ is the time of
each measurement. The amplitude spectra after prewhitening the first seven frequencies and then all 33 frequencies are presented
in the middle and bottom panels of Figure 5, respectively. The results are presented in Table 2, where the uncertainties were
calculated according to Kallinger, Reegen \& Weiss (2008). The synthetic curve obtained from the 33-frequency fit is displayed in
the lower panel of Figure 4. Some additional peaks still exist in the bottom panel of Figure 5, but their S/N ratios are smaller
than the empirical threshold of 4.0.
As in the case of KIC 4739791 (Lee et al. 2016), to see if the main frequencies detected in this paper are real and stable during
the observing run of $\sim$ 200 d, we re-analysed the light residuals at intervals of about 50 d and examined the frequency
variations with time. The four subsets resulted in slight differences from each other and approximate 20 frequencies were
detected at each subset with the same criterion of S/N$>$4.0. Figure 6 showed the stability of the 14 frequencies. Among these,
the four frequencies ($f_9$, $f_{10}$, $f_{14}$, $f_{21}$) varied significantly and the three frequencies decreased ($f_3$, $f_{16}$)
or increased ($f_{17}$) monotonically. Within the frequency resolution of 0.008 d$^{-1}$ (Loumos \& Deeming 1978), we checked
the frequencies for possible harmonic and combination terms. The result is listed in the last column of Table 2, where the six
($f_6$, $f_8$, $f_{11}$, $f_{20}$, $f_{24}$, $f_{31}$) frequencies are the orbital frequency ($f_{\rm orb}$ = 0.75576 d$^{-1}$)
and its harmonics.
\section{DISCUSSION AND CONCLUSIONS}
In this paper, we studied the physical properties of KIC 6220497 from detailed analyses of the {\it Kepler} time-series data
obtained during Quarters 14 and 15. The ${\it Kepler}$ light curve was satisfactorily modelled for two cases: including and
removing pulsations. The binary parameters between them are consistent with each other, which indicates that the photometric
analysis of KIC 6220497 is almost unaffected by pulsations. The light-curve synthesis presented in this paper demonstrates that
the system is a semi-detached EB with parameters of $q$ = 0.243$\pm$0.001, $i$ = 77.3$\pm$0.3 deg, and
$\Delta T$ = 3,372$\pm$58 K. The detached primary component is about 1.6 times larger than the lobe-filling secondary and
fills its inner critical lobe by about 87\%, which is one of the largest filling factors for pulsating EBs. In Figure 7,
a comparison of the KIC 6220497 parameters with the mass-radius, mass-luminosity, and HR diagrams shows that the primary star
lies in the main-sequence band, while the secondary is highly evolved and its radius and luminosity are more than four times
oversized and about 26 times overluminous, respectively, compared with dwarf stars of the same mass. In these diagrams,
the locations of both components conform to the general pattern of semi-detached Algols (\. Ibano\v{g}lu et al. 2006).
For the detection of the pulsation frequencies in KIC 6220497, we removed the binarity effects from the observed {\it Kepler} data
and performed a frequency analysis in the entire light residuals, excluding the data around the primary minima. As a consequence,
we detected the 33 frequencies including a dominant oscillation found at $f_{1}$ = 8.51752$\pm$0.00003 d$^{-1}$, corresponding to
0.1174051$\pm$0.0000004 d. Among these, the four ($f_1$, $f_2$, $f_5$, $f_7$) frequencies in the $p$-mode region were highly stable
during the observing runs of about 200 d. On the contrary, the other frequencies varied with time and/or may be orbital harmonics
and combination terms. The frequency ratio of $f_1$ and $f_5$ is 0.984, very close to 1.0, and these frequencies could not be
interpreted by two radial modes of $\delta$ Sct stars (Breger 1979). We checked the ratios of the other frequencies and found
one candidate to be identified as radial modes, i.e. the ratio of 0.848 between $f_2$ and $f_7$ is nearly the same with
the period ratio of 0.845 for the second (2H) and third (3H) overtone radial modes of the $\delta$ Sct stars. It means that
the longer period $f_7$ could be classified as the second overtone radial mode and the shorter period $f_2$ as the third overtone
radial mode. Applying the physical parameters of the primary component in Table 1 to the equation of
$\log Q_i = -\log f_i + 0.5 \log g + 0.1M_{\rm bol} + \log T_{\rm eff} - 6.456$ (Breger 2000), we obtained the pulsation constants
for the four frequencies ($f_{1,2,5,7}$) to be $Q_1$ = 0.033 d, $Q_2$ = 0.016 d, $Q_5$ = 0.033 d, and $Q_7$ = 0.019 d. We compared
these $Q$ values with the theoretical models with 1.5 $M_\odot$ given by Fitch (1981). The $f_1$ and $f_5$ frequencies could be
identified as fundamental ($n = 0$) modes and the $f_2$ and $f_7$ as the third ($n = 3$) and second ($n = 2$) overtone $p$-modes,
respectively. The ratios, $P_{\rm pul}/P_{\rm orb}$ = 0.042$\sim$0.089, of the pulsational to orbital periods are within
the upper limit of $P_{\rm pul}/P_{\rm orb}$ = 0.09$\pm$0.02 for $\delta$ Sct stars in EBs that could be used to distinguish
approximately whether a binary component pulsates in the $p$-mode (Zhang, Luo \& Fu 2013). These results indicate that
the primary star with a spectral type of A9V can be classified as a $\delta$ Sct variable. Because the system is in
a semi-detached configuration, these results imply that KIC 6220497 would be an oEA star and a mass transfer from the evolved
lobe-filling secondary to the detached primary component could be responsible for the $\delta$ Sct-type pulsations detected in
this paper.
The oEA stars have pulsation features similar to classical $\delta$ Sct stars. However, their pulsations may be influenced
by the tidal interaction and mass transfer between the components, as well as gravitational force from companions. Some pulsations
in KIC 6220497 can be excited by the tidal forces of the secondary companion. Tidally excited modes occur when the orbital
frequency is close to a stellar eigenfrequency in a binary star with an eccentric orbit. The signature of the pulsation modes
is the frequencies at integer multiples of the orbital frequency (Welsh et al. 2011; Thompson et al. 2012; Hambleton et al. 2013).
We detected four ($f_6$, $f_{11}$, $f_{20}$, $f_{24}$) frequencies that are the harmonics of the orbital frequency, which could be
partly affected by imperfect removal of the eclipses from the light curve. Although the orbit of the binary system is circular,
these frequencies could result from tidally induced pulsations (Reyniers \& Smeyers 2003a,b; Southworth et al. 2011).
In the upper panel of Figure 8, we plot the dominant pulsation period versus the orbital period for 74 oEA stars, including
KIC 6220497. The data are taken from the compilations of Zhang, Luo \& Fu (2013; 67 oEA stars) and from more recent literature
(Yang, Wei \& Li 2014 for FR Ori; Zhang et al. 2014 for OO Dra; Zhang, Luo \& Wang 2015 for EW Boo; Zhang et al. 2015 for V392 Ori;
Lee et al. 2016 for KIC 4739791; Soydugan et al. 2016 for XZ Aql). In the panel, we can see that the pulsation period of KIC 6220497
deviated from the general trend of the oEA stars and also the empirical relation between $P_{\rm pul}$ and $P_{\rm orb}$ of
Zhang, Luo \& Fu (2013). The pulsation periods in EBs increase with decreasing gravitational pull exerted by the secondary
companion to the pulsating primary component (Soydugan et al. 2006). For KIC 6220497, the gravitational pull applied to per gram
of the matter on the surface of the pulsating component by the lobe-filling secondary was calculated to be $\log F$ = 2.90 in
cgs units, which is about 3.4 times smaller than the value of 9.89 (again, in cgs units) taken from the relation of
$\log P_{\rm pul} = -0.61 \log F +$ 5.1 calibrated by Soydugan et al. (2006). However, as shown in the lower panels of Figure 8
from 34 oEA stars currently known, the gravitational force $\log F$ for KIC 6220497 matches well with those for the other oEA stars
with similar orbital periods, while the surface gravity $\log g_1$ = 3.78 for the primary component is clearly smaller.
The period-gravity relation is similar to that for radially pulsating stars suggested by Fernie (1995); as the surface gravity
decreases, its pulsation period increases. Further, it is known that the more evolved the star, the slower the pulsations
(cf. Liakos et al. 2012). The pulsation period and the surface gravity of KIC 6220497 indicate that the system might be a more
evolved EB than the other oEA stars. We think that the pulsation periods strongly depend on the surface gravities of the pulsating
components and the evolutionary status of the binary stars.
When the double-lined radial velocities and the multiband light curves are made (e.g., Hong et al. 2015, Koo et al. 2016),
they will help to understand the absolute parameters, identification of pulsation modes, and evolutionary status of KIC 6220497
better than now. These offer us important information for asteroseismology and the study of stellar interior and evolution.
\section*{Acknowledgments}
We appreciate the careful reading and valuable comments of the anonymous referee. This paper includes data collected by
the {\it Kepler} mission. {\it Kepler} was selected as the 10th mission of the Discovery Program. Funding for the {\it Kepler}
mission is provided by the NASA Science Mission directorate. We have used the Simbad database maintained at CDS, Strasbourg,
France. This work was supported by the KASI (Korea Astronomy and Space Science Institute) grant 2016-1-832-01.
|
\section{Introduction}
\juno{} is a Liquid Scintillator Antineutrino Detector (LAND) currently under construction in the south of China (Jiangmen city, Guangdong province).
Once completed, it will be the largest LAND ever built, consisting in a 20~kt target mass made of Linear Alkyl-Benzene (LAB) liquid scintillator (LS),
monitored by roughly 18000 twenty-inch high-QE photomultipliers (PMTs) providing a \textasciitilde80\% photo-coverage.
Large photo-coverage and large QE are two pivotal parameters of the experiment, which allow an unprecedented 3\% energy resolution at 1~MeV.
The current conceptual design report~\cite{CDR} foresees the LS to be contained in an acrylic sphere 12~cm thick and 36~m wide.
The whole detector is immersed in a cylindrical water pool, acting both as a moderator for the environmental radioactivity, and as
a cherenkov detector to tag and veto cosmic muons.
The ultimate control and minimal impact of calorimetry systematics is of maximal importance
to achieve the aforementioned energy resolution.
For this reason, a novel LAND design was introduced, where a
second layer of small PMTs is used to provide a second calorimetry handle with complementary systematic budget,
allowing a combined, more precise and accurate energy scale definition.
This calorimetry redundancy system is still under optimisation considerations in the context of \juno{} physics.
\textsc{Juno}'s main physics goal is to determine the neutrino mass hierarchy by detecting reactor $\overline{\nu}_e$ coming from
two nuclear power plants both 53~km distant from the detector,
but here we focus on \textsc{Juno}'s physics programme relying on all neutrino sources but reactors. In particular,
neutrinos from supernova (SN) burst, solar neutrinos, and geonutrinos. A complete review of \textsc{Juno}'s
physics goals can be found in~\cite{yellow_book}.
\section{Neutrino Physics at JUNO}
\noindent \textsc{ \textbf{Supernova Burst Neutrinos}}\\
\indent A SN is a stellar explosion that briefly outshines an entire galaxy,
radiating as much energy as the Sun or any ordinary star is expected
to emit over its entire life span. During such explosion, 99\% of
the gravitational binding energy of the newly formed neutron
star is emitted in the form of $\nu$.
The observation of SN $\nu$ is expected to play a relevant role both in
particle physics and astrophysics. Here we focus on the latter, where a SN signal might
help answering several fundamental questions, such as (\textsc{i}) what are the conditions
inside massive stars during their evolution?, (\textsc{ii})
what mechanism triggers the SN explosion?, (\textsc{iii})
are SN explosions responsible for the production of heavy chemical elements?, and
(\textsc{iv}) is the compact remnant a neutron star or a black hole?
Each of these questions would deserve a dedicated section, but because of the limited space
we consider (\textsc{i}) as a case study.
The Standard Stellar Evolution Model describes temperature and density of a star
as a function of time and distance from its centre. Optical observations usually provide
benchmark data to test it, but optical observations have little power in constraining
the model of the star's innermost layers. Indeed, the star's high density results in optical photons
propagating mainly via diffusion, hence loosing all the information about the stellar core.
On the contrary, $\nu$s interact weakly with stellar matter, and they represent a
powerful tool to probe the inner structure of the star.
In the case a star close to its collapse, the $\nu$ production is dominated by thermal processes
(mainly $e^{+}$-$e^{-}$ annihilating into $\nu$-$\overline{\nu}$ pairs). That is, the $\nu$
production rate, and the $\nu$ mean energy, both increase significantly with temperature. As a result,
the last stages of the star's nuclear burning produce the most abundant $\nu$ signal
(called pre-SN $\nu$), easier to detect and powerful in describing the stellar evolution.
Fig.~1 shows the simulated inverse beta decay (IBD) event rate in \juno{} for the nearest possible SN progenitor
(the red supergiant Betelgeuse) whose mass is taken to be 20 solar masses (M$_{\odot}$) at a distance of 0.2~kpc.
The sudden drop in the rate around 0.6 day before the SN explosion is ascribable to a drop
of the core temperature, mostly due to the silicon depletion of the core itself.
In the case of a SN explosion, \juno{}'s capability to precisely measure the position of such a dip
could serve as a discriminator for different progenitor star masses.
Moreover, the quick rise starting few hours prior to core collapse
makes \juno{} an ultimate pre-warning system of SN explosion,
extremely valuable to the astrophysics community.
For a typical galactic SN at 10~kpc, there will be more than 5000 signal events solely from the IBD channel.
However, several other $\nu$ interactions contribute to the total event rate.
They differ in terms of total yield, energy spectrum and energy threshold.
Fig.~2 shows all of them together, where $(\mathrm{E_d})$
is the deposited visible energy in the detector, $(\mathrm{E^{th}})$ is the energy threshold of
each process,
$(\nu\text{-p})$ are the neutral current interactions on protons,
$(\nu\text{-e})$ are the elastic scatterings on electrons,
$(^{12}\text{C NC})$ are the neutral-current-mediated carbon excitations,
$(^{12}\text{N CC})$ are the charged current \nue{} interactions on $^{12}$C, and
$(^{12}\text{B CC})$ are the the same charged current interactions initiated by \anue{}.
\noindent\begin{tabular*}{\textwidth}{ p{0.45\textwidth} p{0.02\textwidth} p{0.45\textwidth} }
\includegraphics[height=5cm]{4_8_pdf}
&
&
\includegraphics[height=5cm]{4_2_pdf} \\
{\small \textbf{Figure 1.} Neutrino event rate in JUNO for a massive star (20
M$_{\odot}$) distant 0.2~kpc from the Earth,
the same as the nearest possible SN progenitor Betelgeuse.\vspace{10pt}}
& &
{ \small \textbf{Figure 2.} Neutrino energy spectra
in the \juno{} detector for a SN at 10~kpc, where no neutrino flavor conversions is assumed.
$\mathrm{E_d}$ is the visible energy and $\mathrm{E^{th}}$ the threshold energy.
}\\
\end{tabular*}
\noindent \textsc{ \textbf{Solar Neutrinos}}\\
\indent The Sun is a powerful source of \nue{} with O(1~MeV) energy, produced in the thermonuclear
fusion reactions happening in the solar core. \juno{}'s solar $\nu$ programme focuses on those emitted by the
$^7$Be and $^8$B chains.
Indeed, despite the great achievements of the last decades, there are still important aspects of solar
$\nu$ physics to clarify, and some questions of great relevance for astrophysics and elementary
particle physics waiting for definite solutions. Two of the most important ones are
(\textsc{i}) the solution of the solar metallicity problem,
and (\textsc{ii}) the detailed analysis of the
oscillation-probability energy dependence in the lower end
of the $^8$B $\nu$ spectrum.
(\textsc{i}) The solar metallicity problem emerged when
the former agreement between Standard Solar Model (SSM) and solar data
got compromised by the revision of the solar surface heavy element content,
leading to a discrepancy between the SSM and helioseismology results.
The predictions of different SSM versions differ (also) by the $^8$B and $^7$Be neutrino fluxes.
\juno{}'s capability to determine these fluxes with high accuracy,
together with data (coming from other future experiments) about the CNO fluxes,
could help solving this key issue in nuclear astrophysics.
(\textsc{ii}) According to the Mikheyev-Smirnov-Wolfenstein (MSW) effect, $\nu$ oscillation
parameters are different if $\nu$ propagates through matter or in vacuum.
In the case of solar \nue{},
the transition between the two behaviors
is expected to happen in the 1\textasciitilde3~MeV range, therefore
solar $^8$B $\nu$s ---with their continuous energy spectrum stretching far beyond 3~MeV ---
are a privileged tool to study the MSW-modulated energy dependence.
The theory predicts a smooth transition between
the vacuum and matter related \nue{}-survival probabilities,
namely an up-turn in the spectrum. However,
none of the existing experiments so far observed a clear evidence of this effect.
The only exception is Super-Kamiokande, which got a mild evidence of the up-turn in its data~\cite{Renshaw:2013dzu}.
\juno{}'s capability to perform an independent and high-significance
test of the up-turn existence would be extremely important to confirm the consistency
of the standard LMA-MSW solution, or to indicate any possible deviations from this standard paradigm.
\noindent\begin{tabular*}{\textwidth}{ p{0.45\textwidth} p{0.02\textwidth} p{0.45\textwidth} }[t]
\includegraphics[height=5cm]{6_3b_pdf}
&
&
\includegraphics[height=5cm]{6_4_pdf}
\\
{\small
\textbf{Figure 3.} Energy spectra of singles from natural radioactivity (background) in the \textit{ideal}
radiopurity scenario, together with the $^7$Be $\nu$ signal.\vspace{10pt}}
& &
{ \small \textbf{Figure 4.}
Energy spectra of the $^8$B $\nu$ signal and of the cosmogenic backgrounds.}\\
\end{tabular*}
The challenge in detecting solar \nue{} at \juno{} is that they are detected only by elastic scattering,
which results in an experimental signature (single energy deposition) indistinguishable from most
of the background processes. The two main background sources are natural radioactivity and cosmogenic isotopes.
The first needs to be suppressed by achieving a high-level radiopurity in all the detector components.
\textsc{Juno}'s \textit{baseline} radiopurity scheme envisages a $^{232}$Th, $^{40}$K, $^{14}$C residual contamination at
the level of $10^{-16}$, $10^{-16}$, $10^{-17}$ g/g respectively. As a comparison, the same level of
radiopurity was achieved during KamLAND solar phase, and it would allow \juno{} to achieve a signal/background
ratio of 1/3. A more demanding radiopurity scheme (called \textit{ideal}) requires the previous contamination levels to improve
by one order of magnitude, which would correspond to Borexino phase I, and would allow a 2/1 signal/background ratio.
The energy spectrum of the radioactive background processes in the case of \textit{ideal} radiopurity, together with the signal
$^7$Be \nue{}, are shown in Fig.~3.
Among the cosmogenic isotopes, the most dangerous are the long-lived ones, namely $^{10}$C ($\tau=24.4$~min)
, $^{11}$C ($\tau = 27.8$~s), and $^{11}$Be ($\tau=19.9$~s) , since they cannot be suppressed by a muon veto
without introducing a relevant deadtime. The energy spectrum of these background events is shown in
Fig.~4. The only way to handle them is to tag them via a three-fold coincidence (muon + spallation neutron
+ isotope decay) and subtract them statistically from the total spectrum.
\noindent \textsc{ \textbf{Geoeutrinos}}\\
\indent Over the last half a century, the Earth's surface heat flow has been established to be $46\pm3$~TW.
However the community is still vigorously debating what fraction of this power comes from primordial
versus radioactive sources. This debate touches on the composition of the Earth, the question of chemical
layering in the mantle, the nature of mantle convection, the energy needed to drive plate tectonics,
and the power source of the geodynamo, which powers the magnetosphere that shields the Earth from the harmful cosmic ray flux.
Radioactive beta-decays of heavy elements (such as Th and U) taking place inside the Earth
result in an upwards \anue{} flux (also called geoneutrino flux) which can be detected at \juno{} by means
of IBD reactions. A precise measurement of such flux would allow to accurately define
the absolute abundance of Th and U in the Earth, which in turn would allow to:
(\textsc{i}) define the building blocks, the chondritic meteorites, that formed the Earth,
(\textsc{ii}) discriminate models of parameterised mantle convection that define the thermal evolution of the Earth,
(\textsc{iii}) potentially identify and characterize deep, hidden reservoirs in the mantle, and
(\textsc{iv}) fix the radiogenic contribution to the terrestrial heat flow.
Moreover, such studies can place stringent limits on the power of any natural nuclear reactor in or
near the Earth's core.
The main experimental challenge in detecting a geoneutrino signal is
to disentangle it from the reactor \anue{} signal, which is overwhelming. Such a separation can be done
only via statistical subtraction, and it relies heavily on a precise modeling of the
low-energy reactor \anue{} spectrum. Moreover, to interpret the geoneutrino signal in terms of mantle's
radioactivity, the contribution from the Earth's crust need to be subtracted,
since it's been well established that
the crust surrounding the detector will play a major role in total geoneutrino budget.
Thus, to understand the relative contributions from the crust and mantle to the total
geoneutrino signal at \juno{}, detailed geological, geochemical, and geophysical studies
need to be performed in the areas surrounding the detector.
\section{Conclusion}
\textsc{Juno}'s physics programme is extremely broad, and makes it a genuine general purpose neutrino experiment.
In this poster we presented only some of the topics associated to neutrinos not coming from nuclear reactor,
namely supernova neutrinos, solar neutrinos and geoneutrinos. A complete review of \juno{}'s
physics programme can be found in~\cite{yellow_book}.
\vspace{4pt}
\noindent \textbf{\large Acknowledgments}\\
\indent MG acknowledges support from the CAS President's International Fellowship Initiative grant 2015PM007.
MG wishes to thank Virginia Strati for helpful discussion on geoneutrinos.
|
\section{Introduction}
\label{sec:intro}
Over the past decade it has become clear that in order to understand the formation and evolution of galaxies we need to account for the various feedback processes that are able to both stimulate and truncate star formation. These are generally split into three forms of feedback. In the less massive galaxies supernovae provide the necessary energy to drive the gas from these relatively shallow potential wells, providing negative feedback that is able to truncate the star formation \citep[e.g.][]{Efstathiou2000,DallaVecchia2008}. In more massive galaxies, supernovae do not provide enough energy to drive the gas out of more massive haloes, and therefore an alternative mechanism is required. Negative feedback from active galactic nuclei (AGN) was introduced to models and simulations of galaxy formation in order to halt the star formation in more massive galaxies \citep[e.g.][]{Bower2006,Croton2006,Sijacki2007}. Since the initial introduction of feedback from AGN in simulations, further progress has been made in differentiating between radiative (or cold-mode) feedback and the mechanical (or hot-mode) feedback that results principally from the passage of radio jets and lobes \citep[e.g.][]{Hardcastle2007,BestHeckman2012,Fabian2012,Fernandes2015}. However, it is clear that some level of positive feedback also occurs as a result of AGN activity: radio jets sometimes appear to stimulate star formation as they pass through the interstellar and intergalactic medium \citep[e.g.][]{Croft2006,SilkNusser2010,Silk2013,Kalfountzou2012,Kalfountzou2014}.
To differentiate between these forms of feedback, and to trace their evolution, requires observations that are both deep enough to detect star formation and low-luminosity radio jets in galaxies, and wide enough to obtain the number statistics with which to trace the evolution of the rarer bright sources. Radio continuum observations provide a unique opportunity to study both the black-hole accretion activity in the Universe \citep[e.g.][]{JarvisRawlings2000,Smolcic2009,Rigby2011,McAlpine2013,Best2014}, and the star-formation history of the Universe \citep[e.g.][]{Seymour2008,Smolcic2009sf,Jarvis2015a}, at a wavelength that is free from selection biases due to dust obscuration.
Additionally, it has been suggested that the radio continuum surveys envisaged to be carried out with the new generation of radio telescopes will provide a complementary technique to measure the large-scale structure of the Universe, but to much higher redshifts and volumes than is feasible with optical surveys \citep[e.g.][]{Raccanelli2012,Camera2012,Ferramacho2014,Jarvis2015b}. Although these are predominantly the remit of the very deep all-sky surveys that will be carried out with the Square Kilometre Array \citep{Dewdney2013}, much progress in this area can be made now by obtaining a greater understanding of the properties of radio sources over fields targeted due to their extensive multi-wavelength data.
Various deep-field surveys have been conducted with the (Karl G.~Jansky) Very Large Array \citep[VLA; e.g.][]{Bondi2003,Simpson2006,Schinnerer2007,Miller2013,Heywood2013}, the Westerbork Synthesis Radio Telescope \citep[e.g.][]{devries2002}, the Giant Metrewave Radio Telescope \citep[e.g.][]{Garn2008} and the Australia Telescope Compact Array \citep[e.g.][]{Norris2006,Middelberg2008}. All of these cover areas of the order of 1-10~deg$^2$, which, although very good for detecting the bulk of the typical radio sources that lie close to the flux-density limit, miss the rarer populations, such as the most luminous starbursts and powerful AGN. The volume surveyed also limits the investigation of the environmental influence in triggering or truncating AGN and/or star-formation activity at relatively low redshift \citep[e.g.][]{Peng2010,Burton2013,Karouzos2014,Sabater2014}. For this reason, projects such as the Galaxy and Mass Assembly \citep[GAMA; ][]{Driver2011,Hopkins2013} survey, along with multi-wavelength data at radio \citep{Mauch2013} and far-infrared \citep{Eales2010} wavelengths, provide an important tier in the general wedding-cake structure of extragalactic surveys.
The SDSS Stripe~82 region is quickly becoming the field of choice for surveys covering $\ge 100$~square degrees. This is due to the wealth of multi-wavelength data that has been accrued following the repeat observations in $ugriz$ as part of the SDSS Survey, reaching $g\sim 24.5$ \citep{Jiang2014,Annis2014}. Additional multi-wavelength data include near-infrared imaging ($YJHK$) as part of the UKIRT Infrared Deep Sky Survey \citep[UKIDSS; ][]{Lawrence2007} and more recently the VISTA Hemisphere Survey \citep[VHS;][]{VHS} to $K_{\rm AB}\sim 20$, far-infrared imaging data from the {\em Herschel Space Observatory} as part of the {\em Herschel} Multi-tiered Extragalactic Survey \citep[HerMES;][]{Oliver2012} and {\em Herschel} Stripe 82 Survey \citep[HerS;][]{Viero2014}, and at millimetre-wavelength imaging with the Atacama Cosmology Telescope \citep[ACT;][]{Hincks2010}. A survey covering 92~square degrees within Stripe 82 to a median flux-density limit of 52~$\mu$Jy beam$^{-1}$ at 1.4 GHz has already been carried out with the pre-upgrade VLA in its most extended A-array configuration \citep{Hodge2011}. This dataset provides excellent resolution ($\sim$1.5~arcsec) imaging with sufficient positional accuracy to allow cross-matching to the wealth of ancillary data. However, the A-array data also resolves out the more-diffuse emission from lobes and any possible extended emission from star-forming galaxies. They do not therefore obtain an accurate measure of the total flux density from these systems. Furthermore, objects identified as single point sources in the \citet{Hodge2011} data may in fact turn out to be components of a single extended source.
We have used the Karl G.~Jansky Very Large Array in a compact, hybrid CnB configuration to target the area covered by the observations of \citet{Hodge2011}. This configuration results in excellent sensitivity to diffuse and low surface brightness radio structures, with short baselines sensitive to structures up to $\sim$16 arcmin in extent, making the observations highly complementary to the existing data. The use of 1 GHz of uninterrupted bandwidth means that the new observations can cover the survey area via a series of short snapshot observations and reach depths comparable to those of the existing observations.
\section{Observations and data processing}
\label{sec:selfcal}
The data\footnote{Project code: 13B-272} were taken with the array in the CnB\footnote{This is a hybrid configuration in which the east and west arms of the array are spaced as they are in C configuration and the northern arm of the array is spaced according to the more extended B configuration.} configuration. Standard wide-band mode was employed with the correlator splitting the 1--2 GHz of frequency coverage into 16 spectral windows (SPWs) with 64~$\times$~1~MHz channels each, and an integration time per visibility point of 3 seconds. The programme was divided into sixteen scheduling blocks (SBs) of 4.5 hours duration, each of which was self-contained with both target scans and appropriate calibration observations. The SBs were scheduled dynamically by the observatory. A total of 1368 target pointings were scheduled, 608 and 760 in the eastern and western regions respectively, coincident with the two eastern and western areas of the existing \citet{Hodge2011} data. The observations were designed to follow a standard hexagonal mosaic pattern, with pointings lying at the half-power point of their neighbour in right ascension, such that they are critically sampling the sky at the centre of the frequency band. The total integration time per pointing was approximately 2.5 minutes. 3C48 was observed once per SB for flux calibration, and the strong calibrator J2225-0457 was observed for bandpass calibration. A suitable phase calibrator source within the survey area was visited approximately once per hour. When the array moved from its CnB configuration, 12 of the 16 SBs had been observed resulting in 522 and 504 observed target pointings in the eastern and western regions respectively, and a somewhat unfortunate discontinuity in the coverage of the western patch. The pointing positions and the total sky coverage are shown in Figure \ref{fig:pointings}, with the alternating colours used to demarcate different SBs.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{stripe82_ptgs.eps}
\caption{Coverage of the observed pointings of the survey described in this paper for the eastern (upper) and western (lower) area. The circles show the approximate size of the half-power point of the VLA primary beam at 1.5~GHz. The set of pointings contained within an individual Scheduling Block (SB) are coloured alternately with blue and pink markers. The eastern patch contains 522 pointings. The extra pointing in the western area corresponds to an image derived from the visits to the 2.9~Jy phase calibrator J0059+0006, added to the mosaic to improve the image quality around this source, bringing the total number of pointings contributing to the final western mosaic to 505. Note the discontinuity in the western patch due to one of the Scheduling Blocks not being executed.}
\label{fig:pointings}
\end{center}
\end{figure*}
Initial processing of the data was performed using the NRAO VLA pipeline\footnote{{\tt \tiny \href{https://science.nrao.edu/facilities/vla/data-processing/pipeline}{https://science.nrao.edu/facilities/vla/data-processing/pipeline}}}. This is a set of {\tt CASA}\footnote{\href{http://casa.nrao.edu/}{http://casa.nrao.edu/}} \citep{McMullin2007} scripts that were used to apply Hanning smoothing and make a first pass of flagging bad or unnecessary data. The flagging steps include removal of data from shadowed antennas, the initial few integration points of a scan where not all antennas may be on source and visibility amplitudes that are exactly zero, together with automatic radio frequency interference (RFI) excision using using the sliding time median filtering approach of the {\tt rflag} algorithm. Following these steps the pipeline performs delay and bandpass corrections. Antenna-based complex gain corrections were derived from the calibrator scans and interpolated over the target fields. No post-hoc time or frequency averaging of the 3~s / 1~MHz native resolutions was performed. Retaining these time and frequency resolutions for this array configuration means that time and bandwidth smearing effects are of no concern, the latter mitigated by the imaging method described below. During the observations, visits to a strong source for phase-calibration were scheduled somewhat infrequently in order to minimise overheads, on the grounds that there would be enough flux in any given field to further correct complex gain errors via self-calibration.
Following execution of the NRAO pipeline, each of the target pointings (typically $\sim$90 per SB) was split into a single Measurement Set and a secondary pass of automatic flagging was performed on the calibrated data using the {\tt rflag} algorithm within the {\tt CASA} {\tt flagdata} task. Each pointing was then subjected to a self-calibration procedure.
An estimate of the thermal noise for each pointing was made by measuring the standard deviation of a 4096~$\times$~4096 1.5 arcsec-pixel Stokes-V image, and the dimensions of the fitted restoring beam specific to that pointing were determined. Total intensity images were then made, spanning 4096~$\times$~4096 pixels (again with a 1.5 arcsec pixel scale), encompassing the first sidelobe of the VLA primary beam pattern at the lowest frequency. Multi-term, multi-frequency synthesis (MT-MFS) imaging \citep{Rau2011} was used, as implemented in {\tt CASA}'s {\tt clean} task. Briefly, the clean components are modelled in the frequency dimension by a Taylor polynomial
\begin{equation}
I(\nu)~=~\sum_{t}I_{t}\left(\frac{\nu - \nu_{0}}{\nu_{0}}\right)^{t},
\end{equation}
where $I(\nu)$ is the component brightness as a function of frequency, $I_{t}$ is the image of Taylor polynomial coefficient $t$, and $\nu_{0}$ is the reference frequency about which the expansion is performed, set in this case to the band centre at 1.5 GHz. A typical choice for the spectral model, motivated by the frequency behaviour of astrophysical synchrotron emission, is a power law with a curvature term of the form
\begin{equation}
I(\nu)~=~I_{\nu_{0}}\left(\frac{\nu}{\nu_{0}}\right)^{\alpha+\beta \mathrm{log}(\nu/\nu_{0})}
\label{eq:specmodel}
\end{equation}
where $I_{\nu_{0}}$ is the sky brightness at the reference frequency, $\alpha$ is the power law index (the spectral index\footnote{We adopt the convention $S$~$\propto$~$\nu^{\alpha}$ where $S$ is the flux density, $\nu$ is the frequency and $\alpha$ is the spectral index.}) and $\beta$ is the curvature term. A third-order ($t$~=~0,1,2) Taylor expansion of Equation \ref{eq:specmodel} then returns the total intensity image $I_{\nu_{0}}$~=~$I_{0}$, and allows $\alpha$ and $\beta$ to be determined from linear combinations of the higher-order Taylor coefficient images according to
\begin{equation}
\alpha~=~\frac{I_{0}}{I_{1}}
\label{eq:alpha}
\end{equation}
and
\begin{equation}
\beta~=~\frac{I_{2}}{I_{0}}~-~\frac{\alpha(\alpha - 1)}{2}.
\label{eq:beta}
\end{equation}
A third order Taylor series was employed during deconvolution as, based on tests of a limited subset of the data, it delivered improved dynamic range in the Stokes-I images, particularly for pointings that included off-axis sources with brightnesses in the $\sim$10--100 mJy beam$^{-1}$ range. The full frequency resolution gridding technique that the algorithm adopts renders the effects of bandwidth smearing negligible.
The low declination of Stripe 82, the extended northern arm of the array in CnB configuration and the short snapshot pointings result in a point spread function (PSF) with an elongated main lobe and high sidelobe levels. The use of \citet{Briggs1995} weighting can temper these effects by weighting down the data in the inner part of the $u$,$v$ plane: in broad terms, higher angular resolution and lower sidelobe levels can be achieved at the expense of reduced sensitivity. Down-weighting the short spacings also affects the sensitivity to extended structure, however deconvolution is made significantly easier and more reliable with the application of a weighting function that moves further from natural weighting to uniform weighting. After some trial runs on a limited subset of the data we opted for a robust parameter of 0.3, reasoning that the unique combination of depth and sensitivity to extended emission that these observations have should be retained given the properties of the existing radio observations of Stripe 82.
Deconvolution of a snapshot VLA survey must be done with extreme care. The suppression of source brightnesses in an improperly cleaned image (an effect known as clean or snapshot bias, discussed further in Section \ref{sec:photometry}) is a well known but poorly understood phenomenon. The use of masks to constrain the regions that are subjected to cleaning is known to reduce the effect, as is terminating the cleaning cycle before spurious sidelobe features or noise peaks are cleaned. Thus for each pointing a cleaning mask was generated by convolving the positions of sources from the Faint Images of the Radio Sky at Twenty-cm \citep{Becker1995} catalogue with the Gaussian restoring beam appropriate to that pointing, and converting this into a Boolean mask above the 5$\sigma$ level, where $\sigma$ was the value determined from the initial Stokes-V image.
The images were deconvolved using this mask, terminating the process after 8,000 iterations or when the peak residual dropped below 5$\sigma$. We generated model visibilities (including spectral behaviour) from the Taylor term images that resulted from this process, relying on the use of the cleaning mask to prevent spurious sidelobe features from entering the model. Phase-only gain corrections were then generated for each of the 16 spectral windows using a single solution interval for the 2.5 minute scan. Data that did not meet minimum signal-to-noise or number of baselines criteria when forming a solution were flagged.
The data were then re-imaged using the FIRST-based cleaning mask. For the final images, a second pass of deconvolution was applied with the mask removed, with deconvolution terminating after 2000 iterations or when the peak in the residual image reached 3$\sigma$, whichever was sooner. These final images were restored using a 16 arcsec~$\times$~10 arcsec Gaussian with a position angle (PA) of 90$^{\circ}$ east of north. This is slightly broader than the generally-achievable angular resolution, however it imparts a desirable level of uniformity to the survey products, and it encompasses the variation in the fitted restoring beams that is caused by the dynamic scheduling of the observations. Mean values plus/minus one standard deviation of the distribution of per pointing major axes, minor axes and PAs for the survey are (13.3~$\pm$~2.1) arcsec, (7.5~$\pm$~0.3) arcsec and (88~$\pm$~15) deg respectively.
The {\tt widebandpbcor} task was used to correct for primary beam effects in the image domain. This task works by making a spectral cube of a model of the main lobe of the VLA's primary beam, in this case with one frequency plane for each of the 16 SPWs. This cube represents the directional sensitivity of the array as a function of frequency. The frequency cube is then transformed into 2D images for each Taylor term $t$, in this case for $t$~=~[0,1,2], which are then used to correct the Taylor term images of the sky brightness distribution. The $t$~=~0 image is practically equivalent to the usual narrow-band correction method, i.e. it is divided by a total intensity beam model computed at the band centre. Examining the difference between the {\tt widebandpbcor} $t$~=~0 image and the effective sensitivity image produced by the {\tt clean} task shows deviations of around 1--2\% across the field. Accurate primary beam-corrected flux densities are recovered, aided by the spacing of the pointings and the weighting method used when forming the mosaic (see below). The higher-order Taylor term images should in principle now contain intrinsic spectral index and curvature estimates, with the spectral corruption effects of the primary beam removed. Once the wide band primary beam correction has been applied to the three Taylor term images the spectral index ($\alpha$) and spectral curvature maps ($\beta$) are recomputed as per Equations \ref{eq:alpha} and \ref{eq:beta}. All primary beam corrected images were cut beyond the radius where the normalized, total-intensity primary beam gain as computed at the band centre exceeded 0.4.
At this point the total intensity images were examined to identify pointings for which the automatic calibration procedure had been inadequate. Such datasets invariably fell into one of the following categories: (i) a very strong source inside the imaged area; (ii) a strong source outside of the imaged area, the sidelobes of which were encroaching into the main image; (iii) morphologically complex sources, or very faint extended sources, for which the FIRST-based cleaning mask was inadequate; (iv) residual RFI that had been missed by the automatic flagging.
Case (iv) was remedied by manually editing the visibility data and reprocessing the Measurement Set using the steps described above. Cases (ii) and (iii) were addressed by the addition of outlier fields at the position of the confusing sources, and by setting manual cleaning masks. This was done on the post-NRAO pipelined data, discarding the automatically generated phase-self-calibration solutions. Having generated a model via this process that was certainly more complete than the model used for the automated self-calibration procedure, we executed three iterations of imaging and self-calibration, two passes of phase only self-calibration (with 30 s and 6 s solution intervals), and one pass of amplitude and phase self-calibration with a single correction applied to the whole 2.5 minute scan. Solutions were generated in each of the 16 SPWs. This extended self-calibration procedure was also used to improve case (i) pointings. Primary beam correction was then applied as described above. Of the 1,026 pointings in the survey, 202 (20 percent) of them required manual re-processing, and thus a smaller fraction than this were subjected to amplitude self-calibration. This subset also includes fields for which the automated calibration procedure was sufficient, but had significantly extended sources in the survey which only required a deeper (interactive) deconvolution process than our automated procedure afforded.
Improper amplitude self-calibration can bias the flux density measurements in the resulting map. This is due to the risk of the contribution to the visibility function made by the faintest sources which are typically absent from the sky model being subsumed into the gain solutions. The end result of this is suppression of the unmodelled sources, as well as more subtle, counterintuitive effects, e.g. \citet{grobler2014}, see also \citet{franzen2015} and \citet{heywood2016}. We do not expect to suffer significantly from such self-cal biases, as (i) the model should be largely complete, being formed via interactive, constrained deconvolution which was terminated when the residuals contained only noise, or noise plus artefacts; (ii) our solution interval was conservative, with only a single amplitude and phase correction being applied for the entire scan; and (iii) the VLA is robust against such biases due to the self calibration problem being well constrained, with 27 complex gain terms being derived from 351 equations.
Pointings that were processed manually were re-imaged using their individual clean masks and primary beam corrected following the procedures used in the automated case. Once all the final images were in place, the {\tt Montage}\footnote{{\tt \href{http://montage.ipac.caltech.edu/}{http://montage.ipac.caltech.edu/}}} software was used to construct linear mosaics in total intensity, $\alpha$, $\beta$, and $\alpha$-error. For all mosaics, each pointing is weighted by the assumed variance in the total intensity map (i.e.~the square of the assumed Stokes-I primary beam).
\section{Data products}
\label{sec:results}
The principal products from this snapshot survey are wide area mosaics for the east and west regions, formed directly (Stokes-I, $\alpha$, $\beta$, $\alpha$-error) or derivatively (model, residual and sensitivity) from the individual pointing images as described in Section \ref{sec:selfcal}, and the corresponding lists of components derived from them using automatic source-finding software.
\subsection{Radio mosaics}
\label{sec:mosaic}
Stokes-I mosaics of the east and west regions are shown in Figure \ref{fig:maps}. As each mosaic contains approximately a third of a billion pixels only the brightest sources are obvious in the upper panels showing the full area. The discontinuity in the western field due to the unobserved SBs is apparent as the horizontal gap spanning the width of the field. Representative zooms of a 1$^{\circ}$~$\times$~1$^{\circ}$ region in the eastern field and a 0.5$^{\circ}$~$\times$~0.5$^{\circ}$ region in the western field can be seen in the lower panels. The pixel brightness scale runs from $-$1 to 3~mJy~beam$^{-1}$ for all components of Figure \ref{fig:maps}, as indicated by the colour bar at the top. The image products are examined further in Section \ref{sec:imgcomp}.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{maps.eps}
\caption{Total intensity mosaics showing the full extent of the survey for the eastern and western regions (upper two panels). The pixel scale is linear, is the same for all panels in the image, and runs from -1.0 to 3.0 mJy beam$^{-1}$ as indicated by the colour bar at the top of the figure. The square markers in the eastern and western regions are 1$^{\circ}$~$\times$~1$^{\circ}$ and 0.5$^{\circ}$~$\times$~0.5$^{\circ}$ in extent respectively. A closer view of these regions is shown in the lower panels of the figure. The striations most evident in the western close-up region are associated with the PSF of the observations, and remain in the image due to imperfect deconvolution.}
\label{fig:maps}
\end{center}
\end{figure*}
\subsection{Component catalogue}
\label{sec:catalogue}
The source catalogue for this survey was generated by running the Python Blob Detection and Source Management ({\tt PyBDSM}) source finder \citep{MohanRafferty2015} on each of the wide area radio images presented in Section \ref{sec:mosaic}. A brief overview of the source detection algorithm is as follows. The spatial variation of the image noise was estimated by moving a box across the image in overlapping steps, calculating the root mean square (RMS) of the pixels within the box and interpolating the values measured from each step. The box size was 250 $\times$ 250 pixels, stepped across the image in units of 50 pixels. This procedure provided the source finder with an estimate of the spatial variation of the image noise for detection thresholding purposes. This variation was written to a FITS image in order to provide a map of the sensitivity of the eastern and western fields, and these images are presented in Figure \ref{fig:rms}.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{stripe82_rms.eps}
\caption{Variations in the image noise across the east (upper) and west (lower) mosaic areas shown as a greyscale image. The corresponding values are shown in the colour bars above each panel. The noise is higher at the edge of the mosaic area due to the primary beam correction also raising the image noise, and also around strong or very extended sources where calibration deficiencies or incomplete deconvolution leave PSF-like residuals that effectively raise the noise floor. Components in the catalogue (Section \ref{sec:catalogue}) have the local RMS at their position listed based on the measurement from the above image. The masked rectangular regions visible at the southern edge of the western field (at approximate RAs of 0h57m and 01h25m) correspond to three very extended sources that {\tt PyBDSM} was unable to decompose into a point and Gaussian model (see Section \ref{sec:sources}).}
\label{fig:rms}
\end{center}
\end{figure*}
Image peaks that exceeded 5$\sigma$ where $\sigma$ is the local value of the RMS noise are identified. These detections were then grown into islands, defined as regions where there is contiguous emission above a secondary threshold, in this case 3$\sigma$. The islands were then decomposed into a component list by fitting Gaussian components to them. An attempt is made to measure the intrinsic shape of each component by deconvolving the restoring beam from the component, and then fitting a Gaussian component to the result. If no satisfactory fits are obtained then the component is deemed to be indistinguishable from being point-like.
{\tt PyBDSM} also attempts to identify individual sources within each island via a pairwise examination of the fitted components. The criteria which must be satisfied for components to be part of the same source are (i) the pixels along the line joining the two component centres must all be above the 3$\sigma$ island threshold, and (ii) the length of this line is less than the sum of the full width half maxima of the two components.
Generating a source catalogue from a radio survey in most cases involves a trade-off between maximising the catalogue completeness and minimising the risk of including of spurious features. Although setting a cutoff at 5$\sigma$ will naturally result in the exclusion of numerous faint sources, it has long been established that below this threshold Eddington bias (Eddington, 1913) will preclude sensible interpretation of the results due to the inclusion of very high numbers of spurious noise features \citep{Condon1974,Hogg2001}.
Around strong sources in particular, incomplete deconvolution and residual calibration errors can leave spurious PSF-like features. The Y-shape of the VLA results in linear structure with six-fold symmetry in the PSF for any observation that is not sufficiently long enough to achieve complete Fourier plane coverage, and for snapshot surveys such as this one these features are very pronounced. As mentioned in Section \ref{sec:selfcal}, even with tapered weighting of the visibilities the PSF exhibits sidelobes at 10--20 percent of the peak level. This exacerbates the error patterns around sources that do not deconvolve properly, and in numerous parts of the mosaic components that do not represent genuine astronomical signals will find their way into the source catalogue. The position of every entry in the component list with a peak flux density exceeding 100 mJy beam$^{-1}$ was therefore examined and any visually identified artefacts were removed. Pointings with known calibration deficiencies (typically around the phase calibrator J0059+0006) were also examined for spurious features. In general the use of a local RMS estimate as opposed to fixed threshold results in a catalogue that is not flux limited but does have high reliability. The spectral indices presented in the catalogue are discussed in detail in Section \ref{sec:alphas}.
The final eastern catalogue contains 5,674 individual point and Gaussian components, fitted to 4,602 islands and grouped into 4,354 sources. The western catalogue contains 6,094 components, 4,869 islands and 4,594 sources.
\begin{table*}
\begin{minipage}{170mm}
\centering
\caption{The first ten components from the eastern catalogue, presented here in order to demonstrate the table structure. Note that the local RMS values here are not typical as the first components in the catalogue are extracted from the edge of the mosaic area. Please refer to the text for a detailed description of each column.}\label{tab:components}
\begin{tabular}{rcccccccccc} \hline
&
\multicolumn{1}{c}{ID} &
\multicolumn{1}{c}{R.A.} &
\multicolumn{1}{c}{Decl.} &
\multicolumn{1}{c}{$\sigma_{\textrm{\tiny R.A.}}$} &
\multicolumn{1}{c}{$\sigma_{\textrm{\tiny Decl.}}$} &
\multicolumn{1}{c}{$S_{\textrm{\tiny int}}$} &
\multicolumn{1}{c}{$\sigma_{S_\textrm{\tiny int}}$} &
\multicolumn{1}{c}{$S_{\textrm{\tiny peak}}$} &
\multicolumn{1}{c}{$\sigma_{S_\textrm{\tiny peak}}$} \\
&
&
\multicolumn{1}{c}{[deg]} &
\multicolumn{1}{c}{[deg]} &
\multicolumn{1}{c}{[arcsec]} &
\multicolumn{1}{c}{[arcsec]} &
\multicolumn{1}{c}{[mJy]} &
\multicolumn{1}{c}{[mJy]} &
\multicolumn{1}{c}{[mJy~b$^{-1}$]} &
\multicolumn{1}{c}{[mJy~b$^{-1}$]} \\
&
\multicolumn{1}{c}{(1)} &
\multicolumn{1}{c}{(2)} &
\multicolumn{1}{c}{(3)} &
\multicolumn{1}{c}{(4)} &
\multicolumn{1}{c}{(5)} &
\multicolumn{1}{c}{(6)} &
\multicolumn{1}{c}{(7)} &
\multicolumn{1}{c}{(8)} &
\multicolumn{1}{c}{(9)} \\ \hline
\textcolor{Gray}{\emph{1}}& J232038.0+003139& 350.15832& 0.52752& 0.02 & 0.01 & 80.21& 0.30& 73.98& 0.17& \\
\textcolor{Gray}{\emph{2}}&J232026.9-002731& 350.11217& -0.45875& 0.04& 0.01& 38.78& 0.26& 35.90& 0.15& \\
\textcolor{Gray}{\emph{3}}&J232032.1+002351& 350.13383& 0.39765& 0.06& 0.03& 27.23& 0.29& 24.36& 0.17& \\
\textcolor{Gray}{\emph{4}}&J232049.2+001430& 350.20493& 0.24173& 0.28& 0.14& 7.07& 0.27& 5.25& 0.17& \\
\textcolor{Gray}{\emph{5}}&J232038.9+002718& 350.16226& 0.45522& 0.43& 0.17& 3.23& 0.30& 3.29& 0.16& \\
\textcolor{Gray}{\emph{6}}&J232038.1-005554& 350.15886& -0.93193& 0.48& 0.20& 3.20& 0.26& 2.82& 0.15& \\
\textcolor{Gray}{\emph{7}}&J232034.3-004254& 350.14293& -0.71517& 1.49& 0.45& 1.20& 0.24& 1.00& 0.14& \\
\textcolor{Gray}{\emph{8}}&J232031.3-002624& 350.13030& -0.44017& 1.25& 0.59& 1.10& 0.26& 0.99& 0.15& \\
\textcolor{Gray}{\emph{9}}&J232039.4-001318& 350.16421& -0.22176& 0.99& 1.72& 0.92& 0.27& 0.75& 0.17& \\
\textcolor{Gray}{\emph{10}}&J232039.5-000037& 350.16453& -0.01029& 0.84& 0.40& 0.63& 0.34& 1.11& 0.15& \\
\hline
\end{tabular}
\vspace{2mm}
\begin{tabular}{rcccccccccccc} \hline
&
\multicolumn{1}{c}{Local RMS} &
\multicolumn{1}{c}{$\alpha$} &
\multicolumn{1}{c}{$\theta_{\mathrm{maj}}$} &
\multicolumn{1}{c}{$\sigma_{\theta_{\mathrm{maj}}}$} &
\multicolumn{1}{c}{$\theta_{\mathrm{min}}$} &
\multicolumn{1}{c}{$\sigma_{\theta_{\mathrm{min}}}$} &
\multicolumn{1}{c}{PA} &
\multicolumn{1}{c}{$\sigma_{PA}$} &
\multicolumn{1}{c}{Resolved}&
\multicolumn{1}{c}{ID$_{g}$} &
\multicolumn{1}{c}{ID$_{s}$} &
\multicolumn{1}{c}{ID$_{i}$} \\
&
\multicolumn{1}{c}{[mJy~b$^{-1}$]} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{[arcsec]} &
\multicolumn{1}{c}{[arcsec]} &
\multicolumn{1}{c}{[arcsec]} &
\multicolumn{1}{c}{[arcsec]} &
\multicolumn{1}{c}{[deg]} &
\multicolumn{1}{c}{[deg]} &
&
&
&\\
&
\multicolumn{1}{c}{(10)} &
\multicolumn{1}{c}{(11)} &
\multicolumn{1}{c}{(12)} &
\multicolumn{1}{c}{(13)} &
\multicolumn{1}{c}{(14)} &
\multicolumn{1}{c}{(15)} &
\multicolumn{1}{c}{(16)} &
\multicolumn{1}{c}{(17)} &
\multicolumn{1}{c}{(18)} &
\multicolumn{1}{c}{(19)} &
\multicolumn{1}{c}{(20)} &
\multicolumn{1}{c}{(21)}\\ \hline
\textcolor{Gray}{\emph{1}}& 177.00& -0.12& 4.95& 0.05& 2.53& 0.02& 84.87& 0.13& 1& 1& 20& 1\\
\textcolor{Gray}{\emph{2}}& 150.00& -0.30& 4.60& 0.08& 2.62& 0.03& 97.83& 0.32& 1& 10& 37& 10\\
\textcolor{Gray}{\emph{3}}& 169.00& -1.05& 5.54& 0.14& 2.96& 0.06& 112.10& 0.62& 1& 8& 34& 8\\
\textcolor{Gray}{\emph{4}}& 165.00& -1.47& 8.12& 0.68& 6.23& 0.32& 66.10& 3.91& 1& 0& 0& 0\\
\textcolor{Gray}{\emph{5}}& 173.00& -1.75& 0.00& 0.00& 0.00& 0.00& 0.00& 0.00& 0& 2& 21& 2\\
\textcolor{Gray}{\emph{6}}& 153.00& -0.29& 8.19& 1.17& 0.00& 0.00& 54.30& 5.06& 0& 5& 27& 5\\
\textcolor{Gray}{\emph{7}}& 145.00& -1.42& 10.50& 3.57& 0.00& 0.00& 101.42& 11.46& 0& 7& 33& 7\\
\textcolor{Gray}{\emph{8}}& 152.00& -2.04& 6.69& 3.02& 0.00& 0.00& 138.33& 16.89& 0& 9& 36& 9\\
\textcolor{Gray}{\emph{9}}& 162.00& 1.63& 0.00& 0.00& 0.00& 0.00& 0.00& 0.00& 0& 4& 23& 4\\
\textcolor{Gray}{\emph{10}}& 191.00& 2.67& 0.00& 0.00& 0.00& 0.00& 0.00& 0.00& 0& 3& 22& 3\\
\hline
\end{tabular}
\end{minipage}
\end{table*}
Table \ref{tab:components} lists the first ten components drawn from the eastern catalogue in order to illustrate the structure of the component table. The columns are defined as follows:\\
\noindent
(1) Identifier for the component formed from its HHMMSS.SS+/-DDMMSS.SS right ascension and declination position in J2000 coordinates.\\
\noindent
(2) Right ascension of the component in degrees.\\
\noindent
(3) Declination of the component in degrees.\\
\noindent
(4) The 1$\sigma$ uncertainty in right ascension in arcseconds.\\
\noindent
(5) The 1$\sigma$ uncertainty in declination in arcseconds.\\
\noindent
(6) Integrated flux density of the component in mJy.\\
\noindent
(7) The 1$\sigma$ uncertainty in the integrated flux density of the component in mJy.\\
\noindent
(8) Peak intensity of the component in mJy beam$^{-1}$.\\
\noindent
(9) The 1$\sigma$ uncertainty in the peak intensity of the component in mJy beam$^{-1}$.\\
\noindent
(10) Estimate of the local RMS noise at the position of the component in $\mu$Jy beam$^{-1}$.\\
\noindent
(11) Spectral index ($\alpha$) of the component formed from an error-weighted mean of the pixel values in the spectral index mosaic over the extent of the component, or the extent of the survey restoring beam, whichever is larger.\\
\noindent
(12) Deconvolved major-axis size of the Gaussian fitted to the component in arcseconds. A value of zero indicates that the source finder was unable to determine the intrinsic size, and the component is indistinguishable from being point-like.\\
\noindent
(13) The 1$\sigma$ uncertainty in the major axis of the component in arcseconds.\\
\noindent
(14) Deconvolved minor-axis size of the Gaussian fitted to the component in arcseconds. A value of zero indicates that the component is indistinguishable from being point-like.\\
\noindent
(15) The 1$\sigma$ uncertainty in the minor axis of the component in arcseconds.\\
\noindent
(16) Position angle measured east of north of the Gaussian fitted to the component in degrees.\\
\noindent
(17) The 1$\sigma$ uncertainty in the position angle of the component in degrees.\\
\noindent
(18) Boolean flag to indicate whether or not the component is likely to be reliably spatially resolved, i.e.~does the component lie above the upper envelope on Figure \ref{fig:resolved}.\\
\noindent
(19) Zero-indexed unique identifier for the component.\\
\noindent
(20) Zero-indexed unique identifier for the source.\\
\noindent
(21) Zero-indexed unique identifier for the island.\\
Note that there is not a direct one-to-one correspondence between a component having zeros in columns 12, 14 and 16, and a one in column 18. Both of these methods feature different statistical approaches, and the column 18 method in particular is a somewhat blunt approach that is applied to capture the bulk properties of the source catalogue when determining source counts (see Section \ref{sec:bias} for details). When considering the morphological properties of individual sources, columns 12, 14 and 16 are likely to be more reliable, and the ratio of columns 6 and 8 is also a good indicator.
\section{Results and discussion}
\label{sec:discussion}
In this section the data products presented in Section \ref{sec:results} are examined. We verify the positional accuracy, the photometry and the in-band spectral measurements by making use of existing radio surveys and simulations. We present the differential source counts as derived from the survey, and present radio images of the four bright and extended sources in the survey area that are not included in the component catalogue.
\subsection{Comparison with existing 1.4 GHz radio observations of Stripe 82}
\label{sec:imgcomp}
The radio imaging of Stripe 82 presented in this paper is complementary with existing observations of the field at 1.4~GHz in terms of angular resolution and depth. Our observations lie between the A-array observations of \citet{Hodge2011} and the FIRST survey in terms of depth, and the CnB configuration means that they bridge the FIRST and NVSS datasets in terms of angular resolution and surface brightness sensitivity. The suitability of the new data for morphological classification of extragalactic radio sources is illustrated in Figure \ref{fig:examplesource} which presents the wide angle tail radio source associated with the galaxy SDSS J013412.84-010724.8 ($z$~=~0.078). The first three columns show the A-, B-, and D-array data with the CnB data in the fourth column. Much of the diffuse structure associated with the source is resolved out by the A- and B-array data, and while the NVSS image detects the large scale structure at high significance it is lacking the angular resolution required to disentangle core, jet and hotspot emission, all of which are readily identifiable in the CnB map.
The fifth column shows the morphology of the source as reconstructed by convolving the point (cyan squares) and Gaussian (cyan ellipses) catalogue components with the restoring beam used in the imaging, and the final column shows the residual of the model reconstruction subtracted from the Stokes-I data. The large diffuse feature associated with the northern lobe in the residual map is not detected and modelled by the source-finding software, and is thus absent from the components catalogue. The recovery of diffuse, low surface brightness emission is problematic for most source finders that typically make use of peak-finding algorithms. For this reason total source flux densities determined by summing catalogue components should be considered to be lower limits. The model and residual mosaics form secondary data products along with the total intensity and spectral mosaics. The residual image in particular should be the default product for any continuum stacking experiments \citep[e.g.][]{Zwart2014}.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{examplesource.eps}
\caption{A single radio galaxy is used to illustrate the continuum image products generated by this survey, and present a comparison to previous 1.4~GHz data of the Stripe 82 region. The source in question is a wide angle tail radio galaxy associated with SDSS J013412.84-010724.8 at a redshift of 0.078. The first three columns show the high resolution A-array data of \citet{Hodge2011}, the B-array FIRST data, and the D-array images from NVSS, with the CnB data presented in this paper in the fourth column. The greyscale is linear and runs from -2 to 5 mJy~beam$^{-1}$ for all panels as shown at the base of the figure. The fifth column shows a reconstruction of the source formed by convolving the catalogued point and Gaussian components (the cyan markers) with the restoring beam of the radio mosaics. Subtraction of the reconstructed model image from the data results in the residual image shown in the final column. Note the diffuse feature associated with the northern lobe that is not picked up by the source finder and is therefore absent from the catalogue.}
\label{fig:examplesource}
\end{center}
\end{figure*}
\subsection{Extended sources that are excluded from the components catalogue}
\label{sec:sources}
The survey area contains four bright and extended regions of emission for which {\tt PyBDSM} was unable to obtain a suitable point and Gaussian component fit in a reasonable amount of time. These sources are thus excluded from the component catalogue presented in Section \ref{sec:catalogue}. Total intensity contours for these sources are overlaid on the SDSS $r$-band image in Figure \ref{fig:extended}. The left panel shows the radio source 3C40, which is actually two distinct radio galaxies within the Abell 194 cluster \citep{Sakelliou2008}. 3C40-A is a narrow angle tail (NAT) source associated with NGC 541, and 3C40-B (the brighter and more extended of the two) is a wide angle tail (WAT) source associated with NGC 547. The remaining two panels of Figure \ref{fig:extended} show three sources that are associated with the cluster Abell 119. The centre panel shows two sources that are close to the cluster centre in projection. Higher resolution radio imaging \citep{Feretti1999} reveals these to be a pair of NAT sources. The final source in the right hand panel is 3C 29, a FR-I \citep{Fanaroff1974} source in the periphery of Abell 119.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{resolved_sources.eps}
\caption{Bright and extended radio galaxies that are not included in the component catalogue. These are 3C40-A and 3C40-B in the Abell 194 cluster (left panel), the twin narrow angle tail galaxies (centre panel) and the bright FR-I source 3C29 (right panel). The centre and right panels show sources that are all associated with the cluster Abell 119. The contours trace the total intensity radio emission with levels of (1, $\sqrt{2}$, 2, 2$\sqrt{2}$, 4, 4$\sqrt{2}$, 8, 8$\sqrt{2}$, ...)~$\times$~2 mJy beam$^{-1}$, overlaid on the SDSS $r$-band image. Arcminute scale bars and the size of the restoring beam are provided in each panel.}
\label{fig:extended}
\end{center}
\end{figure*}
\subsection{Survey sensitivity: thermal noise vs effective noise}
\label{sec:sensitivity}
Figure \ref{fig:rms} shows the measured spatial variation in the image sensitivity expressed as the RMS image noise, the determination of which is described in Section \ref{sec:catalogue}. As usual with a primary beam corrected radio image the effective noise is raised at the edges of the field as the off axis source flux densities attenuated by the primary beam response are corrected to their intrinsic values and the image noise is correspondingly raised. Aside from this, regions of elevated noise are otherwise coincident with bright or extended sources where incomplete deconvolution (primarily due to calibration deficiencies) results in residual PSF sidelobes that are interpreted as an effectively higher noise floor.
Figure \ref{fig:cumu_rms} shows a normalized histogram of the pixel RMS values shown in Figure \ref{fig:rms} (pink), as well as those measured from an equivalent mosaic formed from the Stokes-V images mentioned in Section \ref{sec:selfcal} (blue). The RMS values of the Stokes-V mosaic are assumed to provide a reasonable measurement of the thermal noise of the observations. The Stokes-V images are largely empty: any residual peaks in emission are coincident with bright sources, occurring due to (uncorrected) instrumental leakages and manifesting itself as a small fraction of the Stokes-I flux density. The distribution of the Stokes-V measurements peaks at 52~$\mu$Jy~beam$^{-1}$, very close to the target thermal noise of 50~$\mu$Jy~beam$^{-1}$ for which the survey was designed. The values measured in the Stokes-I mosaic are higher, with the distribution peaking at 88~$\mu$Jy~beam$^{-1}$. The difference between the thermal and effective noise is
likely due to the following effects: (i) the conservative and automated approach to deconvolution (see Section \ref{sec:selfcal}) resulting in residual sidelobe structures, particularly around extended sources; (ii) calibration deficiencies for pointings close to very strong sources; (iii) the significant number of pixels that form the perimeter of the survey area having elevated noise levels due to the primary-beam correction.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{sensitivity.eps}
\caption{Normalized histograms of the RMS values from Figure \ref{fig:rms} as determined by the {\tt PyBDSM} source finder and those measured via the formation of Stokes-V images of the individual survey pointings. The Stokes-V measurements peak at 52 $\mu$Jy beam$^{-1}$, however the effective survey ``noise'' as estimated by the source finder is higher, with the distribution peaking at 88~$\mu$Jy beam$^{-1}$. See text in Sections \ref{sec:selfcal} and \ref{sec:sensitivity} for further details.}
\label{fig:cumu_rms}
\end{figure}
\subsection{Astrometric accuracy}
\label{sec:astrometry}
The uncertainties in the measured position of a radio source consist of a statistical component related to the signal to noise ratio (SNR) of the detection as well as the angular resolution of the instrument, and a component associated with systematic astrometric frame errors present in the observation. The latter is generally due to calibration deficiencies \citep{condon1997}. Systematic offsets in the positional measurements of the sources in the components catalogue, as well as an estimate of the level of statistical uncertainty, can be determined by cross matching the source positions with those determined in independent radio surveys. Systematic offsets are typically investigated by selecting sources that are bright enough such that the statistical component is negligible (e.g.~typical phase calibrators) and comparing their measured positions against those obtained with Very Long Baseline Interferometry (VLBI) measurements.
To verify the astrometry of the survey we compare the catalogue positions to matched sources from the A-array data of \citet{Hodge2011} and the B-array data from the FIRST survey \citep{Becker1995}, both of which have superior angular resolution to our CnB data. The component catalogue described in Section \ref{sec:catalogue} is filtered to include only entries where the island identifier is associated with a single component. This produces a subset of components that does not contain very extended or morphologically complex emission (with the exception of the four sources shown in Figure \ref{fig:extended} and discussed in Section \ref{sec:sources}). Component lists from each of the external surveys are then cross referenced with the filtered CnB sample in order to find the nearest associated component, although the cross-match is rejected if the nearest component is separated by more than 5 arcseconds to exclude spurious detections and components that are not detected in both surveys. This results in 2,626 and 2,773 cross matches for the external A- and B-array surveys respectively.
The offsets in right ascension and declination are determined for matched components, and the distribution of these offsets is shown as a two-dimensional histogram with 51~$\times$~51 bins in Figure \ref{fig:astrometry}, with the A- and B-array comparisons in the upper and lower panels respectively. The mean ($\pm$1 standard deviation) values of the distributions in (RA, Dec) in units of arcseconds are ($-$0.018~$\pm$~1.171, $-$0.023~$\pm$~1.018) and ($-$0.03~$\pm$~1.186, $-$0.069~$\pm$~1.094) for the A and B comparisons respectively. The cyan ellipses on each panel are centred on the mean offset position, and the extent of the major and minor axes are one standard deviation of the distribution of offsets in RA and Dec. The distributions also confirm that the $<$5 arcsec separation criterion is not biasing the results.
For the A-array and B-array comparisons respectively, the 1$\sigma$ scatter values are 0.8 arcsec and 0.89 arcsec in right ascension and 0.5 arcsec and 0.6 arcsec in declination. The primary source of the increased scatter in the CnB measurements is likely to be the angular resolution of the survey, noting the increased right ascension offset is consistent with the PSF being elongated in that direction. Scaling the RMS offsets reported by \citet{Hodge2011} by the ratios of the angular resolutions of the A and CnB surveys results in higher scatter than is reported here, however \citet{Hodge2011} reason that bandwidth smearing contributes significantly to their uncertainties. This does not significantly affect our measurements due to our choice of imaging method. Measuring the mean value of the offsets as a function of flux density cut off reveals no systematic bias. Note that the positional uncertainties in the component catalogue are derived from the uncertainties in the fit performed by {\tt PyBDSM}, and are therefore purely statistical.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{astrometry.eps}
\caption{Measured positional differences between catalogued sources from the survey presented in this paper, and those detected by Hodge et al.~(2011; upper panel) and the FIRST survey (Becker, White \& Helfand., 1995; lower panel). The points on the above figure relate only to those entries in the component catalogue that are isolated, i.e.~entries where the island identifier belongs only to one component. The images show the two dimensional histograms (51~$\times$~51 bins) of the offsets in right ascension and declination. The ellipses are centred on the mean offset positions and their extents show the corresponding 1$\sigma$ scatter in the points in right ascension and declination between the CnB survey and the corresponding external survey. The extents of the panels above are one half of the minor axis of the restoring beam used in the CnB survey.}
\label{fig:astrometry}
\end{figure}
\subsection{Photometric accuracy}
\label{sec:photometry}
The calibration of the absolute flux density scale against the model of a strong calibrator source provides the VLA with an excellent measurement accuracy of order one percent \citep{Perley2013}. There are however factors external to the absolute calibration that can bias the flux scale accuracy, e.g.~deficiencies with the subsequent referenced and self-calibration, deconvolution effects and RFI.
Our first check is to compare the flux density measurements of the catalogued components to those of existing surveys. Again we employ the A-array data of \citet{Hodge2011} and the FIRST survey B-array data of \citet{Becker1995}, and additionally the D-array data from the NVSS \citep{Condon1998}. As with the positional matching, islands that contain only a single component are selected from the catalogue, and matches are rejected if the separation between the CnB position and that of the comparison survey exceeds 5 arcseconds. Applying these criteria results in 5,821, 5,616 and 1,145 matches from the A-, B- and D-array data, respectively. These subsets are plotted as grey points on Figure \ref{fig:photometry}, which shows the comparison integrated flux density measurements against those of the CnB components. The 1:1 ratio line is indicated on the plot, and there is increased scatter about this line with decreasing flux density. This is to be expected as the image noise becomes a more significant fraction of the source flux density at fainter levels. A 5$\sigma$ source that is observed multiple times with identical observations will appear to be 20 percent variable by definition.
The three external surveys being used for this comparison have markedly different angular resolutions and sensitivities to extended structure. Despite selecting only isolated components, the differences that this introduces to the flux density comparisons are also evident in Figure \ref{fig:photometry}. For example, the points tend towards a flux excess at the faint end in the high angular resolution A-array data in the upper panel. A similar effect is noted by \citet{Hodge2011} which is interpreted in terms of a combination of resolution effects and Malmquist bias. A component that is resolved in one survey will exhibit a lower peak flux density than its corresponding unresolved measurement. Such a source will be picked up by a peak-finding source detection algorithm only if a positive noise contribution pushes the peak over the threshold. Such sources are therefore biased to high flux densities in the higher resolution data. The turn-up in the distribution of grey points due to this effect is lessened in the B-array data, as expected.
The effects of the mismatched angular resolutions can be further illustrated by requiring components to have peak to integrated flux density ratios of between 0.9 and 1.1 in both samples. This selects sources that are unresolved (or barely resolved) by both the CnB survey and the external reference survey. These components are plotted as blue points on Figure \ref{fig:photometry}, and they sit much closer to the 1:1 line. There is a hint of an excess flux density bias in the faint end of the NVSS measurements. This could also be a resolution effect even for sources that are selected to be point like as extended emission seen by the D-configuration of NVSS may be being resolved out in the CnB data (cf. Figure \ref{fig:examplesource}).
\begin{figure}
\centering
\includegraphics[width= 0.87 \columnwidth]{photometry.eps}
\caption{Comparison of the integrated flux density measurements of our survey with those of existing surveys over the same region, namely the A-array observations of \citet{Hodge2011}, the B-array observations of the FIRST survey \citep{Becker1995} and the D-array NVSS data \citep{Condon1998}. As with Figure \ref{fig:astrometry} only single component islands are selected for comparison in order to minimise the effects of the differing sensitivities to extended structure. The number of matched components is shown on each panel, with the total number shown in grey and the unresolved (in both the CnB catalogue and the reference catalogue) matches shown in blue.}
\label{fig:photometry}
\end{figure}
An additional source of photometric error occurs due to CLEAN (or snapshot) bias. This is an incompletely understood effect related to the deconvolution process, typically manifesting itself as a systematic underestimation of the source brightnesses \citep{Condon1998,Becker1995}. The deficit is typically found to be fixed down to some flux density threshold, below which it becomes flux dependent and has even been shown to affect sources below the thermal noise limit of the images \citep{White2007}. The magnitude of the effect appears to be coupled to the PSF sidelobe level, with the strong linear features in the VLA snapshot PSF thought to be particularly problematic, and unconstrained deconvolution where sidelobe features are more likely to enter the clean model can also cause increased suppression of the field sources. Few attempts have been made to investigate the problem in the context of high fractional bandwidth continuum observations although \citet{Helfand2015} note that the bias appears to correlate with RFI levels.
The deconvolution strategy described in Section \ref{sec:selfcal} was designed to be conservative in order to minimise the effects of clean bias. We estimate the level at which it is present in the final images via the standard technique of injecting artificial sources of known flux density into the calibrated data, re-imaging and comparing the recovered flux density to the known intrinsic flux density. The artificial sources are flat-spectrum point sources placed at random positions within the band centre half power point of the primary beam, with flux densities that span twelve logarithmically-spaced values between 0.32 and 158 mJy inclusive. Such sources were injected into the Measurement Sets of 500 pointings which were then re-imaged, a total of 6,000 independent simulations. If the flux density of the test source exceeded the 1~mJy catalogue limit of the FIRST survey then the cleaning mask was also modified to include that source. These simulations were performed on the Galaxy\footnote{{\tt http://www.pawsey.org.au/our-systems/\\
galaxy-technical-specifications/}} system at the Pawsey Supercomputing Centre.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{clean_bias.eps}
\caption{The results of the simulations used to quantify the level of clean (or snapshot) bias in the survey images. The x-axis shows the known true flux density of the flat-spectrum point source that was injected into the data. The y-axis shows the measured clean blas, i.e.~the measured minus the true flux density. The dashed horizontal lines show $\pm$1$\sigma$ where $\sigma$ is the typical effective RMS noise in the survey (88~$\mu$Jy~beam$^{-1}$) and the vertical line shows the typical detection threshold at five times this value. Statistics on the clean bias are gathered by running the simulation on 500 of the single pointing Measurement Sets from the survey data for each flux density value. The points show the mean value of the clean bias recovered from the simulations, and the error bars show $\pm$1 standard deviation of the distribution.
}
\label{fig:clean_bias}
\end{figure}
Figure \ref{fig:clean_bias} shows the results of these simulations, with the true flux density of the artificial flat-spectrum source inserted into the data along the x-axis, and the clean bias (i.e.~the recovered minus the intrinsic flux density) on the y-axis. The points show the mean values from each flux density bin as measured from the 500 different Measurement Sets into which the artificial sources were inserted. The error bars show $\pm$1 standard deviation of the values in each flux density bin. The dashed horizontal lines are $\pm$1$\sigma$ where $\sigma$ is the typical RMS noise in the survey (88~$\mu$Jy~beam$^{-1}$). The vertical line shows the typical detection threshold at five times the typical RMS noise. As can be seen the lowest bin is below the typical source detection threshold. Since the position of the source was known, a forced fit to the component at that location could be made.
The clean bias expressed as a percentage of the component flux density in the second lowest bin is 5 percent. This is therefore the \emph{average} worst-case error for the clean bias in the \emph{catalogued} component flux densities. The fractional clean-bias errors are $<$1 percent by the fifth flux density bin. Persons making use of the survey images in the sub-5$\sigma$ regime should be mindful that the clean bias errors may lie at the few tens of percent level. The clean bias manifesting itself as an excess at the faint end is also not typically seen in such investigations. This may be an artefact of the MT-MFS algorithm used to deconvolve the broadband data, an effect not seen in previous narrowband studies. Note that although the bias appears to rise significantly with increasing flux density, when expressed as a fraction of the component flux it becomes increasingly insignificant, dropping from 3 percent in the third bin to 0.07 percent in the highest bin. We make no attempt to account for the increase in the fractional-clean bias error that will result from the primary-beam correction: the actual catalogued flux-density measurement for a component will be a weighted sum of its measurements in all the pointings that contribute to the linear mosaic at that position. As this sum is weighted by the assumed noise variance of the image (i.e.~it is coupled to the assumed primary beam pattern) the component flux density measurement will be dominated by those images where the component lies closest to the pointing centre. However in addition to this, and as we will shortly discuss, the clean bias effect appears to vary significantly between pointings, so there is no way to rigorously quantify the bias in terms of the radial separation of a component from the contributing pointing centres.
Unmasked deconvolution with fixed numbers of iterations on a limited subset of the data, followed by a smaller scale round of clean bias simulations, demonstrated a clear advantage for our mask-based deconvolution approach in terms of faithful component photometry. Nonetheless, a significant clean bias effect remains despite our conservative approach. There appears to be no obvious way to mitigate this, short of fully interactive and iterative deconvolution of the data, complete with simulations that monitor the clean bias, allowing an optimum deconvolution strategy to be implemented on a per-pointing basis. The usual statistical approach to quantifying it and correcting it via the injection of artificial sources, as has been implemented here, clearly falls short: the error bars indicate large amounts (typically $\pm$0.2 mJy beam$^{-1}$) of scatter in the measured clean bias. The apparent dependence of the bias on RFI (Helfand, White \& Becker, 2015) has already been mentioned, and residual low level RFI that has been missed by the automatic and manual flagging operations may be present and contributing to the spread. The significant spread in the extents of the main lobes of the per-pointing PSFs (Section \ref{sec:selfcal}) gives another hint, as it is suspected that clean bias is coupled to the PSF sidelobe levels. Despite the very narrow declination range of our survey area, dynamic scheduling of the observations demands a fairly generous range of acceptable hour angle (or local sidereal time) ranges in order to maximise the chances of a Scheduling Block being selected from the queue. The end result is that the elevation of the field can change significantly between observations, and this gives rise to significantly different PSF sidelobe structures. Optimally rather than dynamically scheduled observations are clearly advantageous for large area radio surveys where uniformity of the data products is highly desirable.
A deeper investigation of the clean bias effect in broadband snapshot VLA surveys will be pursued using these data, examining for correlations between the bias and peak sidelobe levels, amount of RFI, and by searching for a frequency dependence by conducting sub-band rather than full-band imaging, with corresponding sets of artificial source simulations. The use of the MT-MFS algorithm, which models the spectral behaviour of clean components as a polynomial pegged to a reference frequency at the band centre, almost certainly plays a role, and the biases that this introduces in the frequency domain is examined in the next section.
\subsection{Spectral index and curvature}
\label{sec:alphas}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{noisy_two_point.eps}
\caption{Spectral index distribution made to illustrate the effect of image noise on the recovered $\alpha$ values, and simulate the distribution of a two-point measurement made by splitting our data into two 512~MHz sub-bands. Catalogued total intensity flux densities at 1.5~GHz are turned into measurements at 1.25 and 1.75~GHz by assigning a spectral index drawn from a normal distribution with a mean of $-$0.71 and a standard deviation of 0.38 \citep{Mauch2013}. This pair of flux density measurements is then perturbed by an appropriate noise value based on the true local RMS of the component. The values of $\alpha$ measured from the noisy flux density measurements are plotted above. Median values of $\alpha$ are plotted as black points, the error bars for which are the median values of the errors in $\alpha$ for that flux density bin. The main purpose of this figure is to illustrate the significant scatter in $\alpha$ that is introduced purely as a result of the SNR of the component decreasing.}
\label{fig:two_point_noisy}
\end{figure}
Estimates of the spectral index ($\alpha$) and curvature ($\beta$) are made for each component in the catalogue by extracting pixels from the $\alpha$ and $\beta$ mosaics at the position of the component, and over an area that is either the fitted size of the Gaussian component or the area of the restoring beam used to form the mosaic, whichever is larger. A corresponding area is extracted from the spectral index error mosaic and $\alpha$ and $\beta$ values are assigned to each component based on the error-weighted mean of the pixels in the extracted area. Note that the spectral index error maps produced by CASA are not necessarily formal uncertainties in the value of $\alpha$, and are intended to be used (as they are here) as a coarse estimate of the reliability of the value of $\alpha$ at any given position in the map. The error estimates are derived from the statistics of the residuals in the Taylor term images, and these are propagated into the calculation of $\alpha$ via Equation \ref{eq:alpha}, resulting in a map of the spectral index error.
The usual approach to determining the spectral index of a radio source is to observe it using two or more separate observations at different radio frequencies, preferably matched in resolution and depth, and fit a spectral index to the flux density measurements \citep[e.g.][]{Mauch2003,Prandoni2006,Ibar2009,Randall2012,Mauch2013,Huynh2015}. Noise in the independent flux density measurements propagates into the spectral index measurement, and reliable spectral index determination requires both high signal-to-noise ratio flux-density measurements, as well as observations that are maximally separated in frequency but remain within the regime where the spectrum is adequately described as a power law (this separation in frequency is sometimes referred to as the spectral 'lever arm'). Catalogues of large numbers of objects derived from radio surveys are typically extracted down to flux density limits of 5$\sigma$, and matching two pairs of such catalogues introduces significant scatter in the spectral index distribution from noise considerations alone. These spectral index distributions are thus best suited to studying the \emph{average} spectral index behaviours of large numbers of radio sources, for example median spectral index as a function of flux density, population type, etc.
The notion that the new generation of broad-band radio interferometers can make accurate in-band spectral index measurements is an attractive one in terms of making economical use of telescope time. Two approaches are available here, namely (i) imaging the data in matched resolution sub-bands and deriving spectral index measurements in a method analogous to the dual-observation approach described in the paragraph above \citep[e.g.][]{heywood2016}; and (ii) using broadband deconvolution algorithms to image the entire band, fitting an estimate of the spectral behaviour of each component in the process. The latter method has the distinct advantage of providing a single image that has the depth afforded by the broad bandwidth, the angular resolution of the approximate band centre frequency and the PSF sidelobe benefits provided by MFS gridding. It is this method that we adopt and verify here, using the MT-MFS algorithm as described in Section \ref{sec:selfcal}.
By way of a benchmark, Figure \ref{fig:two_point_noisy} presents a simulation of the spectral index distribution that would be obtained by dividing the 1--2 GHz of bandwidth into two parts and determining $\alpha$ from flux density measurements at 1.25 and 1.75 GHz. This is achieved by taking the catalogued component flux densities at 1.5~GHz together with their local RMS measurements. A spectral index is assigned to each component drawn from a normal distribution with a mean of $-$0.71 and a standard deviation of 0.38, adopting the distribution measured by \citet{Mauch2013}. The assigned spectral index value is used to compute 1.25 and 1.75~GHz flux densities based on the catalogued 1.5~GHz value, and these are then perturbed by two independent noise values drawn from a normal distribution centred on zero with a standard deviation of $\sqrt{2}$ times the local RMS value. This simulates the increase in noise that would result from imaging the band in two 512~MHz chunks, assuming equal sensitivity across the VLA's 1--2 GHz band. An $\alpha$ value is computed from the two noisy measurements and the resulting distribution is shown in Figure \ref{fig:two_point_noisy}. The upper panel shows a histogram of the $\alpha$ values, colour coded by where the component flux density lies in a set of logarithmically-spaced integrated flux density bins, as indicated at the top of the figure. The lower panel shows the distribution of the component integrated flux densities as a function of $\alpha$, with the black points showing the median $\alpha$ measurements in each flux density bin, with an error bar corresponding to the median error in $\alpha$ for the values in that bin. An error value for $\alpha$ for each individual point is computed by propagating the simulated flux density errors in the measurement at each frequency. Note that as the flux density decreases so does the SNR, and this introduces scatter into the $\alpha$ measurements that goes way beyond that of the assumed intrinsic $\alpha$ distribution. The SNR-related errors start to becoe significant below about 1 mJy.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{alpha_beta_obs.eps}
\caption{The upper row shows histograms of the in-band $\alpha$ (left column) and $\beta$ (right column) values for each component, as derived from the third-order MT-MFS imaging. The lower row shows plots of the integrated flux density of each component as a function of $\alpha$ (left column) and $\beta$ (right column). All points are coloured according to where the component integrated flux density lies in a set of logarithmically-spaced bins as indicated at the top of the figure. The black points in the lower panels indicate the median $\alpha$ and $\beta$ measurements for each integrated flux density bin. Error bars on the median $\alpha$ values are those derived for Figure \ref{fig:two_point_noisy}, intended to provide an indication of the broadening of the distribution due to SNR considerations. The $\alpha$ values indicate a trend towards flatter spectra with decreasing flux density, whereas the $\beta$ values rapidly tend towards unreasonable spectral curvatures as the component flux densities decrease.}
\label{fig:alphas}
\end{center}
\end{figure*}
The distribution of the spectral indices ($\alpha$) in the catalogue (measurement of which is described in Section \ref{sec:catalogue}) are presented in Figure \ref{fig:alphas}, along with the spectral-curvature measurements ($\beta$) which were extracted using the same method. As per Figure \ref{fig:two_point_noisy}, the upper panels show histograms of the $\alpha$ and $\beta$ values, colour coded by the component flux density, as indicated at the top of the figure. Broadening of the distribution with decreasing flux density is again evident, likely due to SNR considerations. The lower panels show the distribution of the component integrated flux densities as a function of $\alpha$ and $\beta$, with the black points showing the median $\alpha$ and $\beta$ measurements in each flux density bin. Error bars on the median values of $\alpha$ are those derived from the simulation used to produce Figure \ref{fig:two_point_noisy}. These show the approximate error introduced purely as a function of SNR, and independent of the underlying true spectral index distribution. The scatter introduced by decreasing SNR starts to become significant below about 1 mJy.
Our first check is to measure the spectral index of components matched with the quasar positions from the Stripe 82 DR7 quasar catalogue \citep{Shen2011}. This gives a median spectral index of $-$0.4 for compact (S$_{peak}$/S$_{int}$ $>$ 0.95) radio components compared to $-$0.58 for the components with extended emission. A Kolmogorov-Smirnov test rejects the hypothesis that the two spectral index samples are drawn from the same distribution ($p$-value $<$2 percent).
There is an apparent flattening of the median spectral index with decreasing flux density, as quantified in Table \ref{tab:medalphas} and shown in Figure \ref{fig:alphas}. The spectral behaviour of radio sources as a function of their flux density has been investigated by several authors. Studies at the bright end are best done with shallow, large-area surveys \citep[e.g.][]{Massardi2011} where a shift between flat spectrum quasars and radio galaxies with steeper spectra is seen at flux densities above $\sim$100~mJy. Moving to fainter flux density levels, \citet{Mauch2003} compare the 843 MHz Molonglo Observatory Synthesis Telescope (MOST) observations of the Sydney University Molonglo Sky Survey (SUMSS) to the 1.4 GHz NVSS measurements. A shift towards flatter spectra with decreasing flux density is observed, with a median spectral index of $-$0.89 above 50 mJy, increasing to $-$0.77 below 20 mJy, likely due to the shift from steep-spectrum Fanaroff-Riley Type-I radio galaxies \citep{Fanaroff1974} to lower luminosity flat spectrum AGN. For comparison, the median spectral index in our catalogue above 50 mJy is $-$0.92, and below 20 mJy it is $-$0.79, consistent with existing dual-frequency measurements.
Figure \ref{fig:alphas} shows a gradual shift towards a flatter median spectral index with decreasing flux density, and the presence of such a trend is not consistently reported in the literature. Previous studies of the spectral indices of sources at $\sim$mJy levels typically involve deeper surveys covering a few primary beam areas, and typically compare 1.4 GHz observations made with the VLA or ATCA with either lower frequency 325, 610 MHz GMRT or 843 MHz MOST observations, or higher frequency 5~GHz VLA or ATCA data. Some studies measure a flattening, e.g.~\citet{Prandoni2006} compare 1.4 and 5 GHz observations, and observe a median spectral index shift from $-$0.71 above a 1.4 GHz flux of 4~mJy to $-$0.53 for the components below this threshold. Authors who claim no evidence for spectral flattening include \citet{Ibar2009}, who present a spectral index measurement of between $-$0.7 and $-$0.6 for a source sample with 1.4 GHz VLA fluxes between $\sim$100~$\mu$Jy and 10 mJy, derived from well-matched GMRT observations at 610~MHz. For comparison, our median spectral index in this flux range is $-$0.8. \citet{Randall2012} find ``no statistically significant evidence for'' (but do not rule out) a systematic flattening based on 843 MHz MOST and 1.4 GHz ATCA observations. \citet{Huynh2015} present a median spectral index of $-$0.58 by cross matching a sample of 187 point sources observed with ATCA at 5 GHz with high resolution 1.4 GHz VLA data \citep{Miller2013}. These median measurements are flatter then the ones we present in Figure \ref{fig:alphas}, however this is likely a result of comparing a 1.4 GHz survey with a second survey at a higher frequency. At frequencies $\sim$10 GHz rest-frame and above, and faint flux densities, one would expect a typical galaxy to have a flatter radio spectrum as thermal (free-free) processes begin to contribute to the radio output at an increasing fraction compared to non-thermal synchrotron emission \citep{Whittam2013,Murphy2015}.
The spectral curvature of extragalactic radio sources is significantly less well studied than the typical spectral index, particularly for faint sources. While certain rare objects are known to exhibit extremely curved, peaked spectra \citep[e.g.][]{Callingham2015}, the amount of curvature exhibited by a typical synchrotron spectrum between 1--2 GHz is expected to be very slight. Guglielmino (2013) examines the spectra of 79 radio sources in the Lockman Hole between 345 MHz and 1.4 GHz. In that study, curvature is defined as the difference between the spectral index measured between 345 and 610 MHz, and that measured between 610 MHz and 1.4 GHz. The mean and median spectral curvatures for the sample are both consistent with zero. (We note however that the median $\alpha$ value shifts from $-$0.82 in the sample with 150 MHz fluxes above 60 mJy to $-$0.65 in the sample below 60 mJy.) In any case, MT-MFS is only expected to be able to fit meaningful $\beta$ values for typical (i.e.~only slightly curved) radio source spectra at very high SNRs ($\mbox{{\raisebox{-0.4ex}{$\stackrel{>}{{\scriptstyle\sim}}$}}}$100). The increasingly significant curvature with decreasing flux density present in our catalogue is suspicious, and will be discussed further below.
A truly consistent picture is yet to emerge from these many observations, however there are clearly many selection biases at play, such as: shot noise effects in the smallest-area, deepest observations capturing a mixture of source populations; the ability of higher-resolution imaging to separate core-jet sources into individual components that are blended in lower resolution studies; and differing survey depths at the two frequencies biasing the matched samples towards steep-spectrum sources at the faint end. The in-band spectral index measurements that are possible with new and upgraded broadband radio telescopes may offer considerable advantages in terms of minimising the differences present when determining spectral index measurements from observations made at two widely separated frequencies with two different instruments, for example resolution mismatches and different calibration systematics. In this paper we have opted for use of the MT-MFS deconvolution method to estimate the spectral properties of the components. However, the reliability of these measurements must be determined.
\begin{figure*}
\begin{center}
\setlength{\unitlength}{1cm}
\includegraphics[width = \textwidth]{alpha_beta_sim.eps}
\caption{The upper row shows histograms of the recovered values of $\alpha$ and $\beta$ derived from approximately 25,000 imaging simulations where sources of known spectral shape were inserted into the VLA data. The sources had $\alpha$ values of either $-$0.7, $-$1.0 or $-$1.4, and no spectral curvature ($\beta$~=~0) in each case. These three $\alpha$ values are shown by the coloured lines (as per the inset legend) on the lower plots which show the behaviour of $\alpha$ and $\beta$ as a function of the flux density of the artificial source. The black points show the median $\alpha$ and $\beta$ values as derived from the observations and presented in Figure \ref{fig:alphas}, and in the case of the spectral indices, Table \ref{tab:medalphas}. This figure shows that the observed trend towards flatter spectra with decreasing flux density is at least partially an artefact of the MT-MFS algorithm as it was applied to these data, and that the curvature values should be disregarded. The crosses in the lower left panel show a smaller scale simulation whereby $\alpha$~=~$-$0.7 sources in the lowest three flux density bins were imaged using only a second-order Taylor expansion. The artificial flattening in this case is much less severe, however please refer to the text for a full discussion of this.}
\label{fig:alphas_sim}
\end{center}
\end{figure*}
To achieve this we expand on the simulation method used to determine the clean bias in Section \ref{sec:photometry}. Point sources were once again injected into Measurement Sets of 500 of the survey pointings; however, each source was assigned a spectral index value of either 0.0, $-$0.7 or $-$1.4. Spectral curvature in each case was zero. The same twelve logarithmically-spaced flux density bins were retained, but extra iterations were required to build up the statistics at the faint end as sometimes the reconstruction of the synthetic source in the higher order Taylor terms was not satisfactory and no spectral index could be fit. In total approximately 25,000 viable simulations were executed on the Galaxy supercomputer.
Figure \ref{fig:alphas_sim} recreates Figure \ref{fig:alphas} but presents the reconstructed $\alpha$ and $\beta$ measurements for the simulated sources. On the lower two panels the lines coloured blue, green and red track the recovered flux density-dependent spectral indices for the three intrinsic values shown in the inset legend. Large points are the median $\alpha$ and $\beta$ values, again coloured by flux density as indicated at the top of the figure. The black points show the median measurements from the catalogue in each flux density bin, as per Figure \ref{fig:alphas} and Table \ref{tab:medalphas}.
Regarding the $\alpha$ values, the simulations reveal something potentially insidious, in that the apparent flattening of the median spectral index may at least in part be an artefact of the broadband deconvolution approach. A shift towards flatter spectra begins below $\sim$3 mJy, but the rate at which this flattening occurs seems to be a function of the intrinsic steepness of the spectrum. As mentioned in Section \ref{sec:selfcal} a third-order Taylor series was used during imaging. For comparison we run a smaller simulation whereby $\alpha$~=~$-$0.7 sources are simulated for the lowest three flux density bins, but the data are imaged with a second-order Taylor expansion. The results are shown by the crosses on Figure \ref{fig:alphas_sim} (n~=~2 in the legend). Some flattening still occurs but it is less pronounced than in the third-order data. The flux densities at which this flattening begins means that it also cannot be explained via SNR considerations, as our simple two-band simulation illustrates (Figure \ref{fig:two_point_noisy}). This appears to present us with a trade-off: the spectra at the faint end can be recovered more faithfully, at the expense of reducing the dynamic range of the survey images (and therefore increasing the effective noise floor, see Section \ref{sec:sensitivity}). In our limited trial imaging runs increasing the number of terms to 3 significantly improved the quality of the Stokes-I images, particularly around brighter sources. We ascribe this to the significant variation in the extent of the primary beam across the 1--2 GHz of frequency coverage, which imparts pronounced artificial spectral shapes in off-axis sources that are more faithfully deconvolved by the higher-order approach.
The simulations also confirm that the suspicious behaviour of the observed $\beta$ values is also likely to be an artefact of fitting a higher-order polynomial, with faint sources exhibiting unreasonable levels of curvature in their reconstructed model spectra. For this reason we omit the curvature measurements from the component catalogue. Note that the steepest $\alpha$~=~$-$1.4 simulated sources all have a negative curvature measurement even at the brightest flux densities. While it is reasonable that the $\beta$ measurements derived from the data for the very strongest sources may be accurate, the persistent offset between the measured and simulated curvature values, and the significantly different derivative of the curve is a mystery. We speculate that this may be related to the initial self-calibration step described in Section \ref{sec:selfcal}. A single per-antenna phase-only gain correction is made for each $\sim$2.5 minute snapshot, however the gain corrections are derived by producing a visibility model derived from the components in the Taylor term images, constrained by the FIRST-based cleaning mask. This might give rise to a positive feedback situation whereby anomalous spectral shapes are reinforced by the self-calibration, and then enhanced further in the subsequent round of MT-MFS imaging. This would remain hidden in a flux density comparison (e.g. Figure \ref{fig:photometry}) as MT-MFS was fixing the flux density at the band centre. If this proves to be the case then it suggests that a sub-band self-calibration approach is probably better for such large fractional bandwidths, with independent sky models derived from images formed from smaller fractional bandwidth frequency chunks which can be reliably imaged using 1 or 2 Taylor terms to model the spectra. As with the clean bias we also cannot rule out effects such as survey-wide RFI persisting in certain regions of the spectrum that is not excised by the automatic flagging from playing a part. A spectral window with a flux scale that was skewed high would introduce discontinuities that would affect the reconstructed spectrum.
Unfortunately, although these simulations have proved enlightening in terms of identifying the point where the spectral index measurements may become unreliable, they are unable to provide any heuristic corrections to nullify the biases, cf. the FIRST and NVSS approach to clean bias. The artificial flattening of $\alpha$ with decreasing flux density is clearly dependent on the intrinsic spectral shape of the component being modelled, and since genuine astronomical radio sources will have a range of intrinsic spectral shapes, no single correction factor will be valid.
\begin{table}
\centering
\caption{Median spectral index measurements for components with integrated flux densities between $S_{1}$ and $S_{2}$. The number of components in each bin is denoted by N. These values are plotted as black markers on the lower panel of Figure \ref{fig:alphas}. See the caveats as to the reliability of these measurements at fainter flux densities in Section \ref{sec:alphas}.}
\begin{tabular}{ccccc} \hline
$S$ & $S_{1}$ & $S_{2}$ & $\alpha_{\mathrm{med}}$ & $N$ \\
$[$mJy b$^{-1}]$ & $[$mJy b$^{-1}]$ & $[$mJy b$^{-1}]$ & & \\ \hline
0.49 & 0.4 & 0.58 & $-$0.37 & 496 \\
0.71 & 0.58 & 0.85 & $-$0.53 & 1176\\
1.04 & 0.85 & 1.23 & $-$0.72 & 1775\\
1.51 & 1.23 & 1.79 & $-$0.78 & 1741\\
2.2 & 1.79 & 2.61 & $-$0.82 & 1446\\
3.21 & 2.61 & 3.8 & $-$0.86 & 1264\\
4.67 & 3.8 & 5.53 & $-$0.88 & 1048\\
6.79 & 5.53 & 8.05 & $-$0.86 & 746\\
9.89 & 8.05 & 11.72 & $-$0.94 & 601\\
14.39 & 11.72 & 17.06 & $-$0.92 & 424\\
20.95 & 17.06 & 24.83 & $-$0.94 & 327\\
30.49 & 24.83 & 36.14 & $-$0.9 & 252\\
44.37 & 36.14 & 52.6 & $-$0.95 & 146\\
64.58 & 52.6 & 76.56 & $-$0.92 & 114\\
93.99 & 76.56 & 111.43 & $-$0.98 & 65\\
136.8 & 111.43 & 162.18 & $-$0.92 & 46\\
199.11 & 162.18 & 236.04 & $-$0.97 & 21\\
289.79 & 236.04 & 343.54 & $-$0.72 & 8\\
421.77 & 343.54 & 500.0 & $-$0.95 & 9\\ \hline
\end{tabular}
\label{tab:medalphas}
\end{table}
\subsection{Eddington and resolution bias corrections, visibility areas, and the differential source counts}
\label{sec:bias}
This section describes measurements of the Euclidean-normalized differential source counts as constructed from the component catalogues, together with estimates of relevant bias corrections. For an overview of the subject of radio source counts we refer the reader to the review by \citet{deZotti2010}. Although our measurements do not break new ground in this area, determining the source counts and associated bias corrections are an additional effective way to validate the results of a radio survey.
The source counts are determined in logarithmically-spaced flux density bins as per Table \ref{tab:counts}. The counts are expressed per unit area, therefore the first correction we determine is the so-called visibility area of each flux density bin in order to compensate for the non-uniform depth of the survey. The visibility area ($A_{vis}$) is simply the sky area over which the source detection procedure will be sensitive to a source with a flux density at or above the lower edge of the bin. This is computed using Figure \ref{fig:rms} by summing the total pixel area with RMS values exceeding one-fifth of the lower bin edge (recalling the 5$\sigma$ peak threshold described in Section \ref{sec:catalogue}). The visibility areas are listed for each bin in Table \ref{tab:counts}. Since the lowest visibility area corresponding to the faintest flux density bin is 40.4 square degrees, the effects of field to field variations on the source counts are negligible \citep{HeywoodJarvisCondon2013}.
\begin{table}
\centering
\caption{Tabulated differential source counts in each flux density bin, centered at $S_{\mathrm{bin}}$, in both raw (N) and Euclidean-normalized form. The uncertainties for the Euclidean-normalized counts are based on the Poisson ($\sqrt{N}$) uncertainties of the raw counts. Visibility areas ($A_{vis}$) and corrections due to Eddington bias ($f_{\mathrm{E}}$) are also provided.}
\begin{tabular}{cccccc} \hline
$S_{\mathrm{bin}}$ & $A_{vis}$ & $f_{\mathrm{E}}$ & $N$ & $S^{2.5} dN/dS$ & Uncertainty \\
$[$mJy$]$ & $[$sq.deg.$]$ & & & $[$Jy$^{1.5}$~sr$^{-1}]$ & $[$Jy$^{1.5}$~sr$^{-1}]$ \\ \hline
0.68 & 40.4 & 1.4 & 334 & 4.0 & 0.1 \\
0.95 & 76.9 & 1.2 & 869 & 5.4 & 0.1 \\
1.34 & 98.1 & 1.0 & 1508 & 7.5 & 0.2 \\
1.89 & 104.3 & 1.0 & 1647 & 10.4 & 0.3 \\
2.66 & 105.3 & 1.0 & 1485 & 14.6 & 0.4 \\
3.74 & 105.8 & 1.0 & 1244 & 21.4 & 0.6 \\
5.27 & 105.9 & 1.0 & 1106 & 28.9 & 1.0 \\
7.42 & 105.9 & 1.0 & 892 & 35.9 & 1.4 \\
10.46 & 105.9 & 1.0 & 656 & 49.5 & 2.1 \\
14.73 & 105.9 & 1.0 & 554 & 60.3 & 3.0 \\
20.74 & 105.9 & 1.0 & 391 & 78.5 & 4.5 \\
29.22 & 105.9 & 1.0 & 310 & 103.8 & 6.6 \\
41.15 & 105.9 & 1.0 & 243 & 111.9 & 8.9 \\
57.96 & 105.9 & 1.0 & 159 & 149.4 & 13.3 \\
81.64 & 105.9 & 1.0 & 126 & 144.8 & 16.9 \\
114.99 & 105.9 & 1.0 & 73 & 169.0 & 23.7 \\
161.96 & 105.9 & 1.0 & 51 & 193.9 & 32.8 \\
228.11 & 105.9 & 1.0 & 35 & 138.9 & 35.9 \\
321.29 & 105.9 & 1.0 & 15 & 108.3 & 40.9 \\
452.54 & 105.9 & 1.0 & 7 & 207.0 & 73.2 \\
637.39 & 105.9 & 1.0 & 8 & 173.0 & 86.5 \\ \hline
\end{tabular}
\label{tab:counts}
\end{table}
\citet{Eddington1913} bias is a by-product of image noise that causes a redistribution of the source counts in the fainter bins. As the value of a flux density bin approaches the noise level of the image the noise becomes an increasingly significant fraction of the component flux density. Thus there is an increased chance of a random noise spike coinciding with the position of a component and pushing it into a neighbouring bin. Eddington bias is also the reason that it is not advisable to count sources below a threshold of 5$\sigma$, as its effects quickly become devastating below this threshold. Even at 5$\sigma$ one can expect $\sim$6 spurious features in a 10,000~$\times$~10,000 pixel image.
Quantifying its effects in the catalogue of a survey such as this, derived using position-dependent thresholds, is not trivial. To estimate the Eddington bias correction for each flux density bin we make use of a simulation of the extragalactic radio source population \citep{Wilman2008,Wilman2010} and the effective survey sensitivity map (Figure \ref{fig:rms}). The 1.4 GHz component flux densities are extracted from a random 100 square degree subset of the simulation. The flux density of each component is then perturbed by an additive noise value drawn from a normal distribution with a mean value of zero, the RMS of which is selected from a weighted distribution based on the survey RMS measurements given in Figure \ref{fig:rms}. This process is repeated a thousand times and the mean `perturbed' source count distribution is measured from the resulting catalogue using the same set of bins as listed in Table \ref{tab:counts}. This process simulates the Eddington bias: the effect of the noise perturbation is to redistribute the counts between bins in a way that becomes increasingly significant with the decreasing SNR of the component. Since we can determine the true source counts from the unperturbed simulation, the factor $f_{\mathrm{E}}$ required to remove the effects of Eddington bias can be estimated by taking the ratio of the true counts to the perturbed counts. This does assume that the model counts are correct, however they agree very well with observed values in the flux density regime we are considering. The correction factors only deviate from unity for the lowest two flux density bins as indicated in Table \ref{tab:counts}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{resolved.eps}
\caption{The ratio of integrated ($S_{\mathrm{int}}$) to peak ($S_{\mathrm{peak}}$) flux density for each catalogued component as a function of its integrated flux density expressed in units of the local RMS noise ($\sigma$). In the high signal to noise regime the ratio $S_{\mathrm{int}}$~/~$S_{\mathrm{peak}}$ being greater than 1 is a robust indicator that a component is spatially resolved. The dashed lines indicate the `reliability locus' above which a component is deemed to be resolved (the blue points).}
\label{fig:resolved}
\end{figure}
The final effect we investigate is the resolution bias (distinct from the resolution mismatch discussed in Section \ref{sec:photometry}). Resolution bias is defined here as a underestimation of source counts in a particular flux density bin due to a resolved component having a lower peak flux density than an unresolved component with equivalent total (integrated) flux density. Examination of the effect in other radio surveys suggests that at the 16 arcsec~$\times$~10 arcsec resolution of our images the resulting catalogue will not be significantly affected. However we verify this via the analytic method used by e.g.~\citet{Prandoni2001} and \citet{Huynh2005}.
If the ratio of the integrated ($S_{\mathrm{int}}$) to peak ($S_{\mathrm{peak}}$) flux density of a detected component is greater than unity then this is a robust indicator that at the component is spatially resolved, with
\begin{equation}
\frac{S_{\mathrm{int}}}{S_{\mathrm{peak}}}~=~\frac{\theta_{\mathrm{maj}}\theta_{\mathrm{min}}}{b_{\mathrm{maj}}b_{\mathrm{min}}}
\label{eq:s_int}
\end{equation}
where $\theta_{\mathrm{maj}}$ and $\theta_{\mathrm{min}}$ are the (deconvolved) major and minor axes of the component and $b_{\mathrm{maj}}$ and $b_{\mathrm{min}}$ are the major and minor axis of the restoring beam. The reliability of this metric depends on the SNR ($S_{\mathrm{int}}$~/~$\sigma$) of the detection, i.e.~a source detected at low significance would require $S_{\mathrm{int}}$~/~$S_{\mathrm{peak}}$~$\gg$~1. The first step is to determine a `reliability envelope' on a $S_{\mathrm{int}}$~/~$S_{\mathrm{peak}}$ vs SNR plot, as illustrated on Figure \ref{fig:resolved}. Components below the horizontal line corresponding to an integrated-to-peak flux density ratio of 1 exhibit scatter that increases with decreasing SNR. These components are taken to indicate the region on the plot for which the integrated-to-peak flux density ratio is an unreliable indicator that a component is resolved as it may be dominated by noise in the measurement. Reasoning that there is an equivalent distribution of sources above the unity line, a function below the line is defined that encompasses 90 percent of the points, as indicated by the dashed line. This boundary is mirrored above the plot, represented by the function
\begin{equation}
\frac{S_{\mathrm{int}}}{S_{\mathrm{peak}}}~=~1~+~\left[\frac{c}{\mathrm{SNR}^{1.4}}\right]
\label{eq:resolved}
\end{equation}
where $c$ is determined by iteration to be 35. Using Equation \ref{eq:s_int} it follows that for a survey with a peak flux density limit of 5$\sigma$
\begin{equation}
\frac{S_{\mathrm{int}}}{5\sigma}~=~\frac{\theta_{\mathrm{max}}^{2}}{b_{\mathrm{maj}}b_{\mathrm{min}}},
\end{equation}
i.e.~for a given $S_{\mathrm{int}}$ the maximum angular-size that a source can have before its peak flux density drops below the detection threshold is
\begin{equation}
\theta_{\mathrm{max}}~=~\left[\frac{S_{\mathrm{int}}}{5\sigma}(b_{\mathrm{maj}}b_{\mathrm{min}})\right]^{0.5}.
\label{eq:thetamax}
\end{equation}
Conversely the envelope defined by Equation \ref{eq:resolved} can be combined with Equation \ref{eq:s_int} to give
\begin{equation}
\theta_{\mathrm{min}}~=~\left[(b_{\mathrm{maj}}b_{\mathrm{min}})\left(1~+~\left[\frac{c}{\mathrm{SNR}^{1.4}}\right]\right)\right]^{0.5}
\label{eq:thetamin}
\end{equation}
where $\theta_{\mathrm{min}}$ is the minimum angular size a source can have before it can be reliably deemed to be resolved as a function of its signal to noise ratio. Using the two angular size limits defined by Equations \ref{eq:thetamax} and \ref{eq:thetamin} together with an assumed true source angular size distribution allows estimation of the incompleteness corrections required for the data due to component resolution biases. Following \citet{Windhorst1990} the true size distribution $h(\theta)$ is assumed to follow the exponential function
\begin{equation}
h(\theta)~=~\mathrm{exp}\left[-\mathrm{ln}~2\left(\frac{\theta_{\mathrm{lim}}}{\theta_{\mathrm{med}}}\right)^{0.62}\right]
\end{equation}
where
\begin{equation}
\theta_{\mathrm{lim}}~=~\mathrm{max}(\theta_{\mathrm{min}},\theta_{\mathrm{max}})
\end{equation}
and
\begin{equation}
\theta_{\mathrm{med}}~=~2S^{0.3}
\end{equation}
where S is the flux density in mJy. The correction factor $f_{\mathrm{R}}$ for the counts is then given by
\begin{equation}
f_{\mathrm{R}}~=~\frac{1}{1 - h(\theta)}.
\end{equation}
In all the above calculations an area-weighted mean value of $\sigma$ is used. The resulting resolution bias correction factors are unity for all flux density bins. As expected, the counts do not need to be adjusted to compensate for this effect.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{source_counts.eps}
\caption{Euclidean-normalized differential source counts derived from the Stripe 82 survey (blue points) overlaid on other observationally-derived measurements by various authors, as tabulated by \citet{deZotti2010}.}
\label{fig:source_counts}
\end{figure}
Figure \ref{fig:source_counts} shows the final values of the differential source counts, normalised to a Euclidean universe, with the visibility area and Eddington bias corrections applied. The raw and normalised counts, along with the relevant correction parameters are listed in Table \ref{tab:counts}. The values derived agree with existing observations. The significant discrepancies at the bright end are due to the relatively small area of the survey in terms of studying these rare and bright objects. We have made no attempts to compensate for this incompleteness.
\section{Conclusions}
\label{sec:conclusions}
We have used the Karl G.~Jansky Very Large Array to conduct a snapshot survey covering $\sim$100 square degrees of the SDSS Stripe 82 region, reaching a thermal noise level of approximately 50~$\mu$Jy beam$^{-1}$, but with a typical 1$\sigma$ effective depth of 88~$\mu$Jy beam$^{-1}$. The use of the CnB configuration results in a survey with good sensitivity to low surface brightness and diffuse emission, highly synergistic with existing 1.4 GHz radio imaging of the field. Our component catalogue contains 11,782 point and Gaussian components, associated with approximately 8,948 unique radio sources. Photometric and astrometric performance of the survey is excellent, as verified by comparison with measurements from three external radio surveys of varying resolution and depth covering the same field. The differential source counts derived from the catalogue are in good agreement with previous studies.
We used the full 1--2 GHz bandwidth available with the VLA's L-band system, and in doing so have produced one of the first large area radio surveys to feature in-band spectral index estimates in the catalogue without having to cross match the results with an external survey at a second frequency. Spectral measurements were derived by forming images from the entire bandwidth using the MT-MFS algorithm, identifying components in the Stokes-I images and then extracting error-weighted measurements of $\alpha$ and $\beta$ for each component. The clean bias effect remains an issue for snapshot surveys with the VLA, despite our best efforts to minimise it with a conservative approach to deconvolution. However it also appears that there are significant biases in the spectral domain associated with the MT-MFS algorithm. The apparent trend towards flatter spectra with decreasing flux density is shown to be at least partially artificial, and dependent on the intrinsic spectral shape of the source, and the behaviour of the spectral curvature estimates is not deemed to be reliable, although this is expected for anything but the brightest components. Above a threshold of $\sim$3~mJy we deem our median $\alpha$ measurements to be reliable, and in good agreement with existing measurements derived by cross-matching the results of pairs of independent observing programs at different frequencies.
The use of large scale simulations to investigate the biases in the imaging process has proven to be very informative, however unlike surveys such as FIRST and NVSS it does not seem practical to derive a single correction factor for the clean bias, which appears to vary significantly from pointing to pointing. A significant contributing factor is likely to be the changes in the PSF shape that arise due to the dynamic scheduling of the observations, which impart an undesirable level of non-uniformity to the survey. A deeper examination of how the clean bias correlates with factors such as the PSF shape and the level of residual RFI will be pursued. Although single correction factors to the Stokes-I flux density and the spectral measurements cannot be determined by running large-scale simulations on a per-observation basis, as was the method for FIRST and NVSS, an estimate of the overall level of the biases can be obtained, and indeed should be, in order to properly understand the limitations of survey products. The simulations were computationally expensive and required the use of a supercomputer, although we note that this may be a natural application for cloud computing platforms.
Finally, the vastly increased survey potential of the now complete VLA upgrade cannot be understated. The sub-millijansky radio sky can now be probed with a snapshot survey with very modest integration times per pointing, and the potential for the instantaneous characterisation of the bulk spectral behaviour of radio sources without the need for conducting two surveys (with mismatched characteristics) is huge, provided the associated systematics, examples of which are highlighted in this paper, can be understood and mitigated.
Our images and catalogues are publicly available, and can be retrieved from the Australia Telescope National Facility's FTP server: {\tt \href{fftp://ftp.atnf.csiro.au/pub/people/hey036/Stripe82}{ftp://ftp.atnf.csiro.au/pub/people/hey036/Stripe82}}.
\section*{Acknowledgements}
\addcontentsline{toc}{section}{Acknowledgements}
We thank the anonymous referee and the MNRAS editorial staff for providing very useful comments on this paper. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This work was supported by resources provided by the Pawsey Supercomputing Centre with funding from the Australian Government and the Government of Western Australia. IH thanks the Rhodes Centre for Radio Astronomy Techniques and Technologies (RATT) for the provision of computing facilities. This research has made use of NASA's Astrophysics Data System. This research made use of Montage. It is funded by the National Science Foundation under Grant Number ACI-1440620, and was previously funded by the National Aeronautics and Space Administration's Earth Science Technology Office, Computation Technologies Project, under Cooperative Agreement Number NCC5-626 between NASA and the California Institute of Technology. Some figures in this paper were created using the Python package APLpy, an open-source plotting package for Python hosted at {\tt \href{http://aplpy.github.com}{http://aplpy.github.com}}. IH acknowledges useful discussions with Natasha Maddox. IH thanks the participants of the SAGE workshop, and SKA South Africa for their hospitality during this event.
|
\section{Introduction}
In \cite{Atiyah}, Atiyah introduced $L^2$-Betti numbers for manifolds with
cocompact free $G$-action for a discrete group $G$ (later generalized to
finite free $G$-CW-complexes). There, he asked \cite{Atiyah}*{p.~72} about the
possible values these can assume.
This question was later popularized in precise form as the so-called ``strong
Atiyah conjecture''. One easily sees that the possible values depend on $G$.
For a finite subgroup of order $n$ in $G$, a free cocompact $G$-manifold
with $L^2$-Betti number $1/n$ can be constructed.
For certain groups $G$ which contain finite subgroups of arbitrarily large
order, with quite some effort manifolds $M$ with $\pi_1(M)=G$ and with
transcendental $L^2$-Betti numbers have been constructed
\cites{Austin,Grabowski,Pichot-Schick-Zuk}. In the following, we will
therefore concentrate on $G$ with a bound on the orders of finite subgroups.
The $L^2$-Betti numbers are defined using the $L^2$-chain complex.
The chain groups there are of the form $l^2(G)^d$, and the differentials are
given by convolution multiplication with a matrix over $\mathbb{Z}[G]$.
The strong Atiyah conjecture for free finite $G$-CW-complexes
is equivalent to the following (with $K=\mathbb{Z}$):
\begin{definition}\label{DAtiyah}
Let $G$ be a group with a bound on the orders of finite
subgroups and let $\lcm(G)\in\mathbb{N}$ (the positive integers)
denote the least common multiple of these orders.
Let $K\subset\mathbb{C}$ be a subring.
We say that $G$ satisfies the \emph{strong Atiyah conjecture over $K$,
or $K[G]$ satisfies the strong Atiyah conjecture} if for every
$n\in\mathbb{N}$ and every $A\in \Mat_n(K[G])$
\begin{equation*}
\dim_G(\ker(A)):=\tr_G(\pr_{\ker A}) \in \frac{1}{\lcm(G)}\mathbb{Z}.
\end{equation*}
Here, as before, we consider $A\colon l^2(G)^n\to l^2(G)^n$ as a
bounded operator, acting by left convolution multiplication --- the
continuous extension of the left multiplication action on the group
ring to $l^2(G)$. $\tr_G$ is the canonical trace on
$\Mat_n(\NG)$, i.e.~the extension (using the matrix trace) of
$\tr_G\colon \NG\to\mathbb{C}$; $a\mapsto \innerprod{a
\delta_e,\delta_e}_{l^2(G)}$, where $\NG$, the weak closure of
$\field{C}[G]\subset \mathcal{B}(\ell^2(G))$ is the group von Neumann algebra.
If $G$ contains arbitrarily large finite subgroups, we set $\lcm(G):=+\infty$.
\end{definition}
A projection $P$ will always be a self adjoint idempotent, so $P =
P^2 = P^*$, where $^*$ indicates the involution on $\mathcal{N}(G)$.
If $E$ is an idempotent, then $E$ is similar to a
projection $P$ and then $\tr_G(E) = \tr_G(P)$. Also a central
idempotent is always a projection. Note that if $G$ is
an infinite group, then \remts{the set
$\{\tr_G(P)\}$, where $P$ runs through the projectors in
$\Mat_n(\mathcal{N}(G))$, $n\in\mathbb{N}$ consists of all non-negative real numbers}. The strong Atiyah conjecture
predicts, \remts{on the other hand,} that the $L^2$-Betti numbers take values in
the subgroup of $\mathbb{R}$ generated by traces of projectors defined already
over $\mathbb{Q}[H]$ for the finite subgroups $H$ of $G$: the projector
$p_H=(\sum_{h\in H} h)/\abs{H}$ satisfies $\tr_G(p_H)=1/\abs{H}$. And by the
Chinese remainder theorem, the additive subgroup of $\mathbb{R}$ generated by the
$\abs{H}^{-1}$ is exactly $\frac{1}{\lcm(G)}\mathbb{Z}$.
We now turn to the center-valued refinements of the above statements. The
center-valued $L^2$-Betti numbers are obtained by replacing the canonical
(com\-plex-valued) trace $\tr_G$ by the center-valued trace $\tr^u_G% _{\NG$
(see Definition \ref{def:traces}), taking
values in the center of $\NG$. Note
that by general theory \cite{Kadison2}*{Chapter 8}, as every finite von
Neumann algebra has a unique
normalized center-valued trace, this is a powerful invariant: two finitely
generated projective Hilbert $\NG$-modules are isomorphic if and only if their
center-valued dimensions coincide. The center of a ring $R$ will be
denoted $\Zen(R)$.
\begin{definition}\label{def:LG_and_Atiyahconj}
Let $G$ be a group with $\lcm(G)<\infty$,
let $K$ be a subring of $\mathbb{C}$, let $F$ be the field of
fractions of $K$, and assume that $F$ is closed under complex
conjugation. Let $\remts{L_K}(G)$ be the additive
subgroup of $\Zen(\NG)$ generated by $\tr^u_G% _{\NG(P)\in
\Zen(\mathbb{C}[G])\subset \Zen(\NG)$ where $P$ runs through projections
\remts{$P\in F[H]$ with $H\leq G$} a finite subgroup.
We say that $G$ satisfies the \emph{center-valued Atiyah conjecture
over $K$,
or $K[G]$ satisfies the center-valued conjecture} if for every
$n\in\mathbb{N}$ and every $A\in \Mat_n(K[G])$ we have
$\dim^u_G% _{\NG(\ker(A)):=\tr^u_G% _{\NG(\pr_{\ker A})\in \remts{L_K}(G)$.
\end{definition}
Observe that \remts{$G$
satisfies the center-valued Atiyah conjecture over $K$ if and only if
$G$ satisfies the center-valued conjecture over its field of fractions $F$}.
Indeed the ``only
if'' is obvious. On the other hand if $A \in \Mat_n(F[G])$, then \remts{(``clearing
denominators'')} there
exists $0 \ne k \in K$ such that $kA \in \Mat_n(K[G])$, and $\ker A =
\ker kA$, which verifies the ``if" part.
\remts{\begin{proposition}
If a group $G$ satisfies the center-vlaued Atiyah conjecture over $K$ of
Definition \ref{def:LG_and_Atiyahconj}, then $G$ also satisfies the
(classical) strong Atiyah conjecture over $K$ of Definition \ref{DAtiyah}.
\end{proposition}
\begin{proof}
By the universal property of the center-valued trace \cite[Chapter 8]{Kadison2},
$\tr_G=\tr_G\circ\tr^u_G% _{\NG$. We therefore only have to check that $\tr_G(a)\in
\frac{1}{\lcm(G)}\mathbb{Z}$ for all $a\in L_K(G)$. By the definition of
$L_K(G)$, we just have to show that $\tr_G(P)\in\frac{1}{\lcm(G)}\mathbb{Z}$
for each projector $P\in F[H]$, where $H\leq G$ is an arbitrary finite
subgroup. This is of course well known to be true, it follows e.g.~from
the fact that finite groups satisfy the strong Atiyah conjecture over $K$.
\end{proof}
}
\begin{proposition}[compare Corollary \ref{corol:discrete}]\label{prop:LGdiscrete}
If $\lcm(G)<\infty$ then $\remts{L_K}(G)\subset \Zen(\NG)$ is discrete. In particular, the
center-valued Atiyah conjecture predicts a ``quantization'' of the
center-valued $L^2$-Betti numbers.
\end{proposition}
\begin{remark}
As for the ordinary strong Atiyah conjecture, the center-valued Atiyah conjecture
over $\mathbb{Z} [G]$ is equivalent to the statement that the center-valued
$L^2$-Betti numbers for finite free $G$-CW-complexes take values in $\remts{L_\mathbb{Z}}(G)$.
The center-valued $L^2$-Betti numbers have been introduced and used in
\cite{MR1474192}.
\end{remark}
The strong Atiyah conjecture has many applications. Most interesting are
those for a torsion-free group $G$, i.e.~if $\lcm(G)=1$. This is exemplified
by the following surprising result of Linnell \cite{MR1242889}. We first recall
the notion of ``the'' division closure of $K[G]$.
\begin{definition}
Let $G$ be a discrete group and let $K\subset \mathbb{C}$ be a
subring. Let $\mathcal{U}(G)$ denote
the ring of unbounded operators on
$l^2(G)$ affiliated to $\NG$ (algebraically, $\mathcal{U}(G)$ is the Ore localization
of $\NG$ at the set of all non-zero-divisors).
Define the \emph{division
closure} $D(K[G])$ to be the smallest subring of $\mathcal{U}(G)$ containing $K[G]$ which
is closed under taking inverses in $\mathcal{U}(G)$.
\end{definition}
\begin{theorem}\label{theo:skew_field}
Let $G$ be a discrete group with $\lcm(G)=1$ and let $K$ be a
subring of $\mathbb{C}$. Then
$K[G]$ satisfies the strong Atiyah conjecture if and only if $D(K[G])$ is
a skew field.
\end{theorem}
The appealing feature of this theorem is that it provides a canonical
over-ring, namely $D(K[G])$ of $K[G]$ which should be a skew field, provided $G$ is
torsion free. Observe that this implies in particular that $K[G]$ has no
non-trivial zero-divisors. For more information on this, see
\cite{Lueck09}*{Remark 4.11}.
Part of the motivation for the work at hand was the question of how to generalize
Theorem \ref{theo:skew_field} if $\lcm(G)>1$. It turns out that one expects
that $D(K[G])$ is semisimple Artinian. In the situation at hand this means that
$D(K[G])$ is a finite direct sum of matrix rings over skew fields. This is
proved in many cases e.g.~in \cite{MR1242889}.
The present paper \remts{gives} a very precise (conjectural) description of
$D(K[G])$, and if it is satisfied we call $D(K[G])$ \emph{Atiyah-expected
Artinian}: the
lattice of finite subgroups and their $K$-linear representations
give a precise prediction into which matrix summands $D(K[G])$ decomposes
and the size of the corresponding matrices. The precise formula is a bit
cumbersome, so we don't give it here but refer to Definition
\ref{def:Atiyah-Artinean}.
One of our main theorems is the precise generalization of Theorem
\ref{theo:skew_field}.
\begin{theorem}
Let $G$ be a discrete group with $\lcm(G)<\infty$ and let $K$ be a
subfield of $\mathbb{C}$ closed under complex conjugation.
Then $K[G]$ satisfies the center-valued Atiyah conjecture if and only if
$D(K[G])$ is Atiyah-expected Artinian.
\end{theorem}
Indeed, we show in Theorem \ref{con:at} that these two properties are also
equivalent to the property that $K_0(D(K[G]))$ is generated by the images of
$K_0(K[H])$ as $H$ runs over the finite subgroups of $G$.
\begin{definition}
Given a discrete group $G$ with $\lcm(G)<\infty$, let $\Delta^+(G)$
denote the
maximal finite normal subgroup, and let $\Delta(G)$ denote the \emph{finite conjugacy
center}, i.e.~the set of those elements of $G$ which have only a finite
number of conjugates.
Indeed, by \cite{Passman}*{\S 1}, $\Delta(G)$ is a normal subgroup of
$G$. Recall that the product of two normal subgroups is a normal subgroup,
therefore, as
$\lcm(G)<\infty$, $\Delta^+(G)$ makes sense. Note that $\Delta^+(G)\subset
\Delta(G)$, indeed, using \cite{Passman}*{Lemma 19.3} it is exactly the
subset of all elements of finite order in $\Delta(G)$.
\end{definition}
In the special case $\Delta^+(G)=\{1\}$, we have that $D(K[G])$ is Atiyah-expected
Artinian if and only
if it is an $\lcm(G)\times\lcm(G)$-matrix ring over a
skew field, and by Theorem \ref{con:at} this is equivalent to the
center-valued Atiyah conjecture
(which in this case is implied by the usual Atiyah conjecture, as the relevant
part of $\Zen(\NG)$ is $\mathbb{C}[\Delta^+(G)]$). This special case (and slightly
more general situations) have already been covered in \cite{L+S2}, but without
the use of the center-valued trace. It turns out that the general case
requires this more refined dimension function. However, much of our arguments
for Theorem \ref{con:at} follow closely the arguments of \cite{L+S2}.
In \cite{L+S2}, a variant of the division closure, namely the ring $\mathcal{E}(K[G])$ is
introduced and used (compare Definition \ref{def:EKG}). It is closed under
adding central idempotents in $\mathcal{U}(G)$
which generated the same submodules as elements already in the ring. We expect
that this actually coincides with $D(K[G])$.
\begin{theorem}\label{theo:EKG_equal_DKG}
If $\lcm(G)<\infty$, $K$ is a subfield of $\mathbb{C}$ which is
closed under complex conjugation
and $G$ satisfies the center-valued Atiyah conjecture, then $\mathcal{E}(K[G])=D(K[G])$.
\end{theorem}
As the second main result of the paper we establish the center-valued Atiyah
conjecture for
certain classes of groups (namely almost all for which the original Atiyah
conjecture is known). The algebraic closure of $\mathbb{Q}$ will be
denoted $\overline{\mathbb{Q}}$.
\begin{theorem}\label{theo:good_groups}
Let $K$ be a subfield of $\mathbb{C}$ which is closed
under complex conjugation.
The center-valued Atiyah conjecture \remts{over $K$} is
true for the following groups $G$:
\begin{enumerate}
\item\label{item:linnellsC} all groups $G$ which belong to Linnell's class
of groups $\mathfrak{C}$ of Definition \ref{def:CCC}, in
particular all free by elementary amenable groups $G$.
\item\label{item:braids_and_others} if $K$ is contained
in $\overline{\mathbb{Q}}$, all elementary amenable extensions
of
\begin{itemize}
\item pure braid groups
\item right-angled Artin groups
\item primitive link
groups
\item virtually cocompact special groups, where a ``cocompact special groups''
is a fundamental group of a compact special cube complex ---this class of
groups contains Gromov
hyperbolic groups which act cocompactly and properly on CAT(0) cube complexes,
fundamental groups of compact hyperbolic $3$-manifolds with empty or
toroidal boundary, and Coxeter groups
without a Euclidean triangle Coxeter subgroup,
\item or of products of the above.
\end{itemize}
\end{enumerate}
\end{theorem}
\begin{question}
Missing in the above list are congruence subgroups of $\SL_n(\mathbb{Z})$ and
finite extensions thereof. Note that the usual Atiyah conjecture for these
groups,
as long as they are torsion free, is proved in \cite{MR2279234}. For
torsion-free groups, the center-valued Atiyah conjecture is not stronger
than the usual Atiyah conjecture. However, it would be interesting to
generalize the work of \cite{MR2279234} to certain extensions which are not
torsion free, and then (or along the way) to deal with the center-valued
Atiyah conjecture for these.
\end{question}
Recall that the center-valued Atiyah conjecture for a group $G$ only makes an
assertion when $\lcm(G)<\infty$. For the proof of \ref{item:linnellsC} of
Theorem \ref{theo:good_groups} we closely follow the method of
\cite{MR1242889}, making use of the equivalent algebraic formulations of the
Atiyah conjecture of Theorem \ref{con:at}. Indeed, we show that the conjecture
is stable under extensions by torsion-free elementary amenable groups. We
actually show (and use) slightly more refined stability properties.
For \ref{item:braids_and_others} of Theorem \ref{theo:good_groups} we use the
approximation theorem for the center-valued $L^2$-Betti numbers,
\cite{Knebusch}*{Theorem 3.2}. Because of the discreteness of the possible
center-valued $L^2$-Betti numbers, the Atiyah conjecture for a suitable
sequence of quotients implies the Atiyah conjecture for the group itself. We
follow here the general idea as already applied in \cite{Schick} and for more
general coefficient rings in \cite{DLMSY}. That this idea can be used for the
class of groups listed in \ref{item:braids_and_others} was shown for the pure
braid groups in \cite{L+S1}, for primitive link groups in \cite{MR2415028} and
for right-angled Coxeter and Artin groups in \cite{LinnellOkunSchick}, and for
cocompact special groups by Schreve in \cite{Schreve} (who uses
fundamentally the geometric insights of Haglund-Wise \cite{Haglund-Wise}, and
develops further the methods of \cite{LinnellOkunSchick}). Agol \cite{Agol}
shows in breakthrough work that
Gromov hyperbolic cocompact CAT(0) cube groups are virtually cocompact
special; with Bergeron-Wise' construction of a cocompact action of a
hyperbolic $3$-manifold group on a CAT(0) cube complex \cite{Bergeron-Wise}
this implies that hyperbolic $3$-manifold groups are virtually cocompact
special.
\section{Preliminaries \remts{on rings associated to groups}}\label{sec:prelim}
\subsubsection*{$\mathcal{U}(G)$, $D(K[G])$ and traces on these}
\begin{definition}\label{def:traces}
Let $G$ be a discrete group. The center-valued-trace is
the uniquely
defined $\mathbb{C}$-linear map
\[\tr^u_G% _{\NG:\NG\rightarrow\Zen(\NG)\] such that for
$a,b\in \NG$ and $c\in\mathcal{Z}{(\NG)}$, we have
\begin{itemize}
\item $\tr^u_G% _{\NG(ab)=\tr^u_G% _{\NG(ba)$;
\item $\tr^u_G% _{\NG(c)=c$;
\item $\tr^u_G% _{\NG(a)\in (\Zen(\mathcal{N}(G)))^+$ if
$a\in (\mathcal{N}(G))^+$.
\end{itemize}
The trace can be extended to $\Mat_d(\NG)$ by taking
$\tr^u_G% _{\NG:=\tr^u_G% _{\NG\otimes\tr_{\Mat_d(\field{C})}$
(by abuse of notation), with $\tr_{\Mat_d(\field{C})}$ the non-normalized
trace on $\Mat_d(\field{C})$.
If $P\in \Mat_d(\NG)$ is a projector with image the (Hilbert
$\NG$-module) $V$, set $\dim^u_G% _{\NG(V):=\tr^u_G% _{\NG(P)$.
\end{definition}
That a unique such trace exists is established e.g.~in
\cite{Kadison2}*{Chapter 8}.
Later, we want to apply the trace also for the division closure. Recall that
we have (by definition) the following diagram of inclusions of rings
\begin{equation*}
\begin{CD}
K[G] @>>> \NG\\
@VVV @VVV \\
D(K[G]) @>>> \mathcal{U}(G).
\end{CD}
\end{equation*}
Given a finitely presented $K[G]$-module $M$, represented by $A\in \Mat_{k\times
l}(K[G])$, i.e.~with exact sequence $K[G]^l\xrightarrow{A} K[G]^k \to M\to 0$, the
induced modules $M\otimes_{K[G]} \NG$, $M\otimes_{K[G]}\mathcal{U}(G)$, $M\otimes_{K[G]}D(K[G])$
are also finitely presented with the same presenting matrix $A$. The standard
theory of Hilbert $\NG$-modules gives a center-valued dimension for each
finitely presented $\NG$-module, in particular for $M\otimes_{K[G]} \NG$, and
$\dim^u_G% _{\NG(M\otimes_{K[G]} \NG) =k-\dim^u_G% _{\NG(\ker(A))$ in the above situation (compare
\cite{MR1474192}). In \cite{MR1885124}, this dimension is extended to finitely
presented $\mathcal{U}(G)$-modules, of course in such a way that the value is unchanged
if we induce up from $\NG$ to $\mathcal{U}(G)$. More precisely, \cite{MR1885124} describes
the extension of dimensions based on arbitrary $\mathbb{C}$-valued traces on
$\NG$, this implies easily the corresponding extension for $\dim^u_G% _{\NG$.
\subsubsection*{The central idempotent division closure $\mathcal{E}(K[G])$}
\begin{definition}\label{def:EKG}
Let $R$ be a subring of the ring $S$ and let $C = \{e \in S \mid e$ is a
central idempotent of $S$ and $eS = rS$ for some $r \in R\}$. Then we
define
\[
\mathcal{C}(R,S) = \sum_{e\in C} eR,
\]
a subring of $S$. In the case $S = \mathcal{U}(G)$, we write
$\mathcal{C}(R)$ for $\mathcal{C}(R,\mathcal{U}(G))$. For each ordinal
$\alpha$, define $\mathcal{E}_{\alpha}(R,S)$ as follows:
\begin{itemize}
\item $\mathcal{E}_0(R,S) = R$;
\item $\mathcal{E}_{\alpha+1}(R,S) =
\mathcal{D}(\mathcal{C}(\mathcal{E}_{\alpha}(R,S),S),S)$;
\item $\mathcal{E}_{\alpha}(R,S) = \bigcup_{\beta < \alpha}
\mathcal{E}_{\beta}(R,S)$ if $\alpha$ is a limit ordinal.
\end{itemize}
Then $\mathcal{E}(R,S) = \bigcup_{\alpha} \mathcal{E}_{\alpha}(R,S)$. Also
in the case $R = K[G]$ where $G$ is a group and $K$ is a subfield of
$\mathbb{C}$, we write $\mathcal{E}(K[G])$ for
$\mathcal{E}(K[G],\mathcal{U}(G))$.
\end{definition}
\begin{conjecture}
Let $G$ be a discrete group and $K\subset\mathbb{C}$ a subfield. Then
$D(K[G])=\mathcal{E}(K[G])$, at least if $\lcm(G)<\infty$.
\end{conjecture}
We cite some properties of $\mathcal{E}(K[G])$ from \cite{L+S2} which will be useful
later. Indeed, we generalize from the canonical trace to the center-valued
trace, but the proofs literally also cover this more general situation.
\begin{lemma}\label{lem:range_of_dim_on_EKG}
(cf.~\cite{L+S2}*{Lemma 2.4})
The following additive subgroups of $\Zen(\NG)$ coincide:
\begin{multline*}
\langle \dim^u_G% _{\NG(x\mathcal{U}(G)^n)\mid x\in \Mat_n(K[G]), n\in\mathbb{N}\rangle\\
=
\langle \dim^u_G% _{\NG(x\mathcal{U}(G)^n)\mid x\in \Mat_n(\mathcal{E}(K[G])), n\in\mathbb{N}\rangle
\end{multline*}
\end{lemma}
This has as an immediate corollary that $\mathcal{E}(K[G])=D(K[G])$ if $K[G]$ satisfies the
center-valued Atiyah conjecture:
\begin{proof}[Proof of Theorem \ref{theo:EKG_equal_DKG}]
Let $e\in \mathcal{E}(K[G])$ be a central idempotent of $\mathcal{U}(G)$. Then all the spectral
projections of $e$ lie in $\Zen(\NG)$, therefore $e$ is affiliated to
$\Zen(\NG)$. Being an idempotent, even $e\in\Zen(\NG)$. Therefore, on the
one hand,
$\tr^u_G% _{\NG(e)=e$ while, on the other hand by Lemma \ref{lem:range_of_dim_on_EKG},
$\tr^u_G% _{\NG(e)=\dim^u_G% _{\NG(e\mathcal{U}(G))\in \remts{L_K}(G)$, in particular $e\in
\Zen(K[\Delta^+])\subset K[\Delta^+]$.
\end{proof}
\begin{remark}
The proof just given didn't need the full force of the center-valued Atiyah
conjecture, only the statement that $\dim^u_G% _{\NG(x\mathcal{U}(G)^n)\in
\Zen(\NG)$ is supported only on elements of finite order, i.e.~lies in
$\Zen(K[\Delta^+])$.
\end{remark}
\subsubsection*{Approximation of the center-valued trace}
The following is a special case of
\cite{Knebusch}*{Theorem 3.2} which will be used in the next section.
\begin{theorem}\label{theo:approxi}
Let $G$ be a discrete group with a sequence $G=G_0\geq G_1
\geq \cdots$ of
normal subgroups with $\bigcap_{i\in \mathbb{N}} G_i=\{1\}$.
Let $A\in \Mat_d(\overline{\field{Q}}[G])$ and $g\in\Delta(G)$. Let
$A[i]\in \Mat_d(\overline \mathbb{Q} [G/G_i])$ be the image
of $A$ under the map induced by the projection
$\pr_i\colon G\to G/G_i$.
Assume that all $G/G_i$ satisfy the determinant bound
property \cite{DLMSY}*{Definition 3.1}, e.g.~are
elementary amenable (or more
generally belong to the class $\mathcal{G}$ of groups
introduced in \cite{DLMSY}*{Definition 1.8} and corrected
in the errata to \cite{MR1828605} at arXiv:math/9807032,
or are
sofic, compare \cite{Elek+Szabo} and
\cite{Knebusch}*{Theorem 4.1}). Then
\begin{equation*}
\lim_{i\to\infty}\langle
\dim^u_{\mathcal{N}(G/G_i)}(\ker(A[i]))
,\pr_i(g) \rangle_{l^2(G/G_i)}=\langle
\dim^u_G% _{\NG\ker(A),g\rangle_{l^2(G)}.
\end{equation*}
\end{theorem}
\subsubsection*{Linnell's class $\mathfrak{C}$}
\begin{definition}\label{def:CCC} Let $\mathfrak{C}$ denote the smallest class of groups which
\begin{enumerate}
\item contains all free groups,
\item is closed under directed unions,
\item satisfies $G\in \mathfrak{C}$ whenever $H\lhd G$\,, $H\in\mathfrak{C}$ and $G/H$ is elementary amenable.
\end{enumerate}
\end{definition}
\section{Reformulation of the center-valued Atiyah conjecture}
Let $G$ be a group with $\lcm(G) <\infty$. We shall assume
that $K$ is
a subfield of $\mathbb{C}$ which is closed under complex
conjugation. Many of the arguments given below don't require this
assumption; however if $K$ is a subfield closed under complex
conjugation and $e$ is a central idempotent in $K[G]$, then $e$ is a
projection \cite{BPR10}*{Lemma 9.2(i)}. Furthermore if $G$ is a
finite group and $A \in \Mat_n(K[G])$, then $\pr_{\ker A} \in
\Mat_n(K[G])$ (use \cite{BPR10}*{Proposition 9.3});
it is here where we are using the property that $K$ is
closed under complex conjugation.
Recall that $\Delta^+$ is the (finite) normal subgroup consisting
of all elements of finite order and having only finitely many
conjugates.
\begin{lemma}\label{lemma:xyz} Let $K\subset\mathbb{C}$ be
a subfield which contains all $|\Delta^+|$-th roots of 1, and let
$c_G$ denote the number of finite conjugacy classes of elements of
finite order in $G$,
i.e.~the dimension of $\Zen(\NG)\cap \Zen(K[\Delta^+])$.
There is a finite set of primitive central projections
$\{U^1,\dots,
U^{c_G}\}$
of $\Zen(\NG)\cap \Zen(K[\Delta^+])\subset \Zen(K[G])$, given by
\begin{equation*}
U^i:=\sum_{ k \text{ s.t. }\exists g\in G:
gu_ig^{-1}=u_k }u_k,
\end{equation*}
where $u_i$ are the primitive central idempotents of the semisimple
Artinian ring $K[\Delta^+]$.
Furthermore $u_i=\frac{n_i}{|\Delta^+|}\sum_{s\in G}
\chi_i(s^{-1})s$,
with $n_i$ the dimensions of the irreducible representations (over
$\mathbb{C}$) of $\Delta^+$ and $\chi_i$ the corresponding characters
(extended by $0$ to all of $G$).
Moreover the $U^j$ form
an orthogonal basis of the vector space $\Zen(\NG)\cap
\Zen(\mathbb{C}[\Delta^+])$.
\end{lemma}
\begin{proof}
By Maschke's theorem of standard representation theory, the algebra
$K[\Delta^+]$ is semisimple
Artinian, compare \cite{Lang1}*{XVIII, Theorem 1.2}. Therefore it has finitely
many primitive central idempotents $u_i$.
Any algebra automorphism must permute the $u_i$,
in particular the conjugation action of $G$. An element of $\Zen(K[\Delta^+])$
belongs to the
center of $K[G]$ (and then also of $\NG$) if and only if it is invariant under
conjugation by elements of $G$. It follows
immediately that the $U^i$ are the primitive central idempotents of
$\Zen(\NG)\cap \Zen(K[\Delta^+])$, and furthermore they form an
orthogonal basis for $\Zen(\NG)\cap
\Zen(\mathbb{C}[\Delta^+])$.
The formula for the $u_i$
is also standard, \cite{Lang1}*{XVIII, Proposition 4.4 and Theorem
11.4}.
\end{proof}
\begin{lemma}\label{lemma:centralidempotents}
Let $K$ be a subfield of $\mathbb{C}$ and let $L/K$ be a finite
Galois extension of $K$ with Galois group $F$. Let $G$ be a finite
group, let $\{e_1,\dots,e_n\}$ denote the primitive central
idempotents of $K[G]$, and let $\{u_1, \dots, u_m\}$ denote the
primitive central idempotents of $L[G]$. Then $F$ acts as
automorphisms on $L[G]$ according to the rule $\theta \sum_{g \in G}
a_gg = \sum_{g \in G}\theta(a_g)g$ for $\theta \in F$.
The $u_i$ form an orthogonal set and $\langle
u_i,1\rangle = \langle \theta u_i,1\rangle$ for all $i$. For each
$i$, define $N_i = \{j \in \mathbb{N} \mid e_iu_j = u_j\}$. Then $F$
acts transitively on $\{u_j \mid j \in N_i\}$ and $e_i =
\sum_{j \in N_i} u_j$.
\end{lemma}
\begin{proof}
This is well-known, and follows from Galois descent.
Note that $u_ie_j$ is a central idempotent in $L[G]$ and $u_i = u_ie_j
+ (1-e_j)u_i$. It follows for all $i,j$, either $u_ie_j = 0$ or
$u_ie_j = u_i$, because $u_i$ is primitive. It follows easily that
$e_i = \sum_{j\in N_i} u_j$. Also $F$ acts on $\{u_j \mid j\in
N_i\}$, and the sum of the $u_j$ in an orbit is fixed by $F$ and is
therefore in $K[G]$. Since $e_i$ is primitive, it follows that this
orbit must be the whole of $N_i$. Finally if $e = \sum_{g\in G}
e_gg \in L[G]$ is an idempotent, then $e_1 \in \mathbb{Q}$ (by the character
formula of Lemma \ref{lemma:xyz}) and we see
that $\langle u_i,1\rangle = \langle \theta u_i,1\rangle$ for all $i$.
\end{proof}
\begin{lemma}\label{lemma:xyz1} Let $K\subset\mathbb{C}$ be a subfield,
let $\omega$ be a primitive $|\Delta^+|$-root of 1 and
set $L = K(\omega)$. Let $F$ denote the Galois group of $L$ over
$K$, and let $U^1,\dots, U^{c_{L[G]}}$ be the primitive central
projections of $\Zen(\NG)\cap \Zen(L[\Delta^+])\subset \Zen(L[G])$ as
described above in Lemma~\ref{lemma:xyz}. There is a finite set of
primitive central projections
$\{P^1,\dots,
P^{C_{K[G]}}\}$
of $\Zen(\NG)\cap \Zen(K[\Delta^+])\subset \Zen(K[G])$, given by
\begin{equation*}
P^i:=\sum_{ k \text{ s.t. }\exists g\in G:
gp_ig^{-1}=p_k }p_k,
\end{equation*}
where $p_i$ are the primitive central idempotents of the
semisimple
Artinian ring $K[\Delta^+]$. Set $N_i = \{j \in \mathbb{N} \mid P^i
U^j = U^j\}$. Then
\[
P^i = \sum_{j \in N_i} U^j
\]
and $F$ acts transitively on $\{U^j \mid j \in N_i\}$.
\end{lemma}
\begin{proof}
This follows from Lemmas \ref{lemma:xyz} and
\ref{lemma:centralidempotents}.
\end{proof}
\begin{lemma}\label{lemma:Atiyah}
Let $H$ be a finite subgroup of $G$ which contains $\Delta^+$.
For an irreducible projection $Q\in K[H]$ (in the sense that if
$Q=Q_1+Q_2$ with projections in $Q_1,Q_2\in K[H]$ satisfying
$Q_1Q_2=0$ then either $Q_1=0$ or $Q_2=0$) we have $\tr^u_G% _{\NG(Q)\in
\Zen(\NG)\cap \Zen(K[\Delta^+])\subset \Zen(\NG)$. More precisely, using
the central projections $P^i$ of Lemma \ref{lemma:xyz1} we have
\begin{equation}\label{eq:Atiyah}
\tr^u_G% _{\NG(Q)
\frac{\dim_\mathbb{C}(Q\cdot
\mathbb{C} [H])\cdot |\Delta^+|}{|H|\cdot
\dim_\mathbb{C}(P^i\cdot
\mathbb{C}[\Delta^+])}P^i
=\frac{\dim_{\NG}(Q\cdot l^2(G))}{\dim_{\NG}(P^i\cdot l^2(G))}P^i
\end{equation}
where $P^i$ is characterized by the
property $QP^i=Q$.
\end{lemma}
\begin{proof}
Let $\omega$ be a primitive $|\Delta^+|$-th root of 1, let $L =
K(\omega)$ and let $F$ denote the Galois group of $L/K$.
The center-valued trace is obtained by orthogonal projection from
$l^2(G)$ to the subspace of $l^2(\Delta)$ spanned by functions which
are constant on $G$-conjugacy classes, using the standard embedding of
$\NG$
into $l^2(G)$. For $Q$, which is supported on group
elements of finite order, therefore $\tr^u_G% _{\NG(Q)\in \mathbb{C}[\Delta^+]$.
Let $U^1,\dots, U^{c_G}$ and $P^1,\dots,P^{C_{K[G]}}$ be the primitive
central projections as described in Lemma \ref{lemma:xyz1}.
Using the standard inner product on $\mathbb{C} [H]$ we obtain,
using that $(U^1,\dots, U^{c_G})$ is an orthogonal basis of
$\Zen(\NG)\cap
\Zen(L[H]) = \Zen(\NG)\cap \Zen(L [\Delta^+])$
\begin{equation}\label{eq:trGQ}
\tr^u_G% _{\NG(Q)=\sum_j \frac{\innerprod{Q,U^j}}{\innerprod{U^j,U^j}} U^j.
\end{equation}
Moreover, we have for each $j$ that $QP^j+Q(1-P^j)=Q$ and
$QP^jQ(1-P^j)=0$, the latter because $P^j$ is central. Since
$Q$ is irreducible,
we get either $QP^j=Q$ or $QP^j=0$.
If $QP^i=Q$ we have $Q\sum_{j \in N_i}U^j = Q$ and $QU^j = 0$ for $j
\notin N_i$. Also if $j_1,j_2 \in N_i$, then $\theta (QU^{j_1}) =
QU^{j_2}$ for some $\theta \in F$ and we see that $\langle QU^{j_1},
1\rangle =\langle QU^{j_2},1\rangle$, consequently $\langle
Q,U^j\rangle$ is independent of $j$ for $j\in N_i$. Similarly
$\langle U^j,U^j\rangle$ is independent of $j$ for $j \in N_i$.
Thus $\langle Q,P^i\rangle = |N_i| \langle Q,U^j\rangle$,
$\langle P^i,P^i \rangle = |N_i| \langle U^j,U^j\rangle$ for $j \in
N_i$, hence
\[
\frac{\innerprod{Q,U^j}}{\innerprod{U^j,U^j}} =
\frac{\innerprod{Q,P^i}}{\innerprod{P^i,P^i}}.
\]
Substitute this in equation \eqref{eq:trGQ} together with
\begin{align*}
\langle Q,P^j\rangle & =\langle QP^j,1\rangle =\langle
Q,1\rangle =\frac{\dim_\field{C}(Q \cdot \mathbb{C}[H])}{|H|}\\
\langle P^j,P^j\rangle &
= \innerprod{P^j,1}=\frac{\dim_\field{C}(P^j \cdot \mathbb{C}[\Delta^+])}{|\Delta^+|}.
\end{align*}
These formulas follow from the character formula for
projections or are directly obtained as follows: for a
projection $P\in\mathbb{C} [E]$
and a finite group $E$ we have $\innerprod{P,1}_{l^2(E)}
=\innerprod{Ph,h}_{l^2(E)}$ for all $h\in E$, therefore
$\dim_{\mathbb{C}}(P\cdot \mathbb{C} [E])=\tr(P)=\sum_{h\in E}
\innerprod{Ph,h}=\abs{E}\cdot \innerprod{P,1}$.
Note, finally, that $\frac{\dim_\mathbb{C}(Q\cdot \mathbb{C}
[H])}{\abs{H}} = \dim_{\mathcal{N}(H)}(Q\cdot l^2(H)) = \dim_{\NG}(Q\cdot
l^2(G))$ by the induction rule for von Neumann dimensions.
\end{proof}
\begin{corollary}\label{corol:discrete}
The additive subgroup $\remts{L_K}(G)$ of $\Zen(\NG)$ of Definition
\ref{def:LG_and_Atiyahconj} is discrete.
\end{corollary}
\begin{proof}
\remts{Recall that $F$ denotes the relevant subfield of $\mathbb{C}$ in the
setup of Definition \ref{def:LG_and_Atiyahconj}}, namely $F$ is
the field of fractions of $K$. Given a finite subgroup $H$ of $G$ and a \remts{projection $P\in F[H]$},
$\tr^u_G% _{\NG(P)$ is a positive integral linear combination of $\tr^u_G% _{\NG(Q_\alpha)$
where $Q_\alpha \in \remts{F[H]}$ are irreducible projections, corresponding to
the decomposition of $\im(P)$ into irreducible $\remts{F[H]}$-modules.
It therefore suffices to check that the additive subgroup of
$\Zen(\NG)$
generated by $\tr^u_G% _{\NG(Q)$ is discrete, where $Q$ runs through the
irreducible projections in $\remts{F[H]}$ and $H$ runs through the finite
subgroups of $G$. Increasing the field and increasing the finite
subgroup has the only potential effect that an irreducible projection
breaks up as a sum of new irreducible projections and therefore the
subgroup generated by their center-valued
traces increases. Therefore we may assume that these subgroups
contain $\Delta^+$ and that $\remts{F = \mathbb{C}}$.
By Lemma \ref{lemma:Atiyah}, these are all integer
multiples of $\lcm(G)^{-1}P^i$ with the orthogonal basis
$(P^1,\dots,
P^{c_G})$, therefore span a discrete subgroup of $\Zen(\NG)$.
\end{proof}
\begin{definition}\label{def:Atiyah-Artinean}
Assume that $G$ is a discrete group with $\lcm(G)<\infty$ and that
$K$ is a subfield of $\mathbb{C}$ which is closed under complex
conjugation.
We say that $D(K[G])$ is \emph{Atiyah-expected Artinian} if it is a
semisimple Artinian ring such that its primitive central idempotents are
the central idempotents $P^1,\dots,P^{C_{K[G]}} \in
K[\Zen(K[\Delta^+])]$ of Lemma
\ref{lemma:xyz1},
and if each direct summand $P^jD(K[G])P^j$ is an $L_j\times L_j$ matrix
ring over
a skew field.
Here, $L_j$ is determined as follows: consider all irreducible
sub-projections $Q_\alpha\in K[H_\alpha]$ of
$P^j$ (i.e.~those satisfying $Q_\alpha P^j=Q_\alpha$), where $H_\alpha$ runs
through all finite subgroups of $G$ containing $\Delta^+$.
By Lemma \ref{lemma:Atiyah},
$\tr^u_G% _{\NG(Q_\alpha)=q_\alpha P^j$ for some rational number $q_\alpha$. Because
there are only
finitely many isomorphism classes of finite subgroups of $G$,
formula \eqref{eq:Atiyah} shows that the collection of these rational
numbers is finite. $L_j$ is the smallest integer such that each
$q_\alpha$ is
an integer multiple of $\frac{1}{L_j}$. Explicitly,
\begin{equation*}
L_j=\frac{\dim_\field{C}(P^j\cdot
\mathbb{C}[\Delta^+])\lcm(G)}{
\gcd\left( \dim_\field{C}(P^j\cdot
\mathbb{C}[\Delta^+])\lcm(G),
\dim_\mathbb{C}(Q_\alpha\cdot\mathbb{C}
[H_\alpha])\frac{\lcm(G)}{|H_\alpha|} \abs{\Delta^+}\mid
\alpha\right) } \in\Z.
\end{equation*}
\end{definition}
\begin{proof}
We have to show that the two descriptions of $L_j$ coincide, using
Equation \eqref{eq:Atiyah}, i.e.~we have to find the smallest common
denominator of all these fractions. We expand the denominators to the common
value $\lcm(G)\cdot \dim_\mathbb{C}(P^j\cdot \mathbb{C}[\Delta^+])$, then we
have to divide this by the greatest common divisor of this number and of
all the new numerators.
\end{proof}
\begin{theorem}\label{con:at}
Let $G$ be a discrete group, with $\lcm(G) < \infty$ and let $K\subset
\field{C}$ be a subfield closed under complex conjugation.
The following statements are equivalent.
\begin{enumerate}
\item\label{item:At-art} $D(K[G])$ is Atiyah-expected Artinian as in
Definition \ref{def:Atiyah-Artinean}.
\item\label{item:KKG} $\phi\colon \bigoplus_{E\le G\,:\,|E|<\infty} K_0(K[E])
\rightarrow K_0(D(K[G]))$ is surjective and $D(K[G])$ is semisimple Artinian.
\item\label{item:GKG} $\phi\colon\bigoplus_{E\le G\,;\,|E|<\infty}
G_0(K[E])\rightarrow G_0(D(K[G]))$ is surjective.
\item\label{item:quant}
$KG$ satisfies the center-valued Atiyah conjecture.
\end{enumerate}
Recall here that, for a ring $R$, $K_0(R)$ is the Grothendieck group
of finitely generated projective $R$-modules, whereas $G_0(R)$ is
the Grothendieck group of arbitrary finitely generated
$R$-modules.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{con:at}]
\ref{item:At-art} $\implies$ \ref{item:KKG}: We use the notation of
Definition \ref{def:Atiyah-Artinean}. Using the row projectors of
matrix rings, there are projections
$x_1,\dots,x_{C_{K[G]}}\in
D(K[G])$ which represent a $\mathbb{Z}$-basis of the free abelian group
$K_0(D(K[G]))$, and $[P^i]=L_i[x_i]$ in $K_0(D(K[G]))$. We have to show
that each $x_i$ is an integer linear combination of images of elements
of $K_0(K[H_\alpha])$ with $H_\alpha$ finite. If $Q_\alpha\in K[H_\alpha]$
is an irreducible sub-projection of $P^i$, then $\phi([Q_\alpha])$ is a
multiple of $[x_i]$ in $K_0(D(K[G]))$, namely (comparing the
center-valued dimensions which are defined for finitely generated
projective $D(K[G])$-modules by the discussion of Section
\ref{sec:prelim}) $\phi([Q_\alpha])=q_\alpha [P^i]$ if
$\tr^u_G% _{\NG(Q_\alpha)=q_\alpha P^i$. By the Chinese remainder theorem and
the definition of $L_i$ as the smallest integers such that all the
$q_\alpha$ are integer multiples of $L_i^{-1}$, also
$[x_i]=L_i^{-1}[P^i]$ belongs to the image of $\phi$.
\ref{item:KKG} $\implies$ \ref{item:GKG}: For a semisimple Artinian ring
every finitely generated module
is projective, therefore $G_0 = K_0$ under the assumptions we make.
\ref{item:GKG} $\implies$ \ref{item:quant}: Let $M$ be a finitely
presented $K[G]$-module with presentation $K[G]^l\xrightarrow{A} K[G]^n\to
M\to 0$, $A\in M_{n\times l}(K[G])$. Then $M\otimes_{K[G]}
D(K[G])$ is finitely generated, therefore by the assumption stably
isomorphic to an
integer linear combination $\bigoplus a_i x_iD(K[G])$ with $x_i$
projectors defined over finite subgroups $E$ of $G$ --- note that
$G_0(K[E])=K_0(K[E])$ for any finite group $E$, as $K[E]$ is semisimple
Artinian. Inducing further
to $\mathcal{U}(G)$ and using that the dimension function extends to finitely
presented $\mathcal{U}(G)$-modules (which is additive, so that we can leave out
the stabilization summands),
we read off that
\begin{equation*}
\dim^u_G% _{\NG(M)=\dim^u_G% _{\NG(\bigoplus a_i x_i\mathcal{U}(G))
=\sum a_i
\dim^u_G% _{\NG(x_i\mathcal{U}(G)) \in \remts{L_K}(G)
\end{equation*}
by definition of $\remts{L_K}(G)$. Finally, by additivity of the von Neumann dimension $\dim^u_G% _{\NG(\ker(A)) =
n-\dim^u_G% _{\NG(M)\in \remts{L_K}(G)$.
\ref{item:quant} $\implies$ \ref{item:At-art}: Here, we follow closely the argument of the proof of
\cite{L+S2}*{Proposition 2.14}.
Our assumption implies by Theorem \ref{theo:EKG_equal_DKG} that
$\mathcal{E}(K[G])=D(K[G])$. Because the center-valued Atiyah conjecture implies that
the ordinary $L^2$-Betti numbers are contained in a finitely generated
subgroup of $\mathbb{Q}$ (generated by $\tr_G(P^j)/L_j$), by
\cite{L+S2}*{Theorem 2.7} $D(K[G])$ is a semisimple Artinian ring.
The $P^j$ are central idempotents in $D(K[G])$. We have
to show that they are primitive central idempotents, and that each is
the sum of exactly $L_j$ orthogonal sub-idempotents which are
themselves irreducible. The structure theory of rings then implies
that each $P^jD(K[G])P^j$ is simple Artinian and an $L_j\times
L_j$-matrix ring over a skew field.
Fix, as in Definition \ref{def:Atiyah-Artinean}, the (finite) collection of
sub-projections $Q_\alpha$ of $P^j$,
where the $Q_\alpha$ are irreducible projections supported on
$K[H_\alpha]$ and $H_\alpha$ runs through the (isomorphism classes of)
finite extensions of $\Delta^+(G)$ inside $G$. Then
$\tr^u_G% _{\NG(Q_\alpha)=\frac{n_\alpha}{L_j} P^j$ with integers $n_\alpha$, and by
definition of $L_j$ we have $\gcd_\alpha(n_\alpha)=1$. Set
$d:=\lcm_\alpha(n_\alpha)$.
Consider now $P^j\mathcal{U}(G)^d$. Because
$$\dim^u_G% _{\NG(P^j\mathcal{U}(G)^d)=dP^j=\dim^u_G% _{\NG(Q_\alpha\mathcal{U}(G)^{L_jd/n_\alpha})$$ by
\cite{Lueck02}*{Theorem 9.13(1)} then $P^j\mathcal{U}(G)^d\cong
Q_\alpha\mathcal{U}(G)^{L_jd/n_\alpha}$, so we find $L_jd/n_\alpha$ mutually orthogonal
projections in $\Mat_d(\mathcal{U}(G))$ corresponding to the copies of $Q_\alpha$. Because
the center-valued trace of each of those equals
$\frac{n_\alpha}{L_j}P^j=\tr^u_G% _{\NG(Q_\alpha)$, by \cite{Sterling}*{Exercise
13.15A}, there exist $L_jd/n_\alpha$
similarities (i.e.~self-adjoint unitaries)
$u_i \in \mathcal{U}(G)$ with $u_1 = 1$ such that these projections can be
written as $u_iP'_\alpha u_i$ (where $P_\alpha'$ is the diagonal matrix with
first entry $P_\alpha$ and all other entries $0$).
Then, exactly as in the proof of \cite{L+S2}*{Proposition 2.14} we can replace
the $u_i$ by $\tilde u_{i}\in \Mat_d(D(K[G]))$ which are invertible and such that we
still have a direct sum decomposition
\begin{equation}\label{eq:dirsum}
P^jD(K[G])^d = \bigoplus_{i=1}^{L_jd/n_\alpha} \tilde u_{i} P_\alpha' D(K[G])^d.
\end{equation}
This uses
the Kaplansky density theorem, the quantization of the center-valued trace and
\cite{L+S2}*{Lemma 2.12}.
Let us now take a central idempotent $\epsilon$ in $D(K[G])$ which is a sub-projection of
$P^j$ (i.e.~$\epsilon P^j=\epsilon$). We have to show that $\epsilon=0$ or
$\epsilon=P^j$. To do this, we
compute $\tr^u_G% _{\NG(\epsilon)$. Note that all the modules $\epsilon \tilde
u_iP_\alpha'\mathcal{U}(G)^d$ are isomorphic, therefore by Equation \eqref{eq:dirsum}
\begin{equation}\label{eq:tru_eps}
d\tr^u_G% _{\NG(\epsilon)=\dim^u_G% _{\NG(\epsilon\mathcal{U}(G)^d) = \frac{L_jd}{n_\alpha} \dim^u_G% _{\NG(\epsilon
P_\alpha' \mathcal{U}(G)^d).
\end{equation}
By Lemma \ref{lem:range_of_dim_on_EKG} and the assumption \ref{item:quant},
$L_j\cdot \dim^u_G% _{\NG(\epsilon P_\alpha'\mathcal{U}(G)^d)$ is an integer multiple of
$P^j$. Therefore, rearranging Equation \eqref{eq:tru_eps}
\begin{equation*}
n_\alpha\tr^u_G% _{\NG(\epsilon) \in \mathbb{Z} P^j.
\end{equation*}
As this holds for all $\alpha$, even
\begin{equation*}
\epsilon= \tr^u_G% _{\NG(\epsilon) =\lcm_\alpha(n_\alpha)\tr^u_G% _{\NG(\epsilon)\in \mathbb{Z} P^j.
\end{equation*}
So we can indeed conclude that $P^j$ is a primitive central
idempotent and
therefore $P^jD(K[G])$ is an $l\times l$ matrix ring over a skew field. It
follows that $P^jD(K[G])^{n_\alpha d}$ is the direct sum of $n_\alpha d l$
copies of an irreducible submodule. On the other hand, $P^j
D(K[G])^{n_\alpha d}$ is the direct sum of $L_jd$ isomorphic summands
for every $\alpha$. As
$\lcm_\alpha(n_\alpha)=1$ we conclude that $L_j\mid l$. On the other hand,
by the assumption \ref{item:quant} and Lemma \ref{lem:range_of_dim_on_EKG},
the center-valued dimension of the irreducible submodule (which is generated
by one projector as $P^jD(K[G])$ is Artinian) is an integer multiple of
$L_j^{-1}P^j$ and therefore $L_j\mid l$. It follows that $l=L_j$ as
claimed.
\end{proof}
\section{Special cases and inheritance properties of the center-valued Atiyah conjecture}
Throughout this section, we assume that $K$ is a subfield of
$\mathbb{C}$ which is closed under complex conjugation.
\begin{lemma}\label{lem:finite}
The center-valued Atiyah conjecture is true for finitely generated
virtually free groups.
\end{lemma}
\begin{proof}
This follows from the proof of \cite{MR1242889}*{Proposition 5.1(i)
and Lemma 5.2(ii)} (in
which $\mathbb{C}$ can be replaced by any subfield of $\mathbb{C}$)
and Theorem \ref{con:at}\ref{item:KKG}.
\end{proof}
\begin{lemma}\label{prop:union}
If $G$ is a directed union of groups $G_i$ and the center-valued Atiyah
conjecture over $K$ is true for all groups $G_i$, then it is also true for
$G$.
\end{lemma}
\begin{proof}
By \cite{MR1242889}*{Lemma 5.3}, $D(K[G])$ is the directed union of the
$D(K[G_i])$. Any matrix $A$ over $D(K[G])$ is therefore already a matrix over
$D(K[G_i])$ for some $i$, with $\dim^u_{G_i}(\ker(A))\in \remts{L_K}(G_i)$. Composition
with the center-valued trace for $G$ gives (by the induction formula for von
Neumann dimensions) $\dim^u_G% _{\NG(\ker(A)) \in \tr^u_G% _{\NG(\remts{L_K}(G_i))\subset \remts{L_K}(G)$.
\end{proof}
\begin{proposition}\label{prop:extension}
Assume that we have an extension $1\to H\to G\xrightarrow{\pi} E\to 1$ where
$E$ is
elementary amenable and for each finite subgroup $F\leq E$,
$\pi^{-1}(F)\leq G$ satisfies the center-valued Atiyah conjecture over
$K$ . Then
also $K[G]$ satisfies the center-valued Atiyah conjecture.
\end{proposition}
\begin{proof}
By transfinite induction, the statement is a formal consequence of the same
assertion where $E$ is finitely generated virtually abelian, as explained
e.g.~in the proof of \cite{Schick}*{Proposition 3.1} or in \cite{MR1242889}.
If $E$ is finitely generated virtually abelian then in the proof of
\cite{MR1242889}*{Lemma 5.3} it is shown that
$$\bigoplus_{F\leq E \text{ finite}} G_0(D(K[\pi^{-1}(F)]))\to
G_0(D(K[G]))$$
is onto, using Moody's induction
theorem \cite{Moody}*{Theorem 1}. Since by assumption
$\bigoplus_{U\leq
\pi^{-1}(F)\text{ finite}}G_0(K[U])\to G_0(D(K[\pi^{-1}(F)]))$ is
onto for each such $F$ and the composition of surjective maps is
surjective we conclude that
$$\bigoplus_{F\in \mathcal{F}(G)} G_0(K[F])\to G_0(D(K[G]))$$
is onto and \ref{item:GKG} of Theorem \ref{con:at} is established.
\end{proof}
\begin{proposition}\label{prop:inverse_limit}
Let $K$ be a subfield of $\overline{\mathbb{Q}}$ which is closed
under complex conjugation.
Assume that $G$ is a group with a sequence $G \geq
G_1\geq \cdots$ of
normal subgroups such that $\bigcap_{i\in\mathbb{N}} G_i=\{1\}$. Assume
moreover that for each $i\in \mathbb{N}$ and each finite subgroup $F\leq
G/G_i$ there is a finite subgroup $F'\leq G$ which is mapped
isomorphically to $F$ by the projection $G\to G/G_i$.
Finally, assume that each $G/G_i$ satisfies the determinant bound conjecture
and the center-valued Atiyah
conjecture over $K$. Then $K[G]$ satisfies the center-valued Atiyah
conjecture.
\end{proposition}
\begin{proof}
As the statement is empty if $\lcm(G)=\infty$, we assume that
$\lcm(G)<\infty$.
We first show that, if $i$ is large enough,
$\pi_i$ induces an isomorphism $\pi_i\colon \Delta^+(G)\to
\Delta^+(G/G_i)$. Dropping finitely many terms in the sequence we can then
assume that this is the case for all $i\in\mathbb{N}$. To prove the
assertion, choose a finite
subgroup $M$ of $G$ with maximal order (possible since
$\lcm(G)<\infty$). Note that the product $\Delta^+M$ is also a finite
subgroup,
therefore by maximality equal to $M$, consequently $\Delta^+\leq M$. Then
choose finitely many
$g_1,\dots,g_n\in G$ such that $\Delta^+(G)=\bigcap_{k=1}^n
M^{g_k}$ (where $M^g$ denotes the conjugate $gMg^{-1}$),
which is possible by the descending chain condition for finite sets.
Finally, choose $r>0$ such that $\pi_r\colon G\to G/G_r$ is injective when
restricted to $\bigcup_{k=1}^n M^{g_k}$, which is possible because
$\bigcap_{i} G_i=\{1\}$.
Because $\pi_r$ is surjective, $\pi_r(\Delta^+(G))$ is a finite normal
subgroup
of $G/G_i$ and therefore $\pi_r(\Delta^+(G))\leq \Delta^+(G/G_r)$. On the
other hand, $\pi_r(M)$ is a finite subgroup with maximal order in $G/G_r$
(because $\pi_r|_M$ is injective and every finite subgroup of $G/G_r$ is an
isomorphic image of a finite subgroup of $G$), therefore
$\Delta^+(G/G_r)\leq \pi_r(M)$, by normality even
$\Delta^+(G/G_r)\leq \bigcap_{k=1}^n \pi_r(M)^{\pi_r(g)}$. As
$\bigcap_{k=1}^n M^g =\Delta^+(G)$ and by injectivity of $\pi_r$ on
$\bigcup_{k=1}^n M^g$ we finally get
\begin{equation*}
\Delta^+(G/G_r)\leq \bigcap_{k=1}^n \pi_r(M)^{\pi_r(g)} =
\pi_r(\Delta^+(G)) \leq \Delta^+(G/G_r).
\end{equation*}
This implies the statement for all $i\ge r$.
Secondly, given $g\in G$ of infinite order, for all sufficiently large $i$,
the restriction of $\pi_i$ to $\{1,g,g^2,\dots, g^{\lcm(G)}\}$ is injective
and therefore, as by assumption the orders of finite subgroups of $G/G_i$ are
bounded by $\lcm(G)$, $\pi_i(g)$ also has infinite order.
Fix now $A\in \Mat_d(K[G])$ and denote by $Q_i$ the projection onto the kernel
of $A[i]:=p_i(A)$. Recall that
\begin{equation*}
\tr^u_G% _{\NG(Q_i)=\dim^u_G% _{\NG(\ker(A))=\sum_{g\in
G}\innerprod{\dim^u_G% _{\NG(\ker(A)),g}_{l^2(G)}g,
\end{equation*}
and we denote by
$\innerprod{\dim^u_G% _{\NG(\ker(A)),g}$ the \emph{coefficient of $g$} in
$\dim^u_G% _{\NG(\ker(A))$, and correspondingly for $\tr^u_G% _{\NG(Q_i)$.
The center-valued Atiyah conjecture for $K[G/G_i]$ implies in particular that
$\tr^u_G% _{\NG(Q_i)$ is contained in $K[\Delta^+(G/G_i)]$, therefore supported only on
elements of finite order. Consequently, if $g\in G$ has infinite order, then
$\innerprod{\tr^u_G% _{\NG(Q_i),\pr_i(g)}=0$ for sufficiently large $i$ and, by Theorem
\ref{theo:approxi}, $\innerprod{\dim^u_G% _{\NG(\ker(A)),g}=0$. This implies that $\dim^u_G% _{\NG(\ker(A))$ is
supported on elements of finite order, i.e.~is contained in $\Zen(\NG)\cap
K[\Delta^+(G)]$.
As explained above, we can use $\pi_i$ to identify $\Delta^+(G)$ and
$\Delta^+(G/G_i)$ and consider $\tr^u_G% _{\NG(Q_i)$ as an element of $K[\Delta^+(G)]$. By
Theorem \ref{theo:approxi}, for each $g\in \Delta^+(G)$,
\begin{equation*}
\innerprod{\dim^u_G% _{\NG(\ker(A)),g} =\lim_{i\to\infty} \innerprod{\tr^u_G% _{\NG(Q_i),g}.
\end{equation*}
Since all the (finitely many) coefficients converge, we even have
\begin{equation*}
\lim_{i\to\infty} \tr^u_G% _{\NG(Q_i) = \dim^u_G% _{\NG(\ker(A))\in \Zen(\NG)\cap
K[\Delta^+(G)].
\end{equation*}
Because the sets of isomorphism classes of finite
subgroups
of $G/G_i$ and of $G$ are identified by $\pi_i$, we get exactly the
same relevant irreducible projections defined over finite subgroups and the
same central idempotents in the formulas of Lemma \ref{lemma:xyz} and Lemma
\ref{lemma:Atiyah} for $\remts{L_K}(G)$ and $\remts{L_K}(G/G_i)$. Consequently, $\pi_i$ identifies
$\remts{L_K}(G)$ and
$\remts{L_K}(G/G_i)$. Finally, observe that by assumption about the Atiyah
conjecture for $G/G_i$ we have $\tr^u_G% _{\NG(Q_i)\in \remts{L_K}(G)$. As the
latter is a discrete subset of $\Zen(\NG)$, we finally observe that
$\dim^u_G% _{\NG(\ker(A))\in \remts{L_K}(G)$, i.e.~$K[G]$ satisfies the center-valued Atiyah
conjecture.
\end{proof}
\begin{theorem}\label{theo:LiC}
The center-valued Atiyah conjecture is true for all groups $G\in\mathfrak{C}$.
\end{theorem}
\begin{proof}
In the proof of
\cite{MR1242889}*{Lemma 4.9} it is shown that the assertion
follows (by transfinite induction) directly from Lemma
\ref{lem:finite},
Lemma \ref{prop:union} and Proposition \ref{prop:extension}.
\end{proof}
\begin{corollary}\label{corollary}
Let $K$ be a subfield of $\overline{\mathbb{Q}}$ which is closed
under complex conjugation.
Then the center-valued Atiyah conjecture is true for all elementary amenable extensions of pure braid groups, of
right-angled Artin groups, of primitive link groups, of cocompact special
groups, or of products of such.
\end{corollary}
\begin{proof}
Each of the groups in the list has a sequence of normal subgroups with
trivial intersection and with elementary amenable quotients such that in
addition the condition of Proposition \ref{prop:inverse_limit} is met. This
is shown for the extensions of pure braid groups in \cite{L+S1}, for
primitive link groups in \cite{MR2415028} and for right-angled Coxeter and
Artin groups in \cite{LinnellOkunSchick}, and combining \cite{Schreve} with
\cite{LinnellOkunSchick} it also follows for special cocompact
groups. Combining Theorem \ref{theo:LiC}
and Proposition \ref{prop:inverse_limit}, the assertion
follows.
\end{proof}
\begin{bibdiv}
\begin{biblist}
\bib{Agol}{article}{
AUTHOR = {Agol, Ian},
TITLE = {The virtual {H}aken conjecture},
NOTE = {With an appendix by Agol, Daniel Groves, and Jason Manning},
JOURNAL = {Doc. Math.},
FJOURNAL = {Documenta Mathematica},
VOLUME = {18},
YEAR = {2013},
PAGES = {1045--1087},
ISSN = {1431-0635},
MRCLASS = {20F67 (57Mxx)},
MRNUMBER = {3104553},
MRREVIEWER = {Thomas Koberda},
}
\bib{arveson}{book}{
author={Arveson, William},
title={An invitation to $C\sp*$-algebras},
note={Graduate Texts in Mathematics, No. 39},
publisher={Springer-Verlag},
place={New York},
date={1976},
pages={x+106},
}
\bib{Atiyah}{article}{
author={Atiyah, Michael F.},
title={Elliptic operators, discrete groups and von Neumann algebras},
conference={
title={Colloque ``Analyse et Topologie'' en l'Honneur de Henri Cartan
(Orsay, 1974)},
},
book={
publisher={Soc. Math. France},
place={Paris},
},
date={1976},
pages={43--72. Ast\'erisque, No. 32-33},
review={\MR{0420729 (54 \#8741)}},
}
\bib{Austin}{article}{
author={Austin, Tim},
title={Rational group ring elements with kernels having irrational
dimension},
journal={Proc. Lond. Math. Soc. (3)},
volume={107},
date={2013},
number={6},
pages={1424--1448},
issn={0024-6115},
review={\MR{3149852}},
doi={10.1112/plms/pdt029},
}
\bib{Bergeron-Wise}{article}{
AUTHOR = {Bergeron, Nicolas and Wise, Daniel T.},
TITLE = {A boundary criterion for cubulation},
JOURNAL = {Amer. J. Math.},
FJOURNAL = {American Journal of Mathematics},
VOLUME = {134},
YEAR = {2012},
NUMBER = {3},
PAGES = {843--859},
ISSN = {0002-9327},
CODEN = {AJMAAN},
MRCLASS = {20F67},
MRNUMBER = {2931226},
DOI = {10.1353/ajm.2012.0020},
URL = {http://dx.doi.org/10.1353/ajm.2012.0020},
}
\bib{BPR10}{article}{
author={Birkenmeier, Gary F.},
author={Park, Jae Keol},
author={Rizvi, S. Tariq},
title={A theory of hulls for rings and modules},
conference={
title={Ring and module theory},
},
book={
series={Trends Math.},
publisher={Birkh\"auser/Springer Basel AG, Basel},
},
date={2010},
pages={27--71},
review={\MR{2744041 (2012c:16001)}},
doi={10.1007/978-3-0346-0007-1-2},
}
\bib{Sterling}{book}{
author={Berberian, Sterling K.},
title={Baer {$*$}-rings},
note={Die Grundlehren der mathematischen Wissenschaften, Band 195},
publisher={Springer-Verlag},
place={New York},
date={1972},
pages={xiii+296},
review={\MR{0429975 (55 \#2983)}},
}
\bib{MR2415028}{article}{
author={Blomer, Inga},
author={Linnell, Peter A.},
author={Schick, Thomas},
title={Galois cohomology of completed link groups},
journal={Proc. Amer. Math. Soc.},
volume={136},
date={2008},
number={10},
pages={3449--3459},
issn={0002-9939},
review={\MR{2415028 (2009i:20059)}},
doi={10.1090/S0002-9939-08-09395-7},
}
\bib{DLMSY}{article}{
author={Dodziuk, J{\'o}zef},
author={Linnell, Peter},
author={Mathai, Varghese},
author={Schick, Thomas},
author={Yates, Stuart},
title={Approximating $L^2$-invariants and the Atiyah conjecture},
note={Dedicated to the memory of J\"urgen K. Moser},
journal={Comm. Pure Appl. Math.},
volume={56},
date={2003},
number={7},
pages={839--873},
issn={0010-3640},
review={\MR{1990479 (2004g:58040)}},
doi={10.1002/cpa.10076},
}
\bib{Elek+Szabo}{article}{
author={Elek, G{\'a}bor},
author={Szab{\'o}, Endre},
title={Hyperlinearity, essentially free actions and $L^2$-invariants.
The sofic property},
journal={Math. Ann.},
volume={332},
date={2005},
number={2},
pages={421--441},
issn={0025-5831},
review={\MR{2178069 (2007i:43002)}},
doi={10.1007/s00208-005-0640-8},
}
\bib{MR2279234}{article}{
author={Farkas, Daniel R.},
author={Linnell, Peter A.},
title={Congruence subgroups and the Atiyah conjecture},
conference={
title={Groups, rings and algebras},
},
book={
series={Contemp. Math.},
volume={420},
publisher={Amer. Math. Soc.},
place={Providence, RI},
},
date={2006},
pages={89--102},
review={\MR{2279234 (2008b:16033)}},
}
\bib{Grabowski}{article}{
author={Grabowski, Lukasz},
title={On Turing dynamical systems and the Atiyah problem},
journal={Invent. Math.},
volume={198},
date={2014},
number={1},
pages={27--69},
issn={0020-9910},
review={\MR{3260857}},
doi={10.1007/s00222-013-0497-5},
}
\bib{Haglund-Wise}{article}{
AUTHOR = {Haglund, Fr{\'e}d{\'e}ric and Wise, Daniel T.},
TITLE = {Special cube complexes},
JOURNAL = {Geom. Funct. Anal.},
FJOURNAL = {Geometric and Functional Analysis},
VOLUME = {17},
YEAR = {2008},
NUMBER = {5},
PAGES = {1551--1620},
ISSN = {1016-443X},
CODEN = {GFANFB},
MRCLASS = {20F36 (20F55 20F67)},
MRNUMBER = {2377497},
MRREVIEWER = {Patrick Bahls},
DOI = {10.1007/s00039-007-0629-4},
URL = {http://dx.doi.org/10.1007/s00039-007-0629-4},
}
\bib{Kadison2}{book}{
author={Kadison, Richard V.},
author={Ringrose, John R.},
title={Fundamentals of the theory of operator algebras. Vol. II},
series={Pure and Applied Mathematics},
volume={100},
note={Advanced theory},
publisher={Academic Press Inc.},
place={Orlando, FL},
date={1986},
pages={i--xiv and 399--1074},
isbn={0-12-393302-1},
review={\MR{859186 (88d:46106)}},
}
\bib{Knebusch}{article}{
author ={Knebusch, Anselm},
title={Approximation of center-valued Betti numbers},
journal={Houston Journal of Mathematics},
volume={37}, pages={161--179}, year={2011},
}
\bib{Lang1}{book}{
author={Lang, Serge},
title={Algebra},
series={Graduate Texts in Mathematics},
volume={211},
edition={3},
publisher={Springer-Verlag},
place={New York},
date={2002},
pages={xvi+914},
isbn={0-387-95385-X},
review={\MR{1878556 (2003e:00003)}},
doi={10.1007/978-1-4613-0041-0},
}
\bib{MR1242889}{article}{
author={Linnell, Peter A.},
title={Division rings and group von Neumann algebras},
journal={Forum Math.},
volume={5},
date={1993},
number={6},
pages={561--576},
issn={0933-7741},
review={\MR{1242889 (94h:20009)}},
doi={10.1515/form.1993.5.561},
}
\bib{LinnellOkunSchick}{article}{
author={Linnell, Peter},
author={Okun, Boris},
author={Schick, Thomas},
title={The strong Atiyah conjecture for right-angled Artin and
Coxeter
groups},
journal={Geom. Dedicata},
volume={158},
date={2012},
pages={261--266},
issn={0046-5755},
review={\MR{2922714}},
doi={10.1007/s10711-011-9631-y},
}
\bib{L+S1}{article}{
author={Linnell, Peter},
author={Schick, Thomas},
title={Finite group extensions and the Atiyah conjecture},
journal={J. Amer. Math. Soc.},
volume={20},
date={2007},
number={4},
pages={1003--1051 (electronic)},
issn={0894-0347},
review={\MR{2328714 (2008m:58041)}},
doi={10.1090/S0894-0347-07-00561-9},
}
\bib{L+S2}{article}{
author={Linnell, Peter},
author={Schick, Thomas},
title={The Atiyah conjecture and Artinian rings},
journal={Pure and Applied Mathematics Quarterly},
volume={8}, year={2012},
pages={313--328},
note={arXiv:0711.3328},
date={2007},
}
\bib{MR1474192}{article}{
author={L{\"u}ck, Wolfgang},
title={Hilbert modules and modules over finite von Neumann algebras and
applications to $L^2$-invariants},
journal={Math. Ann.},
volume={309},
date={1997},
number={2},
pages={247--285},
issn={0025-5831},
review={\MR{1474192 (99d:58169)}},
doi={10.1007/s002080050112},
}
\bib{Lueck02}{book}{
author={L{\"u}ck, Wolfgang},
title={$L^2$-invariants: theory and applications to geometry and
$K$-theory},
series={Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematics [Results in Mathematics and
Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]},
volume={44},
publisher={Springer-Verlag},
place={Berlin},
date={2002},
pages={xvi+595},
isbn={3-540-43566-2},
review={\MR{1926649 (2003m:58033)}},
}
\bib{Lueck09}{article}{
author={L{\"u}ck, Wolfgang},
title={$L^2$-invariants from the algebraic point of view},
conference={
title={Geometric and cohomological methods in group theory},
},
book={
series={London Math. Soc. Lecture Note Ser.},
volume={358},
publisher={Cambridge Univ. Press},
place={Cambridge},
},
date={2009},
pages={63--161},
review={\MR{2605176 (2011f:46090)}},
}
\bib{Moody}{article}{
author={Moody, John Atwell},
title={Brauer induction for $G_0$ of certain infinite groups},
journal={J. Algebra},
volume={122},
date={1989},
number={1},
pages={1--14},
issn={0021-8693},
review={\MR{994933 (90b:18014)}},
doi={10.1016/0021-8693(89)90235-4},
}
\bib{Passman}{book}{
author={Passman, Donald S.},
title={Infinite group rings},
note={Pure and Applied Mathematics, 6},
publisher={Marcel Dekker Inc.},
place={New York},
date={1971},
pages={viii+149},
review={\MR{0314951 (47 \#3500)}},
}
\bib{Pichot-Schick-Zuk}{article}{
author={Pichot, Mika{\"e}l},
author={Schick, Thomas},
author={\.Zuk, Andrzej},
title={Closed manifolds with transcendental $L^2$-Betti numbers},
journal={J. Lond. Math. Soc. (2)},
volume={92},
date={2015},
number={2},
pages={371--392},
issn={0024-6107},
review={\MR{3404029}},
doi={10.1112/jlms/jdv026},
}
\bib{MR1885124}{article}{
author={Reich, Holger},
title={On the $K$- and $L$-theory of the algebra of operators affiliated
to a finite von Neumann algebra},
journal={$K$-Theory},
volume={24},
date={2001},
number={4},
pages={303--326},
issn={0920-3036},
review={\MR{1885124 (2003m:46103)}},
doi={10.1023/A:1014078228859},
}
\bib{MR1828605}{article}{
author={Schick, Thomas},
title={$L^2$-determinant class and approximation of $L^2$-Betti
numbers},
journal={Trans. Amer. Math. Soc.},
volume={353},
date={2001},
number={8},
pages={3247--3265 (electronic)},
issn={0002-9947},
review={\MR{1828605 (2002f:58056)}},
doi={10.1090/S0002-9947-01-02699-X},
note={\textit{Erratum} at arXiv:math/9807032},
}
\bib{Schick}{article}{
author={Schick, Thomas},
title={Integrality of $L^2$-Betti numbers},
journal={Math. Ann.},
volume={317},
date={2000},
number={4},
pages={727--750},
issn={0025-5831},
review={\MR{1777117 (2002k:55009a)}},
doi={10.1007/PL00004421},
note={\textit{Erratum} in vol.~\textbf{322}, 421--422}
}
\bib{Schreve}{article}{
AUTHOR = {Schreve, Kevin},
TITLE = {The strong {A}tiyah conjecture for virtually cocompact special
groups},
JOURNAL = {Math. Ann.},
FJOURNAL = {Mathematische Annalen},
VOLUME = {359},
YEAR = {2014},
NUMBER = {3-4},
PAGES = {629--636},
ISSN = {0025-5831},
MRCLASS = {20F65},
MRNUMBER = {3231009},
MRREVIEWER = {Qin Wang},
DOI = {10.1007/s00208-014-1007-9},
URL = {http://dx.doi.org/10.1007/s00208-014-1007-9},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
\label{s_introduction}
Toric varieties appear frequently in algebraic geometry. This is surprising as the
definition of a toric variety is so restrictive; $X$ is normal and there is an open subset
isomorphic to a torus such that the action of the torus on itself extends to $X$. On the
other hand the appearance of toric varieties is very useful as many geometric problems are
reduced to straightforward combinatorics. We are interested in explaining why toric
varieties appear so often and to give additional criteria for their appearance.
One approach is to try to give a simple characterisation of toric varieties. We give a
characterisation that only involves invariants coming from log pairs:
\begin{definition}\label{d_complexity} Let $X$ be a proper variety of dimension $n$
and let $(X,\Delta)$ be a log pair. A \textbf{decomposition} of $\Delta$ is an expression
of the form
\[
\sum a_iS_i\leq \Delta,
\]
where $S_i\geq 0$ are $\mathbb{Z}$-divisors and $a_i\geq 0$, $1\leq i\leq k$. The
\textbf{complexity} of this decomposition is $n+r-d$, where $r$ is the rank of the vector
space spanned by $\llist S.k.$ in the space of Weil divisors modulo algebraic equivalence
and $d$ is the sum of $\llist a.k.$.
The \textbf{complexity} $c=c(X,\Delta)$ of $(X,\Delta)$ is the infimum of the complexity
of any decomposition of $\Delta$.
\end{definition}
Note that we don't require that the divisors $S_i$ are prime divisors (since the
components of $S_i$ might span a larger vector space). On the other hand in practice the
smallest complexity is often achieved by taking $\llist S.k.$ to be prime divisors. In
the special case when the coefficients of $D=\Delta=\sum S_i$ are all one, then $d$ is the
number of components of $D$. It is well known that for a toric pair, that is, a toric
variety together with the sum of the invariant divisors, we have $d=n+r$, so that $c=0$.
We introduce some ad hoc but very convenient notation. If
\[
\Delta=\sum a_iD_i,
\]
is a boundary, that is, a divisor whose coefficients $a_i\in (0,1]$ then
\[
\sship \Delta.=\sum_{i:a_i>1/2} D_i=\rfdown \Delta.+\rfup 2\Delta.-\rfdown 2\Delta..
\]
We give a characterisation of toric pairs involving the complexity:
\begin{theorem}\label{t_toric} Let $X$ be a proper variety of dimension $n$ and let
$(X,\Delta)$ be a log canonical pair such that $-(K_X+\Delta)$ is nef.
If $\sum a_iS_i$ is a decomposition of complexity $c$ less than one then there is a
divisor $D$ such that $(X,D)$ is a toric pair, where $D\geq \sship \Delta.$ and all but
$\rdown 2c.$ components of $D$ are elements of the set $\{\, S_i \,|\, 1\leq i\leq k \,\}$.
\end{theorem}
\eqref{t_toric} is a special case of a conjecture of Shokurov, cf. \cite{Shokurov00},
which is stated in the relative case. Here are two simple corollaries of \eqref{t_toric}:
\begin{corollary}\label{c_complexity} Let $X$ be a proper variety and let $(X,\Delta)$
be a log canonical pair such that $-(K_X+\Delta)$ is nef.
Then the complexity is non-negative.
\end{corollary}
\begin{corollary}\label{c_span} Let $X$ be a proper variety of dimension $n$ and let
$(X,\Delta)$ be a log canonical pair such that $-(K_X+\Delta)$ is nef.
If the complexity is less than one then the components of $\Delta$ span the N\'eron-Severi
group.
\end{corollary}
One can extend \eqref{t_toric} to the case of any field of characteristic zero:
\begin{corollary}\label{c_non} Let $k$ be a field of characteristic zero.
Let $X$ be a proper variety over $k$ and let $(X,\Delta)$ be a log canonical pair such
that $-(K_X+\Delta)$ is nef.
If $\sum a_iS_i$ is a decomposition of complexity $c$ less than one then there is a
divisor $D$ such that $(X,D)$ is a toric pair, where $D\geq \sship \Delta.$ and all but
$\rdown 2c.$ components of $D$ are elements of the set $\{\, S_i \,|\, 1\leq i\leq k \,\}$.
\end{corollary}
We are able to prove that log pairs with small complexity have a simple birational
structure:
\begin{theorem}\label{t_form} Let $(X,\Delta)$ be a divisorially log terminal pair
where $X$ is a $\mathbb{Q}$-factorial projective variety.
If $-(K_X+\Delta)$ is nef then we may find an ample divisor $A$ and a divisor
$0\leq \Delta_0\leq \Delta$ such that the numerical dimension of $K_X+A+\Delta_0$ is at
most the complexity of $(X,\Delta)$.
In particular if $\rmap X.Z.$ is the maximal rationally connected fibration then the
dimension of $Z$ is at most the complexity.
\end{theorem}
Toric varieties are special as they are rational. We are able to give a rationality
criterion in terms of a slightly different version of the complexity:
\begin{definition}\label{d_absolute} Let $X$ be a proper variety of dimension $n$
and let $(X,\Delta)$ be a log pair. The \textbf{absolute complexity}
$\gamma=\gamma(X,\Delta)$ of $(X,\Delta)$ is $n+\rho-d$, where $\rho$ is the rank of the
group of Weil divisors modulo algebraic equivalence and $d$ is the sum of the coefficients
of $\Delta$.
\end{definition}
If $X$ is $\mathbb{Q}$-factorial then $\rho$ is the Picard number.
\begin{theorem}\label{t_rational} Let $X$ be a proper variety. Suppose that
$(X,\Delta)$ is log canonical and $-(K_X+\Delta)$ is nef.
If $\gamma(X,\Delta)<\frac 32$ then there is a proper finite morphism $\map Y.X.$ of
degree at most two, which is \'etale outside a closed subset of codimension at least two,
such that $Y$ is rational.
In particular if $A_{n-1}(X)$ contains no $2$-torsion then $X$ is rational.
\end{theorem}
The condition on torsion in the class group is necessary and we give an example of this in
\S \ref{s_example}. Note that most rationality criteria are used to establish
irrationality. There are relatively few criteria to show rationality.
\eqref{t_toric} was proved for surfaces in \cite{KM99} for Picard number one (based
heavily on ideas of Shokurov) and in \cite{Shokurov00} in general. Both proofs use
Shokurov's theory of complements. Cheltsov, in unpublished work, proved \eqref{t_toric}
when $X$ is $\mathbb{Q}$-factorial projective and the Picard number is one. The technique
he uses is the basis of our proof, which we will explain below. \cite{Prokhorov01}
contains a proof of \eqref{t_toric} for threefolds in some special cases. The method of
proof is to run the MMP. \cite{Yao13} has a proof of \eqref{t_toric} when $X$ is a smooth
projective variety, $\Delta=\sum D_i$ has global normal crossings and $K_X+\Delta$ is
numerically trivial. The method of proof is quite different from the other papers and
uses ideas coming from mirror symmetry and the powerful methods of Gross, Hacking, Keel
and Siebert, cf. \cite{GHKS16}. \cite{Karzhemanov13} contains work related to both
\eqref{t_toric} and \eqref{t_rational}.
There are some examples to show \eqref{t_toric} and \eqref{t_rational} are sharp. First
an example to show that not every invariant divisor is a component of $\Delta$:
\begin{example}\label{e_but} Consider $(X=\pr 2.,\Delta=L_1+L_2+1/2C)$ where $L_1$ and $L_2$
are two lines and $C$ is a conic, in general position. Then $(X,\Delta)$ is divisorially
log terminal, $K_X+\Delta \sim_{\mathbb{Q}}0$ and the complexity is
\[
c=2+1-5/2=1/2.
\]
Note that $\sship \Delta.=L_1+L_2$. Let $L_3$ be a third line in general position. Then
$(\pr 2.,L_1+L_2+L_3)$ is a toric pair and two of the three invariant divisors are
components of $\Delta$ but not all three.
\end{example}
It is also not hard to see that it is crucial that $(X,\Delta)$ is log canonical:
\begin{example}\label{e_lc} Take $X=\Hz n.$ the unique $\pr 1.$-bundle over $\pr 1.$
with a curve $E_{\infty}$ of self-intersection $-n$. Let $\Delta=2E_{\infty}+\sum F_i$,
where $\llist F.n+2.$ are $n+2$ fibres. Then $K_X+\Delta\sim 0$ and the complexity
\[
c=2+2-(n+4)=-n,
\]
is arbitrarily large and negative. Note that if one contracts $E_{\infty}$ then the image
of $\Delta$ is a boundary and the complexity is $c=1-n$.
\end{example}
One can also see that one cannot relax nef to pseudo-effective:
\begin{example}\label{e_pe} If we replace $\Delta$ by $E_{\infty}+\sum F_i$ in
\eqref{e_lc} then $(X,\Delta)$ is log canonical and $-(K_X+\Delta)$ is pseudo-effective
but the complexity is again $1-n$.
\end{example}
We also have an example where $X$ is smooth and all the coefficients are one:
\begin{example}\label{e_quadric} Let $Q=(XY-ZW=0)\subset \pr 4.$ be a rank four quadric
threefold. Pick a small resolution $\map X.Q.$ with exceptional locus $L$ isomorphic to
$\pr 1.$. Note that any hyperplane through the vertex of $Q$, which intersects the
quadric at infinity in two lines, intersects $Q$ in two planes through the vertex. By
adjunction the sum of three such pairs is an element of $|-K_Q|$.
If $D=D_1+D_2+D_3+D_4+D_5+D_6$ is the strict transform of these six divisors then
$K_X+D\sim 0$ and the complexity
\[
c=3+2-6=-1.
\]
On the other hand three components of $D$ contain the curve $L$, so that $(X,D)$ is not
log canonical, even though $X$ is smooth and every component of $D$ has coefficient one.
\end{example}
It is also easy to see that we need to work with the absolute complexity for
unirrationality and that \eqref{t_toric} is sharp:
\begin{example}\label{e_elliptic} If $X=E$ is an elliptic curve and we take $\Delta=0$ then
$K_E \sim 0$ and the complexity is $1$. On the other hand $E$ is not unirational. Note
that the absolute complexity is $2$.
\end{example}
In fact if one works over a non-algebraically field, it is easy to see that we need to
allow an extension of degree two for rationality:
\begin{example}\label{e_real} If $C=V(x^2+y^2+z^2)$ is a smooth conic over $\mathbb{R}$
without a real point then we may find a divisor $D$ of degree one such that $K_X+D \sim 0$
so that the absolute complexity is one. On the other hand $C$ is irrational but $C$
becomes rational if we replace $\mathbb{R}$ with $\mathbb{C}$.
\end{example}
We give an example in \S \ref{s_example} to show that we need a cover of degree two to
achieve rationality, cf. \eqref{t_rational}. This example is in some sense a geometric
realisation of \eqref{e_real}.
Let us turn to a description of the proof of \eqref{t_toric}. The first step is to
replace $(X,\Delta)$ by a divisorially log terminal model $(Y,\Gamma)$. This means that
$Y$ is projective, $\mathbb{Q}$-factorial and $(Y,\Gamma)$ is divisorially log terminal.
There is a birational contraction map $\pi\colon\rmap Y.X.$ and the only exceptional
divisors have log discrepancy zero. If $X$ is projective then we can take $\pi$ to be a
morphism and this is a standard reduction step (by a result of Hacon, see for example,
\cite[3.1]{KK09}). If $X$ is not projective then there are examples which show it is not
always possible to arrange for $\pi$ to be a morphism.
For example, take $X$ to be any smooth proper variety which is not projective and take
$\Delta$ to be empty. Let $\pi\colon\rmap Y.X.$ be a divisorially log terminal model of
$X$. As $X$ is smooth it is kawamata log terminal and so $\pi$ is small. $\pi$ is not
the identity morphism as $Y$ is projective and $X$ is not. Therefore $\pi$ is not a
morphism as $X$ is $\mathbb{Q}$-factorial.
For a concrete example, consider the smooth toric threefold $X$ on page 71 of
\cite{Fulton93}. It is not projective as it has no ample divisors. It is easy to see
that if one flops an invariant curve $\rmap X.Y.$, corresponding to a diagonal edge of the
slanted faces of the tetrahedron, then $Y$ is projective and the induced birational map
$\pi\colon\rmap Y.X.$ is a divisorially log terminal model.
We prove the existence of divisorially log terminal models in \eqref{p_dlt}, contingent on
the existence of a nef divisor $M$ such that $K_X+\Delta+M$ is nef. This covers the case
when either $K_X+\Delta$ or $-(K_X+\Delta)$ is nef and the latter is sufficient for our
purposes. We check in \eqref{l_dlt} that the complexity of $(Y,\Gamma)$ is at most the
complexity of $(X,\Delta)$; this is straightforward since every exceptional divisor
extracted by $\pi$ is a component of $\Gamma$ of coefficient one. Finally it is not hard
to see that it is enough to work with $(Y,\Gamma)$, cf. \eqref{l_persist}.
Thus we may assume that $X$ is projective, $\mathbb{Q}$-factorial and $(X,\Delta)$ is
divisorially log terminal. The next step is to proceed based on the assumption that $X$
is a Mori dream space.
To explain this step we first describe Cheltsov's argument which applies when the Picard
number is one. In this case $K_X$ and all the components of $\Delta$ are proportional to
a very ample divisor $H$. If we let $(Y,\Gamma)$ be the cone over $(X,\Delta)$ under the
embedding given by $H$ then $(Y,\Gamma)$ is log canonical and by construction every
component of $\Gamma$ is $\mathbb{Q}$-Cartier and passes through the vertex $p$ of the
cone.
The goal is then to prove \eqref{l_flipsabundance}, a local version of \eqref{t_toric}.
The proof of \eqref{l_flipsabundance} is based on the proof of \cite[18.22]{Kollaretal},
which establishes that the sum of the coefficients of $\Gamma$, which is precisely the sum
of the coefficients of $\Delta$, is no more than the dimension of $Y$. Passing to a
composition of cyclic covers, we may assume that both $K_Y$ and every component of
$\Gamma$ is Cartier and in this case it suffices to check that $Y$ is smooth. If we
replace components of $\Gamma$ whose coefficients sum to one by a general element of the
linear system they span we can apply adjunction and induction to conclude that $Y$ is
smooth. Since the original variety is a quotient by a product of cyclic groups, it is not
hard to see that the original variety $Y$ is toric. Since the only way to get a toric
variety as a cone is to start with a toric variety we see that $X$ must be toric; indeed
$X$ is isomorphic to the exceptional divisor of the blow up of the cone at the vertex $p$.
Unfortunately the naive generalisation of this argument does not apply if the Picard
number is not one. The problem is that the cone over a variety of Picard number at least
two is not even $\mathbb{Q}$-factorial; for example the quadric cone which is the cone
over $\pr 1.\times \pr 1.$ is not $\mathbb{Q}$-factorial.
Instead of working with a cone we work with the affine variety $Y$ associated to the Cox
ring of $X$. $X$ is a Mori dream space if and only if the Cox ring is finitely generated.
The Cox ring is naturally graded by the class group, the group of Weil divisors modulo
linear equivalence. As usual this grading corresponds to the action on $Y$ of an
algebraic group $H$, the spectrum of the group algebra associated to the class group,
which is the product of a torus and a finite abelian group. We can recover $X$ as the
quotient of $Y$ by $H$. In the case when the class group is isomorphic to $\mathbb{Z}$
(so that, in particular, the Picard number is one), $Y$ is a cone and $H$ is a one
dimensional torus, acting in the usual way on the lines of the cone. As in the case of a
cone, there is a natural log pair $(Y,\Gamma)$ associated to $X$ and every component of
$\Gamma$ passes through the same point $p$. $(Y,\Gamma)$ is log canonical if and only if
$(X,\Delta)$ is log canonical by \cite{Brown13}, \cite{GOST15}, and \cite{KO12}. Mori
dream spaces were introduced in the very influential paper \cite{HK00}. We actually use a
more sophisticated version of the Cox ring, which was introduced in \cite{Hausen08}. It
has the advantage that every Weil divisor on $X$ corresponds to a Cartier divisor on $Y$,
so that we don't even need to take any cyclic covers.
The main point at this step of the proof is to bound the dimension of $Y$. The dimension
of $Y$ is the dimension of $X$ plus the Picard number. By assumption the sum of the
coefficients of a decomposition $\sum a_iD_i$ of $\Delta$ is at least the dimension of
$X$, minus one, plus the dimension $r$ of the space spanned by the components
$\llist D.k.$. So we have to show that $r=\rho$, that is, the components $\llist D.k.$
generate the vector space of divisors modulo linear equivalence.
We prove this result by induction on $r$. We start with the case that $\llist D.k.$ span
the same ray of the cone of divisors. It is easy to show that the Picard number of $X$ is
one. Consider for example the case that $X$ is a smooth projective surface and $K_X+D$ is
numerically trivial. If the Picard number is not one then either there is a $-1$-curve
$\Sigma$ or a $\pr 1.$-bundle $\map X.C.$. If $\Sigma$ is a $-1$-curve then $K_X$ is
negative on $\Sigma$ so that $D$ is positive on $\Sigma$. As the components of $D$ are
proportional to each other it follows that every component of $D$ intersects $\Sigma$. As
the sum of the coefficients of $D$ is at least three, $D\cdot \Sigma\geq 3$, which is
impossible as $K_X\cdot \Sigma=-1$. If $\map X.C.$ is a $\pr 1.$-bundle and $\Sigma$ is a
general fibre we have $D\cdot\Sigma\geq 3$ and $K_X\cdot \Sigma=-2$, which is again
impossible. In the general case we run an appropriate MMP. After finitely many flips we
either get a divisorial contraction or a Mori fibre space and both cases we can rule out,
using a similar argument, cf. \eqref{l_sum}.
Otherwise we may pick two components $D_1$ and $D_2$ of $D$ such that neither
$P_1=m_1D_1-m_2D_2$ nor $P_2=m_2D_2-m_1D_1$ is pseudo-effective. In this case consider
the $\pr 1.$-bundle given by the direct sum of the line bundles corresponding to $P_1$ and
$P_2$. $Y$ is a Mori dream space and the two sections corresponding to $P_1$ and $P_2$
are contractible, $\rmap Y.Z.$. In this case we proceed by induction on the rank $r$,
cf. \eqref{t_decomp}. The details of this step are in \S \ref{s_local}.
To reduce to the case when $X$ is a Mori dream space we have to pass to a different model
$Y$ such that $-(K_Y+\Gamma)$ is ample for some kawamata log terminal pair $(Y,\Gamma)$.
Note that in this case $K_Y+B+\Gamma$ is numerically trivial, where $B=-(K_Y+\Gamma)$ is
ample. So we look for divisors $0\leq \Delta_0\leq \Delta$ and ample divisors $A$ such
that $K_X+A+\Delta_0$ has numerical dimenson zero. In this case $Y$ is a log terminal
model of $(X,A+\Delta_0)$.
If the numerical dimension is not zero then there is a non-trivial fibration $\map Y.Z.$.
Not every component of $D$ dominates $Z$, since otherwise the complexity of the general
fibre is less than zero, cf. \eqref{l_fibration}. On the other hand it is not hard to
decrease the numerical dimension if there is a component of $D$ which does not dominate,
cf. \eqref{l_decrease_dim}. To finish off, we replace $A+\Delta_0$ by a convex linear
combination of $A+\Delta_0$ and $M+\Delta$, where $M=-(K_X+\Delta)$, and cancel off common
components of $\Delta_0$ and exceptional divisors of $f\colon\rmap X.Y.$ so that the
complexity of $(X,A+\Delta_0)$ is close to the complexity of $(X,\Delta)$ and $f$ does not
contract any components of $\Delta$, cf. \eqref{l_minimal}. The details are in \S
\ref{s_small}.
To finish the proof of \eqref{t_toric}, we need to know that if $Y$ is toric then so is
$X$. The key point is to reduce to the case that $N=0$. The first step is to pass to a
model such that no centre of $(X,\Delta)$ is contained in the exceptional locus of $f$,
cf. \eqref{l_up}. We then perturb $\Gamma$ so that it is more singular along at least one
exceptional divisor, cf. \eqref{l_less}. Taking a convex linear combination of
$\Delta_0$, a divisor supported on $N$ and the perturbed divisor we may decrease the
number of components of $N$ and we are done by induction on the number of components of
$N$. The details are in \S \ref{s_reduction}.
Now we turn to the proof of \eqref{t_rational}. The proof follows similar lines to the
proof of \eqref{t_toric}. We may assume that $X$ is projective, $\mathbb{Q}$-factorial
and $(X,\Delta)$ is divisorially log terminal and by \eqref{c_numerical} we may assume
that $X$ is a Mori dream space. If the absolute complexity is less than two then we can
conclude that the affine variety $Y$ associated to the Cox ring of $X$ has compound Du Val
singularities, meaning that there is a surface section with Du Val singularities. If we
further assume that the absolute complexity is less than $3/2$ then we can conclude that
$Y$ has a compound $A_l$ singularity, meaning that a surface section of $Y$ has an $A_l$
singularity.
It follows that $Y$ is a hypersurface in affine space $\af m.$ given by a polynomial $q$
whose quadratic part has rank two. The action of $H$ on $Y$ extends to $\af m.$. The
quotient of $\af m.$ by $H$ is a toric variety and $X$ is birational to the image of $Y$
in this toric variety. If $xy\in q$, that is, $xy$ is a monomial with non-zero
coefficient in $q$, then it is not hard to check that there is a one dimensional torus
whose general orbit intersects $X$ in a single point. Thus $X$ is birational to an
invariant divisor so that $X$ is rational. Otherwise after rescaling we may assume that
the quadratic part of $f$ has the form $x^2+y^2$. If $x$ and $y$ have the same
multidegree then we may change variable and reduce to the previous case. Otherwise there
must be torsion in the class group and there is a cover $\map Y.X.$ of degree two. The
details are in \eqref{p_quadricmds}.
In \S \ref{s_example} we exhibit log canonical pairs $(X,\Delta)$ of absolute
complexity one such that $X$ is irrational. The idea is to start with a conic bundle of
relative Picard number two over $\pr 1.\times \pr 1.$ and take a $\mathbb{Z}_2$-quotient
to achieve relative Picard number one. The key observation is that the discriminant
curve, the locus of reducible fibres, makes no contribution in Kawamata's canonical bundle
formula. Thus we can arrange for the discriminant curve to have arbitrarily large genus,
in which case $X$ is irrational.
We suspect that if the absolute complexity is less than two then we may always find a
cover so that $X$ is rational. In this case we have to consider the extra possibility
that $Y$ has a compound singularity of type $D_l$, $E_6$, $E_7$, or $E_8$. However we
were unable to see how to proceed in this case.
\makeatletter
\renewcommand{\thetheorem}{\thesubsection.\arabic{theorem}}
\@addtoreset{theorem}{subsection}
\makeatother
\section{Preliminaries}
In this section we will collect some definitions and preliminary results. We work over a
field of characteristic zero which is algebraically closed unless otherwise stated.
\subsection{Notation and Conventions}
\label{s_notation}
Let $X$ be a proper variety. $\rho(X)$ is the rank of the Picard group of $X$. We denote
the class group, the group of Weil divisors modulo linear equivalence, by $A_{n-1}(X)$.
We will follow the terminology from \cite{KM98}. In particular we only consider
valuations $\nu$ of $X$ whose centre on some birational model $Y$ of $X$ is a divisor. A
\textit{log canonical place} of a log canonical pair $(X,\Delta)$ is any valuation $\nu$
whose log discrepancy is zero.
Suppose that $f\colon\rmap X.Y.$ is a rational map whose domain is an open subset $U$
whose complement has codimension at least two. In this case if $D$ is an
$\mathbb{R}$-Cartier divisor on $Y$ we may define $f^*D$ as the $\mathbb{R}$-Weil divisor
whose restriction to $U$ is the usual pullback.
We say a proper morphism $f\colon\map X.Y.$ is a \textit{contraction morphism} if
$f_*\ring X.=\ring Y.$. Let $f\colon\rmap X.Y.$ be a proper rational map of normal
quasi-projective varieties and let $p\colon\map W.X.$ and $q\colon\map W.Y.$ be a common
resolution of $f$. We say that $f$ is a \textit{rational contraction} if $q$ is a
contraction morphism and the image of every $p$-exceptional divisor has codimension two or
more in $Y$. We say that a prime divisor $P$ on $X$ is \textit{horizontal} if the image
of the generic point of $P$ is the generic point of $Y$. We say that $P$ is
\textit{vertical} if it is not horizontal.
We say that a birational map $f\colon\rmap X.Y.$ is a \textit{birational contraction} if
$f$ is a rational contraction, so that every $p$-exceptional divisor is $q$-exceptional.
If $D$ is an $\mathbb{R}$-Cartier divisor on $X$ such that $D':=f_*D$ is
$\mathbb{R}$-Cartier then we say that $f$ is $D$-\textit{non-positive} (resp.
$D$-\textit{negative}) if we have $p^*D=q^*D'+E$ where $E\geq 0$ and $E$ is
$q$-exceptional (respectively $E$ is $q$-exceptional and the support of $E$ contains the
strict transform of the $f$-exceptional divisors).
Now suppose that $f\colon\rmap X.Y.$ is a birational contraction of projective varieties.
If $X$ is $\mathbb{Q}$-factorial and $(X,\Delta)$ is a divisorially log terminal pair such
that $f$ is $(K_X+\Delta)$-negative, $K_Y+\Gamma$ is nef and $Y$ is
$\mathbb{Q}$-factorial, where $\Gamma=f_*\Delta$, then we say that $f\colon\rmap X.Y.$ is
a \textit{log terminal model} of $K_X+\Delta$. If the ring
\[
R(X,K_X+\Delta):=\bigoplus_{m\geq 0}H^0(X,\ring X.(m(K_X+\Delta)))
\]
is finitely generated then $\rmap X.Z.$ is called the \textit{ample model} of $(X,\Delta)$,
where
\[
Z=\operatorname{Proj} R(X,K_X+\Delta).
\]
Let $D$ be an $\mathbb{R}$-Cartier divisor on a projective variety $X$. Let $C$ be a
prime divisor. If $D$ is big then
\[
\sigma_C(D)=\inf \{\, \operatorname{mult}_C(D') \,|\, D'\sim_\mathbb{R} D, D'\geq 0\,\}.
\]
More generally if $D$ is simply pseudo-effective we extend the definition of $\sigma_C$ as
follows. Let $A$ be any ample $\mathbb{Q}$-divisor. Following \cite{Nakayama04}, let
\[
\sigma_C(D)=\lim_{\epsilon\to 0} \sigma_C(D+\epsilon A).
\]
Then $\sigma_C(D)$ exists and is independent of the choice of $A$. There are only
finitely many prime divisors $C$ such that $\sigma_C(D)>0$ and the $\mathbb{R}$-divisor
$N_\sigma(X,D)=\sum _C\sigma_C(D)C$ is determined by the numerical equivalence class of
$D$, cf. \cite[3.3.1]{BCHM10} and \cite{Nakayama04} for more details. If we put
\[
P_\sigma(X,D)=D-N_\sigma(X,D)
\]
then we will call
\[
D=P_\sigma(X,D)+N_\sigma(X,D),
\]
\textit{Nakayama's Zariski decomposition}.
Following \cite{Nakayama04} we define \textit{the numerical dimension}
\[
\kappa_{\sigma}(X,D)=\max_{H\in {\operatorname{Pic}}(X)}\{\, k\in \mathbb{N} \,|\, \limsup_{m\to \infty} \frac{h^0(X,\ring X.(mD+H))}{m^k} >0\,\},
\]
where $H$ is an ample divisor on $X$. If $D$ is nef then this is the same as
\[
\nu(X,D)=\max \{\, k\in \mathbb{N} \,|\, H^{n-k}\cdot D^k>0 \,\}.
\]
Let $f\in K[\llist x.n.]$ be a polynomial. If $\mu$ is a monomial in $\llist x.n.$ then
we write $\mu\in f$ if and only if the coefficient of $\mu$ in $f$ is non-zero.
If $k$ is a field and $\bar k$ is the algebraic closure of $k$ then bars will denote
extension of schemes to $\bar k$.
\subsection{Birational Geometry}
\begin{lemma}\label{l_orientation} Let $X$ be a $\mathbb{Q}$-factorial projective variety
and let $(X,\Delta)$ be a kawamata log terminal pair. Suppose that $\Delta$ is big and
$K_X+\Delta$ is pseudo-effective. Let $\pi\colon\rmap X.Z.$ be the ample model and let
$D$ be a prime divisor.
Then $K_X+\Delta-dD$ is pseudo-effective for $d$ sufficiently small, if and only if,
either $D$ does not dominate $Z$ or the support of $D$ lies in the support of the stable
base locus of $K_X+\Delta$.
\end{lemma}
\begin{proof} Let $K_X+\Delta=P+N$ be Nakayama's Zariski decomposition. Then the
components of $N$ are the prime divisors in the stable base locus of $K_X+\Delta$.
If the support of $D$ lies in the support of the stable base locus of $K_X+\Delta$ then we
may find $d>0$ such that $dD\leq N$ and in this case
\[
K_X+\Delta-dD=P+(N-dD)\geq P
\]
is pseudo-effective.
Let $H$ be the ample divisor on $Z$ corresponding to $K_X+\Delta$. If $D$ does not
dominate $Z$ then we can pick $d>0$ and $H' \sim_{\mathbb{R}} H$ such that
$\pi^*H'\geq dD$, so that $K_X+\Delta-dD$ is pseudo-effective.
Now suppose $D$ dominates $Z$ and let $F$ be the general fibre of $\pi$. Then $P|_F=0$.
Therefore if $K_X+\Delta-dD$ is pseudo-effective then $dD\leq N$. But then the support of
$D$ lies in the support of the stable base locus of $K_X+\Delta$.
\end{proof}
We will need a version of the MMP for log canonical pairs.
\begin{lemma}\label{l_mmp} Let $X$ be a $\mathbb{Q}$-factorial kawamata log terminal
projective variety and let $(X,\Delta)$ be a log canonical pair.
If $K_X+\Delta$ is not pseudo-effective then we may run the $(K_X+\Delta)$-MMP until we
arrive at a Mori fibre space.
\end{lemma}
\begin{proof} Pick an ample divisor $A$ such that $K_X+A+\Delta$ is not pseudo-effective.
Since $X$ is $\mathbb{Q}$-factorial kawamata log terminal we may find a divisor
$\Delta'\sim_{\mathbb{R}}A+\Delta$ such that $(X,\Delta')$ is kawamata log terminal.
In particular, \cite[1.3.3]{BCHM10} implies that the $(K_X+\Delta')$-MMP with scaling of
$A$ always terminates with a Mori fibre space. On the other hand any run of the
$(K_X+\Delta')$-MMP with scaling of $A$ is automatically a run of the $(K_X+\Delta)$-MMP.
\end{proof}
We will need divisorially log terminal models in the case when $X$ is proper but not
necessarily projective. In this case we need to relax the requirement that we have a
morphism. To emphasize this point we use the term model rather than modification.
\begin{definition}\label{d_dlt} Let $(X,\Delta)$ be a log canonical pair,
where $X$ is a proper variety.
A \textbf{divisorially log terminal model} is a divisorially log terminal pair
$(Y,\Gamma)$, where $Y$ is a projective $\mathbb{Q}$-factorial variety, together with a
birational contraction $\pi\colon\rmap Y.X.$ such that
\[
K_Y+\Gamma=\pi^*(K_X+\Delta),
\]
and the only divisors contracted by $\pi$ have log discrepancy zero with respect to
$(X,\Delta)$.
\end{definition}
We are only able to prove the existence of divisorially log terminal models in very
special cases:
\begin{proposition}\label{p_dlt} Let $(X,\Delta)$ be a log canonical pair where
$X$ is a proper variety.
If $M$ is a nef divisor such that $K_X+\Delta+M$ is nef then we may find a divisorially
log terminal model such that both $N=\pi^*M$ and $K_Y+\Gamma+\lambda N$, for some
$\lambda\geq 1$, are nef.
In particular if $M=\pm (K_X+\Delta)$ then $\pm(K_Y+\Gamma)$ is nef.
\end{proposition}
We will need some preliminary results, which are simple extensions of results by Shokurov,
cf. Addendum 4 of \cite{Shokurov06}.
\begin{lemma}\label{l_combinatorial} Let $\llist m.k.$ be positive real numbers and let
$m$ and $r$ be positive integers.
Then there is a positive constant $\hbar$ such that if
\[
a\in \{\, \sum \frac{a_i m_i}r \,|\, \llist a.k.\in \mathbb{Z}, a_i\geq -m\,\}
\]
and $a>0$ then $a\geq \hbar$.
\end{lemma}
\begin{proof} Clear.
\end{proof}
\begin{lemma}\label{l_decompose} Let $(X,\Delta)$ be a kawamata log terminal pair
where $X$ is a $\mathbb{Q}$-factorial projective variety and $\Delta$ is a big
$\mathbb{R}$-divisor.
If $M$ is a nef $\mathbb{R}$-divisor then we may find a positive constant $\hbar$ with the
following property:
If $f\colon\rmap X.Y.$ is any sequence of $(K_X+\Delta)$-flips which are $M$-trivial and
$C$ is any curve spanning a $(K_Y+\Gamma)$-extremal ray of the cone of curves of $Y$ then
either $N\cdot C\geq \hbar$ or $N\cdot C=0$, where $N=f_*M$ and $\Gamma=f_*\Delta$.
\end{lemma}
\begin{proof} We may write
\[
M=\sum m_iM_i,
\]
where $\llist m.k.$ are positive real numbers and $\llist M.k.$ are $\mathbb{Q}$-Cartier
divisors. Pick this decomposition minimal with this property, so that $\llist m.k.$ are
independent over $\mathbb{Q}$. Pick $M_i$ sufficiently close to $M$ so that we may find
\[
\Phi_i \sim_{\mathbb{R}} \Delta+M_i,
\]
where $(X,\Phi_i)$ is kawamata log terminal. Pick a positive integer $r$ so that $rM_i$
is Cartier, for all indices $1\leq i\leq k$.
We first check that all of these properties are preserved by any sequence
$f\colon\rmap X.Y.$ of $(K_X+\Delta)$-flips which are $M$-trivial. By induction we are
reduced to the case of one flip. If $R$ is the corresponding $(K_X+\Delta)$-extremal ray
then $R$ is spanned by a rational curve $C$. As $M\cdot C=0$ and $\llist m.k.$ are
independent over $\mathbb{Q}$, we must have $M_i\cdot C=0$. Thus $N$ is nef and $rN_i$ is
Cartier. It is clear that
\[
N=\sum m_iN_i \qquad \text{and} \qquad \Psi_i=f_*\Phi_i \sim_{\mathbb{R}} \Gamma+N_i,
\]
since $f$ is an isomorphism in codimension one. The pair $(Y,\Psi_i)$ is kawamata log
terminal as $f$ is a $(K_X+\Phi_i)$-flip.
Thus there is no harm in assuming that $f$ is the identity. Suppose that $R$ is a
$(K_X+\Delta)$-extremal ray. Then \cite[1]{Kawamata91} implies that $R$ is spanned by a
rational curve $C$ such that
\[
-2n\leq (K_X+\Phi_i)\cdot C\leq N_i\cdot C=\frac{a_i}r,
\]
for some integer $a_i$. Now apply \eqref{l_combinatorial} with $m=2nr$.
\end{proof}
\begin{proof}[Proof of \eqref{p_dlt}] As a first approximation, let $\pi\colon\map Y.X.$
be a log resolution of $(X,\Delta)$ such that $Y$ is projective. We may write
\[
K_Y+\Gamma=K_Y+\widetilde \Delta+E=\pi^*(K_X+\Delta)+F
\]
where $\widetilde \Delta$ is the strict transform of $\Delta$, $E=\sum E_i$ is the sum of
the exceptional divisors and $F\geq 0$ is exceptional. This model would be a divisorially
log terminal model provided $F=0$. Our goal is to contract $F$ using the MMP, preserving
the condition that $N$ is nef.
We may write
\[
K_Y+\Gamma+N=\pi^*(K_X+\Delta+M)+F.
\]
Note that $K_Y+\Gamma+N$ is pseudo-effective and the diminished base locus of
$K_Y+\Gamma+N$ is equal to the support of $F$. Pick an ample divisor $A$ so that the
support of the stable base locus of $K_Y+A+\Gamma+N$ is equal to the support of $F$. Then
the stable base locus of $K_Y+A+\Gamma+tN$ is equal to the support of $F$ for any
$t\geq 1$.
Let
\[
\lambda=\max(1,\frac{2n}{\hbar})>0,
\]
where $\hbar$ is defined in \eqref{l_decompose}. Let $f\colon\rmap Y.Y'.$ be a step of
the $(K_Y+A+\Gamma+\lambda N)$-MMP with scaling of $A$. If $R$ is the corresponding
extremal ray then $R$ is spanned by a rational curve $C$ such that
$(K_Y+A+\Gamma)\cdot C>-2n$ so that $N\cdot R=0$. In particular $f_*N$ is nef. If $f$
contracts a divisor then this divisor is a component of $F$ so that $f$ only contracts
divisors which are exceptional for $\pi$. Therefore we are free to replace $Y$ by $Y'$.
Note that we might lose the property that $\pi$ is a morphism, when $f$ is a flip, but we
retain the property that $\pi$ is a birational contraction.
Now suppose that $g\colon\rmap Y.Y'.$ is a sequence of flips which are $N$-trivial. By
\eqref{l_decompose} these are all steps of the $(K_Y+A+\Gamma+\lambda N)$-MMP with scaling
of $A$. Since this MMP always terminates, after finitely many steps we construct a model
on which $F=0$.
\end{proof}
\subsection{Toric Geometry}
We say that $X$ is a \textit{toric variety} if $X$ is a normal variety over a field $k$
(not necessarily algebraically closed), there is a dense open subset $U$ isomorphic to
$\mathbb{G}^n_m$ such that the natural action of $U$ on itself extends to the whole of
$X$. (Note that this is stronger than the usual definition in the literature which only
requires that $U$ is isomorphic to $\mathbb{G}^n_m$ after passing to the algebraic
closure). We will say that a log pair $(X,D)$ is \textit{toric} if $X$ is a toric variety
and $D$ is the sum of the invariant divisors.
Every toric variety has a description in terms of fans. We will use the notation of
\cite{Fulton93}.
\begin{lemma}\label{l_missing} Let $X$ be a $\mathbb{Q}$-factorial projective toric variety
of dimension $n$ and let $V$ be a closed irreducible invariant subset. Let $D$ be a fixed
invariant divisor.
Then we may find a divisor $B\geq 0$ on $X$, supported on the invariant divisors which
contain neither $D$ nor $V$, such that $A=B|_V$ is very ample and every element of the
linear system $|A|$ lifts to $X$.
\end{lemma}
\begin{proof} We may as well assume that $D$ does not contain $V$. If
$F\subset N_{\mathbb{R}}$ is the fan corresponding to $X$ then $D$ is given by a one
dimensional cone $\rho$ in $F$. If $P_{\mathbb{R}}$ is the quotient vector space of
$N_{\mathbb{R}}$ corresponding to $V$ then the image of $\rho$ in $P_{\mathbb{R}}$ is
either a ray or zero. Let $W$ be the closed invariant subset of $V$ determined by the
smallest cone which contains the image.
Let $A\geq 0$ be a very ample divisor on $V$ supported on the invariant divisors which do
not contain $W$. $A$ determines a continuous piecewise linear function $f_A$ on
$P_{\mathbb{R}}$, which is non-negative as $A\geq 0$. By composition we get a continuous
piecewise linear function $g$ on $N_{\mathbb{R}}$ which in turn corresponds to a divisor
$B$ supported on the invariant divisors. $B\geq 0$ as $g$ is non-negative and the
restriction to $V$ is $A$, as $g$ is the composition of the natural projection and $f_A$.
It is enough to lift every invariant element $A'\in |A|$. Note that, in the notation of
\cite{Fulton93}, $M(\sigma)\subset M$ is naturally the space of monomials on $V$, where
$\sigma$ is the cone corresponding to $V$. We may find $u\in M(\sigma)$ such that
\[
A'=A+(\chi^u).
\]
On the other hand the zeroes and poles of $\chi^u$, as a rational function on $X$, don't
contain $V$. Note that $f'=f+u$ is the continuous piecewise linear function corresponding
to $A'$ and $f'$ takes only non-negative values. Then $g'=g+u$ is the composition of the
naturally projection and $f'$, and so $g'$ only take non-negative values. Hence
\[
B'=B+(\chi^u)\in |B|
\]
is a divisor on $X$ which restricts to $A'$. \end{proof}
We will need the next couple of results in the case when the groundfield is not
necessarily algebraically closed:
\begin{lemma}\label{l_persist} Let $k$ be any field. Let $X$ and $Y$ be two proper
varieties and let $(X,D)$ and $(Y,G)$ be two log pairs. Let $\pi\colon\rmap X.Y.$ be a
birational contraction and $G=\pi_*D$.
\begin{enumerate}
\item If $(X,D)$ is toric and $Y$ is projective then both $(Y,G)$ and $\pi$ are toric.
\item If $(Y,G)$ is toric, $X$ is projective and the exceptional divisors of $\pi$ are
components of $D$ that correspond to toric valuations of $Y$ then both $(X,D)$ and $\pi$
are toric.
\end{enumerate}
\end{lemma}
\begin{proof} Suppose $(X,D)$ is toric. If $H$ is an ample divisor on $Y$ then $\pi^*H$
is linearly equivalent to an invariant divisor. As $Y=\operatorname{Proj}(X,\pi^*H)$ then both $(Y,G)$
and $\pi$ are toric. This is (1).
Now suppose $(Y,G)$ is toric, the exceptional divisors of $\pi$ are components of $D$ and
correspond to toric valuations of $Y$. We may find a toric pair $(Z,H)$ and a birational
morphism $f\colon\map Z.Y.$ whose only exceptional divisors correspond to these toric
valuations. As the induced birational map $\rmap X.Z.$ is an isomorphism in codimension
one, it is a birational contraction. Thus (2) follows from (1).
\end{proof}
We will need an extension of \eqref{l_persist} to the case when $X$ and $Y$ are not
projective, only proper. We start with:
\begin{lemma}\label{l_non} Let $(X,D)$ be a log pair over a field $k$ and let
bars denote extension to the algebraic closure $\bar k$ of $k$.
Then $(X,D)$ is toric if and only if $U=X-D$ is isomorphic to $\mathbb{G}_m^n$ and
$(\bar X,\bar D)$ is toric.
\end{lemma}
\begin{proof} One direction is clear.
Otherwise if $U$ is a torus then it acts on itself and we get a morphism
\[
\map U \times U.U..
\]
Now $U\times U$ is birational to $U\times X$ and so we get a rational map
\[
f\colon\rmap U \times X.X..
\]
This induces a rational map
\[
\bar f\colon\rmap \bar U \times \bar X.\bar X..
\]
As $\bar X$ is toric, $\bar f$ is in fact a morphism. But then $f$ is a morphism.
\end{proof}
\begin{lemma}\label{l_morphism} Let $k$ be any field. Let $Y$ be a proper variety
and let $\pi\colon\map Y.X.$ be a birational morphism of normal varieties.
If $Y$ is toric then both $X$ and $\pi$ are toric.
\end{lemma}
\begin{proof} We first prove this result using the additional hypothesis that $k$ is
algebraically closed.
Replacing $Y$ by a toric resolution, we may assume that $Y$ is
smooth and projective. In particular $\pi$ is projective. Let $U\subset Y$ be the torus.
By assumption there is an action
\[
\map U\times Y.Y. \qquad \text{given by} \qquad \map (u,y).u\cdot y..
\]
By composition there is a morphism
\[
f\colon\map U\times Y.X..
\]
Since $U\times X$ is birational to $U\times Y$ there is an induced rational map
\[
g\colon\rmap U\times X.X..
\]
We check that $g$ is a morphism.
Suppose that $y_1$ and $y_2$ are two points of $Y$, with the same image in $X$. It
suffices to check that $f(u,y_1)=f(u,y_2)$ for all points $u\in U$. As $\pi$ is
projective and birational and $X$ is normal the fibres of $\pi$ are connected. Then $y_1$
and $y_2$ are connected by a chain of curves $C$ in $Y$ which are contracted by $\pi$. As
the torus $U$ is connected the components of $C$ and of $u\cdot C$ are numerically
equivalent. But then $u\cdot y_1$ and $u\cdot y_2$ belong to the connected curve
$u\cdot C$ which is contracted by $\pi$. Thus $f(u,y_1)=f(u,y_2)$, for all $u\in U$ and
so there is an induced morphism $g$.
It is clear that $g$ defines an action of $U$ on $X$. As $\pi$ is birational
$\map U.\pi(U).$ is an isomorphism. Thus $X$ contains a torus and the natural action of
the torus extends to $U$. Therefore $X$ is a toric variety.
Now suppose that $k$ is not algebraically closed. Let $U$ be the open subset of $Y$
isomorphic to $\mathbb{G}_m^n$. As $\bar \pi \colon\map \bar Y.\bar X.$ is a toric
morphism the restriction of $\bar \pi$ to $\bar U$ is an isomorphism, so that
$\bar \pi(\bar U)$ is an open subset of $\bar X$. But then the restriction of $\pi$ to
$U$ is an isomorphism and so $\pi(U)$ is an open subset of $X$ isomorphic to
$\mathbb{G}_m^n$. It follows that $X$ is toric by \eqref{l_non} and it is easy to
conclude that $\pi$ is toric.
\end{proof}
We now return to assuming that the groundfield is algebraically closed.
\begin{lemma}\label{l_proper} Let $X$ be a proper variety and let $(X,\Delta)$ be a
log canonical pair of complexity less than one such that $-(K_X+\Delta)$ is nef. Suppose
that $(Y,\Gamma)$ is a divisorially log terminal model of $(X,\Delta)$,
$\pi\colon\rmap Y.X.$.
If $(Y,G)$ is a toric pair, where $G\geq \sship\Gamma.$ then $(X,D)$ is a toric pair,
where $D=\pi_*G\geq \sship\Delta.$.
\end{lemma}
\begin{proof} It suffices to prove that $(X,D)$ is a toric pair. Note that $(X,D)$ is log
canonical, $K_X+D$ is numerically trivial and
\[
K_Y+G=\pi^*(K_X+D).
\]
In particular a valuation $\nu$ is a log canonical place of $(Y,G)$ if and only if it is a
log canonical place of $(X,D)$.
Let $(Z,L)$ be a toric resolution of $(Y,G)$. Then the exceptional divisors of
$\map Z.Y.$ have log discrepancy zero, so that the induced birational map $\rmap Z.X.$ is
a divisorially log terminal model of $(X,D)$. Replacing $(Y,G)$ by $(Z,L)$ we may assume
that $Y$ is smooth. Let $\map W.X.$ be a divisorially log terminal modification of
$(X,D)$. We may write
\[
K_W+C=f^*(K_X+D),
\]
where $C$ is the strict transform of $D$ plus the exceptionals. By \eqref{l_morphism} it
suffices to prove that $(W,C)$ is toric. As $f$ only extracts divisors of log discrepancy
zero which also have log discrepancy zero for $(Y,G)$, possibly blowing up $Y$, we may
assume that the induced rational map $\rmap Y.W.$ is a birational contraction. Replacing
$(X,D)$ by $(W,C)$ we may assume that $X$ is $\mathbb{Q}$-factorial and $X$ is kawamata
log terminal.
Let $\llist \nu.k.$ be the set of valuations corresponding to the exceptional divisors of
$\pi$. Then the centres of $\llist \nu.k.$ are components of $G$ and so $\llist \nu.k.$
are log canonical places. We may find a modification $f\colon\map W.X.$ such that the
exceptional divisors of $f$ are precisely the centres of $\llist \nu.k.$, where $W$ is
$\mathbb{Q}$-factorial and kawamata log terminal. Replacing $(X,D)$ by $(W,C)$ once again
we may assume that $X$ is isomorphic to $Y$ in codimension one.
The result now follows by \cite[Corollary 2]{Fine89}.
\end{proof}
\subsection{Calculus of the complexity}
In \S \ref{s_introduction} we defined the complexity $c(X,\Delta)$ and the absolute
complexity $\gamma(X,\Delta)$ for any log pair $(X,\Delta)$. It is not hard to see that
the infimum is achieved for the complexity as there are only finitely many partitions of
the set of prime divisors contained in the support of $\Delta$. It is immediate from the
definitions that
\[
c(X,\Delta) \leq \gamma(X,\Delta).
\]
\begin{lemma}\label{l_dlt} Let $X$ be a proper variety and let $(X,\Delta)$ be a log
canonical pair.
If $\pi\colon\rmap Y.X.$ is a divisorially log terminal model,
\[
K_Y+\Gamma=\pi^*(K_X+\Delta),
\]
then the complexity (respectively absolute complexity) of $(Y,\Gamma)$ is at most the
complexity (respectively absolute complexity) of $(X,\Delta)$.
\end{lemma}
\begin{proof} Let $\sum_{i=1}^m a_iS_i$ be a decomposition of $\Delta$. Let $R_i$ be the
strict transform of $S_i$, $1\leq i\leq m$ and let $\llist E.k.$ be the exceptional
divisors.
Let
\[
T_i=\begin{cases} R_i & \text{if $1\leq i\leq m$} \\
E_{i-m} & \text{if $m<i\leq m+k$}
\end{cases}
\]
and
\[
b_i=\begin{cases} a_i & \text{if $1\leq i\leq m$} \\
1 & \text{if $m<i\leq m+k$.}
\end{cases}
\]
Then $\sum b_iT_i$ is a decomposition of $\Gamma$. The sum $e$ of the coefficients of
$\sum b_iT_i$ is $d+k$. $\llist T.m+k.$, modulo algebraic equivalence, span a vector
space of dimension at most $r+k$. Thus the complexity of the decomposition $\sum b_iT_i$
is at most
\[
n+(r+k)-(d+k)=n+r-d,
\]
which is the complexity of the decomposition given by $\sum a_iS_i$. Thus the complexity
of $(Y,\Gamma)$ is at most the complexity of $(X,\Delta)$. The absolute case is similar
and easier.
\end{proof}
\begin{definition}\label{d_local-complexity} Let $(x\in X,\Delta)$ be the germ of a log pair.
A \textbf{local decomposition} of $\Delta$ is an expression of the form
\[
\sum a_iD_i\leq \Delta,
\]
where $D_i\geq 0$ are integral $\mathbb{Q}$-Cartier divisors and $a_i\geq 0$,
$1\leq i\leq k$. The \textbf{local complexity} of this decomposition is $n-d$, where $n$
is the dimension of $X$ and $d$ is the sum of $\llist a.k.$.
\end{definition}
The following lemma establishes a local version of \eqref{t_toric}. The proof is adapted
from the proof of \cite[18.22]{Kollaretal}:
\begin{lemma}\label{l_flipsabundance} Let $(x \in X,\Delta)$ be the germ of a log canonical
pair where $X$ has dimension $n$ and let $\sum a_iD_i\leq \Delta$ be a local
decomposition. Assume that $K_X$ and $\llist D.k.$ are Cartier.
If $\gamma=n-\sum a_i=n-d$ is the local complexity then
\begin{enumerate}
\item $\gamma \geq 0$.
\item If $\gamma < 1$ then, possibly re-ordering $\llist D.k.$,
\[
(X,\alist D.+.m.)
\]
is log smooth, where $m=n-\rfdown 2\gamma.$. In addition
\[
\sship \Delta.\leq \alist D.+.m..
\]
\item If $\gamma < \frac 32$ then either $X$ is smooth at $x$ or has a $cA_l$ singularity
at $x$.
\end{enumerate}
\end{lemma}
\begin{proof} We proceed by induction on $n$. All claims are clear for $n=1$ and so
we assume that $n\geq 2$.
Fix a log resolution $\pi\colon\map Y.X.$ of $(X,\Delta)$, with exceptional divisors
$\llist E.l.$. Let $f$ be a general linear combination of $\llist g.k.$, the functions
defining $\llist D.k.$. Let $S$ be the divisor cut out by $f$. As $S$ specialises to
$D_i$, for each $i$, it follows that
\[
\operatorname{mult}_{E_j} S\leq \operatorname{mult}_{E_j} D_i.
\]
for each $1\leq i\leq k$ and $1\leq j\leq l$. It also follows that $\pi$ is a log
resolution of $(X,\Delta+S)$. For any $0 \leq b_i \leq a_i$ such that $\sum b_i=b\leq 1$,
it follows that the pair $(X,\Phi=bS+\sum (a_i-b_i)D_i)$ is log canonical, and the local
complexity of the indicated decomposition is $\gamma$.
Suppose that $0<b_i\neq a_i$, and $b=1$, so that
\[
\sum_{i:b_i\neq 0} a_i>1.
\]
Let $V\subset X$ be a codimension two subset. As the pair $(X,\Delta)$ is log canonical
in a neighbourhood of the generic point of $V$ there is an index $i$ such that $b_i\neq 0$
and either $V$ is not contained in $D_i$ or $D_i$ is smooth at the generic point of $V$.
In this case $S$ is normal. In particular if $d>1$ we may pick $\llist b.k.$ so that
$b=1$ and $S$ is normal.
As $S$ is Cartier and normal, $X$ is smooth in codimension two along $S$. Therefore we
may write
\[
(K_X+\Phi)|_S=K_S+\Psi,
\]
where $(S,\sum (a_i-b_i)D_i|_S\leq \Psi)$ is log canonical and the local complexity is at
most $\gamma$.
Now suppose that $\gamma<1$. As $n\geq 2$ then $d>1$ and so we may choose $\llist b.k.$
so that $S$ is normal. By induction $S$ is smooth. As $S$ is Cartier $X$ is smooth.
Then $\operatorname{mult} _x\Delta\leq n$ as $(X,\Delta)$ is log canonical. In particular every
component of $\rdown\Delta.$ is smooth.
(1) and (2) follow by induction on $n$.
Now suppose that $\gamma<3/2$. If $n\geq 3$ then $d>1$. By definition of compound
singularities it suffices to prove that $S$ has a $cA_l$ singularity. By induction we may
assume that $n=2$ and we have to show that $X$ has an $A_l$ singularity. As $K_X$ is
Cartier and $X$ is a normal surface, $X$ is Gorenstein. As $\Delta\neq 0$ it follows that
$X$ is kawamata log terminal so that $X$ is canonical. Thus $X$ has du Val singularities.
We may also assume that $\Delta=dD$, where $D=S$ is a prime Cartier divisor.
If $\pi\colon\map Y.X.$ is the minimal desingularisation of the surface $X$ then
$K_Y=\pi^*K_X$. Let $G$ be the strict transform of $D$ and let $\llist E.l.$ be the
exceptional divisors. Since $D$ is Cartier, we have
\[
f^*D=G+\sum m_iE_i
\]
where $\llist m.l.$ are positive integers.
The log discrepancy of $E_i$ with respect to $K_X+\Delta=K_X+dS$ is
\[
1-dm_i.
\]
As $(X,\Delta)$ is log canonical and $d>\frac 12$, we must have $m_i=1$, for all
$1\leq i\leq l$. Hence
\[
0=f^*D\cdot E_j = (G +\sum E_i) \cdot E_j\geq \delta(E_j)-2
\]
where $\delta(E_j)$ is the degree of the vertex corresponding to $E_j$ in the dual graph
of the resolution. It follows that every vertex in the dual graph has degree at most $2$
and so $X$ has an $A_l$ singularity.
\end{proof}
\subsection{Mori Dream Spaces}
Recall, cf. \cite{HK00},
\begin{definition}\label{d_mds} Let $X$ be a $\mathbb{Q}$-factorial normal projective variety. We say
that $X$ is a \textbf{Mori dream space} if the following conditions hold:
\begin{enumerate}
\item X is $\mathbb{Q}$-factorial and $\operatorname{Pic}(X)_\mathbb{Q}= N^1(X)_\mathbb{Q}$;
\item the cone of nef divisors, $\operatorname{Nef}(X)$, is the affine hull of finitely many semi-ample divisors;
\item there exist finitely many small birational maps $f_i\colon\rmap X.X_i.$, such that
each $X_i$ satisfies (1) and (2) and the closure of the cone of movable divisors,
$\operatorname{Mov}(X)$, is the union of the cones $f_i^*\operatorname{Nef}(X_i)$.
\end{enumerate}
\end{definition}
The Cox ring of a variety with finitely generated class group was originally defined in
\cite{HK00}; it is unique but ignores torsion in the class group. Subsequently
\cite{Hausen08} gave a refined definition which takes into account torsion in the class
group. As we would like to allow torsion we will use this definition of the Cox ring.
We will need some of the basic properties of the Cox ring, cf. \cite{Hausen08} for more
details and proofs. The most important result is that $X$ is a Mori dream space if and
only if the ring $R=\operatorname{Cox}(X)$ is a finitely generated $\mathbb{C}$-algebra. One can use
this to give many examples of Mori dream spaces:
\begin{lemma}\label{l_mori} Let $X$ be a projective variety.
The following are equivalent
\begin{enumerate}
\item We may find a kawamata log terminal pair $(X,\Delta)$ such that
$-(K_X+\Delta)$ is ample.
\item We may find a kawamata log terminal pair $(X,\Delta)$ such that
$-(K_X+\Delta)$ is big and nef.
\item We may find a kawamata log terminal pair $(X,\Delta)$ such that
$K_X+\Delta$ is numerically trivial and $\Delta$ is big.
\end{enumerate}
In particular, if $X$ is $\mathbb{Q}$-factorial then $X$ is a Mori dream space.
\end{lemma}
\begin{proof} (1) clearly implies (2). If $(X,\Delta)$ is kawamata log terminal and
$-(K_X+\Delta)$ is big and nef then $-(K_X+\Delta)$ is semiample. Then we may find
$B\geq 0$, $B\sim_{\mathbb{R}} -(K_X+\Delta)$ such that $(X,\Delta+B)$ is kawamata log
terminal. Thus (2) implies (3).
Suppose that $(X,\Delta)$ is kawamata log terminal, $K_X+\Delta$ is numerically trivial
and $\Delta$ is big. We may find an ample $\mathbb{Q}$-divisor $A$ and a divisor
$B\geq 0$ such that
\[
\Delta \sim_{\mathbb{R}} A+B.
\]
Pick $\epsilon>0$ such that $(X,\Delta+\epsilon B)$ is kawamata log terminal. Then
\[
-(K_X+(1-\epsilon)\Delta+\epsilon B) \sim_{\mathbb{R}} \epsilon A.
\]
As $(X,(1-\epsilon)\Delta+\epsilon B)$ is kawamata log terminal, (3) implies (1).
The last assertion is \cite[1.3.2]{BCHM10}.
\end{proof}
If $X$ is a Mori dream space then let $Y=\operatorname{Spec} R$. If $D$ is a prime divisor on $X$ one
can associate a Cartier divisor $G$ on $Y$. The ring $R$ is naturally graded by the class
group $A_{n-1}(X)$. There is a unique closed point $p\in Y$ corresponding to the unique
maximal homogeneous ideal and $p\in G$. The grading corresponds to an action of the
diagonalisable group $H=\operatorname{Spec} \mathbb{C}[A_{n-1}(X)]$. $X$ is a geometric quotient of $Y$
by the action of $H$ and the divisor $D$ is naturally the image of the associated Cartier
divisor $G$ on $Y$.
We will need a small strengthening of \cite[2.10]{HK00}:
\begin{theorem}\label{t_polynomial-ring} Let $X$ be a $\mathbb{Q}$-factorial projective
variety.
Then $X$ is toric if and only if the Cox ring is a polynomial ring generated by
$\operatorname{dim} X+\rho(X)$ variables, in which case the invariant divisors correspond to the
coordinate hyperplanes.
\end{theorem}
\begin{proof} If $X$ is a toric variety then the Cox ring is the homogeneous coordinate
ring of $X$ and the Cox ring is a polynomial ring with $\operatorname{dim} X+\rho(X)$ variables, which
correspond to the invariant divisors on $X$, cf. the discussion after the proof of
\cite[2.2]{Hausen08}.
Now suppose that the Cox ring is a polynomial ring. Then $X$ is a Mori dream space. In
particular its divisor class group $A_{n-1}(X)$ is a finitely generated abelian group and
the Cox ring is graded by the class group. In this case $X$ is the GIT quotient of affine
space $\af m.$ by a diagonalisable group $H$, the product of a torus and a finite abelian
group, \cite[2.2]{Hausen08}. Therefore $X$ is a toric variety.
\end{proof}
We will also need:
\begin{lemma}\label{l_cartier} Let $X$ be a $\mathbb{Q}$-factorial projective variety.
Suppose that $X$ is a Mori dream space and let $R=\operatorname{Cox} (X)$ be the Cox ring.
If $Y=\operatorname{Spec} R$ then $K_Y$ is Cartier. In particular if $Y$ is Cohen-Macaulay then $Y$ is
Gorenstein.
\end{lemma}
\begin{proof} Let $H=\operatorname{Spec} \mathbb{C}[A_{n-1}(X)]$. According to \cite[2.2]{Hausen08} it
suffices to check that $K_Y$ is $H$-invariant. It also follows from \cite[2.2]{Hausen08}
that there is a universal $H$-torsor $q\colon\map \hat X.X.$ and it suffices to prove that
$K_{\hat X}$ is $H$-invariant.
The group $H$ decomposes as a torus and a finite abelian group. The morphism $q$ then
decomposes as a torus bundle followed by an \'etale cover. It follows that $K_{\hat X}=q^*K_X$
so that $K_{\hat X}$ is $H$-invariant.
\end{proof}
\makeatletter
\renewcommand{\thetheorem}{\thesection.\arabic{theorem}}
\@addtoreset{theorem}{section}
\makeatother
\section{Local to global}
\label{s_local}
\begin{theorem}\label{t_mdscase} Let $X$ be a $\mathbb{Q}$-factorial projective variety
with kawamata log terminal singularities and let $(X,\Delta)$ be a log canonical pair.
Suppose that $-(K_X+\Delta)$ is nef and $\sum a_iS_i$ is a decomposition of complexity $c$
less than one for $\Delta$.
If $X$ is a Mori dream space then there is a divisor $D$ such that $(X,D)$ is a toric
pair, where $D\geq \sship \Delta.$ and all but $\rdown 2c.$ components of $D$ are elements
of the set $\{\, S_i \,|\, 1\leq i\leq k \,\}$.
\end{theorem}
\begin{theorem}\label{t_decomp} Let $X$ be a $\mathbb{Q}$-factorial kawamata log terminal
projective variety. Suppose that $(X,\Delta)$ is a log canonical pair such that
$K_X+\Delta$ is numerically trivial. Let $\sum a_i S_i$ be a decomposition of $\Delta$
with complexity less than $1$.
If $X$ is a Mori dream space then $\llist S.k.$ generate $A_{n-1}(X)_{\mathbb{Q}}$.
\end{theorem}
\begin{lemma}\label{l_sum} Let $X$ be a $\mathbb{Q}$-factorial kawamata log terminal
projective variety of dimension $n$ and let $(X,\Delta)$ be a log canonical pair. Let
$D=\sum a_iS_i\leq \Delta$ be a decomposition of $\Delta$.
If
\begin{enumerate}
\item $K_X+\Delta$ is numerically trivial,
\item $d=\sum_{i=1}^k a_i>n$, and
\item $\llist S.k.$ all span the same ray of the cone of effective divisors
\end{enumerate}
then the Picard number of $X$ is one.
\end{lemma}
\begin{proof} Let $\Theta=\Delta-D$. We run the $(K_X+\Theta)$-MMP with scaling of some
ample divisor.
Let $f\colon\rmap X.Y.$ be a step of this MMP. $f$ is $D$-positive and as the components
of $S$ span the same ray of the cone of effective divisors, it follows that $f$ is
$S_i$-positive, for every $1\leq i\leq k$. Let $T_i=f_*S_i$.
Suppose that $f$ is a divisorial contraction. If $V$ is the image of the exceptional
divisor $E$ then $T_i$ contains $V$. If $\Gamma=f_*\Delta$ then $(Y,\Gamma)$ is log
canonical and the local complexity about a point of $V$ is negative. This is not possible
by (1) of \eqref{l_flipsabundance}.
If $f$ is a flip then $\rho(X)=\rho(Y)$ and $\llist T.k.$ all span the same ray of the
cone of effective divisors. We replace $X$ by $Y$ in this case. \eqref{l_mmp} implies
that after finitely many flips $f$ must be a Mori fibre space. Let $F$ be the general
fibre and let $\Sigma$ be the restriction of $\Delta$ to $F$. Then $(F,\Sigma)$ is log
canonical. As $\llist S.k.$ dominate $Y$, the sum of the coefficients of $\Sigma$ is
greater than $n$. \eqref{l_flipsabundance} implies that $F$ has dimension $n$. But then
$Y$ is a point and $X$ has Picard number one.
\end{proof}
\begin{proof}[Proof of \eqref{t_decomp}] We proceed by induction on the dimension $r$ of
the span of $\llist S.k.$ in $A_{n-1}(X)_{\mathbb{Q}}$. If $r=1$ then we may apply
\eqref{l_sum}.
Otherwise, we may assume that $S_1$ and $S_2$ are linearly independent in
$A_{n-1}(X)_{\mathbb{Q}}$. Pick integers $m_1$ and $m_2$ such that $m_1S_1$ and $m_2S_2$
are Cartier, and neither $m_1S_1-m_2S_2$ nor $m_2S_2-m_1S_1$ is pseudo-effective.
Consider the $\pr 1.$-bundle
\[
Y=\proj{\ring X.(m_1D_1)\oplus \ring X.(m_2D_2)}..
\]
Let $f\colon\map Y.X.$ be the structure morphism. Then $Y$ is a $\mathbb{Q}$-factorial
projective variety with kawamata log terminal singularities. There are two distinguished
sections, which we will call $E_0$ and $E_\infty$. Set
$\Gamma=f^* \Delta+ E_0 + E_{\infty}$. Adjunction implies that $(Y,\Gamma)$ is a log
canonical pair, and that $K_Y+\Gamma$ is numerically trivial. Note that
$\rho(Y)=\rho(X)+1$. Finally, $Y$ is a Mori Dream Space because the Cox ring of $Y$ is
isomorphic as a ring to the Cox ring of $X$ with two variables adjoined, corresponding to
the sections $E_0$ and $E_{\infty}$, cf. \cite[3.2]{Brown13}.
As both $m_1S_1-m_2S_2$ and $m_2S_2-m_1S_1$ are not pseudo-effective, $E_0|_{E_0}$ and
$E_{\infty}|_{E_\infty}$ are not pseudoeffective. Thus $D=E_0+E_{\infty}$ has Kodaira
dimension zero. As $Y$ is a Mori dream space we may run $g\colon\rmap X.Z.$ the $D$-MMP
and the image of $D$ is semiample. Thus the birational map $g$ contracts $E_0$ and
$E_{\infty}$. $Z$ is a $\mathbb{Q}$-factorial projective variety with kawamata log
terminal singularities.
Note that $K_Z+\Psi$ is numerically trivial and $(Z,\Psi=g_*\Gamma)$ is log canonical.
Note that
\[
\rho(Z)=\rho(Y)-2=\rho(X)-1.
\]
If $T_i=f^*S_i$ then the dimension of the space spanned by $\llist T.k.$ in
$A_{n-1}(Y)_{\mathbb{Q}}$ is equal to $r$. Let $C_i=g_*T_i$. As $m_1C_1-m_2C_2$ is
linearly equivalent to zero in $Z$, $\llist C.k.$ span a vector space of dimension $r-1$
in $A_{n-1}(Z)_{\mathbb{Q}}$.
$Z$ is a Mori dream space, as $Y$ is a Mori dream space. By induction $\llist C.k.$
generate $A_{n-1}(Z)_{\mathbb{Q}}$. We can identify $A_{n-1}(Z)_{\mathbb{Q}}$ with
\[
\frac{A_{n-1}(X)_{\mathbb{Q}}}{\langle m_1S_1-m_2S_2\rangle}
\]
and so $\llist S.k.$ span $A_{n-1}(X)_{\mathbb{Q}}$. \end{proof}
\begin{proof}[Proof of \eqref{t_mdscase}] Since $-(K_X+\Delta)$ is nef, and $X$ is a Mori
dream space, we can find $B\sim_{\mathbb{R}} -(K_X+\Delta)$ such that $(X,\Delta+B)$ is
log canonical. By \eqref{t_decomp}, the components of any decomposition of $(X,\Delta+B)$
with complexity less than $1$ generate $A_{n-1}(X)_{\mathbb{Q}}$. The pair $(X,\Delta)$
has a decomposition with complexity less than $1$, and this is a decomposition of
$(X,\Delta+B)$. Thus $r=\rho$, where $r$ is the rank of the group generated by the
$\llist S.k.$ and $\rho$ is the Picard number.
Let $Y=\operatorname{Spec} R$ where $R=\operatorname{Cox}(X)$ is the Cox ring. Then $Y$ has dimension $n+\rho$. Let
$T_i$ be the divisor corresponding to $S_i$ and let $\Gamma=\sum a_iT_i$. Then $T_i$ is a
Cartier divisor and every component $\llist T.k.$ contains the point $p$ corresponding to
the unique maximal ideal which is homogeneous, \cite[2.2]{Hausen08}. By \cite[1.1]{KO12}
the pair $(Y,\Gamma)$ is log canonical (as observed in \cite[2.5]{KO12} their result
applies to the Cox ring, as defined in \cite{Hausen08}).
\eqref{l_flipsabundance} implies that $Y$ is smooth, every component of $\Gamma$ of
coefficient greater than $1/2$ is smooth and at least $\operatorname{dim} Y-\rfdown 2c.$ components of
$\llist T.k.$ are smooth at $p$ and intersect transversally.
The result now follows from \eqref{t_polynomial-ring}.
\end{proof}
\section{Log divisors of small numerical dimension}
\label{s_small}
\begin{proposition}\label{p_numerical} Assume \eqref{t_toric}$_{n-1}$, that is,
assume \eqref{t_toric} when $X$ has dimension $n-1$.
Let $(X,\Delta)$ be a divisorially log terminal pair where $X$ is a $\mathbb{Q}$-factorial
projective variety of dimension $n$.
If $-(K_X+\Delta)$ is nef then we may find an ample divisor $A$ and a divisor
$0\leq \Delta_0\leq \Delta$ such that $K_X+A+\Delta_0$ is pseudo-effective, no component
of $N_{\sigma}(X,K_X+A+\Delta_0)$ is a component of $\Delta_0$, and the numerical
dimension of $K_X+A+\Delta_0$ is at most the complexity of $(X,\Delta)$.
\end{proposition}
\begin{corollary}\label{c_numerical} Assume \eqref{t_toric}$_{n-1}$.
Let $X$ be a $\mathbb{Q}$-factorial projective variety of dimension $n$. Suppose
$(X,\Delta)$ is a divisorially log terminal pair such that $-(K_X+\Delta)$ is nef.
Let $\gamma_0\in (0,2)$.
If the absolute complexity $\gamma(X,\Delta)<\gamma_0$ then there is a log canonical pair
$(Y,\Gamma)$ such that $-(K_Y+\Gamma)$ is ample, $\gamma(Y,\Gamma)<\gamma_0$ and $Y$ is a
$\mathbb{Q}$-factorial projective variety birational to $X$.
\end{corollary}
\begin{lemma}\label{l_fibration} Assume \eqref{t_toric}$_{n-1}$.
Let $(X,\Delta)$ be a divisorially log terminal pair where $X$ is a $\mathbb{Q}$-factorial
projective variety of dimension $n$. Let $A$ be an ample divisor such that $K_X+A+\Delta$
is pseudo-effective and let $\phi\colon\rmap X.Z.$ be the ample model of $K_X+A+\Delta$.
Assume that no component of $N=N_{\sigma}(X,K_X+A+\Delta)$ is a component of $\Delta$.
If the dimension of $Z$ is greater than the complexity of $(X,\Delta)$ then we may find a
component $P$ of $\Delta$ which is vertical.
\end{lemma}
\begin{proof} Let $f\colon\rmap X.Y.$ be a log terminal model of $K_X+A+\Delta$. Then
there is a contraction morphism $g\colon\map Y.Z.$. The divisors contracted by $f$ are
the components of $N$ and so $f$ does not contract any components of $\Delta$. If
$B=f_*A$ and $\Gamma=f_*\Delta$ then $(Y,B+\Gamma)$ is divisorially log terminal.
If $F$ is the general fibre of $g$ and $\Theta$ is the restriction of $B+\Gamma$ to $F$
then $(F,\Theta)$ is log canonical and $K_F+\Theta$ is numerically trivial. Let
$\sum a_iS_i$ be a decomposition of $(X,\Delta)$ which computes the complexity. Let $C_i$
be the restriction to $F$ of the image of $S_i$. Then $\sum a_iC_i$ is a decomposition of
$(F,\Theta)$, where the sum ranges over the indices $i$ such that at least one component
of $S_i$ is horizontal. The rank of the span of $\llist C.k.$ is at most the rank of the
span of $\llist S.k.$, the sum $h$ of the coefficients of $\llist C.k.$ is at least the
sum of the coefficients of the horizontal components of $\llist S.k.$ and the dimension of
$F$ is equal to the dimension of $X$ minus the dimension of $Z$.
As we are assuming \eqref{t_toric}$_{n-1}$, which implies \eqref{c_complexity}$_{n-1}$,
the complexity of the pair $(F,\Theta)$ is non-negative. Thus $h<d$ so that there is an
index $i$ such that every component of $S_i$ is vertical. In particular at least one
component $P$ of $\Delta$ is vertical.
\end{proof}
\begin{lemma}\label{l_decrease_dim} Let $(X,\Delta)$ be a divisorially log terminal
pair, where $X$ is a $\mathbb{Q}$-factorial projective variety and $-(K_X+\Delta)$ is nef.
Let $A_0$ be an ample divisor and let $0\leq \Delta_0\leq \Delta$ be a divisor such that
$K_X+A_0+\Delta_0$ has numerical dimension $k$. Suppose $\Delta$ has a component $P$
which is vertical for the ample model $\phi\colon\rmap X.Z_0.$ of $K_X+A_0+\Delta_0$.
Then there is an ample divisor $A_1$ and a divisor $0\leq \Delta_1 \leq \Delta$ such that
$K_X+A_1+\Delta_1$ is pseudo-effective and the numerical dimension of $K_X+A_1+\Delta_1$
is less than $k$.
\end{lemma}
\begin{proof} Set $M=-(K_X+\Delta)$. Let $p$ be the coefficient of $P$ in $\Delta$. Pick
$\lambda$ minimal so that
\begin{align*}
K_X+A_1+\Delta_1 &=K_X+\lambda A_0+(1-\lambda)M+\lambda\Delta_0+(1-\lambda)(\Delta-pP)\\
&=\lambda(K_X+A_0+\Delta_0)+(1-\lambda)(K_X+M+\Delta-pP)\\
&=\lambda(K_X+A_0+\Delta_0)-(1-\lambda)pP
\end{align*}
is pseudo-effective, where
\[
A_1=\lambda A_0+(1-\lambda)M \qquad \text{and} \qquad \Delta_1=\lambda\Delta_0+(1-\lambda)(\Delta-pP).
\]
Note that $\lambda>0$. In particular $A_1$ is ample and $0\leq \Delta_1\leq \Delta$.
Let $A_t=(1-t)A_0+tA_1$ and $\Delta_t=(1-t)\Delta_0+t\Delta_1$. Let $Z_t$ be the ample
model of $K_X+A_t+\Delta_t$. Note that $K_X+A_t+\Delta_t$ is a convex linear combination
of $K_X+A_0+\Delta_0$ and $-P$. Therefore if $P$ is in the stable base locus of
$K_X+A_t+\Delta_t$ then $Z_t=Z_0$ and so $P$ is vertical over $Z_t$. If $P$ is not in the
stable base locus and $t<1$ then \eqref{l_orientation} implies that $P$ is vertical over
$Z_t$.
By \cite[3.3.2]{HM10} we may find $\delta>0$ such that $Y=Z_t$ is independent of
$t\in (1-\delta,1)$ and there is a contraction morphism $f\colon\map Y.Z_1.$.
By what we just proved $P$ is vertical for $Y$. On the other hand \eqref{l_orientation}
implies that $P$ is horizontal for $Z_1$. Thus $f$ is not birational and so the dimension
of $Z_1$ is less than the dimension of $Y$. In particular the numerical dimension of
$K_X+A_1+\Delta_1$ is less than $k$.
\end{proof}
\begin{lemma}\label{l_minimal} Let $(X,\Delta)$ be a divisorially log terminal pair
where $X$ is a $\mathbb{Q}$-factorial projective variety and $M=-(K_X+\Delta)$ is nef.
Let $\delta>0$ be any positive real number.
Let $A$ be an $\mathbb{R}$-divisor and let $0\leq \Delta_0\leq \Delta$ (respectively
$\Delta_0\geq \alpha\Delta$ some $\alpha$). Suppose that $K_X+A+\Delta_0$ is
pseudo-effective. Let
\[
K_X+A+\Delta_0=P+N=P_{\sigma}(X,K_X+A+\Delta_0)+N_{\sigma}(X,K_X+A+\Delta_0)
\]
be Nakayama's Zariski decomposition. We may write $N=N_0+N_1$, where every component of
$N_0$ is a component of $\Delta$ and no component of $N_1$ is a component of $\Delta$.
Then we may find $t>0$ and $\Delta_1\leq \Delta\leq (1+\delta)\Delta_1$ (respectively
$\Delta_1\geq (1-\delta)\Delta$) such that
\[
P_{\sigma}(X,K_X+A_t+\Delta_1)=tP \qquad \text{and} \qquad N_{\sigma}(X,K_X+A_t+\Delta_1)=tN_1,
\]
where
\[
A_t=tA+(1-t)M.
\]
\end{lemma}
\begin{proof} We have
\begin{align*}
K_X+A+\Delta_0 &\sim_{\mathbb{R}} P+N \\
K_X+M+\Delta &\sim_{\mathbb{R}} 0.
\end{align*}
Given $t\in (0,1]$,
\begin{align*}
K_X+A_t+t\Delta_0+(1-t)\Delta &=t(K_X+A+\Delta_0)+(1-t)(K_X+M+\Delta) \\
&\sim_{\mathbb{R}}tP+tN.
\end{align*}
It $t>0$ is sufficiently small then
\[
\Delta_1=t\Delta_0+(1-t)\Delta-tN_0\geq 0 \qquad \text{and} \qquad \Delta\leq (1+\delta)\Delta_1.\qedhere
\]
\end{proof}
\begin{proof}[Proof of \eqref{p_numerical}] Consider divisors of the form
$K_X+A+\Delta_0$, where $0\leq \Delta_0\leq \Delta$ and $A$ is ample. Let $k$ be the
minimum of the numerical dimension of pseudo-effective divisors of this form.
Then $k<\infty$, since we can always pick $A$ so that $K_X+A$ is ample. Pick
$0\leq \Delta_0\leq \Delta$ and an ample divisor $A$ such that $K_X+A+\Delta_0$ has
numerical dimension $k$.
Suppose that $k$ is bigger than the complexity of $(X,\Delta)$. By \eqref{l_minimal} we
may assume that no component of $\Delta$ is a component of $N$ and that the complexity of
$(X,A+\Delta_0)$ is less than $k$. By \eqref{l_fibration} we may find a component $P$ of
$\Delta$ which is vertical for the ample model of $(X,A+\Delta_0)$. Then by
\eqref{l_decrease_dim} we can find an ample divisor $A_1$ and a divisor
$0\leq \Delta_1 \leq \Delta$ such that $K_X+A_1+\Delta_1$ is pseudo-effective and the
numerical dimension is less than $k$, a contradiction. \end{proof}
\begin{proof}[Proof of \eqref{c_numerical}] Consider divisors of the form
$K_X+A+\Delta_0$, where $0\leq \Delta_0\leq \Delta$, $A$ is ample and no component of
$N_{\sigma}(X,K_X+A+\Delta_0)$ is a component of $\Delta_0$. Let $k$ be the minimal
numerical dimension of pseudo-effective divisors of this form. As
\[
c(X,\Delta)\leq \gamma(X,\Delta)<2,
\]
by \eqref{p_numerical} we have $k\leq 1$.
Pick $0\leq \Delta_0\leq \Delta$ and an ample divisor $A$ such that $K_X+A+\Delta_0$ has
numerical dimension $k$. By \eqref{l_minimal} we may assume that that the absolute
complexity of $(X,A+\Delta_0)$ is less than $\gamma_0$. Pick $\delta>0$ such that
$A+\delta\Delta_0$ is ample. Replacing $A$ by $A+\delta\Delta_0$ and $\Delta_0$ by
$(1-\delta)\Delta_0$ we may assume that $(X,A+\Delta_0)$ is kawamata log terminal.
Let $\phi\colon\rmap X.Z.$ be the ample model of $K_X+A+\Delta_0$ and let
$f\colon\rmap X.Y.$ be a log terminal model of $K_X+A+\Delta_0$. Then there is a
contraction morphism $g\colon\map Y.Z.$. If $\Gamma=f_*(\Delta_0+A)$ then
$\gamma(Y,\Gamma) <\gamma_0$.
Suppose that $k=1$, that is, suppose $Z$ is a curve. If $F$ is a general fibre of $g$ and
$\Theta=\Gamma|_F$ then $(F,\Theta)$ is log canonical and $K_F+\Theta$ is numerically
trivial. The natural map which assigns to a divisor on $Y$ its restriction to $F$ has a
non-trivial kernel, since $F$ restricts to zero. Therefore the dimension $r$ of the span
of the components of $\Theta$ is at most $\rho(Y)-1$. \eqref{l_decrease_dim} implies that
every component of $\Gamma$ dominates $Z$. Therefore the sum $t$ of the coefficients of
$\Theta$ is at least the sum $d$ of the coefficients of $\Delta$. Hence
\begin{align*}
c(F,\Theta)&\leq \operatorname{dim} F+r-t\\
&\leq (\operatorname{dim} Y-1)+(\rho(Y)-1)-d\\
&=\gamma(Y,\Gamma) -2\\
&<0,
\end{align*}
a contradiction. Hence $k=0$ and we may apply \eqref{l_mori}.
\end{proof}
\section{Reduction to Mori dream spaces}
\label{s_reduction}
\begin{lemma}\label{l_up} Let $X$ be a $\mathbb{Q}$-factorial projective variety.
Suppose that $(X,\Delta)$ is a divisorially log terminal pair and $-(K_X+\Delta)$ is nef.
Suppose that we may find a big and nef $\mathbb{Q}$-divisor divisor $A$ and a kawamata log
terminal pair $(X,\Delta_0)$ such that $K_X+\Delta_0+A \sim_{\mathbb{R}} N\geq 0$ has
numerical dimension zero.
Then we may find a divisorially log terminal modification $\pi\colon\map Y.X.$ of
$(X,\Delta)$, a big and nef $\mathbb{Q}$-divisor $B$ and a kawamata log terminal pair
$(Y,\Gamma_1)$ such that $K_Y+\Gamma_1+B \sim_{\mathbb{R}} L\geq 0$ has numerical
dimension zero, $\Gamma_1$ and $L$ have no common components and no non kawamata log
terminal centre of $(Y,\Gamma)$ is contained in the support of $L$.
\end{lemma}
\begin{proof} \eqref{l_minimal} implies that replacing $\Delta_0$ and $N$ we may assume
that $\Delta_0$ and $N$ have no common components. Pick $t>0$ such that $\rfdown tN.=0$.
Let $\pi\colon\map Y.X.$ be a divisorially log terminal modification of $(X,\Delta+tN)$.
Then $\pi$ has finitely many exceptional divisors. If $E$ is an exceptional divisor which
is not a log canonical place of $(X,\Delta)$ then $E$ is not a log canonical place of
$(X,\Delta+\delta N)$ for $\delta>0$ sufficiently small. Thus replacing $t$ by $\delta>0$
sufficiently small, we may assume that $\pi$ is also a divisorially log terminal
modification of $(X,\Delta)$.
If we write
\[
K_Y+\Gamma=\pi^*(K_X+\Delta),
\]
then $(Y,\Gamma)$ is divisorially log terminal and $-(K_Y+\Gamma)$ is nef. We may also
write
\[
K_Y+\Gamma_0=\pi^*(K_X+\Delta_0) \qquad \text{and} \qquad B=\pi^*A.
\]
If $L=\pi^*N$ then $\operatorname{mult}_EL>0$ for every exceptional divisor $E$ whose centre $V$ is
contained in the support of $N$. As
\[
K_Y+\Gamma_0+B \sim_{\mathbb{R}} L,
\]
by \eqref{l_minimal} we may find $B_1$ big and nef and
$\Gamma_1\geq (1-\delta)\Gamma\geq 0$ such that
\[
K_Y+\Gamma_1+B_1 \sim_{\mathbb{R}} tL_1,
\]
where $L=L_0+L_1$ and $L_1$ has no common components with $\Gamma$. But then no non
kawamata log terminal centre of $(Y,\Gamma)$ is contained in the support of $L$.
\end{proof}
\begin{lemma}\label{l_less} Let $X$ be a $\mathbb{Q}$-factorial projective toric variety.
Let $B\geq 0$ be an $\mathbb{R}$-Cartier divisor whose support contains all but one
invariant divisor. Let $\nu$ be a valuation which is not toric.
Then we may find a divisor $0\leq B' \sim_{\mathbb{R}} B$ such that $\nu(B')>\nu(B)$
whilst $\mu(B')\leq \mu(B)$ for every toric valuation $\mu$.
If further $(X,B)$ is a log canonical pair such that every log canonical place is toric
then we may pick $B'$ such that $(X,B')$ is log canonical and the only log canonical
places of $(X,B')$ are toric valuations.
\end{lemma}
\begin{proof} We prove the first statement.
Let $\pi\colon\map Y.X.$ be a birational morphism of toric varieties. As the support of
$B$ contains every invariant subset of codimension at least two, it follows that
$\pi^*B\geq 0$ is an $\mathbb{R}$-Cartier divisor whose support contains all but one
invariant divisor. Thus we are free to replace $X$ by $Y$ and $B$ by $\pi^*B$. In
particular we may assume that $X$ is smooth. We are also free to replace $B$ by a
multiple.
Let $W$ be the centre of $\nu$ and let $V$ be the smallest invariant subset of $X$ which
contains $W$. If $W=V$ then let $\pi\colon\map Y.X.$ blow up $V$. Replacing $X$ by $Y$
and repeating this procedure finitely many times we reduce to the case $W\neq V$.
\cite[1.2]{Tevelev07} implies that we may find a birational morphism of toric varieties
$\map V'.V.$ such that the strict transform $W'$ of does not contain any invariant subsets
of $V'$. Let $\map Y.X.$ be a birational morphism of toric varieties, which is an
isomorphism at the generic point of $V$, such that if $U$ is the strict transform of $V$
then the birational morphism $\map U.V.$ factors through $\map V'.V.$.
\cite[1.2]{Tevelev07} implies that, replacing $X$ by $Y$, we may reduce to the case when
$W$ does not contain any invariant subsets.
Note that by \eqref{l_missing} we may find a divisor $C\geq 0$ supported on the invariant
components of $B$ not containing $V$ such that $A=C|_V$ is very ample and we can lift
elements of the linear system $|A|$.
Pick a birational morphism $\pi\colon\map Y.X.$ of toric varieties such that the mobile
part of $\pi^*C$ is base point free. Replacing $X$ by $Y$ we may assume that the mobile
part $B_0$ of $C$ is base point free. Note that $B_0$ and $C$ are the same in a
neighbourhood of $V$. Replacing $B$ by a multiple we may assume $B_1=B-B_0\geq 0$.
Let $f\colon\map X.Z.$ be the contraction morphism associated to $B_0$. Then the
restriction of $f$ to $V$ is an isomorphism. As $W$ does not contain any invariant
subsets, it follows that $f(W)$ does not contain any invariant subsets and so $f(W)$ does
not contain the image of any invariant subvariety of $X$.
As $B_0$ is the pullback of very ample divisor from $Z$, we may pick
$0\leq B_0'\sim_{\mathbb{R}} B_0$ such that $B'_0$ contains $W$ and
$\mu(B_0')\leq \mu(B_0)$ for all toric valuations $\mu$. If $B'=B'_0+B_1$ then
$0\leq B' \sim_{\mathbb{R}} B$, $\nu(B')>\nu(B)$, whilst $\mu(B')\leq \mu(B)$ for every
toric valuation $\mu$.
Now suppose that $(X,B)$ is log canonical. If $B_t=tB'+(1-t)B$ then
$0\leq B_t\sim_{\mathbb{R}} B$, $\nu(B_t)>\nu(B)$ for $t>0$, $(X,B_t)$ is log canonical if
$t$ is sufficiently small and the only log canonical places are toric valuations.
\end{proof}
\begin{lemma}\label{l_dream} Assume \eqref{t_toric}$_{n-1}$.
Let $X$ be a $\mathbb{Q}$-factorial projective variety of dimension $n$. Suppose that
$(X,\Delta)$ is a divisorially log terminal pair and $-(K_X+\Delta)$ is nef.
If the complexity of $(X,\Delta)$ is less than one then $X$ is of Fano type. In
particular $X$ is a toric variety.
\end{lemma}
\begin{proof} \eqref{p_numerical} implies that we may find a big and nef
$\mathbb{Q}$-divisor divisor $A$ and a kawamata log terminal pair $(X,\Delta_0)$ such that
$K_X+\Delta_0+A \sim_{\mathbb{R}} N\geq 0$ has numerical dimension zero.
By \eqref{l_up} possibly replacing $X$ by a higher model we may assume that $\Delta$ and
$N$ have no common componentd and that no non kawamata log terminal centre of $(X,\Delta)$
is contained in the support of $N$. In particular we may pick $\epsilon>0$ such that no
non kawamata log terminal centre of $(X,\Delta+\epsilon N)$ is contained in $N$.
Let $\pi\colon\rmap X.Y.$ be a log terminal model of $(X,\Delta_0+A)$. Set $B=\pi_*A$,
$\Gamma_0=\pi_*\Delta_0$ and $\Gamma=\pi_*\Delta$. Then $(Y,\Gamma_0+B)$ is a kawamata
log terminal pair, $\Gamma_0+B$ is big and $K_Y+\Gamma_0+B$ is numerically trivial so that
$Y$ is of Fano type. In particular \eqref{t_mdscase} implies that $Y$ is a toric variety.
If $\pi$ does not contract any divisors then $N=0$, $K_X+\Delta_0+A$ is numerically
trivial and so $\pi$ is an isomorphism.
Pick an exceptional divisor $E$ of $\pi$ and let $\nu$ be the corresponding valuation. If
$\nu$ is a toric valuation of $Y$ for every exceptional divisor $E$ then $X$ is log Fano.
So we may assume that $\nu$ is not toric. $E$ is a component of $N$ and so $\nu$ is not a
log canonical place of $(Y,\Gamma)$. \eqref{l_less} implies that we may find
$0\leq \Gamma' \sim_{\mathbb{R}} \Gamma$ such that $\nu(\Gamma')>\nu(\Gamma)$,
$(Y,\Gamma')$ is log canonical and the only log canonical places are toric valuations.
Let $\Delta'$ be the strict transform of $\Gamma'$. Then certainly $(X,\Delta')$ is log
canonical outside of the support of $N$, since the indeterminancy locus of $\pi$ is
contained in the support of $N$.
Let $\Delta_s=s\Delta'+(1-s)\Delta$. As $N$ contains no non kawamata log terminal centre
of $(X,\Delta)$, it follows that $N$ contains no non kawamata log terminal centre of
$(X,\Delta_s+\epsilon N)$ for $s$ sufficiently close to zero. In particular
$(X,\Delta_s)$ is log canonical for $s$ sufficiently close to zero. Replacing
$(Y,\Gamma')$ by $(Y,\Gamma_s=\pi_*\Delta_s)$ for $s$ sufficiently close to zero, we may
assume that $(X,\Delta')$ is log canonical and $N$ contains no non kawamata log terminal
centre of $(X,\Delta'+\epsilon N)$. We may write
\begin{align*}
K_X+\Delta&=\pi^*(K_Y+\Gamma)+F\\
K_X+\Delta'&=\pi^*(K_Y+\Gamma')+F'
\end{align*}
where $F$ and $F'$ are exceptional. Note that the coefficients of $F'$ are linear
functions of $s$ and so we may pick $s$ sufficiently close to zero so that
$F-F'\leq \epsilon N$.
Let
\[
E_t=t(F'-F)+(1-t)N,
\]
Decompose $E_t$ as $E^+_t-E^-_t$, where $E^{\pm}_t\geq 0$ and $E_t^+$ and $E_t^-$ have no
common components. If we put
\[
\Delta_t=(1-t)\Delta_0+t\Delta'+E_t^- \qquad \text{and} \qquad A_t=(1-t)A+tM
\]
then $(X,\Delta_t)$ is kawamata log terminal for $t\in [0,1)$ and we have
\begin{align*}
K_X+\Delta_t+A_t &=(1-t)(K_X+\Delta_0+A)+t(M+K_X+\Delta')+E_t^-\\
&=(1-t)(K_X+\Delta_0+A)+t(\Delta'-\Delta)+E_t^-\\
&\sim_{\mathbb{R}} (1-t)N+t(F'-F)+E_t^-\\
&=E_t+E_t^-\\
&=E_t^+.
\end{align*}
By assumption $\operatorname{mult}_EF'<\operatorname{mult}_EF$, so that $E_t^-\neq 0$ for $t$ sufficiently close to
$1$. Thus $E_t^+$ has fewer components than $N$ and we are done by induction on the
number of components of $N$.
\end{proof}
\section{Rationality via the Cox ring}
\label{s_rational}
\begin{proposition}\label{p_quadricmds} Let $X$ be a $\mathbb{Q}$-factorial projective variety.
Suppose that the Cox ring of $X$ is a polynomial ring modulo a single relation $Q$,
\[
\operatorname{Cox}(X)=\frac{k[\llist x.n.]}{\langle Q\rangle},
\]
where $Q$ and $\llist x.n.$ are homogenous elements of $\operatorname{Cox}(X)$.
If the rank of the quadratic part of $Q$ is at least two then there is a proper finite
morphism $\map Y.X.$ of degree at most two, which is \'etale outside a closed subset of
codimension at least two, such that $Y$ is rational.
In particular if $A_{n-1}(X)$ has no $2$-torsion then $X$ is rational.
\end{proposition}
\begin{proof} $R=\operatorname{Cox}(X)$ is a multigraded ring, and this grading corresponds to the
action of a diagonalisable group $H$ on $\operatorname{Spec} R$. $X$ is a GIT quotient of $\operatorname{Spec} R$ by
$H$. The action of $H$ extends to the polynomial ring $k[\llist x.n.]$ and the GIT
quotient of the corresponding affine space is a toric variety $Z$ which contains $X$ as a
divisor; the relation $Q$ is homogeneous for this action.
Let $T$ be the torus of $Z$. The monomials in the coordinate ring $k[M]$ of the torus are
Laurent monomials in the variables $\llist x.n.$ of multi-degree $0$ in the grading.
Suppose that $x_ix_j\in Q$ for $i\neq j$. Possibly permuting the coordinates we may
assume that $x_1x_2\in Q$. Collecting together all of the terms divisible by $x_1$, we
may write
\[
Q=x_1(x_2+q_0)+q_1,
\]
where $q_1$ is a polynomial in $x_2$, $x_3$, \dots, $x_n$. After the homogeneous change
of variable,
\[
\map x_i..\begin{cases} x_2+q_0 & \text{if $i=2$} \\
x_i & \text{otherwise},
\end{cases}
\]
we may write
\[
Q=x_1x_2-q
\]
where $q$ is a polynomial in $x_2$, $x_3$, \dots, $x_n$.
As $Q$ is homogeneous and $q$ is not equal to zero, we may find a monomial $\nu\in q$ in
the variables $x_2$, $x_3$, \dots, $x_n$, with the same multi-degree as $x_1x_2$. If we
set
\[
\mu=\frac{x_1x_2}{\nu}
\]
then $\mu$ is a Laurent monomial in the variables $\llist x.n.$ of multi-degree zero.
Therefore $\mu\in k[M]\subset k[\mathbb{Z}^n]$. As the degree of $x_1$ is one in $\mu$ it
follows that $\mu=\mu_1$ corresponds to a primitive element $m_1$ of the lattice $M$ and
we may extend it to a basis $\llist m.p.$ of $M$, where the first coordinate of $m_i$ is
zero, for $i>1$. If $\llist \mu.p.$ are the Laurent monomials corresponding to
$\llist m.p.$ of $M$, then $\mu_i$, $i>1$ are Laurent monomials in the variables $x_2$,
$x_3$, \dots, $x_n$.
Now let $G=\mathbb{G}_m$ act by $t$ on $\mu$ and trivially on all other basis elements.
Let $U$ be the open subset of the torus where $q\neq 0$. On $X$ we have
\[
\mu=\frac{x_1x_2}{\nu}=\frac q{\nu}.
\]
As the RHS is invariant under the action of $G$ and the LHS is not, it follows that the
orbits of $G$ intersect $X\cap U$ in a unique point. Thus $X$ is birational to $U/G$ and
so $X$ is rational.
Otherwise assume $x_ix_j\notin Q$, for all $i\neq j$. Since the quadratic part of $Q$ has
rank at least two, $x_i^2$, $x_j^2\in Q$, for $i\neq j$. Possibly permuting the
coordinates we may assume that $x_1^2$, $x_2^2\in Q$. Rescaling we may assume that
\[
Q=x_1^2-x_2^2+q,
\]
where the quadratic part of $q$ is a polynomial in the variables $x_3$, $x_4$, \dots,
$x_n$.
As $Q$ is homogeneous, $x_1^2$ and $x_2^2$ have the same multi-degree. If $x_1$ and $x_2$
have the same multi-degree then both $x_1+x_2$ and $x_1-x_2$ are homogeneous and after the
change of coordinates
\[
\map x_i..\begin{cases} x_1+x_2 & \text{if $i=1$} \\
x_1-x_2 & \text{if $i=2$} \\
x_i & \text{otherwise},
\end{cases}
\]
$x_1x_2\in Q$ so that $X$ is rational by what we already proved.
Otherwise $x_1/x_2$ is torsion of degree two. By definition of the Cox ring, there is a
Weil divisor $D$ on $X$ such that $2D\sim 0$. $D$ defines a proper finite morphism
$\map Y.X.$ of degree two, which is \'etale outside a closed subset of codimension at
least two.
On the other hand, $E$ lifts to a Weil divisor on $Z$ such that $2E\sim 0$. $E$ also
defines a proper finite morphism $\map W.Z.$ of degree two, which is also \'etale outside
a closed subset of codimension at least two, and induces the original cover $\map Y.X.$.
$W$ is a toric variety. Moreover $W$ has the same Cox ring as $Z$ but with a grading
given by setting the class of $D$ equal to zero. This grading corresponds to the action
of a diagonalisable group $G$ on $\operatorname{Spec} R$ and $W$ is a GIT quotient of $\operatorname{Spec} R$ by $G$.
$Y$ is a divisor in $W$, defined by the same equation as $X$ inside $Z$. On $W$, however,
$x_1$ and $x_2$ have the same multi-degree and so $Y$ is rational by what we have already
proved. \end{proof}
Note that $Y$ does not necessarily have the same Cox ring as $X$, since $Y$ might have
more divisors than $X$, as happens in the example in \S \ref{s_example}.
\section{An irrational example}
\label{s_example}
In this section we give an example of an irrational projective threefold $X$, together
with a log canonical pair $(X,\Delta)$ of absolute complexity $1$ such that $K_X+\Delta$
numerically trivial. In particular the condition on torsion in $A_{n-1}(X)$ in
\eqref{p_quadricmds} is necessary. We will construct $X$ as a $\mathbb{Z}_2$-quotient of
a conic bundle $Y$ over $T=\pr 1.\times \pr 1.$.
Pick bihomogeneous coordinates $([y_0:y_1],[z_0:z_1])$ on $T$. If $q$ is a general
polynomial of degree $(2d,2d)$ in the monomials $y_i^2$, $z_i^2$ and $y_iz_i$,
$i\in \{\,0,1\,\}$, where $d>3$, then the zero locus of $q$ is a smooth curve $D$ which
contains none of the invariant points. The equation
\[
x_0^2-x_1^2=q(y_0,y_1,z_0,z_1)x_2^2,
\]
defines a divisor $Y$ inside the projectivisation $W$ of
\[
\ring {\pr 1.\times \pr 1.}.\oplus \ring {\pr 1.\times \pr 1.}. \oplus \ring {\pr 1.\times \pr 1.}.(-d,-d).
\]
$Y$ is a conic bundle $f\colon\map Y.T.$ over $T$. Let $\Theta$ be the divisor on $W$ given by
the sum of the vanishing of $x_2$, together with the pullbacks of the torus invariant
divisors from $\pr 1. \times \pr 1.$, and set $\Gamma=\Theta|_Y$.
\begin{lemma}\label{l_y}\
\begin{enumerate}
\item $(Y,\Gamma)$ is log smooth,
\item $K_Y+\Gamma=0$, and
\item $f\colon\map Y.T.$ has relative Picard number two.
\end{enumerate}
\end{lemma}
\begin{proof} To prove (1) and (2), by adjunction it suffices to check that $(W,Y+\Theta)$
is log smooth and $K_W+Y+\Theta=0$.
Consider the linear series on $W$ spanned by $x_0^2$, $x_1^2$, and $x_2^2m$, where $m$
ranges over all monomials in $y_i^2$, $z_i^2$ and $y_iz_i$, $i\in \{\,0,1\,\}$, of degree
$(2d,2d)$. Rescaling $x_0$ and $x_1$ we see that $Y$ is a general member of this linear
series. On the other hand this linear series is base point free and so $Y$ is smooth and
intersects all torus invariant strata of $W$ transversely. This gives (1) and (2).
For (3) note that the fibres of $f$ are irreducible except over the curve $D$, where the
fibres have two components. It follows that the relative Picard number is at most two.
On the other hand, the inverse image of $D$ consists of two prime divisors
$D_1=V(x_0+x_1,q)$ and $D_2=V(x_0-x_1,q)$. This is (3).
\end{proof}
Consider the $\mathbb{Z}_2$ action on $W$ sending
\[
\map (y_0,y_1,z_0,z_1,x_0,x_1,x_2).(y_0,-y_1,z_0,-z_1,x_0,-x_1,x_2)..
\]
This action also defines an action on $T$ and under this action both $Y$ and $D$ are
invariant. If $X$ is the quotient of $Y$ and $S$ is the quotient of $T$ then there is a
commutative diagram
\[
\begin{diagram}
Y & \rTo & X \\
\dTo^f & & \dTo^g \\
T & \rTo & S.
\end{diagram}
\]
Note that $g\colon\map X.S.$ is a conic bundle. Note also that the action on $T$ is
toric, fixing only the four torus invariant points and so $S$ is also a toric surface with
four $A_1$ singularities.
Let $\Delta$ be the image of $\Gamma$.
\begin{proposition} \
\begin{enumerate}
\item $(X,\Delta)$ is log canonical
\item $K_X+\Delta=0$,
\item the absolute complexity $\gamma$ is one, and
\item $X$ is irrational.
\end{enumerate}
\end{proposition}
\begin{proof} Note that as $\map Y.X.$ is \'etale in codimension one, (1) and (2) follow
easily.
The action of $\mathbb{Z}_2$ switches the divisors $D_1$ and $D_2$. Thus $X$ has relative
Picard number $1$ over $S$ and so the absolute complexity of $(X,\Delta)$ is
\[
\gamma=\operatorname{dim} X+\rho(X)-d=3+3-5=1.
\]
This is (3).
If $X$ is rational then the Griffiths component of the intermediate Jacobian must be
trivial. If $C$ is the discriminant curve of the conic bundle $g\colon\map X.S.$ and the
Griffiths component is trivial then $C$ is hyperelliptic, trigonal, or isomorphic to a
plane quintic, cf. \cite{Shokurov83}.
$C$ is certainly not a plane quintic. On the other hand $C$ is a smooth quotient of $D$
by the fixed point-free action of $\mathbb{Z}_2$. It suffices to check that $D$ has no
$g_6^1$ (since if $D$ has a $g^1_3$ it has a $g^1_6$). But it follows from a theorem of
Martens \cite{Martens96} that if $D$ has a $g^1_k$ then $k\geq 2d$.
\end{proof}
\section{Proofs}
\label{s_proofs}
\begin{proof}[Proof of \eqref{t_toric}] We proceed by induction on the dimension $n$ of
$X$.
\eqref{p_dlt} implies that we may find a divisorially log terminal model
$\pi\colon\rmap Y.X.$, such that if we write
\[
K_Y+\Gamma=\pi^*(K_X+\Delta),
\]
then $-(K_Y+\Gamma)$ is nef. \eqref{l_dlt} implies that the complexity of $(Y,\Gamma)$ is
at most the complexity of $(X,\Delta)$. \eqref{l_persist} implies that if $(Y,G)$ is a
toric pair then $(X,D)$ is a toric pair, where $D=\pi_*G$. If $G\geq \sship \Gamma.$ then
$D\geq \sship \Delta.$ and if all but $\rfdown 2c.$ invariant divisors are components of
$\Gamma$ then all but $\rfdown 2c.$ invariant divisors are components of $\Delta$.
Replacing $(X,\Delta)$ by $(Y,\Gamma)$ we may assume that $X$ is a $\mathbb{Q}$-factorial
projective variety and $(X,\Delta)$ is divisorially log terminal.
\eqref{l_dream} implies that $X$ is of Fano type. Thus $X$ is a Mori dream space and we
are done by \eqref{t_mdscase}.
\end{proof}
\begin{proof}[Proof of \eqref{c_complexity}] It it immediate from \eqref{t_toric} that
$\rdown 2c.\geq 0$.
\end{proof}
\begin{proof}[Proof of \eqref{c_span}] If $c<1$ then \eqref{t_toric} implies that $X$ is
toric and all but one invariant divisor is a component of $\Delta$. On the other hand the
invariant divisors span the N\'eron-Severi group and any invariant divisor is in the span
of the other invariant divisors.
\end{proof}
\begin{proof}[Proof of \eqref{c_non}] Let $\bar k$ be the algebraic closure of $k$ and let
bars denote extension to the algebraic closure.
Then $(\bar X,\bar \Delta)$ is log canonical and $-(K_{\bar X}+\bar \Delta)$ is nef. If
$\sum a_iS_i$ is a decomposition of $\Delta$ of complexity less than one then
$\sum a_i\bar S_i$ is a decomposition of $\bar \Delta$ of complexity less than one.
\eqref{t_toric} implies that there is a divisor $\bar D$ such that $(\bar X,\bar D)$ is
toric.
Let $m$ be the number of invariant divisors. Possibly reordering, we may assume that
$\llist \bar S.m-1.$ are invariant divisors. In particular $\llist S.m-1.$ are prime
divisors. Consider the linear system
\[
|-(K_X+\sum_{i\leq m-1} S_i)|.
\]
If $\bar D_m$ is the last invariant divisor on $\bar X$ then
\[
\bar D_m\in |-(K_{\bar X}+\sum_{i\leq m-1} \bar S_i)|.
\]
Thus the linear system
\[
|-(K_X+\sum _{i\leq m-1} S_i)|
\]
is non-empty and we may find $S_m$ such that $K_X+D$ is linearly equivalent to zero and
$(X,D=\sum _{i\leq m} S_i)$ is log canonical.
In this case $(\bar X,\bar D)$ is toric, again by \eqref{t_toric}. Replacing $(X,\Delta)$
by $(X,D)$ we may assume that $\Delta=D$. In this case every component of $\bar D$ is an
invariant divisor and $\llist S.m.$ are all prime divisors. By induction
$(S_i,(D-S)|_{S_i})$ is a toric pair. In particular the strata of $D$ are geometrically
irreducible.
We may find $\bar\pi\colon\map \bar Y.\bar X.$ a birational morphism of toric varieties
such that $\bar Y$ is projective and there is a birational morphism to projective space,
$\bar g\colon\map \bar Y.{\gpr n.\bar k.}.$. As the strata of $D$ are geometrically
irreducible, there is a birational morphism $\pi\colon\map Y.X.$ which extracts only
divisors of log discrepancy zero. If we write
\[
K_Y+G=\pi^*(K_X+D),
\]
then $G$ is the sum of the strict transform of $D$ and the exceptional divisors. It is
enough to prove that $(Y,G)$ is toric by \eqref{l_morphism}.
Replacing $(X,D)$ by $(Y,G)$ we may assume that there is a birational morphism
$\bar f\colon\map \bar X.{\gpr n.\bar k.}.$. Pulling back an invariant hyperplane, this
linear system is given by a sum of invariant divisors $\sum b_j\bar D_j$. Consider the
linear system $|\sum b_jD_j|$. This is base point free, has dimension $n$ and separates
points. Thus we get a birational map to projective space $f\colon\map X.{\gpr n.k.}.$
such that $\bar f$ is toric.
In particular $f$ only extracts divisors of log discrepancy zero. \eqref{l_persist}
implies that $(X,D)$ is a toric pair.
\end{proof}
\begin{proof}[Proof of \eqref{t_form}] As \eqref{t_toric} holds in all dimensions
\eqref{p_numerical} implies the first statement.
Let $c$ be the complexity of $(X,\Delta)$. Pick $\delta>0$ such that
$A+(1-\delta)\Delta_0$ is ample. Replacing $\Delta_0$ by $(1-\delta)\Delta_0$ and $A$ by
$A+(1-\delta)\Delta_0$ we may assume that $(X,\Delta_0)$ is kawamata log terminal. Let
$f\colon\rmap X.Y.$ be a log terminal model of $(X,A+\Delta_0)$. Replacing $X$ by $Y$ we
may assume that $K_X+A+\Delta_0$ is nef. In this case $K_X+A+\Delta_0$ is semiample. Let
$f\colon\map X.W.$ be the induced model. Then $-(K_X+\Delta_0)$ is ample over $W$.
\cite{HM05a} implies that the fibres of $f$ are rationally connected. Thus $f$ factors
through the maximal rationally connected fibration $\rmap X.Z.$. It follows that
\[
\operatorname{dim} Z\leq \operatorname{dim} Y\leq \nu(X,A+\Delta_0) \leq c. \qedhere
\]
\end{proof}
\begin{proof}[Proof of \eqref{t_rational}] \eqref{p_dlt} implies that we may find a
divisorially log terminal model $\pi\colon\rmap Y.X.$, such that if we write
\[
K_Y+\Gamma=\pi^*(K_X+\Delta),
\]
then $-(K_Y+\Gamma)$ is nef. \eqref{l_dlt} implies that the absolute complexity of
$(Y,\Gamma)$ is at most the absolute complexity of $(X,\Delta)$. Replacing $(X,\Delta)$
by $(Y,\Gamma)$ we may assume that $X$ is a $\mathbb{Q}$-factorial projective variety and
$(X,\Delta)$ is divisorially log terminal.
By \eqref{c_numerical} we may find a divisorially log terminal pair $(Y,\Gamma)$ such that
$-(K_Y+\Gamma)$ is ample, the absolute complexity is less than two and $Y$ is birational
to $X$. Replacing $(X,\Delta)$ by $(Y,\Gamma)$ we may assume that $-(K_X+\Delta)$ is
ample. \eqref{l_mori} implies that $X$ is a Mori dream space. Pick
$B \sim_{\mathbb{R}} -(K_X+\Delta)$ such that $(X,B+\Delta)$ is divisorially log terminal.
Replacing $(X,\Delta)$ by $(X,B+\Delta)$ we may assume that $K_X+\Delta$ is numerically
trivial.
Let $R=\operatorname{Cox}(X)$ be the Cox ring of $X$, $Y=\operatorname{Spec} R$, and $\Gamma$ the divisor on $Y$
corresponding to $\Delta$. Then every component of $\Gamma$ is Cartier and $K_Y$ is
Cartier. \cite{KO12} implies that $(Y,\Gamma)$ is log canonical as $(X,\Delta)$ is log
canonical.
By \eqref{l_flipsabundance} $Y$ has a $cA_l$ singularity at the point $p$. If $Y$ is
smooth then $X$ is a toric variety and there is nothing to prove. Otherwise $\operatorname{Cox}(X)$ is
a polynomial ring modulo a single relation $Q$, where the rank of the quadratic part of
$Q$ is at least two. Thus we may apply \eqref{p_quadricmds}. \end{proof}
\bibliographystyle{/home/mckernan/Jewel/Tex/hamsplain}
|
\section{INTRODUCTION}
\label{sec:intro}
Following the trend in many other fields, the advent of high-volume and weakly-supervised data is driving increased interest in \emph{large-scale} sentiment studies in Affective Computing \cite{picard_1997}.
However, directly studying affect, or emotion, in a dimensional representation (e.g.~valence-arousal-dominance) or even discrete semantics (e.g.~sad, happy, angry, etc.) tend to suffer from the problem of data sparsity since such specialized psychology terminology are less likely to be found in large volume from the Web.
Inspired by similar work in Computer Vision and Multimedia, several works have since proposed affective mid-level representations to bridge the \emph{affective gap} between low-level features and high-level affect states.
One such attractive mid-level semantic representation is the \emph{adjective-noun pair} (ANP) \cite{borth_2013_vso}, where `nouns' are used to provide visual grounding to a detectable object and `adjectives' are used to sentimentally bias the object.
In a recent work, we presented a large-scale multilingual visual sentiment ontology (MVSO) \cite{jou_2015} consisting of 15,630 such ANPs across 12 languages with an associated dataset of over 7.36 million images.
As part of this initial work, we developed a set of language-specific visual concept detector banks that could identify the presence of ANPs in six of the major languages in the MVSO.
These detector banks were shown to be useful in cross-lingual sentiment analysis \cite{jou_2015}, and recently, in diversifying image query expansion \cite{li_2016} and facial imagery understanding \cite{pappas_2016}.
Yet in the original development of the ANP detectors, many of the original training parameter settings were defaulted and left as-is for a proof-of-concept.
Even though using such default settings show that the multilingual ANP detectors can achieve reasonable detection performance already, they do not represent the most near-optimal networks that can be trained for the given architectures used.
In this report, we detail and release
(1) higher accuracy models for detecting adjective-noun pairs (ANPs) in six languages from the same image pool in the original Multilingual Visual Sentiment Ontology (MVSO) release \cite{jou_2015} and new detector banks fine-tuned using a different more modern network architecture with benchmark comparisons,
(2) a sub-corpus of MVSO based on tag-restricted queries for higher fidelity labels, and
(3) adjective-noun pair detectors based on this tag-restricted subset.
The model and data release can be found through the original MVSO website at \url{http://mvso.cs.columbia.edu}.
\section{BRIEF OVERVIEW OF MVSO}
\label{sec:mvso_overview}
The Multilingual Visual Sentiment Ontology (MVSO) \cite{jou_2015}, a substantial extension and improvement on the Visual Sentiment Ontology (VSO) \cite{borth_2013_vso}, consists of over 15.6K sentiment-biased mid-level visual concepts using the semantic construct of adjective-noun pairs (ANPs).
The ontology was constructed by first using seed emotion keywords from ``Plutchik's Wheel of Emotions'' \cite{plutchik_1980} where were translated into 12 different languages by native speakers.
For each language, these keywords were used to query the Flickr API\footnote{\url{https://www.flickr.com/services/api}} to retrieve a large corpus of images with related tags and other metadata.
From the tags and metadata, we mined popularly used adjectives and nouns, identified by language-dependent part-of-speech taggers, combinatorially formed adjective-noun pair (ANP) candidates, and then filtered them on a number of criteria including semantic correctness, sentiment strength, popular usage on Flickr, and coverage in Flickr uploader diversity.
After this candidate filtering, the MVSO consisted of 15,630 ANP concepts representative of 12 different languages (language codes in brackets): Arabic {\bf [ar]}, Chinese {\bf [zh]}, Dutch {\bf [nl]}, English {\bf [en]}, French {\bf [fr]}, German {\bf [de]}, Italian {\bf [it]}, Persian {\bf [fa]}, Polish {\bf [pl]}, Russian {\bf [ru]}, Spanish {\bf [es]} and Turkish {\bf [tr]}.
These ANP concepts were then used to query the Flickr API to retrieve associated images and metadata, and it was later found that our ANPs also have geo-reference coverage of over 235 countries \cite{jou_2016_senticart}.
To develop an actual ontology with a semantic tree structure, we proposed two construction methods: \emph{exact matching} and \emph{approximate matching} \cite{jou_2015,pappas_2016}.
In `exact matching,' we use a pivot language where all ANPs are automatically translated into English and ANPs become naturally grouped by their translations.
This yields high cluster ``precision,'' but low ``recall'' since nuances like plurals and synonyms, and even related concepts are not grouped together.
In `approximate matching,' we extract word feature vectors from the ANPs and perform traditional distance-based clustering.
This trades of some cluster consistency for increased cluster coverage.
These clustering approaches have proven useful in uncovering some multilingual geographic social phenomena \cite{jou_2016_senticart} as well as in cross-lingual sentiment analysis and query expansion \cite{jou_2015,li_2016}.
Samples images from MVSO for an ANP cluster with concepts like `old house' are shown in Figure \ref{fig:old_house}.
The final MVSO image corpus consists of 7,368,364 images from which language-specific ANP detector banks were trained.
Six of the major languages were chosen to train these detectors, resulting in a coverage of 9,918 total ANPs across 6,994,319 images in English {\bf [en]}, Spanish {\bf [es]}, Italian {\bf [it]}, French {\bf [fr]}, German {\bf [de]} and Chinese {\bf [zh]}.
For more details on MVSO, we refer the interested reader to the related peer-review publications \cite{jou_2015,jou_2016_senticart,li_2016,pappas_2016}.
\begin{figure}[h]
\centering
\includegraphics[width=5.3in]{old_house.png}
\caption{Example MVSO Images in Four Sampled Languages from a Multilingual ANP Cluster. The cluster, formed by approximate matching, portrays concepts related to the ``old'' or ``ancient'' nature of some ``houses''.}
\label{fig:old_house}
\end{figure}
\section{DEEP VISUAL SENTIMENT CONCEPT DETECTOR BANKS}
In our original MVSO publication \cite{jou_2015}, we trained six detector banks, fine-tuning using an AlexNet-styled network architecture where there are five convolutional layers and three fully-connected layers \cite{krizhevsky_2012} with pooling and normalization layers swapped to create a variant called CaffeNet \cite{jia_2014}.
We initialized our networks weights from a CaffeNet fine-tuned network called DeepSentiBank \cite{chen_2014_dsb}, which was trained to detect 2,089 ANPs (in English) from the images of the VSO dataset \cite{borth_2013_vso} and improving on its original proposed SVM-based detector banks, SentiBank \cite{borth_2013_sb}.
In fine-tuning these six detector banks, we followed the traditional fine-tuning strategy of resizing the final fully-connected layer to have the same number of output units as the number of classes with random initialization (in this case, Gaussian), started optimization at a lower base learning rate (in this case, 0.001), and used a higher learning rate multiplier for the last fully-connected layer to allow for more aggressive updates on that layer compared to the other layers that were pretrained (in this case, $\times$2).
During optimization, we used a batch size of 256 and fixed the learning rate, gradually reducing it by a factor of 10 whenever we saw training loss plateau.
Every other setting in the network architecture, e.g.~dropout rates and output sizes of other fully-connected layers, and training routines, e.g.~optimization technique, was kept the same as when training the original ILSVRC \cite{imagenet} model\footnote{See \url{https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_caffenet} for parameter details.}.
We note that in these MVSO detectors, the original ILSVRC12 mean image was used for mean subtraction.
All training was originally performed on a single NVIDIA GeForce GTX 980 GPU and was implemented with Caffe \cite{jia_2014}.
The performances in the original MVSO release \cite{jou_2015} are copied in Table \ref{tab:old_classifiers}.
To further improve the detection performance of ANPs, we experimented with a much deeper and more complex architecture called GoogLeNet, or Inception \cite{szegedy_2015}.
The original GoogLeNet design consists of mini-networks, also called Inception modules, that are composed of 1$\times$1\emph{C}, 1$\times$1\emph{C}-3$\times$3\emph{C}, 1$\times$1\emph{C}-5$\times$5\emph{C}, and 3$\times$3\emph{MP}-1$\times$1\emph{C} towers where \emph{C} corresponds to convolutions and \emph{MP} denotes max pooling, i.e.~so 1$\times$1\emph{C}-3$\times$3\emph{C} denotes one path in the modules that has a 1$\times$1 convolution followed by a 3$\times$3 convolution.
While AlexNet-styled architectures are only 8 layers deep, the GoogLeNet architecture is 22 layers deep while still using less parameters\footnote{Note that here we use the first iteration of GoogLeNet/Inception, sometimes referred to as Inception-v1, though there have been several proposed improvements since that use double 3$\times$3 layers or that integrate residual learning.}.
These fan-in-fan-out modules are stacked in an architecture referred to as network-in-network \cite{lin_2014}.
GoogLeNet additionally uses two auxiliary branches in the network to prevent gradients from vanishing during backpropagation \cite{szegedy_2015}.
For training, we preserved the same training setting as in the original MVSO release \cite{jou_2015} using the same train/test split of images.
Unlike before though we pre-train from a ILSVRC12 model since the architecture is significantly different from DeepSentiBank \cite{chen_2014_dsb} and no comparable Inception-like network was previously trained on VSO.
In addition, since the output space for languages like English and Spanish is large, in the auxiliary heads of Inception, we widened the second-to-last fully-connected layers which were originally 1,024 to 4,352 and 4,096, respectively, since they would otherwise become a representational bottleneck.
For English and Spanish, we found that using the second auxiliary head as the final output actually yielded higher performance due to this layer widening and so we truncate the network after this second auxiliary head for the final network.
For all other languages, the auxiliary heads were truncated and the usual penultimate layers were preserved in the final network.
We used SGD with a sigmoid decay learning rate decay with a base learning rate of 0.001, a batch size of 64 and a decay factor of 0.1 and unlike before use the true training data mean image during mean subtraction.
\section{EXPERIMENTS \& DISCUSSION}
All new experiments used a single NVIDIA GeForce GTX Titan X GPU and were implemented with Caffe \cite{jia_2014}.
\begin{table}[t]
\centering
\begin{tabular}{lccccccc}
\textbf{Language} & \textbf{\#ANPs} & \textbf{\#params} & \textbf{\#train} & \textbf{\#test} & \textbf{time} & \textbf{top-1} & \textbf{top-5} \\ \hline
English & 4,342 & 74.66 & 3,236,728 & 807,447 & 95 & 10.13\% & 21.06\% \\
Spanish & 2,382 & 66.63 & 1,085,678 & 270,400 & 45 & 12.35\% & 25.36\% \\
Italian & 1,561 & 63.26 & 602,424 & 149,901 & 30 & 17.01\% & 30.93\% \\
French & 1,115 & 61.44 & 462,522 & 115,112 & 26 & 17.66\% & 35.46\% \\
German & 275 & 57.99 & 108,744 & 27,048 & 12 & 30.11\% & 52.78\% \\
Chinese & 243 & 57.86 & 102,740 & 25,575 & 15 & 27.07\% & 45.06\%
\end{tabular}
\caption{Classification accuracies of the ANP detector banks (\%) from the original MVSO release\cite{jou_2015}, where images were collected from a hybrid pool of tag-restricted and free text searches. This table essentially mirrors Table 4 in the original publication in ACM Multimedia 2015\cite{jou_2015} except now to two-decimal precision and with estimated walltimes (hours) instead of CPU/GPU process timing and number of network parameters (millions).}
\label{tab:old_classifiers}
\end{table}
\subsection{Hybrid-pool Multilingual ANP Detectors}
\label{ssec:hybridpool}
In the original MVSO image mining \cite{jou_2015}, when we downloaded data, we first queried the Flickr API using our multilingual ANPs.
Generally, there are two types of ways to search for photos on Flickr: asking the Flickr API to search for the query in the (1) tags of a photo, or (2) anywhere in the photo text data, including photo title, description and tags.
We call these ``tag search'' and ``free-text search,'' respectively.
Tag search yields less but more precise results, and free-text search will give more but noisier data, i.e.~leading to weaker supervision.
The intuition is that if a user uses an ANP as a tag, then it is more likely to describe what is actually visually present in the image, while an ANP that occurs in, say the description, may be relevant, but not visually present.
In our data collection, we used an upper bound of 1,000 images per ANP where we first queried via tag search and if the upper bound had not been reached, we then pulled results from free-text search on Flickr.
As a result, the original MVSO image dataset constitutes a \emph{hybrid-pool} of Flickr images with respect to how the querying was performed.
The six ANP detector banks presented in the original release \cite{jou_2015} were based on this very hybrid-pool of images.
In Table \ref{tab:old_classifiers}, we show the top-$k$ accuracies of the original \cite{jou_2015} CaffeNet-structured \cite{jia_2014}, DeepSentiBank-fine-tuned \cite{chen_2014_dsb} detector banks using this hybrid-pool of images.
Likewise, in Table \ref{tab:new_classifiers}, we show the top-$k$ performances of the new Inception-styled architectures for the same six languages on the same hybrid-pool of images.
Top-$k$ accuracy refers to the percentage of classifications for which the true class is in the top $k$ predicted ranks.
\begin{table}[h]
\centering
\begin{tabular}{lccccccc}
\textbf{Language} & \textbf{\#ANPs} & \textbf{\#params} & \textbf{\#train} & \textbf{\#test} & \textbf{time} & \textbf{top-1} & \textbf{top-5} \\ \hline
English & 4,342 & 30.50 & 3,236,728 & 807,447 & 370 & 13.64\% & 26.63\% \\
Spanish & 2,382 & 20.84 & 1,085,678 & 270,400 & 106 & 13.86\% & 27.98\% \\
Italian & 1,561 & \,\,\,7.57 & 602,424 & 149,901 & 57 & 17.46\% & 32.35\% \\
French & 1,115 & \,\,\,7.12 & 462,522 & 115,112 & 36 & 16.76\% & 34.27\% \\
German & 275 & \,\,\,6.26 & 108,744 & 27,048 & 11 & 31.08\% & 54.10\% \\
Chinese & 243 & \,\,\,6.22 & 102,740 & 25,575 & 10 & 25.96\% & 47.11\%
\end{tabular}
\caption{Classification accuracies of the new ANP detector banks (\%) using an Inception-based architecture\cite{szegedy_2015} with training walltimes (hours) and number of parameters (millions). The classes, training and testing sets all match those from Table \ref{tab:old_classifiers}, i.e.~those used in the original MVSO release\cite{jou_2015}.}
\label{tab:new_classifiers}
\end{table}
Compared to the previous detector banks in Table \ref{tab:old_classifiers} which we fine-tuned from an affectively biased set of network weights \cite{chen_2014_dsb}, training takes noticeably longer in most cases since we now fine-tune from a model trained on ILSVRC.
Nonetheless, we observe that we are able to get improved ANP classification rates on most languages with Inception \cite{szegedy_2015}, particularly when there is more data for the networks to take advantage of; for example, with English ANP detection, we get a 35.05\% relative improvement at top-1.
Overall, except for French, we see a consistent improvement at top-5, indicating that more semantically relevant ANPs are being surfaced into the top ranks by Inception compared to the CaffeNet model.
While we do not see this same consistency at top-1, given that these detectors currently treat all ANPs as if they came from a flat taxonomy, many of the ANPs may in fact be semantically close to each other as indicated by the hierarchical grouping studied in the original ontology construction process \cite{jou_2015}.
As a result, from a purely empirical standpoint, a top-1 accuracy metric may treat a `pretty flower' prediction for a `beautiful flower'-labeled image to be a misclassification.
A top-5 metric in these ontology-structure-agnostic settings then may be a more reasonable indicator of detector bank performance than a top-1 performance metric.
\subsection{Tag-pool Multilingual ANP Detectors}
\label{ssec:tagpool}
Since the image corpus in the hybrid-pool setting, where images come from both tag and free-text search on Flickr, can be noisy, we perform a separate set of experiments where we train our ANP detector banks just from the subset of images that come from tag-restricted queries on Flickr.
We call this the \emph{tag-pool} of MVSO images.
Though it is a subset, each ANP still has a maximum of 1,000 images to learn from and we split the dataset as before 80/20\% per ANP.
However, since we still enforce the requirement that there are at least 125 images per ANP, some languages experienced significant decrease in ANP coverage.
In Figure \ref{fig:mvso_tag_anp_imgs}, we show the change in ANP and image coverage when going to a tag-restricted subset in MVSO.
We observe that English experiences a small 9.92\% loss in ANPs and a 63.04\% loss in image count when restricting to tag-only queries while all other languages experience about 80\% coverage loss for both ANPs and images.
Though the ANPs in MVSO are useful given their pervasive occurrence in social media, this indicates that there are a fair number of images in hybrid-pool have a considerable amount of weak supervision.
At the same time, it is worth noting that it is wrong to assume all images from the free-text search are not useful for visual recognition of adjective-noun pairs.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[height=1.7in]{tag_anps.png}
\caption{}
\label{fig:mvso_tag_anps}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[height=1.7in]{tag_images.png}
\caption{}
\label{fig:mvso_tag_images}
\end{subfigure}
\caption{Hybrid \texttt{vs.}~Tag-pool MVSO Coverage. In \textbf{(a)}, we show the coverage of ANPs and in \textbf{(b)}, images, across the six major languages in MVSO\cite{jou_2015}. English suffers the least from the tag restriction while all other language experience a large drop in ANP and image coverage.}
\label{fig:mvso_tag_anp_imgs}
\end{figure}
In Table \ref{tab:tag_classifiers} below, we show the detector bank performances for the six major languages using the tag-restricted subset of images with CaffeNet-styled \cite{jia_2014} architectures.
To train these tag-pool ANP detectors, we follow training scheme from the hybrid-pool ANP detectors in the original MVSO release \cite{jou_2015} for 1-to-1 comparison, pre-training with DeepSentiBank \cite{chen_2014_dsb} using Caffe \cite{jia_2014} with the same optimization strategies, with the only slight difference being that we use a different GPU, i.e.~980 versus Titan X, and use the true training mean image instead of the ILSVRC mean.
Again, note that the number of ANP classes each detector learns over has reduced compared to those seen earlier in Table \ref{tab:old_classifiers} due to the tag-based pool restriction.
\begin{table}[h]
\centering
\begin{tabular}{lccccccc}
\textbf{Language} & \textbf{\#ANPs} & \textbf{\#params} & \textbf{\#train} & \textbf{\#test} & \textbf{time} & \textbf{top-1} & \textbf{top-5} \\ \hline
English & 3,911 & 72.89 & 2,294,411 & 255,192 & 85 & 19.00\% & 33.81\% \\
Spanish & 502 & 58.92 & 164,119 & 18,276 & 12 & 29.07\% & 52.86\% \\
Italian & 402 & 58.52 & 151,088 & 16,807 & 9 & 33.69\% & 55.76\% \\
French & 222 & 57.78 & 80,790 & 8,986 & 6 & 34.70\% & 63.32\% \\
German & 90 & 57.24 & 32,195 & 3,578 & 2.5 & 47.04\% & 74.62\% \\
Chinese & 60 & 57.11 & 19,044 & 2,122 & 1.5 & 45.05\% & 71.35\%
\end{tabular}
\caption{Classification performances of the CaffeNet-based\cite{jia_2014} ANP detectors (\%) using images retrieved by restricting Flickr API queries to tag-based search. Note that walltimes (hours) are for running a fixed 80 epochs.}
\label{tab:tag_classifiers}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{lc|ccc||ccc}
\textbf{Language} & \textbf{\#ANPs} & \textbf{\#test} & \textbf{top-1} & \textbf{top-5} & \textbf{\#test} & \textbf{top-1} & \textbf{top-5} \\ \hline
English & 3,911 & 756,243 & 10.47\% & 21.67\% & 51,062 & 12.04\% & 24.58\% \\
Spanish & 502 & 91,002 & 14.86\% & 30.48\% & 3,740 & 17.91\% & 34.39\% \\
Italian & 402 & 64,668 & 22.51\% & 39.44\% & 3,386 & 24.25\% & 42.85\% \\
French & 222 & 38,107 & 21.78\% & 43.10\% & 1,835 & 25.29\% & 46.54\% \\
German & 90 & 13,249 & 36.18\% & 58.99\% & 769 & 39.14\% & 62.68\% \\
Chinese & 60 & 9,093 & 29.47\% & 48.78\% & 412 & 34.95\% & 51.94\%
\end{tabular}
\\
\hspace{1.53in}(a)\hspace{1.6in}(b)
\caption{Classification performances of the original CaffeNet-based \cite{jia_2014} ANP detectors (\%) from the ACM Multimedia 2015 release \cite{jou_2015}, i.e.~same detectors as in Table \ref{tab:old_classifiers}, but now evaluated only on the test images \textbf{(a)} of the tag-pool in Table \ref{tab:tag_classifiers}, and \textbf{(b)} intersection of both the hybrid-pool and tag-pool test sets.}
\label{tab:hybrid_on_tag_eval}
\end{table}
At first glance, the top-1 and top-5 accuracy rates seem much higher than those we observed with a similar architecture in Table \ref{tab:old_classifiers}.
However, it is still difficult to determine if restricting to the tag-pool actually contributed to more reliable labels and thus better performing ANP detectors, especially since the number of output classes has changed.
As a result, we evaluated the hybrid-pool ANP detectors from \S\ref{ssec:hybridpool} on the test set of the tag-restricted pool of images as well as the the intersection of the test sets.
Since random shuffling was done during train/test splitting for both hybrid- and tag-pool image datasets, the latter ensures that no images in the test set (\#test) were used in either model's training.
The results from this evaluation are shown in Table \ref{tab:hybrid_on_tag_eval}.
We observe that in both cases, the classification performances are both lower than those in Table \ref{tab:tag_classifiers} by about 10\% absolute on each language, indicating that the tag-pool ANP detectors do indeed achieve a comparably higher accuracy on the same set of ANPs.
In addition, for an even closer comparison, we tested the trained tag-pool ANP detectors on the intersection of the hybrid-pool and test-pool test sets from Tables \ref{tab:old_classifiers} and \ref{tab:tag_classifiers}, respectively, as shown in Table \ref{tab:classifiers4}.
The results here are thus most comparable to those found in Table \ref{tab:hybrid_on_tag_eval}(b).
Although there are significantly less test images than with Table \ref{tab:old_classifiers}, \ref{tab:tag_classifiers}, or \ref{tab:hybrid_on_tag_eval}(a), we observe that the top-1 classification rates for the tag-pool English ANP detector bank is a relative 59.21\% improvement over the hybrid-pool English ANP detector bank on a comparable dataset.
ANP detection in all other language likewise see an improvement in the evaluation using the tag-pool ANP detectors.
From the evaluations in Tables \ref{tab:tag_classifiers}-\ref{tab:classifiers4}, we believe we can conclude that training ANP detector banks using the tag-pool have led to significantly better ANP detectors.
The trade-off here, however, is that we have dramatically less images to train on and also a significantly reduced coverage of ANPs per language, except in the case of English.
Nonetheless, all this is serves as concrete evidence that tag-based labels are more reliable than those mined from the metadata overall since users are more likely to use tags to describe content that is visually present in the image.
While this does not mean that all labeled images gathered from free-text metadata context are wrong, it does reduce the trust that should be placed on their ability to represent elements visually present in the associated image.
We believe that in the future development of ANP detectors factoring such label reliability will allow for a greater coverage of ANPs without sacrificing in the area of data sparsity or the classification performance of ANP detector banks.
\begin{table}[h]
\centering
\begin{tabular}{lcccc}
\textbf{Language} & \textbf{\#ANPs} & \textbf{\#test} & \textbf{top-1} & \textbf{top-5} \\ \hline
English & 3,911 & 51,062 & 19.17\% & 34.01\% \\
Spanish & 502 & 3,740 & 28.85\% & 52.75\% \\
Italian & 402 & 3,386 & 32.75\% & 54.90\% \\
French & 222 & 1,835 & 35.80\% & 64.20\% \\
German & 90 & 769 & 49.41\% & 74.51\% \\
Chinese & 60 & 412 & 41.50\% & 67.23\%
\end{tabular}
\caption{Classification performances of the CaffeNet-based\cite{jia_2014} ANP detectors (\%) using tag-restricted images but now evaluated only on the set of test images intersecting the hybrid pool of Table \ref{tab:old_classifiers}, i.e.~original MVSO data release\cite{jou_2015}, and tag-restricted pool of Table \ref{tab:tag_classifiers}. These tag-restricted performances are thus most comparable to the results reported in Table \ref{tab:hybrid_on_tag_eval}(b).}
\label{tab:classifiers4}
\end{table}
\section{CONCLUSION}
\label{sec:conclusion}
In this technical report, we have presented new ANP detector banks for MVSO for six major languages using a more modern network architecture than in the original MVSO release.
These detectors have a reduced model memory footprint compared to their predecessors and achieve very competitive or even significantly better ANP detection performance, e.g.~a 35.05\% improvement for English ANP detection.
In addition, we presented ANP detectors that use a restricted subset of images and ANPs that originate only from tag-based Flickr API retrieval results, which resulted in much higher ANP classification performances due to improved label quality.
We also publicly release the new Inception-based ANP detector banks, tag-pool CaffeNet ANP detector banks, and the list of image URLs corresponding to the tag-restricted MVSO image subset used in this report.
In the future, we would like to investigate additional fine-tuning strategies for training ANP detectors.
For example, in the original release \cite{jou_2015}, we noted that fine-tuning from DeepSentiBank \cite{chen_2014_dsb} was done in the hopes that we initialized the models from an affectively biased source; however, it is not empirically clear whether such a biasing actually helps or hurts detector performance.
In addition, we would like to investigate and benchmark our ANP detectors against other popular network architectures.
Also, we plan to investigate how we might integrate the semantic hierarchy in MVSO to train ANP detectors that exploit the tree-structure.
\acknowledgments
We would like to thank Miriam Redi, Mercan Topkara, Nikolaos Pappas and Tao Chen from Multilingual Visual Sentiment Ontology (MVSO) team for their continued support and insightful discussions.
Also, we thank Margaret Yuying Qian for performing some of the early tag-restricted subset partitioning and statistics.
|
\section{Introduction}
High resolution imaging plays a crucial role in atomic physics experiments, from trapping and imaging single atoms \cite{schlosser01,sherson10,piotrowicz13,nogrette14} to creating complex arbitrary optical potentials for quantum gas studies \cite{ramanathan11,zimmermann11,gaunt12}. These experiments are performed under vacuum requiring imaging through a glass window which introduces significant spherical aberrations and substantially reducing the resolution of an uncompensated optical system. One approach is to use lenses mounted inside the vacuum system such that collimated light passes through the viewport \cite{sortais07,tey08} however adjustment of the optical system is not possible once under vacuum and can additionally limit the lowest achievable pressure.
\begin{figure}[t]
\includegraphics{fig1.eps}
\caption{(Color online). ATOM objective, optimized for high NA fluorescence detection for single atom readout at 852~nm.\label{fig1}}
\end{figure}
An alternative is to design a compound objective lens with the vacuum window included as the final element as first demonstrated by W. Alt \cite{alt02} which requires four singlet lenses to provide diffraction limited performance. A number of similar designs have since been published, either requiring one or more custom optics to be fabricated \cite{bucker09,zimmermann11} or using catalog optics but a short focal length ($\lesssim35$~mm) optimal for use with a small glass vacuum cell \cite{bennie13}. This precludes use with stainless steel vacuum systems that typically have much larger working distances.
In this paper we present a pair of optimized objective lenses with long working distances compatible for use with a standard stainless steel chamber. These are designed to be diffraction limited when used with standard conflat viewports and are comprized of standard catalog optics available with a range of anti-reflection coatings to provide a simple and cost-effective solution. Whilst designed for use at 802 and 852~nm for trapping and imaging neutral Cs atoms, the lenses remain diffraction limited at 780~nm for use with Rb without modification. For other elements, the lenses can be adapted to shorter or longer wavelengths by simply re-optimising the relative spacings.
\begin{figure*}[ht!]
\includegraphics{fig2.eps}
\caption{(Color online). TRAP objective, optimized for diffraction limited performance at 802~nm.\label{fig2}}
\end{figure*}
\vspace{-0.1cm}
\section{Design}
The objectives presented in this paper are designed to provide infinity-corrected and diffraction limited performance whilst maximising the available numerical aperture given the minimum working distances imposed by our vacuum chamber geometry. For simplicity all elements are constrained to be 2~inch diameter BK7 singlets. The objectives are iteratively designed using ray tracing software (Zemax) to numerically optimise the lens spacing and curvatures to minimise the RMS wavefront error. One by one each element is then replaced by the closest matching catalog singlet and the remaining elements curvature and separation re-optimized until convergence is obtained. Interestingly, the objective designs converge to the same lens shapes as used originally by W. Alt \cite{alt02} suggesting this provides the optimal aberration compensation when the final element is a flat glass window. The objectives are then assembled in a standard 2 inch long lens tube (Thorlabs SM2L02) using spacing rings machined from delrin to avoid damaging the lenses.
The first objective (`ATOM') is designed to provide high efficiency fluorescence collection for imaging a single Cs atom at 852~nm through a standard CF4.5" fused silica viewport (thickness 6.4~mm) with a minimal working distance of 65~mm, giving a maximum achievable numerical aperture 0.3. Figure~\ref{fig1} shows the final lens configuration along with spacing ring parameters, with the complete prescription given in Table~\ref{ATOM}. This lens has an effective focal length of 67.4~mm and $\mathrm{NA}=0.291$ offering a collection efficiency of 2.1~\% and a diffraction limited modulation transfer function (MTF) on axis with an Airy radius of $r_\mathrm{A}=1.8~\mu$m.
\begin{table}[t]
\begin{ruledtabular}
\begin{tabular}{c c c c}
Surface & Curvature (mm) & Thickness (mm) & Material \\
\hline
1 & $\infty$ & 4 & BK7\\
2 & 77.19 & 19.235 & Air \\
3 & 179.14 & 6.62 & BK7\\
4 & -179.14 & 0.2 & Air \\
5 & 77.26 & 7.29 & BK7\\
6 & $\infty$ & 0.2 & Air\\
7 & 47.87 & 7.29 & BK7\\
8 & 119.32 & 14.66 & Air\\
9 & $\infty$ & 6.4 & Silica\\
10 & $\infty$ & 49.2 & Vacuum
\end{tabular}
\end{ruledtabular}
\caption{ATOM Objective prescription - lenses from left to right are Thorlabs LC1611, LB1607, LA1417 and LE1418. The effective focal length at 852~nm is 67.4 mm.\label{ATOM}}
\end{table}
\begin{table}[t]
\begin{ruledtabular}
\begin{tabular}{c c c c}
Surface & Curvature (mm) & Thickness (mm) & Material \\
\hline
1 & $\infty$ & 2.5 & BK7\\
2 & 129.2 & 15.6 & Air\\
3 & 205.0 & 6.2 & BK7\\
4 & -205.0 & 0.5 & Air\\
5 & 154.5 & 5.1 & BK7\\
6 & $\infty$ & 0.5 & Air\\
7 & 100.1 & 5.1&BK7\\
8 & 279.1 & 20.8 & Air\\
9 & $\infty$ & 3.3 & Silica\\
10 & $\infty$ & 98.0 & Vacuum
\end{tabular}
\end{ruledtabular}
\label{TRAP}
\caption{TRAP Objective prescription - lenses from left to right are Newport KPC067, Thorlabs LB1199, LA1256 and LE1985. The effective focal length at 802~nm is 120 mm.}
\end{table}
The second objective (`TRAP') is designed for use with light at 802~nm to create an optical trapping potential. Here the lens is used in conjuction with a CF2.75" fused silica viewport (thickness 3.3~mm) with a working distance $\ge110$~mm and numerical aperture limited by the internal diameter of the viewport to $\mathrm{NA}=0.17$. The optimized configuration is shown in Fig.~\ref{fig2}, with complete prescription in Table~\ref{TRAP}. This design has an effective focal length of 119.5~mm and $\mathrm{NA}=0.172$, resulting in a an Airy radius of $2.85~\mu$m.
In addition to considering the on-axis performance of the objectives, it is also possible to extract the diffraction-limited field of view (FOV). This is defined as the distance from the axis in the image space ($\Delta$) for which the Strehl ratio $S\ge0.8$ \cite{gross07}, and is plotted in Fig.~\ref{fig3} giving FOV diameters of $0.87$ and $0.62$~mm respectively.
Since the objectives are comprized of a single glass type they suffer from chromatic aberration. However, for small departures from the design wavelength no degradation in performance is observed, only a slight change in the effective focal length. For example, the ATOM objective at 780~nm has an effective focal length of 67.2~mm whilst at 852~nm the TRAP objective changes to 119.75~mm with a slight increase in $\mathrm{NA}$ to 0.175. Larger wavelength changes require re-optimisation of the lens separation, and this has been successfully applied to the TRAP objective to obtain diffraction-limited performance at 410~nm for imaging Ho atoms.
\begin{figure}[b]
\includegraphics{fig3.eps}
\caption{(Color online). Calculated Strehl ratio $S$ for objectives. Diffraction limit defined as $S\ge0.8$.}
\label{fig3}
\end{figure}
\newpage
\section{Performance}
\begin{figure}[b]
\includegraphics{fig4.eps}
\caption{(Color online). Calibration of imaging magnification and resolution for the ATOM objective using (a) 1951 USAF test card and (b) a 5~$\mu$m pinhole. Camera pixels are $8\times8~\mu$m.}
\label{fig4}
\end{figure}
To characterise the performance of the objective lenses, they are tested \textit{ex vacuo} using catalog AR coated fused-silica viewports. A relay telescope comprised of the objective lens in conjunction with an $f=1~$m singlet is used to image a USAF 1951 resolution target onto a CCD camera (Andor Luca R) using light at 852~nm. From these images the magnification and resolution are extracted by fitting the observed pattern to a square wave convolved with a gaussian point-spread function (PSF) characterised by standard deviation $\sigma$. Figure~\ref{fig4}(a) shows the test card image for the ATOM objective giving magnification $M=-14.9\pm0.1$ and $\sigma=0.8\pm0.1~\mu$m respectively, in excellent agreement with simulation ($M=-15$, $\sigma=0.62~\mu$m). For the TRAP objective $M=-8.4$ and $\sigma=1.8\pm0.2~\mu$m ($M=-8.36$, $\sigma=1.1~\mu$m). This large discrepancy in the PSF arises due to the imperfections in the optical quality of the viewport, as the full clear aperture of the window is used with substantial reduction in the optical flatness around the edges of the glass to metal seal. Improved results are possible either by post-selecting viewports or using home-made viewports from optical flats \cite{weatherill09,gupta12}. This does not affect the ATOM objective as this uses a larger viewport. Additionally, the ATOM objective is used to image a 5~$\mu$m pinhole which can again be treated as a top-hat profile convolved with the point spread function yielding $\sigma=0.75\pm0.05~\mu$m as shown in \ref{fig4}(b).
The objectives have subsequently been used for their designed application. Figure~\ref{fig5}(a) shows an image of the 802~nm trap configuration created using the TRAP objective to obtain with diffraction limited waists of 3~$\mu$m and 6~$\mu$m separation created using the Gaussian beam array (GBA) technique \cite{piotrowicz13}. This is achieved by combining a pair of frequency-shifted beams with orthogonal polarisations on a polarising beam splitter, then rotating the polarisation by 45 degrees and using a calcite to displace the beams to create four spots of equal intensity where each focus is a different frequency and polarisation from its neighbour to avoid interference effects at the centre of the beams. This creates a blue-detuned trapping potential for neutral Cs atoms which are localized within the saddle point at the centre of trap, with sufficiently small volume to enable single atom loading \cite{schlosser02} and trapping of both ground and Rydberg states \cite{zhang11}, ideal for application in neutral atom quantum computing \cite{saffman10}. For a total power of 350~mW, this gives a trap depth of $U_0=750~\mu$K and radial (axial) trap frequencies of $\omega_{r,z}/2\pi=25,1.5$~kHz.
Figure~\ref{fig5}(b) shows single atom readout measurements performed using the ATOM objective to provide high collection efficiency (2.2\%). The atom is trapped in a $2.4~\mu$m red-detuned dipole trap at 935~nm focused through the objective lens using a convergent input beam to compensate the 205~$\mu$m chromatic shift and overlap the 935~nm waist with the 852~nm focal plane. The atom is loaded from a standard magneto-optical trap using light assisted collisions to ensure either 0 or 1 atom in the trap \cite{schlosser02}. Readout is performed using light at 852~nm detuned by $\Delta/\Gamma=-4.4$ from the $6S_{1/2}~F=4\rightarrow6P_{3/2}~F'=5$ cooling transition with a total intensity of 10~mW/cm$^2$, corresponding to a scattering rate of 750~kHz. To overcome the differential AC Starkshift in the trap, the cooling light is chopped out of phase with the trap light at 1~MHz with a 50~\% duty cycle. The collected fluorescence is then collimated by the objective and re-imaged into a 1550~nm single mode fiber connected to a single photon counting module (Perkin Elmer SPCM-ARQH-13) with quantum efficiency of 45~\% at 852~nm. Transmission through the optical system (including transmission losses from dichroic optics and fiber coupling) is calibrated at 30~\%, giving an expected single atom signal of $\sim1$~counts/ms in excellent agreement with the observed histogram data which shows clear discrimination between 0 and 1 atom in the trap with an average signal of 100 counts above background and demonstrating single atom resolution with the ATOM objective.
\begin{figure}[t]
\includegraphics{fig5.eps}
\caption{(Color online). (a) Single-atom trap configuration using TRAP objective to create a single site GBA trap with four Gaussian spots with 3~$\mu$m waist separated by 6~$\mu$m. (b) Single atom fluorescence readout using the ATOM objective to demonstrate atom number discrimination, solid line shows fit to binomial distribution.}
\label{fig5}
\end{figure}
\section{Conclusion}
We have presented a pair of objective lenses offering diffraction-limited performance and long-working distance for use with standard optical viewports suitable for single atom trapping and readout. The objectives are highly versatile, using standard catalog optics to provide a simple and economical approach to obtaining high resolution objective lenses. The design can be adapted to work over a range of wavelengths and can be used at multiple wavelengths simultaneously if the convergence of the secondary wavelength is adjusted to compensate the chromatic shifts as demonstrated above. Zemax lens files for the two objectives are included as supplemental material\cite{supp}.
\begin{acknowledgments}
This work was supported by funding from the NSF award PHY-1212448 and the University of Wisconsin Graduate School.
\end{acknowledgments}
|
\section{Introduction}
There have recently been many proposals for ways to induce energy propagation along a metallic nanoparticle chain. Such propagation occurs in the form of traveling waves, which are driven by electric dipole-dipole coupling between localized plasmon modes on neighboring nanoparticles\cite{Maier2003,Brongersma2000}. The dispersion relations of such waves have been calculated by several authors within the quasistatic approximation~\cite{Maier2003,Park2004,Pike2013,Pike2016}. This approximation, which assumes that the dipole-dipole coupling is accurately described by electrostatics, is believed accurate when the spacing between particles is small compared to the wavelength of light at a typical plasmon frequency. The resulting propagating plasmonic waves have been observed in recent experiments~\cite{Hossain2012,Li2011}. These calculations have also been extended to included dynamical effects beyond the quasistatic approximation~\cite{Weber2004}.
In the present paper, we investigate the analogous problem of waves propagating along chains of {\it magnetic} particles coupled via their magnetic dipole moments. We consider particles of the well known magnetic insulator Yttrium Iron Garnet (YIG). The magnetic dipole-dipole interaction between localized magnetic excitations on these particles leads to propagating magnetic waves, or magnons. These waves can carry both magnetization and energy and thus may produce wireless power transmission~\cite{Karenowska2015,Damon1965,Kreisel2009}. Specifically, we calculate the dispersion relations of propagating magnon waves along chains of spherical or cylindrical YIG particles within the quasistatic approximation. These calculations are analogous to the extensive earlier work, mentioned above, on wave propagation along chains of small metal particles, also within the quasistatic approximation. In this latter work, the waves are able to propagate because of the electric dipole-dipole interactions between the plasmonic excitations in the individual nanoparticles. By contrast, the relevant interactions for the YIG particles are {\it magnetic} dipole-dipole interactions. Here, we use the magnetic properties of YIG, together with equations of motion describing the interactions of localized magnetic excitations on the particles with the excitations on neighboring particles, to calculate the dispersion relations of propagating magnon waves along chains of spherical or cylindrical YIG particles.
Recently, other types of magnon waves have been studied theoretically at the interface between a two-dimensional (2D) magneto-optical photonic crystal and a regular 2D gapped photonic crystal~\cite{Wang2008a}. These studies show that one-way edge modes can propagate at the interface and within the band gap of the bulk modes. They also show that these edge modes are confined to within a few lattice constants of the boundary between the two crystals, and are thus basically one-dimensional. Work by Khanikaev {\it et al.}~\cite{Khanikaev2007} in metallic systems containing holes filled with the magneto-optical material Bi:YIG suggest that the Faraday rotation angle within the material is enhanced because of plasmonic excitations at the surface of the magneto-optic material.
The goal of transmitting power wirelessly is, of course, very old. In the early $20^{th}$ century, Tesla~\cite{Tesla1914,Tesla1891} attempted to transmit power wirelessly even before the dawn of the electrical power grid. His proposal, like many others~\cite{Li2011a,Javadi2013}, requires the generation of large electric fields in order for a reasonable amount of power to be transmitted. In order to avoid the need for large electric fields, recent efforts have turned towards magnetic systems. In particular, it has been shown~\cite{Kur2007} that one can use a coil of wire, in which a current is driven at a fixed frequency, to transmit power, via magnetic induction, to another coil of wire a few meters away, with approximately 40\% efficiency. Likewise, one can transmit power using split-ring resonators~\cite{Shiffler2013} or many other patterned devices. In the present work, we describe an approach which requires only single crystals of YIG~\cite{Kimura1977,Linares1965}, patterned as spheres or rods and arranged in a periodic chain, in order to produce wireless power transmission.
We turn now to the body of the paper. In Section~\ref{model}, we describe the geometry of the system and derive equations for the coupled magnetic dipole moments within the quasistatic approximation, using an approach analogous to some earlier work~\cite{Brongersma2000,Maier2003,Pike2013}. In Section~\ref{three}, we write down the propagating wave solutions for chains of spherical or cylindrical YIG particles, and illustrate the solutions with some numerical results. Finally, in Section~\ref{four}, we discuss the results of our calculations and provide some concluding remarks. A short Appendix provides further algebraic details about the calculation of the dispersion relations for chains of YIG spheres.
\section{Model}\label{model}
We will consider periodic chains of spherical or cylindrical YIG particles. The spacing between the centers is denoted $d$, while the radius of the spheres or cylinders is $a$. For both spheres and cylinders, the chains run along the $z$ axis, while for the cylinders, the cylinder is parallel to the $x$ axis. In all cases, we assume that $d = 3 a$. A cartoon image of the assumed geometries is shown in Fig.~\ref{fig:one}.
We will be interested in periodic waves of magnetic dipole moments traveling along the chain. As shown elsewhere in the case of coupled electric dipole moments~\cite{Pike2013,Maier2001,Maier2003,Koenderink2006,Brongersma2000,Maier2003}, these magnetic dipole waves can be obtained from coupled linear equations involving the dipole moments on each particle within the quasistatic approximation. For the electric dipole case, the dispersion relations are controlled by electric dipole-dipole coupling between plasmons centered on neighboring metallic particles. In the case of YIG particles, the dispersion relations describe propagating waves of magnetic dipoles and are generated by magnetic dipole-dipole coupling between magnetic plasmon excitations centered on neighboring YIG particles. The solutions to the coupled set of equations lead to the dispersion relations for the three branches of propagating magnon waves.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Theory_setup8.png}
\caption{(a) Considered geometry for a chain of YIG spheres of radius $a$ and separation $d$ periodically arranged along the $z$ axis. (b) A chain of very long YIG rods of radius $a$, arranged periodically along the $z$ axis with separation $d$. In both cases, we assume that the static magnetic field ${\bf H}_0$ and the saturation magnetization ${\bf M}_s$ are parallel to one another and lie along either the $x$ or $z$ axis.}
\label{fig:one}
\end{figure}
To obtain the coupled linear equations it is convenient to proceed using equations analogous to Eq.\ (9) of Ref.~\citenum{Pike2013}. We assume that the chain of particles (either spheres or cylinders) is parallel to the $z$ axis as shown in Fig.~\ref{fig:one}. Writing out the coupled equations explicitly, and assuming, first, that the magnetic particles are spheres of radius $a$, we obtain
\begin{equation}\label{eq:mag_moment3D}
{\bf m}_n = -\frac{4\pi a^3}{3}\hat{t}\sum_{n^\prime \neq n}\hat{\cal G}^{s}({\bf x}_n - {\bf x}_{n^\prime})\cdot {\bf m}_{n^\prime},
\end{equation}
where ${\bf m}_n$ is the magnetic dipole on the n$^{th}$ sphere. If the particles are cylinders of radius $a$, we find
\begin{equation}
\label{eq:mag_moment2D}
{\bf m}_n = -\pi a^2\hat{t}\sum_{n^\prime \neq n}\hat{\cal G}^{c}({\bf x}_n - {\bf x}_{n^\prime})\cdot {\bf m}_{n^\prime},
\end{equation}
where the quantity ${\bf m}_n$ is now the magnetic moment per unit cylinder length of the n$^{th}$ cylinder.
In Eqs.~\eqref{eq:mag_moment3D} and~\eqref{eq:mag_moment2D} we find that $\hat{\cal G}^{s}$ or $\hat{\cal G}^c$ are related to the Greens functions of a point charge in two or three dimensions. Specifically, ${\cal G}_{ij}^s({\bf x}-{\bf x}^\prime) = \partial_i^\prime\partial_j G^s({\bf x}-{\bf x}^\prime)$, where $G^s({\bf x} - {\bf x}^\prime)= -1/[4\pi|{\bf x}-{\bf x}^\prime|]$, while ${\cal G}_{ij}^c({\bf x}-{\bf x}^\prime)= \partial_i^\prime\partial_jG^c({\bf x}-{\bf x}^\prime)$, where $G^c(({\bf x}-{\bf x}^\prime) = 1/(2\pi)\ln(|{\bf x}-{\bf x}^\prime|)$. From these expressions, we find the nonzero components of the matrices $\hat{\cal G}^s$ and $\hat{\cal G}^c$ to be
\begin{equation}\label{eq:G_sphere_z}
{\cal G}^s_{xx}({\bf x}_n-{\bf x}_{n^\prime}) = {\cal G}^s_{yy}({\bf x}_n - {\bf x}_{n^\prime}) =
-\frac{1}{4\pi}\frac{1}{|z_n-z_{n^\prime}|^3}
\end{equation}
and
\begin{equation}\label{eq:G_sphere_z2}
{\cal G}^s_{zz}({\bf x}_n-{\bf x}_{n^\prime}) = \frac{1}{2\pi}\frac{1}{|z_n-z_{n^\prime}|^3}
\end{equation}
for the case of a spheres in vacuum, and
\begin{equation}\label{eq:G_rods_z}
{\cal G}^c_{yy}=-{\cal G}^c_{zz}=\frac{1}{2\pi}\frac{1}{|z_n-z_n^\prime|^2}.
\end{equation}
for the case of a cylinders in vacuum, with all other elements in both ${\cal G}^s$ or ${\cal G}^c$ equal to zero.
The matrix $\hat{t}$ is equal to
\begin{equation}\label{eq:t_matrix}
\hat{t} = \hat{\delta\mu}(\hat{I}-\hat{\Gamma}\hat{\delta\mu})^{-1},
\end{equation}
where
\begin{equation}
\hat{\delta\mu} = \hat{\mu} - \hat{I},
\end{equation}
$\hat{I}$ is the identity matrix (since we assume that the host has permeability equal to unity) and the permeability tensor $\hat{\mu}$, discussed below, depends on the orientation of the magnetic field.
The demagnetization matrices $\hat{\Gamma}$ for spheres and cylinders in vacuum are also well known~\cite{Osborn1945}. For spheres,
\begin{equation}\label{eq:gamma_sphere}
\hat{\Gamma} = -\frac{1}{3}\hat{I}
\end{equation}
while for cylinders parallel to $\hat{z}$, the nonzero elements are
\begin{align}\label{eq:gamma_rods}
\Gamma_{yy} &= \Gamma_{zz} =-\frac{1}{2} \nonumber \\
\Gamma_{xx} &= 0.
\end{align}
If the applied magnetic field ${\bf H}_0$ and the saturation magnetization ${\bf M}_s$ are both parallel to $\hat{z}$, the permeability tensor $\hat{\mu}$ (in Gaussian units) takes the form~\cite{Pozar1990, Moorish1965}
\begin{equation}\label{eq:permeability_tensor}
\hat{ \mu} = \begin{pmatrix} \mu_1 & i \mu_2 & 0 \\
- i \mu_2 & \mu_1 & 0\\
0 & 0 & \mu_f
\end{pmatrix}.
\end{equation}
If both ${\bf H}_0$ and ${\bf M}_s$ are parallel to $\hat{x}$, then
\begin{equation}\label{eq:permeability_tensor_2}
\hat{ \mu} = \begin{pmatrix} \mu_f & 0& 0 \\
0 & \mu_1 & i\mu_2 \\
0 & -i\mu_2 & \mu_1
\end{pmatrix}.
\end{equation}
In both cases, for YIG, the components of the permeability tensor $\hat{\mu}$ are given by
\begin{equation}\label{eq:mu1}
\mu_1 = \mu_f\left( 1+\frac{\omega_0 \omega_1}{\omega_0^2-\omega^2}\right)
\end{equation}
and
\begin{equation}\label{eq:mu2}
\mu_2 = \mu_f \frac{\omega\omega_1}{\omega_0^2 - \omega^2},
\end{equation}
with $\omega_0 = \gamma | {\bf H}_0| - i\alpha\omega$ and $\omega_1 = 4\pi\gamma M_s/\mu_f$. Here $M_s$ is the magnitude of the saturation magnetization, $\gamma$ is the gyromagnetic ratio, and $\alpha$ is the phenomenological damping coefficient~\cite{Pozar1990}. For our purposes, we will assume~\cite{Cunha2015,Yu2014} that $4 \pi M_s = 1760\ G$ and $\gamma = 2.8 \ MHz/G$ and we take $\alpha = 5.0\times10^{-5}$. With these parameters and the equations given in Ref.~\citenum{Yu2014}, we calculate $\mu_f$ to be $1.40$. We use Gaussian units throughout this article.
The $\hat{t}$-matrices for the four cases presented here are easily found by combining the previous equations. For a chain of spheres along the $z$ axis with ${\bf H}_0 \| {\bf M}_s \| \hat{z}$ the non-vanishing components of the $\hat{t}$-matrix are found to be
\begin{align}\label{eq:t_matrix_sphere_z}
t_{xx} &= t_{yy} = \frac{\mu_1 -2 +\mu_1^2-\mu_2^2}{D} \nonumber \\
t_{xy} &= - t_{yx} = \frac{ 3 \mu_2}{D} \nonumber \\
t_{zz} &= \frac{3(\mu_f -1)}{\mu_f +1}
\end{align}
where $D = \frac{1}{3}(4+4\mu_1+\mu_1^2-\mu_2^2)$.
For a chain of spheres along the $z$ with ${\bf H}_0 \| {\bf M}_s \| \hat{x}$ the non-vanishing components of the $\hat{t}$-matrix are
\begin{align}\label{eq:t_matrix_sphere_x}
t_{xx}&= \frac{3(\mu_f -1)}{\mu_f+1}\nonumber \\
t_{yy} &= t_{zz} = \frac{\mu_1 -2 +\mu_1^2-\mu_2^2}{D} \nonumber \\
t_{yz} &= - t_{zy} = \frac{ 3 \mu_2}{D }
\end{align}
where $D$ is the same in Eq.~\eqref{eq:t_matrix_sphere_z}.
For a chain of cylinders along $z$ with cylinder axes parallel to $\hat{x}$ (see Fig.~\ref{fig:one}) with ${\bf H}_0\|{\bf M}_s \| \hat{z}$ the non-vanishing components of the $\hat{t}$-matrix are
\begin{align}\label{eq:t_matrix_rod_z}
t_{xx} &=\frac{\mu_1^2-4\mu_1 -\mu_2^2 +3}{D_1} \nonumber \\
t_{yy} &= \frac{2( \mu_1-1)}{D_1} \nonumber \\
t_{zz} &= \frac{2(\mu_f-1)}{\mu_f+1} \nonumber \\
-t_{xy} &= t_{yx}= \frac{2 \mu_2}{D_1}
\end{align}
where $D_1= \mu_1 - 3$.
Finally, for a chain of cylinders along $z$ with cylinder axes parallel to $\hat{x}$ with ${\bf H}_0 \| {\bf M}_s \| \hat{x}$ the non-vanishing components of the $\hat{t}$-matrix are
\begin{align}\label{eq:t_matrix_rod_x}
t_{xx} & = \mu_f -1 \nonumber \\
t_{yy} & = \frac{-\mu_1^2 +\mu_2^2+2}{D_2} \nonumber \\
t_{zz} & = \frac{\mu^2_1-4\mu_1 -\mu_2^2 +3}{D_2} \nonumber \\
t_{yz} &= -t_{zy}= \frac{2 \mu_2}{D_2}.
\end{align}
where $D_2 = \frac{1}{2}(\mu_1^2-2 \mu_1-\mu_2^2-3)$.
Having calculated the elements of the $\hat{t}$-matrices, the $\hat{{\cal G}}$ matrices, and the demagnetization matrix $\hat{\Gamma}$, we can now calculate the dispersion relations using Eq.~\eqref{eq:mag_moment3D} or~\eqref{eq:mag_moment2D}.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Chain_par_Ho_trans8.png}
\caption{The real part of the calculated dispersion relations $\omega(k)$ of the transverse branch for a chain of YIG spheres parallel to $\hat{z}$, for the case ${\bf H}_0 \|{\bf M}_s \|\hat{z}$. We plot $\omega/(\gamma H_0)$ versus $kd$. The dispersion relation (full line) is obtained from Eq.~\eqref{eq:disp_sphere_z}, using the parameters $t_{xx}$ and $t_{xy}$ as given in Eq.~\eqref{eq:t_matrix_sphere_z}. Only one real solution occurs for this geometry, as described in the appendix. The solution is symmetric about k = 0.}
\label{fig:two}
\end{figure}
\vspace{1.0in}
\section{Propagating Wave Solutions for Spherical and Cylindrical Particles}~\label{three}
\subsection*{ Spherical Particles}
Using Eqs.~\eqref{eq:mag_moment3D},~\eqref{eq:G_sphere_z},~\eqref{eq:G_sphere_z2}, and either~\eqref{eq:t_matrix_sphere_z} or~\eqref{eq:t_matrix_sphere_x}, we can obtain dispersion relations for propagating transverse ($T$) or longitudinal ($L$) waves for the cases ${\bf H}_0 \| {\bf M}_s \| \hat{z}$ and ${\bf H}_0 \| {\bf M}_s \| \hat{x}$. In the former case, we find that the equations for the two transverse components of ${\bf m}$ are coupled, leading to solutions which are left- or right-circularly polarized waves. The $L$ branch is
found to be independent of $k$ in the quasistatic approximation.
To obtain the dispersion relations for the $T$ and $L$ branches, with ${\bf H}_0 \| {\bf M}_s \| \hat{z}$, we assume that ${\bf m}_n ={\bf m}_0 e^{ikn}$. We can then write down the matrix equation for the two coupled $T$ branches as
\begin{equation}\label{eq:18}
{\bf m}_0 = \frac{2a^3}{3d^3}\begin{pmatrix}
t_{xx} & it_{xy} \\
-it_{xy} & t_{xx}
\end{pmatrix}{\bf m}_0Cl_3(kd).
\end{equation}
The equation for the $L$ branch is given as
\begin{equation}\label{eq:19}
{\bf m}_{0z} = -\frac{4 a^3}{3 d^3} t_{zz}{\bf m}_{0z}Cl_3(kd).
\end{equation}
In both these expressions,
\begin{equation}
Cl_s(z) = \sum^\infty_{n' =1,2,...}\frac{\cos{n' z}}{n'^s}
\nonumber
\end{equation}
is known as the Clausen function~\cite{Clausen1832}. We have obtained these expressions by making a change of variables to rewrite the sum in Eq.~\eqref{eq:mag_moment3D} in terms of a single summand. Since $t_{zz}$ is independent of $\omega$ we find that the $L$ branch is independent of wave number within the quasistatic approximation.
In principle the sum in the Clausen function can be evaluated numerically, which would allow the calculation of dispersion relations including the effects of all neighbors. In practice this may be difficult, since the elements of both the $\hat{t}$-matrix and the Clausen function are multivalued functions of $\omega$ and of $kd$. Hence, we will only include only the nearest-neighbor contributions for the remainder of this paper.
In the case of only nearest-neighbor interactions, i.\ e., including only the term $n' = 1$, the equation for the $T$ branches simplifies to
\begin{equation}\label{eq:disp_sphere_z}
\left[1- \frac{2a^3}{3d^3}\begin{pmatrix}
t_{xx} & it_{xy} \\
-it_{xy} & t_{xx}
\end{pmatrix}\cos{kd}\right]{\bf m}_0 = 0.
\end{equation}
The solutions to this equation, which are found by setting the determinant of the matrix of coefficients of ${\bf m}_0$ equal to zero, give the dispersion relations for the coupled transverse modes when the chain and magnetic field both lie in the $\hat{z}$ direction. Since the $\hat{t}$-matrix depends on the frequency $\omega$ this equation represents an implicit relation between $\omega$ and $k$ for the coupled waves. In this geometry, as already mentioned, the two solutions are left- and right-circularly polarized $T$ waves. Since these $T$ waves have different dispersion relations, a linearly polarized $T$ wave will undergo a rotation as it propagates along the chain in a manner similar to the Faraday effect in a bulk homogeneous magnetic material~\cite{Fu2008}.
In Fig.~\ref{fig:two}, and all following figures, we plot the real part of the dispersion relation for the coupled waves for the various cases presented here. In Fig.~\ref{fig:two} we plot the dispersion relations for the case ${\bf H}_0 \| {\bf M}_s \| \hat{z}$ (and also the chain parallel to $\hat{z}$) as functions of $\omega/(\gamma H_0)$, taking $H_0 = 1 \ T$ and $\gamma = 2.8$ MHz/Gauss. As explained in the appendix, only a single solution exists for the geometry considered here and we have numerically verified that, upon changing the strength of the off-diagonal matrix element, we find two coupled solutions for the $T$ waves whose dispersion is dependent on the strength of the off-diagonal matrix element. Therefore, the Faraday rotation angle per unit chain length, a measure of the difference in wave vector for circularly polarized waves, is also strongly dependent on the off-diagonal element. The Faraday rotation angle is~\cite{Pike2016} $\theta(\omega) = \frac{1}{2}\left(k_1(\omega)-k_2(\omega)\right)$ where $k_1$ and $k_2$ are the wave numbers of the two circularly polarized solutions at frequency $\omega$.
Next, we consider the case where ${\bf H}_0 \| {\bf M}_s \| \hat{x}$, while the chain of magnetic spheres is again parallel to $\hat{z}$. In this case, the permeability tensor is given by Eq.~\eqref{eq:permeability_tensor_2} and the $\hat{t}$-matrix has components given by Eq.~\eqref{eq:t_matrix_sphere_x}. From the $\hat{t}$-matrix it is clear that the $T$ waves polarized in the $x$ direction are independent of $k$ within the quasistatic approximation, and are decoupled from waves polarized in the $y$ and $z$ directions.
The coupled $y$ and $z$ components once again satisfy equations of motion which are analogous to Eq.~\eqref{eq:disp_sphere_z} but with the permeability tensor now given by Eq.~\eqref{eq:permeability_tensor_2}. As in the previous case we can assume ${\bf m}_n ={\bf m}_0 e^{ikn}$ and, considering only nearest-neighbor coupling, can obtain the dispersion relations for the coupled $y$ and $z$ waves. The matrix equation for these two components is found to be
\begin{equation}
\left[1 - \frac{2a^3}{3d^3} \cos{kd}\begin{pmatrix}
t_{yy} & -2it_{yz} \\
-it_{yz} & -2t_{yy}
\end{pmatrix}\right] {\bf m}_0 = 0.
\end{equation}
Once again, we can determine the dispersion relations for the two coupled branches by setting the determinant of the matrix of coefficients of ${\bf m}_0$ equal to zero and solving for $k(\omega)$.
In Fig.~\ref{fig:three} we present the resulting dispersion relations $\omega(k)$ for a chain of YIG spheres oriented along the $\hat{z}$ axis with ${\bf H}_0\| {\bf M}_s \| \hat{x}$. In this geometry the two coupled $L$ and $T$ wave are elliptically polarized. The frequency of the second $T$ branch (not shown) is independent of $k$ within the quasistatic approximation.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Chain_perp_Ho8.png}
\caption{The real part of the calculated dispersion relations $\omega/(\gamma H_0)$ for a periodic array of spheres parallel to $\hat{z}$ with ${\bf H}_0 \| {\bf M}_s \| \hat{x}$. We showing the coupled $L$ and $T$ branches as functions of $kd$. The third branch (not shown) is a pure $T$ mode whose frequency is independent of $k$ within the quasistatic approximation. We use the same parameters as in Fig.~\ref{fig:two}.}
\label{fig:three}
\end{figure}
\subsection*{Cylindrical Particles}
Next, we consider the case of cylindrical particles whose long axes is parallel to $x$ in a chain which is periodically arranged parallel to $z$ (see Fig.~\ref{fig:one}). Using Eqs.~\eqref{eq:mag_moment2D},~\eqref{eq:G_rods_z},~\eqref{eq:gamma_rods}, and either~\eqref{eq:t_matrix_rod_z} or~\eqref{eq:t_matrix_rod_x} we can determine the coupled equations for the dipole moments in either of two cases: $ {\bf H}_0\|{\bf M}_s \| \hat{z}$ and ${\bf H}_0\| {\bf M}_s \| \hat{x}$. In the former case, we find from the form of the $\hat{t}$-matrix that the frequency of the $L$ branch (i.\ e., that polarized parallel to $\hat{z}$), is independent of $k$ while the two $T$ branches, which are parallel to the $x$ and $y$ axes, are coupled. The equation for these two coupled branches can be written as
\begin{equation}
{\bf m}_0 = \frac{a^2}{d^2}\begin{pmatrix}
0 & it_{xy} \\
0 & -t_{yy}
\end{pmatrix}{\bf m}_0 Cl_2(kd),
\end{equation}
or, if we include only nearest-neighbor interactions, i.e. $n' =1$,
\begin{equation}\label{eq:cyl_h0_z}
{\bf m}_0 = \frac{a^2}{d^2}\begin{pmatrix}
0 & it_{xy} \\
0 & -t_{yy}
\end{pmatrix}{\bf m}_0\cos{kd}.
\end{equation}
We may write out Eq.~\eqref{eq:cyl_h0_z} as two coupled algebraic equations:
\begin{equation}
\left[1+\frac{a^2}{d^2}t_{yy}(\omega)\cos(kd)\right]m_{y0} = 0
\label{eq:my0}
\end{equation}
and
\begin{equation}
m_{x0} - \frac{a^2}{d^2} it_{xy}(\omega)\cos(kd)m_{y0} = 0.
\label{eq:mxy0}
\end{equation}
This pair of equations may, in principle, have two solutions for each $k$, corresponding to the two possible transverse branches. We consider each in turn.
For the first solution, $m_{y0} \neq 0$. In that case, according to Eq.~\eqref{eq:my0}, the dispersion relation is determined by the implicit equation $1 + (a^2/d^2)t_{yy}\cos(kd) = 0$, since $t_{xx}$ (and $t_{xy}$) depend on $\omega$. The $k$-dependent polarization of the corresponding mode is obtained from Eq.~\eqref{eq:mxy0} and satisfies $m_{x0}/m_{y0} = (a^2/d^2)it_{xy}\cos(kd)$. Using $\cos(kd) = -(d^2/a^2)[1/t_{yy}]$, we obtain $m_{x0}/m_{y0} = -it_{xy}/t_{yy}$ along this branch. Using Eq.~\eqref{eq:t_matrix_rod_z} for $t_{xx}$ and $t_{xy}$, we find that $m_{x0}$ and $m_{y0}$ are $\pi/2$ out of phase, so this wave is elliptically polarized.
In the second solution, we take $m_{y0} = 0$, so the mode would be polarized along $\hat{x}$. Eq.~\eqref{eq:my0} is then automatically satisfied, provided that $t_{yy}$ is finite. In order for Eq.~\eqref{eq:mxy0} to be satisfied with $m_{y0} = 0$ and $m_{x0} \neq 0$, we must have $t_{xy} = \infty$ as well as $t_{yy}$ finite. But if we consider the expressions given in Eq.~\eqref{eq:t_matrix_rod_z} for $t_{xy}$ and $t_{yy}$, we find that there exists no frequency for which both $t_{xy}(\omega) = \infty$ and $t_{yy}(\omega)$ is finite. Therefore, we conclude that there is only a single propagating branch for this geometry, which is elliptically polarized, propagating along $\hat{z}$, and described by the dispersion relation $\omega(k)$ given implicitly by setting the quantity in square brackets in Eq.~\eqref{eq:my0} equal to zero.
In Fig.~\ref{fig:four} we show the dispersion relation for a chain of YIG cylinders oriented along the $x$ axis with ${\bf H}_0 \| {\bf M}_s \| \hat{z}$. In this geometry, and within the quasistatic approximation, as mentioned above, we find a branch elliptically polarized in the $xy$ plane. There is also a $z$-polarized ($L$) mode, not shown in the figure, whose frequency is independent of $k$ and which is uncoupled to the $xy$ polarized branch.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Chain_disp_cyl_z8.png}
\caption{The real part of the calculated frequency, $\omega/(\gamma H_0)$, as a function of $kd$ for the elliptically polarized $T$ branch given by Eq.~\eqref{eq:my0} when the YIG cylinder axes is parallel to $\hat{x}$ and ${\bf H}_0 \|{\bf M}_s \| \hat{z}$. }
\label{fig:four}
\end{figure}
We now consider the final case, in which the long axis of the cylinders are aligned along the $x$ axis and ${\bf H}_0 \| {\bf M}_s \|\hat{x}$. In this case we notice from the $\hat{t}$-matrix that the $x$-polarized branch, which is one of the transverse branches, decouples from the $y$ and $z$ branches and has a frequency independent of $k$. The remaining two branches obey, in the nearest-neighbor approximation, the equation
\begin{equation}
{\bf m}_0 = \frac{a^2}{d^2}\begin{pmatrix}
-t_{yy} & it_{yz} \\
it_{yz} & t_{zz}
\end{pmatrix}{\bf m}_0\cos{kd}.
\end{equation}
This equation can be solved by the methods described above to give the dispersion relations of the coupled $y$ and $z$ branches.
In Fig.~\ref{fig:five} we show the dispersion relations for a chain of YIG cylinders oriented with their long axes parallel to the $\hat{x}$ axis with ${\bf H}_0 \|{\bf M}_s \| \hat{x}$. In this geometry the coupled $L$ and $T$ waves are elliptically polarized. The third branch, which is a $T$ branch, polarized parallel to $\hat{x}$, is independent of wave number within the quasistatic approximation and is not plotted in the figure.
\section{Discussion}~\label{four}
While we show numerical results only for nearest-neighbor interactions, we have found numerically that including further neighbors does not qualitatively change these results. We do not show these numerical results in the paper, but have presented the relevant formal expressions in Eqs.~\eqref{eq:18} and~\eqref{eq:19}. We also note that the spacing between the particles in our calculations was chosen to limit the effects of higher-order dipole moments. If the particles are spaced closer together than about $a/d = 1/3$, one must consider additional contributions from magnetic quadrupole and higher modes. If these were included, the dispersion relations would quantitatively change, but the qualitative results would remain similar for the lowest bands. We have also shown that there are usually several branches of magnon waves with dispersion relations that depend on the external magnetic field magnitude and orientation. These waves propagate along the chain with different polarization- and wave-number-dependent group velocities.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Chain_disp_cyl_x8.png}
\caption{The real part of the calculated frequencies, $\omega/(\gamma H_0)$, for the coupled $L$ and $T$ modes as functions of $kd$ when the axis of the YIG cylinders is parallel to $\hat{x}$ and ${\bf H}_0 \|{\bf M}_s \| \hat{x}$. }
\label{fig:five}
\end{figure}
As mentioned earlier, these magnon waves will transmit power along the chains. The energy density is proportional to the square of the absolute wave amplitude, and the transmitted power is therefore equal to the energy density multiplied by the group velocity of the wave. The group velocity $v_{gn}(k)$ of the n$^{th}$ polarization can be computed from the dispersion relation via the relation $v_{ng}(k) = d\omega_n(k)/dk$. Just as for plasmonic waves on metallic particle chains, the transmitted power can be controlled by introducing geometries such as T-junctions, which allow the power to be split into two parts with relative magnitudes depending on frequency~\cite{Brongersma2000}. The calculations of the splitting will be more complicated, however, because the waves may be circularly or elliptically polarized. Similarly, one can calculate the magnetization current due to these waves.
The dispersion relations of magnons on chains of either spheres or cylinders calculated here include the effects of Gilbert damping to describe the relaxation of the magnetic moments within each particle. Additionally, one could also include the effects of a magnetic torque, which would further couple the three polarization modes, or one could include the effects of crystalline anisotropy or demagnetization fields. Since both crystalline anisotropy and demagnetization fields are frequency independent they would directly affect the permeability matrix elements $\mu_1$ and $\mu_2$~\cite{Moorish1965}. Therefore, the effects of these terms will quantitatively change the dispersion relations calculated here, but will not affect the qualitative results. In addition, since the crystal anisotropy is a sensitive function of temperature~\cite{Moorish1965} one can modify the dispersion relations for different polarizations in a controllable manner by varying the temperature. It should also be possible to include effects beyond the quasistatic approximation by extending the approach of Weber and Ford~\cite{Weber2004} to magnon waves propagating via magnetic dipole-dipole interactions.
To summarize, we have calculated the dispersion relations for magnon waves propagating along chains of YIG spheres and cylinders, in which only the quasistatic coupling between magnetic dipoles is included. We found that, depending on the orientation of the static magnetization and applied magnetic field relative to the chain, these waves are either circularly, elliptically, or linearly polarized. In the first two cases, an incident linearly polarized wave will undergo Faraday rotation, analogous to that seen in bulk magnetic compounds~\cite{Fu2008}, and this rotation can be tuned in a controllable way. These waves also carry magnetization current (magnon current) along the chain. Thus, it should be possible to use chains of YIG particle to transmit power wirelelssly in a meso- or nanoscale circuit without generation of large electric potentials and within a diameter small compared to the wavelength,
\begin{acknowledgments}
N. P was supported by the Belgian Fonds National de la Recherche Scientifique FNRS under grant number PDR T.1077.15-1/7 and both authors acknowledge funds from the Center for Emerging Materials at The Ohio State University, an NSF MRSEC (Grant No.\ DMR-1420451).
\end{acknowledgments}
|
\section{ Introduction}
The foundational tenets and concepts of quantum mechanics (QM) significantly differ from classical ideas and intuitions. A seminal contribution to quantum concepts was provided by demonstrating quantum nonlocality through Bell's inequality \cite{Bell,CHSH} used for showing an incompatibility between quantum mechanics (QM) and the notion of local realism underpinning Bell's inequality. Soon after the discovery of Bell's inequality, a different formulation of local realist inequality was provided by Wigner \cite{Wigner}. This was based upon the assumption of the existence of joint probability distributions in the underlying stochastic hidden variable (HV) space pertaining to the occurrence of different possible combinations of outcomes for the measurements of the relevant observables, and these joint probability distributions are taken to yield all the observable marginal probabilities by satisfying the locality condition. However, Wigner's original formulation was restricted in showing the QM incompatibility with local realism for the bipartite qubit singlet states.
Subsequently, among the few studies using Wigner's approach are its use in the case of entangled neutral kaons \cite{Domenico,Bramon}, and a study of its implication for quantum key distribution \cite{gw}. Only recently, Wigner's formalism has been generalized for $N$-partite qubit states by deriving generalized Wigner inequalities (GWI) \cite{GWI}, and in another recent work, the temporal version of GWI, namely, Wigner's form of the Leggett-Garg inequality has been derived \cite{wlgi}. Apart from these investigations, surprisingly, Wigner's approach has remained largely unexplored.
Against this backdrop, the motivation underlying the present paper is to extend the significance of Wigner's approach in the context of bipartite qutrit systems by developing a framework for local realist inequalities based on the assumption of the existence of joint probability distributions. Here it needs to be mentioned that investigations related to QM violations of local realist inequalities for arbitrary dimensional systems have steadily acquired much interest over the years \citep{t1, t2, t3, t4, t5, t6, t8, t9, t10, home1, cabello, t11}. In this context, we should also recall that qutrit systems are of special interest due to their experimental relevance in the areas of atomic and laser physics, as well as because of a number of foundational and information theoretic applications of qutrit systems \cite{bq1, bq2, bq6, bq7, bq10, bq11, bq12, bq13, bq14, bq15, bq16}.
For the purpose of probing quantum nonlocality of bipartite qutrit systems, particularly noteworthy is the QM incompatibility with local realism for bipartite qutrit isotropic states as studied by Collins et. al. \citep{cglmp} using the local realist inequality derived by them (known as the Collins-Gisin-Linden-Massar-Popescu inequality or CGLMP inequality). While experimental violation of the CGLMP inequality has been demonstrated for non-maximally entangled states of bipartite qutrits \citep{bq8}, here it needs to be mentioned that isotropic and singlet qutrit states are regarded to be particularly relevant in quantum information processing \cite{qic}. Hence in this paper, these states are used for studying the QM violation of the derived forms of GWI in the context of the six-port beam splitter and spin-$1$ component observables. Note that the forms of GWI for bipartite qutrits derived in this paper do not reduce to Bell-CHSH (Bell-Clauser-Horne-Shimony-Holt) type inequalities by clubbing any two outcomes.
An important point to stress is that the efficacy of any local realist inequality for demonstrating its incompatibility with QM is restricted in practical situations that are usually far from ideal. Hence the robustness of the QM violation of any local realist inequality in the presence of white noise in a given state is a key issue. The present paper provides a comparative study of the robustness of the QM violation of both the GWI and CGLMP inequality in the presence of white noise incorporated in the qutrit states considered. Results obtained in this paper demonstrate that for six-port beam splitter observables, the QM violation of CGLMP inequality is more robust against white noise for bipartite qutrit \textit{isotropic} states than that obtained by using GWI. On the other hand, for bipartite qutrit \textit{singlet} states, the QM violation of GWI is more robust against white noise than that pertinent to the CGLMP inequality. The corresponding calculations are also done for the spin-$1$ component observables. It is found that for both these types of states, the QM violation of GWI is more robust against white noise than that of the CGLMP inequality.
Interestingly, it may happen that the maximum QM violation of a local realist inequality is not obtained for maximally entangled qutrit state. In order to probe this, the maximum QM violation of GWI and the corresponding robustness against white noise present in the state has been calculated. It has been found that the maximum QM violation of GWI occurs for non-maximally asymmetric entangled qutrit state if one uses six-port beam splitter or spin-$1$ component observables. The maximum robustness of the QM violation of GWI against white noise present in a state is also compared with that of CGLMP inequality \cite{cglmp}.
Since all pure bipartite entangled states violate Bell-type local realist inequalities \cite{gisin}, it was believed that entanglement is equivalent to such violation. After the work of Werner \cite{werner}, it turned out that all mixed entangled states do not violate Bell-type local realist inequalities. Thus the issue of QM violations of local realism by mixed states is worth to be investigated. Motivated by this fact, we have also discussed the QM violation of GWI for some specific classes of mixed bipartite qutrit states contingent upon using six-port beam splitter as well as spin-$1$ component observables.
The plan of this paper is as follows. In the next Section II we briefly outline the original derivation by Wigner applicable for bipartite qubit singlet state. Then, in Section III, we present the derivation of GWI for bipartite qutrit systems, followed by Sections IV and V where it is shown that the derived GWI is violated by isotropic and singlet qutrit states using six-port beam splitter and spin-$1$ component observables respectively. In Sections VI and VIII, contingent upon using six-port beam splitter and spin-$1$ observables respectively, we compare the robustness of QM violations of GWI with that of CGLMP inequality for the case of isotropic and singlet qutrit states. We consider the introduction of white noise to the pure states considered in order to perform the comparative study of robustness of GWI and CGLMP inequality corresponding to the above mentioned two categories of entangled qutrit states. In Sections VII and IX we have shown the maximum QM violations of GWI using six-port beam splitter and spin-$1$ component observables respectively, and the corresponding maximum robustness of the QM violations of GWI against white noise present in the states. QM violations of GWI for mixed bipartite qutrits have been discussed in Section X. Section XI contains a summary of the results obtained in this paper and we make some concluding remarks.
\section{Recapitulating Wigner's original derivation}
In the scenario considered by Wigner \citep{Wigner}, two spin-$1/2$ particles are prepared in a singlet state and are then spatially separated. The spin components of the particles, respectively, are measured along three directions, say, $a,b$ and $c$. Then, in this context, considering the individual outcomes ($\pm 1$) of nine possible pairs of measurements, Wigner's original inequality can be derived as follows.
Assuming the locality condition and an underlying stochastic HV distribution corresponding to a quantum state specified by a wave function, one can infer in the HV space, according to the reality condition, the existence of overall joint probabilities for the individual outcomes of measuring the pertinent observables, from which the observable marginal probabilities can be obtained. Thus, corresponding to an underlying stochastic HV, say $\lambda$, one can define $p_\lambda(v_1(a),v_1(b),v_1(c);v_2(a),v_2(b),v_2(c))$ as the overall joint probability of occurrence of the outcomes, where $v_1(a)$ represents an outcome ($\pm 1$) of the measurement of the observable $a$ for the first particle, and so on. For example, $p_\lambda(-,+,-;+,+,-)$ expresses the overall joint probability of occurrence of the outcomes $v_1(a)=-1,v_1(b)=+1,v_1(c)=-1$ for the first particle, and $v_2(a)=+1,v_2(b)=+1,v_2(c)=-1$ for the second particle. Then, the joint probability, say, $v_1(a)=+1$ and $v_2(b)=+1$ for the first and the second particle respectively can be written, using the perfect anti-correlation property of the singlet state, as $ p_\lambda(a+,b+)=p_\lambda(+,-,+;-,+,-)+p_\lambda(+,-,-;-,+,+)$. Similarly, writing $p_\lambda(c+,b+)$ and $p_\lambda(a+,c+)$ as marginals, and assuming non-negativity of the overall joint probability distributions in the HV space, it can be shown that
\begin{equation}
\label{equ0}
p_\lambda(a+,b+)\leq p_\lambda(a+,c+)+p_\lambda(c+,b+)
\end{equation}
Subsequently, by integrating over the hidden variable space for an arbitrary distribution, one can obtain the original form of Wigner's inequality
\begin{equation}
\label{equ1}
p(a+,b+)\leq p(a+,c+)+p(c+,b+).
\end{equation}
where $p(a+,b+)$ is the observable joint probability of getting $+1$ for both the outcomes if the observables $a$ and $b$ are measured on the first and the second particle respectively, and so on.
If the respective angles between $a$ and $b$, $a$ and $c$, $b$ and $c$ are $\theta_{12},\theta_{13}$ and $\theta_{23}$, then substituting the QM expressions for the relevant joint probabilities in the inequality given by Eq.(\ref{equ1}) one obtains $\frac{1}{2} \sin^2(\theta_{12}/2) \leq \frac{1}{2} \sin^2(\theta_{13}/2) +\frac{1}{2} \sin^2(\theta_{23}/2)$ - a relation which is not valid for arbitrary values of $\theta_{12},\theta_{13}, \theta_{23}$. This shows an incompatibility between QM and Wigner's form of inequality given by Eq.(\ref{equ1}), restricted for the singlet state in the bipartite case. Note that, the above argument is within the framework of stochastic HV theory, subject to the locality condition, and the notion of determinism has not been used here.
\section{Generalized Wigner inequalities for bipartite qutrit systems}
Now, in order to generalise the above argument for deriving GWI for arbitrary bipartite qutrit systems, we proceed as follows. Note that in the following derivation we are not using the assumption of perfect anti-correlation embodied in the singlet states that was used in Wigner's original derivation. Let us consider that pairs of trichotomic observables $a^1$ or $a^2$ and $b^1$ or $b^2$ are measured on the first and the second particle respectively. We assume an underlying HV distribution given by $\rho(\lambda)$ such that for $3^4$ possible combinations of pairs of outcomes, each such pair of outcomes occur with a certain probability in the HV space. Thus, corresponding to an underlying stochastic HV, say $\lambda$, one can define $p_{\lambda}(v_1(a^1), v_1(a^2); v_2(b^1), v_2(b^2))$ as overall joint probability of occurrence of the outcomes, where $v_1(a^1)$ represents an outcome ($+1$, or $0$, or $-1$) of the measurement of the observable $a^1$ for the first particle, and so on. For example, $p_{\lambda}(+,0;-,+)$ expresses the overall joint probability of occurrence of the outcomes $v_1(a^1) = +1$ and $v_1(a^2) = 0$ for the first particle and $v_2(b^1) = -1$ and $v_2(b^2) = +1$ for the second particle. Then, consistent with the locality condition, the joint probability of, say $v_1(a^1) = 0$ and $v_2(b^1)=-$ for the first and second particle, respectively, can be obtained as a marginal of the overall joint probabilities in the HV space, given by the following expression
\begin{equation}
p_{\lambda}(a^1 0 ,b^1 -) = \sum_{v_1(a^2)=+,0,-} \sum_{v_2(b^2)=+,0,-} p_{\lambda}(0, v_1(a^2); -, v_2(b^2)) \nonumber
\end{equation}
Similarly, writing $p_{\lambda}(a^1 -, b^1 -)$, $p_{\lambda}(a^2 0, b^1 -)$, $p_{\lambda}(a^2 -, b^1 -)$, $p_{\lambda}(a^1 0, b^2 -)$, $p_{\lambda}(a^1 -, b^2 -)$, $p_{\lambda}(a^2 +, b^2 +)$ and $p_{\lambda}(a^2 +, b^2 0)$ as marginals, and assuming non-negativity of the overall joint probability distributions in the HV space, it can be shown that
\begin{equation}
p_{\lambda}(a^1 0, b^1 -) - p_{\lambda}(a^2 0, b^1 -) - p_{\lambda}(a^2 -, b^1 -)- p_{\lambda}(a^1 0, b^2 -) - p_{\lambda}(a^1 -, b^2 -) - p_{\lambda}(a^2 +, b^2 +) - p_{\lambda}(a^2 +, b^2 0) + p_{\lambda}(a^1 -, b^1 -) \leq 0 \nonumber
\end{equation}
Subsequently, integrating over the HV space for an arbitrary distribution, one can obtain the following form of GWI for bipartite qutrit systems:
\begin{equation}
\label{equgwi}
p(a^1 0, b^1 -) - p(a^2 0, b^1 -) - p(a^2 -, b^1 -) - p(a^1 0, b^2 -) - p(a^1 -, b^2 -) - p(a^2 +, b^2 +) - p(a^2 +, b^2 0) + p(a^1 -, b^1 -) \leq 0
\end{equation}
Similarly, other forms of $8$-term GWI can be derived by using various combinations of the observable joint probabilities. Such forms of GWI (including the above form mentioned in Eq.(\ref{equgwi})) can be expressed by the following two inequalities:
\begin{equation}
p(a^1=m_1, b^1 = m_2) - p(a^2 = m_1, b^1 = m_2) - p(a^2 = m_1, b^1 = m_1) - p(a^1 = m_1, b^2 = m_2) - p(a^1 = m_1, b^2 = m_1) \nonumber
\end{equation}
\begin{equation}
\label{equgwi2}
- p(a^2 = m_3, b^2 = m_3) - p(a^2 = m_2, b^2 = m_3) + p(a^1 = m_1, b^1 = m_1) \leq 0
\end{equation}
and
\begin{equation}
p(a^1 = m_1, b^1 = m_2) - p(a^2 = m_1, b^1 = m_2) - p(a^2 = m_2, b^1 = m_2) - p(a^1 = m_1, b^2 = m_2) - p(a^1 = m_2, b^2 = m_2) \nonumber
\end{equation}
\begin{equation}
\label{equgwi3}
- p(a^2 = m_3, b^2 = m_3) - p(a^2 = m_3, b^2 = m_1) + p(a^1 = m_2, b^1 = m_2) \leq 0
\end{equation}
There are six permutations of the set $\{ m_1, m_2, m_3 \}$, namely: $(+1, 0, -1)$, $(+1, -1, 0)$, $(0, +1, -1)$, $(0, -1, +1)$, $(-1, +1, 0)$ and $(-1, 0, +1)$, which produce twelve GWI from the inequalities (\ref{equgwi2}) and (\ref{equgwi3}) (including the GWI mentioned in Eq.(\ref{equgwi})).
Now, interchanging $a \leftrightarrow a'$, or $b \leftrightarrow b'$, or interchanging both other three sets of twelve such $8$-term GWI can be obtained for the bipartite qutrit system. QM violations of all the aforementioned GWIs for bipartite qutrit system are quantified by the positive value of the left hand side of each inequality.
Here it needs to be stressed that this set of inequalities is such that none of these inequalities can be reduced to equivalent classes of Bell-CHSH inequalities by grouping any two outcomes (for details, see Appendix A).
\section{QM violations of GWI by bipartite qutrit isotropic and singlet states using six-port beam splitter}
The phenomenon of spontaneous parametric down-conversion can be used to obtain an optical analog of the maximally entangled state for two correlated spins of arbitrary magnitudes \cite{spdc}. Next, to make measurements of nondichotomic observables, it is experimentally more convenient to use six-port (or, multi port) beam splitters than spin component observables.
The properties of the unbiased six-port beam splitter (three input and three output ports) have been demonstrated in detail in several works \cite{tt1, tt2, cglmp, six, six1, six2, six3}. One considers the following settings: first the two parties apply unitary operations on each subsystem with non-zero diagonal terms equal to $e^{i \phi_a(j)}$ and $e^{i \varphi_b(j)}$ for the first and second particle respectively, and all off-diagonal terms being equal to zero. These unitary operations are denoted by $U(\vec{\phi_a})$, where $\vec{\phi_a} \equiv [\phi_a(0), \phi_a(1), \phi_a(2)]$ for the first particle and $U(\vec{\varphi_b})$, where $\vec{\varphi_b} \equiv [\varphi_b(0), \varphi_b(1), \varphi_b(2))]$ for the second particle. The freedom of choice of the measurement of both the particles is given by this unitary transformation. Then, a discrete Fourier transformation $U_{FT}$ is carried out on the first particle and $U_{FT}^*$ is carried out on the second particle. The matrix element of the discrete Fourier transformation is given by, ${(U_{FT})}_{jk}$ = $exp[(j-1)(k-1)i2\pi /3]$ and finally measurement is done in the basis in which the initial shared state is prepared. Here the observables $a^1$, $a^2$, $b^1$ and $b^2$ denote unitary transformations $U(\vec{\phi_{a^1}})$, $U(\vec{\phi_{a^2}})$, $U(\vec{\varphi_{b^1}})$ and $U(\vec{\varphi_{b^2}})$ respectively, where $\vec{\phi_{a^1}} \equiv [\phi_{a^1}(0), \phi_{a^1}(1), \phi_{a^1}(2)]$, $\vec{\phi_{a^2}} \equiv [\phi_{a^2}(0), \phi_{a^2}(1), \phi_{a^2}(2)]$, $\vec{\varphi_{b^1}} \equiv [\varphi_{b^1}(0), \varphi_{b^1}(1), \varphi_{b^1}(2))]$ and $\vec{\varphi_{b^2}} \equiv [\varphi_{b^2}(0), \varphi_{b^2}(1), \varphi_{b^2}(2))]$.
\subsection{QM violation of GWI for bipartite qutrit isotropic state using six-port beam splitter}
Let us consider the pure isotropic qutrit state given by
\begin{eqnarray}
\label{state1}
\vert\psi_1\rangle = \frac{\vert 00 \rangle + \vert 11 \rangle + \vert 22 \rangle}{\sqrt{3}}
\end{eqnarray}
where $\vert 0 \rangle$, $\vert 1 \rangle$ and $\vert 2 \rangle$ are three mutually orthonormal states. In the case of six-port beam splitter, each of these states defines the state of photon passing through one of the three input ports or one of the three output ports of the six-port beam splitter. On the other hand, in the case of spin-$1$ component observables, $\vert 0 \rangle$, $\vert 1 \rangle$ and $\vert 2 \rangle$ are the eigenstates of spin angular momentum operator along z-direction corresponding to the eigenvalues $+1$, $0$ and $-1$ respectively (assuming $\hbar = 1$).
If measurements defined by the six-port beam splitter are performed on two particles of the state given by Eq.(\ref{state1}), the left hand side of the GWI given by Eq.(\ref{equgwi}) becomes
\begin{equation}
\label{gwilhs11}
W = \frac{1}{27}[-12 - 2 ( \sum_{i=1}^{2} \sum_{j=1}^{2} \sum_{k=0}^{2} (-1)^{\delta_{i+j,2}} \big[ sin(\frac{\pi}{6}-\alpha_{ik}-\beta_{jk}) \big]
\end{equation}
where, $\delta_{i+j,2}$ is the Kronecker delta function; $\alpha_{ik}=\big[ \phi_{a^i}(k)-\phi_{a^i}(k+1 \text{ mod } 3) \big]$; $\beta_{jk} = \big[ \varphi_{b^j}(k)-\varphi_{b^j}(k+1 \text{ mod } 3) \big]$. In order to obtain the maximum QM violation of GWI given by Eq.(\ref{equgwi}) for pure isotropic qutrit state, we have to maximize the right hand side of Eq.(\ref{gwilhs11}), where $0 \leq \phi_{a^i} (j) \leq 2 \pi$ and $0 \leq \varphi_{b^i} (j) \leq 2 \pi$ ($i=1,2$; and $j=0,1,2$). For this maximization we have used a numerical procedure (analytical maximization is too difficult, because one has to find the global maximum of a twelve-variable function defined on some bounded twelve-dimensional domain) based on the downhill simplex method (so-called Nelder-Mead method or amoeba method) \cite{nm}. If the dimension of the domain of a function is $D$ (in our case $D=12$), the procedure first randomly generates $D+1$ points. In this way it creates the vertices of a starting the simplex. Next it calculates the value of the function at the vertices and starts exploring the space by stretching and contracting the simplex. In every step, when it finds vertices where the value of the function is higher than in others, it goes in this direction \cite{nm}.
Following this numerical procedure we observe that for the set of measurement settings ($\phi_{a^1}(0)$, $\phi_{a^1}(1)$, $\phi_{a^1}(2)$, $\phi_{a^2}(0)$, $\phi_{a^2}(1)$, $\phi_{a^2}(2)$, $\varphi_{b^1}(0)$, $\varphi_{b^1}(1)$, $\varphi_{b^1}(2)$, $\varphi_{b^2}(0)$, $\varphi_{b^2}(1)$, $\varphi_{b^2}(1) $)= (4.62, 3.02, 3.93, 2.46, 1.80, 0.81, 0.43, 4.80, 4.64, 4.01, 3.04, 0.98) in radians, the maximum QM violation of the GWI (\ref{equgwi}) occurs, and the magnitude of this maximum violation is found to be 0.12949.
\subsection{QM violation of GWI for bipartite qutrit singlet state using six port beam splitter}
Similar to the way discussed above, it can be shown that if the trichotomic measurements labeled by ($a^1; a^2; b^1; b^2$) denoting observables using six-port beam splitter are performed on the $3\otimes 3$-dimensional pure singlet state given by
\begin{eqnarray}
\label{state2}
\vert\psi_2\rangle = \frac{\vert 02 \rangle - \vert 11 \rangle + \vert 20 \rangle}{\sqrt{3}}
\end{eqnarray}
Following the numerical procedure based on the downhill simplex method \cite{nm} as described earlier, we obtain that the left hand side of the GWI given by, Eq.(\ref{equgwi}) has the maximum value 0.12949 for the measurement settings ($\phi_{a^1}(0)$, $\phi_{a^1}(1)$, $\phi_{a^1}(2)$, $\phi_{a^2}(0)$, $\phi_{a^2}(1)$, $\phi_{a^2}(2)$, $\varphi_{b^1}(0)$, $\varphi_{b^1}(1)$, $\varphi_{b^1}(2)$, $\varphi_{b^2}(0)$, $\varphi_{b^2}(1)$, $\varphi_{b^2}(1) $)= (4.05, 0.11, 4.45, 3.02, 0.03, 2.47, 3.53, 1.87, 2.50, 6.20, 0.17, 6.13) (in radian) corresponding to the maximum QM violation of the GWI. Numerical calculations show that GWI mentioned in Eq.(\ref{equgwi}) gives the maximum QM violation for both bipartite qutrit isotropic state and bipartite qutrit singlet state among all the GWIs derived in this paper. Henceforth, we would, therefore, consider only the GWI given by Eq.(\ref{equgwi}) in case of observables using six-port beam splitters.
\section{QM violations of GWI by bipartite qutrit isotropic and singlet states using spin-$1$ component observables}
Let us assume, $a^i$ denotes measurements of spin component of the first particle in the directions $\hat{n_i^a} = sin\theta_i^{a} cos\phi_i^{a} \hat{x} + sin\theta_i^{a} sin\phi_i^{a} \hat{y} + cos\theta_i^{a} \hat{z}$ ($i=1,2$). Similarly, $b^j$ denotes measurements of spin component of the second particle in the directions $\hat{n_j^{b}} = sin\theta_j^{b} cos\phi_j^{b} \hat{x} + sin\theta_j^{b} sin\phi_j^{b} \hat{y} + cos\theta_j^{b} \hat{z}$ ($j=1,2$), where $\theta_i^a$, $\theta_j^b$ ($i,j=1,2$) are the polar angle; $\phi_i^a$, $\phi_j^b$ ($i,j=1,2$) are the azimuthal angle; $\hat{x}$, $\hat{y}$, and $\hat{z}$ are the unit vectors in Cartesian coordinates.
\subsection{QM violation of GWI for bipartite qutrit isotropic state using spin-$1$ component observables}
If the measurements of spin-$1$ components in arbitrary directions are performed on the isotropic state (\ref{state1}), the left hand side of the GWI given by Eq.(\ref{equgwi}) becomes
\begin{equation}
\label{gwilhs1}
\begin{split}
& W = \sum_{i=1}^{2} \sum_{j=1}^{2} \frac{(-1)}{12} (-1)^{\delta_{i+j,2}} \Big[ sin^{2}( \theta_i^a + \theta_j^b ) + 2 (1 - sin \theta_i^a sin \theta_j^b ) cos \theta_i^a cos \theta_j^b + 2 (1 - cos \theta_i^a cos \theta_j^b ) cos (\phi_i^a + \phi_j^b ) sin \theta_i^a sin \theta_j^b \\&
+ sin^2 (\phi_i^a + \phi_j^b ) sin^2 \theta_i^a sin^2 \theta_j^b + 2 \Big]
\end{split}
\end{equation}
Here, $\delta_{i+j,2}$ is the Kronecker delta function; $0 \leq \theta_i^a \leq \pi$, $0 \leq \theta_j^b \leq \pi$, $0 \leq \phi_i^a \leq 2 \pi$ and $0 \leq \phi_j^b \leq 2 \pi$ ($i,j = 1,2$). Following the numerical procedure based on the downhill simplex method \cite{nm} as described earlier, it has been observed that for the set of measurement settings ($\theta_1^a,\phi_1^a;\theta_2^a,\phi_2^a;\theta_1^b,\phi_1^b;\theta_2^b,\phi_2^b$)= (1.52, 3.88; 2.60, 3.84; 0.03, 0.76; 1.08, 5.56) in radians, the maximum QM violation of the GWI (\ref{equgwi}) occurs, and the magnitude of this maximum violation is found to be 0.12077.
\subsection{QM violation of GWI for bipartite qutrit singlet state using spin-$1$ component observables}
Similar to the way discussed above, using the numerical procedure based on the downhill simplex method \cite{nm}, it can be shown that if the trichotomic measurements using spin-$1$ component observables are performed on the $3\otimes 3$-dimensional pure singlet state given by Eq.(\ref{state2}), the left hand side of the GWI given by Eq.(\ref{equgwi}) has the maximum value 0.12077 for the measurement settings ($\theta_1^a,\phi_1^a;\theta_2^a,\phi_2^a;\theta_1^b,\phi_1^b;\theta_2^b,\phi_2^b$) = (1.09, 0.05; 0.02, 0.01; 0.52, 3.19; 0.56, 0.05) (in radian) corresponding to the maximum QM violation of the GWI. Numerical calculations based on the downhill simplex method \cite{nm} show that GWI mentioned in Eq.(\ref{equgwi}) gives the maximum QM violation for both bipartite qutrit isotropic state and bipartite qutrit singlet state among all the GWIs derived in this paper. Henceforth, we would, therefore, consider only the GWI given by Eq.(\ref{equgwi}) in case of spin-$1$ component observables.
\section{Comparison of GWI with the CGLMP inequality for bipartite qutrits contingent upon using six-port beam splitter}
In order to show the efficacy of GWI derived here, we will now make a comparative analysis of the QM violation obtained through GWI with that obtained by using the CGLMP inequality for bipartite qutrits using six-port beam splitters.
The CGLMP inequality \cite{cglmp} is derived based on a constraint that the correlations exhibited by a local realist theory must satisfy. This inequality has not been derived from the assumption of existence of a joint probability distribution (JPD) in the HV space. The CGLMP inequality for bipartite $3$-dimensional system has the following form
\begin{equation}
\label{cglmp}
I_3 = P(a^1=b^1)+P(b^1=a^2+1)+P(a^2=b^2)+P(b^2=a^1) - P(a^1=b^1-1)-P(b^1=a^2)-P(a^2=b^2-1)-P(b^2=a^1-1) \leq 2
\end{equation}
where, $P(a^i = b^j + k)$ denotes the probability that the measurements $A^i$ and $B^j$ have outcomes that differ, modulo $3$, by $k$. The QM violation of the CGLMP inequality is quantified by $(I_3 - 2)$.
Here it may be noted that Wu et. al. had suggested another local realist inequality \cite{wu} which is derived based on the assumption of a local HV model satisfying the factorizability condition, using a few algebraic theorems and basic concepts of probability theory. This inequality has the following form
\begin{equation}
\label{wu}
S = P (a^1 +, b^1 +) - P (a^1 +, b^2 +) + P (a^2 +, b^2 +) + P (a^2 0, b^1 0) + P (a^2 0, b^1 -) + P (a^2 -, b^1 0) + P (a^2 -, b^1 -) \leq 1
\end{equation}
However, an important point is that this inequality (\ref{wu}) reduces to just a version of CHSH (Clauser-Horne-Shimony-Holt) inequality \cite{CHSH} after one has grouped the outcomes ``$0$" and ``$-$" so that the inequality becomes a two outcome (``$+$" and ``not $+$") inequality. Now, it is well known that in a $2 \times 2 \times 2$ experiment (2 parties, 2 measurement settings per party, 2 outcomes per settings), all generalised Bell inequalities are simply re-writings of the CHSH inequality, obtained by linear combinations of the CHSH inequality with the appropriate normalisation conditions. Thus, the inequality (\ref{wu}) is equivalent to CHSH inequality. We will, therefore, not consider this inequality for probing efficacy of GWI for bipartite qutrits.
Before computing the effects of white noise incorporated in the states considered to GWI and CGLMP inequality, let us first obtain the maximum QM violations of CGLMP inequality for the bipartite qutrit isotropic and singlet states respectively using six-port beam splitter. If the left hand side of CGLMP inequality (\ref{cglmp}) is evaluated for isotropic state in terms of the four aforementioned trichotomic observables $a^1, a^2, b^1, b^2$ denoting observables using six-port beam splitters, then the inequality is maximally violated for the choice of measurement settings ($\phi_{a^1}(0)$, $\phi_{a^1}(1)$, $\phi_{a^1}(2)$, $\phi_{a^2}(0)$, $\phi_{a^2}(1)$, $\phi_{a^2}(2)$, $\varphi_{b^1}(0)$, $\varphi_{b^1}(1)$, $\varphi_{b^1}(2)$, $\varphi_{b^2}(0)$, $\varphi_{b^2}(1)$, $\varphi_{b^2}(1) $) = (0, 3.13, 2.64, 2.51, 4.60, 6.19, 3.73, 0.07, 1.62, 2.14, 5.81, 5.26) (in radian), and the magnitude of the maximum violation is given by 0.87293 \cite{cglmp, tt2} (Following the numerical procedure based on the downhill simplex method \cite{nm}). On the other hand, for singlet states given by Eq.(\ref{state2}), CGLMP inequality (\ref{cglmp}) is \textit{not} violated for \textit{arbitrary} choice of measurement settings.
In order to probe the efficacy of the derived GWI, we now compare the tolerances of GWI and CGLMP inequality against white noise present in a state pertaining to measurement of observables using six-port beam splitters. For this, let us consider the bipartite qutrit mixed state given by,
\begin{eqnarray}
\label{mixedstate}
\rho = p \vert \psi \rangle \langle \psi \vert + (1-p) \frac{\mathbb{I}_3\otimes \mathbb{I}_3}{3^2}
\end{eqnarray}
where, $p$ is the visibility parameter which changes the pure state $\vert \psi \rangle$ into a mixed state $\rho$ and $(1-p)$ denotes the amount of white noise present in the state $| \psi \rangle$ (Here we take $|\psi\rangle$ to be either the isotropic state (\ref{state1}) or the singlet state (\ref{state2})). $p=0$ denotes the maximally mixed separable state.
Now, we first consider Eq.(\ref{mixedstate}) by taking $|\psi\rangle$ as the isotropic state given by Eq.(\ref{state1}), and compute respectively the left hand side of the various local realist inequalities for the pure state $|\psi\rangle$ pertaining to measurement using six-port beam splitters. Subsequently, we repeat the computation by only taking the white noise part of Eq.(\ref{mixedstate}). After applying appropriate weightage using the visibility parameter, we obtain the various expressions of the left hand sides of the local realist inequalities corresponding to the mixed state $\rho$ mentioned in Eq.(\ref{mixedstate}) in terms of the parameter $p$. The same procedure is followed for the case of the singlet state (\ref{state2}). The minimum values of $p$ for which QM violates local realist inequalities signify the maximum amounts of white noise that can be present in the given state for the persistence of the QM violation of the relevant local realist inequality, and this value of $p$ is known as the threshold visibility pertaining to the given local realist inequality. In the Table (\ref{tab3}), the threshold visibilities of GWI and CGLMP inequality pertaining to the bipartite qutrit isotropic and singlet states, using six-port beam splitters, are shown.
The above mentioned Table clearly shows that for the six-port beam splitter case, the CGLMP inequality given by Eq.(\ref{cglmp}) is more robust than GWI for the persistence of the QM violation in the presence of white noise incorporated in qutrit isotropic states. On the other hand, GWI given by Eq.(\ref{equgwi}) is more robust than the CGLMP inequality given by Eq.(\ref{cglmp}) for the persistence of the QM violation in the presence of white noise incorporated in qutrit singlet states. Moreover, CGLMP inequality is not violated at all by QM for qutrit singlet states using six-port beam splitter.
\begin{center}
\begin{table}
\begin{tabular}{|*{3}{c|}}
\hline
{\textit{\textbf{State}}} & \multicolumn{2}{|c|}{\textit{\textbf{Threshold visibility of}}}\\
\cline{2-3}
& \textit{\textbf{GWI given by Eq.(\ref{equgwi})}} & \textit{\textbf{CGLMP inequality}} \\
\hline
\hline
Isotropic & $0.774$ & $0.696$ \\
\hline
Singlet & $0.774$ & $---$ \\
\hline
\end{tabular}
\caption{Threshold visibilities of GWI and CGLMP inequality for the bipartite qutrit isotropic state and singlet state using six-port beam splitter.} \label{tab3}
\end{table}
\end{center}
\section{Maximal violations of GWI and CGLMP inequality contingent upon using six-port beam splitter}
It may happen that the maximum violation of a local realist inequality is not obtained for maximally entangled states, like singlet states or isotropic states, but is obtained rather for non-maximally entangled states. One can derive the Bell operator corresponding to a local realist inequality, when observables using six-port beam splitters are measured. Any typical joint probability, say, $P(a^1=+, b^1=-)$ of obtaining outcomes $+$ and $-$ respectively, when the observable $a^1$ is measured on the first particle and the observable $b^1$ in measured on the second particle, and the initial state is $|\psi \rangle \in \mathbb{C}^{n}$, is given by,
\begin{equation}
P(a^1 =+, b^1 =-) =\langle \psi |( \{ V( \vec{\phi_{a^1}})\dagger \otimes V( \vec{\varphi_{b^1}})\dagger \} \{ |+ \rangle \langle +| \otimes |- \rangle \langle -| \} \{V( \vec{\phi_{a^1}}) \otimes V( \vec{\varphi_{b^1}})\}) | \psi \rangle
\end{equation}
where $V( \vec{\phi_{a^1}}) = U_{FT} U( \vec{\phi_{a^1}})$ and $V( \vec{\varphi_{b^1}}) = U_{FT}^* U( \vec{\varphi_{b^1}})$.
Similarly, evaluating other joint probabilities, the left hand side of any local realist inequality can be expressed for the initial state $|\psi \rangle \in \mathbb{C}^{n}$ as $\langle \psi | B | \psi \rangle$, where $B$ is the Bell operator associated with the respective local realist inequality for bipartite qutrits corresponding to using six-port beam splitter. $B$ is a $3 \otimes 3$ Hermitian Matrix. Now, for the purpose of finding the maximum eigenvalue of the Bell operator associated with a particular local realist inequality, we use the Min-Max Theorem of functional analysis and linear algebra. According to Min-Max theorem, the largest and smallest eigenvalues of a Hermitian matrix $\hat{A} \in \mathbb{C}^{n \otimes n}$ can be found as, $\lambda_{max}$ = $ \max_{\forall x \in \mathbb{C}^{n}, x \neq 0} \frac{\langle x| \hat{A} | x \rangle}{\langle x| x \rangle}$ and $\lambda_{min}$ = $ \min_{\forall x \in \mathbb{C}^{n}, x \neq 0} \frac{\langle x| \hat{A} | x \rangle}{\langle x| x \rangle}$ respectively.
Using the above mentioned procedure it is found that, contingent upon using six-port beam splitter, the maximum QM violation of GWI given by Eq.(\ref{equgwi}) is 0.20711, which is larger than the maximum QM violations of GWI given by Eq.(\ref{equgwi}) for bipartite qutrit isotropic and singlet states. Its corresponding eigenvector is a non-maximally entangled state of two qutrits, which has the following form
\begin{equation}
\label{mvgw}
| \psi_{gwi} \rangle = -0.35| 00 \rangle + 0.35| 01 \rangle + 0.09| 02 \rangle + 0.35| 10 \rangle -0.35 | 11 \rangle - 0.09 | 12 \rangle + 0.09 | 20 \rangle - 0.09 | 21 \rangle + 0.70 | 22 \rangle
\end{equation}
Therefore, the threshold visibility of GWI given by Eq.(\ref{equgwi}) for the state given by Eq.(\ref{mvgw}) is 0.682.
The maximum QM violation of CGLMP inequality given by Eq.(\ref{cglmp}) is 0.9149 \cite{tt2}, which is a bit larger than the maximum QM violations of CGLMP inequality for bipartite qutrit isotropic and much larger than that of qutrit singlet states. Its corresponding eigenvector is a non-maximally entangled state of two qutrits, which has the following form
\begin{equation}
\label{mvcg}
| \psi_{c} \rangle = \frac{1}{\sqrt{2 + (0.792)^2}} ( | 00 \rangle + (0.792) |11 \rangle + |22 \rangle)
\end{equation}
The threshold visibility of CGLMP inequality given by Eq.(\ref{cglmp}) for the state given by Eq.(\ref{mvcg}) is 0.686.
Hence the maximum threshold visibility, contingent upon using six-port beam splitter, of GWI given by Eq.(\ref{equgwi}) corresponding to a non-maximally entangled state is smaller than that of CGLMP inequality.
\section{Comparison of GWI with the CGLMP inequality for bipartite qutrits contingent upon using spin-$1$ component observables}
Now, again in order to show the efficacy of the derived GWI, contingent upon using spin-$1$ component observables, we will now make a comparative analysis of the QM violation obtained through GWI with that obtained by CGLMP inequality. We perform the required comparison by inserting white noise to the pure isotropic (mentioned in Eq.(\ref{state1})) and singlet (mentioned in Eq.(\ref{state2})) states.
Before computing the effect of white noise incorporated in the states considered, let us first obtain the maximum QM violations of CGLMP inequality for the bipartite qutrit isotropic and singlet states respectively contingent upon using spin-$1$ component observables. If LHS of the CGLMP inequality (\ref{cglmp}) is evaluated in terms of the four trichotomic observables $a^1, a^2, b^1, b^2$ denoting spin-$1$ components in arbitrary directions for the isotropic state, then (from numerical procedure based on the downhill simplex method \cite{nm}) it can be shown that for the choice of measurement settings ($\theta_1^a,\phi_1^a;\theta_2^a,\phi_2^a;\theta_1^b,\phi_1^b;\theta_2^b,\phi_2^b$) = (0.45, 6.28; 1.35, 6.28; 0, 1.07; 0.90, 6.28) (in radian), CGLMP inequality is maximally violated and the magnitude of the maximum violation is given by 0.52951. On the other hand, it can be shown that if the measurements of trichotomic observables $a^1, a^2, b^1, b^2$ denoting spin-$1$ components in arbitrary directions are performed on $3\otimes 3$-dimensional singlet state given by, Eq.(\ref{state2}), then (from numerical procedure based on the downhill simplex method \cite{nm}) it can be shown that the maximum QM violation of the CGLMP inequality is 0.52951. This occurs for the choice of measurement settings ($\theta_1^a,\phi_1^a;\theta_2^a,\phi_2^a;\theta_1^b,\phi_1^b;\theta_2^b,\phi_2^b$) = (0.78, 5.72; 0.88, 4.47; 2.12, 3.08; 2.41, 1.92) (in radian).
In order to probe the efficacy of the derived GWI, we now compare the tolerances of GWI against white noise present in a state with that of CGLMP inequality, contingent upon using spin-$1$ component observables, in a similar way described in Section VI. In the Table (\ref{tab5}), the threshold visibilities of GWI and CGLMP inequality pertaining to the bipartite qutrit isotropic and singlet states respectively are shown.
The above mentioned Table clearly shows that the GWI given by Eq.(\ref{equgwi}) is more robust than the CGLMP inequality for the persistence of the QM violation in the presence of white noise incorporated in both the qutrit isotropic and singlet states.
\begin{center}
\begin{table}
\begin{tabular}{|*{3}{c|}}
\hline
{\textit{\textbf{State}}} & \multicolumn{2}{|c|}{\textit{\textbf{Threshold visibility of}}}\\
\cline{2-3}
& \textit{\textbf{GWI given by Eq.(\ref{equgwi})}} & \textit{\textbf{CGLMP inequality}} \\
\hline
\hline
Isotropic & $0.786$ & $0.791$ \\
Singlet & $0.786$ & $0.791$ \\
\hline
\end{tabular}
\caption{Threshold visibilities of GWI and CGLMP inequality for the bipartite qutrit isotropic state and singlet state using spin-$1$ component observables.} \label{tab5}
\end{table}
\end{center}
\section{Maximal violations of GWI and CGLMP inequality contingent upon using spin-$1$ component observables}
As discussed in Section VII, the maximum QM violations of GWI and CGLMP inequality using spin-$1$ component observables can be evaluated using the Min-Max theorem as stated before. Here any typical joint probability, say, $P(a^1=+, b^1=-)$ of obtaining outcomes $+$ and $-$ respectively, when the observable $a^1$ is measured on the first particle and the observable $b^1$ in measured on the second particle, and the initial state is $|\psi' \rangle \in \mathbb{C}^{n}$, is given by,
\begin{equation}
P(a^1=+, b^1=-) =\langle \psi' |( \vert + \rangle_{(\theta_{1},\phi_{1})} \langle + \vert \otimes \vert - \rangle_{(\theta_{3},\phi_{3})} \langle - \vert) | \psi' \rangle
\end{equation}
Similarly, evaluating other joint probabilities, the LHS of any local realist inequality can be expressed for the initial state $|\psi' \rangle \in \mathbb{C}^{n}$ as $\langle \psi' | B' | \psi' \rangle$, where $B'$ is the Bell operator associated with the respective local realist inequality for bipartite qutrits corresponding to using spin-$1$ component observables. $B'$ is a $3 \otimes 3$ Hermitian Matrix. The largest eigenvalue of $B'$ will be the maximum QM violation of the corresponding local realist inequality using spin-$1$ component observables and using the aforementioned Min-Max theorem, one can find the largest eigenvalues of the Bell operators associated with different local realist inequalities for bipartite qutrits.
We have found that the maximum QM violation of GWI given by Eq.(\ref{equgwi}) using spin-$1$ component observables is 0.20711, which is larger than the maximum QM violations of GWI given by Eq.(\ref{equgwi}) for bipartite qutrit isotropic and singlet states. Its corresponding eigenvector is a non-maximally entangled state of two qutrits, which has the following form
\begin{equation}
\label{statemvg}
| \psi'_{gwi} \rangle = -0.01| 00 \rangle - 0.01 | 01 \rangle + 0.67 | 02 \rangle - 0.18 | 10 \rangle - 0.40 | 11 \rangle - 0.19 | 12 \rangle + 0.23 | 20 \rangle + 0.51 | 21 \rangle - 0.13 | 22 \rangle
\end{equation}
Therefore, the threshold visibility of GWI given by Eq.(\ref{equgwi}) for the state given by Eq.(\ref{statemvg}) is 0.682.
The maximum QM violation of CGLMP inequality given by Eq.(\ref{cglmp}) using spin-$1$ component observables is 0.62877, which is a bit larger than the maximum QM violations of CGLMP inequality for bipartite qutrit isotropic and singlet states. Its corresponding eigenvector is a non-maximally entangled state of two qutrits, which has the following form
\begin{equation}
| \psi'_{c} \rangle = (0.51 - 0.15 i)| 00 \rangle - (0.22 + 0.28 i)| 01 \rangle + (0.05 + 0.13 i)| 02 \rangle -(0.28 + 0.12 i)| 10 \rangle - (0.11 + 0.31 i)| 11 \rangle \nonumber
\end{equation}
\begin{equation}
\label{statemvc}
- (0.12 - 0.23 i)| 12 \rangle + 0.21 i | 20 \rangle - (0.17 - 0.22 i)| 21 \rangle - 0.42 | 22 \rangle
\end{equation}
The threshold visibility of CGLMP inequality given by Eq.(\ref{cglmp}) for the state given by Eq.(\ref{statemvc}) is 0.761.
Hence the maximum threshold visibility, contingent upon using spin-$1$ component observables, of GWI given by Eq.(\ref{equgwi}) corresponding to a non-maximally entangled state is much smaller than that of CGLMP inequality.
To summarize, while the maximum QM violation of GWI occurs for the state given by Eq.(\ref{statemvg}), maximum QM violation of CGLMP inequality occurs for the state given by Eq.(\ref{statemvc}) when one uses spin-$1$ component observables. Note that none of the states is a maximally entangled state. Interestingly, the maximum QM violations of GWI given by Eq.(\ref{equgwi}) are the same whether one uses spin-$1$ component observables or six-port beam splitter. But the maximum QM violations of CGLMP inequality given by Eq.(\ref{cglmp}) differ for the two different types of observables stated above.
\section{QM violation of GWI by mixed bipartite qutrit states}
\begin{figure}[t]
\centering
\begin{minipage}{.45\textwidth}
\centering
\begin{framed}
\includegraphics[width=7.5cm,height=6cm]{plot_sixport_plot2.pdf}
\captionof{figure}{Red plot represents the $W_{Q}^{max}$ versus $p$ for $\rho_1$ and $\rho_2$ using six-port beam splitter. Blue plot represents the $W_{Q}^{max}$ versus $p$ for $\rho_3$ using six-port beam splitter.}
\label{fig1}
\end{framed}
\end{minipage}%
\begin{minipage}{.45\textwidth}
\begin{framed}
\centering
\includegraphics[width=7.5cm,height=5.6cm]{spin-1_plot2.pdf}
\captionof{figure}{Red plot represents the $W_{Q}^{max}$ versus $p$ for $\rho_1$ and $\rho_2$ using spin-$1$ component observables. Blue plot represents the $W_{Q}^{max}$ versus $p$ for $\rho_3$ using spin-$1$ component observables.}
\label{fig2}
\end{framed}
\end{minipage}
\end{figure}
It is well known that all pure bipartite entangled states violate local realist inequalities \cite{gisin}. This is, however, not true for mixed states. Hence, the relation between entanglement and QM violations of local realist inequalities for mixed states is not clear. The connection between entanglement and mixedness of the state, and the amount of QM violation of different local realist inequalities by that state is another area of interest. There are instances demonstrating that to produce an equal amount of Bell-CHSH (Bell-Clauser-Horne-Shimony-Holt) violation some states require more entanglement than others. In Ref. \cite{mixed1}, it was suggested that with the increase in mixedness of bipratite qubit state, higher degree of entanglement is required for it to violate the Bell-CHSH inequality. However, there are counter-examples showing the existence of states with equal amount of Bell-CHSH violation and entanglement, but one of them is more mixed than the other. The reason as to why equal amount of Bell-CHSH violation requires different amounts of entanglement cannot be explained by mixedness alone \cite{mixed2}. For a class of bipartite qubit mixed states it has been shown that it is not possible to discriminate between states violating or not violating Bell-CHSH inequalities, knowing only their entanglement and mixedness \cite{mixed3}.
In this Section we have investigated the QM violations of GWI by four different classes of mixed bipartite qutrit states which are given by
\begin{equation}
\rho_1 = p \vert \psi_1 \rangle \langle \psi_1 \vert + (1-p) \frac{\mathbb{I}_3\otimes \mathbb{I}_3}{3^2}
\end{equation}
where $\vert\psi_1\rangle = \frac{\vert 00 \rangle + \vert 11 \rangle + \vert 22 \rangle}{\sqrt{3}}$ is the bipartite qutrit pure isotropic states; $0 \leq p \leq 1$.
\begin{equation}
\rho_2 = p \vert \psi_2 \rangle \langle \psi_2 \vert + (1-p) \frac{\mathbb{I}_3\otimes \mathbb{I}_3}{3^2}
\end{equation}
where $\vert\psi_2\rangle = \frac{\vert 02 \rangle - \vert 11 \rangle + \vert 20 \rangle}{\sqrt{3}}$ is the bipartite qutrit pure singlet states; $0 \leq p \leq 1$.
\begin{equation}
\rho_3 = p \vert \psi_1 \rangle \langle \psi_1 \vert + (1-p) \vert \psi_2 \rangle \langle \psi_2 \vert
\end{equation}
where $\vert\psi_1\rangle$ and $\vert\psi_2\rangle$ are the bipartite qutrit pure isotropic state and singlet state respectively; $0 \leq p \leq 1$.
\begin{equation}
\rho_4 = p\Big( q \vert \psi_1 \rangle \langle \psi_1 \vert + (1-q) \vert \psi_2 \rangle \langle \psi_2 \vert \Big) + (1-p) \frac{\mathbb{I}_3\otimes \mathbb{I}_3}{3^2}
\end{equation}
where $\vert\psi_1\rangle$ and $\vert\psi_2\rangle$ are the bipartite qutrit pure isotropic state and singlet state respectively; $0 \leq p \leq 1$; $0 \leq q \leq 1$.
Varying over measurement settings, we have calculated numerically (based on the downhill simplex method \cite{nm}) the maximum QM violations of GWI given by Eq.(\ref{equgwi}) using six-port beam splitter and by using spin-$1$ component observables for different values of the state parameters. In Fig.(\ref{fig1}) and Fig.(\ref{fig2}) we have plotted the maximum values of the left hand side of GWI given by Eq.(\ref{equgwi}) ($W^{max}_{Q}$) for the states $\rho_1$, $\rho_2$ and $\rho_3$ for different values of the state parameter $p$ using six-port beam splitter and using spin-$1$ component observables respectively. In Fig.(\ref{fig3}) and Fig.(\ref{fig4}) we have plotted the maximum values of the left hand side of GWI given by Eq.(\ref{equgwi}) ($W^{max}_{Q}$) for the states $\rho_4$ for $q=0.3$ and $q=0.7$ for different values of the state parameter $p$ using six-port beam splitter and using spin-$1$ component observables respectively. For other values of $q$ the plot is similar, but we have not shown them in the Figures.
\begin{figure}[t]
\centering
\begin{minipage}{.45\textwidth}
\centering
\begin{framed}
\includegraphics[width=7.5cm,height=6cm]{plot_sixport_plot1.pdf}
\captionof{figure}{Red and blue plot represent the $W_{Q}^{max}$ versus $p$ for $\rho_4$ with $q=0.7$ and with $q=0.3$, respectively, using six-port beam splitter.}
\label{fig3}
\end{framed}
\end{minipage}%
\begin{minipage}{.45\textwidth}
\begin{framed}
\centering
\includegraphics[width=7.5cm,height=5.6cm]{spin-1_plot1.pdf}
\captionof{figure}{Red and blue plot represent the $W_{Q}^{max}$ versus $p$ for $\rho_4$ with $q=0.7$ and with $q=0.3$, respectively, using spin-$1$ component observables..}
\label{fig4}
\end{framed}
\end{minipage}
\end{figure}
Fig.(\ref{fig1}), Fig.(\ref{fig2}), Fig.(\ref{fig3}) and Fig.(\ref{fig4}) show that each of $\rho_1$, $\rho_2$ and $\rho_4$ (with $q=0.3$ and $q=0.7$) violates the GWI given by Eq.(\ref{equgwi}) above a particular value of $p$ using six-port beam splitter as well as using spin-$1$ component observable. Above this particular value of $p$, the magnitude of QM violation of GWI by each of these states increases linearly with increasing values of $p$. Moreover, the magnitudes of QM violations of GWI for $\rho_1$ and $\rho_2$ are the same for any value of $p$.
Fig.(\ref{fig1}) indicates that the state $\rho_3$ violates the GWI given by Eq.(\ref{equgwi}) using six-port beam splitter for any values of $p$. However, the magnitude of the QM violation of GWI by the state $\rho_3$ is not a linear function of $p$. Fig.(\ref{fig2}) indicates that the state $\rho_3$ violates the GWI given by Eq.(\ref{equgwi}) using spin-$1$ component observable in some particular range of $p$. In this case, the magnitude of the QM violation of GWI by the state $\rho_3$ is not a linear function of $p$.
\section{Conclusion}
In this work we have extended Wigner's approach \cite{Wigner} by deriving generalized Wigner type local realist inequalities (GWI) for bipartite qutrit systems based on the assumption of the existence of the overall joint probability distributions in the underlying stochastic HV space for the measurement outcomes pertaining to the relevant trichotomic observables, that satisfy the locality condition, and yield the measurable marginal probabilities. An important point to stress here is that the expressions of GWIs that have been derived here do \textit{not} reduce to that of Bell-CHSH inequalities by grouping any two outcomes. This feature distinguishes GWIs considered here from the other local realist inequalities for bipartite qutrits; for example, the one suggested by Wu et. al. \cite{wu}. Also, note that the factorizability condition for hidden variables used in deriving Bell-CHSH inequalities is not required for deriving GWI. In this context, it should be mentioned that the role of factorizability condition for stochastic hidden variables has been subjected to a critical examination \cite{Alex}. Our work based on GWI serves to validate the notion that assuming the existence of overall joint probabilities in any stochastic HV theory yielding the measurable marginal probabilities is sufficient to demonstrate for the bipartite qutrit systems an incompatibility between QM and a class of stochastic HV theories satisfying the locality condition.
Efficacy of the derived GWI for the bipartite qutrit systems has been probed by analysing the robustness of its QM violation against white noise incorporated in the states considered. A comparative study of GWI in these contexts has been performed with respect to CGLMP inequality \cite{cglmp}, using the two widely used maximally entangled states, {\it viz.} the singlet state, and the isotropic state which are also considered to be relevant in quantum information processing. Using six-port beam splitter, we have found that CGLMP inequality has lower threshold visibility compared to GWI when white noise is incorporated in the qutrit isotropic states. On the other hand, GWI has lower threshold visibility compared to CGLMP inequality when white noise is introduced in qutrit singlet states. In fact, CGLMP inequality is not violated by QM for qutrit singlet states when six-port beam splitter is used. It is also found that, contingent upon using spin-$1$ component observables, GWI has a lower threshold visibility than CGLMP inequality when white noise is introduced in both isotropic and singlet states. We can, therefore, state that for showing QM incompatibility of qutrit \textit{isotropic} states with local realism using six-port beam splitter, the CGLMP inequality is more efficient in terms of the robustness of its QM violation against white noise than GWI; whereas, for qutrit \textit{singlet} states, GWI is more efficient than the CGLMP inequality in showing QM incompatibility with local realism using six-port beam splitter. On the other hand, if one uses spin-$1$ component observables, GWI is more efficient in showing QM incompatibility with local realism for \textit{both} qutrit isotropic and singlet states in the presence of white noise in these states, compared to the CGLMP inequality.
Another significant result obtained in this paper is that, for both six-port beam splitter and spin-$1$ component observables, the maximum QM violations of GWI and the CGLMP inequality occur for non-maximally entangled states. Further, it is found that the maximum QM violation of GWI is the same whether one uses spin-$1$ component observables or the observables pertaining to the six-port beam splitter. On the other hand, the maximum QM violation of the CGLMP inequality differs for the two different types of observables stated above.
QM violations of GWI for different classes of mixed bipartite qutrit states using the two aforementioned types of observables have also been addressed.
It requires to be studied what interesting results such comparison between GWI and the CGLMP inequality would yield when extended in the context of non maximally entangled bipartite qutrit pure and mixed states. Finally, it should be worth probing the possibility of any information theoretic application of GWI similar to that of the more familiar Bell-CHSH inequalities, for example, in the context of device independent quantum key generation \cite{diqkd}, and for developing robust multipartite multilevel quantum protocols \cite{ns}.
\begin{acknowledgements}
DD acknowledges the financial support from University Grants Commission (UGC), Govt. of India. DH, ASM and SG acknowledge support from the project SR/S2/LOP-08/2013 of DST, India. The research of DH is also supported by the Center for Science, Kolkata. SD acknowledges support from DST-INSPIRE fellowship.
\end{acknowledgements}
|
\section{Introduction} \label{sec1}
Most recent advancements in the waveform tomography of discontinuity surfaces reside in the context of acoustic and electromagnetic inverse scattering. Spurred by the early study in~\cite{Kress1995}, such developments include: i) the Factorization Method (FM)~\cite{Bouk2013,Cham2014}; ii) the Linear Sampling Method (LSM)~\cite{Has2013, Fiora2003} and MUSIC algorithms~\cite{Park2009, Park2015(2)}; iii) the subspace migration technique~\cite{Park2015}, and iv) the method of Topological Sensitivity (TS)~\cite{Guz2004, Bonnet2011, Park2013}. In general, the LSM and FM techniques are applicable to a wide class of interfacial conditions and inherently carry a superior localization property -- potentially leading to high-fidelity geometric reconstruction. These methods, however, may suffer from the sensitivity to measurement uncertainties. In contrast the TS approach, that is inherently robust to noisy data, fails to adequately recover the shape of a scatterer at long illuminating wavelengths. The subspace migration methods offer another alternative for a high-fidelity reconstruction, even from partial-aperture data, while requiring some a priori knowledge about the geometry of a discontinuity surface. Among the aforementioned methods, the LSM has been applied to the problem of elastic-wave imaging of fractures with homogeneous (traction-free) boundary condition~\cite{Bour2013}, while the TS approach was recently extended to cater for qualitative elastodynamic sensing of fractures endowed with a more general class of contact laws~\cite{Bellis2013, Fatemeh2015}. In geophysics, major strides~\cite{Willis2006,Zheng2013,Minato2013,Minato2014,Fang2014} have been made toward a robust reconstruction of fractures via seismic waveform tomography. So far the proposed methods, often reliant upon a rudimentary parameterization of the fracture geometry (e.g. planar fractures) and nonlinear minimization, entail a number of impediments including: i) high computational cost; ii) sensitivity to the assumed parametrization; iii) computational instabilities~\cite{Minato2014}, and iv) major restrictions in terms of the seismic sensing configuration~\cite{Fang2014,Minato2013}, namely the location of sources and receivers relative to the (planar) fracture surface. One recent study aiming to mitigate such limitations can be found in~\cite{Zheng2013} that makes use of focused Gaussian beams emitted from the surface source/receiver arrays to non-iteratively assess the orientation, spacing, and compliance of systems of parallel planar fractures.
This work aims to develop a non-iterative, full-waveform approach to 3D elastic-wave imaging of fractures with non-trivial (generally heterogeneous and dissipative) interfacial condition. To this end, the sought indicator map -- targeting \emph{geometric} fracture reconstruction -- is preferably (i) agnostic with respect to the fracture's interfacial condition, (ii) robust against measurement errors, and (iii) flexible in terms of sensing parameters, e.g. the illumination frequency. This is pursued by drawing from the theories of inverse scattering~\cite{Fiora2008, Col1992} and, in particular, by building upon the Factorization Method~\cite{Kirsch2008, Bouk2013} and the recently developed Generalized Linear Sampling Method (GLSM)~\cite{Audibert2015, Audibert2014} which completes the theoretical foundation of its LSM~predecessor. First, the inverse problem is formulated in the frequency domain where the illuminating wavefield is described by the elastic Herglotz wave function~\cite{Dassios1995} with its inherent compressional (P) and shear (S) wave components. On characterizing the induced scattered wavefield in terms of its far-field P- and S-wave patterns~\cite{Martin1993}, the far-field operator~$F$ is then defined as a map from the Herglotz densities to the far-field measurements. In this setting, the GLSM indicator functional is introduced as in~\cite{Audibert2014} on the basis of (i) a custom factorization of the far-field operator, and (b) a sequence of approximate solutions to the LSM integral equation, seeking Herglotz densities whose far-field pattern matches that of a point-load solution radiating from the sampling point. The latter sequence is essentially a set of penalized least-squares misfit functionals -- aimed at producing nearby solutions to the LSM equation, where the penalty term is constructed using a factorization component of~$F$. Minimizing this class of cost functionals in their most general form requires an optimization procedure~\cite{Audibert2014}. Thanks to the premise of a linear contact law, however, this study takes advantage of the so-called $F_\sharp$-factorization~\cite{Kirsch2008, Bouk2013} of the far-field operator to formulate the penalty term. This results in a sequence of \emph{convex} GLSM cost functionals whose minimizers can be computed without iterations.
\section{Problem statement}\label{PS}
With reference to Fig.~\ref{fig1}(a), consider the elastic-wave sensing of a partially closed fracture $\Gamma \subset \mathbb{R}^3$ embedded in a homogeneous, isotropic, elastic solid endowed with mass density~$\rho$ and Lam\'{e} parameters $\mu$ and~$\lambda$. The fracture is characterized by a heterogeneous contact condition synthesizing the spatially-varying nature of its rough and/or multi-phase interface. Next, let $\Omega$ denote the unit sphere centered at the origin. For a given triplet of vectors $\bd\in\Omega$ and~$\bq_p,\bq_s\!\in\mathbb{R}^3$ such that $\bq_p\!\parallel\bd$ and~$\bq_s\!\perp\!\bd$, the obstacle is illuminated by a combination of compressional and shear plane waves
\begin{equation}\label{plwa}
\bu^{\text{\tiny f}}(\bxi) ~=~ \bq_p \exs e^{\textrm{i} k_p \bxi \cdot \bd} \:+\: \bq_s \exs e^{\textrm{i} k_s \bxi \cdot \bd}
\end{equation}
propagating in direction~$\bd$, where $k_p$ and $k_s=k_p\sqrt{(\lambda\!+\!2\mu)/\mu}$ denote the respective wave numbers. The interaction of $\bu^{\text{\tiny f}}$ with $\Gamma$ gives rise to the scattered field $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3\backslash\Gamma)^3$, solving
\begin{equation}\label{GE}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \bv) \,+\, \rho \exs \omega^2\bv ~=~ \boldsymbol{0} \quad &\text{in}& \quad {\mathbb{R}^3}\backslash\Gamma, \\*[1mm]
&\bn \cdot \bC \exs \colon \! \nabla \bv~=~ \mathcal{L}(\dbv) \,-\, \bt^{\text{\tiny f}} \quad &\text{on}& \quad \Gamma,
\end{aligned}
\end{equation}
where $\omega^2=k_s^2 \mu/\rho$ is the frequency of excitation; $\dbv=[\bv^+\!-\bv^-]$ is the jump in~$\bv$ across~$\Gamma$, hereon referred to as the fracture opening displacement \textcolor{black}{(FOD)};
\begin{equation}\label{bC}
\bC \:=\: \lambda\,\bI_2\!\otimes\bI_2 \:+\: 2\mu\,\bI_4
\end{equation}
is the fourth-order elasticity tensor; $\bI_m \,(m\!=\!2,4)$ denotes the $m$th-order symmetric identity tensor; \mbox{$\bt^{\text{\tiny f}} = \bn \cdot \bC \colon \! \nabla \bu^{\text{\tiny f}}$} is the free-field traction vector; $\bn = \bn^-$ is the unit normal on~$\Gamma$, and $\mathcal{L}: H^{1/2}(\Gamma)^3\to H^{-1/2}(\Gamma)^3$ represents a heterogeneous bijective contact law over the fracture surface, physically relating the displacement jump to surface traction. In many practical situations, the fracture's contact law is \emph{linearized} about a dynamic equilibrium state as
\begin{equation}
\label{contact} \mathcal{L}(\dbv) \:=\: \bK(\bxi) \dbv, \qquad \bxi \in \Gamma,
\end{equation}
where $\bK=\bK(\bxi)$ is a \emph{symmetric} (due to reciprocity considerations) and possibly \emph{complex-valued} matrix of specific stiffness coefficients.
\begin{figure}[tp]
\center\includegraphics[width=0.94\linewidth]{figures/fig1_v7.jpg} \vspace*{0mm}
\caption{Direct scattering problem. The fracture boundary $\Gamma$ is arbitrarily extended to a piecewise smooth, simply connected, closed surface $\partial D$ of a bounded domain~$D$.} \label{fig1}
\end{figure}
\begin{rem}
\textcolor{black}{In what follows, the analysis is based on the linear contact condition~(\ref{contact}) over~$\Gamma$. Under the premise of bijectivity, most of the ensuing developments (except for the $F_\sharp$ factorization method) can be adapted to handle nonlinear contact laws; such extension, however, is beyond the scope of this study.}
\end{rem}
The formulation of the direct scattering problem can now be completed by requiring that~$\bv$ satisfies the Kupradze radiation condition at infinity \cite{Kuprad1979}. On uniquely decomposing the scattered field into an irrotational part and a solenoidal part as $\bv = \bv^p + \bv^{s}$ \textcolor{black}{where
\begin{equation}\label{vpvs}
\bv^p = \frac{1}{k_s^2\!-\!k_p^2}(\Delta+k_s^2)\bv, \qquad \bv^{s} = \frac{1}{k_p^2\!-\!k_s^2}(\Delta+k_p^2)\bv,
\end{equation}}
the Kupradze condition can be stated as
\begin{equation}\label{KS}
\frac{\partial\bv^p}{\partial r} - \text{i} k_p \bv^p = o\big(r^{-1}\big) \quad \mbox{ and } \quad
\frac{\partial\bv^s}{\partial r} - \text{i} k_s \bv^ s = o\big(r^{-1}\big) \qquad \text{as} ~~r:=|\bxi|\to\infty,
\end{equation}
uniformly with respect to $\hat\bxi:=\bxi/r$.
\paragraph*{Dimensional platform.}
In what follows, all quantities are rendered \emph{dimensionless} by taking $\rho$, $\mu$, and ${\sf R}$ -- the characteristic size of a region sampled for fractures -- as the respective scales for mass density, elastic modulus, and length -- which amounts to setting $\rho = \mu = {\sf R} = 1$~\cite{Scaling2003}.
\paragraph*{Function spaces.}
To assist the ensuing analysis, the fracture surface $\Gamma$ is arbitrarily extended, as shown in Fig.~\ref{fig1}(b), to a piecewise smooth, simply connected, closed surface $\partial D$ of a bounded domain $D$ such that the normal vector $\bn$ to the fracture surface $\Gamma$ coincides with the outward normal vector to $\partial D$ -- likewise denoted by $\bn$. We also assume that $\Gamma$ is an open set (relative to $\partial D$) with positive surface measure.
Following \cite{McLean2000}, we define
\begin{equation}\label{funS}
\begin{aligned}
&H^{\pm \frac{1}{2}}(\Gamma) ~:=~\big\lbrace f\big|_\Gamma \colon \,\,\, f \in H^{\pm \frac{1}{2}}(\partial D) \big\rbrace, \\*[1 mm]
& \tilde{H}^{\pm \frac{1}{2}}(\Gamma) ~:=~\big\lbrace f \in H^{\pm\frac{1}{2}}(\partial D) \colon \,\,\, \text{supp}(f) \subset \overline{\Gamma} \big\rbrace,
\end{aligned}
\end{equation}
and recall that $H^{-1/2}(\Gamma)$ and $\tilde{H}^{-1/2}(\Gamma)$ are respectively the dual spaces of $\tilde{H}^{1/2}(\Gamma)$ and $H^{1/2}(\Gamma)$. Accordingly, the following embeddings hold
\begin{equation}\label{embb}
\tilde{H}^{\frac{1}{2}}(\Gamma) \,\subset\, H^{\frac{1}{2}}(\Gamma) \,\subset\, L^2(\Gamma) \,\subset\, \tilde{H}^{-\frac{1}{2}}(\Gamma) \,\subset\, H^{-\frac{1}{2}}(\Gamma).
\end{equation}
\begin{rem} \textcolor{black}{In the context of fracture mechanics, it is well known that $\dbv(\bxi)\to \boldsymbol{0}$ continuously as $\Gamma\!\ni\!\bxi\to\partial\Gamma$ (typically as $d^\alpha$, $0\!<\!\alpha\!\leqslant\!\tfrac{1}{2}$~\cite{Ueda2006} where $d$ is a normal distance to $\partial\Gamma$ when $\partial\Gamma$ is smooth), which lends credence to the assumption $\dbv\in\tilde{H}^{1/2}(\Gamma)^3$ used hereon.}
\end{rem}
\section{On the well-posedness of the forward scattering problem} \label{WP}
Serving as a prerequisite for the analysis of the inverse scattering problem, this section investigates the well-posedness of the direct scattering problem \eqref{GE}--\eqref{KS}. Let $R>0$ be sufficiently large so that the ball $B_R$ of radius $R$ contains $\Gamma$, and consider the Dirichlet-to-Neumann operator ${\mathcal T}_R: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3$ associated with the scattering problem in $\mathbb{R}^3 \backslash B_R$, namely
$$
{\mathcal T}_R({\boldsymbol{\varphi}})(\bxi) := \hat{\bxi}\cdot \bC \colon \! \nabla \btu_{{\boldsymbol{\varphi}}}(\bxi), \qquad \bxi\in\partial B_R,
$$
where $\btu_{\boldsymbol{\varphi}} \in H^1_\mathrm{loc}(\mathbb{R}^3 \backslash B_R)^3$ is the unique radiating solution, satisfying~\eqref{KS}, of
\begin{equation}\label{TR}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \btu_{\boldsymbol{\varphi}}) \,+\, \rho \exs \omega^2 \btu_{\boldsymbol{\varphi}} ~=~ \boldsymbol{0} \quad &\text{in}& \quad {\mathbb{R}^3}\backslash B_R, \\*[1mm]
& \btu_{\boldsymbol{\varphi}}~=~{\boldsymbol{\varphi}} \quad &\text{on}& \quad \partial B_R.
\end{aligned}
\end{equation}
The scattering problem \eqref{GE}--\eqref{KS} can now be equivalently written in terms of $\bv\in H^1(B_R\backslash\Gamma)^3$ as
\begin{equation}\label{GEbis}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \bv) \,+\, \rho \exs \omega^2\bv ~=~ \boldsymbol{0} \quad &\text{in}& \quad {\mathbb{R}^3}\backslash\Gamma, \\*[1mm]
&\bn \cdot \bC \exs \colon \! \nabla \bv ~=~ \bK\sip\dbv \,-\, \bt^{\text{\tiny f}} \quad &\text{on}& \quad \Gamma, \\*[1mm]
&\bn \cdot \bC \colon \! \nabla \bv = {\mathcal T}_R(\bv) \quad &\text{on}& \quad \partial B_R,
\end{aligned}
\end{equation}
where $\bn(\bxi)=\hat\bxi$ on~$\partial B_R$. This problem can be written variationally in terms of $\bv\in H^1(\mathcal{B}_2\backslash\Gamma)^3$ as
\begin{equation}\label{Wik-GE}
\begin{aligned}
&- \rho \exs \omega^2 \! \int_{\mathcal{B}_2 \backslash \Gamma} \overline\bw \cdot \bv \,\exs \textrm{d}V_{\bxi} ~+\, \int_{\mathcal{B}_2 \backslash \Gamma} \nabla \exs \overline{\bw} \colon \bC \colon \nabla \bv \, \textrm{d}V_{\bxi} ~+\, \dualGA{\bK \sip \dbv}{\llbracket\bw\rrbracket}~-\, \\*[1 mm]
& \quad\,\, \dualBR{{\mathcal T}_R(\bv)}{\bw} ~=\, \int_{\Gamma} \exs \overline{\llbracket \bw \rrbracket} \cdot \bt^{\text{\tiny f}} \, \textrm{d}S_{\bxi}, \qquad \quad \forall \exs \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\end{aligned}
\end{equation}
where $\dualGA{\cdot}{\cdot}$ and $\dualBR{\cdot}{\cdot}$ respectively denote the $\big\langle H^{-1/2}(\Gamma)^3, \tilde{H}^{1/2}(\Gamma)^3 \big\rangle$ and $\big\langle H^{-1/2}(\partial B_R)^3, {H}^{1/2}(\partial B_R)^3 \big\rangle$ \emph{duality products} that extend $L^2$ inner products. The analysis of the forward scattering problem is based on the following properties of the Dirichlet-to-Neumann operator ${\mathcal T}_R$ (see also \cite{BramblePasciak}). \textcolor{black}{For clarity, we will use an abbreviated notation of relevant vector norms where e.g. $\|\boldsymbol{\cdot}\|_{H^{1/2}(\Gamma)^3}$ is denoted by~ $\|\boldsymbol{\cdot}\|_{H^{1/2}(\Gamma)}$ and so on.}
\begin{lemma} \label{LemmaTR}
There exists a bounded, non-negative and self-adjoint operator ${\mathcal T}_R^0\!: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3$ such that ${\mathcal T}_R+{\mathcal T}_R^0\!: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3 $ is compact. Moreover,
\begin{equation}
\Im \dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\varphi}}} > 0 \qquad \forall {\boldsymbol{\varphi}} \in H^{1/2}(\partial B_R)^3:~ {\boldsymbol{\varphi}} \neq 0.
\end{equation}
\end{lemma}
\begin{proof} Let $R_{\circ} > R$ and ${\boldsymbol{\varphi}}, {\boldsymbol{\psi}} \in H^{1/2}(\partial B_R)^3 $. Multiplying the first equation in~\eqref{TR} by $\overline{\btu_{\boldsymbol{\psi}}}$ and
integrating by parts on $B_{R_{\circ}} \backslash B_R$ yields
$$
\dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\psi}}} ~=~ \rho\exs\omega^2 \! \int_{B_{R_{\circ}}\!\backslash B_R} \overline{\btu_{\boldsymbol{\psi}}}\cdot \btu_{\boldsymbol{\varphi}} \,\exs \textrm{d}V_{\bxi} ~- \int_{B_{R_{\circ}}\!\backslash B_R} \nabla \exs \overline{\btu_{\boldsymbol{\psi}}} \colon \bC \colon \nabla \btu_{\boldsymbol{\varphi}}
\, \textrm{d}V_{\bxi} ~+ \int_{\partial{B_{R_{\circ}}}} \exs\overline{\btu_{\boldsymbol{\psi}}} \cdot \bt({\boldsymbol{\varphi}}) \, \textrm{d}S_{\bxi},
$$
where $\bt({\boldsymbol{\varphi}})(\bxi):= \hat{\bxi}\cdot \bC \colon \! \nabla \btu_{{\boldsymbol{\varphi}}}(\bxi)$ for $\bxi \in \partial B_{R_{\circ}}$. Using the well-posedness of \eqref{TR} and the Riesz representation theorem, we define ${\mathcal T}_R^0$ by
$$
\dualBR{{\mathcal T}_R^0({\boldsymbol{\varphi}})}{{\boldsymbol{\psi}}} := \int_{B_{R_{\circ}}\!\backslash B_R} \nabla\exs\overline{\btu_{\boldsymbol{\psi}}} \colon \bC \colon \nabla \btu_{\boldsymbol{\varphi}}
\, \textrm{d}V_{\bxi}.
$$
\textcolor{black}{On demonstrating that $\|({\mathcal T}_R+{\mathcal T}_R^0)({\boldsymbol{\varphi}})\|_{H^{-1/2}(\Gamma)}\leqslant C (\|\btu_{{\boldsymbol{\varphi}}}\|_{L^2(B_{R_{\circ}}\!\backslash B_R)}+\|\bt({\boldsymbol{\varphi}})\|_{L^2(\partial B_{R_{\circ}}\!)})$ for some constant~$C\!>\!0$ independent of~${\boldsymbol{\varphi}}$, the compactness of ${\mathcal T}_R+{\mathcal T}_R^0$ then follows from the compactness of mapping ${\boldsymbol{\varphi}}\to\btu_{\boldsymbol{\varphi}}$ (resp. ${\boldsymbol{\varphi}}\to\bt({\boldsymbol{\varphi}})$) from $H^{1/2}(\partial B_R)$ into~$L^2(B_{R_{\circ}}\!\backslash B_R)$ (resp.~$L^2(\partial B_{R_{\circ}})$) thanks to the compact embedding of $H^1(B_{R_{\circ}}\!\backslash B_R)$ into~$L^2(B_{R_{\circ}}\!\backslash B_R)$ and the standard regularity results for scattering problems~\cite{McLean2000}, which can be recovered from the boundary integral representation of $\btu_{\boldsymbol{\varphi}}$ in $\mathbb{R}^3\backslash B_R$ in terms of boundary data on~$\partial B_R$. As shown in~\ref{lem1-p}, the sign of the imaginary part of ${\mathcal T}_R$ is a consequence of the asymptotic behavior of $\btu_{\boldsymbol{\varphi}}$ at infinity~\cite{Kuprad1979} which implies}
\begin{equation}\label{kup1}
\Im \dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\varphi}}} = \Im \lim_{R_{\circ} \to \infty} \int_{\partial{B_{R_{\circ}}}} \exs\overline{\btu_{\boldsymbol{\varphi}}} \cdot \bt({\boldsymbol{\varphi}}) \,
\textrm{d}S_{\bxi} \;=\; \lim_{R_{\circ} \to \infty} \int_{\partial{B_{R_{\circ}}}} \Big\lbrace k_p (\lambda+2\mu) |\btu_{\boldsymbol{\varphi}}^{\mathrm{p}} \exs |^2 \,+\, k_s\mu\exs |\btu_{\boldsymbol{\varphi}}^{\mathrm{s}}|^2 \Big\rbrace \,\, \textrm{d}S_{\bxi}.
\end{equation}
The \textcolor{black}{sign-definiteness} of the imaginary part is a consequence of the Rellich lemma \cite{Col1992} applied to $\btu_{\boldsymbol{\varphi}}^{\mathrm{p}}$ and $\btu_{\boldsymbol{\varphi}}^{\mathrm{s}}$, which requires that $\btu_{\boldsymbol{\varphi}} = \btu_{\boldsymbol{\varphi}}^{\mathrm{p}} + \btu_{\boldsymbol{\varphi}}^{\mathrm{s}}=0$ whenever $\Im \dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\varphi}}}=0$.
\end{proof}
\begin{theorem} \label{maindirect}
Assume that $\bt^{\text{\tiny f}} \in H^{-1/2}(\Gamma)^3$ and that $\bK \in L^\infty(\Gamma)^{3\times 3}$ is symmetric such that $\Im\bK\leqslant\boldsymbol{0}$ on $\Gamma$, i.e. that $\overline{\boldsymbol{\theta}} \sip \Im \bK(\bxi)\sip \boldsymbol{\theta} \leqslant 0$, $\forall\boldsymbol{\theta}\in\mathbb{C}^3$ and a.e.~on~$\Gamma$. Then problem~\eqref{Wik-GE} has a unique solution that continuously depends on $\bt^{\text{\tiny f}} \in H^{-1/2}(\Gamma)^3$.
\end{theorem}
\begin{proof}
Since $\bt^{\text{\tiny f}} \in H^{-1/2}(\Gamma)$, the antilinear form $\int_{\Gamma} \exs \overline{\llbracket \bw \rrbracket} \cdot \bt^{\text{\tiny f}} \, \textrm{d}S_{\bxi}$ may be understood as a duality pairing $\dualGA{\cdot}{\cdot}$. The continuity of this form comes from the continuity of the trace mapping $\bw \to \llbracket \bw \rrbracket$ from $ H^1(\mathcal{B}_2 \backslash \Gamma)^3$ into $\tilde{H}^{1/2}(\Gamma)^3$.
On the basis of the adopted dimensional platform i.e.~$\rho = \mu = 1$ (see Section~\ref{PS}), the sesquilinear form on the left hand side of \eqref{Wik-GE} can be decomposed into a coercive part
\begin{equation}\label{cv}
\begin{aligned}
&\textrm{A}(\bv,\bw) ~=\, \! \int_{\mathcal{B}_2 \backslash \Gamma} \overline\bw \cdot \bv \,\exs \textrm{d}V_{\bxi} ~+\, \int_{\mathcal{B}_2 \backslash \Gamma} \!\! \nabla \exs \overline{\bw} \colon \! \bC \colon \! \nabla \bv \, \textrm{d}V_{\bxi} ~+\, \dualBR{{\mathcal T}_R^0(\bv)}{\bw} , \quad \forall \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\end{aligned}
\end{equation}
and a compact part
\begin{equation}\label{compact}
\text{B}(\bv,\bw) ~=\, - (1+k_s^2) \int_{\mathcal{B}_2 \backslash \Gamma} \overline\bw
\cdot \bv \,\, \textrm{d}V_{\bxi} ~+\, \dualGA{\bK \sip \dbv}{\llbracket\bw\rrbracket}~ -\,
\dualBR{({\mathcal T}_R+{\mathcal T}_R^0)(\bv)}{\bw} , \quad \forall \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3.
\end{equation}
The coercivity of $\textrm{A}(\bv,\bw)$ follows from the Korn inequality~\cite{McLean2000} and the non negative sign of ${\mathcal T}_R^0$ (Lemma \ref{LemmaTR}). Now, in order to prove that the antilinear form $\text{B}$ defines a compact perturbation of $\textrm{A}(\bv,\bw)$, one may observe that
\[
| \text{B}(\bv,\bw) | ~\leqslant~\! \text{c}_2 \exs \big\lbrace \nxs
\norms{\!\bv\!}_{L^2(\mathcal{B}_2 \backslash \Gamma)} \, \norms{\!\bw\!}_{L^2(\mathcal{B}_2
\backslash \Gamma)} \,+\, \norms{\! \llbracket \bv \rrbracket
\!}_{L^2(\Gamma)} \, \norms{\! \llbracket \bw \rrbracket \!}_{L^2(\Gamma)}
\! \big\rbrace + \norms{({\mathcal T}_R+{\mathcal T}_R^0)(\bv)}_{H^{-1/2}(\partial B_R)}\norms{\bw}_{H^{1/2}(\partial B_R)}
\]
for a constant $\text{c}_2$ independent of $\bv$ and $\bw$. The claim then follows from Lemma \ref{LemmaTR}, the compact embedding of $H^1(\mathcal{B}_2 \backslash \Gamma)$ into $L^2(\mathcal{B}_2 \backslash \Gamma)$ and the compactness of the trace operator $\bv \rightarrow \llbracket \bv \rrbracket$ as an application from $H^1(\mathcal{B}_2 \backslash \Gamma)$ into ${L^2(\Gamma)}$ where the latter comes from the compact embedding of $\tilde H^{1/2}(\Gamma)$ into $L^2(\Gamma)$.
Problem \eqref{Wik-GE} is then of Fredholm type, and is therefore well-posed as soon as the uniqueness of a solution is guaranteed. Assume that $\bt^{\text{\tiny f}} = 0$. Then
$$
\Im \dualBR{{\mathcal T}_R(\bv)}{\bv} =\, \dualGA{\Im \bK \sip \dbv}{\llbracket\bv\rrbracket} \,\leqslant\: 0
$$
by premise of the Theorem. Thanks to Lemma \ref{LemmaTR}, this requires that $\bv\!=\!\boldsymbol{0}$ on $\partial B_R$ and thus $\bv\!=\!\boldsymbol{0}$ in $\mathcal{B}_2 \backslash \Gamma$ by the unique continuation principle.
\end{proof}
\section{Elements of the inverse scattering solution}\label{Prelim}
\renewcommand{\Omega_{\bd}}{{\Omega}}
\renewcommand{\boldsymbol{u}^{\textit{i}}}{\btu}
This section is devoted to the introduction of the \emph{far-field operator} -- relevant to the scattering problem~(\ref{GE}), and the derivation of its first and second factorizations. In the sequel, we assume that the hypotheses of Theorem \ref{maindirect} hold.
\paragraph*{Elastic Herglotz wave function.} For given density $\bg \in L^2(\Omega_{\bd})^3$, we consider the unique decomposition
\begin{equation}\label{herden}
\bg \;:=\; \bg_p \oplus\, \bg_s
\end{equation}
such that $\bg_p(\bd)\!\parallel\!\bd$ and $\,\bg_s(\bd)\!\perp\!\bd$, $\,\bd\in\Omega_{\bd}$. In dyadic notation, one has
\begin{equation}\label{freef}
\bg_p(\bd) := (\bd \nxs \otimes \nxs \bd \exs) \cdot \bg(\bd) \quad ~~\text{and}~~ \quad \bg_s(\bd) := (\bI - \bd \nxs \otimes \nxs \bd \exs) \cdot \bg(\bd).
\end{equation}
Next, we define the elastic Herglotz wave function~\cite{Dassios1995} as
\begin{equation}\label{HW}
\boldsymbol{u}^{\textit{i}}_\bg(\bxi) ~: =~ \int_{\Omega_{\bd}} \bg_p(\bd) \exs e^{\textrm{i} k_p \bd \cdot \bxi} \,\, \text{d}S_{\bd} \,+\, \int_{\Omega_{\bd}} \bg_s(\bd) \exs e^{\textrm{i} k_s \bd \cdot \bxi} \,\, \text{d}S_{\bd}, \qquad \bxi \in \mathbb{R}^3
\end{equation}
in terms of {the compressional and shear wave densities~$\bg_p$ and~$\bg_s$.
\paragraph*{The far-field pattern.}
As shown in~\cite{Martin1993}, any scattered wave $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ solving \eqref{GE}-\eqref{KS} has the asymptotic
expansion
\begin{equation}\label{vinf}
\bv(\bxi) ~=~ \exs \frac{e^{\text{i}k_p r}\!}{4 \pi(\lambda\!+\!2\mu)r} \exs \bv^\infty_p(\hat\bxi) \:+\:
\frac{e^{\text{i}k_s r}\!}{4\pi\mu r} \exs \bv^\infty_s(\hat\bxi) \:+\: O(r^{-2}) \quad~~ \text{as} \quad r:=|\bxi|\to\infty,
\end{equation}
where $\hat\bxi$ is the unit direction of observation, while $\bv^\infty_p$ and $\bv^\infty_s$ denote respectively the far-field patterns of $\bv^p$ and $\bv^s$ -- see~\eqref{vpvs}, which satisfy $\bv^\infty_p\!\parallel\hat\bxi$ and $\bv^\infty_s\!\perp\hat\bxi$. In this setting, we define the far-field pattern of~$\bv$ by
\begin{equation}\label{far-field}
\bv^\infty := \bv^\infty_p \oplus \bv^\infty_s.
\end{equation}
By way of the integral representation theorem in elastodynamics~\cite{Bon1999} and the far-field representation of the elastodynamic fundamental stress tensor (see Appendix), one can show that if $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ satisfies~\eqref{GE}-\eqref{KS}, then
\begin{equation}\label{vinf2}
\begin{aligned}
&\bv_p^\infty(\hat\bxi) ~=~- \text{i} k_p \exs \hat\bxi \int_\Gamma \Big\lbrace \lambda \, \llbracket \bv \rrbracket \sip \bn + 2\mu \big(\bn \sip \hat\bxi \big) \exs \llbracket \bv \rrbracket \sip \hat\bxi \exs \Big\rbrace \, e^{-\text{i}k_p \hat\bxi \cdot \bx} \,\, \text{d}S_{\bx}, \\*[1 mm]
& \bv_s^\infty(\hat\bxi) ~=~ -\text{i}k_s \exs \hat\bxi \exs \times \int_{\Gamma} \Big\lbrace \mu \big( \llbracket \bv \rrbracket \!\times\! \hat\bxi \exs \big)(\bn \sip \hat\bxi \exs ) \,+\, \mu \big( \bn \!\times \hat\bxi \big) (\llbracket \bv \rrbracket \sip \hat\bxi) \Big\rbrace \, e^{-\text{i}k_s \hat\bxi \cdot \bx} \,\, \text{d}S_{\bx}.
\end{aligned}
\end{equation}
\paragraph*{The far-field operator.}
\begin{defn}\label{deffarf}
We define the far-field operator $F: L^2(\Omega_{\bd})^3 \to L^2(\Omega_{\bd})^3$ by
\begin{equation}\label{ffo0}
F(\bg) ~=~ \textcolor{black}{\bv_{\bg_\Omega}^\infty,}
\end{equation}
where~$\bv_{\bg_\Omega}^\infty$ is the far-field pattern~\eqref{far-field} of~$\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ solving~\eqref{GE}-\eqref{KS} with data $\bu^{\text{\tiny f}} = \boldsymbol{u}^{\textit{i}}_\bg$, see~\eqref{HW}.
\end{defn}
When the contact law specified by~$\mathcal{L}(\dbv)$ is linear as in~(\ref{contact}), the far-field operator can be expressed as a linear integral operator. To examine this case, \textcolor{black}{consider an incident plane wave~(\ref{plwa}) propagating in direction $\bd\in\Omega_{\bd}$ with amplitude $\bq=\bq_p\oplus\bq_s$}, and denote the induced far-field pattern~(\ref{far-field}) by \textcolor{black}{$\bv^\infty_\bq(\bd, \cdot)=\bv^\infty_{\bq_p}\oplus \bv^\infty_{\bq_s}$}. Next, let us define the far-field kernel \textcolor{black}{$\bW^\infty(\bd,\hat\bxi)\in\mathbb{C}^{6\times 6}$} so that
\begin{equation}\label{w-inf}
\bW^\infty(\bd,\hat\bxi) \sip \bq ~:=~ \bv^\infty_\bq(\bd, \hat\bxi).
\end{equation}
Then one easily verifies that
\begin{equation}\label{ffo2}
F(\bg)\textcolor{black}{(\hat\bxi)} ~=\, \int_{\Omega_{\bd}} \bW^\infty(\bd,\hat\bxi) \sip \bg(\bd) \,\, \text{d}S_{\bd}.
\end{equation}
\begin{lemma}\label{recip}
The far-field kernel $\bW^\infty(\bd,\hat\bxi)$ satisfies the reciprocity identity
\begin{equation}\label{W-recip}
\textcolor{black}{
\bW^\infty(\bd,\hat\bxi) ~=~ \overline{\bW^{\infty *}}(-\hat\bxi,-\bd), \qquad \forall\bd,\hat\bxi\!\in\Omega.}
\end{equation}
\end{lemma}
\begin{proof}
\textcolor{black}{See~\ref{Recip}}.
\end{proof}
\section{Key properties for the application of sampling methods}\label{SFS}
\paragraph*{Factorization of the far-field operator $F$.}
Consider the Herglotz operator $\mathcal{H} \colon L^2(\Omega_{\bd})^3 \rightarrow H^{-1/2}(\Gamma)^3$ given by
\begin{equation}\label{oH}
\mathcal{H}(\bg) ~:=~ \bn\cdot\bC\exs\colon\!\nabla\boldsymbol{u}^{\textit{i}}_\bg \quad~~ \text{on}\quad \Gamma,
\end{equation}
where $\boldsymbol{u}^{\textit{i}}_\bg$ is the Herglotz wave function~(\ref{HW}). Next, define $\mathcal{G} \colon H^{-1/2}(\Gamma)^3 \rightarrow L^2(\Omega_{\bd})^3$ as the map taking the traction vector $\bt^{\text{\tiny f}}$ over $\Gamma$ to the far-field pattern, $\bv^\infty$, of $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ satisfying \eqref{GE}-\eqref{KS}. Then from Definition~\ref{deffarf}, the far-field operator~\eqref{ffo0} becomes
\begin{equation}\label{fac1}
F ~=~ \mathcal{G} \mathcal{H}.
\end{equation}
\begin{lemma}\label{H*}
With reference to decomposition~\eqref{far-field}, the adjoint Herglotz operator $\mathcal{H}^* \colon \tilde{H}^{1/2}(\Gamma)^3 \rightarrow L^2(\Omega_{\bd})^3$ takes the form
\begin{equation}\label{Hstar}
\begin{aligned}
\mathcal{H}^*\nxs(\ba)(\hat\bxi) ~=~ - \Big( \, & \text{\emph{i}} k_p \, \hat\bxi \exs \int_\Gamma \, \big\lbrace \lambda \exs (\ba \sip \bn) \,+\, 2\mu \exs (\bn \sip \hat\bxi) ( \ba \sip \hat\bxi) \big\rbrace \, e^{-\text{\emph{i}}k_p \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \\*[1 mm]
& \textcolor{black}{\oplus}~ \text{\emph{i}} k_s \, \hat\bxi\times\! \int_\Gamma \big\lbrace \mu \exs(\ba \times \hat\bxi)(\bn\sip\hat\bxi) \,+\, \mu \exs (\bn \times \hat\bxi) (\ba \sip \hat\bxi) \big\rbrace \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \Big).
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
\textcolor{black}{see~\ref{H*pruf}}.
\end{proof}
On the basis of~\eqref{vinf2} and~(\ref{Hstar}), map $\mathcal{G}$ can be further decomposed as $\mathcal{G}=\mathcal{H}^* T$ where the middle operator $T\colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ is given by
\begin{equation}\label{T}
T(\bt^{\text{\tiny f}})(\bxi) ~:=~ \llbracket \bv(\bxi) \rrbracket, \qquad \bxi \in \Gamma
\end{equation}
such that~$\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ satisfies~\eqref{GE}-\eqref{KS} or equivalently \eqref{Wik-GE}. Thanks to this new decomposition of $\mathcal{G}$, the second factorization of $F \colon L^2(\Omega_{\bd})^3 \rightarrow L^2(\Omega_{\bd})^3$ is obtained
\begin{equation}\label{fact}
F ~=~ \mathcal{H}^* \exs T \exs \mathcal{H},
\end{equation}
which provides the second important ingredient for the ensuing analysis.
\paragraph*{Properties of the Herglotz operator $\mathcal{H}$.}
\begin{lemma}\label{H*p}
Operator $\mathcal{H^*}:\tilde{H}^{1/2}(\Gamma)^3\rightarrow L^2(\Omega_{\bd})^3$ in Lemma~\ref{H*} is compact and injective.
\end{lemma}
\begin{proof}
Integral operator $\mathcal{H}^*$ has a smooth kernel and is therefore compact from $\tilde{H}^{1/2}(\Gamma)^3$ into $L^2(\Omega_{\bd})^3$. Next, suppose that there exists $\ba \in \tilde{H}^{1/2}(\Gamma)^3$ such that $\mathcal{H^*}(\ba) = \boldsymbol{0}$. In light of~(\ref{vinf}) and~(\ref{vinf2}), it is apparent that $\mathcal{H^*}$ is nothing else but the far-field operator stemming from the double-layer potential
\begin{equation}\label{Dlp}
\bV(\ba)(\bxi) ~=~ \int_{\Gamma} \ba(\by) \cdot \bfT(\bxi,\by) \, \text{d}S_{\by}, \qquad
\bfT(\bxi,\by) ~=~ \bn(\by)\cdot\bSig(\bxi,\by), \qquad \bxi \in \mathcal{B}_2 \backslash \Gamma,
\end{equation}
where
$\bSig(\bxi,\by)$ is the (third-order) elastodynamic fundamental stress tensor \textcolor{black}{given in~\ref{stress-fund}}. By virtue of definition~(\ref{vinf}), vanishing far-field pattern of $\bV(\ba)$ implies, by the Rellich Lemma and the unique continuation principle, that $\bV(\ba) = \boldsymbol{0}$ in $\mathbb{R}^3 \backslash \Gamma$. Owing to the fundamental jump property of double-layer potentials by which~$\llbracket \bV \rrbracket=\ba$, one obtains $\ba = \boldsymbol{0}$ which guarantees the injectivity of $\mathcal{H}^*$.
\end{proof}
One additional property that is needed for the analysis of sampling methods is the densness of the range of~$\mathcal{H}^*$, which is equivalent to the injectivity of $\mathcal{H}$. Unfortunately the latter cannot be guaranteed in general, and one has to impose this property as an assumption on $\Gamma$ and $\omega$.
\begin{assumption}\label{Inject-H}
We assume that $\Gamma$ and $\omega$ are such that the Herglotz operator $\mathcal{H}\!:L^2(\Omega_{\bd})^3\rightarrow H^{-1/2}(\Gamma)^3$ is injective, i.e. that $\,\mathcal{H}^*\!:\tilde{H}^{1/2}(\Gamma)^3\rightarrow L^2(\Omega_{\bd})^3$ has a dense range.
\end{assumption}
The following lemma indicates why we expect that for a given fracture geometry $\Gamma$, Assumption \ref{Inject-H} holds for all $\omega\!>\!0$ \textcolor{black}{possibly} excluding a discrete set of values without finite accumulation points.
\begin{lemma}\label{Dense-H}
Assume that \textcolor{black}{$\exs\Gamma$ contains $M\!\geqslant\!1$ (possibly disjoint) analytic surfaces~$\Gamma_m\!\subset\Gamma$, $m=1,\ldots M$, and consider the unique analytic continuation $\partial D_m$ of $\:\Gamma_m$ identifying ``interior'' domain~$D_m\!\subset\mathbb{R}^3$}. Then Assumption~\ref{Inject-H} holds as soon as \textcolor{black}{for any such~$m$, $\omega\!>\!0$} is not a ``Neumann'' eigenfrequency of the Navier equation in $D_m$, i.e. as long as every function $\boldsymbol{u}^{\textit{i}} \in H^1(D_m)^3$ satisfying
\begin{equation}\label{uiH}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \boldsymbol{u}^{\textit{i}}) \,+\, \rho \exs \omega^2 \boldsymbol{u}^{\textit{i}} ~=~ \boldsymbol{0} \quad &\textrm{in}~ D_m, \\*[1 mm]
&\bn \sip \bC \colon \! \nxs \nabla \boldsymbol{u}^{\textit{i}} ~=~ \boldsymbol{0} \quad &\textrm{on}~ \partial D_m
\end{aligned}
\end{equation}
\textcolor{black}{vanishes identically in $D_m$. Further if~$D_m$ is bounded, the real eigenfrequencies of~\eqref{uiH} form a discrete set.}
\end{lemma}
\begin{proof}
\textcolor{black}{Let~$\Gamma_m$ denote the $m$th analytic piece of~$\Gamma$. Recalling~\eqref{HW} and invoking the analyticity of $\bn\cdot\bC\colon \!\nabla \boldsymbol{u}^{\textit{i}}_\bg$ with respect to the surface coordinates on $\partial D_m$, we deduce that if $\bn \sip \bC\colon \!\nabla \boldsymbol{u}^{\textit{i}}_\bg=\boldsymbol{0}$ on $\Gamma_m\subset\partial D_m$} then
$$
\bn \cdot \bC \exs \colon \!\nabla\boldsymbol{u}^{\textit{i}}_\bg =\boldsymbol{0} \quad \mbox{ on}~ \partial D_m.
$$
This means that $\boldsymbol{u}^{\textit{i}}_\bg =\boldsymbol{0}$ in $D_m$ since $\omega$ is not a ``Neumann'' eigenvalue of the Navier equation in $D_m$. The unique continuation principle then implies that $\boldsymbol{u}^{\textit{i}}_\bg\!=\!\boldsymbol{0}$ in $\mathbb{R}^3$. Accordingly, we deduce that the Herglotz density vanishes, i.e. that~$\bg\!=\!\boldsymbol{0}$ as in the scalar case \cite{Col1992}. \textcolor{black}{The proof of discreteness of the set of real eigenfrequencies characterizing~\eqref{uiH} when~$D_m$ is bounded can be found in~\cite{Kuprad1979}, Chapter~7, Theorem~1.4}.
\end{proof}
\paragraph*{Properties of the middle operator $T$.}
\begin{lemma}\label{I{T}>0}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ in~(\ref{T}) is bounded and satisfies
\begin{equation}\label{pos-IT}
\Im \dualGA{{\boldsymbol{\varphi}}}{T{\boldsymbol{\varphi}}} <0\, \qquad \forall \exs {\boldsymbol{\varphi}} \in H^{-1/2}(\Gamma)^3:~ {\boldsymbol{\varphi}} \neq \boldsymbol{0}.
\end{equation}
\end{lemma}
\begin{proof}
The boundedness of $T$ stems from the well-posedness of problem~\eqref{Wik-GE} and classical trace theorems. Next, let ${\boldsymbol{\varphi}} \in H^{-1/2}(\Gamma)^3$ and consider $\bv$ satisfying \eqref{Wik-GE} with $\bt^{\text{\tiny f}}= {\boldsymbol{\varphi}}$. Taking $\bw = \bv$ in \eqref{Wik-GE} we get
\begin{equation}\label{T-bound}
\Im \dualGA{{\boldsymbol{\varphi}}}{T{\boldsymbol{\varphi}}} ~=~ \dualGA{\Im \bK \sip \dbv}{\llbracket\bv\rrbracket}-\Im \dualBR{{\mathcal T}_R(\bv)}{\bv}.
\end{equation}
By virtue of~\eqref{T-bound}, the claim of the theorem follows immediately from Lemma~\ref{LemmaTR} and earlier hypothesis that $\Im\bK<0$.
\end{proof}
\begin{lemma}\label{Cp-Cr}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow\tilde{H}^{1/2}(\Gamma)^3$ can be decomposed into a compact part ${\sf T}_{\!c}$ and a coercive and self-adjoint part ${\sf T}_{\!\text{\tiny o}}$ such that $T = {\sf T}_{\! c} + {\sf T}_{\!\text{\tiny o}}$. The coercive part ${\sf T}_{\!\text{\tiny o}} \colon H^{-1/2}(\Gamma)^3 \rightarrow\tilde{H}^{1/2}(\Gamma)^3$ is defined by
\begin{equation}\label{To}
{\sf T}_{\!\text{\tiny o}}({\boldsymbol{\varphi}}) ~:=~ \llbracket \bu^{\text{\tiny o}}\rrbracket \quad \text{on}~~ \Gamma,
\end{equation}
where $\bu^{\text{\tiny o}} \in H^1(\mathcal{B}_2 \backslash \Gamma)$ is a solution to
\begin{equation}\label{Wuo}
\textrm{A}(\bu^{\text{\tiny o}},\bw) = \dualGA{{\boldsymbol{\varphi}}}{\llbracket\bw \rrbracket} \quad \forall \exs \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\end{equation}
$\textrm{A}$ being the coercive sesquilinear form defined by \eqref{cv}.
\end{lemma}
\begin{proof}
We first observe from~\eqref{cv} that
\begin{equation}\label{uoG}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \bu^{\text{\tiny o}}) \,-\, \bu^{\text{\tiny o}} ~=~ \boldsymbol{0} \qquad &\text{in} \quad \mathcal{B}_2 \backslash \Gamma,& \\
& \bn \cdot \bC \colon \! \nxs \nabla \bu^{\text{\tiny o}} ~=~ \! -{\boldsymbol{\varphi}} \qquad &\text{on} \quad \Gamma,& \\
&\bn \cdot \bC \colon \! \nxs \nabla\bu^{\text{\tiny o}} ~=~ \textcolor{black}{\mathcal{S}_R(\bu^{\text{\tiny o}})} \qquad &\text{on} \quad \partial{B}_R&,
\end{aligned}
\end{equation}
\textcolor{black}{where $\mathcal{S}_R\!: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3$ is a Dirichlet-to-Neumann operator, $\mathcal{S}_R({\boldsymbol{\psi}}):= \bn\cdot\bC\colon\!\nabla \btu_{{\boldsymbol{\psi}}}$, stemming from the \emph{elastostatic problem} in $B_{R_\circ}\!\backslash B_R$ with Dirichlet data $\btu_{{\boldsymbol{\psi}}}={\boldsymbol{\psi}}$ on~$\partial B_{R}$ and homogeneous ``Neumann'' data $\bn\cdot\bC\colon\!\nabla \btu_{{\boldsymbol{\psi}}}=\boldsymbol{0}$ on~$\partial B_{R_\circ}$.}
Using standard trace theorems for vector fields with square-integrable divergence~\cite{Monk2003}, one finds that
\begin{equation}\label{ToTr}
\norms{\nxs{\boldsymbol{\varphi}} \nxs}_{H^{-\frac{1}{2}}(\Gamma)}\! \:=\: \norms{\nxs\bn\cdot\bC\colon\!\!\nabla\bu^{\text{\tiny o}}\!}_{H^{-\frac{1}{2}}(\Gamma)}
\,\,\leqslant\: \norms{\nxs\bn\cdot\bC\colon\!\!\nabla\bu^{\text{\tiny o}}\!}_{H^{-\frac{1}{2}}(\partial D)}
\,\,\leqslant\: c \left(\nxs\norms{\!\nabla\sip(\bC\colon\!\!\nabla \bu^{\text{\tiny o}})\!}_{L^2(D)} +
\norms{\!(\bC\colon\!\!\nabla\bu^{\text{\tiny o}})\!}_{L^2(D)} \right)
\end{equation}
for a positive constant $c$ independent from $\bu^{\text{\tiny o}}$. Thanks to the first equation in \eqref{uoG} we then deduce
$$
\norms{\nxs{\boldsymbol{\varphi}} \nxs}_{H^{-\frac{1}{2}}(\Gamma)} ~\leqslant~ c_1 \norms{\bu^{\text{\tiny o}} \!}_{H^1(D)}
$$
for some $c_1\!>\!0$ independent from $\bu^{\text{\tiny o}}$. On taking $\bw=\bu^{\text{\tiny o}}$ in \eqref{Wuo}, deploying the coercivity of $A$, \textcolor{black}{and recalling from~\eqref{cv} that $\Im A(\bv,\bv)=0$}, we find
\begin{equation}\label{ToCo}
\dualGA{{\boldsymbol{\varphi}}}{{\sf T}_{\!\text{\tiny o}}{\boldsymbol{\varphi}}} = \textrm{A}(\bu^{\text{\tiny o}},\bu^{\text{\tiny o}}) \geqslant c_2 \norms{\nxs{\boldsymbol{\varphi}} \nxs}_{H^{-\frac{1}{2}}(\Gamma)}^2
\end{equation}
for a positive constant $c_2$ independent from ${\boldsymbol{\varphi}}$, which establishes the coercivity of ${\sf T}_{\!\text{\tiny o}}$. The self-adjointness of~${\sf T}_{\!\text{\tiny o}}$ follows immediately from that of~$\textrm{A}$.
To complete the argument, consider the compactness of ${\sf T}_c \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$, given by
\[
{\sf T}_c({\boldsymbol{\varphi}}) \:=\: \llbracket\bv^c\rrbracket, \quad \bv^c \:=\: \bv - \bu^{\text{\tiny o}} \quad \text{on}~ \Gamma
\]
where $\bv$ solves~\eqref{Wik-GE}. On subtracting~(\ref{Wuo}) from~(\ref{Wik-GE}) with~$\bt^{\text{\tiny f}}={\boldsymbol{\varphi}}$, one finds that
\[
\textrm{A}(\bv^c, \bw) \:= - \textrm{B}(\bv,\bw) \qquad \forall \exs \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\]
where $\textrm{A}$ is coercive while $\text{B}$, given by~(\ref{compact}), is compact on $H^1(\mathcal{B}_2 \backslash \Gamma)^3$. As a result, the induced mapping $\bv \rightarrow \bv^c$ from $H^1(\mathcal{B}_2 \backslash \Gamma)^3$ into $H^1(\mathcal{B}_2 \backslash \Gamma)^3$ is \emph{compact}, whereby the compactness of ${\sf T}_c$ follows directly from the continuity of $\bv \in H^1(\mathcal{B}_2 \backslash \Gamma)^3$ with respect to ${\boldsymbol{\varphi}} \in H^{-\frac{1}{2}}(\Gamma)^3$ and the trace theorem.
\end{proof}
\textcolor{black}{
\begin{lemma}\label{T-invs0}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ has a bounded (and thus continuous) inverse.
\end{lemma}
\begin{proof}
The idea is to show that $T$, given by~\eqref{T}, is injective and Fredholm of index zero. The second claim follows immediately from Lemma~\ref{Cp-Cr}. To demonstrate the injectivity of~\eqref{T}, one may recall a double-layer potential representation of elastodynamic fields solving~\eqref{GE}-\eqref{KS} which demonstrates that for any ${\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)$, one has
\[
\bv({\boldsymbol{\varphi}})(\bxi) ~=~ \int_{\Gamma} T({\boldsymbol{\varphi}}) \cdot \bfT(\bxi,\by) \, \text{d}S_{\by}, \qquad
\bfT(\bxi,\by) ~=~ \bn(\by)\cdot\bSig(\bxi,\by), \qquad \bxi \in \mathbb{R}^3 \backslash \Gamma,
\]
where $\dbv=T({\boldsymbol{\varphi}})$ on~$\Gamma$ thanks to the fundamental property of double-layer potentials. Thus, on assuming that there exists ${\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)$ so that $T({\boldsymbol{\varphi}})=\boldsymbol{0}$, one finds that $\bv\!=\!\boldsymbol{0}$ in~$\mathbb{R}^3\backslash\Gamma$ and consequently, by the second of~\eqref{GE} and trace theorems, that $\|{\boldsymbol{\varphi}}\|_{H^{-1/2}(\Gamma)}=\|\bn\sip\bC\!:\!\nabla\bv\|_{H^{-1/2}(\Gamma)}=0$.
\end{proof}}
\textcolor{black}{
\begin{lemma}\label{T-coerc0}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ is coercive, i.e. there exists constant $c\!>\!0$ independent of~${\boldsymbol{\varphi}}$ such that
\begin{equation}\label{co-T0}
|\langle {\boldsymbol{\varphi}}, \, T ({\boldsymbol{\varphi}}) \rangle| \,\,\geqslant\,\, \textrm{c} \nxs \norms{{\boldsymbol{\varphi}}}_{H^{-\frac{1}{2}}(\Gamma)}^2, \qquad \forall{\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)^3.
\end{equation}
\end{lemma}
\begin{proof}
Lemma~\ref{I{T}>0} demonstrates that the duality product $\big\langle {\boldsymbol{\varphi}} ,\, T({\boldsymbol{\varphi}}) \big\rangle\in\mathbb{C}\setminus(-\infty,\infty)$ for all nonzero ${\boldsymbol{\varphi}}\!\in H^{-1/2}(\Gamma)^3$. Due to Lemma~\ref{Cp-Cr}, on the other hand, decomposition $T = {\sf T}_{\! c} + {\sf T}_{\!\text{\tiny o}}$ exists where ${\sf T}_{\! c}$ is compact and~${\sf T}_{\!\text{\tiny o}}$ is such that $\langle \exs {\boldsymbol{\varphi}}, {\sf T}_{\!\text{\tiny o}}({\boldsymbol{\varphi}}) \exs \rangle\in \mathbb{R}$ satisfies the coercivity condition~(\ref{ToCo}) $\forall \exs {\boldsymbol{\varphi}}\!\in\!H^{-1/2}(\Gamma)^3$. With such results in place, claim~\eqref{co-T} follows immediately by Lemma~1.17 in~\cite{Kirsch2008}.
\end{proof}}
\section{Application of sampling methods} \label{SSA}
\subsection{Linear sampling method (LSM)}
The essential idea behind the LSM~\cite{Fiora2003} and also the factorization method (FM)~\cite{Bouk2013} for geometrical obstacle reconstruction stems from the particular nature of an approximate solution, \textcolor{black}{$\bg =\bg_p\oplus\bg_s$}, to the far-field equation
\begin{equation}\label{FF}
F \bg ~=~ \boldsymbol{\Phi}_L^\infty, \qquad F ~=~ \mathcal{G} \mathcal{H} ~=~ \mathcal{H}^* \exs T \exs \mathcal{H},
\end{equation}
where $\boldsymbol{\Phi}_L^\infty$ is the far-field pattern \textcolor{black}{of a trial radiating field}, see Definition~\ref{phi-infinity}. In this setting, the behavior of $\bg$ in the sampling region is exposed by characterizing the range of $\mathcal{G}$ or $\mathcal{H}^*$, \textcolor{black}{which then forms the basis for approximating the characteristic function of a scatterer}. This section presents an adaptation of the key LSM results for the problem of elastic-wave imaging of heterogeneous fractures, which provides a foundation for the GLSM developments in Section~\ref{GLSMM}.
\begin{defn} \label{phi-infinity}
\textcolor{black}{With reference to~\eqref{Hstar}, for every admissible FOD profile $\ba\!\in\!\tilde{H}^{1/2}(L)$ specified over a smooth, non-intersecting trial fracture $L\!\subset\!\mathcal{B}_2$, the induced far-field pattern $\boldsymbol{\Phi}_L^\infty \colon \tilde{H}^{1/2}(L) \rightarrow L^2(\Omega_{\bd})^3$ is given by}
\begin{equation}\label{Phi-inf}
\begin{aligned}
\boldsymbol{\Phi}_{L}^\infty(\ba)(\hat\bxi) ~=~ - \Big( \, & \text{\emph{i}} k_p \, \hat\bxi \exs \int_L \, \Big\lbrace \lambda \exs (\ba \sip \bn) \,+\, 2\mu \exs (\bn \sip \hat\bxi) ( \ba \sip \hat\bxi) \Big\rbrace \, e^{-\text{\emph{i}}k_p \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \\*[1 mm]
& \textcolor{black}{\oplus}~ \text{\emph{i}} k_s \, \hat\bxi\times\! \int_L \Big\lbrace \mu \exs(\ba \times \hat\bxi)(\bn\sip\hat\bxi) \,+\, \mu \exs (\bn \times \hat\bxi) (\ba \sip \hat\bxi) \Big\rbrace \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \Big).
\end{aligned}
\end{equation}
and $\bn$ is the unit normal on~$L$.
\end{defn}
\begin{rem}\label{LSMrem}
\textcolor{black}{On the basis of Definition~\ref{phi-infinity}, one may interpret the LSM reconstruction philosophy as follows. Let ${\sf L}\!\subset\!\mathbb{R}^3$ (containing the origin) denote a reference fracture surface whose characteristic size is small relative to the length scales describing the forward scattering problem, and let $L=\bz\!+\bR{\sf L}$ where $\bz\!\in\mathbb{R}^3$ and $\bR\!\in\!U(3)$ is a unitary rotation matrix. Given an admissible FOD profile $\ba\!\in\!\tilde{H}^{1/2}({\sf L})$, solving the far-field equation~\eqref{FF} over a grid of trial pairs $(\bz,\bR)$ sampling $\mathbb{R}^3\!\times U(3)$ is simply an effort to probe the far-field kernel~\eqref{w-inf} -- through synthetic rearrangement of the illuminating plane waves -- for fingerprints in terms of~$\boldsymbol{\Phi}_L^\infty$. As shown by Theorems~\ref{TR2}, \ref{GLSM1} and~\ref{GLSM2}, such fingerprint is found in the data if and only if~$L\subset\Gamma$. Otherwise, the norm of any approximate solution to~(\ref{FF}) can be made arbitrarily large, which then provides a criterion for the reconstruction of~$\Gamma$.}
\end{rem}
\begin{theorem}\label{TR1}
Provided that $\omega$ is \emph{not} a ``Neumann'' eigenvalue of the Navier equation~(\ref{uiH}) and that~$\bK^{-1} \!\! \in \! L^{\infty}(\Gamma)$, for \emph{every} smooth and non-intersecting trial crack $L\subset\mathcal{B}_2$ and some density function $\ba(\bxi)\!\in\!\tilde{H}^{1/2}(L)$, one has
\[
\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*) ~~ \iff ~~ L \subset \Gamma.
\]
\end{theorem}
\begin{proof}
Consider the following:
\begin{itemize}
\item If~$L \subset \Gamma$, then $\tilde{H}^{1/2}(L)^3 \subset \tilde{H}^{1/2}(\Gamma)^3$. By extending the domain of $\ba\in\tilde{H}^{1/2}(L)^3$ from $L$ to $\Gamma$ through zero padding, one immediately obtains $\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*)$ thanks to~\eqref{Hstar} and~\eqref{Phi-inf}.
\item Assume that~$L\not\subset\Gamma$ and that $\boldsymbol{\Phi}_L^\infty\!\in\!Range(\mathcal{H}^*)$. Then there exists $\bb\!\in\!\tilde{H}^{1/2}(\Gamma)^3$ such that
\begin{equation} \notag
\begin{aligned}
\boldsymbol{\Phi}_L^\infty(\bb)(\hat\bxi) ~=~ -\Big(& \text{\emph{i}} k_p \, \hat\bxi \exs \int_\Gamma \, \Big\lbrace \lambda \exs (\bb \sip \bn) \,+\, 2\mu \exs (\bn \sip \hat\bxi) ( \bb \sip \hat\bxi) \Big\rbrace e^{-\text{\emph{i}}k_p \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \\
& \textcolor{black}{\oplus}~ \text{\emph{i}} k_s \, \hat\bxi \exs \times \int_{\Gamma} \Big\lbrace \mu \exs(\bb \times \hat\bxi)(\bn \sip \hat\bxi) \,+\, \mu \exs (\bn \times \hat\bxi) (\bb \sip \hat\bxi) \Big\rbrace \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \Big),
\end{aligned}
\end{equation}
associated with the layer potential
\begin{equation}\label{Dlpb}
\boldsymbol{\Phi}_{\exs \Gamma}(\bxi) ~=~ \int_{\Gamma} \bb(\by) \cdot \bfT(\bxi,\by) \, \text{d}S_{\by}, \qquad \bfT(\bxi,\by) ~=~ \bn(\by)\cdot\bSig(\bxi,\by), \qquad \bxi \in \mathcal{B}_2 \backslash \Gamma.
\end{equation}
On the other hand, owing to Definition~\ref{phi-infinity} of $\boldsymbol{\Phi}_L^\infty(\hat\bxi)$, potential $\boldsymbol{\Phi}_{\exs \Gamma}(\bxi)$ should coincide with
\begin{equation}\label{Pb}
\boldsymbol{\Psi}_{\! L}(\bxi) ~=~ \int_L \ba(\by) \cdot \bfT(\bxi,\by) \,\, \text{d}S_{\by}, \qquad \bxi \in \mathcal{B}_2 \backslash L,
\end{equation}
over $ \bxi \in \mathcal{B}_2 \backslash (L \cup \Gamma)$. Now, let $\Gamma\not\ni\bxio\!\in L$ and let $\mathcal{B}_\epsilon$ be a small ball centered at $\bxio$ such that $\mathcal{B}_\epsilon \cap \Gamma = \emptyset $. In this case $\boldsymbol{\Phi}_{\exs \Gamma}$ is analytic in $\mathcal{B}_\epsilon$, while $\boldsymbol{\Psi}_{\! L}$ has a singularity at $\bxio\!\in\mathcal{B}_\epsilon$ -- which by contradiction completes the proof.
\end{itemize}
\end{proof}
On the basis of the above result, one arrives at the following statement which inspires most of the LSM-based indicator functionals.
\begin{theorem}\label{TR2}
Under the assumptions of \textcolor{black}{Lemma~\ref{Dense-H} and~Theorem~\ref{TR1}},
\begin{itemize}
\item~If $L\!\subset\!\Gamma$, there exists a Herglotz density vector $\bg_\epsilon^L\!\in L^2(\Omega_{\bd})^3$ such that $\|F\bg_\epsilon^L-\boldsymbol{\Phi}_L^\infty\|_{L^2(\Omega_{\bd})} \leqslant\epsilon$ and $\limsup\limits_{\epsilon \rightarrow 0} \|\mathcal{H}\bg_\epsilon^L\|_{H^{-1/2}(\Gamma)}<\infty$.
\item~If $L \not\subset \Gamma$, then $\forall \bg_\epsilon^L\!\in L^2(\Omega_{\bd})^3$ such that $\norms{\nxs F\bg_\epsilon^L-\boldsymbol{\Phi}_L^\infty \nxs}_{L^2(\Omega_{\bd})} \, \leqslant\epsilon$, one has $\,\lim\limits_{\epsilon \rightarrow 0} \norms{\mathcal{H}\bg_\epsilon^L}_{H^{-1/2}(\Gamma)} \,\,=\infty$.
\end{itemize}
\end{theorem}
\begin{proof}
\begin{description}
Let us first assume $L \subset \Gamma$, whereby $\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*)$ thanks to Theorem~\ref{TR1}. Then, by definition, there exists $\ba^L\!\in\!\overline{Range(T)}$ such that $\mathcal{H}^*\ba^L = \boldsymbol{\Phi}_L^\infty$. By invoking \textcolor{black}{Lemma~\ref{T-invs0}} on the boundedness i.e.~continuity of $T^{-1}$ and Lemma~\ref{H*p} which \textcolor{black}{(by the injectivity of~$\mathcal{H}^*$)} guarantees the range denseness of $\mathcal{H}$, one finds that $\forall \epsilon>0$, $\exists \, \bg_\epsilon^L\!\in L^2(\Omega_{\bd})^3$ such that $\norms{T^{-1} \ba^L\!-\!\mathcal{H}\bg_\epsilon^L}_{H^{-1/2}(\Gamma)} \,\leqslant \epsilon$. \textcolor{black}{Thanks to (i) the continuity of~$\mathcal{H}^* T$ and (ii) the fact that~$\ba^L\!\in\tilde{H}^{1/2}(\Gamma)^3$, this establishes the first part of the claim}.
Next, consider the case where $L\!\not\subset\!\Gamma$ and consequently $\boldsymbol{\Phi}_L^\infty \not\in Range(\mathcal{H}^*)$ \textcolor{black}{by Theorem~\ref{TR1}}. Then, thanks to \textcolor{black}{Lemma~\ref{Dense-H} which implies the denseness of~$Range(\mathcal{H}^*)$}, \textcolor{black}{for every $\epsilon\!>\!0$ and some regularization parameter $0\!<\!\alpha\!<\!C\epsilon$ where~$C$ is a constant independent of~$\epsilon$}, a nearby solution $\ba^L_\epsilon \in \tilde{H}^{1/2}(\Gamma)^3$ can be built e.g.~via Tikhonov regularization~\cite{Kress1999} such that \textcolor{black}{$\norms{\!\boldsymbol{\Phi}_L^\infty-\mathcal{H}^* \ba_\epsilon^L\!}_{L^2(\Omega_{\bd})} \,\leqslant \nxs \epsilon\,$} and $\,\lim\limits_{\epsilon \rightarrow 0} \norms{\nxs \ba^L_\epsilon \nxs}_{\tilde{H}^{1/2}(\Gamma)} \,=\infty$ -- due to the compactness of $\mathcal{H}^*$ \textcolor{black}{established in Lemma~\ref{H*p}}. At this point, the same argument as in the first part of the proof -- deploying the continuity of $T^{-1}$ and the range denseness of~$\mathcal{H}$ -- can be used to show establish the second claim.
\end{description}
\end{proof}
\vspace{-7 mm}
\subsection{Factorization method (FM)}
To facilitate the ensuing developments, we recall elements of the factorization method~\cite{Kirsch2008} as they pertain to our inverse problem.
\begin{defn}\label{DFs}
The self-adjoint operator $F_\sharp \colon L^2(\Omega_{\bd})^3 \rightarrow L^2(\Omega_{\bd})^3$ is defined by
\begin{equation}\label{Fs}
F_\sharp \,\colon \!\!\!=\, |\Re{F}| \:+\: \Im{F},
\vspace{-2 mm}
\end{equation}
where $F\!:L^2(\Omega_{\bd})^3 \to L^2(\Omega_{\bd})^3$ is given by~\eqref{ffo2}, and
\begin{equation}\label{ReIm}
\Re{F} \,=\, \tfrac{1}{2} (F+F^*), \qquad \Im{F} \,=\, \tfrac{1}{2 \textrm{\emph{i}}} (F-F^*).
\end{equation}
\end{defn}
\begin{rem}\label{TsD}
In line with decomposition~(\ref{fact}) of the far-field operator $F$, there exists factorization
\begin{equation}\label{facts}
F_\sharp ~=~ \mathcal{H}^* \exs T_\sharp \exs \mathcal{H}
\end{equation}
of~\eqref{Fs}, where the middle operator $T_\sharp \colon H^{-1/2}(\Gamma)^3\rightarrow\tilde{H}^{1/2}(\Gamma)^3$ is given by
\begin{equation}\label{Tsdef}
T_\sharp :=\, \Re{T}(Q^+\!-Q^-) +\, \Im{T};
\end{equation}
$Q^+$ and $Q^-\!$ are bounded projectors such that $Q^{+}\!+Q^{-}=I$; $Q^{+}\!-Q^{-}$ is an isomorphism, and $Q^-\!$ has a finite rank. See Theorem 2.15 in~\cite{Kirsch2008} for derivation.
\end{rem}
\vspace*{-5mm}
\textcolor{black}{
\begin{theorem}\label{TR3}
Under the assumptions of Theorem~\ref{TR1}, operator $F_\sharp$ in~\eqref{Fs} has the following properties:
\begin{itemize}
\item~Operator $F_\sharp$ is positive.
\item~The ranges of~$\mathcal{H}^* \colon \tilde{H}^{1/2}(\Gamma)^3\rightarrow L^2(\Omega_{\bd})^3$ and $F_\sharp^{1/2}\colon L^2(\Omega_{\bd})^3\rightarrow L^2(\Omega_{\bd})^3$ coincide.
\item~$\boldsymbol{\Phi}_L^\infty \in Range(F_\sharp^{1/2}) ~~ \iff ~~ L \subset \Gamma$.
\end{itemize}
\end{theorem}
\begin{proof}
The first two claims follow directly from Theorem~2.15 in~\cite{Kirsch2008}, its extended version (Theorem~3.2) in~\cite{Bouk2013}, Lemma~\ref{H*p} Lemma~\ref{I{T}>0}, and Lemma~\ref{Cp-Cr}. With such result in place, the last claim is immediately established by Theorem~\ref{TR1}.
\end{proof}}
\vspace*{-5mm}
\textcolor{black}{
\begin{lemma}\label{T-invs}
Operator~$T_\sharp \colon H^{-1/2}(\Gamma)^3\rightarrow\tilde{H}^{1/2}(\Gamma)^3$ in the factorization~\eqref{facts} has the following properties:
\begin{itemize}
\item~$T_\sharp$ has a bounded (and thus continuous) inverse.
\item~$T_\sharp$ is selfadjoint and is positively coercive, i.e. there exists a constant~$c\!>\!0$ independent of~${\boldsymbol{\varphi}}$ so that
\begin{equation}\label{co-T}
\big({\boldsymbol{\varphi}}, \, T_\sharp ({\boldsymbol{\varphi}}) \big)_{H^{-\frac{1}{2}}(\Gamma)} \,\,\geqslant\,\, c \nxs \norms{{\boldsymbol{\varphi}}}_{H^{-\frac{1}{2}}(\Gamma)}^2, \qquad
\forall \,{\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)^3.
\end{equation}
\end{itemize}
\end{lemma}
\begin{proof}
See Appendix~A in~\cite{Bouk2013} and the proof of Theorem 2.15, part E in~\cite{Kirsch2008}.
\end{proof}}
\textcolor{black}{On the basis of Theorem~\ref{TR2}, one sees that $F_\sharp^{1/2}$ can be used to characterize $\Gamma$ from the far-field measurements. In what follows, it is in particular shown that the GLSM cost functionals based on $F_\sharp$ (i)~are convex, (ii)~have closed-form minimizers, and (iii)~enable fast and robust reconstruction of $\Gamma$ -- especially when the data (and thus the far-field operator) are contaminated by noise.}
\subsection{Generalized Linear Sampling Method (GLSM)} \label{GLSMM}
\noindent Theorem~\ref{TR2} of the linear sampling method poses two fundamental challenges in that:~i) the featured anomaly indicator $\norms{\!\mathcal{H}\bg_\epsilon^L\!}_{H^{-1/2}(\Gamma)}$ inherently depends on the unknown fracture support $\Gamma$, and ii) construction of the Herglotz density vector $\bg_\epsilon^L \in L^2(\Omega_{\bd})^3$ is implicit in the theorem~\cite{Audibert2014}. Conventionally, these issues are addressed by replacing $\norms{\!\mathcal{H}\bg_\epsilon^L\!}_{H^{-1/2}(\Gamma)}\!$ with $\norms{\nxs\bg_\epsilon^L\!}_{L^2(\Omega_{\bd})}$ which is, in turn, computed by way of Tikhonov regularization~\cite{Kress1999}. Such treatment, however, has proven to be particularly sensitive to perturbations in the data due to e.g. measurement errors.
To help meet the challenge, the GLSM takes advantage of the second factorization (\ref{fact}) of the far-field operator and the mathematical properties of its components to properly construct a \emph{stable} approximate solution to the far-field equation (\ref{FF}). This is accomplished through a \emph{sequence} of \textcolor{black}{penalized least-squares problems} where the principal ingredient of the penalty term is $\norms{\!\mathcal{H}\bg_\epsilon^L\!}_{H^{-1/2}(\Gamma)}$, reformulated in a computable way in terms of the far-field operator~$F$. More specifically, by invoking factorizations~(\ref{fact}) and~(\ref{facts}), one may observe that
\[
\begin{aligned}
&(\bg_\epsilon^L, \exs F\bg_\epsilon^L)_{L^2(\Omega)} \,\exs ~=~ \big \langle \mathcal{H}\bg_\epsilon^L, \, T\mathcal{H}\bg_\epsilon^L \big \rangle_{\Gamma}, \\*[1 mm]
& (\bg_\epsilon^L, \exs F_\sharp \exs \bg_\epsilon^L)_{L^2(\Omega)} ~=~ \big \langle \mathcal{H}\bg_\epsilon^L, \, T_\sharp \exs \mathcal{H}\bg_\epsilon^L \big \rangle_{\Gamma}, \qquad \forall \bg_\epsilon^L \in L^2(\Omega_{\bd})^3
\end{aligned}
\]
where~$(\cdot,\cdot)_{L^2(\Omega)}:=(\cdot,\cdot)_{L^2(\Omega)^3}$ denotes the usual~$L^2$ inner product on~$\Omega$. \textcolor{black}{Then, thanks to the coercivity of the middle operator $T$ (see Lemma~\ref{T-coerc0})}, quantity $|(\bg_\epsilon^L,F\bg_\epsilon^L)_{L^2(\Omega)}|$ -- which is computable without prior knowledge of~$\Gamma$ -- may be safely substituted for $\norms{\!\mathcal{H}\bg_\epsilon^L\!}^2_{H^{-1/2}(\Gamma)}$ \textcolor{black}{in constructing a penalty term for the GLSM cost functional. Similarly, the positive coercivity $T_\sharp$ (See Lemma~\ref{T-invs}) and factorization~\eqref{facts} of $F_\sharp$ demonstrate that $|(\bg_\epsilon^L, \, F_\sharp \exs \bg_\epsilon^L)_{L^2(\Omega_{\bd})}| = \,\, \norms{F_\sharp^{1/2}\bg_\epsilon^L}^2$ may serve as a replacement for $\norms{\!\mathcal{H}\bg_\epsilon^L\!}^2_{H^{-1/2}(\Gamma)}$}, giving birth to a \emph{convex} GLSM cost functional whose minimizer can be computed without iterations. This shines light on the GLSM approach to elastodynamic reconstruction of heterogeneous fractures, whose specificities are presented next.
\paragraph*{GLSM cost functional.} \label{cost}
\begin{itemize}
\item~\emph{Unperturbed (noise-free) operators.}~Let $\alpha\!>\!0$ be a regularization parameter, and consider the far-field pattern~$\boldsymbol{\Phi}_L^\infty\!\in L^2(\Omega_{\bd})^3$ as in Definition~\ref{phi-infinity}. Then the GLSM cost functional is defined by a sequence of penalized least-squares misfit functionals $J_\alpha(\boldsymbol{\Phi}_L^\infty;\, \cdot)\colon \, L^2(\Omega_{\bd})^3 \rightarrow \mathbb{R}$, namely
\begin{equation} \label{J-alph}
J_\alpha(\boldsymbol{\Phi}_L^\infty; \, \bg) ~ \colon \!\!\! =~ \! \norms{\nxs F\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+\,\,\exs \alpha \nxs \norms{F_\sharp^{\frac{1}{2}} \bg}^2, \qquad \bg \in L^2(\Omega_{\bd})^3,
\end{equation}
whose minimizers~$\bg^L_\alpha \in L^2(\Omega_{\bd})^3$ can be computed \emph{non-iteratively} by solving
\begin{equation} \label{min-J}
F^*(F\bg^L_\alpha \,-\, \boldsymbol{\Phi}_L^\infty) ~+~ \alpha \exs (F_\sharp^{\frac{1}{2}} )^* F_\sharp^{\frac{1}{2}} \, \bg^L_\alpha~=~ \boldsymbol{0}.
\end{equation}
For completeness, a more general form $\mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty; \, \cdot)\colon \, L^2(\Omega_{\bd})^3 \rightarrow \mathbb{R}$ of the GLSM cost functional, namely
\begin{equation} \label{fJ-alph}
\mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\, \bg) ~ \colon \!\!\! =~ \! \norms{\nxs F\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+\,\,\exs \alpha \exs | ( \bg, \exs F \bg ) | , \qquad \bg \in L^2(\Omega_{\bd})^3,
\end{equation}
is also considered. \textcolor{black}{Note that~\eqref{fJ-alph} does not demand $F_\sharp$ to be applicable (see Theorem~\ref{TR3}), and thus may cater for a wider class of contact laws, $\mathcal{L}\dbv$, over the fracture surface in~(\ref{GE})}.
\begin{rem}\label{App-sol}
In general, $\mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\,\bg)$ does not have a minimizer; however, one may define
\[
j_\alpha(\boldsymbol{\Phi}_L^\infty) ~ \colon \!\!\! = \inf\limits_{\bg \in L^2(\Omega_{\bd})^3} \! \mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty; \, \bg).
\]
Thanks to the range denseness of $F$ (see Lemma~\ref{FF_op}), one has that $j_\alpha \rightarrow 0$ as $\alpha \rightarrow 0$. Accordingly, an optimized nearby solution can be constructed by following the algorithm described in~\cite{Audibert2014}.
\end{rem}
\item~\emph{Perturbed operators.}~When the measurements are contaminated with noise (e.g.~sensing errors, fluctuations in the medium properties), one has to deal with noisy operators $F^\delta\!$ and $F_\sharp^\delta$ satisfying
\begin{equation}\label{Ns-op}
\norms{\nxs F^\delta - F \nxs} \,\,\, \leqslant \,\, \delta , \qquad \norms{\nxs F^\delta_\sharp - F_\sharp \nxs} \,\,\, \leqslant \,\, \delta ,
\end{equation}
where $\delta\!>\!0$ is a measure of perturbation in data -- independent of $F$ and $F_\sharp$. Assuming that $F^\delta\!$ and $F_\sharp^\delta$ are compact, a regularized version $J_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\,\cdot)\colon \, L^2(\Omega_{\bd})^3 \rightarrow \mathbb{R}$ of the GLSM cost functional is defined in spirit of the Tikhonov regularization method as
\begin{equation} \label{RJ-alph}
J_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\, \bg) ~ \colon \!\!\! =~ \! \norms{\nxs F^\delta\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 + \,\, \alpha \exs \big(\!\norms{\nxs(F^\delta_\sharp)^{\frac{1}{2}} \exs \bg \nxs}^2 +\,\,\exs \delta \! \norms{\nxs \bg \nxs}^2 \! \big), \qquad \bg\in L^2(\Omega_{\bd})^3.
\end{equation}
Note that $J_\alpha^\delta$ is again convex and that its minimizer $\bg^L_{\alpha,\delta} \in L^2(\Omega_{\bd})^3$ solves the linear system
\begin{equation} \label{min-RJ}
F^{\delta *}(F^\delta \bg^L_{\alpha,\delta} \,-\, \boldsymbol{\Phi}_L^\infty) ~+~ \alpha \exs \big( \exs (F_\sharp^\delta)^{\nxs\frac{1}{2}*} (F_\sharp^\delta)^{\nxs\frac{1}{2}} \, \bg^L_{\alpha,\delta} \,+\, \delta \, \bg^L_{\alpha,\delta} \exs \big) ~=~ \boldsymbol{0}.
\end{equation}
In this vein, the (regularized) cost functional affiliated with the general form~(\ref{fJ-alph}) may be recast as
\begin{equation}\label{RfJ-alph}
\mathcal{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\,\bg) ~ \colon \!\!\! =~ \! \norms{\nxs F^\delta\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+~ \alpha \exs \big( \exs | (\bg, \, F^\delta \bg ) | \,+\, \delta \! \norms{\nxs \bg \nxs}^2 \! \big), \qquad \bg\in L^2(\Omega_{\bd})^3.
\end{equation}
\begin{rem}
\textcolor{black}{In (\ref{RJ-alph}) and (\ref{RfJ-alph}), $\delta$ signifies both a measure of perturbation in $F$ and a regularization parameter that, along with $\alpha$, is designed to create a robust fracture indicator functional via a sequence of the GLSM minimizers (see the proof of Theorem~\ref{GLSM2}).}
\end{rem}
\end{itemize}
With the above definitions in place, the main GLSM theorems are based on the following lemmas.
\begin{lemma}\label{comp_G}
Operator $\mathcal{G}=\mathcal{H}^* T \colon H^{-1/2}(\Gamma)^3 \rightarrow L^2(\Omega_{\bd})^3$ is compact over $H^{-1/2}(\Gamma)^3$.
\end{lemma}
\begin{proof}
The claim follows immediately from Lemmas~\ref{H*p} and~\ref{I{T}>0} establishing, respectively, the compactness of $\mathcal{H}^*$ and the boundedness of $T$.
\end{proof}
\begin{lemma}\label{FF_op}
The far-field operator $F \colon L^2(\Omega_{\bd})^3 \rightarrow L^2(\Omega_{\bd})^3$ is injective, compact and, under the assumptions of Lemma~\ref{Dense-H}, has a dense range.
\end{lemma}
\begin{proof}
\emph{Injectivity.}~Let $F(\bg) = \boldsymbol{0}$. \textcolor{black} {Then, recalling the factorization $F = \mathcal{H}^* T \mathcal{H}$ and the injectivity of $\mathcal{H}^*$ and~$T$ (due respectively to Lemma~\ref{Dense-H} and Lemma~\ref{T-invs0}), one finds that $\mathcal{H}(\bg):= \bn\cdot\bC\colon\!\nabla\boldsymbol{u}^{\textit{i}}_\bg=\boldsymbol{0}$ on~$\Gamma$. Under the assumptions of Lemma~\ref{Dense-H}, this requires that $\boldsymbol{u}^{\textit{i}}_\bg=\boldsymbol{0}$ in~$\mathbb{R}^3$, i.e. that~$\bg=\boldsymbol{0}$ which establishes the first claim.}
\emph{Compactness.}~ The compactness of~$F$ follows immediately from the compactness of $\mathcal{H}^*$ -- and thus that of $\mathcal{H}$ (Lemma~\ref{H*p}), and the boundedness of~$T$ (Lemma~\ref{I{T}>0}).
\emph{Range densenes.}~\textcolor{black}{This claim is conveniently verified by establishing the injectivity of~$F^*$. To this end, recall~(\ref{ffo2}) and consider the $L^2$-inner product
\begin{equation}\label{Dp-F*}
\big(F(\bg),\ba\big)_{L^2(\Omega)} ~=~ \int_{\Omega} \bar\ba(\hat\bxi) \cdot \bv_{\bg_\Omega}^\infty(\hat\bxi) \, \text{d}S_{\hat\bxi} ~=~ \int_{\Omega_{\bd}} \bg(\bd) \cdot \overline{\int_{\Omega} \bW^{\infty*}(\bd,\hat\bxi)\cdot \ba(\hat\bxi) \, \text{d}S_{\hat\bxi}} \,\, \text{d}S_{\bd},
\end{equation}
where~$\ba\in L^2(\Omega)^3$. Thanks to the reciprocity identity~(\ref{W-recip}), inner product~\eqref{Dp-F*} exposes the adjoint far-field operator as
\begin{equation}\label{F*}
F^*(\ba)(\bd) ~=~ \int_{\Omega} \bW^{\infty*}(\bd,\hat\bxi)\cdot \ba(\hat\bxi) \, \text{d}S_{\hat\bxi} ~=~
\overline{\int_{\Omega} \bW^{\infty}(\hat\bxi,-\bd)\cdot \overline\ba(-\hat\bxi) \, \text{d}S_{\hat\bxi}} ~=~ \overline{F}(\tilde{\ba})(-\bd), \quad~ \bd\in\Omega,
\end{equation}
where $\tilde\ba(\hat\bxi)\!:=\!\overline{\ba}(-\hat\bxi)$ on~$\Omega$. Owing to the injectivity of $F$, one finds from~\eqref{F*} that setting $F^*(\ba)\!=\!\boldsymbol{0}$ necessitates~$\tilde\ba=\boldsymbol{0}$ and thus $\ba=\boldsymbol{0}$.}
\end{proof}
We are now in position to establish the main result of the GLSM approach, given by Theorem~\ref{GLSM1} and Theorem~\ref{GLSM2}, catering for the elastodynamic reconstruction of heterogeneous fractures.
\begin{theorem} \label{GLSM1}
\textcolor{black}{Consider the GLSM cost functional $\mathfrak{J}_\alpha$ unifying~\eqref{J-alph} and~\eqref{fJ-alph} with unperturbed operators $F^\delta$ and~$F_\sharp^\delta$, namely}
\begin{equation}\label{GCf}
\mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\,\bg) ~:=~ \norms{\nxs F\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}_{L^2(\Omega)}^2 \,+~ \alpha\,|(\bg, B\bg)|, \qquad \bg\in L^2(\Omega_{\bd})^3,
\end{equation}
\textcolor{black}{where $\alpha>0$ and~$B$, denoting either~$F$ or~$F_\sharp$, admits the factorization}
\begin{equation}\label{Bdf}
B ~=~ \mathcal{H^*} \exs \mathfrak{T} \exs \mathcal{H}, \qquad \mathfrak{T} ~=~ T,\,T_\sharp.
\end{equation}
Since $\mathfrak{J}_\alpha\geqslant 0$, define the infimum
\[
\mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty) ~\colon\!\!=~\! \inf\limits_{\bg \in L^2(\Omega_{\bd})^3} \! \mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\bg),
\]
and let $\bg_\alpha^L \in L^2(\Omega_{\bd})^3$ denote a nearby solution such that
\[
\mathfrak{J}_{\alpha}(\boldsymbol{\Phi}_L^\infty;\bg_\alpha^L) \,\,\leqslant \,\, \mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty) + \mu \alpha,
\]
$\mu>0$ being a constant independent of $\alpha$. Then,
\[
\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*) ~\iff~ \Big\{\limsup\limits_{\alpha \rightarrow 0} |( \bg_\alpha^L, B \bg_\alpha^L ) | < \infty ~\iff~ \liminf\limits_{\alpha \rightarrow 0} |( \bg_\alpha^L, B \bg_\alpha^L ) | < \infty\Big\}.
\]
\end{theorem}
\begin{proof} See the proof of Theorem~3 in~\cite{Audibert2014}, synthesized in~\ref{GLSM*pruf} using present notation.
\end{proof}
\begin{lemma}\label{min-Jad}
Consider the regularized GLSM cost functional $\mathfrak{J}_\alpha^\delta$ unifying~\eqref{RJ-alph} and~\eqref{RfJ-alph} with perturbed operators $F^\delta$ and~$F_\sharp^\delta$, namely
\begin{equation}\label{GCfn}
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg) ~:=~ \norms{\nxs F^\delta\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+~ \alpha \exs \big( \exs | ( \bg, B^\delta \bg ) | \,+\, \delta \! \norms{\nxs \bg \nxs}^2 \! \big), \qquad \bg \in L^2(\Omega_{\bd})^3
\end{equation}
where $\alpha,\beta\!>\!0$ and $B^\delta$ denotes either~$F^\delta$ or~$F^\delta_\sharp$. Assuming that~$B^\delta$ is compact, $\mathfrak{J}_\alpha^\delta$ has a minimizer $\bg_{\alpha,\delta}^L \in L^2(\Omega_{\bd})^3$ satisfying
\begin{equation}\label{limlim}
\lim\limits_{\alpha \rightarrow 0} \limsup\limits_{\delta \rightarrow 0} \exs \mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\,\bg_{\alpha,\delta}^L) ~=~ 0.
\end{equation}
\end{lemma}
\begin{proof}
\emph{Existence of a minimizer.}~For any $\alpha, \delta > 0$ and $\boldsymbol{\Phi}_L^\infty\!\in L^2(\Omega_{\bd})^3$ given by~\eqref{Phi-inf}, any sequence $(\bg^n)$ constructed to minimize $\mathfrak{J}_\alpha^\delta$ is bounded in $L^2(\Omega_{\bd})^3$, and thus weakly convergent to some $\bg_{\alpha,\delta}^L \in L^2(\Omega_{\bd})^3$. Thanks to the lower semi-continuity of a norm with respect to the weak convergence and the postulated compactness of $B^\delta$, one has
\begin{equation}\label{ming}
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta}^L) \,\, \leqslant \,\, \liminf\limits_{n \rightarrow \infty} \exs \mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg^n) \,\, \leqslant \,\, \!\! \inf\limits_{\bg \in L^2(\Omega_{\bd})^3} \! \mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg),
\end{equation}
which proves that $\bg_{\alpha,\delta}^L$ is a minimizer of $\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg)$ in~$L^2(\Omega_{\bd})^3$.
\emph{Limiting behavior.} \textcolor{black}{ Let us first observe from~\eqref{Ns-op}, \eqref{GCf} and~\eqref{GCfn} that
\begin{equation}\label{JadJa}
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg) \,\,\leqslant \,\, \mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\bg) ~+~
\delta \big\{2 \alpha\|\bg\|^2 \,+\, \delta\|\bg\|^2 \,+\, 2\|F\bg-\boldsymbol{\Phi}_L^\infty\|\exs \|\bg\|\big\}, \qquad \forall \bg \in L^2(\Omega_{\bd})^3.
\end{equation}
For any $\delta\!>\!0$ ($\alpha$ fixed), on can chose $\bg_{\alpha,\delta}$ such that $|\mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta})- \mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty)| \, \leqslant \, \delta$. Then by the definition of~$\bg_{\alpha,\delta}^L$ one finds via triangle inequality that
\[
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta}^L) \, \leqslant \
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta}) \, \leqslant \, \mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty) +
\delta \big\{1+ 2 \alpha\|\bg_{\alpha,\delta}\|^2 \,+\, \delta\|\bg_{\alpha,\delta}\|^2 \,+\, 2\|F\bg_{\alpha,\delta}-\boldsymbol{\Phi}_L^\infty\|\exs \|\bg_{\alpha,\delta}\|\big\}.
\]
The proof of~\eqref{limlim} is now completed by noting that (i) given~$\alpha$, the term inside the brackets is bounded for any~$\delta$, and (ii) $\lim\limits_{\alpha\to 0} \mathfrak{j}_\alpha=0$. }
\end{proof}
\begin{theorem}\label{GLSM2}
Under the assumptions of Theorem~\ref{GLSM1} and an additional hypothesis that $B^\delta$ (denoting either~$F^\delta$ or~$F^\delta_\sharp$) is compact, one has
\begin{multline}\notag
\boldsymbol{\Phi}_L^\infty \,\in\, Range(\mathcal{H}^*) ~~\iff~~
\Big\{\limsup\limits_{\alpha \rightarrow 0}\limsup\limits_{\delta \rightarrow 0}\big(\exs |(\bg_{\alpha,\delta}^L, B^\delta \bg_{\alpha,\delta}^L)| \,+\, \delta \! \norms{\! \bg_{\alpha,\delta}^L \!}^2 \nxs \big) \,<\, \infty \\
\iff~ \liminf\limits_{\alpha \rightarrow 0}\liminf\limits_{\delta \rightarrow 0}\big(\exs |( \bg_{\alpha,\delta}^L, B^\delta \bg_{\alpha,\delta}^L)| \,+\, \delta \! \norms{\! \bg_{\alpha,\delta}^L \!}^2 \nxs \big) \,<\, \infty\Big\},
\end{multline}
where~$\bg_{\alpha,\delta}^L$ is a minimizer of the perturbed GLSM cost functional~\eqref{GCfn} in the sense of~\eqref{limlim}.
\end{theorem}
\begin{proof} See the proof of Theorem~5 in~\cite{Audibert2014}, also summarized in~\ref{GLSM*pruf}.
\end{proof}
\vspace{-7 mm}
\subsubsection{The GLSM criteria for imaging heterogeneous fractures} \label{GLSM_C}
On the basis of Theorem~\ref{GLSM2}, a robust GLSM-based criterion for the elastic-wave reconstruction of heterogeneous fractures can be designed as
\begin{equation}\label{GLSMg}
I^{\mathcal{G}}(L) \,\, \colon \!\!\! = \,\, \dfrac{1}{\sqrt{|( \bg_{\alpha,\delta}^L, B^\delta \bg_{\alpha,\delta}^L)| \exs+\exs \delta \! \norms{ \bg_{\alpha,\delta}^L \!}^2}}, \qquad B^\delta ~=~ F^\delta, F^\delta_\sharp,
\end{equation}
where $\bg_{\alpha,\delta}^L$ is a minimizer of~\eqref{GCfn} in the sense of~\eqref{limlim}. In this setting, it is particularly instructive to focus on the case where $B^\delta = F^\delta_\sharp$, since $\bg_{\alpha,\delta}^L$ in this case can be obtained \emph{non-iteratively} by explicitly solving~(\ref{min-RJ}). Accordingly, the GLSM indicator functional used is the sequel is taken as
\begin{equation}\label{GLSMgs}
I^{{\mathcal{G}}_\sharp}(L) \,\, = \,\, \dfrac{1}{\sqrt{\norms{\!(F^\delta_\sharp)^{\frac{1}{2}} \exs \bg_{\alpha,\delta}^L \nxs}^2 \exs+\,\, \delta \! \norms{ \bg_{\alpha,\delta}^L \!}^2}}.
\end{equation}
For future reference, let us also recall the classical LSM/FM solution $\bg_\epsilon^L \in L^2(\Omega_{\bd})^3$ (see Theorem~\ref{TR2}) obtained by way of Tikhonov regularization~\cite{Kress1999}, namely
\begin{equation}\label{LSMg}
\textcolor{black}{
\bg_\epsilon^L \,\, \colon \!\!\! = \,\, \min_{\bg \in L^2(\Omega_{\bd})^3} \big\lbrace \! \norms{F^\delta \bg \,-\, \boldsymbol{\Phi}_L^\infty}^2 + \,\, \beta \! \norms{\bg}^2 \! \big\rbrace,}
\end{equation}
where~$\beta$ is a regularization parameter computable by the Morozov discrepancy principle.
\begin{rem}
It is worth noting that the GLSM characterization of $\Gamma$ from the far-field data (via the range of $F$) is deeply rooted in geometrical considerations, so that the fracture indicator functionals~(\ref{GLSMg}) and~(\ref{GLSMgs}) may exhibit only a minor dependence on its heterogeneous contact condition -- given by the distribution of~$\bK$ on~$\Gamma$. This behavior can be traced back to Remark~\ref{LSMrem}, where the opening displacement profile $\ba \in \tilde{H}^{1/2}(L)$ -- intimately related to the interface law -- is deemed arbitrary (within the constraints of admissibility). This quality makes the GLSM imaging paradigm particularly attractive in situations where the fracture's contact law is unknown beforehand, which opens up possibilities for the sequential geometrical reconstruction and interfacial characterization of partially-closed fractures.
\end{rem}
\section{Computational treatment and results} \label{numerics}
To illustrate the theoretical developments, this section examines the performance of~\eqref{GLSMgs} through a set of numerical experiments and compares the results of the GLSM reconstruction to those obtained by two alternative approaches, namely the linear sampling method (LSM)~\cite{Fiora2003} and the method of topological sensitivity (TS)~\cite{Fatemeh2015}. In what follows the synthetic sensory data, namely the far-field patterns~\eqref{vinf2} over the unit sphere, are generated by way of an elastodynamic boundary integral method~\cite{Fatemeh2015}.
\begin{figure}[tp]
\center\includegraphics[width=0.88\linewidth]{figures/SetNum4.jpg} \vspace*{0mm}
\caption{Elastic-wave sensing setup (left), position of the cutting plane (middle), and ``zebra'' pattern of the fracture's heterogeneous contact condition~(right).} \label{SetNum}
\end{figure}
\emph{Testing configuration.}~The sensing setup, shown in Fig.~\ref{SetNum}, features a ``true'' cylindrical fracture $\Gamma$ of length $L = 0.7$ and radius $R = 0.35$. The fracture is endowed with a piecewise-constant (``zebra'') distribution of interfacial stiffness~$\bK(\bxi)$ on~$\Gamma$, alternating between $\bK_1$ and~$\bK_2$, where
\[
\begin{aligned}
\bK_1 \:=\: (1-0.25\textrm{i}) \, \bn \otimes \bn \,\,+\, (4 - 2\textrm{i}) \, \be_1 \nxs \otimes \be_1 \,+\, (4 - 2\textrm{i}) \, \be_2 \nxs \otimes \be_2, \qquad \bK_2 \:=\: \boldsymbol{0}
\end{aligned}
\]
in terms of the orthonormal basis~$(\be_1, \be_2,\bn)$ shown in the figure. The shear modulus, mass density, and Poisson's ratio of the background solid are taken as $\mu = 1$, $\rho = 1$ and $\nu = 0.35$, whereby the shear and compressional wave speeds read $c_s = 1$ and $c_p = 2.08$, respectively. The interaction of $\Gamma$ with incident (P- and S-) plane waves, propagating in direction $\bd$, gives rise to the scattered wavefield $\bv$ solving~\eqref{GE} -- whose far-field pattern $\bv^\infty$ is then computed on the basis of~(\ref{vinf2}).
\emph{Far-field operator.}~For both illumination and sensing purposes, the unit sphere $\Omega$ is sampled by a uniform grid of $N_\theta \!\times\! N_\phi$ observation directions, specified by the polar ($\theta_j,\, j\!=\!1,\ldots N_\theta$) and azimuthal~($\phi_k, \,k\!=\!1,\ldots N_\phi$) angle values. With reference to~\eqref{mat1}, note that both the polarization vector $\bq\!=\!\bq_p\!\oplus\bq_s$ of an incident plane wave and the far-field pattern $\bv^\infty_\bq=\bv^\infty_{\bq_p}\!\oplus \bv^\infty_{\bq_s}$ of the scattered wave each have \emph{only three} nontrivial components. In this setting, the discretized far-field operator $\textrm{\bf{F}}$ is represented as a $3N\!\times 3N$ matrix ($N\!=\!N_\theta N_\phi$) with components
\begin{equation}\label{DF}
\textrm{\bf{F}}(3k\nxs+\nxs1\!:\!3k\nxs+\nxs3, \,3j\nxs+\nxs1\!:\!3j\nxs+\nxs3) ~=~ \textrm{\bf{W}}^\infty (\bd_j,\hat\bxi_k), \qquad j,k = 0,\ldots N-1,
\end{equation}
where
\begin{equation}\label{mat2}
\textrm{\bf{W}}^\infty (\bd_j,\hat\bxi_k) ~=~
\left[\begin{array}{cccccc}
W^{\infty}_{11} & W^{\infty}_{12} & W^{\infty}_{13} \\
W^{\infty}_{21} & W^{\infty}_{22} & W^{\infty}_{23} \\
W^{\infty}_{31} & W^{\infty}_{32} & W^{\infty}_{33}
\end{array}\right] (\bd_j,\hat\bxi_k),
\end{equation}
and~$W^{\infty}_{kj}$ $(j,k\!=\!1,2,3)$ are specified in~\eqref{mat1}. Unless stated otherwise, we assume $N_\theta=50$ and $N_\phi=25$.
\emph{Noisy data.}~To account for the presence of noise in measurements, we consider the perturbed far-field operator
\begin{equation}\label{DFN}
\textrm{\bf{F}}^\delta \,\, \colon \!\!\!= \, (\boldsymbol{I} + \boldsymbol{N}_{\!\epsilon} ) \exs \textrm{\bf{F}},
\end{equation}
where $\boldsymbol{I}$ is the $3N \times 3N$ identity matrix, and $\boldsymbol{N}_{\!\epsilon}$ is the noise matrix of commensurate dimension whose components are uniformly-distributed (complex) random variables in $[-\epsilon, \, \epsilon]^2$. On the basis of definition~(\ref{Ns-op}), one has $\delta = \norms{\!\boldsymbol{N}_{\!\epsilon} \exs \textrm{\bf{F}}\!}$ which in the sequel takes values of up to~$20\%$. With reference to Remark~\ref{LSMrem}, the region of interest
\emph{Trial far-field pattern.} With reference to Remark~\ref{LSMrem}, the GLSM indicator map~\eqref{GLSMgs} is constructed by solving~\eqref{min-RJ} for the minimizer of~(\ref{RJ-alph}) over a grid of trial infinitesimal fractures $L=\bz\!+\bR{\sf L}$, where~$\bz$ denotes the sampling point and $\bR$ is a unitary rotation matrix. In what follows, this is accomplished by taking~{\sf L} to be a vanishing penny-shaped fracture with unit normal~$\bn_\circ$, i.e. by setting the FOD in~(\ref{Phi-inf}) as $\ba(\by) = \delta (\by-\bz) \bR\bn_\circ$. Writing for brevity $\textrm{\bf{n}}=\bR\bn_\circ$, one in particular finds that
\begin{equation}\label{Phi-inf-num}
\boldsymbol{\Phi}_L^\infty(\hat\bxi) ~=~ - \Big(\text{i} k_p \, \hat\bxi \exs \big[ \exs \lambda+2\mu \exs (\textrm{\bf{n}} \cdot \hat\bxi)^2 \exs \big] \exs e^{-\text{i}k_p \hat\bxi \cdot \bz} \;\oplus\;
2 \text{i} \mu \exs k_s \,\hat\bxi \times \nxs (\textrm{\bf{n}} \times\hat\bxi)\exs (\textrm{\bf{n}}\cdot\hat\bxi) \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \bz} \Big).
\end{equation}
Recalling~\eqref{mat1}, one may note that for each observation direction~$\hat\bxi_k$, \eqref{Phi-inf-num} has only three non-trivial components in the reference $(\hat\bxi_k,\boldsymbol{\theta}_k,\boldsymbol{\phi}_k)$ orthonormal basis, which are for consistency with~\eqref{mat2} arranged as a $3N\!\times\!1$ vector
\begin{equation}\label{Phi-inf-Dnum}
\boldsymbol{\Phi}_{\bz,\textrm{\bf{n}}}^\infty(3k+1\!:\!3k+3) ~=\,
\left[\begin{array}{c} \text{i} k_p \big[ \exs \lambda + 2\mu \exs (\textrm{\bf{n}} \sip \hat\bxi_k)^2 \exs \big] e^{-\text{i}k_p \hat\bxi_k \sip \bz} \\
2 \text{i} \mu \exs k_s (\textrm{\bf{n}}\sip\boldsymbol{\theta}_k) (\textrm{\bf{n}}\sip\hat\bxi_k) \, e^{-\textrm{\emph{i}} k_s \hat\bxi_k \sip \bz} \\
2 \text{i} \mu \exs k_s (\textrm{\bf{n}}\sip\boldsymbol{\phi}_k) (\textrm{\bf{n}}\sip\hat\bxi_k) \, e^{-\textrm{\emph{i}} k_s \hat\bxi_k \sip \bz}
\end{array} \right], \qquad k=0,\ldots N-1.
\end{equation}
Accordingly, the far-field equation~\eqref{FF} takes the discretized form
\begin{equation}\label{Dff}
\textrm{\bf{F}}^\delta \bg_{\bz,\textrm{\bf{n}}} ~=~ \boldsymbol{\Phi}_{\bz,\textrm{\bf{n}}}^\infty,
\end{equation}
thus forming the basis for computing GLSM and LSM indicator functionals.
\subsection{Fracture indicators} \label{RRM}
As shown in Fig.~\ref{SetNum}, the search area i.e.~the sampling region is a \emph{cube of side 2} where the featured (GLSM and LSM) indicator functionals are evaluated. The resulting distributions are plotted either in three dimensions, or in the mid-section of the ``true'' cylindrical fracture (see Fig.~\ref{SetNum}).
\emph{Sampling.}~In what follows, the search cube $[-1,1]^3\subset\mathbb{R}^3$ is probed by a uniform $40 \!\times\! 40 \!\times\! 40$ grid of sampling points~$\bz$, while the unit sphere -- spanning possible fracture orientations -- is sampled by a $24 \!\times\! 6$ grid of trial normal directions $\textrm{\bf{n}}=\bR\bn_\circ$. Accordingly, the fracture indicator map is constructed by solving~(\ref{Dff}) for a total of $M = 64000 \!\times\! 144$ trial pairs $(\bz,\textrm{\bf{n}})$.
\emph{GLSM indicator.}~With reference to~\eqref{min-RJ} and~\eqref{DF}-\eqref{Dff}, a discretized version of the GLSM solution vector, $\bg^{\mbox{\tiny{GLSM}}}_{\bz,\textrm{\bf{n}}}$, is computed by solving the linear system
\begin{equation} \label{min-DRJ}
\Big( \textrm{\bf{F}}^{\delta *}\textrm{\bf{F}}^\delta + \alpha_{\bz,\textrm{\bf{n}}} \exs (\textrm{\bf{F}}_\sharp^\delta)^{\nxs\frac{1}{2}*} (\textrm{\bf{F}}_\sharp^\delta)^{\nxs\frac{1}{2}} + \alpha_{\bz,\textrm{\bf{n}}} \delta \exs \boldsymbol{I} \Big) \exs \bg^{\mbox{\tiny{GLSM}}}_{\bz,\textrm{\bf{n}}} ~=~ \textrm{\bf{F}}^{\delta *} \boldsymbol{\Phi}^\infty_{\bz,\textrm{\bf{n}}},
\end{equation}
where $(\cdot)^*$ is the Hermitian operator; $\textrm{\bf{F}}_\sharp^\delta$ is evaluated on the basis of definitions~(\ref{Fs}) and~(\ref{ReIm}); and, following~\cite{Audibert2014},
\begin{equation}\label{Alph}
\alpha_{\bz,\textrm{\bf{n}}} \,\, \colon \!\!\! = \,\, \frac{\eta_{\bz,\textrm{\bf{n}}}}{\norms{\textrm{\bf{F}}^\delta\!} + \,\, \delta}.
\end{equation}
Here $\eta_{\bz,\textrm{\bf{n}}}$ is a regularization parameter of the classical LSM solution~\eqref{lssm1}, computed via the Morozov discrepancy principle~\cite{Kress1999}. With reference to~(\ref{GLSMgs}), the GLSM indicator function is then obtained as
\begin{equation}\label{GLSM-Dgs}
I^{\mathcal{G}_\sharp}(\bz) \,\, = \,\, \dfrac{1}{\sqrt{\norms{\!(\textrm{\bf{F}}^\delta_\sharp)^{\frac{1}{2}} \exs \bg^{\mbox{\tiny{GLSM}}}_{\bz} \nxs}^2 \exs+\,\, \delta \! \norms{\bg^{\mbox{\tiny{GLSM}}}_{\bz} \!}^2}}, \qquad
\textcolor{black}{
\bg^{\mbox{\tiny{GLSM}}}_{\bz} \,\,\colon \!\!= \,\, \text{argmin}_{\bg^{\mbox{\tiny{GLSM}}}_{\bz, \textrm{\bf{n}}}} \norms{\bg^{\mbox{\tiny{GLSM}}}_{\bz,\textrm{\bf{n}}}}^2_{L^2(\Omega)}, ~ \textrm{\bf{n}}\in\Omega.}
\end{equation}
\emph{LSM indicator.}~To gain better insight into the effectiveness of the proposed approach, the GLSM reconstruction is compared to a corresponding~LSM map. The latter is computed on the basis of a Tikhonov-regularized solution $\bg^{\mbox{\tiny{LSM}}}_{\bz,\textrm{\bf{n}}}$ to~(\ref{Dff}), namely
\begin{equation}\label{lssm1}
\bg^{\mbox{\tiny{LSM}}}_{\bz,\textrm{\bf{n}}} \,\,\colon \!\!= \,\, \text{argmin}_{\bg_{\bz, \textrm{\bf{n}}}} \Big\{ \norms{\textrm{\bf{F}}^\delta \bg_{\bz,\textrm{\bf{n}}} \,-\, \boldsymbol{\Phi}_{\bz,\textrm{\bf{n}}}^\infty}^2_{L^2(\Omega)} \,+\,\,\, \eta_{\bz,\textrm{\bf{n}}} \norms{\bg_{\bz,\textrm{\bf{n}}}}^2_{L^2(\Omega)}\Big\},
\end{equation}
where the regularization parameter $\eta_{\bz,\textrm{\bf{n}}}$ is obtained by way of Morozov discrepancy principle~\cite{Kress1999}. On the basis of~\eqref{lssm1}, the LSM indicator functional is constructed following~\cite{Fiora2003} as
\begin{equation}\label{LSM}
I^{\mathcal{L}}(\bz) \,\, := \,\, \frac{1}{\norms{\bg^{\mbox{\tiny{LSM}}}_{\bz}}^2}, \qquad
\textcolor{black}{
\bg^{\mbox{\tiny{LSM}}}_{\bz} \,\,\colon \!\!= \,\, \text{argmin}_{\bg^{\mbox{\tiny{LSM}}}_{\bz, \textrm{\bf{n}}}} \norms{\bg^{\mbox{\tiny{LSM}}}_{\bz,\textrm{\bf{n}}}}^2_{L^2(\Omega)}, ~ \textrm{\bf{n}}\in\Omega.}
\end{equation}
\subsection{Results} \label{Comp}
In the sequel, the arclength ($\ell \!=\! 0.55$) of a ``true'' cylindrical fracture in its mid-plane, see Fig.~\ref{SetNum}, is used as a reference length to gauge the illuminating shear wavelength $\lambda_s = 2\pi/k_s$.
\emph{Density of the sensing grid}. Taking $\lambda_s/\ell = 0.7$, Fig.~\ref{NN} illustrates the sensitivity of the GLSM indicator~\eqref{GLSM-Dgs} to the spatial density of sensory data, given by $N_\theta \!\times\! N_\phi$ incident/observation directions over the unit sphere. This is done by gradual downsampling of the default $50 \!\times\! 25$ sensing grid. From the panels, it is apparent that for satisfactory geometric reconstruction, the sensing grid should carry at least 100 test directions over $\Omega$. In what follows, the (full-aperture) reconstructions are implemented using a $50 \!\times\! 25$ grid.
\emph{Sensitivity to measurement noise}. Assuming full-aperture illumination and sensing, the GLSM and LSM indicators are next compared in terms of their robustness against noise in the far-field data. With reference to~\eqref{DFN}, the levels of ``white'' noise used to contaminate the boundary integral simulations of the forward scattering problem are taken $\delta = \norms{\!\boldsymbol{N}_{\!\epsilon} \exs \textrm{\bf{F}}\!} \in\{0, 0.1, 0.2\} \| \textrm{\bf{F}}\|$. On focusing the comparison on the mid-section~$\Pi$ of a ``true'' fracture, the results are shown in Figs.~\ref{F1}, \ref{F2}, and \ref{F4} assuming the illuminating wavelengths of~$\lambda_s/\ell = 1.3$, $0.7$, and $0.3$, respectively. Note that $\delta\%:=\delta/\|\textrm{\bf{F}}\|$. As can be seen from the display, the GLSM indicator~\eqref{GLSM-Dgs} inherits the superior localization ability of its LSM predecessor~\eqref{LSM}, while carrying far greater robustness to noise in the sensory data.
\begin{figure}[h]
\center\includegraphics[width=0.64\linewidth]{figures/NN.jpg} \vspace*{0mm}
\caption{Full-aperture GLSM reconstruction of a cylindrical fracture in its mid-section, $\Pi$, for $\lambda_s/\ell = 0.7$: effect of density of the $N_\theta \!\times\! N_\phi$ sensing grid of illumination/observation directions spanning the unit sphere.} \label{NN}
\end{figure}
\begin{figure}[h!]
\center\includegraphics[width=0.66\linewidth]{figures/F1_v3.jpg} \vspace*{-2mm}
\caption{Sensitivity to measurement noise for $\lambda_s/\ell = 1.3$: Full-aperture reconstruction of a cylindrical fracture, mid-section~$\Pi$, by the LSM indicator (top panels) and its GLSM counterpart (bottom panels).} \label{F1} \vspace*{2mm}
\center\includegraphics[width=0.66\linewidth]{figures/F2_v3.jpg} \vspace*{-2mm}
\caption{Sensitivity to measurement noise for $\lambda_s/\ell = 0.7$: Full-aperture reconstruction of a cylindrical fracture, mid-section~$\Pi$, by the LSM indicator (top panels) and its GLSM counterpart (bottom panels).} \label{F2} \vspace*{2mm}
\center\includegraphics[width=0.66\linewidth]{figures/F4_v3.jpg} \vspace*{-2mm}
\caption{Sensitivity to measurement noise for $\lambda_s/\ell = 0.3$: Full-aperture reconstruction of a cylindrical fracture, mid-section~$\Pi$, by the LSM indicator (top panels) and its GLSM counterpart (bottom panels).} \label{F4}
\end{figure}
\emph{Effect of the sensing aperture}. The ramifications of an incomplete aperture on the quality of fracture reconstruction are illustrated in Figs.~\ref{H1} and~\ref{H4}, where only \textcolor{black}{\emph{the ``upper'' half} of $\Omega$ in Fig.~\ref{SetNum}} is available for the purposes of illumination and observation. More specifically, Figs.~\ref{H1} and~\ref{H4} depict the GLSM and LSM fields in the mid-section of $\Gamma$ at ``long'' ($\lambda_s/\ell = 1.3$) and ``short'' ($\lambda_s/\ell = 0.3$) excitation wavelengths, respectively, constructed from the half-aperture sensory data. While the loss of resolution in both GLSM and LSM maps is clear relative to Figs.~\ref{F1} and~\ref{F4}, it is noted that (for the problem under consideration) the GLSM indicator offers far better robustness to noise, providing acceptable reconstruction of~$\Gamma$ for~$\delta$ as high as~$0.1\|\textrm{\bf{F}}\|$.
\begin{figure}[h!]
\center\includegraphics[width=0.66\linewidth]{figures/H1_v3.jpg} \vspace*{-2mm}
\caption{Half-aperture reconstruction of a cylindrical fracture, mid-section $\Pi$, for $\lambda_s/\ell = 1.3$:~sensitivity of the LSM indicator (top panels) and its GLSM counterpart (bottom panels) to noise in the measurements.} \label{H1} \vspace*{0mm}
\center\includegraphics[width=0.66\linewidth]{figures/H4_v3.jpg} \vspace*{-2mm}
\caption{Half-aperture reconstruction of a cylindrical fracture, mid-section $\Pi$, for $\lambda_s/\ell = 0.3$:~sensitivity of the LSM indicator (top panels) and its GLSM counterpart (bottom panels) to noise in the measurements.} \label{H4} \vspace*{3mm}
\center\includegraphics[width=0.7\linewidth]{figures/3D_v3.png} \vspace*{-2mm}
\caption{Full-aperture 3D GLSM reconstruction for $\{\lambda_s/\ell = 1.3,\delta\% = 0.1\}$ (top) and $\{\lambda_s/\ell = 0.7,\delta\% = 0.05\}$ (bottom):~ GLSM indicator~\eqref{GLSM-Dgs} thresholded at 10\% (left), fracture surface as reconstructed from the 3D cloud of points (middle), and fracture reconstruction after the application of a mean filter (right).} \label{3D} \vspace*{-4mm}
\end{figure}
\emph{3D reconstruction}. For completeness, Fig.~\ref{3D} illustrates the full-aperture GLSM reconstruction of $\Gamma$ inside the sampling region~$[-1,1]^3$, assuming $\lambda_s/\ell = 1.3$ and $\delta\% = 10$ (top panels) and $\lambda_s/\ell = 0.7$ and $\delta\% = 5$ (bottom panels). For clarity, the indicator maps are thresholded by $10\%$, i.e.~only the sampling points whose~$I^{\mathcal{G}_\sharp}(\bz)$ values are higher than ten percent of the global maximum value are shown (left panels). Then, a scattered interpolant is constructed based on thus obtained 3D cloud of points, giving an optimal reconstruction of the fracture surface. The latter is generated by (i) projecting the thresholded GLSM map onto a reference plane (the $X-Y$ plane in this example), and (ii) defining a suitable grid of points covering the projected area. This forms the sought-for input for the scattered interpolant providing a 3D reconstruction of the fracture interface, as shown in the middle panels of Fig.~\ref{3D}. Due in part to a scattered nature of the interpolant, thus obtained fracture surface will suffer from some artificial roughness -- that depends for example on the density of sampling points and an ad-hoc thresholding parameter. This issue may be mitigated by implementing a suitable spatial (e.g. moving average) filter, as shown in the right panels of Fig.~\ref{3D}.
\section{Conclusions} \label{Conc}
The Generalized Linear Sampling Method (GLSM) combined with the $F_\sharp$-factorization technique form a fast, yet robust, platform for the geometric reconstruction of heterogeneous (and dissipative) discontinuity surfaces from scattered wavefield data. It is illustrated that the GLSM indicator possesses little sensitivity to (the reasonable levels of) measurement noise -- that is comparable to the robustness of TS, while inheriting the top-tier localization property of the classical LSM, which guarantees a high-quality geometric characterization of the fracture -- notwithstanding the frequency regime of excitation and the unknown (generally heterogeneous) interfacial stiffness $\bK$. Such attributes carries a remarkable potential for developing a GLSM-based hybrid approach for not only geometric reconstruction of hidden fractures, but also identification of their interfacial condition (e.g.~retrieval of $\bK$ in the present work) from scattered field data. Furthermore, this approach may be naturally and rigorously extended to other sensing configurations and to more sophisticated background-domain geometries. It should also be noted that the analysis in this study does not require the fracture surface to be \emph{connected}, so one should be able to use the GLSM for simultaneous imaging of multiple fractures in the medium.
\section{Introduction} \label{sec1}
Most recent advancements in the waveform tomography of discontinuity surfaces reside in the context of acoustic and electromagnetic inverse scattering. Spurred by the early study in~\cite{Kress1995}, such developments include: i) the Factorization Method (FM)~\cite{Bouk2013,Cham2014}; ii) the Linear Sampling Method (LSM)~\cite{Has2013, Fiora2003} and MUSIC algorithms~\cite{Park2009, Park2015(2)}; iii) the subspace migration technique~\cite{Park2015}, and iv) the method of Topological Sensitivity (TS)~\cite{Guz2004, Bonnet2011, Park2013}. In general, the LSM and FM techniques are applicable to a wide class of interfacial conditions and inherently carry a superior localization property -- potentially leading to high-fidelity geometric reconstruction. These methods, however, may suffer from the sensitivity to measurement uncertainties. In contrast the TS approach, that is inherently robust to noisy data, fails to adequately recover the shape of a scatterer at long illuminating wavelengths. The subspace migration methods offer another alternative for a high-fidelity reconstruction, even from partial-aperture data, while requiring some a priori knowledge about the geometry of a discontinuity surface. Among the aforementioned methods, the LSM has been applied to the problem of elastic-wave imaging of fractures with homogeneous (traction-free) boundary condition~\cite{Bour2013}, while the TS approach was recently extended to cater for qualitative elastodynamic sensing of fractures endowed with a more general class of contact laws~\cite{Bellis2013, Fatemeh2015}. In geophysics, major strides~\cite{Willis2006,Zheng2013,Minato2013,Minato2014,Fang2014} have been made toward a robust reconstruction of fractures via seismic waveform tomography. So far the proposed methods, often reliant upon a rudimentary parameterization of the fracture geometry (e.g. planar fractures) and nonlinear minimization, entail a number of impediments including: i) high computational cost; ii) sensitivity to the assumed parametrization; iii) computational instabilities~\cite{Minato2014}, and iv) major restrictions in terms of the seismic sensing configuration~\cite{Fang2014,Minato2013}, namely the location of sources and receivers relative to the (planar) fracture surface. One recent study aiming to mitigate such limitations can be found in~\cite{Zheng2013} that makes use of focused Gaussian beams emitted from the surface source/receiver arrays to non-iteratively assess the orientation, spacing, and compliance of systems of parallel planar fractures.
This work aims to develop a non-iterative, full-waveform approach to 3D elastic-wave imaging of fractures with non-trivial (generally heterogeneous and dissipative) interfacial condition. To this end, the sought indicator map -- targeting \emph{geometric} fracture reconstruction -- is preferably (i) agnostic with respect to the fracture's interfacial condition, (ii) robust against measurement errors, and (iii) flexible in terms of sensing parameters, e.g. the illumination frequency. This is pursued by drawing from the theories of inverse scattering~\cite{Fiora2008, Col1992} and, in particular, by building upon the Factorization Method~\cite{Kirsch2008, Bouk2013} and the recently developed Generalized Linear Sampling Method (GLSM)~\cite{Audibert2015, Audibert2014} which completes the theoretical foundation of its LSM~predecessor. First, the inverse problem is formulated in the frequency domain where the illuminating wavefield is described by the elastic Herglotz wave function~\cite{Dassios1995} with its inherent compressional (P) and shear (S) wave components. On characterizing the induced scattered wavefield in terms of its far-field P- and S-wave patterns~\cite{Martin1993}, the far-field operator~$F$ is then defined as a map from the Herglotz densities to the far-field measurements. In this setting, the GLSM indicator functional is introduced as in~\cite{Audibert2014} on the basis of (i) a custom factorization of the far-field operator, and (b) a sequence of approximate solutions to the LSM integral equation, seeking Herglotz densities whose far-field pattern matches that of a point-load solution radiating from the sampling point. The latter sequence is essentially a set of penalized least-squares misfit functionals -- aimed at producing nearby solutions to the LSM equation, where the penalty term is constructed using a factorization component of~$F$. Minimizing this class of cost functionals in their most general form requires an optimization procedure~\cite{Audibert2014}. Thanks to the premise of a linear contact law, however, this study takes advantage of the so-called $F_\sharp$-factorization~\cite{Kirsch2008, Bouk2013} of the far-field operator to formulate the penalty term. This results in a sequence of \emph{convex} GLSM cost functionals whose minimizers can be computed without iterations.
\section{Problem statement}\label{PS}
With reference to Fig.~\ref{fig1}(a), consider the elastic-wave sensing of a partially closed fracture $\Gamma \subset \mathbb{R}^3$ embedded in a homogeneous, isotropic, elastic solid endowed with mass density~$\rho$ and Lam\'{e} parameters $\mu$ and~$\lambda$. The fracture is characterized by a heterogeneous contact condition synthesizing the spatially-varying nature of its rough and/or multi-phase interface. Next, let $\Omega$ denote the unit sphere centered at the origin. For a given triplet of vectors $\bd\in\Omega$ and~$\bq_p,\bq_s\!\in\mathbb{R}^3$ such that $\bq_p\!\parallel\bd$ and~$\bq_s\!\perp\!\bd$, the obstacle is illuminated by a combination of compressional and shear plane waves
\begin{equation}\label{plwa}
\bu^{\text{\tiny f}}(\bxi) ~=~ \bq_p \exs e^{\textrm{i} k_p \bxi \cdot \bd} \:+\: \bq_s \exs e^{\textrm{i} k_s \bxi \cdot \bd}
\end{equation}
propagating in direction~$\bd$, where $k_p$ and $k_s=k_p\sqrt{(\lambda\!+\!2\mu)/\mu}$ denote the respective wave numbers. The interaction of $\bu^{\text{\tiny f}}$ with $\Gamma$ gives rise to the scattered field $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3\backslash\Gamma)^3$, solving
\begin{equation}\label{GE}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \bv) \,+\, \rho \exs \omega^2\bv ~=~ \boldsymbol{0} \quad &\text{in}& \quad {\mathbb{R}^3}\backslash\Gamma, \\*[1mm]
&\bn \cdot \bC \exs \colon \! \nabla \bv~=~ \mathcal{L}(\dbv) \,-\, \bt^{\text{\tiny f}} \quad &\text{on}& \quad \Gamma,
\end{aligned}
\end{equation}
where $\omega^2=k_s^2 \mu/\rho$ is the frequency of excitation; $\dbv=[\bv^+\!-\bv^-]$ is the jump in~$\bv$ across~$\Gamma$, hereon referred to as the fracture opening displacement \textcolor{black}{(FOD)};
\begin{equation}\label{bC}
\bC \:=\: \lambda\,\bI_2\!\otimes\bI_2 \:+\: 2\mu\,\bI_4
\end{equation}
is the fourth-order elasticity tensor; $\bI_m \,(m\!=\!2,4)$ denotes the $m$th-order symmetric identity tensor; \mbox{$\bt^{\text{\tiny f}} = \bn \cdot \bC \colon \! \nabla \bu^{\text{\tiny f}}$} is the free-field traction vector; $\bn = \bn^-$ is the unit normal on~$\Gamma$, and $\mathcal{L}: H^{1/2}(\Gamma)^3\to H^{-1/2}(\Gamma)^3$ represents a heterogeneous bijective contact law over the fracture surface, physically relating the displacement jump to surface traction. In many practical situations, the fracture's contact law is \emph{linearized} about a dynamic equilibrium state as
\begin{equation}
\label{contact} \mathcal{L}(\dbv) \:=\: \bK(\bxi) \dbv, \qquad \bxi \in \Gamma,
\end{equation}
where $\bK=\bK(\bxi)$ is a \emph{symmetric} (due to reciprocity considerations) and possibly \emph{complex-valued} matrix of specific stiffness coefficients.
\begin{figure}[tp]
\center\includegraphics[width=0.94\linewidth]{figures/fig1_v7.jpg} \vspace*{0mm}
\caption{Direct scattering problem. The fracture boundary $\Gamma$ is arbitrarily extended to a piecewise smooth, simply connected, closed surface $\partial D$ of a bounded domain~$D$.} \label{fig1}
\end{figure}
\begin{rem}
\textcolor{black}{In what follows, the analysis is based on the linear contact condition~(\ref{contact}) over~$\Gamma$. Under the premise of bijectivity, most of the ensuing developments (except for the $F_\sharp$ factorization method) can be adapted to handle nonlinear contact laws; such extension, however, is beyond the scope of this study.}
\end{rem}
The formulation of the direct scattering problem can now be completed by requiring that~$\bv$ satisfies the Kupradze radiation condition at infinity \cite{Kuprad1979}. On uniquely decomposing the scattered field into an irrotational part and a solenoidal part as $\bv = \bv^p + \bv^{s}$ \textcolor{black}{where
\begin{equation}\label{vpvs}
\bv^p = \frac{1}{k_s^2\!-\!k_p^2}(\Delta+k_s^2)\bv, \qquad \bv^{s} = \frac{1}{k_p^2\!-\!k_s^2}(\Delta+k_p^2)\bv,
\end{equation}}
the Kupradze condition can be stated as
\begin{equation}\label{KS}
\frac{\partial\bv^p}{\partial r} - \text{i} k_p \bv^p = o\big(r^{-1}\big) \quad \mbox{ and } \quad
\frac{\partial\bv^s}{\partial r} - \text{i} k_s \bv^ s = o\big(r^{-1}\big) \qquad \text{as} ~~r:=|\bxi|\to\infty,
\end{equation}
uniformly with respect to $\hat\bxi:=\bxi/r$.
\paragraph*{Dimensional platform.}
In what follows, all quantities are rendered \emph{dimensionless} by taking $\rho$, $\mu$, and ${\sf R}$ -- the characteristic size of a region sampled for fractures -- as the respective scales for mass density, elastic modulus, and length -- which amounts to setting $\rho = \mu = {\sf R} = 1$~\cite{Scaling2003}.
\paragraph*{Function spaces.}
To assist the ensuing analysis, the fracture surface $\Gamma$ is arbitrarily extended, as shown in Fig.~\ref{fig1}(b), to a piecewise smooth, simply connected, closed surface $\partial D$ of a bounded domain $D$ such that the normal vector $\bn$ to the fracture surface $\Gamma$ coincides with the outward normal vector to $\partial D$ -- likewise denoted by $\bn$. We also assume that $\Gamma$ is an open set (relative to $\partial D$) with positive surface measure.
Following \cite{McLean2000}, we define
\begin{equation}\label{funS}
\begin{aligned}
&H^{\pm \frac{1}{2}}(\Gamma) ~:=~\big\lbrace f\big|_\Gamma \colon \,\,\, f \in H^{\pm \frac{1}{2}}(\partial D) \big\rbrace, \\*[1 mm]
& \tilde{H}^{\pm \frac{1}{2}}(\Gamma) ~:=~\big\lbrace f \in H^{\pm\frac{1}{2}}(\partial D) \colon \,\,\, \text{supp}(f) \subset \overline{\Gamma} \big\rbrace,
\end{aligned}
\end{equation}
and recall that $H^{-1/2}(\Gamma)$ and $\tilde{H}^{-1/2}(\Gamma)$ are respectively the dual spaces of $\tilde{H}^{1/2}(\Gamma)$ and $H^{1/2}(\Gamma)$. Accordingly, the following embeddings hold
\begin{equation}\label{embb}
\tilde{H}^{\frac{1}{2}}(\Gamma) \,\subset\, H^{\frac{1}{2}}(\Gamma) \,\subset\, L^2(\Gamma) \,\subset\, \tilde{H}^{-\frac{1}{2}}(\Gamma) \,\subset\, H^{-\frac{1}{2}}(\Gamma).
\end{equation}
\begin{rem} \textcolor{black}{In the context of fracture mechanics, it is well known that $\dbv(\bxi)\to \boldsymbol{0}$ continuously as $\Gamma\!\ni\!\bxi\to\partial\Gamma$ (typically as $d^\alpha$, $0\!<\!\alpha\!\leqslant\!\tfrac{1}{2}$~\cite{Ueda2006} where $d$ is a normal distance to $\partial\Gamma$ when $\partial\Gamma$ is smooth), which lends credence to the assumption $\dbv\in\tilde{H}^{1/2}(\Gamma)^3$ used hereon.}
\end{rem}
\section{On the well-posedness of the forward scattering problem} \label{WP}
Serving as a prerequisite for the analysis of the inverse scattering problem, this section investigates the well-posedness of the direct scattering problem \eqref{GE}--\eqref{KS}. Let $R>0$ be sufficiently large so that the ball $B_R$ of radius $R$ contains $\Gamma$, and consider the Dirichlet-to-Neumann operator ${\mathcal T}_R: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3$ associated with the scattering problem in $\mathbb{R}^3 \backslash B_R$, namely
$$
{\mathcal T}_R({\boldsymbol{\varphi}})(\bxi) := \hat{\bxi}\cdot \bC \colon \! \nabla \btu_{{\boldsymbol{\varphi}}}(\bxi), \qquad \bxi\in\partial B_R,
$$
where $\btu_{\boldsymbol{\varphi}} \in H^1_\mathrm{loc}(\mathbb{R}^3 \backslash B_R)^3$ is the unique radiating solution, satisfying~\eqref{KS}, of
\begin{equation}\label{TR}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \btu_{\boldsymbol{\varphi}}) \,+\, \rho \exs \omega^2 \btu_{\boldsymbol{\varphi}} ~=~ \boldsymbol{0} \quad &\text{in}& \quad {\mathbb{R}^3}\backslash B_R, \\*[1mm]
& \btu_{\boldsymbol{\varphi}}~=~{\boldsymbol{\varphi}} \quad &\text{on}& \quad \partial B_R.
\end{aligned}
\end{equation}
The scattering problem \eqref{GE}--\eqref{KS} can now be equivalently written in terms of $\bv\in H^1(B_R\backslash\Gamma)^3$ as
\begin{equation}\label{GEbis}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \bv) \,+\, \rho \exs \omega^2\bv ~=~ \boldsymbol{0} \quad &\text{in}& \quad {\mathbb{R}^3}\backslash\Gamma, \\*[1mm]
&\bn \cdot \bC \exs \colon \! \nabla \bv ~=~ \bK\sip\dbv \,-\, \bt^{\text{\tiny f}} \quad &\text{on}& \quad \Gamma, \\*[1mm]
&\bn \cdot \bC \colon \! \nabla \bv = {\mathcal T}_R(\bv) \quad &\text{on}& \quad \partial B_R,
\end{aligned}
\end{equation}
where $\bn(\bxi)=\hat\bxi$ on~$\partial B_R$. This problem can be written variationally in terms of $\bv\in H^1(\mathcal{B}_2\backslash\Gamma)^3$ as
\begin{equation}\label{Wik-GE}
\begin{aligned}
&- \rho \exs \omega^2 \! \int_{\mathcal{B}_2 \backslash \Gamma} \overline\bw \cdot \bv \,\exs \textrm{d}V_{\bxi} ~+\, \int_{\mathcal{B}_2 \backslash \Gamma} \nabla \exs \overline{\bw} \colon \bC \colon \nabla \bv \, \textrm{d}V_{\bxi} ~+\, \dualGA{\bK \sip \dbv}{\llbracket\bw\rrbracket}~-\, \\*[1 mm]
& \quad\,\, \dualBR{{\mathcal T}_R(\bv)}{\bw} ~=\, \int_{\Gamma} \exs \overline{\llbracket \bw \rrbracket} \cdot \bt^{\text{\tiny f}} \, \textrm{d}S_{\bxi}, \qquad \quad \forall \exs \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\end{aligned}
\end{equation}
where $\dualGA{\cdot}{\cdot}$ and $\dualBR{\cdot}{\cdot}$ respectively denote the $\big\langle H^{-1/2}(\Gamma)^3, \tilde{H}^{1/2}(\Gamma)^3 \big\rangle$ and $\big\langle H^{-1/2}(\partial B_R)^3, {H}^{1/2}(\partial B_R)^3 \big\rangle$ \emph{duality products} that extend $L^2$ inner products. The analysis of the forward scattering problem is based on the following properties of the Dirichlet-to-Neumann operator ${\mathcal T}_R$ (see also \cite{BramblePasciak}). \textcolor{black}{For clarity, we will use an abbreviated notation of relevant vector norms where e.g. $\|\boldsymbol{\cdot}\|_{H^{1/2}(\Gamma)^3}$ is denoted by~ $\|\boldsymbol{\cdot}\|_{H^{1/2}(\Gamma)}$ and so on.}
\begin{lemma} \label{LemmaTR}
There exists a bounded, non-negative and self-adjoint operator ${\mathcal T}_R^0\!: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3$ such that ${\mathcal T}_R+{\mathcal T}_R^0\!: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3 $ is compact. Moreover,
\begin{equation}
\Im \dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\varphi}}} > 0 \qquad \forall {\boldsymbol{\varphi}} \in H^{1/2}(\partial B_R)^3:~ {\boldsymbol{\varphi}} \neq 0.
\end{equation}
\end{lemma}
\begin{proof} Let $R_{\circ} > R$ and ${\boldsymbol{\varphi}}, {\boldsymbol{\psi}} \in H^{1/2}(\partial B_R)^3 $. Multiplying the first equation in~\eqref{TR} by $\overline{\btu_{\boldsymbol{\psi}}}$ and
integrating by parts on $B_{R_{\circ}} \backslash B_R$ yields
$$
\dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\psi}}} ~=~ \rho\exs\omega^2 \! \int_{B_{R_{\circ}}\!\backslash B_R} \overline{\btu_{\boldsymbol{\psi}}}\cdot \btu_{\boldsymbol{\varphi}} \,\exs \textrm{d}V_{\bxi} ~- \int_{B_{R_{\circ}}\!\backslash B_R} \nabla \exs \overline{\btu_{\boldsymbol{\psi}}} \colon \bC \colon \nabla \btu_{\boldsymbol{\varphi}}
\, \textrm{d}V_{\bxi} ~+ \int_{\partial{B_{R_{\circ}}}} \exs\overline{\btu_{\boldsymbol{\psi}}} \cdot \bt({\boldsymbol{\varphi}}) \, \textrm{d}S_{\bxi},
$$
where $\bt({\boldsymbol{\varphi}})(\bxi):= \hat{\bxi}\cdot \bC \colon \! \nabla \btu_{{\boldsymbol{\varphi}}}(\bxi)$ for $\bxi \in \partial B_{R_{\circ}}$. Using the well-posedness of \eqref{TR} and the Riesz representation theorem, we define ${\mathcal T}_R^0$ by
$$
\dualBR{{\mathcal T}_R^0({\boldsymbol{\varphi}})}{{\boldsymbol{\psi}}} := \int_{B_{R_{\circ}}\!\backslash B_R} \nabla\exs\overline{\btu_{\boldsymbol{\psi}}} \colon \bC \colon \nabla \btu_{\boldsymbol{\varphi}}
\, \textrm{d}V_{\bxi}.
$$
\textcolor{black}{On demonstrating that $\|({\mathcal T}_R+{\mathcal T}_R^0)({\boldsymbol{\varphi}})\|_{H^{-1/2}(\Gamma)}\leqslant C (\|\btu_{{\boldsymbol{\varphi}}}\|_{L^2(B_{R_{\circ}}\!\backslash B_R)}+\|\bt({\boldsymbol{\varphi}})\|_{L^2(\partial B_{R_{\circ}}\!)})$ for some constant~$C\!>\!0$ independent of~${\boldsymbol{\varphi}}$, the compactness of ${\mathcal T}_R+{\mathcal T}_R^0$ then follows from the compactness of mapping ${\boldsymbol{\varphi}}\to\btu_{\boldsymbol{\varphi}}$ (resp. ${\boldsymbol{\varphi}}\to\bt({\boldsymbol{\varphi}})$) from $H^{1/2}(\partial B_R)$ into~$L^2(B_{R_{\circ}}\!\backslash B_R)$ (resp.~$L^2(\partial B_{R_{\circ}})$) thanks to the compact embedding of $H^1(B_{R_{\circ}}\!\backslash B_R)$ into~$L^2(B_{R_{\circ}}\!\backslash B_R)$ and the standard regularity results for scattering problems~\cite{McLean2000}, which can be recovered from the boundary integral representation of $\btu_{\boldsymbol{\varphi}}$ in $\mathbb{R}^3\backslash B_R$ in terms of boundary data on~$\partial B_R$. As shown in~\ref{lem1-p}, the sign of the imaginary part of ${\mathcal T}_R$ is a consequence of the asymptotic behavior of $\btu_{\boldsymbol{\varphi}}$ at infinity~\cite{Kuprad1979} which implies}
\begin{equation}\label{kup1}
\Im \dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\varphi}}} = \Im \lim_{R_{\circ} \to \infty} \int_{\partial{B_{R_{\circ}}}} \exs\overline{\btu_{\boldsymbol{\varphi}}} \cdot \bt({\boldsymbol{\varphi}}) \,
\textrm{d}S_{\bxi} \;=\; \lim_{R_{\circ} \to \infty} \int_{\partial{B_{R_{\circ}}}} \Big\lbrace k_p (\lambda+2\mu) |\btu_{\boldsymbol{\varphi}}^{\mathrm{p}} \exs |^2 \,+\, k_s\mu\exs |\btu_{\boldsymbol{\varphi}}^{\mathrm{s}}|^2 \Big\rbrace \,\, \textrm{d}S_{\bxi}.
\end{equation}
The \textcolor{black}{sign-definiteness} of the imaginary part is a consequence of the Rellich lemma \cite{Col1992} applied to $\btu_{\boldsymbol{\varphi}}^{\mathrm{p}}$ and $\btu_{\boldsymbol{\varphi}}^{\mathrm{s}}$, which requires that $\btu_{\boldsymbol{\varphi}} = \btu_{\boldsymbol{\varphi}}^{\mathrm{p}} + \btu_{\boldsymbol{\varphi}}^{\mathrm{s}}=0$ whenever $\Im \dualBR{{\mathcal T}_R({\boldsymbol{\varphi}})}{{\boldsymbol{\varphi}}}=0$.
\end{proof}
\begin{theorem} \label{maindirect}
Assume that $\bt^{\text{\tiny f}} \in H^{-1/2}(\Gamma)^3$ and that $\bK \in L^\infty(\Gamma)^{3\times 3}$ is symmetric such that $\Im\bK\leqslant\boldsymbol{0}$ on $\Gamma$, i.e. that $\overline{\boldsymbol{\theta}} \sip \Im \bK(\bxi)\sip \boldsymbol{\theta} \leqslant 0$, $\forall\boldsymbol{\theta}\in\mathbb{C}^3$ and a.e.~on~$\Gamma$. Then problem~\eqref{Wik-GE} has a unique solution that continuously depends on $\bt^{\text{\tiny f}} \in H^{-1/2}(\Gamma)^3$.
\end{theorem}
\begin{proof}
Since $\bt^{\text{\tiny f}} \in H^{-1/2}(\Gamma)$, the antilinear form $\int_{\Gamma} \exs \overline{\llbracket \bw \rrbracket} \cdot \bt^{\text{\tiny f}} \, \textrm{d}S_{\bxi}$ may be understood as a duality pairing $\dualGA{\cdot}{\cdot}$. The continuity of this form comes from the continuity of the trace mapping $\bw \to \llbracket \bw \rrbracket$ from $ H^1(\mathcal{B}_2 \backslash \Gamma)^3$ into $\tilde{H}^{1/2}(\Gamma)^3$.
On the basis of the adopted dimensional platform i.e.~$\rho = \mu = 1$ (see Section~\ref{PS}), the sesquilinear form on the left hand side of \eqref{Wik-GE} can be decomposed into a coercive part
\begin{equation}\label{cv}
\begin{aligned}
&\textrm{A}(\bv,\bw) ~=\, \! \int_{\mathcal{B}_2 \backslash \Gamma} \overline\bw \cdot \bv \,\exs \textrm{d}V_{\bxi} ~+\, \int_{\mathcal{B}_2 \backslash \Gamma} \!\! \nabla \exs \overline{\bw} \colon \! \bC \colon \! \nabla \bv \, \textrm{d}V_{\bxi} ~+\, \dualBR{{\mathcal T}_R^0(\bv)}{\bw} , \quad \forall \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\end{aligned}
\end{equation}
and a compact part
\begin{equation}\label{compact}
\text{B}(\bv,\bw) ~=\, - (1+k_s^2) \int_{\mathcal{B}_2 \backslash \Gamma} \overline\bw
\cdot \bv \,\, \textrm{d}V_{\bxi} ~+\, \dualGA{\bK \sip \dbv}{\llbracket\bw\rrbracket}~ -\,
\dualBR{({\mathcal T}_R+{\mathcal T}_R^0)(\bv)}{\bw} , \quad \forall \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3.
\end{equation}
The coercivity of $\textrm{A}(\bv,\bw)$ follows from the Korn inequality~\cite{McLean2000} and the non negative sign of ${\mathcal T}_R^0$ (Lemma \ref{LemmaTR}). Now, in order to prove that the antilinear form $\text{B}$ defines a compact perturbation of $\textrm{A}(\bv,\bw)$, one may observe that
\[
| \text{B}(\bv,\bw) | ~\leqslant~\! \text{c}_2 \exs \big\lbrace \nxs
\norms{\!\bv\!}_{L^2(\mathcal{B}_2 \backslash \Gamma)} \, \norms{\!\bw\!}_{L^2(\mathcal{B}_2
\backslash \Gamma)} \,+\, \norms{\! \llbracket \bv \rrbracket
\!}_{L^2(\Gamma)} \, \norms{\! \llbracket \bw \rrbracket \!}_{L^2(\Gamma)}
\! \big\rbrace + \norms{({\mathcal T}_R+{\mathcal T}_R^0)(\bv)}_{H^{-1/2}(\partial B_R)}\norms{\bw}_{H^{1/2}(\partial B_R)}
\]
for a constant $\text{c}_2$ independent of $\bv$ and $\bw$. The claim then follows from Lemma \ref{LemmaTR}, the compact embedding of $H^1(\mathcal{B}_2 \backslash \Gamma)$ into $L^2(\mathcal{B}_2 \backslash \Gamma)$ and the compactness of the trace operator $\bv \rightarrow \llbracket \bv \rrbracket$ as an application from $H^1(\mathcal{B}_2 \backslash \Gamma)$ into ${L^2(\Gamma)}$ where the latter comes from the compact embedding of $\tilde H^{1/2}(\Gamma)$ into $L^2(\Gamma)$.
Problem \eqref{Wik-GE} is then of Fredholm type, and is therefore well-posed as soon as the uniqueness of a solution is guaranteed. Assume that $\bt^{\text{\tiny f}} = 0$. Then
$$
\Im \dualBR{{\mathcal T}_R(\bv)}{\bv} =\, \dualGA{\Im \bK \sip \dbv}{\llbracket\bv\rrbracket} \,\leqslant\: 0
$$
by premise of the Theorem. Thanks to Lemma \ref{LemmaTR}, this requires that $\bv\!=\!\boldsymbol{0}$ on $\partial B_R$ and thus $\bv\!=\!\boldsymbol{0}$ in $\mathcal{B}_2 \backslash \Gamma$ by the unique continuation principle.
\end{proof}
\section{Elements of the inverse scattering solution}\label{Prelim}
\renewcommand{\Omega_{\bd}}{{\Omega}}
\renewcommand{\boldsymbol{u}^{\textit{i}}}{\btu}
This section is devoted to the introduction of the \emph{far-field operator} -- relevant to the scattering problem~(\ref{GE}), and the derivation of its first and second factorizations. In the sequel, we assume that the hypotheses of Theorem \ref{maindirect} hold.
\paragraph*{Elastic Herglotz wave function.} For given density $\bg \in L^2(\Omega_{\bd})^3$, we consider the unique decomposition
\begin{equation}\label{herden}
\bg \;:=\; \bg_p \oplus\, \bg_s
\end{equation}
such that $\bg_p(\bd)\!\parallel\!\bd$ and $\,\bg_s(\bd)\!\perp\!\bd$, $\,\bd\in\Omega_{\bd}$. In dyadic notation, one has
\begin{equation}\label{freef}
\bg_p(\bd) := (\bd \nxs \otimes \nxs \bd \exs) \cdot \bg(\bd) \quad ~~\text{and}~~ \quad \bg_s(\bd) := (\bI - \bd \nxs \otimes \nxs \bd \exs) \cdot \bg(\bd).
\end{equation}
Next, we define the elastic Herglotz wave function~\cite{Dassios1995} as
\begin{equation}\label{HW}
\boldsymbol{u}^{\textit{i}}_\bg(\bxi) ~: =~ \int_{\Omega_{\bd}} \bg_p(\bd) \exs e^{\textrm{i} k_p \bd \cdot \bxi} \,\, \text{d}S_{\bd} \,+\, \int_{\Omega_{\bd}} \bg_s(\bd) \exs e^{\textrm{i} k_s \bd \cdot \bxi} \,\, \text{d}S_{\bd}, \qquad \bxi \in \mathbb{R}^3
\end{equation}
in terms of {the compressional and shear wave densities~$\bg_p$ and~$\bg_s$.
\paragraph*{The far-field pattern.}
As shown in~\cite{Martin1993}, any scattered wave $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ solving \eqref{GE}-\eqref{KS} has the asymptotic
expansion
\begin{equation}\label{vinf}
\bv(\bxi) ~=~ \exs \frac{e^{\text{i}k_p r}\!}{4 \pi(\lambda\!+\!2\mu)r} \exs \bv^\infty_p(\hat\bxi) \:+\:
\frac{e^{\text{i}k_s r}\!}{4\pi\mu r} \exs \bv^\infty_s(\hat\bxi) \:+\: O(r^{-2}) \quad~~ \text{as} \quad r:=|\bxi|\to\infty,
\end{equation}
where $\hat\bxi$ is the unit direction of observation, while $\bv^\infty_p$ and $\bv^\infty_s$ denote respectively the far-field patterns of $\bv^p$ and $\bv^s$ -- see~\eqref{vpvs}, which satisfy $\bv^\infty_p\!\parallel\hat\bxi$ and $\bv^\infty_s\!\perp\hat\bxi$. In this setting, we define the far-field pattern of~$\bv$ by
\begin{equation}\label{far-field}
\bv^\infty := \bv^\infty_p \oplus \bv^\infty_s.
\end{equation}
By way of the integral representation theorem in elastodynamics~\cite{Bon1999} and the far-field representation of the elastodynamic fundamental stress tensor (see Appendix), one can show that if $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ satisfies~\eqref{GE}-\eqref{KS}, then
\begin{equation}\label{vinf2}
\begin{aligned}
&\bv_p^\infty(\hat\bxi) ~=~- \text{i} k_p \exs \hat\bxi \int_\Gamma \Big\lbrace \lambda \, \llbracket \bv \rrbracket \sip \bn + 2\mu \big(\bn \sip \hat\bxi \big) \exs \llbracket \bv \rrbracket \sip \hat\bxi \exs \Big\rbrace \, e^{-\text{i}k_p \hat\bxi \cdot \bx} \,\, \text{d}S_{\bx}, \\*[1 mm]
& \bv_s^\infty(\hat\bxi) ~=~ -\text{i}k_s \exs \hat\bxi \exs \times \int_{\Gamma} \Big\lbrace \mu \big( \llbracket \bv \rrbracket \!\times\! \hat\bxi \exs \big)(\bn \sip \hat\bxi \exs ) \,+\, \mu \big( \bn \!\times \hat\bxi \big) (\llbracket \bv \rrbracket \sip \hat\bxi) \Big\rbrace \, e^{-\text{i}k_s \hat\bxi \cdot \bx} \,\, \text{d}S_{\bx}.
\end{aligned}
\end{equation}
\paragraph*{The far-field operator.}
\begin{defn}\label{deffarf}
We define the far-field operator $F: L^2(\Omega_{\bd})^3 \to L^2(\Omega_{\bd})^3$ by
\begin{equation}\label{ffo0}
F(\bg) ~=~ \textcolor{black}{\bv_{\bg_\Omega}^\infty,}
\end{equation}
where~$\bv_{\bg_\Omega}^\infty$ is the far-field pattern~\eqref{far-field} of~$\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ solving~\eqref{GE}-\eqref{KS} with data $\bu^{\text{\tiny f}} = \boldsymbol{u}^{\textit{i}}_\bg$, see~\eqref{HW}.
\end{defn}
When the contact law specified by~$\mathcal{L}(\dbv)$ is linear as in~(\ref{contact}), the far-field operator can be expressed as a linear integral operator. To examine this case, \textcolor{black}{consider an incident plane wave~(\ref{plwa}) propagating in direction $\bd\in\Omega_{\bd}$ with amplitude $\bq=\bq_p\oplus\bq_s$}, and denote the induced far-field pattern~(\ref{far-field}) by \textcolor{black}{$\bv^\infty_\bq(\bd, \cdot)=\bv^\infty_{\bq_p}\oplus \bv^\infty_{\bq_s}$}. Next, let us define the far-field kernel \textcolor{black}{$\bW^\infty(\bd,\hat\bxi)\in\mathbb{C}^{6\times 6}$} so that
\begin{equation}\label{w-inf}
\bW^\infty(\bd,\hat\bxi) \sip \bq ~:=~ \bv^\infty_\bq(\bd, \hat\bxi).
\end{equation}
Then one easily verifies that
\begin{equation}\label{ffo2}
F(\bg)\textcolor{black}{(\hat\bxi)} ~=\, \int_{\Omega_{\bd}} \bW^\infty(\bd,\hat\bxi) \sip \bg(\bd) \,\, \text{d}S_{\bd}.
\end{equation}
\begin{lemma}\label{recip}
The far-field kernel $\bW^\infty(\bd,\hat\bxi)$ satisfies the reciprocity identity
\begin{equation}\label{W-recip}
\textcolor{black}{
\bW^\infty(\bd,\hat\bxi) ~=~ \overline{\bW^{\infty *}}(-\hat\bxi,-\bd), \qquad \forall\bd,\hat\bxi\!\in\Omega.}
\end{equation}
\end{lemma}
\begin{proof}
\textcolor{black}{See~\ref{Recip}}.
\end{proof}
\section{Key properties for the application of sampling methods}\label{SFS}
\paragraph*{Factorization of the far-field operator $F$.}
Consider the Herglotz operator $\mathcal{H} \colon L^2(\Omega_{\bd})^3 \rightarrow H^{-1/2}(\Gamma)^3$ given by
\begin{equation}\label{oH}
\mathcal{H}(\bg) ~:=~ \bn\cdot\bC\exs\colon\!\nabla\boldsymbol{u}^{\textit{i}}_\bg \quad~~ \text{on}\quad \Gamma,
\end{equation}
where $\boldsymbol{u}^{\textit{i}}_\bg$ is the Herglotz wave function~(\ref{HW}). Next, define $\mathcal{G} \colon H^{-1/2}(\Gamma)^3 \rightarrow L^2(\Omega_{\bd})^3$ as the map taking the traction vector $\bt^{\text{\tiny f}}$ over $\Gamma$ to the far-field pattern, $\bv^\infty$, of $\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ satisfying \eqref{GE}-\eqref{KS}. Then from Definition~\ref{deffarf}, the far-field operator~\eqref{ffo0} becomes
\begin{equation}\label{fac1}
F ~=~ \mathcal{G} \mathcal{H}.
\end{equation}
\begin{lemma}\label{H*}
With reference to decomposition~\eqref{far-field}, the adjoint Herglotz operator $\mathcal{H}^* \colon \tilde{H}^{1/2}(\Gamma)^3 \rightarrow L^2(\Omega_{\bd})^3$ takes the form
\begin{equation}\label{Hstar}
\begin{aligned}
\mathcal{H}^*\nxs(\ba)(\hat\bxi) ~=~ - \Big( \, & \text{\emph{i}} k_p \, \hat\bxi \exs \int_\Gamma \, \big\lbrace \lambda \exs (\ba \sip \bn) \,+\, 2\mu \exs (\bn \sip \hat\bxi) ( \ba \sip \hat\bxi) \big\rbrace \, e^{-\text{\emph{i}}k_p \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \\*[1 mm]
& \textcolor{black}{\oplus}~ \text{\emph{i}} k_s \, \hat\bxi\times\! \int_\Gamma \big\lbrace \mu \exs(\ba \times \hat\bxi)(\bn\sip\hat\bxi) \,+\, \mu \exs (\bn \times \hat\bxi) (\ba \sip \hat\bxi) \big\rbrace \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \Big).
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
\textcolor{black}{see~\ref{H*pruf}}.
\end{proof}
On the basis of~\eqref{vinf2} and~(\ref{Hstar}), map $\mathcal{G}$ can be further decomposed as $\mathcal{G}=\mathcal{H}^* T$ where the middle operator $T\colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ is given by
\begin{equation}\label{T}
T(\bt^{\text{\tiny f}})(\bxi) ~:=~ \llbracket \bv(\bxi) \rrbracket, \qquad \bxi \in \Gamma
\end{equation}
such that~$\bv\in H^1_{\mathrm{loc}}(\mathbb{R}^3 \backslash \Gamma)^3$ satisfies~\eqref{GE}-\eqref{KS} or equivalently \eqref{Wik-GE}. Thanks to this new decomposition of $\mathcal{G}$, the second factorization of $F \colon L^2(\Omega_{\bd})^3 \rightarrow L^2(\Omega_{\bd})^3$ is obtained
\begin{equation}\label{fact}
F ~=~ \mathcal{H}^* \exs T \exs \mathcal{H},
\end{equation}
which provides the second important ingredient for the ensuing analysis.
\paragraph*{Properties of the Herglotz operator $\mathcal{H}$.}
\begin{lemma}\label{H*p}
Operator $\mathcal{H^*}:\tilde{H}^{1/2}(\Gamma)^3\rightarrow L^2(\Omega_{\bd})^3$ in Lemma~\ref{H*} is compact and injective.
\end{lemma}
\begin{proof}
Integral operator $\mathcal{H}^*$ has a smooth kernel and is therefore compact from $\tilde{H}^{1/2}(\Gamma)^3$ into $L^2(\Omega_{\bd})^3$. Next, suppose that there exists $\ba \in \tilde{H}^{1/2}(\Gamma)^3$ such that $\mathcal{H^*}(\ba) = \boldsymbol{0}$. In light of~(\ref{vinf}) and~(\ref{vinf2}), it is apparent that $\mathcal{H^*}$ is nothing else but the far-field operator stemming from the double-layer potential
\begin{equation}\label{Dlp}
\bV(\ba)(\bxi) ~=~ \int_{\Gamma} \ba(\by) \cdot \bfT(\bxi,\by) \, \text{d}S_{\by}, \qquad
\bfT(\bxi,\by) ~=~ \bn(\by)\cdot\bSig(\bxi,\by), \qquad \bxi \in \mathcal{B}_2 \backslash \Gamma,
\end{equation}
where
$\bSig(\bxi,\by)$ is the (third-order) elastodynamic fundamental stress tensor \textcolor{black}{given in~\ref{stress-fund}}. By virtue of definition~(\ref{vinf}), vanishing far-field pattern of $\bV(\ba)$ implies, by the Rellich Lemma and the unique continuation principle, that $\bV(\ba) = \boldsymbol{0}$ in $\mathbb{R}^3 \backslash \Gamma$. Owing to the fundamental jump property of double-layer potentials by which~$\llbracket \bV \rrbracket=\ba$, one obtains $\ba = \boldsymbol{0}$ which guarantees the injectivity of $\mathcal{H}^*$.
\end{proof}
One additional property that is needed for the analysis of sampling methods is the densness of the range of~$\mathcal{H}^*$, which is equivalent to the injectivity of $\mathcal{H}$. Unfortunately the latter cannot be guaranteed in general, and one has to impose this property as an assumption on $\Gamma$ and $\omega$.
\begin{assumption}\label{Inject-H}
We assume that $\Gamma$ and $\omega$ are such that the Herglotz operator $\mathcal{H}\!:L^2(\Omega_{\bd})^3\rightarrow H^{-1/2}(\Gamma)^3$ is injective, i.e. that $\,\mathcal{H}^*\!:\tilde{H}^{1/2}(\Gamma)^3\rightarrow L^2(\Omega_{\bd})^3$ has a dense range.
\end{assumption}
The following lemma indicates why we expect that for a given fracture geometry $\Gamma$, Assumption \ref{Inject-H} holds for all $\omega\!>\!0$ \textcolor{black}{possibly} excluding a discrete set of values without finite accumulation points.
\begin{lemma}\label{Dense-H}
Assume that \textcolor{black}{$\exs\Gamma$ contains $M\!\geqslant\!1$ (possibly disjoint) analytic surfaces~$\Gamma_m\!\subset\Gamma$, $m=1,\ldots M$, and consider the unique analytic continuation $\partial D_m$ of $\:\Gamma_m$ identifying ``interior'' domain~$D_m\!\subset\mathbb{R}^3$}. Then Assumption~\ref{Inject-H} holds as soon as \textcolor{black}{for any such~$m$, $\omega\!>\!0$} is not a ``Neumann'' eigenfrequency of the Navier equation in $D_m$, i.e. as long as every function $\boldsymbol{u}^{\textit{i}} \in H^1(D_m)^3$ satisfying
\begin{equation}\label{uiH}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \boldsymbol{u}^{\textit{i}}) \,+\, \rho \exs \omega^2 \boldsymbol{u}^{\textit{i}} ~=~ \boldsymbol{0} \quad &\textrm{in}~ D_m, \\*[1 mm]
&\bn \sip \bC \colon \! \nxs \nabla \boldsymbol{u}^{\textit{i}} ~=~ \boldsymbol{0} \quad &\textrm{on}~ \partial D_m
\end{aligned}
\end{equation}
\textcolor{black}{vanishes identically in $D_m$. Further if~$D_m$ is bounded, the real eigenfrequencies of~\eqref{uiH} form a discrete set.}
\end{lemma}
\begin{proof}
\textcolor{black}{Let~$\Gamma_m$ denote the $m$th analytic piece of~$\Gamma$. Recalling~\eqref{HW} and invoking the analyticity of $\bn\cdot\bC\colon \!\nabla \boldsymbol{u}^{\textit{i}}_\bg$ with respect to the surface coordinates on $\partial D_m$, we deduce that if $\bn \sip \bC\colon \!\nabla \boldsymbol{u}^{\textit{i}}_\bg=\boldsymbol{0}$ on $\Gamma_m\subset\partial D_m$} then
$$
\bn \cdot \bC \exs \colon \!\nabla\boldsymbol{u}^{\textit{i}}_\bg =\boldsymbol{0} \quad \mbox{ on}~ \partial D_m.
$$
This means that $\boldsymbol{u}^{\textit{i}}_\bg =\boldsymbol{0}$ in $D_m$ since $\omega$ is not a ``Neumann'' eigenvalue of the Navier equation in $D_m$. The unique continuation principle then implies that $\boldsymbol{u}^{\textit{i}}_\bg\!=\!\boldsymbol{0}$ in $\mathbb{R}^3$. Accordingly, we deduce that the Herglotz density vanishes, i.e. that~$\bg\!=\!\boldsymbol{0}$ as in the scalar case \cite{Col1992}. \textcolor{black}{The proof of discreteness of the set of real eigenfrequencies characterizing~\eqref{uiH} when~$D_m$ is bounded can be found in~\cite{Kuprad1979}, Chapter~7, Theorem~1.4}.
\end{proof}
\paragraph*{Properties of the middle operator $T$.}
\begin{lemma}\label{I{T}>0}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ in~(\ref{T}) is bounded and satisfies
\begin{equation}\label{pos-IT}
\Im \dualGA{{\boldsymbol{\varphi}}}{T{\boldsymbol{\varphi}}} <0\, \qquad \forall \exs {\boldsymbol{\varphi}} \in H^{-1/2}(\Gamma)^3:~ {\boldsymbol{\varphi}} \neq \boldsymbol{0}.
\end{equation}
\end{lemma}
\begin{proof}
The boundedness of $T$ stems from the well-posedness of problem~\eqref{Wik-GE} and classical trace theorems. Next, let ${\boldsymbol{\varphi}} \in H^{-1/2}(\Gamma)^3$ and consider $\bv$ satisfying \eqref{Wik-GE} with $\bt^{\text{\tiny f}}= {\boldsymbol{\varphi}}$. Taking $\bw = \bv$ in \eqref{Wik-GE} we get
\begin{equation}\label{T-bound}
\Im \dualGA{{\boldsymbol{\varphi}}}{T{\boldsymbol{\varphi}}} ~=~ \dualGA{\Im \bK \sip \dbv}{\llbracket\bv\rrbracket}-\Im \dualBR{{\mathcal T}_R(\bv)}{\bv}.
\end{equation}
By virtue of~\eqref{T-bound}, the claim of the theorem follows immediately from Lemma~\ref{LemmaTR} and earlier hypothesis that $\Im\bK<0$.
\end{proof}
\begin{lemma}\label{Cp-Cr}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow\tilde{H}^{1/2}(\Gamma)^3$ can be decomposed into a compact part ${\sf T}_{\!c}$ and a coercive and self-adjoint part ${\sf T}_{\!\text{\tiny o}}$ such that $T = {\sf T}_{\! c} + {\sf T}_{\!\text{\tiny o}}$. The coercive part ${\sf T}_{\!\text{\tiny o}} \colon H^{-1/2}(\Gamma)^3 \rightarrow\tilde{H}^{1/2}(\Gamma)^3$ is defined by
\begin{equation}\label{To}
{\sf T}_{\!\text{\tiny o}}({\boldsymbol{\varphi}}) ~:=~ \llbracket \bu^{\text{\tiny o}}\rrbracket \quad \text{on}~~ \Gamma,
\end{equation}
where $\bu^{\text{\tiny o}} \in H^1(\mathcal{B}_2 \backslash \Gamma)$ is a solution to
\begin{equation}\label{Wuo}
\textrm{A}(\bu^{\text{\tiny o}},\bw) = \dualGA{{\boldsymbol{\varphi}}}{\llbracket\bw \rrbracket} \quad \forall \exs \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\end{equation}
$\textrm{A}$ being the coercive sesquilinear form defined by \eqref{cv}.
\end{lemma}
\begin{proof}
We first observe from~\eqref{cv} that
\begin{equation}\label{uoG}
\begin{aligned}
&\nabla \sip (\bC \colon \! \nabla \bu^{\text{\tiny o}}) \,-\, \bu^{\text{\tiny o}} ~=~ \boldsymbol{0} \qquad &\text{in} \quad \mathcal{B}_2 \backslash \Gamma,& \\
& \bn \cdot \bC \colon \! \nxs \nabla \bu^{\text{\tiny o}} ~=~ \! -{\boldsymbol{\varphi}} \qquad &\text{on} \quad \Gamma,& \\
&\bn \cdot \bC \colon \! \nxs \nabla\bu^{\text{\tiny o}} ~=~ \textcolor{black}{\mathcal{S}_R(\bu^{\text{\tiny o}})} \qquad &\text{on} \quad \partial{B}_R&,
\end{aligned}
\end{equation}
\textcolor{black}{where $\mathcal{S}_R\!: H^{1/2}(\partial B_R)^3 \to H^{-1/2}(\partial B_R)^3$ is a Dirichlet-to-Neumann operator, $\mathcal{S}_R({\boldsymbol{\psi}}):= \bn\cdot\bC\colon\!\nabla \btu_{{\boldsymbol{\psi}}}$, stemming from the \emph{elastostatic problem} in $B_{R_\circ}\!\backslash B_R$ with Dirichlet data $\btu_{{\boldsymbol{\psi}}}={\boldsymbol{\psi}}$ on~$\partial B_{R}$ and homogeneous ``Neumann'' data $\bn\cdot\bC\colon\!\nabla \btu_{{\boldsymbol{\psi}}}=\boldsymbol{0}$ on~$\partial B_{R_\circ}$.}
Using standard trace theorems for vector fields with square-integrable divergence~\cite{Monk2003}, one finds that
\begin{equation}\label{ToTr}
\norms{\nxs{\boldsymbol{\varphi}} \nxs}_{H^{-\frac{1}{2}}(\Gamma)}\! \:=\: \norms{\nxs\bn\cdot\bC\colon\!\!\nabla\bu^{\text{\tiny o}}\!}_{H^{-\frac{1}{2}}(\Gamma)}
\,\,\leqslant\: \norms{\nxs\bn\cdot\bC\colon\!\!\nabla\bu^{\text{\tiny o}}\!}_{H^{-\frac{1}{2}}(\partial D)}
\,\,\leqslant\: c \left(\nxs\norms{\!\nabla\sip(\bC\colon\!\!\nabla \bu^{\text{\tiny o}})\!}_{L^2(D)} +
\norms{\!(\bC\colon\!\!\nabla\bu^{\text{\tiny o}})\!}_{L^2(D)} \right)
\end{equation}
for a positive constant $c$ independent from $\bu^{\text{\tiny o}}$. Thanks to the first equation in \eqref{uoG} we then deduce
$$
\norms{\nxs{\boldsymbol{\varphi}} \nxs}_{H^{-\frac{1}{2}}(\Gamma)} ~\leqslant~ c_1 \norms{\bu^{\text{\tiny o}} \!}_{H^1(D)}
$$
for some $c_1\!>\!0$ independent from $\bu^{\text{\tiny o}}$. On taking $\bw=\bu^{\text{\tiny o}}$ in \eqref{Wuo}, deploying the coercivity of $A$, \textcolor{black}{and recalling from~\eqref{cv} that $\Im A(\bv,\bv)=0$}, we find
\begin{equation}\label{ToCo}
\dualGA{{\boldsymbol{\varphi}}}{{\sf T}_{\!\text{\tiny o}}{\boldsymbol{\varphi}}} = \textrm{A}(\bu^{\text{\tiny o}},\bu^{\text{\tiny o}}) \geqslant c_2 \norms{\nxs{\boldsymbol{\varphi}} \nxs}_{H^{-\frac{1}{2}}(\Gamma)}^2
\end{equation}
for a positive constant $c_2$ independent from ${\boldsymbol{\varphi}}$, which establishes the coercivity of ${\sf T}_{\!\text{\tiny o}}$. The self-adjointness of~${\sf T}_{\!\text{\tiny o}}$ follows immediately from that of~$\textrm{A}$.
To complete the argument, consider the compactness of ${\sf T}_c \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$, given by
\[
{\sf T}_c({\boldsymbol{\varphi}}) \:=\: \llbracket\bv^c\rrbracket, \quad \bv^c \:=\: \bv - \bu^{\text{\tiny o}} \quad \text{on}~ \Gamma
\]
where $\bv$ solves~\eqref{Wik-GE}. On subtracting~(\ref{Wuo}) from~(\ref{Wik-GE}) with~$\bt^{\text{\tiny f}}={\boldsymbol{\varphi}}$, one finds that
\[
\textrm{A}(\bv^c, \bw) \:= - \textrm{B}(\bv,\bw) \qquad \forall \exs \bw \in H^1(\mathcal{B}_2 \backslash \Gamma)^3,
\]
where $\textrm{A}$ is coercive while $\text{B}$, given by~(\ref{compact}), is compact on $H^1(\mathcal{B}_2 \backslash \Gamma)^3$. As a result, the induced mapping $\bv \rightarrow \bv^c$ from $H^1(\mathcal{B}_2 \backslash \Gamma)^3$ into $H^1(\mathcal{B}_2 \backslash \Gamma)^3$ is \emph{compact}, whereby the compactness of ${\sf T}_c$ follows directly from the continuity of $\bv \in H^1(\mathcal{B}_2 \backslash \Gamma)^3$ with respect to ${\boldsymbol{\varphi}} \in H^{-\frac{1}{2}}(\Gamma)^3$ and the trace theorem.
\end{proof}
\textcolor{black}{
\begin{lemma}\label{T-invs0}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ has a bounded (and thus continuous) inverse.
\end{lemma}
\begin{proof}
The idea is to show that $T$, given by~\eqref{T}, is injective and Fredholm of index zero. The second claim follows immediately from Lemma~\ref{Cp-Cr}. To demonstrate the injectivity of~\eqref{T}, one may recall a double-layer potential representation of elastodynamic fields solving~\eqref{GE}-\eqref{KS} which demonstrates that for any ${\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)$, one has
\[
\bv({\boldsymbol{\varphi}})(\bxi) ~=~ \int_{\Gamma} T({\boldsymbol{\varphi}}) \cdot \bfT(\bxi,\by) \, \text{d}S_{\by}, \qquad
\bfT(\bxi,\by) ~=~ \bn(\by)\cdot\bSig(\bxi,\by), \qquad \bxi \in \mathbb{R}^3 \backslash \Gamma,
\]
where $\dbv=T({\boldsymbol{\varphi}})$ on~$\Gamma$ thanks to the fundamental property of double-layer potentials. Thus, on assuming that there exists ${\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)$ so that $T({\boldsymbol{\varphi}})=\boldsymbol{0}$, one finds that $\bv\!=\!\boldsymbol{0}$ in~$\mathbb{R}^3\backslash\Gamma$ and consequently, by the second of~\eqref{GE} and trace theorems, that $\|{\boldsymbol{\varphi}}\|_{H^{-1/2}(\Gamma)}=\|\bn\sip\bC\!:\!\nabla\bv\|_{H^{-1/2}(\Gamma)}=0$.
\end{proof}}
\textcolor{black}{
\begin{lemma}\label{T-coerc0}
Operator $T \colon H^{-1/2}(\Gamma)^3 \rightarrow \tilde{H}^{1/2}(\Gamma)^3$ is coercive, i.e. there exists constant $c\!>\!0$ independent of~${\boldsymbol{\varphi}}$ such that
\begin{equation}\label{co-T0}
|\langle {\boldsymbol{\varphi}}, \, T ({\boldsymbol{\varphi}}) \rangle| \,\,\geqslant\,\, \textrm{c} \nxs \norms{{\boldsymbol{\varphi}}}_{H^{-\frac{1}{2}}(\Gamma)}^2, \qquad \forall{\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)^3.
\end{equation}
\end{lemma}
\begin{proof}
Lemma~\ref{I{T}>0} demonstrates that the duality product $\big\langle {\boldsymbol{\varphi}} ,\, T({\boldsymbol{\varphi}}) \big\rangle\in\mathbb{C}\setminus(-\infty,\infty)$ for all nonzero ${\boldsymbol{\varphi}}\!\in H^{-1/2}(\Gamma)^3$. Due to Lemma~\ref{Cp-Cr}, on the other hand, decomposition $T = {\sf T}_{\! c} + {\sf T}_{\!\text{\tiny o}}$ exists where ${\sf T}_{\! c}$ is compact and~${\sf T}_{\!\text{\tiny o}}$ is such that $\langle \exs {\boldsymbol{\varphi}}, {\sf T}_{\!\text{\tiny o}}({\boldsymbol{\varphi}}) \exs \rangle\in \mathbb{R}$ satisfies the coercivity condition~(\ref{ToCo}) $\forall \exs {\boldsymbol{\varphi}}\!\in\!H^{-1/2}(\Gamma)^3$. With such results in place, claim~\eqref{co-T} follows immediately by Lemma~1.17 in~\cite{Kirsch2008}.
\end{proof}}
\section{Application of sampling methods} \label{SSA}
\subsection{Linear sampling method (LSM)}
The essential idea behind the LSM~\cite{Fiora2003} and also the factorization method (FM)~\cite{Bouk2013} for geometrical obstacle reconstruction stems from the particular nature of an approximate solution, \textcolor{black}{$\bg =\bg_p\oplus\bg_s$}, to the far-field equation
\begin{equation}\label{FF}
F \bg ~=~ \boldsymbol{\Phi}_L^\infty, \qquad F ~=~ \mathcal{G} \mathcal{H} ~=~ \mathcal{H}^* \exs T \exs \mathcal{H},
\end{equation}
where $\boldsymbol{\Phi}_L^\infty$ is the far-field pattern \textcolor{black}{of a trial radiating field}, see Definition~\ref{phi-infinity}. In this setting, the behavior of $\bg$ in the sampling region is exposed by characterizing the range of $\mathcal{G}$ or $\mathcal{H}^*$, \textcolor{black}{which then forms the basis for approximating the characteristic function of a scatterer}. This section presents an adaptation of the key LSM results for the problem of elastic-wave imaging of heterogeneous fractures, which provides a foundation for the GLSM developments in Section~\ref{GLSMM}.
\begin{defn} \label{phi-infinity}
\textcolor{black}{With reference to~\eqref{Hstar}, for every admissible FOD profile $\ba\!\in\!\tilde{H}^{1/2}(L)$ specified over a smooth, non-intersecting trial fracture $L\!\subset\!\mathcal{B}_2$, the induced far-field pattern $\boldsymbol{\Phi}_L^\infty \colon \tilde{H}^{1/2}(L) \rightarrow L^2(\Omega_{\bd})^3$ is given by}
\begin{equation}\label{Phi-inf}
\begin{aligned}
\boldsymbol{\Phi}_{L}^\infty(\ba)(\hat\bxi) ~=~ - \Big( \, & \text{\emph{i}} k_p \, \hat\bxi \exs \int_L \, \Big\lbrace \lambda \exs (\ba \sip \bn) \,+\, 2\mu \exs (\bn \sip \hat\bxi) ( \ba \sip \hat\bxi) \Big\rbrace \, e^{-\text{\emph{i}}k_p \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \\*[1 mm]
& \textcolor{black}{\oplus}~ \text{\emph{i}} k_s \, \hat\bxi\times\! \int_L \Big\lbrace \mu \exs(\ba \times \hat\bxi)(\bn\sip\hat\bxi) \,+\, \mu \exs (\bn \times \hat\bxi) (\ba \sip \hat\bxi) \Big\rbrace \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \Big).
\end{aligned}
\end{equation}
and $\bn$ is the unit normal on~$L$.
\end{defn}
\begin{rem}\label{LSMrem}
\textcolor{black}{On the basis of Definition~\ref{phi-infinity}, one may interpret the LSM reconstruction philosophy as follows. Let ${\sf L}\!\subset\!\mathbb{R}^3$ (containing the origin) denote a reference fracture surface whose characteristic size is small relative to the length scales describing the forward scattering problem, and let $L=\bz\!+\bR{\sf L}$ where $\bz\!\in\mathbb{R}^3$ and $\bR\!\in\!U(3)$ is a unitary rotation matrix. Given an admissible FOD profile $\ba\!\in\!\tilde{H}^{1/2}({\sf L})$, solving the far-field equation~\eqref{FF} over a grid of trial pairs $(\bz,\bR)$ sampling $\mathbb{R}^3\!\times U(3)$ is simply an effort to probe the far-field kernel~\eqref{w-inf} -- through synthetic rearrangement of the illuminating plane waves -- for fingerprints in terms of~$\boldsymbol{\Phi}_L^\infty$. As shown by Theorems~\ref{TR2}, \ref{GLSM1} and~\ref{GLSM2}, such fingerprint is found in the data if and only if~$L\subset\Gamma$. Otherwise, the norm of any approximate solution to~(\ref{FF}) can be made arbitrarily large, which then provides a criterion for the reconstruction of~$\Gamma$.}
\end{rem}
\begin{theorem}\label{TR1}
Provided that $\omega$ is \emph{not} a ``Neumann'' eigenvalue of the Navier equation~(\ref{uiH}) and that~$\bK^{-1} \!\! \in \! L^{\infty}(\Gamma)$, for \emph{every} smooth and non-intersecting trial crack $L\subset\mathcal{B}_2$ and some density function $\ba(\bxi)\!\in\!\tilde{H}^{1/2}(L)$, one has
\[
\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*) ~~ \iff ~~ L \subset \Gamma.
\]
\end{theorem}
\begin{proof}
Consider the following:
\begin{itemize}
\item If~$L \subset \Gamma$, then $\tilde{H}^{1/2}(L)^3 \subset \tilde{H}^{1/2}(\Gamma)^3$. By extending the domain of $\ba\in\tilde{H}^{1/2}(L)^3$ from $L$ to $\Gamma$ through zero padding, one immediately obtains $\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*)$ thanks to~\eqref{Hstar} and~\eqref{Phi-inf}.
\item Assume that~$L\not\subset\Gamma$ and that $\boldsymbol{\Phi}_L^\infty\!\in\!Range(\mathcal{H}^*)$. Then there exists $\bb\!\in\!\tilde{H}^{1/2}(\Gamma)^3$ such that
\begin{equation} \notag
\begin{aligned}
\boldsymbol{\Phi}_L^\infty(\bb)(\hat\bxi) ~=~ -\Big(& \text{\emph{i}} k_p \, \hat\bxi \exs \int_\Gamma \, \Big\lbrace \lambda \exs (\bb \sip \bn) \,+\, 2\mu \exs (\bn \sip \hat\bxi) ( \bb \sip \hat\bxi) \Big\rbrace e^{-\text{\emph{i}}k_p \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \\
& \textcolor{black}{\oplus}~ \text{\emph{i}} k_s \, \hat\bxi \exs \times \int_{\Gamma} \Big\lbrace \mu \exs(\bb \times \hat\bxi)(\bn \sip \hat\bxi) \,+\, \mu \exs (\bn \times \hat\bxi) (\bb \sip \hat\bxi) \Big\rbrace \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \by} \,\, \text{d}S_{\by} \Big),
\end{aligned}
\end{equation}
associated with the layer potential
\begin{equation}\label{Dlpb}
\boldsymbol{\Phi}_{\exs \Gamma}(\bxi) ~=~ \int_{\Gamma} \bb(\by) \cdot \bfT(\bxi,\by) \, \text{d}S_{\by}, \qquad \bfT(\bxi,\by) ~=~ \bn(\by)\cdot\bSig(\bxi,\by), \qquad \bxi \in \mathcal{B}_2 \backslash \Gamma.
\end{equation}
On the other hand, owing to Definition~\ref{phi-infinity} of $\boldsymbol{\Phi}_L^\infty(\hat\bxi)$, potential $\boldsymbol{\Phi}_{\exs \Gamma}(\bxi)$ should coincide with
\begin{equation}\label{Pb}
\boldsymbol{\Psi}_{\! L}(\bxi) ~=~ \int_L \ba(\by) \cdot \bfT(\bxi,\by) \,\, \text{d}S_{\by}, \qquad \bxi \in \mathcal{B}_2 \backslash L,
\end{equation}
over $ \bxi \in \mathcal{B}_2 \backslash (L \cup \Gamma)$. Now, let $\Gamma\not\ni\bxio\!\in L$ and let $\mathcal{B}_\epsilon$ be a small ball centered at $\bxio$ such that $\mathcal{B}_\epsilon \cap \Gamma = \emptyset $. In this case $\boldsymbol{\Phi}_{\exs \Gamma}$ is analytic in $\mathcal{B}_\epsilon$, while $\boldsymbol{\Psi}_{\! L}$ has a singularity at $\bxio\!\in\mathcal{B}_\epsilon$ -- which by contradiction completes the proof.
\end{itemize}
\end{proof}
On the basis of the above result, one arrives at the following statement which inspires most of the LSM-based indicator functionals.
\begin{theorem}\label{TR2}
Under the assumptions of \textcolor{black}{Lemma~\ref{Dense-H} and~Theorem~\ref{TR1}},
\begin{itemize}
\item~If $L\!\subset\!\Gamma$, there exists a Herglotz density vector $\bg_\epsilon^L\!\in L^2(\Omega_{\bd})^3$ such that $\|F\bg_\epsilon^L-\boldsymbol{\Phi}_L^\infty\|_{L^2(\Omega_{\bd})} \leqslant\epsilon$ and $\limsup\limits_{\epsilon \rightarrow 0} \|\mathcal{H}\bg_\epsilon^L\|_{H^{-1/2}(\Gamma)}<\infty$.
\item~If $L \not\subset \Gamma$, then $\forall \bg_\epsilon^L\!\in L^2(\Omega_{\bd})^3$ such that $\norms{\nxs F\bg_\epsilon^L-\boldsymbol{\Phi}_L^\infty \nxs}_{L^2(\Omega_{\bd})} \, \leqslant\epsilon$, one has $\,\lim\limits_{\epsilon \rightarrow 0} \norms{\mathcal{H}\bg_\epsilon^L}_{H^{-1/2}(\Gamma)} \,\,=\infty$.
\end{itemize}
\end{theorem}
\begin{proof}
\begin{description}
Let us first assume $L \subset \Gamma$, whereby $\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*)$ thanks to Theorem~\ref{TR1}. Then, by definition, there exists $\ba^L\!\in\!\overline{Range(T)}$ such that $\mathcal{H}^*\ba^L = \boldsymbol{\Phi}_L^\infty$. By invoking \textcolor{black}{Lemma~\ref{T-invs0}} on the boundedness i.e.~continuity of $T^{-1}$ and Lemma~\ref{H*p} which \textcolor{black}{(by the injectivity of~$\mathcal{H}^*$)} guarantees the range denseness of $\mathcal{H}$, one finds that $\forall \epsilon>0$, $\exists \, \bg_\epsilon^L\!\in L^2(\Omega_{\bd})^3$ such that $\norms{T^{-1} \ba^L\!-\!\mathcal{H}\bg_\epsilon^L}_{H^{-1/2}(\Gamma)} \,\leqslant \epsilon$. \textcolor{black}{Thanks to (i) the continuity of~$\mathcal{H}^* T$ and (ii) the fact that~$\ba^L\!\in\tilde{H}^{1/2}(\Gamma)^3$, this establishes the first part of the claim}.
Next, consider the case where $L\!\not\subset\!\Gamma$ and consequently $\boldsymbol{\Phi}_L^\infty \not\in Range(\mathcal{H}^*)$ \textcolor{black}{by Theorem~\ref{TR1}}. Then, thanks to \textcolor{black}{Lemma~\ref{Dense-H} which implies the denseness of~$Range(\mathcal{H}^*)$}, \textcolor{black}{for every $\epsilon\!>\!0$ and some regularization parameter $0\!<\!\alpha\!<\!C\epsilon$ where~$C$ is a constant independent of~$\epsilon$}, a nearby solution $\ba^L_\epsilon \in \tilde{H}^{1/2}(\Gamma)^3$ can be built e.g.~via Tikhonov regularization~\cite{Kress1999} such that \textcolor{black}{$\norms{\!\boldsymbol{\Phi}_L^\infty-\mathcal{H}^* \ba_\epsilon^L\!}_{L^2(\Omega_{\bd})} \,\leqslant \nxs \epsilon\,$} and $\,\lim\limits_{\epsilon \rightarrow 0} \norms{\nxs \ba^L_\epsilon \nxs}_{\tilde{H}^{1/2}(\Gamma)} \,=\infty$ -- due to the compactness of $\mathcal{H}^*$ \textcolor{black}{established in Lemma~\ref{H*p}}. At this point, the same argument as in the first part of the proof -- deploying the continuity of $T^{-1}$ and the range denseness of~$\mathcal{H}$ -- can be used to show establish the second claim.
\end{description}
\end{proof}
\vspace{-7 mm}
\subsection{Factorization method (FM)}
To facilitate the ensuing developments, we recall elements of the factorization method~\cite{Kirsch2008} as they pertain to our inverse problem.
\begin{defn}\label{DFs}
The self-adjoint operator $F_\sharp \colon L^2(\Omega_{\bd})^3 \rightarrow L^2(\Omega_{\bd})^3$ is defined by
\begin{equation}\label{Fs}
F_\sharp \,\colon \!\!\!=\, |\Re{F}| \:+\: \Im{F},
\vspace{-2 mm}
\end{equation}
where $F\!:L^2(\Omega_{\bd})^3 \to L^2(\Omega_{\bd})^3$ is given by~\eqref{ffo2}, and
\begin{equation}\label{ReIm}
\Re{F} \,=\, \tfrac{1}{2} (F+F^*), \qquad \Im{F} \,=\, \tfrac{1}{2 \textrm{\emph{i}}} (F-F^*).
\end{equation}
\end{defn}
\begin{rem}\label{TsD}
In line with decomposition~(\ref{fact}) of the far-field operator $F$, there exists factorization
\begin{equation}\label{facts}
F_\sharp ~=~ \mathcal{H}^* \exs T_\sharp \exs \mathcal{H}
\end{equation}
of~\eqref{Fs}, where the middle operator $T_\sharp \colon H^{-1/2}(\Gamma)^3\rightarrow\tilde{H}^{1/2}(\Gamma)^3$ is given by
\begin{equation}\label{Tsdef}
T_\sharp :=\, \Re{T}(Q^+\!-Q^-) +\, \Im{T};
\end{equation}
$Q^+$ and $Q^-\!$ are bounded projectors such that $Q^{+}\!+Q^{-}=I$; $Q^{+}\!-Q^{-}$ is an isomorphism, and $Q^-\!$ has a finite rank. See Theorem 2.15 in~\cite{Kirsch2008} for derivation.
\end{rem}
\vspace*{-5mm}
\textcolor{black}{
\begin{theorem}\label{TR3}
Under the assumptions of Theorem~\ref{TR1}, operator $F_\sharp$ in~\eqref{Fs} has the following properties:
\begin{itemize}
\item~Operator $F_\sharp$ is positive.
\item~The ranges of~$\mathcal{H}^* \colon \tilde{H}^{1/2}(\Gamma)^3\rightarrow L^2(\Omega_{\bd})^3$ and $F_\sharp^{1/2}\colon L^2(\Omega_{\bd})^3\rightarrow L^2(\Omega_{\bd})^3$ coincide.
\item~$\boldsymbol{\Phi}_L^\infty \in Range(F_\sharp^{1/2}) ~~ \iff ~~ L \subset \Gamma$.
\end{itemize}
\end{theorem}
\begin{proof}
The first two claims follow directly from Theorem~2.15 in~\cite{Kirsch2008}, its extended version (Theorem~3.2) in~\cite{Bouk2013}, Lemma~\ref{H*p} Lemma~\ref{I{T}>0}, and Lemma~\ref{Cp-Cr}. With such result in place, the last claim is immediately established by Theorem~\ref{TR1}.
\end{proof}}
\vspace*{-5mm}
\textcolor{black}{
\begin{lemma}\label{T-invs}
Operator~$T_\sharp \colon H^{-1/2}(\Gamma)^3\rightarrow\tilde{H}^{1/2}(\Gamma)^3$ in the factorization~\eqref{facts} has the following properties:
\begin{itemize}
\item~$T_\sharp$ has a bounded (and thus continuous) inverse.
\item~$T_\sharp$ is selfadjoint and is positively coercive, i.e. there exists a constant~$c\!>\!0$ independent of~${\boldsymbol{\varphi}}$ so that
\begin{equation}\label{co-T}
\big({\boldsymbol{\varphi}}, \, T_\sharp ({\boldsymbol{\varphi}}) \big)_{H^{-\frac{1}{2}}(\Gamma)} \,\,\geqslant\,\, c \nxs \norms{{\boldsymbol{\varphi}}}_{H^{-\frac{1}{2}}(\Gamma)}^2, \qquad
\forall \,{\boldsymbol{\varphi}}\in H^{-1/2}(\Gamma)^3.
\end{equation}
\end{itemize}
\end{lemma}
\begin{proof}
See Appendix~A in~\cite{Bouk2013} and the proof of Theorem 2.15, part E in~\cite{Kirsch2008}.
\end{proof}}
\textcolor{black}{On the basis of Theorem~\ref{TR2}, one sees that $F_\sharp^{1/2}$ can be used to characterize $\Gamma$ from the far-field measurements. In what follows, it is in particular shown that the GLSM cost functionals based on $F_\sharp$ (i)~are convex, (ii)~have closed-form minimizers, and (iii)~enable fast and robust reconstruction of $\Gamma$ -- especially when the data (and thus the far-field operator) are contaminated by noise.}
\subsection{Generalized Linear Sampling Method (GLSM)} \label{GLSMM}
\noindent Theorem~\ref{TR2} of the linear sampling method poses two fundamental challenges in that:~i) the featured anomaly indicator $\norms{\!\mathcal{H}\bg_\epsilon^L\!}_{H^{-1/2}(\Gamma)}$ inherently depends on the unknown fracture support $\Gamma$, and ii) construction of the Herglotz density vector $\bg_\epsilon^L \in L^2(\Omega_{\bd})^3$ is implicit in the theorem~\cite{Audibert2014}. Conventionally, these issues are addressed by replacing $\norms{\!\mathcal{H}\bg_\epsilon^L\!}_{H^{-1/2}(\Gamma)}\!$ with $\norms{\nxs\bg_\epsilon^L\!}_{L^2(\Omega_{\bd})}$ which is, in turn, computed by way of Tikhonov regularization~\cite{Kress1999}. Such treatment, however, has proven to be particularly sensitive to perturbations in the data due to e.g. measurement errors.
To help meet the challenge, the GLSM takes advantage of the second factorization (\ref{fact}) of the far-field operator and the mathematical properties of its components to properly construct a \emph{stable} approximate solution to the far-field equation (\ref{FF}). This is accomplished through a \emph{sequence} of \textcolor{black}{penalized least-squares problems} where the principal ingredient of the penalty term is $\norms{\!\mathcal{H}\bg_\epsilon^L\!}_{H^{-1/2}(\Gamma)}$, reformulated in a computable way in terms of the far-field operator~$F$. More specifically, by invoking factorizations~(\ref{fact}) and~(\ref{facts}), one may observe that
\[
\begin{aligned}
&(\bg_\epsilon^L, \exs F\bg_\epsilon^L)_{L^2(\Omega)} \,\exs ~=~ \big \langle \mathcal{H}\bg_\epsilon^L, \, T\mathcal{H}\bg_\epsilon^L \big \rangle_{\Gamma}, \\*[1 mm]
& (\bg_\epsilon^L, \exs F_\sharp \exs \bg_\epsilon^L)_{L^2(\Omega)} ~=~ \big \langle \mathcal{H}\bg_\epsilon^L, \, T_\sharp \exs \mathcal{H}\bg_\epsilon^L \big \rangle_{\Gamma}, \qquad \forall \bg_\epsilon^L \in L^2(\Omega_{\bd})^3
\end{aligned}
\]
where~$(\cdot,\cdot)_{L^2(\Omega)}:=(\cdot,\cdot)_{L^2(\Omega)^3}$ denotes the usual~$L^2$ inner product on~$\Omega$. \textcolor{black}{Then, thanks to the coercivity of the middle operator $T$ (see Lemma~\ref{T-coerc0})}, quantity $|(\bg_\epsilon^L,F\bg_\epsilon^L)_{L^2(\Omega)}|$ -- which is computable without prior knowledge of~$\Gamma$ -- may be safely substituted for $\norms{\!\mathcal{H}\bg_\epsilon^L\!}^2_{H^{-1/2}(\Gamma)}$ \textcolor{black}{in constructing a penalty term for the GLSM cost functional. Similarly, the positive coercivity $T_\sharp$ (See Lemma~\ref{T-invs}) and factorization~\eqref{facts} of $F_\sharp$ demonstrate that $|(\bg_\epsilon^L, \, F_\sharp \exs \bg_\epsilon^L)_{L^2(\Omega_{\bd})}| = \,\, \norms{F_\sharp^{1/2}\bg_\epsilon^L}^2$ may serve as a replacement for $\norms{\!\mathcal{H}\bg_\epsilon^L\!}^2_{H^{-1/2}(\Gamma)}$}, giving birth to a \emph{convex} GLSM cost functional whose minimizer can be computed without iterations. This shines light on the GLSM approach to elastodynamic reconstruction of heterogeneous fractures, whose specificities are presented next.
\paragraph*{GLSM cost functional.} \label{cost}
\begin{itemize}
\item~\emph{Unperturbed (noise-free) operators.}~Let $\alpha\!>\!0$ be a regularization parameter, and consider the far-field pattern~$\boldsymbol{\Phi}_L^\infty\!\in L^2(\Omega_{\bd})^3$ as in Definition~\ref{phi-infinity}. Then the GLSM cost functional is defined by a sequence of penalized least-squares misfit functionals $J_\alpha(\boldsymbol{\Phi}_L^\infty;\, \cdot)\colon \, L^2(\Omega_{\bd})^3 \rightarrow \mathbb{R}$, namely
\begin{equation} \label{J-alph}
J_\alpha(\boldsymbol{\Phi}_L^\infty; \, \bg) ~ \colon \!\!\! =~ \! \norms{\nxs F\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+\,\,\exs \alpha \nxs \norms{F_\sharp^{\frac{1}{2}} \bg}^2, \qquad \bg \in L^2(\Omega_{\bd})^3,
\end{equation}
whose minimizers~$\bg^L_\alpha \in L^2(\Omega_{\bd})^3$ can be computed \emph{non-iteratively} by solving
\begin{equation} \label{min-J}
F^*(F\bg^L_\alpha \,-\, \boldsymbol{\Phi}_L^\infty) ~+~ \alpha \exs (F_\sharp^{\frac{1}{2}} )^* F_\sharp^{\frac{1}{2}} \, \bg^L_\alpha~=~ \boldsymbol{0}.
\end{equation}
For completeness, a more general form $\mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty; \, \cdot)\colon \, L^2(\Omega_{\bd})^3 \rightarrow \mathbb{R}$ of the GLSM cost functional, namely
\begin{equation} \label{fJ-alph}
\mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\, \bg) ~ \colon \!\!\! =~ \! \norms{\nxs F\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+\,\,\exs \alpha \exs | ( \bg, \exs F \bg ) | , \qquad \bg \in L^2(\Omega_{\bd})^3,
\end{equation}
is also considered. \textcolor{black}{Note that~\eqref{fJ-alph} does not demand $F_\sharp$ to be applicable (see Theorem~\ref{TR3}), and thus may cater for a wider class of contact laws, $\mathcal{L}\dbv$, over the fracture surface in~(\ref{GE})}.
\begin{rem}\label{App-sol}
In general, $\mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\,\bg)$ does not have a minimizer; however, one may define
\[
j_\alpha(\boldsymbol{\Phi}_L^\infty) ~ \colon \!\!\! = \inf\limits_{\bg \in L^2(\Omega_{\bd})^3} \! \mathcal{J}_\alpha(\boldsymbol{\Phi}_L^\infty; \, \bg).
\]
Thanks to the range denseness of $F$ (see Lemma~\ref{FF_op}), one has that $j_\alpha \rightarrow 0$ as $\alpha \rightarrow 0$. Accordingly, an optimized nearby solution can be constructed by following the algorithm described in~\cite{Audibert2014}.
\end{rem}
\item~\emph{Perturbed operators.}~When the measurements are contaminated with noise (e.g.~sensing errors, fluctuations in the medium properties), one has to deal with noisy operators $F^\delta\!$ and $F_\sharp^\delta$ satisfying
\begin{equation}\label{Ns-op}
\norms{\nxs F^\delta - F \nxs} \,\,\, \leqslant \,\, \delta , \qquad \norms{\nxs F^\delta_\sharp - F_\sharp \nxs} \,\,\, \leqslant \,\, \delta ,
\end{equation}
where $\delta\!>\!0$ is a measure of perturbation in data -- independent of $F$ and $F_\sharp$. Assuming that $F^\delta\!$ and $F_\sharp^\delta$ are compact, a regularized version $J_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\,\cdot)\colon \, L^2(\Omega_{\bd})^3 \rightarrow \mathbb{R}$ of the GLSM cost functional is defined in spirit of the Tikhonov regularization method as
\begin{equation} \label{RJ-alph}
J_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\, \bg) ~ \colon \!\!\! =~ \! \norms{\nxs F^\delta\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 + \,\, \alpha \exs \big(\!\norms{\nxs(F^\delta_\sharp)^{\frac{1}{2}} \exs \bg \nxs}^2 +\,\,\exs \delta \! \norms{\nxs \bg \nxs}^2 \! \big), \qquad \bg\in L^2(\Omega_{\bd})^3.
\end{equation}
Note that $J_\alpha^\delta$ is again convex and that its minimizer $\bg^L_{\alpha,\delta} \in L^2(\Omega_{\bd})^3$ solves the linear system
\begin{equation} \label{min-RJ}
F^{\delta *}(F^\delta \bg^L_{\alpha,\delta} \,-\, \boldsymbol{\Phi}_L^\infty) ~+~ \alpha \exs \big( \exs (F_\sharp^\delta)^{\nxs\frac{1}{2}*} (F_\sharp^\delta)^{\nxs\frac{1}{2}} \, \bg^L_{\alpha,\delta} \,+\, \delta \, \bg^L_{\alpha,\delta} \exs \big) ~=~ \boldsymbol{0}.
\end{equation}
In this vein, the (regularized) cost functional affiliated with the general form~(\ref{fJ-alph}) may be recast as
\begin{equation}\label{RfJ-alph}
\mathcal{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\,\bg) ~ \colon \!\!\! =~ \! \norms{\nxs F^\delta\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+~ \alpha \exs \big( \exs | (\bg, \, F^\delta \bg ) | \,+\, \delta \! \norms{\nxs \bg \nxs}^2 \! \big), \qquad \bg\in L^2(\Omega_{\bd})^3.
\end{equation}
\begin{rem}
\textcolor{black}{In (\ref{RJ-alph}) and (\ref{RfJ-alph}), $\delta$ signifies both a measure of perturbation in $F$ and a regularization parameter that, along with $\alpha$, is designed to create a robust fracture indicator functional via a sequence of the GLSM minimizers (see the proof of Theorem~\ref{GLSM2}).}
\end{rem}
\end{itemize}
With the above definitions in place, the main GLSM theorems are based on the following lemmas.
\begin{lemma}\label{comp_G}
Operator $\mathcal{G}=\mathcal{H}^* T \colon H^{-1/2}(\Gamma)^3 \rightarrow L^2(\Omega_{\bd})^3$ is compact over $H^{-1/2}(\Gamma)^3$.
\end{lemma}
\begin{proof}
The claim follows immediately from Lemmas~\ref{H*p} and~\ref{I{T}>0} establishing, respectively, the compactness of $\mathcal{H}^*$ and the boundedness of $T$.
\end{proof}
\begin{lemma}\label{FF_op}
The far-field operator $F \colon L^2(\Omega_{\bd})^3 \rightarrow L^2(\Omega_{\bd})^3$ is injective, compact and, under the assumptions of Lemma~\ref{Dense-H}, has a dense range.
\end{lemma}
\begin{proof}
\emph{Injectivity.}~Let $F(\bg) = \boldsymbol{0}$. \textcolor{black} {Then, recalling the factorization $F = \mathcal{H}^* T \mathcal{H}$ and the injectivity of $\mathcal{H}^*$ and~$T$ (due respectively to Lemma~\ref{Dense-H} and Lemma~\ref{T-invs0}), one finds that $\mathcal{H}(\bg):= \bn\cdot\bC\colon\!\nabla\boldsymbol{u}^{\textit{i}}_\bg=\boldsymbol{0}$ on~$\Gamma$. Under the assumptions of Lemma~\ref{Dense-H}, this requires that $\boldsymbol{u}^{\textit{i}}_\bg=\boldsymbol{0}$ in~$\mathbb{R}^3$, i.e. that~$\bg=\boldsymbol{0}$ which establishes the first claim.}
\emph{Compactness.}~ The compactness of~$F$ follows immediately from the compactness of $\mathcal{H}^*$ -- and thus that of $\mathcal{H}$ (Lemma~\ref{H*p}), and the boundedness of~$T$ (Lemma~\ref{I{T}>0}).
\emph{Range densenes.}~\textcolor{black}{This claim is conveniently verified by establishing the injectivity of~$F^*$. To this end, recall~(\ref{ffo2}) and consider the $L^2$-inner product
\begin{equation}\label{Dp-F*}
\big(F(\bg),\ba\big)_{L^2(\Omega)} ~=~ \int_{\Omega} \bar\ba(\hat\bxi) \cdot \bv_{\bg_\Omega}^\infty(\hat\bxi) \, \text{d}S_{\hat\bxi} ~=~ \int_{\Omega_{\bd}} \bg(\bd) \cdot \overline{\int_{\Omega} \bW^{\infty*}(\bd,\hat\bxi)\cdot \ba(\hat\bxi) \, \text{d}S_{\hat\bxi}} \,\, \text{d}S_{\bd},
\end{equation}
where~$\ba\in L^2(\Omega)^3$. Thanks to the reciprocity identity~(\ref{W-recip}), inner product~\eqref{Dp-F*} exposes the adjoint far-field operator as
\begin{equation}\label{F*}
F^*(\ba)(\bd) ~=~ \int_{\Omega} \bW^{\infty*}(\bd,\hat\bxi)\cdot \ba(\hat\bxi) \, \text{d}S_{\hat\bxi} ~=~
\overline{\int_{\Omega} \bW^{\infty}(\hat\bxi,-\bd)\cdot \overline\ba(-\hat\bxi) \, \text{d}S_{\hat\bxi}} ~=~ \overline{F}(\tilde{\ba})(-\bd), \quad~ \bd\in\Omega,
\end{equation}
where $\tilde\ba(\hat\bxi)\!:=\!\overline{\ba}(-\hat\bxi)$ on~$\Omega$. Owing to the injectivity of $F$, one finds from~\eqref{F*} that setting $F^*(\ba)\!=\!\boldsymbol{0}$ necessitates~$\tilde\ba=\boldsymbol{0}$ and thus $\ba=\boldsymbol{0}$.}
\end{proof}
We are now in position to establish the main result of the GLSM approach, given by Theorem~\ref{GLSM1} and Theorem~\ref{GLSM2}, catering for the elastodynamic reconstruction of heterogeneous fractures.
\begin{theorem} \label{GLSM1}
\textcolor{black}{Consider the GLSM cost functional $\mathfrak{J}_\alpha$ unifying~\eqref{J-alph} and~\eqref{fJ-alph} with unperturbed operators $F^\delta$ and~$F_\sharp^\delta$, namely}
\begin{equation}\label{GCf}
\mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\,\bg) ~:=~ \norms{\nxs F\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}_{L^2(\Omega)}^2 \,+~ \alpha\,|(\bg, B\bg)|, \qquad \bg\in L^2(\Omega_{\bd})^3,
\end{equation}
\textcolor{black}{where $\alpha>0$ and~$B$, denoting either~$F$ or~$F_\sharp$, admits the factorization}
\begin{equation}\label{Bdf}
B ~=~ \mathcal{H^*} \exs \mathfrak{T} \exs \mathcal{H}, \qquad \mathfrak{T} ~=~ T,\,T_\sharp.
\end{equation}
Since $\mathfrak{J}_\alpha\geqslant 0$, define the infimum
\[
\mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty) ~\colon\!\!=~\! \inf\limits_{\bg \in L^2(\Omega_{\bd})^3} \! \mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\bg),
\]
and let $\bg_\alpha^L \in L^2(\Omega_{\bd})^3$ denote a nearby solution such that
\[
\mathfrak{J}_{\alpha}(\boldsymbol{\Phi}_L^\infty;\bg_\alpha^L) \,\,\leqslant \,\, \mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty) + \mu \alpha,
\]
$\mu>0$ being a constant independent of $\alpha$. Then,
\[
\boldsymbol{\Phi}_L^\infty \in Range(\mathcal{H}^*) ~\iff~ \Big\{\limsup\limits_{\alpha \rightarrow 0} |( \bg_\alpha^L, B \bg_\alpha^L ) | < \infty ~\iff~ \liminf\limits_{\alpha \rightarrow 0} |( \bg_\alpha^L, B \bg_\alpha^L ) | < \infty\Big\}.
\]
\end{theorem}
\begin{proof} See the proof of Theorem~3 in~\cite{Audibert2014}, synthesized in~\ref{GLSM*pruf} using present notation.
\end{proof}
\begin{lemma}\label{min-Jad}
Consider the regularized GLSM cost functional $\mathfrak{J}_\alpha^\delta$ unifying~\eqref{RJ-alph} and~\eqref{RfJ-alph} with perturbed operators $F^\delta$ and~$F_\sharp^\delta$, namely
\begin{equation}\label{GCfn}
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg) ~:=~ \norms{\nxs F^\delta\bg\,-\,\boldsymbol{\Phi}_L^\infty \nxs}^2 \,+~ \alpha \exs \big( \exs | ( \bg, B^\delta \bg ) | \,+\, \delta \! \norms{\nxs \bg \nxs}^2 \! \big), \qquad \bg \in L^2(\Omega_{\bd})^3
\end{equation}
where $\alpha,\beta\!>\!0$ and $B^\delta$ denotes either~$F^\delta$ or~$F^\delta_\sharp$. Assuming that~$B^\delta$ is compact, $\mathfrak{J}_\alpha^\delta$ has a minimizer $\bg_{\alpha,\delta}^L \in L^2(\Omega_{\bd})^3$ satisfying
\begin{equation}\label{limlim}
\lim\limits_{\alpha \rightarrow 0} \limsup\limits_{\delta \rightarrow 0} \exs \mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\,\bg_{\alpha,\delta}^L) ~=~ 0.
\end{equation}
\end{lemma}
\begin{proof}
\emph{Existence of a minimizer.}~For any $\alpha, \delta > 0$ and $\boldsymbol{\Phi}_L^\infty\!\in L^2(\Omega_{\bd})^3$ given by~\eqref{Phi-inf}, any sequence $(\bg^n)$ constructed to minimize $\mathfrak{J}_\alpha^\delta$ is bounded in $L^2(\Omega_{\bd})^3$, and thus weakly convergent to some $\bg_{\alpha,\delta}^L \in L^2(\Omega_{\bd})^3$. Thanks to the lower semi-continuity of a norm with respect to the weak convergence and the postulated compactness of $B^\delta$, one has
\begin{equation}\label{ming}
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta}^L) \,\, \leqslant \,\, \liminf\limits_{n \rightarrow \infty} \exs \mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg^n) \,\, \leqslant \,\, \!\! \inf\limits_{\bg \in L^2(\Omega_{\bd})^3} \! \mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg),
\end{equation}
which proves that $\bg_{\alpha,\delta}^L$ is a minimizer of $\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg)$ in~$L^2(\Omega_{\bd})^3$.
\emph{Limiting behavior.} \textcolor{black}{ Let us first observe from~\eqref{Ns-op}, \eqref{GCf} and~\eqref{GCfn} that
\begin{equation}\label{JadJa}
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg) \,\,\leqslant \,\, \mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\bg) ~+~
\delta \big\{2 \alpha\|\bg\|^2 \,+\, \delta\|\bg\|^2 \,+\, 2\|F\bg-\boldsymbol{\Phi}_L^\infty\|\exs \|\bg\|\big\}, \qquad \forall \bg \in L^2(\Omega_{\bd})^3.
\end{equation}
For any $\delta\!>\!0$ ($\alpha$ fixed), on can chose $\bg_{\alpha,\delta}$ such that $|\mathfrak{J}_\alpha(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta})- \mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty)| \, \leqslant \, \delta$. Then by the definition of~$\bg_{\alpha,\delta}^L$ one finds via triangle inequality that
\[
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta}^L) \, \leqslant \
\mathfrak{J}_\alpha^\delta(\boldsymbol{\Phi}_L^\infty;\bg_{\alpha,\delta}) \, \leqslant \, \mathfrak{j}_\alpha(\boldsymbol{\Phi}_L^\infty) +
\delta \big\{1+ 2 \alpha\|\bg_{\alpha,\delta}\|^2 \,+\, \delta\|\bg_{\alpha,\delta}\|^2 \,+\, 2\|F\bg_{\alpha,\delta}-\boldsymbol{\Phi}_L^\infty\|\exs \|\bg_{\alpha,\delta}\|\big\}.
\]
The proof of~\eqref{limlim} is now completed by noting that (i) given~$\alpha$, the term inside the brackets is bounded for any~$\delta$, and (ii) $\lim\limits_{\alpha\to 0} \mathfrak{j}_\alpha=0$. }
\end{proof}
\begin{theorem}\label{GLSM2}
Under the assumptions of Theorem~\ref{GLSM1} and an additional hypothesis that $B^\delta$ (denoting either~$F^\delta$ or~$F^\delta_\sharp$) is compact, one has
\begin{multline}\notag
\boldsymbol{\Phi}_L^\infty \,\in\, Range(\mathcal{H}^*) ~~\iff~~
\Big\{\limsup\limits_{\alpha \rightarrow 0}\limsup\limits_{\delta \rightarrow 0}\big(\exs |(\bg_{\alpha,\delta}^L, B^\delta \bg_{\alpha,\delta}^L)| \,+\, \delta \! \norms{\! \bg_{\alpha,\delta}^L \!}^2 \nxs \big) \,<\, \infty \\
\iff~ \liminf\limits_{\alpha \rightarrow 0}\liminf\limits_{\delta \rightarrow 0}\big(\exs |( \bg_{\alpha,\delta}^L, B^\delta \bg_{\alpha,\delta}^L)| \,+\, \delta \! \norms{\! \bg_{\alpha,\delta}^L \!}^2 \nxs \big) \,<\, \infty\Big\},
\end{multline}
where~$\bg_{\alpha,\delta}^L$ is a minimizer of the perturbed GLSM cost functional~\eqref{GCfn} in the sense of~\eqref{limlim}.
\end{theorem}
\begin{proof} See the proof of Theorem~5 in~\cite{Audibert2014}, also summarized in~\ref{GLSM*pruf}.
\end{proof}
\vspace{-7 mm}
\subsubsection{The GLSM criteria for imaging heterogeneous fractures} \label{GLSM_C}
On the basis of Theorem~\ref{GLSM2}, a robust GLSM-based criterion for the elastic-wave reconstruction of heterogeneous fractures can be designed as
\begin{equation}\label{GLSMg}
I^{\mathcal{G}}(L) \,\, \colon \!\!\! = \,\, \dfrac{1}{\sqrt{|( \bg_{\alpha,\delta}^L, B^\delta \bg_{\alpha,\delta}^L)| \exs+\exs \delta \! \norms{ \bg_{\alpha,\delta}^L \!}^2}}, \qquad B^\delta ~=~ F^\delta, F^\delta_\sharp,
\end{equation}
where $\bg_{\alpha,\delta}^L$ is a minimizer of~\eqref{GCfn} in the sense of~\eqref{limlim}. In this setting, it is particularly instructive to focus on the case where $B^\delta = F^\delta_\sharp$, since $\bg_{\alpha,\delta}^L$ in this case can be obtained \emph{non-iteratively} by explicitly solving~(\ref{min-RJ}). Accordingly, the GLSM indicator functional used is the sequel is taken as
\begin{equation}\label{GLSMgs}
I^{{\mathcal{G}}_\sharp}(L) \,\, = \,\, \dfrac{1}{\sqrt{\norms{\!(F^\delta_\sharp)^{\frac{1}{2}} \exs \bg_{\alpha,\delta}^L \nxs}^2 \exs+\,\, \delta \! \norms{ \bg_{\alpha,\delta}^L \!}^2}}.
\end{equation}
For future reference, let us also recall the classical LSM/FM solution $\bg_\epsilon^L \in L^2(\Omega_{\bd})^3$ (see Theorem~\ref{TR2}) obtained by way of Tikhonov regularization~\cite{Kress1999}, namely
\begin{equation}\label{LSMg}
\textcolor{black}{
\bg_\epsilon^L \,\, \colon \!\!\! = \,\, \min_{\bg \in L^2(\Omega_{\bd})^3} \big\lbrace \! \norms{F^\delta \bg \,-\, \boldsymbol{\Phi}_L^\infty}^2 + \,\, \beta \! \norms{\bg}^2 \! \big\rbrace,}
\end{equation}
where~$\beta$ is a regularization parameter computable by the Morozov discrepancy principle.
\begin{rem}
It is worth noting that the GLSM characterization of $\Gamma$ from the far-field data (via the range of $F$) is deeply rooted in geometrical considerations, so that the fracture indicator functionals~(\ref{GLSMg}) and~(\ref{GLSMgs}) may exhibit only a minor dependence on its heterogeneous contact condition -- given by the distribution of~$\bK$ on~$\Gamma$. This behavior can be traced back to Remark~\ref{LSMrem}, where the opening displacement profile $\ba \in \tilde{H}^{1/2}(L)$ -- intimately related to the interface law -- is deemed arbitrary (within the constraints of admissibility). This quality makes the GLSM imaging paradigm particularly attractive in situations where the fracture's contact law is unknown beforehand, which opens up possibilities for the sequential geometrical reconstruction and interfacial characterization of partially-closed fractures.
\end{rem}
\section{Computational treatment and results} \label{numerics}
To illustrate the theoretical developments, this section examines the performance of~\eqref{GLSMgs} through a set of numerical experiments and compares the results of the GLSM reconstruction to those obtained by two alternative approaches, namely the linear sampling method (LSM)~\cite{Fiora2003} and the method of topological sensitivity (TS)~\cite{Fatemeh2015}. In what follows the synthetic sensory data, namely the far-field patterns~\eqref{vinf2} over the unit sphere, are generated by way of an elastodynamic boundary integral method~\cite{Fatemeh2015}.
\begin{figure}[tp]
\center\includegraphics[width=0.88\linewidth]{figures/SetNum4.jpg} \vspace*{0mm}
\caption{Elastic-wave sensing setup (left), position of the cutting plane (middle), and ``zebra'' pattern of the fracture's heterogeneous contact condition~(right).} \label{SetNum}
\end{figure}
\emph{Testing configuration.}~The sensing setup, shown in Fig.~\ref{SetNum}, features a ``true'' cylindrical fracture $\Gamma$ of length $L = 0.7$ and radius $R = 0.35$. The fracture is endowed with a piecewise-constant (``zebra'') distribution of interfacial stiffness~$\bK(\bxi)$ on~$\Gamma$, alternating between $\bK_1$ and~$\bK_2$, where
\[
\begin{aligned}
\bK_1 \:=\: (1-0.25\textrm{i}) \, \bn \otimes \bn \,\,+\, (4 - 2\textrm{i}) \, \be_1 \nxs \otimes \be_1 \,+\, (4 - 2\textrm{i}) \, \be_2 \nxs \otimes \be_2, \qquad \bK_2 \:=\: \boldsymbol{0}
\end{aligned}
\]
in terms of the orthonormal basis~$(\be_1, \be_2,\bn)$ shown in the figure. The shear modulus, mass density, and Poisson's ratio of the background solid are taken as $\mu = 1$, $\rho = 1$ and $\nu = 0.35$, whereby the shear and compressional wave speeds read $c_s = 1$ and $c_p = 2.08$, respectively. The interaction of $\Gamma$ with incident (P- and S-) plane waves, propagating in direction $\bd$, gives rise to the scattered wavefield $\bv$ solving~\eqref{GE} -- whose far-field pattern $\bv^\infty$ is then computed on the basis of~(\ref{vinf2}).
\emph{Far-field operator.}~For both illumination and sensing purposes, the unit sphere $\Omega$ is sampled by a uniform grid of $N_\theta \!\times\! N_\phi$ observation directions, specified by the polar ($\theta_j,\, j\!=\!1,\ldots N_\theta$) and azimuthal~($\phi_k, \,k\!=\!1,\ldots N_\phi$) angle values. With reference to~\eqref{mat1}, note that both the polarization vector $\bq\!=\!\bq_p\!\oplus\bq_s$ of an incident plane wave and the far-field pattern $\bv^\infty_\bq=\bv^\infty_{\bq_p}\!\oplus \bv^\infty_{\bq_s}$ of the scattered wave each have \emph{only three} nontrivial components. In this setting, the discretized far-field operator $\textrm{\bf{F}}$ is represented as a $3N\!\times 3N$ matrix ($N\!=\!N_\theta N_\phi$) with components
\begin{equation}\label{DF}
\textrm{\bf{F}}(3k\nxs+\nxs1\!:\!3k\nxs+\nxs3, \,3j\nxs+\nxs1\!:\!3j\nxs+\nxs3) ~=~ \textrm{\bf{W}}^\infty (\bd_j,\hat\bxi_k), \qquad j,k = 0,\ldots N-1,
\end{equation}
where
\begin{equation}\label{mat2}
\textrm{\bf{W}}^\infty (\bd_j,\hat\bxi_k) ~=~
\left[\begin{array}{cccccc}
W^{\infty}_{11} & W^{\infty}_{12} & W^{\infty}_{13} \\
W^{\infty}_{21} & W^{\infty}_{22} & W^{\infty}_{23} \\
W^{\infty}_{31} & W^{\infty}_{32} & W^{\infty}_{33}
\end{array}\right] (\bd_j,\hat\bxi_k),
\end{equation}
and~$W^{\infty}_{kj}$ $(j,k\!=\!1,2,3)$ are specified in~\eqref{mat1}. Unless stated otherwise, we assume $N_\theta=50$ and $N_\phi=25$.
\emph{Noisy data.}~To account for the presence of noise in measurements, we consider the perturbed far-field operator
\begin{equation}\label{DFN}
\textrm{\bf{F}}^\delta \,\, \colon \!\!\!= \, (\boldsymbol{I} + \boldsymbol{N}_{\!\epsilon} ) \exs \textrm{\bf{F}},
\end{equation}
where $\boldsymbol{I}$ is the $3N \times 3N$ identity matrix, and $\boldsymbol{N}_{\!\epsilon}$ is the noise matrix of commensurate dimension whose components are uniformly-distributed (complex) random variables in $[-\epsilon, \, \epsilon]^2$. On the basis of definition~(\ref{Ns-op}), one has $\delta = \norms{\!\boldsymbol{N}_{\!\epsilon} \exs \textrm{\bf{F}}\!}$ which in the sequel takes values of up to~$20\%$. With reference to Remark~\ref{LSMrem}, the region of interest
\emph{Trial far-field pattern.} With reference to Remark~\ref{LSMrem}, the GLSM indicator map~\eqref{GLSMgs} is constructed by solving~\eqref{min-RJ} for the minimizer of~(\ref{RJ-alph}) over a grid of trial infinitesimal fractures $L=\bz\!+\bR{\sf L}$, where~$\bz$ denotes the sampling point and $\bR$ is a unitary rotation matrix. In what follows, this is accomplished by taking~{\sf L} to be a vanishing penny-shaped fracture with unit normal~$\bn_\circ$, i.e. by setting the FOD in~(\ref{Phi-inf}) as $\ba(\by) = \delta (\by-\bz) \bR\bn_\circ$. Writing for brevity $\textrm{\bf{n}}=\bR\bn_\circ$, one in particular finds that
\begin{equation}\label{Phi-inf-num}
\boldsymbol{\Phi}_L^\infty(\hat\bxi) ~=~ - \Big(\text{i} k_p \, \hat\bxi \exs \big[ \exs \lambda+2\mu \exs (\textrm{\bf{n}} \cdot \hat\bxi)^2 \exs \big] \exs e^{-\text{i}k_p \hat\bxi \cdot \bz} \;\oplus\;
2 \text{i} \mu \exs k_s \,\hat\bxi \times \nxs (\textrm{\bf{n}} \times\hat\bxi)\exs (\textrm{\bf{n}}\cdot\hat\bxi) \, e^{-\textrm{\emph{i}} k_s \hat\bxi \cdot \bz} \Big).
\end{equation}
Recalling~\eqref{mat1}, one may note that for each observation direction~$\hat\bxi_k$, \eqref{Phi-inf-num} has only three non-trivial components in the reference $(\hat\bxi_k,\boldsymbol{\theta}_k,\boldsymbol{\phi}_k)$ orthonormal basis, which are for consistency with~\eqref{mat2} arranged as a $3N\!\times\!1$ vector
\begin{equation}\label{Phi-inf-Dnum}
\boldsymbol{\Phi}_{\bz,\textrm{\bf{n}}}^\infty(3k+1\!:\!3k+3) ~=\,
\left[\begin{array}{c} \text{i} k_p \big[ \exs \lambda + 2\mu \exs (\textrm{\bf{n}} \sip \hat\bxi_k)^2 \exs \big] e^{-\text{i}k_p \hat\bxi_k \sip \bz} \\
2 \text{i} \mu \exs k_s (\textrm{\bf{n}}\sip\boldsymbol{\theta}_k) (\textrm{\bf{n}}\sip\hat\bxi_k) \, e^{-\textrm{\emph{i}} k_s \hat\bxi_k \sip \bz} \\
2 \text{i} \mu \exs k_s (\textrm{\bf{n}}\sip\boldsymbol{\phi}_k) (\textrm{\bf{n}}\sip\hat\bxi_k) \, e^{-\textrm{\emph{i}} k_s \hat\bxi_k \sip \bz}
\end{array} \right], \qquad k=0,\ldots N-1.
\end{equation}
Accordingly, the far-field equation~\eqref{FF} takes the discretized form
\begin{equation}\label{Dff}
\textrm{\bf{F}}^\delta \bg_{\bz,\textrm{\bf{n}}} ~=~ \boldsymbol{\Phi}_{\bz,\textrm{\bf{n}}}^\infty,
\end{equation}
thus forming the basis for computing GLSM and LSM indicator functionals.
\subsection{Fracture indicators} \label{RRM}
As shown in Fig.~\ref{SetNum}, the search area i.e.~the sampling region is a \emph{cube of side 2} where the featured (GLSM and LSM) indicator functionals are evaluated. The resulting distributions are plotted either in three dimensions, or in the mid-section of the ``true'' cylindrical fracture (see Fig.~\ref{SetNum}).
\emph{Sampling.}~In what follows, the search cube $[-1,1]^3\subset\mathbb{R}^3$ is probed by a uniform $40 \!\times\! 40 \!\times\! 40$ grid of sampling points~$\bz$, while the unit sphere -- spanning possible fracture orientations -- is sampled by a $24 \!\times\! 6$ grid of trial normal directions $\textrm{\bf{n}}=\bR\bn_\circ$. Accordingly, the fracture indicator map is constructed by solving~(\ref{Dff}) for a total of $M = 64000 \!\times\! 144$ trial pairs $(\bz,\textrm{\bf{n}})$.
\emph{GLSM indicator.}~With reference to~\eqref{min-RJ} and~\eqref{DF}-\eqref{Dff}, a discretized version of the GLSM solution vector, $\bg^{\mbox{\tiny{GLSM}}}_{\bz,\textrm{\bf{n}}}$, is computed by solving the linear system
\begin{equation} \label{min-DRJ}
\Big( \textrm{\bf{F}}^{\delta *}\textrm{\bf{F}}^\delta + \alpha_{\bz,\textrm{\bf{n}}} \exs (\textrm{\bf{F}}_\sharp^\delta)^{\nxs\frac{1}{2}*} (\textrm{\bf{F}}_\sharp^\delta)^{\nxs\frac{1}{2}} + \alpha_{\bz,\textrm{\bf{n}}} \delta \exs \boldsymbol{I} \Big) \exs \bg^{\mbox{\tiny{GLSM}}}_{\bz,\textrm{\bf{n}}} ~=~ \textrm{\bf{F}}^{\delta *} \boldsymbol{\Phi}^\infty_{\bz,\textrm{\bf{n}}},
\end{equation}
where $(\cdot)^*$ is the Hermitian operator; $\textrm{\bf{F}}_\sharp^\delta$ is evaluated on the basis of definitions~(\ref{Fs}) and~(\ref{ReIm}); and, following~\cite{Audibert2014},
\begin{equation}\label{Alph}
\alpha_{\bz,\textrm{\bf{n}}} \,\, \colon \!\!\! = \,\, \frac{\eta_{\bz,\textrm{\bf{n}}}}{\norms{\textrm{\bf{F}}^\delta\!} + \,\, \delta}.
\end{equation}
Here $\eta_{\bz,\textrm{\bf{n}}}$ is a regularization parameter of the classical LSM solution~\eqref{lssm1}, computed via the Morozov discrepancy principle~\cite{Kress1999}. With reference to~(\ref{GLSMgs}), the GLSM indicator function is then obtained as
\begin{equation}\label{GLSM-Dgs}
I^{\mathcal{G}_\sharp}(\bz) \,\, = \,\, \dfrac{1}{\sqrt{\norms{\!(\textrm{\bf{F}}^\delta_\sharp)^{\frac{1}{2}} \exs \bg^{\mbox{\tiny{GLSM}}}_{\bz} \nxs}^2 \exs+\,\, \delta \! \norms{\bg^{\mbox{\tiny{GLSM}}}_{\bz} \!}^2}}, \qquad
\textcolor{black}{
\bg^{\mbox{\tiny{GLSM}}}_{\bz} \,\,\colon \!\!= \,\, \text{argmin}_{\bg^{\mbox{\tiny{GLSM}}}_{\bz, \textrm{\bf{n}}}} \norms{\bg^{\mbox{\tiny{GLSM}}}_{\bz,\textrm{\bf{n}}}}^2_{L^2(\Omega)}, ~ \textrm{\bf{n}}\in\Omega.}
\end{equation}
\emph{LSM indicator.}~To gain better insight into the effectiveness of the proposed approach, the GLSM reconstruction is compared to a corresponding~LSM map. The latter is computed on the basis of a Tikhonov-regularized solution $\bg^{\mbox{\tiny{LSM}}}_{\bz,\textrm{\bf{n}}}$ to~(\ref{Dff}), namely
\begin{equation}\label{lssm1}
\bg^{\mbox{\tiny{LSM}}}_{\bz,\textrm{\bf{n}}} \,\,\colon \!\!= \,\, \text{argmin}_{\bg_{\bz, \textrm{\bf{n}}}} \Big\{ \norms{\textrm{\bf{F}}^\delta \bg_{\bz,\textrm{\bf{n}}} \,-\, \boldsymbol{\Phi}_{\bz,\textrm{\bf{n}}}^\infty}^2_{L^2(\Omega)} \,+\,\,\, \eta_{\bz,\textrm{\bf{n}}} \norms{\bg_{\bz,\textrm{\bf{n}}}}^2_{L^2(\Omega)}\Big\},
\end{equation}
where the regularization parameter $\eta_{\bz,\textrm{\bf{n}}}$ is obtained by way of Morozov discrepancy principle~\cite{Kress1999}. On the basis of~\eqref{lssm1}, the LSM indicator functional is constructed following~\cite{Fiora2003} as
\begin{equation}\label{LSM}
I^{\mathcal{L}}(\bz) \,\, := \,\, \frac{1}{\norms{\bg^{\mbox{\tiny{LSM}}}_{\bz}}^2}, \qquad
\textcolor{black}{
\bg^{\mbox{\tiny{LSM}}}_{\bz} \,\,\colon \!\!= \,\, \text{argmin}_{\bg^{\mbox{\tiny{LSM}}}_{\bz, \textrm{\bf{n}}}} \norms{\bg^{\mbox{\tiny{LSM}}}_{\bz,\textrm{\bf{n}}}}^2_{L^2(\Omega)}, ~ \textrm{\bf{n}}\in\Omega.}
\end{equation}
\subsection{Results} \label{Comp}
In the sequel, the arclength ($\ell \!=\! 0.55$) of a ``true'' cylindrical fracture in its mid-plane, see Fig.~\ref{SetNum}, is used as a reference length to gauge the illuminating shear wavelength $\lambda_s = 2\pi/k_s$.
\emph{Density of the sensing grid}. Taking $\lambda_s/\ell = 0.7$, Fig.~\ref{NN} illustrates the sensitivity of the GLSM indicator~\eqref{GLSM-Dgs} to the spatial density of sensory data, given by $N_\theta \!\times\! N_\phi$ incident/observation directions over the unit sphere. This is done by gradual downsampling of the default $50 \!\times\! 25$ sensing grid. From the panels, it is apparent that for satisfactory geometric reconstruction, the sensing grid should carry at least 100 test directions over $\Omega$. In what follows, the (full-aperture) reconstructions are implemented using a $50 \!\times\! 25$ grid.
\emph{Sensitivity to measurement noise}. Assuming full-aperture illumination and sensing, the GLSM and LSM indicators are next compared in terms of their robustness against noise in the far-field data. With reference to~\eqref{DFN}, the levels of ``white'' noise used to contaminate the boundary integral simulations of the forward scattering problem are taken $\delta = \norms{\!\boldsymbol{N}_{\!\epsilon} \exs \textrm{\bf{F}}\!} \in\{0, 0.1, 0.2\} \| \textrm{\bf{F}}\|$. On focusing the comparison on the mid-section~$\Pi$ of a ``true'' fracture, the results are shown in Figs.~\ref{F1}, \ref{F2}, and \ref{F4} assuming the illuminating wavelengths of~$\lambda_s/\ell = 1.3$, $0.7$, and $0.3$, respectively. Note that $\delta\%:=\delta/\|\textrm{\bf{F}}\|$. As can be seen from the display, the GLSM indicator~\eqref{GLSM-Dgs} inherits the superior localization ability of its LSM predecessor~\eqref{LSM}, while carrying far greater robustness to noise in the sensory data.
\begin{figure}[h]
\center\includegraphics[width=0.64\linewidth]{figures/NN.jpg} \vspace*{0mm}
\caption{Full-aperture GLSM reconstruction of a cylindrical fracture in its mid-section, $\Pi$, for $\lambda_s/\ell = 0.7$: effect of density of the $N_\theta \!\times\! N_\phi$ sensing grid of illumination/observation directions spanning the unit sphere.} \label{NN}
\end{figure}
\begin{figure}[h!]
\center\includegraphics[width=0.66\linewidth]{figures/F1_v3.jpg} \vspace*{-2mm}
\caption{Sensitivity to measurement noise for $\lambda_s/\ell = 1.3$: Full-aperture reconstruction of a cylindrical fracture, mid-section~$\Pi$, by the LSM indicator (top panels) and its GLSM counterpart (bottom panels).} \label{F1} \vspace*{2mm}
\center\includegraphics[width=0.66\linewidth]{figures/F2_v3.jpg} \vspace*{-2mm}
\caption{Sensitivity to measurement noise for $\lambda_s/\ell = 0.7$: Full-aperture reconstruction of a cylindrical fracture, mid-section~$\Pi$, by the LSM indicator (top panels) and its GLSM counterpart (bottom panels).} \label{F2} \vspace*{2mm}
\center\includegraphics[width=0.66\linewidth]{figures/F4_v3.jpg} \vspace*{-2mm}
\caption{Sensitivity to measurement noise for $\lambda_s/\ell = 0.3$: Full-aperture reconstruction of a cylindrical fracture, mid-section~$\Pi$, by the LSM indicator (top panels) and its GLSM counterpart (bottom panels).} \label{F4}
\end{figure}
\emph{Effect of the sensing aperture}. The ramifications of an incomplete aperture on the quality of fracture reconstruction are illustrated in Figs.~\ref{H1} and~\ref{H4}, where only \textcolor{black}{\emph{the ``upper'' half} of $\Omega$ in Fig.~\ref{SetNum}} is available for the purposes of illumination and observation. More specifically, Figs.~\ref{H1} and~\ref{H4} depict the GLSM and LSM fields in the mid-section of $\Gamma$ at ``long'' ($\lambda_s/\ell = 1.3$) and ``short'' ($\lambda_s/\ell = 0.3$) excitation wavelengths, respectively, constructed from the half-aperture sensory data. While the loss of resolution in both GLSM and LSM maps is clear relative to Figs.~\ref{F1} and~\ref{F4}, it is noted that (for the problem under consideration) the GLSM indicator offers far better robustness to noise, providing acceptable reconstruction of~$\Gamma$ for~$\delta$ as high as~$0.1\|\textrm{\bf{F}}\|$.
\begin{figure}[h!]
\center\includegraphics[width=0.66\linewidth]{figures/H1_v3.jpg} \vspace*{-2mm}
\caption{Half-aperture reconstruction of a cylindrical fracture, mid-section $\Pi$, for $\lambda_s/\ell = 1.3$:~sensitivity of the LSM indicator (top panels) and its GLSM counterpart (bottom panels) to noise in the measurements.} \label{H1} \vspace*{0mm}
\center\includegraphics[width=0.66\linewidth]{figures/H4_v3.jpg} \vspace*{-2mm}
\caption{Half-aperture reconstruction of a cylindrical fracture, mid-section $\Pi$, for $\lambda_s/\ell = 0.3$:~sensitivity of the LSM indicator (top panels) and its GLSM counterpart (bottom panels) to noise in the measurements.} \label{H4} \vspace*{3mm}
\center\includegraphics[width=0.7\linewidth]{figures/3D_v3.png} \vspace*{-2mm}
\caption{Full-aperture 3D GLSM reconstruction for $\{\lambda_s/\ell = 1.3,\delta\% = 0.1\}$ (top) and $\{\lambda_s/\ell = 0.7,\delta\% = 0.05\}$ (bottom):~ GLSM indicator~\eqref{GLSM-Dgs} thresholded at 10\% (left), fracture surface as reconstructed from the 3D cloud of points (middle), and fracture reconstruction after the application of a mean filter (right).} \label{3D} \vspace*{-4mm}
\end{figure}
\emph{3D reconstruction}. For completeness, Fig.~\ref{3D} illustrates the full-aperture GLSM reconstruction of $\Gamma$ inside the sampling region~$[-1,1]^3$, assuming $\lambda_s/\ell = 1.3$ and $\delta\% = 10$ (top panels) and $\lambda_s/\ell = 0.7$ and $\delta\% = 5$ (bottom panels). For clarity, the indicator maps are thresholded by $10\%$, i.e.~only the sampling points whose~$I^{\mathcal{G}_\sharp}(\bz)$ values are higher than ten percent of the global maximum value are shown (left panels). Then, a scattered interpolant is constructed based on thus obtained 3D cloud of points, giving an optimal reconstruction of the fracture surface. The latter is generated by (i) projecting the thresholded GLSM map onto a reference plane (the $X-Y$ plane in this example), and (ii) defining a suitable grid of points covering the projected area. This forms the sought-for input for the scattered interpolant providing a 3D reconstruction of the fracture interface, as shown in the middle panels of Fig.~\ref{3D}. Due in part to a scattered nature of the interpolant, thus obtained fracture surface will suffer from some artificial roughness -- that depends for example on the density of sampling points and an ad-hoc thresholding parameter. This issue may be mitigated by implementing a suitable spatial (e.g. moving average) filter, as shown in the right panels of Fig.~\ref{3D}.
\section{Conclusions} \label{Conc}
The Generalized Linear Sampling Method (GLSM) combined with the $F_\sharp$-factorization technique form a fast, yet robust, platform for the geometric reconstruction of heterogeneous (and dissipative) discontinuity surfaces from scattered wavefield data. It is illustrated that the GLSM indicator possesses little sensitivity to (the reasonable levels of) measurement noise -- that is comparable to the robustness of TS, while inheriting the top-tier localization property of the classical LSM, which guarantees a high-quality geometric characterization of the fracture -- notwithstanding the frequency regime of excitation and the unknown (generally heterogeneous) interfacial stiffness $\bK$. Such attributes carries a remarkable potential for developing a GLSM-based hybrid approach for not only geometric reconstruction of hidden fractures, but also identification of their interfacial condition (e.g.~retrieval of $\bK$ in the present work) from scattered field data. Furthermore, this approach may be naturally and rigorously extended to other sensing configurations and to more sophisticated background-domain geometries. It should also be noted that the analysis in this study does not require the fracture surface to be \emph{connected}, so one should be able to use the GLSM for simultaneous imaging of multiple fractures in the medium.
|
\section{Introduction}\label{sec_intro}
We consider the multi-armed bandit problem, which is the most basic example of a sequential decision problem with an exploration-exploitation trade-off. In each time step $t=1,2,\ldots,n$, the player has to play an arm $I_t \in \{1,\ldots,K\}$ from this fixed finite set and receives reward~$x_{I_t}(t) \in[0,1]$ depending on its choice\footnote
We assume that the player knows the total number of time steps~$n$.}.
The player observes only the reward of the chosen arm, but not the rewards of the other arms $x_i(t)$, $i \neq I_t$. The player's goal is to maximize its total reward $\sum_{t=1}^n x_{I_t}(t)$, and this total reward is compared to the best total reward of a single arm, $\sum_{t=1}^n x_i(t)$. To identify the best arm the player needs to explore all arms by playing them, but it also needs to limit this exploration to often play the best arm. The optimal amount of exploration constitutes the exploration-exploitation trade-off.
Different assumptions on how the rewards $x_i(t)$ are generated have led to different approaches and algorithms for the multi-armed bandit problem. In the original formulation \citep{R52} it is assumed that the rewards are generated independently at random, governed by fixed but unknown probability distributions with means~$\mu_i$ for each arm~$i=1,\ldots,K$. This type of bandit problem is called {\em stochastic}. The other type of bandit problem that we consider in this paper is called non-stochastic or {\em adversarial}~\citep{EXP3}. Here the rewards may be selected arbitrarily by an adversary and the player should still perform well for any selection of rewards. An extensive overview of multi-armed bandit problems is given in~\citep{BC12}.
A central notion for the analysis of stochastic and adversarial bandit problems is the regret~$R(n)$, the difference between the total reward of the best arm and the total reward of the player:
\[ R(n) = \max_{1\leq i \leq K} \sum_{t=1}^n x_i(t) - \sum_{t=1}^nx_{I_t}(t) . \]
Since the player does not know the best arm beforehand and needs to do exploration, we expect that the total reward of the player is less than the total reward of the best arm. Thus the regret is a measure for the cost of not knowing the best arm. In the analysis of bandit problems we are interested in high probability bounds on the regret or in bounds on the expected regret. Often it is more convenient, though, to analyze the pseudo-regret
\[ \pseudoR(n) = \max_{1\leq i \leq K} \EE{\sum_{t=1}^n x_i(t) - \sum_{t=1}^nx_{I_t}(t)} \]
instead of the expected regret
\[ \EE{R(n)} = \EE{\max_{1\leq i \leq K} \sum_{t=1}^n x_i(t) - \sum_{t=1}^nx_{I_t}(t)} . \]
While the notion of pseudo-regret is weaker than the expected regret with $\pseudoR(n) \leq \EE{R(n)}$, bounds on the pseudo-regret imply bounds on the expected regret for adversarial bandit problems with {\em oblivious} rewards $x_i(t)$ selected independently from the player's choices. The pseudo-regret also allows for refined bounds in stochastic bandit problems.
\subsection{Previous results}
For adversarial bandit problems, algorithms with high probability bounds on the regret are known~\cite[Theorem 3.3]{BC12}: with probability $1-\delta$,
\[ R_\adv(n) = \bigO{\sqrt{n \log(1/\delta)}}. \]
For stochastic bandit problems, several algorithms achieve logarithmic bounds on the pseudo-regret, e.g.~\cite{UCB1}:
\[ \pseudoR_\sto(n) = \bigO{\log n}. \]
Both of these bounds are known to be best possible.
While the result for adversarial bandits is a worst-case --- and thus possibly pessimistic --- bound that holds for any sequence of rewards, the strong assumptions for stochastic bandits may sometimes be unjustified.
Therefore an algorithm that can adapt to the actual difficulty of the problem is of great interest.
The first such result was obtained by \cite{BS12}, who developed the SAO algorithm that with probability $1-\delta$ achieves
\[ R_\adv(n) \leq \bigO{(\log n)\sqrt{n\log (n/\delta)}} \]
regret for adversarial bandits and
\[ \pseudoR_\sto(n) = \bigO{(\log n)^2} \]
pseudo-regret for stochastic bandits.
It has remained as an open question if a stochastic pseudo-regret of order $\bigO{(\log n)^2}$ is necessary or if the optimal $\bigO{\log n}$ pseudo-regret can be achieved while maintaining an adversarial regret of order~$\sqrt{n}$.
\subsection{Summary of new results}
We give a twofold answer to this open question. We show that stochastic pseudo-regret of order $\bigO{(\log n)^2}$ is necessary for a player to achieve high probability adversarial regret of order~$\sqrt{n}$ against an oblivious adversary, and to even achieve expected regret of order~$\sqrt{n}$ against an adaptive adversary. But we also show that a player can achieve $\bigO{\log n}$ stochastic pseudo-regret and $\tO{\sqrt{n}}$ adversarial {\em pseudo-regret} at the same time. This gives, together with the results of~\citep{BS12}, a quite complete characterization of algorithms that perform well both for stochastic and adversarial bandit problems.
More precisely, for any player with stochastic pseudo-regret bound of order $\bigO{(\log n)^\beta}$, $\beta < 2$, and any $\epsilon >0$, $\alpha < 1$, there is an adversarial bandit problem for which the player suffers $\Omega(n^\alpha)$ regret with probability $\Omega(n^{-\epsilon})$. Furthermore, there is an adaptive adversary against which the player suffers $\Omega(n^\alpha)$ expected regret. Secondly, we construct an algorithm with
\[ \pseudoR_\sto(n) = \bigO{\log n} \]
and
\[ \pseudoR_\adv(n) = \bigO{\sqrt{n \log n}} . \]
At first glance these two results may appear contradictory for $\alpha - \epsilon > 1/2$, as the lower bound seems to suggest a pseudo-regret of $\Omega(n^{\alpha-\epsilon})$. This is not the case, though, since the regret may also be negative. Indeed, consider an adversarial multi-armed bandit that initially gives higher rewards for one arm, and from some time step on gives higher rewards for a second arm. A player that detects this change and initially plays the first arm and later the second arm, may outperform both arms and achieve negative regret. But if the player misses the change and keeps playing the first arm, it may suffer large regret against the second arm.
In our analysis we use both mechanisms. For the lower bound on the pseudo-regret we show that a player with little exploration (which is necessary for small stochastic pseudo-regret) will miss such a change with significant probability and then will suffer large regret. For the upper bound we explicitly compensate possible large regret that occurs with small probability by negative regret that occurs with sufficiently large probability. For the lower bound on the expected regret we construct an adaptive adversary that prevents such negative regret.
Consequently, our results exhibit one of the rare cases where there is a significant gap between the achievable pseudo-regret and the achievable expected regret.
The explicit consideration of negative regret is one of the technical contributions of this work. Another, maybe even more significant contribution, is a weak testing scheme for non-stochastic arms.
This weak testing scheme is necessary since $\bigO{\log n}$ stochastic pseudo-regret allows only for very little exploration.
Each individual weak test has a constant false positive rate (predicting a non-stochastic arm although the arm is stochastic) and a constant false negative rate (missing a non-stochastic arm).
To avoid classifying a stochastic arm as non-stochastic, an arm is classified as non-stochastic only after $\bigO{\log n}$ positive tests. This reduces the false positive rate of a decision to acceptable $\bigO{1/n}$.
Conversely, this delayed detection needs to be accounted for in the regret analysis when the arms are indeed non-stochastic.
\section{Definitions and statement of results}
In a multi-armed bandit problem with arms $i=1,\ldots,K$ the interaction of a player with its environment is governed by the following protocol:
\begin{quote}
For time steps $t=1,\ldots,n$:
\begin{enumerate}
\item The player chooses an arm $I_t \in \{1,\ldots,K\}$, possibly using randomization.
\item The player receives and observes the reward $x_{I_t}(t)$.\\
It does not observe the reward from any other arm $i \neq I_t$.
\end{enumerate}
\end{quote}
The player's choice~$I_t$ may depend only on information available at this time, namely $I_1,\ldots,I_{t-1}$ and $x_{I_1}(1),\ldots,x_{I_{t-1}}(t-1)$.
If the bandit problem is stochastic, then the rewards $x_i(t)$ are generated independently at random. If the bandit problem is adversarial, then the rewards are generated arbitrarily by an adversary. We assume that all rewards $x_i(t) \in [0,1]$ and that the number of time steps~$n$ is known to the player.
\subsection{Stochastic multi-armed bandit problems}
In a stochastic multi-armed bandit problem the rewards for each arm~$i$ are generated by a fixed but unknown probability distribution $\nu_i$ on~$[0,1]$. All rewards $x_i(t)$, $1 \leq i \leq K$, $1 \leq t \leq n$, are generated independently at random with $x_i(t) \sim \nu_i$.
Important quantities are the average rewards of the arms, $\mu_i = \EE{x_i(t)}$, the average reward of the best arm $\mu^*=\max_i \mu_i$, and the resulting gaps $\Delta_i=\mu^*-\mu_i$.
The goal of the player is to achieve low pseudo-regret which for a stochastic bandit problem can be written as
\begin{align*}
\pseudoR_\sto(n)
= \max_{1\leq i \leq K} \EE{\sum_{t=1}^n x_i(t) - \sum_{t=1}^n x_{I_t}(t)}
= \sum_{i=1}^K \Delta_i \EE{T_i(n)} ,
\end{align*}
where $T_i(n) = \#\{1 \leq t \leq n: I_t = i\}$ is the number of plays of arm~$i$. It can be shown~\citep{UCB1} that --- among others --- upper confidence bound algorithms achieve
\[ \EE{T_i(n)} = \bigO{\frac{\log n}{\Delta_i^2}} \]
for any arm $i$ with $\Delta_i > 0$ such that
\[ \pseudoR_\sto(n) = \bigO{\sum_{i: \Delta_i > 0}\frac{\log n}{\Delta_i}} .\]
It can be even shown that for arms $i$ with $\Delta_i > 0$,
\[ T_i(n) = \bigO{\frac{\log (n/\delta)}{\Delta_i^2}} \]
with probability $1-\delta$ when $n$ is known to the player.
\subsection{Adversarial multi-armed bandit problems}
In adversarial bandit problems the rewards are selected by an adversary. If this is done beforehand (before the player interacts with the environment), then the adversary is called {\em oblivious} as the selection of rewards is independent from the arms~$I_t$ chosen by the player. In this case any upper bound on the pseudo-regret that holds for any selection of rewards is also an upper bound on the expected regret.
If the selection of rewards $x_i(t)$, $1 \leq i \leq K$, depends on which arms $I_1,\ldots,I_{t-1}$ the player has chosen in the past, then the adversary is called {\em adaptive}. In this case a bound on the pseudo-regret does not necessarily translate into a bound on the expected regret.
Nevertheless, strong bounds on the regret against an adaptive adversary are known for the \expp\ algorithm~\citep{EXP3}:
\begin{theorem}[{\citealp[Theorem 3.3]{BC12}}] \label{t:EXP3}
When \expp\ is run with appropriate parameters depending on $n$, $K$, and $\delta$, then with probability $1-\delta$ its regret satisfies
\[ R_\ada(n) = \bigO{\sqrt{nK\log(K/\delta)}} . \]
\end{theorem}
\subsection{Results}
First, we state our lower bounds for oblivious and adaptive adversaries.
\begin{theorem} \label{theorem_lower_bound}
Let $\alpha < 1$, $\epsilon > 0$, $\beta < 2$, and $\clower > 0$. Consider a player that achieves pseudo-regret
\[ \pseudoR_\sto(n) \leq \clower (\log n)^\beta \]
for any stochastic bandit problem with two arms and gap~$\Delta=1/8$.
Then for large enough $n$ there is an adversarial bandit problem with two arms and an oblivious adversary such that the player suffers regret
\[ R_\obl(n) \geq n^\alpha/8 - 4\sqrt{n \log n} \]
with probability at least $1/(16n^{\epsilon}) - 2/n^2$. Furthermore, there is an adversarial bandit problem with two arms and an adaptive adversary such that the player suffers expected regret
\[ \EE{R_\ada(n)} \geq \frac{n^{\alpha-\epsilon}}{128} - 3\sqrt{n \log n} . \]
\end{theorem}
In Section~\ref{s:sao} we present our \sapo~algorithm (Stochastic and Adversarial Pseudo-Optimal) that achieves optimal pseudo-regret in stochastic bandit problems and nearly optimal pseudo-regret in adversarial bandit problems. Its performance is summarized in the following theorem.
\begin{theorem}\label{thm:main}
For large enough $n$ and any $\delta > 0$, algorithm \sapo\ achieves the following bounds for suitable constants $C_\sto$, $C_\adv$, and $\Cib$:
\begin{itemize}
\item For stochastic bandit problems with gaps $\Delta_i$
such that\\
$\Cib\sum_{i:\Delta_i > 0} \frac{\log (n/\delta)}{\Delta_i} \leq \sqrt{nK \log (n/\delta)}$,
\[ T_i(n) \leq C_\sto \frac{\log (n/\delta)}{\Delta_i^2} \]
with probability $1-\delta$ for any arm $i$ with $\Delta_i > 0$, and thus
\[ \pseudoR_\sto(n) \leq C_\sto \sum_{i:\Delta_i > 0} \frac{\log (n/\delta)}{\Delta_i} + \delta n . \]
\item For adversarial bandit problems
\[ \pseudoR_\ada(n) \leq C_\adv K \sqrt{n \log (n/\delta)} + \delta n . \]
\end{itemize}
\end{theorem}
\begin{remark}
Our bound for adversarial bandit problems shows a worse dependency on $K$ than Theorem~\ref{t:EXP3}. This is an artifact of our current analysis and can be improved to a bound
$\pseudoR_\ada(n) = \bigO{\sqrt{nK\log(n/\delta)}}$.
\end{remark}
\subsection{Comparison with related work}
\cite{BS12} show for their SAO algorithm that with probability $1-\delta$,
\[ \sum_{i=1}^K \Delta_i T_i(n) \leq \bigO{\frac{K \log K (\log n/\delta)^2}{\Delta}} \]
for stochastic bandits where $\Delta = \min_{i:\Delta_i > 0}\Delta_i$, and
\[ R_\ada(n) \leq \bigO{(\log K)(\log n)\sqrt{nK\log n/\delta}} \]
for adaptive adversarial bandits.
While our bounds in Theorem~\ref{thm:main} are somewhat tighter, in particular showing the optimal dependency on the gaps~$\Delta_i$ for stochastic bandits, we have only a result on the pseudo-regret for adversarial bandits. We conjecture though, that our analysis can be used to construct an algorithm that with probability $1-\delta$ achieves
$T_i(n) \leq \bigO{{(\log n/\delta)^2}/{\Delta_i^2}}$ for stochastic bandits and
$R_\ada(n) \leq \bigO{(\log K)(\log n)\sqrt{nK\log n/\delta}}$ for adaptive adversarial bandits.
Our \sapo\ algorithm follows the general strategy of the SAO algorithm by essentially employing an algorithm for stochastic bandit problems that is equipped with additional tests to detect non-stochastic arms.
A different approach is taken in~\citep{SS14}: here the starting point is an algorithm for adversarial bandit problems that is modified by adding an additional exploration parameter to achieve also low pseudo-regret in stochastic bandit problems. While this approach has not yet allowed for the tight $\bigO{\log n}$ regret bound in stochastic bandit problems (they achieve a $\bigO{\log^3 n}$ bound), the approach is quite flexible and more generally applicable than the SAO and \sapo\ algorithms.
\subsection{Proof sketch of the lower bound (Theorem~\ref{theorem_lower_bound})}
We present here the main idea of the proof. The proof itself is given in Appendix~\ref{app:fullproof-lower}.
We consider a stochastic bandit problem with constant reward $x_1(t)=1/2$ for arm~1 and Bernoulli rewards with $\mu_2=1/2-\Delta$ for arm~2, $\Delta=1/8$. We divide the time steps into phases of increasing length $L_j=3^j n^{\alpha}$, $j=0,\ldots,J$ with $J = \Omega(\log n)$. Since the pseudo-regret of the player is $\bigO{(\log n)^\beta}$, there is a phase $j^*$ where the expected number of plays of arm~2 in this phase is $\bigO{(\log n)^{\beta-1}}$.
We construct an oblivious adversarial bandit by modifying the Bernoulli distribution of arm~2 in phase~$j^*$ and beyond by setting $\mu_2=1/2+\Delta$. By this modification arm~2 gives larger total reward than arm~1.
Because of the limited number of plays in phase~$j^*$, a standard argument shows that the player will not detect this modification during phase~$j^*$ with probability
$\exp\{-O(\log^{\beta-1}\T)\}=\Omega(n^{-\epsilon})$. When the modification is not detected during phase~$j^*$, then in this phase the player suffers roughly regret $\Delta L_{j^*}$ against arm~2. This is not compensated by negative regret against arm~2 in previous phases since
$\Delta \sum_{j=0}^{j^*-1}L_j \leq \Delta L_{j^*}/2$. Thus in this case the overall regret of the player against arm~2 is roughly $\Delta L_{j^*}/2 = \Omega(n^\alpha)$.
In a very similar way we can construct also an adaptive adversarial bandit: As for the oblivious bandit, we set $\mu_2 = 1/2+\Delta$ in phase~$j^*$. If the player chooses arm~2 only $C(\log n)^{\beta-1}$ times in phase~$j^*$, then we keep $\mu_2 = 1/2+\Delta$ also for the remaining phases. As for the oblivious bandit this happens with probability $\Omega(n^{-\epsilon})$ and gives regret $\Omega(n^\alpha)$. To avoid negative regret, we switch back to $\mu_2=1/2-\Delta$, as soon as there more than $C(\log n)^{\beta-1}$ plays of arm~2 in phase~$j^*$. In this case the reward of the algorithm is roughly $n/2 + C\Delta(\log n)^{\beta-1}$ such that in this case
$R(n) \geq -C\Delta(\log n)^{\beta-1}$. Hence the expected regret is
$\EE{R(n)} \geq \Omega(n^{\alpha-\epsilon}) - C\Delta(\log n)^{\beta-1} = \Omega(n^{\alpha-\epsilon})$.
\section{The \sapo\ algorithm}
\label{s:sao}
\input{sapo-algorithm}
In its core the algorithm is an elimination procedure for stochastic bandits that is augmented by tests safeguarding against non-stochastic arms. If there is sufficient evidence for non-stochastic arms, then the algorithm switches to the adversarial bandit algorithm \expp, starting with the current time step.
The algorithm maintains a set of active arms~$\cA$ and a set of supposedly suboptimal ``bad'' arms~$\cB$. For each arm~$i$ it maintains the sample mean~$\hmu_i(s)$,
\begin{align*}
\hmu_i(s) &= \frac{1}{T_i(s)}\sum_{t=1}^{s} x_i(t) \II{I_t=i} , \\
T_i(s) &= \sum_{t=1}^{s} \II{I_t=i} ,
\end{align*}
and also an unbiased estimate to deal with non-stochastic arms,
\begin{align*}
\bmu_i(s) &= \frac{1}{s}\sum_{t=1}^{s} x_i(t) \frac{\II{I_t=i}}{p_i(t)} ,
\end{align*}
where $p_i(t)$ is the probability of choosing arm~$i$ at time~$t$.
Confidence bounds\footnote
We start with $\lcb_i(0)=\blcb_i(0)=0$ and $\bucb_i(0)=1$.}
around the estimated means are used to evict arms from the active set~$\cA$,
\begin{align*}
\lcb_i(s) &= \max\{\lcb_i(s-1),\hmu_i(s) - \width_i(s)\}, \\
\blcb_i(s) &= \max\{\blcb_i(s-1),\bmu_i(s)-\bwidth(s)\}, \\
\bucb_i(s) &= \min\{\bucb_i(s-1),\bmu_i(s)+\bwidth(s)\}, \\
\lcb^*(s) &= \max_{1 \leq i \leq K} \max\{\lcb_i(s),\blcb_i(s)\}, \\
\width_i(s) &= \sqrt{\Cwid \log (n/\delta)/T_i(s)} ,\\
\bwidth(s) &= \sqrt{\Cwid K\log (n/\delta)/s} .
\end{align*}
Note that $\lcb_i(s)$, $\blcb_i(s)$, and $\lcb^*(s)$ are non-decreasing and $\bucb_i(s)$ are non-increasing. This reflects the intuition that confidence intervals should be shrinking and is used to safeguard against non-stochastic arms.
An arm~$i$ is evicted from~$\cA$ in {\bf Step~2.a}, if
it has a sufficient number of plays ($\Ci\cdot\log (n/\delta)$) for reasonably accurate estimates, and if
its sample mean~$\hmu_i(t-1)$ is significantly smaller than the optimal lower confidence bound~$\lcb^*(t-1)$.
The additional distance $\Cgap\cdot\width_i(t-1)$ is used to estimate the gap~$\Delta_i$. For evicted arms, in {\bf Step 2.b} an estimate for the gap $\egp_i$ and the current estimated mean are frozen, $\emu_i=\hmu_i(t-1)$. For stochastic bandits the accuracy of this estimate is proportional to the estimated gap~$\egp_i$. These quantities are used in the tests for detecting non-stochastic arms. Also the starting time~$n_i(t)$ and the length~$L_i(t)=L_i^0$ of the first testing phase (see below), as well as the number of detections $E_i(t)=0$ are set.
Since \sapo\ needs to perform well also against adversaries, all choices of arms are randomized.
In {\bf Step 3} an active arm is chosen uniformly at random, or with some smaller probabilit
a bad arm~$i$ is chosen where the probability depends on the length of its current testing phase~$L_i(t)$.
Choosing also bad arms is necessary to detect non-stochastic arms among the bad arms.
\subsection{Tests for detecting non-stochastic arms}
The most important test is in {Step 4.a} for detecting that a bad arm receives larger rewards than it should if it were stochastic. Such an arm could be optimal if the bandit problem is adversarial. The best way to view this test is by dividing the time steps of an evicted arm~$i$ into testing phases
\[
\tau_{i,1},\ldots,\tau_{i,2}-1;
\tau_{i,2},\ldots,\tau_{i,3}-1;
\tau_{i,3},\ldots,\tau_{i,4}-1;\ldots
\]
The first phase starts when arm $i$ is evicted from $\cA$. A phase~$k$ ends at time $\tau_{i,k+1}-1$ if either the phase has exhausted its length (Step~4.d), or when the test in Step~4.a reports a detection.\footnote
The last phase ends when the total number of time steps~$n$ is exhausted or when the algorithm switches to \expp.}
Thus the length parameter~$L_i(t)$ is only the maximal length of a phase and the phase may end earlier. In the notation of the algorithm $n_i(t)$ denotes the start of the current phase. Within a phase the probability $p_i(t)$ for choosing arm~$i$ is constant since the length parameter~$L_i(t)$ does not change (Step 4.e). For notational convenience we denote by $p_{ik}$ the probability for choosing arm~$i$ in its $k$-th testing phase, and by $L_{ik}$ the corresponding length parameter,
\begin{align*}
p_i(t) &= p_{ik} \text{~~for $i \in \cB(t)$ and $\tau_{i,k} \leq t < \tau_{i,k+1}$},
\\
L_i(t) &= L_{ik} \text{~~for $i \in \cB(t)$ and $\tau_{i,k} \leq t < \tau_{i,k+1}$},
\\
n_i(t) &= \tau_{i,k} \text{~~for $i \in \cB(t)$ and $\tau_{i,k} \leq t < \tau_{i,k+1}$}.
\end{align*}
Now the test in {\bf Step 4.a} checks if a bad arm $i$ has received significantly more rewards in the current phase then expected, given the estimated mean~$\emu_i$, the maximal phase length~$L_i(t)$ and the probability for choosing arm~$i$, $p_i(t)$, where
\[ \hD_i(s_1,s_2) = \sum_{t=s_1}^{s_2} [x_i(t) - \emu_i] \II{I_t=i} .
\]
If arm $i$ is stochastic, then $\EE{\hD_i(s_1,s_2)} =\bigO{L_i(t) \egp_i p_i(t)}$ such that a positive test suggests that the arm is non-stochastic. Since the expected number of plays of arm~$i$ is $L_i^0/K$ in each phase, the test is weak, though, with constant false positive and false negative rates. To avoid incorrectly classifying a stochastic arm as non-stochastic, the test is repeated several times. To make the tests independent, a new phase is started in {\bf Step~4.b} after a detection is reported. To avoid that too much regret accumulates in the case of a non-stochastic arm, the phase length is halved. If there have been~$E^0$ independent detections, then in {\bf Step~4.c} there is sufficient evidence for a non-stochastic arm and the algorithm switches to \expp.
In {\bf Step~4.d} the phase ends because it has exhausted its length. Since the test in Step~4.a has given no detection, arm~$i$ has performed as expected and the algorithm has accumulated negative regret against this bad arm. This negative regret allows to start the next phase with a doubled phase length, even if the arm were non-stochastic. Doubling the phase length is necessary to avoid too many phases for a stochastic arm. (Remember that the expected number of plays of a bad arm is $L_i^0/K$ in each phase.)
In {\bf Step 4.e} none of the above condition is satisfied and the phase continues.
Additional simpler tests for non-stochastic arms are performed in Step~1. {\bf Step~1.a} checks whether for all active arms the unbiased estimates of the means obey the corresponding confidence intervals.
Finally, {\bf Step 1.b} checks if the algorithm receives significantly less reward than expected from the best lower confidence bound. This may happen if a non-stochastic arm first appears close to optimal but then receives less rewards.
\subsection{Choice of constants in the \sapo\ algorithm}
\label{s:constants}
In the algorithm we keep the constant names because we find them easier to read than actual values. Proper values for the constants are as follows: $\Cwid=16$, $\Cib=522$, $\Ci=100/9$, $\Cgap=60$, $\Cpp=1300$, $\Caa=1/10$, and $\ci=15$.
\section{Preliminaries for the analysis of \sapo}\label{sec_pre}
An important tool for our analysis are concentration inequalities, in particular Bernstein's inequality for martingales and a variant of Hoeffding-Azuma's inequality for the maximum of partial sums, $\max_{1\leq s \leq t \leq n} \sum_{i=s}^t Y_i$. These inequalities are given in Appendix~\ref{app:concentration}.
We denote by $\hist_t$ the past up to and including time~$t$.
The next lemma states some properties of algorithm \sapo.
Let
\[ T_i(s_1,s_2) = \#\{t:s_1 \leq t \leq s_2: I_t=i\} \]
denote the number of plays of arm $i$ in time steps $s_1$ to $s_2$, let $\nbi$ be the time when arm~$i$ is evicted from~$\cA$,
\[ i \in \cA(\nbi-1) \text{~~and~~} i \in \cB(\nbi), \]
and let $\ns$ be the time step when \sapo\ switches to \expp.
If \sapo\ never switches to \expp, then $\ns=n$.
\begin{lemma} \label{l:prelim}
\begin{enumerate}[label={\rm(\alph*)},ref=\alph*,nosep]
\iffalse
\item \label{item:diffTi}
If $T_{i}(s,t) \geq \Ci \log (n/\delta)$ and $t < \nbi$, then $T_{i}(s,t) \geq \cv T_{i'}(s,t)$.
\item \label{item:initT}
If $T_{i}(s,t) \geq \Ci \log (n/\delta)$ and $t < \nbi$, then $T_{i}(s,t) \geq (t-s+1)/(2K)$.
\fi
\item \label{item:emu-lcb} If $i \in \cB(t)$ then $\emu_i + \egp_i < \lcb^*(t)$.
\item \label{item:numPhases}
For each arm the number of testing phases $k$, $\tau_{i,k}\cdots\tau_{i,k+1}-1$ is \\at most
$\numPH = \lceil \log_2 n\rceil + 2E^0$.
\item \label{item:TiB}
With probability $1-\bigO{\delta}$, the number of plays of any bad arm $i$ is bounded as\\
$T_i(\nbi,\ns) \leq \frac{101}{100} L_i^0\numPH/K = \bigO{\numPH/\egp_i^2}$.
\end{enumerate}
\end{lemma}
\newcommand{\refP}[1]{\ref{l:prelim}\ref{item:#1}}
\begin{proof}{\bf (Sketch)}
\iffalse
Statement (\ref{item:diffTi}) and (\ref{item:initT}) follow from the fact that active arms are chosen uniformly, with probability at least $1/K$, and with higher probability than bad arms. Thus the probability that an active arm is chosen significantly less often than any other arm is small.
\fi
Statement (\ref{item:emu-lcb}) follows immediately from Step~2 of the algorithm since $\emu_i=\hmu_i(\nbi-1)$, $\egp_i=\Cgap\cdot\width_i(\nbi-1)$, $\hmu_i(\nbi-1)+\Cgap\cdot\width_i(\nbi-1) < \lcb^*(\nbi-1)$, and $\lcb^*(t)$ is non-decreasing.
Statement (\ref{item:numPhases}) follows from the fact that Step~4.b (where the phase length is halved) is executed at most~$E^0$ times. In the other phases the phase length is doubled in Step~4.d. Since the phase length is at most $n$, the number of phases is at most $\log_2 n + 2E^0$.
For statement (\ref{item:TiB}) we observe that by the definition of $p_i(t)$ the expected number of plays in any testing phase of a bad arm~$i$ is $L_i^0/K$. Thus the expected number of plays in all phases is~$L_i^0 M/K$. Since the variance is bounded by the same quantity, an application of Bernstein's inequality gives the result.
Detailed proofs are given in Appendix~\ref{app:prelim}.
\end{proof}
\section{Analysis of \sapo\ for adversarial bandits}\label{sec_adv_analysis_K}
In this section we prove pseudo-regret bounds for \sapo\ against adversarial and possibly adaptive bandits.
Since we know from Theorem~\ref{t:EXP3} that \expp\ suffers small regret, we only need to bound the pseudo-regret of \sapo\ before it switches to \expp.
For the remaining section we fix some arm~$i$.
We have
\begin{eqnarray}
\lefteqn{\sum_{t=1}^\ns x_i(t) - \sum_{t=1}^\ns x_{I_t}(t)
\;=\; {\sum_{t=1}^{\ns} \left[ x_i(t)-\lcb^*(t) \right] }
+ {\sum_{t=1}^{\ns} \left[ \lcb^*(t) - x_{I_t}(t) \right] }} \nonumber \\
&= & {\sum_{t=1}^{\nbi-1} \left[ x_i(t)-\lcb^*(t) \right] }
+ {\sum_{t=\nbi}^{\ns} \left[ x_i(t)-\lcb^*(t) \right] }
+ {\sum_{t=1}^{\ns} \left[ \lcb^*(t) - x_{I_t}(t) \right] } \label{eq:adv-Ri}
\end{eqnarray}
The first sum in (\ref{eq:adv-Ri}) bounds the regret for the time when $i$ is an active arm. For stochastic arms, the best lower confidence bound $\lcb^*(t)$ would be not too far from the rewards of the arms that are still active. For non-stochastic arms, though, we need the tests in \sapo, in particular those in Step~1, to guarantee a similar behavior and achieve
\begin{align}\label{eq:bound-active}
\EE{\sum_{t=1}^{\nbi-1} \left[ x_i(t)-\lcb^*(t) \right]}
= \bigO{\sqrt{Kn\log (n/\delta)}} ,
\end{align}
see Appendix~\ref{app:bound-active}.
The crucial part of the analysis concerns the second sum in (\ref{eq:adv-Ri}) which bounds the regret for the time when $i$ is a bad arm. For its analysis we explicitly track negative regret to compensate for positive regret. In Section~\ref{s:bad} below we sketch the main ideas for handling this sum (formal proofs are given in Appendix~\ref{app:bound-bad}), showing that
\begin{align}\label{eq:bound-bad}
\EE{\sum_{t=\nbi}^{\ns} \left[ x_i(t)-\lcb^*(t) \right]}
= \bigO{\frac{K\log (n/\delta)}{\egp_i}} .
\end{align}
Note that $1/\egp_i=\bigO{\width_i(\nbi-1)} = \bigO{\sqrt{T_i(\nbi)/\log(n/\delta)}}
= \bigO{\sqrt{n/\log(n/\delta)}}$ such that $\bigO{K\log (n/\delta)/\egp_i} = \bigO{K\sqrt{n\log(n/\delta)}}$.
Finally, the third sum can be observed by the algorithm and is taken care of by the test in Step~1.b, such that
\begin{align}\label{eq:bound-1c}
\sum_{t=1}^{\ns} \left[ \lcb^*(t) - x_{I_t}(t) \right]
= \bigO{\sqrt{Kn\log(n/\delta)}}.
\end{align}
Together, inequalities (\ref{eq:adv-Ri})--(\ref{eq:bound-1c}) and the bound on \expp\ in Theorem~\ref{t:EXP3} give the bound on the pseudo-regret in Theorem~\ref{thm:main}.
\subsection{Bounding the regret for bad arms}
\label{s:bad}
If a bad arm is non-stochastic, then it may first appear suboptimal but still be optimal after all. We need to show that the tests of our algorithm, in particular the test in Step~4.a, are sufficient to detect such a situation. Since the algorithm checks arms in $\cB(t)$ only rarely, it will take some time for such detection. In our analysis we explicitly compensate the regret during this delayed detection by the negative regret accumulated while arm~$i$ was performing suboptimally.
We consider the testing phases $k$, $\tau_{i,k} \ldots \tau_{i,k+1}-1$, of arm~$i$, and recall that~$L_{ik}$ is the length parameter for phase~$k$ and $p_{ik}=L_i^0/(K L_{ik})$ is the probability for choosing arm~$i$ in phase~$k$. Furthermore, let~$E_{ik}$ the value of~$E_i(t)$ in phase~$k$.
Note that these quantities may change only when a new phase begins.
We denote by $\PPik{\cdot}=\pr{\cdot|\hist_{\tau_{i,k}-1}}$ and $\EEik{\cdot}=\EE{\cdot|\hist_{\tau_{i,k}-1}}$ the probabilities and expectations conditioned on the past before phase~$k$.
For any phase we have
\begin{align}
\sum_{t=\tau_{i,k}}^{\tau_{i,k+1}-1} \left[ x_i(t)-\lcb^*(t) \right]
&= \sum_{t=\tau_{i,k}}^{\tau_{i,k+1}-1} \left[ x_i(t)-\emu_i+\emu_i-\lcb^*(t) \right] \nonumber
\\&< \sum_{t=\tau_{i,k}}^{\tau_{i,k+1}-1} \left[ x_i(t)-\emu_i\right] - \egp_i[\tau_{i,k+1} - \tau_{i,k}] \label{eq:negRegret}
\end{align}
by Lemma~\refP{emu-lcb}.
Thus we want to prevent that the rewards of arm~$i$ are significantly larger than the estimated mean~$\emu_i$. In particular, the test in Step~4.a is supposed to detect events
$D_i(s_1,s_2) > \CVV \egp_i L_{ik}$ with
\[
D_i(s_1,s_2) := \sum_{t=s_1}^{s_2} [x_i(t) - \emu_i] .
\]
Since on average arm~$i$ is chosen only $L_i^0/K$ times per phase, there is a constant false negative rate $\qadv$ for missing such events. For appropriate $\Cpp$, though, the false negative rate $\qadv$ is sufficiently small, $\qadv \leq 1/25$: Since $\EE{\hD_i(s_1,s_2)} = p_{ik} D_i(s_1,s_2)$ for $\tau_{i,k} \leq s_1 \leq s_2 < \tau_{i,k+1}$, and Step~4.a tests for $\hD_i(s_1,s_2) > \Caa \egp_i L_{ik} p_{ik}$, we can bound $\qadv$ by Bernstein's inequality using that $1 \leq \egp_i^2 L_i^0/(K \Cpp)$ and a bound on the variance,
\[ \VV{\hD_i(s_1,s_2)} \leq L_{ik} p_{ik} = L_i^0/K \leq (\egp_i L_i^0/K)^2/\Cpp = (\egp_i L_{ik} p_{ik})^2/\Cpp . \]
The formal proof is given in Lemma~\ref{l:qadv}.
We use the false negative rate $\qadv$ to bound $\EEik{D_i(\tau_{i.k},\tau_{i.k+1}-1)}$. Each time an event $D_i(s,t) > \CVV \egp_i L_{ik}$ is missed (we consider only non-overlapping such events), $D_i(\tau_{i.k},t)$ has increased by at most $\CVV \egp_i L_{ik} + 1$, and the probability for the $m$-th miss is at most $\qad{m}$. When such an event is detected, then the phase ends and $D_i(\tau_{i.k},t)$ again has increased by at most $\CVV \egp_i L_{ik} + 1$.
Thus (see Lemma~\ref{l:expD} for the formal proof)
\begin{align*}
\EEik{D_i(\tau_{i.k},\tau_{i.k+1}-1)} \leq (\CVV \egp_i L_{ik}+1) \sum_{m \geq 0} \qad{m} = \frac{\CVV \egp_i L_{ik}+1}{1-\qadv}
\end{align*}
which by (\ref{eq:negRegret}) gives
\begin{align}
\EEik{\sum_{t=\tau_{i,k}}^{\tau_{i,k+1}-1} \left[ x_i(t)-\lcb^*(t) \right]}
&< \frac{\CVV \egp_i L_{ik}+1}{1-\qadv} - \egp_i \EEik{\tau_{i,k+1} - \tau_{i,k} }. \label{eq:ED}
\end{align}
Since the bound in (\ref{eq:ED}) is large for large $L_{ik}$, we show that such a large contribution to the regret can be compensated by negative regret in previous phases due
to the term $- \egp_i[\tau_{i,k+1} - \tau_{i,k}]$.
We show by backward induction over the phases that the expected regret starting from phase~$k$ can be bounded,
\begin{align*}
\EEik{\sum_{t=\tau_{i,k}}^{\ns} \left[ x_i(t)-\lcb^*(t) \right]}
\leq \Phi_i(k,L_{ik}) := L_{ik}\egp_i/2 + 3 L_i^0\egp_i(M-k+1)
\end{align*}
where $M$ is the maximal number of phases from Lemma~\refP{numPhases}.
\begin{lemma} \label{l:recursion}
Let
\[ F_{ik} = \sum_{t=\tau_{i,k}}^{\ns} \left[ x_i(t)-\lcb^*(t) \right] . \]
Then
\begin{align*}
\EEik{F_{ik}} \leq \Phi_i(k,L_{ik}) .
\end{align*}
\end{lemma}
\begin{proof}
Let $k_S$ be the last phase before the algorithm switches to \expp\ with $\tau_{k_S+1}-1=\ns$. By Lemma~\ref{l:prelim}\ref{item:numPhases} we have $k_S \leq M$.
For $k=k_S+1$ the lemma holds trivially since $F_{i,k_S+1}=0$.
By (\ref{eq:ED}) we have
\begin{align*}
\EEik{F_{ik}} \leq \frac{\CVV \egp_i L_{ik} + 1}{1-\qadv} +
\EEik{F_{i,k+1} - \egp_i(\tau_{i,k+1} - \tau_{i,k}) } .
\end{align*}
For the expectation on the right hand side we distinguish three cases, depending on the termination condition of phase $k$ and the value of $L_{ik}$.
\smallskip \noindent
\CASE{1}: Phase $k$ is terminated by the condition in Step 4.d.
Then $L_{i,k+1}=2L_{ik}$ and
\begin{align}
\EEik{\left. F_{i,k+1}- \egp_i(\tau_{i,k+1} - \tau_{i,k})\right| \text{\CASE{1}} }
&\leq \Phi_i(k+1,2L_{ik}) - \egp_i L_{ik} \label{eq:case1}
\end{align}
using the induction hypothesis.
\\This is the case where negative regrets accumulate since~$\CVV/(1-\qadv) < 1$.
\smallskip \noindent
\CASE{2}: Phase $k$ is terminated by the condition in Step 4.a (\footnote
If $k$ is the last phase and the phase is terminated by a condition in Step~1, then the same analysis applies but the value of $L_{k+1,i}$ is irrelevant, since $F_{i,k+1}=0$.})
and $L_{ik} > L_i^0$.\\
Then $L_{i,k+1}=L_{ik}/2$ and
\begin{align}
\EEik{\left. F_{i,k+1}- \egp_i(\tau_{i,k+1} - \tau_{i,k})\right| \text{\CASE{2}} }
&\leq \Phi_i(k+1,L_{ik}/2) .
\end{align}
\smallskip \noindent
\CASE{3}: Phase $k$ is terminated by the condition in Step 4.a
and $L_{ik}=L_i^0$.\\
Then $L_{i,k+1}=L_i^0$ and
\begin{align}
\EEik{\left. F_{i,k+1}- \egp_i(\tau_{i,k+1} - \tau_{i,k})\right| \text{\CASE{3}} }
&\leq \Phi_i(k+1,L_i^0) . \label{eq:case3}
\end{align}
To complete the induction proof, we need to show that for all three cases the right hand side of~(\ref{eq:case1})--(\ref{eq:case3}) is upper bounded by
\[
\Phi_i(k,L_{ik}) - \frac{\CVV \egp_i L_{ik} + 1}{1-\qadv} .
\]
This can be verified by straightforward calculation.
\end{proof}
Now (\ref{eq:bound-bad}) follows from Lemma~\ref{l:recursion} for $k=1$:
\[
\EE{\sum_{t=\nbi}^{\ns} \left[ x_i(t)-\lcb^*(t) \right]}
\leq \Phi_i(1,L_i^0) = \bigO{L_i^0 \egp_i M} = \bigO{\frac{K \log (n/\delta)}{\egp_i}} .
\]
\iffalse
By simple calculus we have the following inequalities for $\Phi(L,E)$:
\begin{align}
\Phi(k,L) &\geq \frac{\CVV \egp_i L + 1}{1-q} + 1 - \egp_i L+ \Phi(k+1,2L) ,
\label{eq:Phi-case1}
\\
\Phi(k,L) &\geq \frac{\CVV \egp_i L + 1}{1-q} + 1 + \Phi(k+1,L/2)
&&\text{for $L > \lceil \Cii/\egp_i^2 \rceil$,}
\label{eq:Phi-case2L}
\\
\Phi(k,L) &\geq \frac{\CVV \egp_i L + 1}{1-q} + 1 + \Phi(k+1,L)
&&\text{for $L = \lceil \Cii/\egp_i^2 \rceil$,}
\label{eq:Phi-case2E}
\end{align}
since
\begin{align*}
\frac{2\CVV}{1-q} \leq 1/2 \leq 1-\frac{\CVV}{1-q},
\\
3 \geq \frac{\CVV(\Cii+1)+1}{1-q} + 1.
\end{align*}
\fi
\section{The stochastic analysis}\label{sec_sto_analysis_K}
\iffalse
{\bf Needed?}\\
Denote the accumulated reward of $i$ from $t+1$ to $t'$ by $\GAitt{i}{t}{t'} = \sum_{\tau=t+1}^{t'}g_{I_\tau}(\tau)\Id{I_\tau=i}$.
We further define the accumulated reward and the deviation of arm $i$ from $t$ to $t'$ as $\Gitt{i}{t}{t'} = \sum_{\tau=t+1}^{t'}g_i(\tau)$ and as $\DAitt{i}{t}{t'} = \sum_{\tau=t+1}^{t'}(g_{I_\tau}(\tau)-\emu_i)\Id{I_\tau=i}$, respectively.
\fi
In this section we assume that all arms $i$ are indeed stochastic with means $\mu_i$.
Recall that $\Delta_i = \mu^* - \mu_i$, $\mu^*=\max_i \mu_i$. We show that with high probability the algorithm does not switch to \expp\ and any suboptimal arm $i$ is chosen at most
$\bigO{\log(n/\delta)/\Delta_i^2}$ times.
We already have from Lemma~\refP{TiB} that with probability $1-\bigO{\delta}$,
$T_i(\nbi,\ns) = \bigO{\numPH/\egp_i^2}$ for all arms. Thus we only need to bound the number of plays before an arm is evicted from $\cA$, $T_i(1,\nbi-1)$.
The next lemma summarizes some properties of \sapo\ against stochastic bandits.
\begin{lemma} \label{l:sto-prelim}
With probability $1-\bigO{\delta}$ the following holds for all time steps $t$ and all arms $i$:
\begin{enumerate}[label={\rm(\alph*)},ref=\alph*,nosep]
\item \label{item:wid-bar}
If $i \in \cA(t)$ then $|\bmu_i(t) - \mu_i| \leq \bwidth(t)/2$.
\item \label{item:wid}
If $i \in \cA(t)$ then $|\hmu_i(t) - \mu_i| \leq \width_i(t)/2$.
\item \label{item:ConfInt}
If $i \in \cA(t)$ then $\bmu_i(t),\mu_i \in [\blcb_i(t),\bucb_i(t)]$ and
$\hmu_i(t),\mu_i \geq \lcb_i(t)$.
\item \label{item:lcb*}
If $\Delta_{i^*}=0$ then $i^* \in \cA(t)$. Furthermore, $\mu^* \geq \lcb^*(t)$.
\item \label{item:DeltaB}
If $i \in \cB(t)$ then $\egp_i \leq 2\Delta_i$.
\end{enumerate}
\end{lemma}
\newcommand{\refSP}[1]{\ref{l:sto-prelim}\ref{item:#1}}
\begin{proof}{\bf (Sketch)}
Statements (\ref{item:wid-bar}) and (\ref{item:wid}) follow from Hoeffding-Azuma's inequality.
Details are given in Appendix~\ref{app:sto-prelim}.
For statement (\ref{item:ConfInt}) we observe that by construction there is a time $s \leq t$ with $\bmu_i(s) - \bwidth(s) = \blcb_i(t)$. Thus (\ref{item:wid-bar}) implies
$\bmu_i(t) \geq \mu_i - \bwidth(t)/2 \geq \bmu_i(s) -\bwidth(s)/2 - \bwidth(t)/2
\geq \bmu_i(s) -\bwidth(s) = \blcb_i(t)$. The other inequalities follow analogously.
Statement (\ref{item:lcb*}) is proven by induction on $t$.
Let $i^*$ be an arm with $\mu_{i^*}=\mu^*$.
If $i^* \in \cA(t-1)$ then we have by (\ref{item:ConfInt}) that $\mu^* \geq \lcb^*(t-1)$.
If any arm~$i$ is evicted at time~$t$, then we have by Step~2.a and~(\ref{item:wid}) that
$\Delta_i = \mu^* - \mu_i \geq \lcb^*(t-1) - \hmu_i(t-1) - \width_i(t-1)/2
\geq (\Cgap - 1/2) \width_i(t-1) > 0$. Thus $i \neq i^*$ and $i^* \in \cA(t)$.
This also shows that when arm~$i$ is evicted,
$\egp_i = \Cgap \cdot \width_i(t-1) \leq \Cgap/(\Cgap-1/2)\Delta_i$, which is statement~(\ref{item:DeltaB}).
\end{proof}
To get a bound on $T_i(1,\nbi-1)$, we show that $\egp_i=\Cgap\cdot\width_i(\nbi-1)$ cannot be too small.
\begin{lemma} \label{l:Delta>}
With probability $1-\bigO{\delta}$ it holds for all times $t$ and all arms~$i \in \cA(t)$ with $T_i(t-1) \geq \Ci \log(n/\delta)$, that
\[ \Cgap \cdot \width_i(t-1) \geq \Delta_i/2 . \]
\end{lemma}
The argument behind the lemma is that if $i \in \cA(t)$ then $\Cgap \cdot \width_i(t-1) \geq \lcb^*(t-1) - \hmu_i(t-1)$ where $\lcb^*(t-1)$ is sufficiently close to $\mu^*$ and $\hmu_i(t-1)$ is sufficiently close to $\mu_i$. The proof is given in Appendix~\ref{app:Delta>}.
Since $i \in \cA(\nbi-1)$, we get from Lemma~\ref{l:Delta>} that with probability $1-\bigO{\delta}$,
\begin{align*}
T_i(\nbi-1)
\leq T_i(\nbi-2) +1
=\frac{\Cwid \log(n/\delta)}{[\width_i(\nbi-2)]^2} + 1
\leq \frac{4 \Cwid \Cgap^2 \log(n/\delta)}{\Delta_i^2} +1 .
\end{align*}
Together with Lemma~\refP{TiB} we have with probability $1-\bigO{\delta}$ that for all arms,
\begin{align} \label{eq:Tns}
T_i(\ns)
\leq \frac{101}{100} L_i^0 M/K + \frac{4 \Cwid \Cgap^2 \log(n/\delta)}{\Delta_i^2} +1
= \bigO{\frac{\log(n/\delta)}{\Delta_i^2}} .
\end{align}
Finally, we need to bound the probability the \sapo\ switches to \expp.
Switching in Step 1.a is already handled by Lemma~\refSP{ConfInt}.
Switching in Step~1.b is also unlikely, since it would mean that the algorithm has accumulated large regret. This contradicts the upper bound~(\ref{eq:Tns}).
Lemma~\ref{l:step1b} shows that \sapo\ switches in Step~1.b only with probability $1-\bigO{\delta}$.
The difficult part, though, is to show that the condition in Step~4.a is not triggered too often such that Step~4.c switches to \expp. We first calculate the false positive rate $\qsto$, the probability that during a given phase the condition in Step~4.a is triggered. The false positive rate is again a constant but small, $\qsto \leq 0.21$, see Lemma~\ref{l:qs}.
Now for a fixed arm the probability that in exactly $E\geq E^0$ out of at most $M$ phases the condition in Step~4.a is triggered, is at most ${M \choose E}\qsto^E$. We set $p=\qsto/(1+\qsto)$ and use a tail bound for the binomial distribution to sum over $E=E^0,\ldots,M$:
\begin{align*}
\sum_{E=E^0}^M {M \choose E}\qsto^E &= (1+\qsto)^M \sum_{E=E^0}^M {M \choose E}p^E (1-p)^{M-E}
\\&\leq (1+\qsto)^M \exp\left\{-M \cdot D(E^0/M || p)\right\}
\end{align*}
where $D(a||p) = a \log\frac{a}{p} + (1-a)\log\frac{1-a}{1-p}$ is the relative entropy. Since $\frac{E^0}{M} \geq \frac{\ci}{2\ci+1/\log 2}$, this sum is $\bigO{\delta/n}$ and a union bound over the arms completes the proof.
\acks{We thank the anonymous reviewers for their very valuable comments. The research leading to these results has received funding from the European Community's Seventh
Framework Programme (FP7/2007-2013) under grant agreement
n$^\circ$ 231495 (CompLACS) and from the Austrian Science Fund (FWF) under contract P~26219-N15.
}
\section{Summary of notations and definitions}
\begin{align*}
T_i(s_1,s_2) &= \#\{t:s_1 \leq t \leq s_2: I_t=i\}
&\text{number of plays of arm $i$} \\
T_i(s) &= T_i(1,s) \\
G_i(s_1,s_2) &= \sum_{t=s_1+1}^{s_2-1} x_i(t) \\
G_i(s) &= G_i(0,s) \\
\hG_i(s_1,s_2) &= \sum_{t=s_1+1}^{s_2-1} x_i(t) \II{I_t=i} & \text{rewards received from arm $i$}\\
\hG_i(s) &= \hG_i(0,s)\\
\hmu_i(s) &= \hG_i(s)/T_i(s) & \text{estimated mean}\\
\width_i(s) &= \sqrt{\Cvii\frac{\log (n/\delta)}{T_i(s)}} & \text{confidence bounds}\\
\lcb_i(s) &= \max\{\lcb_i(s-1),\hmu_i(s) - \width_i(s)\},\\
\ucb_i(s) &= \min\{\ucb_i(s-1),\hmu_i(s) + \width_i(s)\}, \\
\lcb^*(t) &= \max_{1 \leq i \leq K} \lcb_i(t), \\
n^A(s) &= \min\{t: \cA(t) = \cA(s)\}, & \text{time step after the last eviction from $\cA$}\\
\hmu_i^A(s) &= \left. \sum_{t=n^A(s)}^{s} x_i(t) \II{I_t=i} \right/ T_i^A(s),
& \text{estimates from the last eviction}\\
T_i^A(s) &= \#\{t:n^A(s) \leq t \leq s: I_t=i\}, \\
\width_i^A(s) &= \sqrt{ \Cvii\frac{\log (n/\delta)}{T_i^A(s)}} ,\\
\lcb_i^A(s) &= \max\{\lcb_i^A(s-1),\hmu_i^A(s) - \width_i^A(s)\} ,\\
\ucb_i^A(s) &= \min\{\ucb_i^A(s-1),\hmu_i^A(s) + \width_i^A(s)\}, \\
\ucb^*(t) &= \max_i \ucb_i(t), \\
D_i(s_1,s_2) &= \sum_{t=s_1+1}^{s_2} [x_i(t) - \emu_i]
\\
D_i(s) &= D_i(n_i(s),s) \\
\hD_i(s_1,s_2) &= \sum_{t=s_1}^{s_2} [x_i(t) - \emu_i] \II{I_t=i}
& \text{deviation from the estimated mean}\\
\hD_i(s) &= D_i(n_i(s),s) \\
L^0_i & = \lceil 2\Ciii K/\egp_i^2 \rceil
& \text{minimal length of testing phase}\\
E^0 & = \lceil \ci \cdot \log(n/\delta) \rceil
& \text{maximal number of detections in Step 4.a}\\
\numPH & = \lceil \log_2 n\rceil + 2E^0
& \text{maximal number of testing phases for any arm}\\
bla
\end{align*}
\begin{eqnarray*}
\emu_i && \text{frozen estimate of the mean}\\
\egp_i && \text{estimate for the gap}\\
\tau_{i,k} && \text{start time of the $k$-th testing phase for bad arm $i$}\\
p_{ik} && \text{probability of choosing bad arm $i$ in its $k$-th testing phase}\\
L_{ik} && \text{length parameter of the $k$-th testing phase for bad arm $i$}\\
L_i(t) && \text{length parameter of bad arm $i$ in the testing phase containing time $t$}\\
\nbi && \text{time when arm $i$ is evicted from $\cA$}\\
\ns && \text{time when \sapo\ switches to \expp}\\
bla
\end{eqnarray*} |
\section{Generalized dimensional reduction}
\label{sec:Intro}
\subsection{Introduction}
\label{sec:introduction}
In 2003 Brydges and Imbrie discovered a remarkable dimensional reduction
formula that equates the pressure of a gas of hard
spheres in $\bb R^{d}$ with the volume of branched polymers in
$\bb R^{d+2}$~\cite{BI}. See \Cref{fig:HCG-BP} for an illustration of
these models, and \Cref{sec:Intro-BP} for precise definitions. Their
result has very interesting corollaries: it implies
exact enumerative formulas for seemingly intractable high-dimensional
integrals, and it relates the critical behaviour of
$(d+2)$-dimensional branched polymers to the critical behaviour of the
$d$-dimensional hard sphere model. For $d=0,1$ this is a very powerful
reduction.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.8]
\draw[gray,dotted] (0,3) -- (6,3);
\draw[gray,dotted] (3,0) -- (3,6);
\draw[very thick,black, rounded corners] (0,0) rectangle (6,6);
\draw[black,thick, fill = black!30] (1.5,1.5) circle (.5cm);
\draw[black,thick, fill = black!30] (1.1,4.2) circle (.5cm);
\draw[black,thick, fill = black!30] (4.1,4.7) circle (.5cm);
\draw[black,thick, fill = black!30] (2.5,3.5) circle (.5cm);
\draw[black,thick, fill = black!30] (4.5,1.7) circle (.5cm);
\node at (1.5,1.5) {$1$};
\node at (1.1,4.2) {$2$};
\node at (4.1,4.7) {$3$};
\node at (2.5,3.5) {$4$};
\node at (4.5,1.7) {$5$};
\end{tikzpicture}
\qquad\qquad
\begin{tikzpicture}[scale=.8]
\draw[gray,dotted] (0,-3) -- (0,3);
\draw[gray,dotted] (3,0) -- (-3,0);
\draw[very thick,black, rounded corners] (-3,-3) rectangle (3,3);
\path ++(0,0) node (pa) {} ++(170:1cm) node (pb) {} ++(65:1cm)
node (pc) {} ++(130:1cm) node (pd) {};
\path ++(0,0) ++(45:1cm) node (pe) {} ++(10:1cm) node (ph) {};
\path ++(0,0) ++(-70:1cm) node (pf) {} ++(-30:1cm) node (pg) {};
\draw[black, fill = black!30] (pa) circle (.5cm);
\draw[black, fill = black!30] (pb) circle (.5cm);
\draw[black, fill = black!30] (pc) circle (.5cm);
\draw[black, fill = black!30] (pd) circle (.5cm);
\draw[black, fill = black!30] (pe) circle (.5cm);
\draw[black, fill = black!30] (pf) circle (.5cm);
\draw[black, fill = black!30] (pg) circle (.5cm);
\draw[black, fill = black!30] (ph) circle (.5cm);
\node at (pa) {$1$};
\node at (pb) {$4$};
\node at (pc) {$3$};
\node at (pd) {$5$};
\node at (pe) {$2$};
\node at (pf) {$8$};
\node at (pg) {$6$};
\node at (ph) {$7$};
\end{tikzpicture}
\caption{Left: a configuration of hard spheres of equal radius in
$\bb R^{2}$. The adjective hard reflects that the interiors are
disjoint. Right: a branched polymer of disks of equal
radius in $\bb R^{2}$. The configuration has an underlying tree
structure: neighbouring vertices in the tree correspond to tangent
disks. All disks have disjoint interiors. The dotted gray
lines indicate the coordinate axes.}
\label{fig:HCG-BP}
\end{figure}
Brydges and Imbrie's proof of dimensional reduction relied on a
non-constructive supersymmetric localization lemma. Looking to
understand the results of~\cite{BI}, Kenyon and Winkler studied
branched polymers for $d=2,3$ by direct methods~\cite{KW}. Their
proofs are based on an \emph{invariance lemma}. To describe this
lemma, suppose the disk labelled $i$ in a planar branched polymer has
radius $r_{i}$. The invariance lemma states that the total volume
of planar branched polymers is unchanged as the radii $\{r_{i}\}$ are
varied. Inspired by this result, M\'esz\'aros and Postnikov introduced
a model of planar $\cc H$-polymers associated to any central
hyperplane arrangement $\cc H$, and showed that these planar polymers
also satisfy invariance lemmas~\cite{MP}. The main result of this
article is a constructive proof of dimensional reduction formulas for
all $d\geq 2$; invariance lemmas play a central role in the proof.
More precisely, this article establishes that (i) given a central
essential complex hyperplane arrangement, there are dimensional
reduction formulas from $2d+2$ dimensions to $2d$ dimensions, and (ii)
given a central essential real hyperplane arrangement, there are
dimensional reduction formulas from $d+2$ dimensions to $d$
dimensions. These results involve two main objects. The first is a
generalization of branched polymers called $\cc H$-polymers, which are
an extension of the planar polymers in~\cite{MP}. The second is a
generalization of the pressure of the hard sphere gas: for any central
hyperplane arrangement we define an analogue of the Mayer expansion,
which is a power series representation of the pressure of the hard
sphere gas. \Cref{sec:HPBP,sec:MMC} contain the precise definitions of
$\cc H$-polymers and the generalized pressure, and
\Cref{sec:Intro-Theorem} contains a more precise statement of the theorem.
The remainder of this introductory section briefly describes new results and
perspectives that follow from these generalized dimensional reduction
formulas and their proof.
The proof presented in this article yields more information than the
non-constructive proof in~\cite{BI}, and the following results are new
even in the case of branched polymers. First, we obtain a precise
description of the law of $d$-dimensional projections of
$(d+2)$-dimensional polymers, see
\Cref{cor:Projection-Law,cor:Projection-Law-Safe}. This last corollary
can be viewed as a generalization of~\cite[Theorem~7]{KW}, which gave
a description of $1$-dimensional projections of $3$-dimensional
branched polymers. Secondly, our proof of dimensional reduction
extends to some non-spherical bodies, see \Cref{thm:ASA-DR}.
The notion of an $\cc H$-polymer is combinatorially natural, but it
does not immediately connect with statistical mechanics, where
dimensional reduction formulas originated. In
\Cref{sec:SBP-DR} we consider \emph{symmetric hard sphere gases} and
show how $\cc H$-polymers naturally arise. Symmetric hard sphere gases
are models in which, for example, the presence of a sphere at $x_{i}$
implies the presence of a sphere at $-x_{i}$. The bulk properties of
these models coincide with the ordinary hard sphere gas, but there are
corrections to the bulk behaviour due to the symmetry
constraint. The corrections satisfy a dimensional reduction formula:
they can be expressed in terms of $\cc H$-polymers, where
the hyperplane arrangement $\cc H$ reflects the symmetry of the
constraint. See \Cref{fig:SHCG}.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=.8]
\draw[gray,dotted] (0,-3) -- (0,3);
\draw[gray,dotted] (3,0) -- (-3,0);
\draw[very thick,black, rounded corners] (-3,-3) rectangle (3,3);
\draw[black,thick, fill = black!30] (1,1) circle (.5cm);
\draw[black,thick, fill = black!30] (-.6,2.3) circle (.5cm);
\draw[black,thick, fill = black!30] (1.7,2.2) circle (.5cm);
\draw[black,thick, fill = black!30] (-2,2) circle (.5cm);
\draw[black,thick, fill = black!30] (-2,.2) circle (.5cm);
\node at (1,1) {$1$};
\node at (-.6,2.3) {$2$};
\node at (1.7,2.2) {$3$};
\node at (-2,2) {$4$};
\node at (-2,.2) {$5$};
\draw[black,thick, fill = black!30] (-1,-1) circle (.5cm);
\draw[black,thick, fill = black!30] (.6,-2.3) circle (.5cm);
\draw[black,thick, fill = black!30] (-1.7,-2.2) circle (.5cm);
\draw[black,thick, fill = black!30] (2,-2) circle (.5cm);
\draw[black,thick, fill = black!30] (2,-.2) circle (.5cm);
\node at (-1,-1) {$1'$};
\node at (.6,-2.3) {$2'$};
\node at (-1.7,-2.2) {$3'$};
\node at (2,-2) {$4'$};
\node at (2,-.2) {$5'$};
\end{tikzpicture}
\qquad\qquad
\begin{tikzpicture}[scale=.8]
\draw[gray,dotted] (0,-3) -- (0,3);
\draw[gray,dotted] (3,0) -- (-3,0);
\draw[very thick,black, rounded corners] (-3,-3) rectangle (3,3);
\path ++(0,0) ++(30:1cm) node (pa) {} ++(150:1cm) node (pb) {} ++(210:1cm)
node (pc) {} ++(130:1cm) node (pd) {} ++(60:1cm) node (pi) {};
\path ++(0,0) ++(210:1cm) node (pap) {} ++(-30:1cm) node (pbp) {} ++(30:1cm)
node (pcp) {} ++(310:1cm) node (pdp) {} ++(240:1cm) node (pip) {};
\path ++(0,0) ++(30:1cm) ++(-15:1cm) node (pe) {};
\path ++(0,0) ++(30:1cm) ++(70:1cm) node (pf) {} ++(50:1cm) node
(pg) {};
\path ++(0,0) ++(210:1cm) ++(165:1cm) node (pep) {};
\path ++(0,0) ++(210:1cm) ++(250:1cm) node (pfp) {} ++(230:1cm) node
(pgp) {};
\draw[black, fill = black!30] (pa) circle (.5cm);
\draw[black, fill = black!30] (pb) circle (.5cm);
\draw[black, fill = black!30] (pc) circle (.5cm);
\draw[black, fill = black!30] (pd) circle (.5cm);
\draw[black, fill = black!30] (pi) circle (.5cm);
\draw[black, fill = black!30] (pap) circle (.5cm);
\draw[black, fill = black!30] (pbp) circle (.5cm);
\draw[black, fill = black!30] (pcp) circle (.5cm);
\draw[black, fill = black!30] (pdp) circle (.5cm);
\draw[black, fill = black!30] (pip) circle (.5cm);
\draw[black, fill = black!30] (pe) circle (.5cm);
\draw[black, fill = black!30] (pf) circle (.5cm);
\draw[black, fill = black!30] (pg) circle (.5cm);
\draw[black, fill = black!30] (pep) circle (.5cm);
\draw[black, fill = black!30] (pfp) circle (.5cm);
\draw[black, fill = black!30] (pgp) circle (.5cm);
\node at (pa) {$1$};
\node at (pb) {$4$};
\node at (pc) {$3$};
\node at (pd) {$5$};
\node at (pe) {$2$};
\node at (pf) {$8$};
\node at (pg) {$6$};
\node at (pi) {$8$};
\node at (pap) {$1'$};
\node at (pbp) {$4'$};
\node at (pcp) {$3'$};
\node at (pdp) {$5'$};
\node at (pep) {$2'$};
\node at (pfp) {$8'$};
\node at (pgp) {$6'$};
\node at (pip) {$8'$};
\end{tikzpicture}
\caption{Left: a symmetric hard sphere gas configuration associated
to the type $D_{n}$ Coxeter arrangement in $\bb R^{2}$. Right: a $\cc
H$-polymer associated to the type $D_{n}$ Coxeter arrangement in
$\bb R^{2}$. The dotted gray lines indicate the coordinate axes. In
each figure the disks labelled $i$ and $i'$ are located at $x_{i}$ and
$-x_{i}$, respectively. }
\label{fig:SHCG}
\end{figure}
\subsubsection{Related literature}
\label{sec:Outlook}
Dimensional reduction formulas as discussed here first arose in the
context of theoretical physics; see~\cite{BI} for a discussion of this
literature. From a combinatorial viewpoint dimensional reduction
formulas have the flavour of combinatorial reciprocity~\cite{Beck}:
the pressure of the hard sphere gas, which \emph{a priori} makes sense
only for $z>0$, is being given an interpretation for $z<0$.
The dimensional reduction formulas discussed in this article decrease
the dimension by $2$; Imbrie~\cite{I} has obtained a dimensional
reduction formula relating \emph{directed} branched polymers in
$\bb R^{d+1}$ to the hard $\ell_{1}$-sphere gas in $\bb R^{d}$. The methods
of this article can be adapted to give another proof of Imbrie's
result. Similar formulas relating directed objects in $d+1$ dimensions
to undirected objects in $d$ dimensions have arisen often in the
context of random walk representations in statistical mechanics, see,
e.g.,~\cite{BrydgesFrohlichSpencer, FernandezFrohlichSokal, Aizenman,
Helmuth}. It is an open and interesting question to understand when
dimensional reduction results are possible. In particular, are there
formulas involving a reduction in dimension by more than two
dimensions?
\subsubsection{Structure of this article}
\label{sec:structure}
The remainder of this introduction first provides some additional
context by introducing the hard sphere gas, branched polymers, and the
Brydges-Imbrie formula. The connection with the braid arrangement is
described to indicate why hyperplane arrangements are involved. The
introduction concludes with an informal statement of our main result
in \Cref{sec:Intro-Theorem}. A precise formulation and proof of the
main result is given in \Cref{sec:DR-Gen} once the necessary
definitions have been introduced in \Cref{sec:HA}. Many of the
definitions introduces are standard, but we have elected to include
them for the ease of readers from non-combinatorial
backgrounds. Applications of the main result are presented in
\Cref{sec:DR-HCG}.
\subsubsection{Notation and conventions}
\label{sec:notation-conventions}
Throughout the article the term \emph{branched polymer} will refer to
the model studied in~\cite{BI,KW}, i.e., the model that corresponds to
the braid arrangement as outlined in \Cref{sec:Intro-BP}. For other
arrangements $\cc H$ we will always write $\cc H$-polymers. $\bb N$ will
denote the positive integers. For a graph $G=(V,E)$ edges $\{i,j\}\in
E$ will be abbreviated to $ij$, and the notation $ij\in G$ will
indicate $ij\in E(G)$. The integer $d$ will be reserved for the
dimension of a space, while $n$ will count points in a
configuration. Configurations are therefore finite point sets in
$\bb R^{dn}$. We write $2^{A}$ for the set of all subsets of a set $A$,
and $\indicatorthat{A}$ for the indicator function of the set $A$.
\subsection{Branched polymers, the hard sphere gas, and the
Brydges-Imbrie formula}
\label{sec:Intro-BP}
\subsubsection{Branched polymers}
\label{sec:Intro-BP-1}
Let $T$ be a (labelled) spanning tree on $K_{n}$, the complete graph
on $\cb{n} \equiv \{1,2,\dots, n\}$. For integers $d\geq 2$ a \emph{branched
polymer of type $T$ in $\bb R^{d}$} is a configuration of $n$ points
$(x_{1}, \dots, x_{n})$, $x_{i}\in \bb R^{d}$, such that
\begin{enumerate}
\item If $ij\in T$ then $\norm{x_{i}-x_{j}}_{2}=1$,
\item If $ij\notin T$ then $\norm{x_{i}-x_{j}}_{2}>1$,
\end{enumerate}
where $\norm{\cdot}_{2}$ is the Euclidean norm on $\bb R^{d}$. A \emph{branched polymer on $\cb{n}$} is a branched polymer of type
$T$ for some tree $T$ spanning $K_{n}$. Two branched polymers will be
considered equivalent if one is a translation of the other, i.e., the
space of branched polymers on $\cb{n}$ is a subset of
$\bb R^{dn}/\bb R^{d}$. Geometrically, the $x_{i}$ are the centers
of spheres of radius $\frac{1}{2}$, no two spheres have overlapping interiors,
and the tree $T$ determines the tangency graph of the spheres. This
definition of branched polymers was first introduced in~\cite{BI}. See
\Cref{fig:HCG-BP}.
Define $I^{T}_{BP}(x)$ to be $1$ if the points $x = (x_{1},\dots,
x_{n})$ form a branched polymer of type $T$, and define
$I^{T}_{BP}(x)$ to be $0$ otherwise. The \emph{volume of branched
polymers of type $T$ in $\bb R^{d}$}, \emph{volume of $T$} for short, is given by
\begin{equation}
\label{eq:Z-T}
Z^{T}(z) = \int_{\ob{\cc S^{d-1}}^{n-1}} I^{T}_{BP}(x) \prod_{ij\in
T} d\Omega^{d-1}(x_{i}-x_{j}),
\end{equation}
where $\cc S^{d-1}$ is the sphere of radius $1$ in $\bb R^{d}$ and
$\Omega^{d-1}$ is the standard surface measure on $\cc S^{d-1}$.
The partition function for branched polymers is defined by
\begin{equation}
\label{eq:Z-BP}
Z_{BP}^{(d)}(z) = \sum_{n\geq 1} \frac{z^{n}}{n!}\sum_{T\in \cc T\cb{n}} Z^{T}(z),
\end{equation}
where $\cc T\cb{n}$ denotes the set of spanning trees on $K_{n}$,
$z\in\bb R$ is the \emph{activity} of the model, and the superscript $d$
indicates that it is the partition function of branched polymers in
$\bb R^{d}$. For $\abs{z}$ sufficiently small this is a convergent power series.
\subsubsection{The hard sphere gas}
\label{sec:BI-HCG}
The \emph{hard sphere gas} in a finite region $\Lambda\subset \bb R^{d}$ is
the model with partition function
\begin{equation}
\label{eq:Z-HC}
Z^{HC}_{\Lambda}(z) = \sum_{n\geq 0} \frac{z^{n}}{n!} \int_{\Lambda^{n}}
I_{HC}(x)\,\prod_{i=1}^{n}dx_{i},
\end{equation}
where $z\in \bb R$ is the \emph{activity}, $x = (x_{1}, \dots, x_{n})\in
\bb R^{dn}$, and $I_{HC}(x) = \prod_{i\neq j} \indicatorthat{
\norm{x_{i}-x_{j}}_{2} \geq 1}$. The hard-core constraint
$I_{HC}(x)$ means that each point $x_{i}$ can be thought of as the
center of a sphere of radius $\frac{1}{2}$, and that the interiors of the
spheres are pairwise disjoint. $\bb R^{0}$ is considered to be a one-point
space, so $Z_{HC}^{\Lambda}(z) = 1 + z$ when $d=0$. Due to the
hard-core constraint $Z_{\Lambda}^{HC}$ is a polynomial in $z$ for any
finite region $\Lambda$.
\subsubsection{The Brydges--Imbrie dimensional reduction formula}
\label{sec:Intro-BI}
Aside from its intrinsic interest, the following theorem has
many interesting consequences for statistical mechanics,
see~\cite{BI}. Note that the left-hand side involves the hard sphere
gas in $\bb R^{d}$, while the right-hand side involves branched polymers
in $\bb R^{d+2}$.
\begin{theorem}[Brydges-Imbrie~\cite{BI}]
\label{thm:BI}
For all $z$ such that the right-hand side converges absolutely,
\begin{equation}
\label{eq:BI}
\lim_{\Lambda\nearrow \bb R^{d}} \frac{1}{\abs{\Lambda}} \log
Z_{HC}^{(d)}(z) = -2\pi Z_{BP}^{(d+2)}(-\frac{z}{2\pi}),
\end{equation}
where the limit is omitted when $d=0$; the superscript indicates the
dimension in which the model is defined.
\end{theorem}
\subsection{The connection with hyperplane arrangements}
\label{sec:Intro-BP-Braid}
The following gives an alternate description of the space of branched
polymers in $\bb R^{d}$ that establishes a connection with hyperplane
arrangements. This connection was first described, for $d=2$,
in~\cite{MP}.
Recall that the braid arrangement $\cc B_{n}$ is the collection of
hyperplanes $\cc H = \{H_{ij}\}_{1\leq i<j\leq n}$, $H_{ij}$
the hyperplane defined by the linear functional $h_{ij}(x) =
x_{i}-x_{j} = 0$. The bases of the braid arrangement can be
identified with spanning trees in $\cc T\cb{n}$.
\begin{proposition}
\label{def:BP-Alternate}
Define
\begin{align*}
\link(ij) &= \{(x_{1}, \dots, x_{n}) \mid
\norm{h_{ij}(x)}_{2}=1\}
\\
\disj(ij) &= \{(x_{1}, \dots, x_{n}) \mid \norm{h_{ij}(x)}_{2}>1\},
\end{align*}
where both $\link(ij)$ and $\disj(ij)$ are subsets of
$\bb R^{nd}$. The space of $d$-dimensional branched polymers is
the set $P_{\cc B_{n}}(d) = \coprod_{T\in \cc T\cb{n}} P_{\cc
B_{n}}^{T}(d)$, where
\begin{equation}
\label{eq:HPBP-intro}
P_{\cc B_{n}}^{T}(d) = \ob{\bigcap_{ij\in T}\link(ij) \cap
\bigcap_{ij\notin T} \disj(ij)}\big/(1,1,\dots,1)\bb R^{d}.
\end{equation}
\end{proposition}
Verifying this proposition is a matter of translation. We will see
that viewing branched polymers from the perspective of hyperplane
arrangements is fruitful.
\subsection{Informal description of the main result}
\label{sec:Intro-Theorem}
As mentioned in \Cref{sec:introduction}, there are two natural
statistical mechanical objects associated to any central hyperplane
arrangement. The first is a space of a $d$-dimensional \emph{$\cc
H$-polymers} for $d\geq 2$, denoted $P_{\cc H}(d)$, on which there
is a natural volume measure $\vol$. The second is the
$d$-dimensional \emph{pressure} $p^{(d)}_{\cc H}$ of an arrangement $\cc
H$ for $d\geq 0$. For precise descriptions of these objects see
\Cref{sec:HPBP,sec:MMC} respectively.
To establish an analogue of \Cref{thm:BI} requires a sequence
$\vec{\cc H} = (\cc H_{n})_{n\in \bb N}$ of central hyperplane
arrangements, where the arrangement $\cc H_{n}$ is in $\bb R^{n}$. Define
the \emph{$\vec{\cc H}$-polymer partition function} to be
\begin{equation}
\label{eq:Intro-HP-P}
Z_{\vec{\cc H}}^{(d)}(z) = \sum_{n\geq 1} \frac{z^{n}}{n!} \vol(P_{\cc H_{n}}(d)),
\end{equation}
and the \emph{pressure of $\vec{\cc H}$} to be
\begin{equation}
\label{eq:Intro-HP-Pressure}
p^{(d)}_{\vec{\cc H}}(z) = \sum_{n\geq 1} \frac{z^{n}}{n!} p^{d}_{\cc H_{n}}.
\end{equation}
Recall that a hyperplane arrangement is called \emph{essential} if a
maximal linearly independent set of normals forms a basis for the
vector space the arrangement lives in.
\begin{theorem}
\label{thm:Intro-Main}
Let $\vec{\cc H} = (\cc H_{n})_{n\geq 1}$ be a sequence of central
essential real hyperplane arrangements, $\cc H_{n}$ an arrangement
in $\bb R^{n}$, and let $d$ be a non-negative integer. As formal power
series in $z$,
\begin{equation}
\label{eq:thm-Intro-Main}
p^{(d)}_{\vec{\cc H}}(z) = Z^{d+2}_{\vec{\cc H}}(-\frac{z}{2\pi}).
\end{equation}
\end{theorem}
The restriction to real hyperplane arrangements in
\Cref{thm:Intro-Main} is not needed; a similar statement holds for
complex arrangements provided $d$ is even. See \Cref{sec:DR-Gen} for
the precise statement.
\begin{remark}
\label{rem:Reduction-BI}
\Cref{thm:BI} is the special case of \Cref{thm:Intro-Main} when
$\cc H_{n} = B_{n+1}$, the braid arrangement in $\bb R^{n+1}/\bb R$.
Details of this specialization are presented in \Cref{sec:BI-DR}.
\end{remark}
\subsection{Acknowledgements}
\label{sec:acknowledgements}
I would like to thank David Brydges and Matthias Beck for helpful
comments. Parts of this work were completed while I was a Postdoctoral
fellow at ICERM and while visiting SFB 1060 at Universit\"at Bonn, and
I would like to thank both for their support and hospitality. This
work was also partially supported by an NSERC postdoctoral fellowship.
\section{Central hyperplane arrangements and associated objects}
\label{sec:HA}
A \emph{finite hyperplane arrangement $\cc H$} is a finite collection
of hyperplanes in $K^{n}$ for $K$ a field and $n\in \bb N$. We will only
be interested in $K=\bb C$ or $K=\bb R$. An arrangement $\cc H$ is
\emph{central} if each hyperplane in $\cc H$ contains the origin. An
arrangement is \emph{essential} if the normals of the arrangement span
$K^{n}$. The remainder of this article will only involve central
arrangements, so the terms arrangement and central arrangement will be
used synonymously.
An arrangement $\cc H$ can be identified with a set $\{h_{e}\}_{e\in
\cc H}$, $h_{e}\in (K^{n})^{\star}$ the linear functional
that defines the hyperplane $e$. Concretely,
\begin{equation}
\label{eq:HD}
h_{e}(x) = \sum_{i=1}^{n}a_{i}x_{i} = 0, \qquad x = (x_{1}, \dots,
x_{n})\in K^{n}
\end{equation}
is the equation defining the hyperplane $e$, where $a_{i}\in K$ for
$i\in \cb{n}$. The remainder of this section introduces the
objects and properties associated to hyperplane
arrangements relevant for dimensional reduction formulas.
\subsection{Matroids and hyperplane arrangements}
\label{sec:HM}
\subsubsection{Matroids}
\label{sec:Matroid}
This section defines matroids and recalls some needed facts. A
general introduction to matroids can be found in either
of~\cite{Oxley,Welsh}. Readers familiar with matroids may wish to skip
to \Cref{sec:HM-Mat} where the matroids that play a role in this
article are introduced.
\begin{definition}
Let $E$ be a finite set. A \emph{matroid $M = (E,\cc I)$} with \emph{ground
set $E$} is a non-empty collection of subsets $\cc I\subset 2^{E}$, the
\emph{independent sets of $M$}, such that
\begin{enumerate}
\item if $A\in \cc I$ and $B\subset A$, then $B\in \cc I$,
\item if $A,B\in \cc I$, $\abs{A}> \abs{B}$, then there is an
$a\in A\setminus B$ such that $B\cup \{a\} \in \cc I$.
\end{enumerate}
\end{definition}
A \emph{base} of a matroid is a maximal independent set. A subset
$S\subset E$ that is not independent is called \emph{dependent}.
\begin{example}
\label{ex:FE-1}
A fundamental example of a matroid is when $E$ is a finite collection
of vectors in a vector space $V$. Independent sets $A\subset E$ are
collections of linearly independent vectors.
\end{example}
A dependent set $S$ such that $S\setminus\{a\}$ is independent for all
$a\in S$ is called a \emph{circuit}. The following elementary fact
about circuits will be needed, see, e.g.~\cite[Corollary~1.2.6]{Oxley} for a proof.
\begin{proposition}
\label{prop:Fund-Circuit}
Given a base $B\subset E$ and an edge $e\in E\setminus B$ there is
a unique circuit in $B\cup\{e\}$.
\end{proposition}
The circuit identified in \Cref{prop:Fund-Circuit} is known as the
\emph{fundamental circuit of $e$ with respect to $B$}.
\begin{example}
\label{ex:FE-2}
A second fundamental example of a matroid is when $E$ is the edge set of
a connected graph $G$. The independent sets are
cycle-free subgraphs, and the bases are spanning trees.
Circuits correspond to cycles in $G$, and given a spanning tree $T$ the
fundamental circuit of an edge $e\notin T$ is the unique cycle in
the graph $T\cup\{e\}$.
\end{example}
Let $\cc B(M)$ denote the set of bases of a matroid $M=(E,\cc I)$. If
$A,B\in \cc B(M)$ the definition of a matroid implies $A$ and $B$ have
the same cardinality. The cardinality of a base is called the
\emph{rank} of $M$, denoted $\rank(M)$. The rank of a matroid extends
to a function $\rank \colon 2^{E}\to \bb N$, $S\mapsto \rank(S)$, as
follows. For $S\subset E$ define a matroid $M_{S} = (S,\cc I_{S})$
where $\cc I_{S} = \{A\in \cc I \mid A\subset S\}$, and define
$\rank (S) \equiv \rank (M_{S})$. A subset $S\subset E$ is said to be
\emph{spanning} if $\rank(S)=\rank(M)$.
\subsubsection{The characteristic polynomial of a matroid}
\label{sec:HM-CP}
Let $r$ be the rank function of a matroid $M$ with ground set $E$. The
\emph{characteristic polynomial $\chi_{M}$ of $M$} is defined to be
\begin{equation}
\label{eq:HM-CP}
\chi_{M}(t) = \sum_{S\subset E}(-1)^{\abs{S}}t^{r(M)-r(S)}.
\end{equation}
In what follows the relevant evaluation of $\chi_{M}$ will be at
$t=0$, so only subsets of full rank contribute to~\eqref{eq:HM-CP}:
\begin{equation}
\label{eq:HM-CP-Spanning}
\chi_{M}(0) = \sum_{S\subset E}(-1)^{\abs{S}}\indicatorthat{r(S)=r(M)}.
\end{equation}
An alternative representation of the characteristic polynomial in
terms of bases will be useful. Fix a linear order $<$ on the elements
of the ground set $E$. Let $S\in \cc B(M)$. An element $e\in M$ is
called \emph{externally active for $S$} if $e\notin S$ and $e$ is the
minimal element in its fundamental circuit. Following~\cite{KW} say
\emph{$S$ is $<$ safe} if $S$ has no externally
active elements according to the order $<$. The following formula for
$\chi_{M}(0)$ is a specialization of a well-known (see,
e.g.,~\cite{Bjorner}) \emph{activity representation} of the Tutte
polynomial, which contains the characteristic polynomial as a special
case.
\begin{equation}
\label{eq:HM-CP-Bases}
\chi_{M}(0) = \sum_{S\in\cc B(M)} \indicatorthat{\textrm{$S$ is $<$ safe}}.
\end{equation}
\subsubsection{A matroid associated to a hyperplane arrangement}
\label{sec:HM-Mat}
As previously described hyperplane arrangements can be represented by
the set of normals to the hyperplanes. This gives rise to a matroid
$M_{\cc H}$ associated to the arrangement $\cc H$: the ground set
$E = E(M_{\cc H})$ of the matroid is the set of hyperplanes, and the
independent sets are the subsets of hyperplanes whose normals are
linearly independent.
\subsubsection{Convenient identifications and conventions}
\label{sec:HM-Ident}
For simplicity subsets of hyperplanes in an arrangement $\cc H$ will
be conflated with subsets of the ground set $E(M_{\cc H})$ of the
associated matroid. Accordingly, we will use matroid terminology for
subsets of hyperplanes, e.g., we will refer to a base of a hyperplane
arrangment. We will define the characteristic polynomial of a
hyperplane arrangement to be the characteristic polynomial of the
associated matroid.
\subsection{Polymers associated to a
hyperplane arrangements}
\label{sec:HPBP}
Throughout this section fix an essential central hyperplane
arrangement $\cc H$ in $\bb C^{n}$, and choose a positive real number
$R_{e}>0$ for each $e\in \cc H$. The construction in this section is
an extension of a construction in~\cite{MP} from $d=2$ to $d\geq
2$. Recall that elements $e\in \cc H$ can be identified with their defining
linear functionals $h_{e}\colon \bb C^{n}\to \bb C$.
\subsubsection{Construction of $\cc H$-polymers}
\label{sec:HPBP-Construct}
We first define $\cc H$-polymers in $\bb R^{2d}\cong \bb C^{d}$. Given
$e\in \cc H$, $x\in \bb C^{dn}$, define $h_{e}(x) \in \bb C^{d}$ as in~\eqref{eq:HD}:
\begin{equation*}
h_{e}(x) = \sum_{i=1}^{n}a_{i}x_{i},
\end{equation*}
where $a_{i}x_{i}$ is the usual pointwise complex-scalar
multiplication of a vector $x_{i}\in \bb C^{d}$. In the next definition
notice that polymers are defined \emph{without} modding out by
translations, in contrast to what was done for branched polymers in
\Cref{sec:Intro-BP-1}.
\begin{definition}
\label{def:HPBP}
Define
\begin{align*}
\link(e) &= \{x\in \bb C^{nd} \mid \norm{h_{e}(x)}_{2}=R_{e}\}, \\
\disj(e) &= \{x\in \bb C^{nd} \mid \norm{h_{e}(x)}_{2}>R_{e}\}.
\end{align*}
The \emph{space of $2d$-dimensional $\cc H$-polymers with radii
$\{R_{e}\}_{e\in \cc H}$} is the set $P_{\cc H}(2d)\subset
\bb C^{dn}$ defined by
\begin{align}
\label{eq:HPBP-1}
P_{\cc H}(2d) &= \coprod_{ S\in \cc B(M_{\cc H})}
P_{\cc H}^{S}(2d), \\
\label{eq:HPBP-2}
P_{\cc H}^{S}(2d) &= \bigcap_{e\in S} \link(e) \cap
\bigcap_{e\notin S} \disj(e).
\end{align}
\end{definition}
There is a natural probability measure on the space $P_{\cc H}(2d)$,
defined as follows. Injectively map $P_{\cc H}^{S}(2d)$ into $(\cc
S^{(2d-1)})^{n}$ by $ x\mapsto \phi(x) = (\phi_{e}(x))_{e\in S}$,
where $\phi_{e}(x)$ is the unit vector in the direction
$h_{e}(x)$. The fact that this is an injection on $P_{\cc H}^{S}$
follows from the hypothesis that $S$ is a base. With this map in hand define
\begin{equation}
\label{eq:HPBP-Vol}
\vol(P_{\cc H}^{S}(2d)) = \vol( \{ \phi(x) \mid x\in P^{S}_{\cc H}(2d)\}),
\end{equation}
where the volume measure on the right-hand side of \Cref{eq:HPBP-Vol}
is the $n$-fold product of surface measure on $\cc
S^{(2d-1)}$.
As each of the sets $P_{\cc H}^{S}(2d)$ are disjoint, the volume of
$P_{\cc H}^{S}(2d)$ is the sum of the volumes of each
$P_{\cc H}^{S}(2d)$. After normalization, this gives a probability
measure on $P_{\cc H}(2d)$. Note that this
measure agrees with the measure on branched polymers given in
\Cref{sec:Intro-BP}. In general an explicit integral formula for this
measure is given by replacing trees in \Cref{eq:Z-T} with bases $S$ of
$M_{\cc H}$, and the function $I^{T}_{BP}$ with its analogue
$I^{S}_{\cc H}$. The law of $\cc H$-polymers will be thought of as the
law of the locations of the points $x_{i}\in \bb C^{d}$.
To construct $\cc H$-polymers in $d$ dimensions for $d$ odd a further
condition is needed. An arrangement is called \emph{complexified} if
each hyperplane $e\in \cc H$ is defined by a \emph{real} linear
functional, i.e., $a_{i}\in \bb R$ in~\eqref{eq:HD}. In this case
\Cref{def:HPBP} also defines branched polymers in $\bb R^{d}$ by
replacing $\bb C^{d}$ with $\bb R^{d}$. When $d$ is even this construction
for a complexified arrangement agrees with the construction in
\Cref{def:HPBP}.
\begin{remark}
In what follows formulas will be written assuming a
complexified arrangement $\cc H$. For non-complexified arrangements
the formulas continue to hold if $\bb R^{d}$ is replaced by $\bb C^{d}$.
\end{remark}
\begin{proposition}
\label{prop:HPBP-Symmetry}
The law of $\cc H$-polymers is invariant under rotations of $\bb R^{d}$.
\end{proposition}
\begin{proof}
The space of $\cc H$-polymers is rotationally invariant, and the
measure on the space is rotationally invariant.
\end{proof}
The right-hand side of \Cref{thm:Intro-Main} can now be made
precise.
\begin{definition}
\label{def:HPoly-Z}
Let $\vec{\cc H} = (\cc H_{n})_{n\geq 1}$ be a sequence of central
essential hyperplane arrangements. The \emph{partition function} of the
sequence $\vec{\cc H}$ in $\bb R^{d}$ is
\begin{equation}
\label{eq:Gen-Pressure}
Z_{\vec{\cc H}}^{(d)}(z) = \sum_{n\geq 1} \frac{z^{n}}{n!}
\vol(P_{\cc H_{n}}(d)),
\end{equation}
which is to be interpreted as a formal power series in $z$ if
convergence is not known.
\end{definition}
\subsubsection{Geometric interpretation of the space of $\cc
H$-polymers}
\label{sec:HP-Geometric}
There is a geometric interpretation of $\cc H$-polymers extending that
given in~\cite{MP}. To each $e\in S$ there is an associated subspace
$\{x \mid h_{e}(x) = 0\} \subset \bb R^{dn}$, and the set of vectors $x$
such that $\norm{h_{e}(x)}_{2} < R_{e}$ is the set of vectors $v$ that
are at distance less than $R_{e}$ from $\{x\mid h_{e}(x)=0\}$. Call
$\{x\mid \norm{h_{e}(x)}_{2}< R_{e}\}$ the \emph{$R_{e}$-thickening}
of this subspace. The space of $\cc H$-polymers is a union of regions
corresponding to bases $S$; the region associated to $S$ is the
boundary of the intersection of (i) the $R_{e}$-thickenings of the
subspaces associated to $e\in S$ and (ii) the complements of the
$R_{e}$-thickenings of the subspaces associated to $e\notin S$.
\subsubsection{The M\'esz\'aros-Postnikov invariance lemma}
\label{sec:MP}
$\cc H$-polymers in $\bb R^{2}$, i.e., two real dimensions, will be
called \emph{planar $\cc H$-polymers}. Planar $\cc H$-polymers possess
an amazing property: the volume of the space of planar
$\cc H$-polymers is independent of the radii $R_{e}$. The following
theorem is a special case of~\cite[Theorem~1]{MP}.
\begin{theorem}[M\'esz\'aros-Postnikov~\cite{MP}]
\label{thm:MP}
Let $\cc H$ be an essential central arrangement in $\bb C^{n}$. The
volume of the space of planar $\cc H$-polymers with radii
$\{R_{e}\}_{e\in \cc H}$ is
\begin{equation}
\label{eq:MP}
\vol(P_{\cc H}^{S}(2)) = (-2\pi)^{n}\chi_{\cc H}(0),
\end{equation}
where $\chi_{\cc H}$ is the characteristic polynomial of the
arrangement $\cc H$.
\end{theorem}
For $\cc H$ the braid arrangement $\cc B_{n}$ \Cref{thm:MP} was first
proven in~\cite{KW}.
\subsection{Mayer coefficients and their generalizations}
\label{sec:MMC}
The dimensional reduction formula, \Cref{thm:BI}, relates the pressure
of a hard sphere gas in $d$ dimensions to the partition function of
branched polymers in $d+2$ dimensions. The
\emph{Mayer expansion} for the pressure, which is recalled in \Cref{sec:BI-DR}
below, gives a relationship between the \emph{Mayer coefficients}
of the hard sphere gas and branched polymers. This section defines a
generalization of Mayer coefficients that are associated to a central
essential hyperplane arrangement $\cc H$.
\subsubsection{Matroidal Mayer coefficients}
\label{sec:MMC-Def}
Let $H\subset E(M_{\cc H})$ be a spanning set of the matroid
$M_{\cc H}$, and assume $\cc H$ is complexified. The $d$-dimensional
matroidal Mayer coefficient (MMC) associated to $H$ is the number
\begin{equation}
\label{eq:MMC}
\int_{\bb R^{dn}} \prod_{e\in H}\ob{
-\indicatorthat{\norm{h_{e}(x)}_{2}\leq R_{e}}} \, dx.
\end{equation}
Note that this is a finite number: because $H$ is a spanning set each
$x_{i}$ is constrained to lie in a bounded subset of $\bb R^{d}$. For
non-complexified arrangements the definition of the matroidal Mayer
coefficients is the same, with $\bb R^{d}$ replaced by $\bb C^{d}$
in~\eqref{eq:MMC}. When $d=0$ the convention that $\bb R^{0}$ is a
one-point space means~\eqref{eq:MMC} is equal to $(-1)^{\abs{H}}$. The
left-hand side of \Cref{thm:Intro-Main} can now be made precise:
\begin{definition}
\label{def:Gen-Pressure}
Let $\vec{\cc H} = (\cc H_{n})_{n\geq 1}$ be a sequence of central
essential hyperplane arrangements. The \emph{pressure} of the
sequence $\vec{\cc H}$ in $\bb R^{d}$ is
\begin{equation}
\label{eq:Gen-Pressure}
p^{(d)}_{\vec{\cc H}}(z) = \sum_{n\geq 1} \frac{z^{n}}{n!}\mathop{\sum_{H\subset
E(M_{\cc H_{n}})}}_{\rank(H) = n} \int_{\bb R^{dn}} \prod_{e\in H}
-\indicatorthat{\norm{h_{e}(x)}_{2}\leq R_{e}} \, dx,
\end{equation}
which is to be interpreted as a formal power series if convergence is
not known.
\end{definition}
\subsubsection{Geometric interpretation}
\label{sec:MMC-Geometric}
Following \Cref{sec:HP-Geometric} there is a natural geometric
interpretation of the $d$-dimensional MMC associated to a subset
$H$. To each hyperplane $e$ is associated the $R_{e}$-thickening of the
subspace $\{x \mid h_{e}(x)=0\}$ in $\bb R^{nd}$. The MMC associated to a
spanning set $H$ is the (signed) volume of the intersection of the
thickened subspaces corresponding to $e\in H$.
\section{Dimensional reduction formulas for $\cc H$-polymers}
\label{sec:DR-Gen}
\subsection{Proof of dimensional reduction formulas}
\label{sec:DR-BP}
In this section we prove a dimensional reduction formula for $\cc
H$-polymers when $\cc H$ is an essential arrangement. We
standardize to $R_{e}=1$ for each $e\in \cc H$.
Let $\Omega^{d-1}_{a}$ denote the surface measure on the sphere of
radius $\sqrt{a}$ in $\bb R^{d}$, and let $\lambda^{d}_{B}$ denote
Lebesgue measure on the unit ball in $\bb R^{d}$. The next lemma, which
states that a codimension $2$ projection of the surface measure on the
unit sphere in $\bb R^{d+2}$ is the uniform measure on the unit ball in
$\bb R^{d}$, is a well-known calculation, and the proof is omitted.
\begin{lemma}[Generalized Archimedes's theorem]
\label{lem:Archimedes}
Let $(w,y)\in \cc S^{d+1}\subset \bb R^{d+2}$, $w\in \bb R^{2}$, $y\in
\bb R^{d}$. Then
\begin{equation}
\label{eq:Archimedes}
d\Omega^{d+1}_{1}(w,y) =
d\lambda^{d}_{B}(y)d\Omega^{1}_{1-\norm{y}_{2}^{2}}(w).
\end{equation}
\end{lemma}
\begin{theorem}
\label{thm:DR}
Let $\cc H$ be an essential central complexified hyperplane
arrangement in $\bb C^{n}$. Then
\begin{equation}
\label{eq:DR}
\int_{\bb R^{dn}} \mathop{\sum_{H \subset E(M_{\cc H})}}_{r(H)=n} \prod_{e\in H}
-\indicatorthat{\norm{h_{e}(x)}_{2}\leq 1} \, dx = (-2\pi)^{n}\vol (P_{\cc H}(d+2)).
\end{equation}
If $\cc H$ is not complexified then~\eqref{eq:DR} still holds
provided $d$ is even and $\bb R^{2dn}$ is identified with $\bb C^{dn}$.
\end{theorem}
\begin{proof}
It will be assumed $\cc H$ is a complexified arrangement;
\emph{mutatis mutandis} the argument applies for arrangements
that are not complexified. The proof manipulates each side of
\Cref{eq:DR} separately and observes that the resulting expressions
are the same.
First, rewrite the left-hand side of \eqref{eq:DR} by subdividing
the region of integration according to whether or not
$\norm{h_{e}(x)}_{2}\leq 1$. Formally,
\begin{align}
\label{eq:DR-ROI}
\bb R^{dn} &= \coprod_{G\subset E(M_{\cc H})} \Gamma_{G}, \\
\label{eq:DR-Regions}
\Gamma_{G} &= \{ x\in \bb R^{dn} \mid
\textrm{$\norm{h_{e}(x)}_{2}\leq 1$ if and only if $e\in G$}\},
\end{align}
with the disjoint union in~\eqref{eq:DR-ROI} being over all subsets of
the ground set of $M_{\cc H}$.
Each summand $H$ in the left-hand side of \Cref{eq:DR} is a spanning
set. The integral over any region $\Gamma_{G}$ is zero when $G$ is
not a spanning set: in this case there is a normal to a hyperplane
$e\in H$ with $e\notin G$, and the integral of
$\indicatorthat{\norm{h_{e}(x)}_{2}\leq 1}$ over
$\{x \mid \norm{h_{e}(x)}_{2}>1\}$ vanishes. The left-hand side of
\Cref{eq:DR} can therefore be rewritten as:
\begin{align}
\label{eq:DR-1a}
&\hspace{-16mm}
\mathop{\sum_{G\subset E(M_{\cc H})}}_{r(G)=n}
\mathop{\sum_{H \subset E(M_{\cc H})}}_{r(H)=n}
\int_{\Gamma_{G}}\prod_{e\in H}
-\indicatorthat{\norm{h_{e}(x)}_{2}\leq 1} \, dx \\
\label{eq:DR-1c}
&=
\mathop{\sum_{G\subset E(M_{\cc H})}}_{r(G)=n}
\mathop{\sum_{H \subset G}}_{r(H)=n}
\int_{\Gamma_{G}} (-1)^{\abs{H}}\, dx \\
\label{eq:DR-1b}
&= \mathop{\sum_{G\subset E(M_{\cc H})}}_{r(G)=n} \vol(\Gamma_{G})
\chi_{G}(0).
\end{align}
The first equality follows as if $H$ is not a subset of $G$ the
integral vanishes. The second equality follows from
\eqref{eq:HM-CP-Spanning}. This concludes the manipulations of the
left-hand side of \Cref{eq:DR}.
The second step is to perform the integrals in the right-hand side
of \Cref{eq:DR}. The idea of how to do this is simple. The volume
is computed by first fixing the last $d$ coordinates of the points
$x_{i}$ in a $\cc H$-polymer, and then integrating over the first
two coordinates of each point. The integral over the first two
coordinates can be expressed as the volume of a generalized planar
polymer, and hence can be computed explicitly by an invariance
lemma. Lastly we integrate over the last $d$ coordinates. This idea
is similar to what was done for branched polymers in three
dimensions in~\cite{KW}.
As in \Cref{lem:Archimedes}, it will be convenient to write
$x\in P_{\cc H}(d+2)$ as $x = (w,y)$, where
$w = (w_{1},w_{2}, \dots, w_{n})\in \bb R^{2n}$,
$y = (y_{1}, y_{2}, \dots, y_{n})\in \bb R^{dn}$, and
$x_{i} = (w_{i},y_{i})\in \bb R^{d+2}$. To each point
$x\in P_{\cc H}(d+2)$ associate a vector
$(R^{\star}_{e}(y))_{e\in \cc H}$ of non-negative radii:
\begin{equation}
\label{eq:DR-2}
R^{\star}_{e}(y)^{2} = (1-\norm{h_{e}(y)}_{2}^{2})\wedge 0,
\end{equation}
where $a\wedge b$ denotes the minimum of $a$ and $b$. As the
$h_{e}$ are linear functionals,
$\norm{h_{e}(w,y)}_{2}^{2} = \norm{h_{e}(w)}_{2}^{2} +
\norm{h_{e}(y)}_{2}^{2}$.
It follows that if $S$ is a base of $\cc H$ and $x\in P_{\cc H}^{S}(d+2)$ then
\begin{align}
\label{eq:DR-2a}
\norm{h_{e}(w)}_{2}^{2}
&= R_{e}^{\star}(y), \quad\,\,\, e\in S, \\
\label{eq:DR-2b}
\norm{h_{e}(w)}_{2}^{2}
&> R_{e}^{\star}(y)^{2}, \quad e\notin S,
\end{align}
since $\norm{h_{e}(w,y)}_{2}^{2}$ equals $1$ for $e\in S$, and is
greater than $1$ for $e\notin S$. \Cref{lem:Archimedes} gives a
concrete expression for $\vol(P^{S}_{\cc H})$ when $S$ is a
base. Letting $d\Omega_{e}(w)$ denote
$d\Omega^{1}_{R^{\star}_{e}(y)^{2}} (h_{e}(w))$ and
$d\lambda_{e}(y)$ denote $d\lambda^{d}_{B}(h_{e}(y))$, the expression is
\begin{equation}
\nonumber
\int_{P^{S}_{\cc H}} \prod_{e\in S} d\Omega^{d+1}_{1}(\phi_{e}(x)) \\
\int_{\bb R^{(d+2)n}}\!\prod_{e\in \cc H\setminus S}\!
\indicatorthat{ \norm{h_{e}(w,y)}_{2}^{2}\geq 1}
\prod_{e\in S} d\Omega_{e}(w)d\lambda_{e}(y).
\end{equation}
As before it is helpful to decompose the region of integration:
\begin{equation}
\label{eq:DR-ROI-2}
\bb R^{(d+2)n} = \coprod_{G\subset E(M_{\cc H})} \bb R^{2n}\times \Gamma_{G},
\end{equation}
where $\Gamma_{G}$ is the subset of $y$-coordinates defined as in
\Cref{eq:DR-Regions}. In the rest of the proof we abbreviate
$G\subset E(M_{\cc H})$ to $G\subset \cc H$. An argument similar
to the one leading to~\eqref{eq:DR-1a} shows that the integral
over a region $\bb R^{2n}\times\Gamma_{G}$ can be non-zero only if
$S\subset G$, and hence the volume is given by
\begin{equation}
\label{eq:DR-6}
\sum_{G\colon S\subset G}
\int_{\bb R^{2n}\times \Gamma_{G}}
\prod_{e\in \cc H \setminus S}
\indicatorthat{ \norm{h_{e}(w)}_{2}^{2}> R_{e}^{\star}(y)^{2}}
\prod_{e\in S} d\Omega_{e}(w) d\lambda_{e}(y).
\end{equation}
Fix $G$ containing $S$, so $G$ is rank $n$. If $e\notin G$ then
$y\in \Gamma_{G}$ implies $R_{e}^{\star}(y)=0$. The non-trivial
constraints in \eqref{eq:DR-6} on $w$ therefore correspond to
hyperplanes $e\in G$. Letting $\cc H_{G}$ denote the hyperplane
arrangement consisting of hyperplanes in $G$ this implies the
first product in \eqref{eq:DR-6} can be restricted to $e\in \cc
H_{G}\setminus S$. Summing \eqref{eq:DR-6} over all bases $S$ to
compute $\vol(P_{\cc H}(d+2))$ results in
\begin{equation}
\label{eq:DR-8}
\mathop{\sum_{G\subset \cc
H}}_{\rank(G) = n} \sum_{S\in \cc B(M_{\cc H_{G}})}
\int_{\bb R^{2n}\times \Gamma_{G}} \prod_{e\in \cc H_{G} \setminus S}
\!\!\indicatorthat{ \norm{h_{e}(w)}_{2}^{2}> R_{e}^{\star}(y)}
\prod_{e\in S} d\Omega_{e}(w) d\lambda_{e}(y).
\end{equation}
The sum over $S\in \cc B(M_{\cc H_{G}})$ of the integrals over $w$
in \eqref{eq:DR-8} are, for any fixed $y$, precisely the volume of
planar $\cc H_{G}$ polymers with radii $R^{\star}_{e}(y)$ for
$e\in \cc H_{G}$. By \Cref{thm:MP}
\begin{align}
\label{eq:DR-10}
\vol(P_{\cc H}(d+2)) &= \mathop{\sum_{G\subset \cc H}}_{\rank(G) = n} \int_{y\in
\Gamma_{G}} (-2\pi)^{n}\chi_{G}(0)
\prod_{e\in S} d\lambda^{d}_{B}(h_{e}(y)) \\
&= \mathop{\sum_{G\subset M_{\cc
H}}}_{\rank(G) = n} (-2\pi)^{n}\chi_{G}(0)\vol(\Gamma_{G}),
\end{align}
which is exactly~\eqref{eq:DR-1b}.
\end{proof}
\begin{proof}[Proof of \Cref{thm:Intro-Main}]
\Cref{thm:Intro-Main} has been made precise by
\Cref{def:HPoly-Z,def:Gen-Pressure}. To prove the theorem, apply
\Cref{thm:DR} to each arrangement $\cc H_{n}$ in the sequence
$(\cc H_{n})_{n\in \bb N}$, multiply each term by $\frac{z^{n}}{n!}$,
and sum over $n$.
\end{proof}
\subsection{Laws of projections}
\label{sec:laws-projections}
The proof of \Cref{thm:DR} established more than was stated. The
entire proof can be conducted without computing the integrals over
the last $d$ coordinates, i.e., with $y$ fixed. This implies the
law of a $d$-dimensional projection of a $(d+2)$-dimensional
$\cc H$-polymer is given by the law of the MMC
coefficients. Formally,
\begin{corollary}
\label{cor:Projection-Law}
Let $g$ be a function of the last $d$ coordinates in $\bb R^{d+2}$,
integrable with respect to the law of $\cc H$-polymers, where $\cc
H$ is an essential central complexified arrangement of rank
$n$. Then
\begin{equation}
\nonumber
\int_{P_{\cc H}(d+2)}g(y)\,d\vol(w,y) =
(-2\pi)^{n}\!\!\! \mathop{\sum_{G\subset E(M_{\cc H})}}_{\rank(G) = n}
\int_{\bb R^{dn}} g(y)\prod_{e\in G} -\indicatorthat{\norm{h_{e}(y)}\leq 1}\,dy.
\end{equation}
\end{corollary}
The previous corollary is unnatural since the MMC are
distributed according to a signed measure. This can be improved. Given
a set $E$, let $E^{<}$ denote the set of linear orders on $E$. We call
a function $f\colon \bb R^{d} \to E^{<}$ an \emph{ordering function}.
\begin{corollary}
\label{cor:Projection-Law-Safe}
Let $g$ be a function of the last $d$ coordinates in $\bb R^{d+2}$,
integrable with respect to the law of $\cc H$-polymers, where $\cc
H$ is an essential central complexified arrangement of rank
$n$. Let $f$ be an ordering function.
\begin{equation}
\label{eq:Projection-Law-Safe}
\int_{P_{\cc H}(d+2)}g(y)\,d\vol(w,y) = (-2\pi)^{n} \sum_{S\in \cc B(M_{\cc H})}
\int_{\bigcup_{G}\Gamma_{G}} g(y) \indicatorthat{\textrm{$S$
is $f(y)$ safe}}\, dy,
\end{equation}
where the union in the region of integration is over all $G\subset
\cc H$ such that $\rank(G)=n$.
\end{corollary}
\begin{proof}
\Cref{eq:HM-CP-Bases} implies that for any ordering function
$f$ and any $y$
\begin{equation}
\label{eq:Safe-x}
\chi_{M}(0) = \sum_{S\in \cc B(M)} \indicatorthat{\textrm{$S$ is
$f(y)$ safe}}.
\end{equation}
Inserting this expression into \Cref{eq:DR-1b} gives the corollary,
as it is a rewriting of \Cref{cor:Projection-Law}.
\end{proof}
\begin{remark}
\label{rem:Proj-1}
For branched polymers in $d=3$ \Cref{cor:Projection-Law-Safe} was
established for a particular ordering function $f$ in~\cite{KW}.
\end{remark}
\begin{remark}
\label{rem:Proj-2}
There is a notion of \emph{embedding activity} for
embedded graphs $G$ due to Bernardi~\cite{Bernardi}, who has shown
that~\eqref{eq:HM-CP-Bases} holds when external activity is replaced
with external embedding activity. Bernardi's result, together with
an argument as in \Cref{cor:Projection-Law-Safe}, yields an
explicit probability law for the $2d$-projection of $4d$-branched
polymers.
\end{remark}
\subsection{Non-spherical bodies}
\label{sec:APP}
\Cref{thm:DR} did not make essential use of the fact that the measure
on $\cc H$-polymers was induced from the surface measure on unit
spheres. The key ingredient was only that the surface measure
factorized into a product of the surface measure on $\cc S^{1}$ and
Lebesgue measure on the codimension $2$ projection. The next
definition introduces a class of non-spherical objects for which the
proof of \Cref{thm:DR} applies. The definition is a specialization of
more general concepts introduced in~\cite{CollDoddHarrison}, which
studies when generalizations of \Cref{lem:Archimedes} hold.
\begin{definition}
\label{def:ASA}
A \emph{spherical array} in $\bb R^{d}$ is a hypersurface
$\asa = \cc S^{1}\times_{\warp}\asabase{\asa}$, where
\begin{equation}
\label{eq:SA}
\cc S^{1}\times_{\warp}\asabase{\asa} = \{ x\in \bb R^{d} \mid x_{1}^{2}
+ x_{2}^{2} = \cb{\warp(x_{3}, \dots, x_{d})}^{2}\}.
\end{equation}
The function $\warp\colon \bb R^{d-2} \to \co{0,\infty}$ is the
\emph{warping function} and $\asabase{\asa}$ is the \emph{bottom}
of $\asa$. Let $\pi_{d-2}$ denote the orthogonal projection from
$\bb R^{2}\times \bb R^{d-2}\to \bb R^{d-2}$ defined by $\pi_{d-2}(w,y)=y$.
A spherical array is an \emph{Archimedean spherical array (ASA)} if
for all measurable $U\subset \asabase{\asa}$
\begin{equation}
\label{eq:APP}
\Omega(\pi^{-1}_{d-2}(U)) = \vol(U),
\end{equation}
where $\Omega$ is the surface measure on $\asa$ induced from
$\bb R^{d}$, and $\vol$ is Lebesgue measure on $\bb R^{d-2}$.
\end{definition}
The warping function of an ASA must be rather special,
see~\cite{CollDoddHarrison}. Several ASAs are well-known.
\begin{example}
\label{ex:ASA-Arch}
Spheres $\cc S^{d-1}$ are ASAs, with bottom
the unit ball $B^{d-2}$ in $\bb R^{d-2}$ and warping function
$\sqrt{1-x_{d-1}^{2}-x_{d}^{2}}$. This is the content of
\Cref{lem:Archimedes}.
\end{example}
\begin{example}
\label{ex:ASA-Cyl}
Cylinders $\cc S^{d-2}\times I$ with $I$ an interval in $\bb R$ are
ASAs with base $B^{d-3}\times I$ the solid cylinder in
$\bb R^{d-2}$. This follows from writing $\cc S^{d-2}$ as a warped
product $\cc S^{1}\times_{\warp} B^{d-3}$ as in the previous
example, and noting this gives a warped product $\cc
S^{1}\times_{\warp} (B^{d-3}\times I)$, where the warping function
is independent of the coordinate in $I$.
\end{example}
\begin{example}
\label{ex:ASA-Spherical-Cyl}
Spherically capped cylinders, i.e., the boundary of
$\cc B^{d-1}\times I$, are ASAs. This follows by combining the last
two examples.
\end{example}
Dimensional reduction formulas for ASAs require defining the
associated spaces of polymers and Mayer coefficients. The remainder of
this section indicates these definitions, with the conclusion being
the next theorem. Once the definitions are given the proof is,
\emph{mutatis mutandis}, the same as the proof of \Cref{thm:DR}, and
hence it is omitted.
\begin{theorem}
\label{thm:ASA-DR}
\Cref{thm:DR} holds for Archimedean spherical arrays.
\end{theorem}
\begin{remark}
In the case of the braid arrangment and open cylinders,
\Cref{thm:ASA-DR} follows by the methods of~\cite{BI}.
\end{remark}
Let $\cc H = \{h_{e}\}_{e\in E}$ be a hyperplane arrangement in
$\bb R^{n}$. Associate to each $e\in E$ an ASA $\asa_{e}$; by a slight
abuse of notation write $\asa$ for this set of ASAs. Define subsets
of $\bb R^{dn} = \{(w,y)\mid w\in \bb R^{2n},y\in \bb R^{(d-2)n}\}$ by
\begin{align*}
\link(e) &= \{(w,y) \mid h_{e}(w,y) \in \asa_{e}\}, \\
\disj(e) &= \{ (w,y) \mid h_{e}(y)\in \asabase{\asa_{e}} \textrm{ if
and only if } \norm{h_{e}(w)}_{2}^{2} > \cb{\warp(h_{e}(y))}^{2}\},
\end{align*}
where the ASAs $\asa_{e}$ are left implicit in the notation. \emph{The
space of $\cc H$-polymers of type $\asa$}, denoted
$\cc P_{\cc H,\asa}$, is defined by
\begin{align}
\label{eq:ASA-Poly}
\cc P_{\cc H,\asa} &= \coprod_{S\in \cc B(\cc H)} \cc P_{S,\asa} \\
\label{eq:ASA-Base}
\cc P_{S, \asa} &= \bigcap_{e\in S} \link(e) \cap \bigcap_{e\notin
S} \disj(e).
\end{align}
$\cc H$-polymers as introduced in \Cref{sec:HPBP} are the special case
of $\cc H$-polymers of type $\asa$ when $\asa_{e} = \cc S^{d-1}$ for
all $e\in E$.
There is a natural measure $\Omega_{\asa}$ on $\cc H$-polymers of type
$\asa$ induced by the surface measures on the hypersurfaces
$\asa_{e}$. Define the measure $\Omega_{\asa}$ on $\cc P_{S,\asa}$ to
be the pushforward of the product measure on the ASAs
$\{\asa_{e}\}_{e\in S}$ under the map $x\mapsto (h_{e}(x))_{e\in S}$,
i.e.,
\begin{equation}
\label{eq:ASA-SM}
d\Omega_{\asa}(w,y) = \prod_{e\in S} d\Omega_{\asa_{e}}(h_{e}(w,y)).
\end{equation}
Normalizing this measure gives a probability measure on $\cc P_{\cc
H,\asa}$-polymers.
The definition of the matroidal Mayer coefficients is essentially the
same as in \Cref{sec:MMC}. For a spanning set $H\subset E(M_{\cc H})$
and a collection of ASAs $\asa$, the \emph{$d$-dimensional MMC of type
$\asa$} is given by
\begin{equation}
\label{eq:MMC-ASA}
\int_{\bb R^{dn}}\prod_{e\in H}(-\indicatorthat{h_{e}(x)\in
\asabase{\asa_{e}}})\, dx.
\end{equation}
\section{Applications}
\label{sec:DR-HCG}
For particular sequences $\vec{\cc H}$ of hyperplane arrangements the
pressure $p_{\vec{\cc H}}^{d}(z)$ arises naturally when studying
models in statistical mechanics. This section provides
examples. \Cref{sec:BI-DR} focuses on the hard sphere gas:
\Cref{sec:BI-DR-Proof} gives a new proof of \Cref{thm:BI}, while
\Cref{sec:BI-Variations} explains multi-type hard sphere gases.
\Cref{sec:SBP-DR} shows that type $D_{n}$ Coxeter arrangements arise
in the statistical mechanics of a symmetrized hard sphere gas, and
briefly describes some variations on this theme.
\subsection{The Brydges-Imbrie dimensional reduction formula}
\label{sec:BI-DR}
\subsubsection{Proof of the Brydges-Imbrie formula}
\label{sec:BI-DR-Proof}
\begin{proof}[Proof of \Cref{thm:BI}]
Mayer's theorem represents the pressure of a statistical mechanical
model in terms of cluster coefficients associated to connected
graphs~\cite{BrydgesSC}. For the hard-core gas of spheres with
radius $\frac{1}{2}$ Mayer's theorem states that
\begin{equation}
\label{eq:HC-Log}
\lim_{\Lambda\nearrow \bb R^{d}} \frac{1}{\abs{\Lambda}} \log Z_{HC}(z)
= \sum_{n\geq 0}\frac{z^{n}}{n!} \int_{\bb R^{dn}/\bb R^{d}} \sum_{H\in
\cc G^{c}\cb{n}} \prod_{ij\in
H}-\indicatorthat{\norm{x_{i}-x_{j}}_{2}\leq 1}\, dx,
\end{equation}
where $\cc G^{c}\cb{n}$ denotes the set of connected graphs on
$\cb{n}$ and $\bb R^{dn}/\bb R^{d}$ indicates translations are modded out.
The right-hand side of \Cref{eq:HC-Log} is the pressure of the braid
arrangement $\cc B_{n}$ in $\bb C^{n}/(1,1,\dots,1)\bb C\cong \bb C^{n-1}$,
as the matroid associated to $\cc B_{n}$ is the graphical matroid of
$K_{n}$. \Cref{thm:BI} therefore follows from \Cref{thm:DR}; the
extra factor of $-2\pi$ arises as the arrangement is rank $(n-1)$.
\end{proof}
\subsubsection{Variations on the theme}
\label{sec:BI-Variations}
Several variations on this result are known to
exist~\cite{BIb,Cardy}. Rather than be exhaustive, we will just
highlight one variation and its phrasing in terms of hyperplane
arrangements.
Consider the following variant of the braid arrangement. Fix $k\in
\bb N$, $n_{i}\in \bb N$ for $i\in \cb{k}$, and let $n=
\sum_{i=1}^{k}n_{i}$. Define an arrangement in $\bb C^{n}$ to be the set
of hyperplanes with normals $h_{i'j'}^{(r)} =
x^{(r)}_{i'}-x^{(s)}_{j'}$ for $i'\in \cb{n_{i}}$, $j'\in \cb{n_{j}}$,
and $r,s\in \cb{k}$, $r\neq s$.
In the statistical mechanics picture this corresponds to a gas of
spheres of $k$ different colours, with $n_{i}$ spheres of colour
$i$. Spheres of the same colour do not interact, while spheres of
distinct colours are required to be disjoint. For $k=2$ this is known
as the \emph{Widom-Rowlinson model}. The corresponding branched
polymers have trees as tangency graphs, and are restricted to (i) have
tangent spheres be of different colours and (ii) have spheres of
different colours be disjoint.
\subsection{Dimensional reduction in the presence of symmetry
constraints}
\label{sec:SBP-DR}
This section describes dimensional reduction formulas for gases of
hard spheres subject to symmetry
constraints. \Cref{sec:SBP-Models,sec:SBP-Mayer,sec:SBP-Bal,sec:SBP-Cn-DR}
gives a detailed account of the type $D_{n}$ Coxeter arrangement;
similar arguments apply to other models which are briefly described in
\Cref{sec:SBP-DR-Variations}.
\subsubsection{Symmetric hard sphere models}
\label{sec:SBP-Models}
The type $D_{n}$ Coxeter arrangement has hyperplanes defined by the
linear functionals
\begin{equation}
\label{eq:SBP-Cn}
h^{\pm}_{ij}(x) = x_{i}\pm x_{j}, \qquad i\neq j \in \cb{n}.
\end{equation}
The \emph{type $D$ hard sphere model} has partition function
\begin{equation}
\label{eq:Cn-HCG-1}
Z^{D}_{\Lambda}(z) = \sum_{n\geq 0} \frac{z^{n}}{n!}\int_{\Lambda^{n}} \prod_{i=1}^{n}
\indicatorthat{\norm{x_{i}-x_{j}}_{2}\geq 1}
\indicatorthat{\norm{x_{i}+x_{j}}_{2}\geq 1}\, dx,
\end{equation}
where $\Lambda\subset \bb R^{d}$ is a box centered at the origin. The
constraint $\norm{x_{i}-x_{j}}_{2}\geq 1$ is the usual constraint for
hard spheres of radius $\frac{1}{2}$. The constraint $\norm{x_{i}+x_{j}}_{2}$
is a hard sphere constraint between the sphere at $x_{i}$ and the
mirror image $-x_{j}$ of the sphere at $x_{j}$. Prosaically, this is a
model of hard spheres that cannot distinguish between other spheres
and the mirror images of other spheres. Recall \Cref{fig:SHCG}.
Alternately, the formula for the partition function in
\Cref{eq:Cn-HCG-1} can be rewritten as a hard sphere gas in the upper
half space $\bb R^{d-1}\times \bb R_{+}$, and the condition
$\norm{x_{i}+x_{j}}_{2}\geq 1$ can be interpreted as a boundary
condition. Since the pressure of a hard sphere gas is independent of
the boundary conditions, the pressure of this model can be represented
as the partition function of branched polymers in $d+2$ dimensions by
\Cref{sec:BI-DR}. This is verified explicitly in
\Cref{sec:SBP-Bal}. More interestingly, there is a $D_{n}$-polymer
representation for the lowest-order finite volume corrections to
$Z^{D}_{\Lambda}$ as $\Lambda\uparrow \bb R^{d}$; this is explained in
\Cref{sec:SBP-Cn-DR}.
\begin{remark}
\label{rem:Dowling-Exp}
The generating function analysis in these sections follows from
general results on \emph{exponential Dowling
structures}~\cite{EhrenborgReaddy}; similarly the analysis of
signed graphs is a special case of results on \emph{gain
graphs}~\cite{Zaslavsky}. We include the analyses for the benefit
of readers unfamiliar with these topics.
\end{remark}
\subsubsection{Mayer expansion for the type $D$ hard core gas}
\label{sec:SBP-Mayer}
Writing $\mathbbm{1}_{A} = 1 - \mathbbm{1}_{A^{c}}$ in~\eqref{eq:Cn-HCG-1} yields
\begin{equation}
\label{eq:Cn-HCG-2}
Z^{D}_{\Lambda}(z) = \sum_{n\geq 0} \frac{z^{n}}{n!}\int_{\Lambda^{n}}
\sum_{G\in \cc G^{\pm}\cb{n}} \prod_{(ij,\pm)\in E(G)} -\indicatorthat{\norm{x_{i}\mp
x_{j}}_{2}\leq 1}\, dx,
\end{equation}
where $G^{\pm}\cb{n}$ denotes the set of \emph{signed graphs} on
$\cb{n}$. These are graphs $G$ together with a \emph{signing}
$\sigma\colon E(G)\to \{\pm\}$. Note that an edge $(ij,+)$ corresponds
to the constraint $\norm{x_{i}-x_{j}}_{2}\leq 1$; this convention
makes signed graphs with all signs $+$ correspond to the ordinary
graphs that arise in the non-symmetrized hard-sphere gas.
Note that signed graphs may have multiple edges: it is possible for
$i$ and $j$ to be connected by an edge labelled $+$ \emph{and} an edge
labelled $-$. Loops are not permitted. A \emph{cycle} of a signed
graph will refer to a cycle of the underlying (multi)-graph. This
means that if $(ij,+)$ and $(ij,-)$ are edges in a signed graph, the
underlying graph contains two copies of the edge $ij$, and there is a
two-cycle that consists of these two edges.
\begin{definition}
\label{def:Coherent}
A cycle $\cc C$ in a signed graph $(G,\sigma)$ is called
\emph{balanced} if the product of the signs $\sigma(e)$ of the edges
in a cycle $\cc C$ is $+1$. A signed graph $(G,\sigma)$ is
\emph{balanced} if every cycle in the graph is balanced. A signed
graph that is not balanced is \emph{unbalanced}.
\end{definition}
A signed graph $(G,\sigma)$ can be partitioned into two vertex
disjoint signed subgraphs $(G_{b},\sigma_{b})$ and
$(G_{u},\sigma_{u})$, the former balanced and the latter
unbalanced. Let $\cc G_{b}^{\pm}\cb{n}$ and $\cc G_{u}^{\pm}\cb{n}$
denote the sets of balanced and unbalanced signed graphs on $\cb{n}$,
respectively. Then
\begin{equation}
\label{eq:Cn-HCG-3}
Z^{D}_{\Lambda}(z) = \sum_{n\geq 0} \mathop{\sum_{m,\ell}}_{m+\ell=n}
\frac{z^{n}}{m!\ell!}\int_{\Lambda^{n}}
\sum_{G\in \cc G^{\pm}_{b}\cb{m}} \sum_{H\in \cc G^{\pm}_{u}\cb{\ell}}
w(G)w(H)\, dx,
\end{equation}
where $w(G) = \prod_{(ij,\pm)\in E(G)} -\indicatorthat{\norm{x_{i}\pm
x_{j}}_{2}\leq 1}$. This is the convolution of two exponential
generating functions, and hence
\begin{equation}
\label{eq:Cn-HCG-4}
Z^{D}_{\Lambda}(z) = Z_{\Lambda}^{u}(z) Z_{\Lambda}^{b}(z),
\end{equation}
where the subscripts $u$ and $b$ indicated unbalanced and balanced,
respectively, e.g.,
\begin{equation}
\label{eq:Cn-HCG-5}
Z^{u}_{\Lambda}(z) = \sum_{n\geq 0} \frac{z^{n}}{n!} \sum_{G\in \cc
G^{\pm}_{u}\cb{n}} \int_{\Lambda^{n}} w(G)\, dx.
\end{equation}
\subsubsection{Calculation of $Z_{\Lambda}^{b}(z)$}
\label{sec:SBP-Bal}
\begin{lemma}
\label{lem:SBP-Bal}
The map from $\cc G^{\pm}_{b}\cb{n}$ to $\cc G\cb{n}$ given by forgetting the
signing of a balanced graph is a $2^{n-1}$-to-$1$ map when
restricted to connected graphs.
\end{lemma}
\begin{proof}
First we note that this map is well-defined: a balanced graph cannot
contain the edges $(ij,+)$, and $(ij,-)$, as this would be an
unbalanced cycle. Forgetting the signing of a balanced graph
therefore does yield an ordinary graph.
Let $G$ be a connected unsigned graph on $\cb{n}$. There is a
one-to-$2^{n}$ map given by arbitrarily labelling each vertex $i$ of
$G$ with $\sigma_{i}\in \{\pm\}$, and then assigning the edge $ij$
the sign $\sigma_{i}\sigma_{j}$. Each such assignment results in a
balanced graph: supposing that $i$ is $+$, trace any cycle
containing $i$. By construction seeing a $-$ edge means the sign of
the next vertex is different; since the cycle ends at $i$ there must
be at least one change of vertex sign after observing the first
$-$. This implies there are an even number of $-$ signs, so the
cycle is balanced.
The lemma follows by counting how many ways distinct signings of
vertices can give rise to the same signing of edges. This is $2$:
for any connected balanced signed graph once the sign of a single
vertex is fixed, the sign of every other vertex is
determined by $\sigma_{ij}=\sigma_{i}\sigma_{j}$. That this rule
signs each vertex consistently follows from the graph being balanced.
\end{proof}
It follows that from \Cref{lem:SBP-Bal} that
\begin{equation}
\label{eq:SBP-Bal-1}
Z^{b}_{\Lambda}(z) = \sum_{n\geq 0}\frac{(2z)^{n}}{n!}
\sum_{H\in \cc G\cb{n}}
\int_{\bb R^{dn}} (\frac{1}{2})^{\# H}w(H)\, dx,
\end{equation}
where $\# H$ is the number of connected components of the graph
$H$. Thus balanced graphs give rise to a cluster-weighted
variant of the hard sphere gas at activity $2z$. The exponential
principle combined with the argument that gives the Mayer expansion
implies
\begin{equation}
\label{eq:SBP-Bal-2}
\lim_{\Lambda\uparrow \bb R^{d}} \log Z^{b}_{\Lambda}(z) = \frac{1}{2}\sum_{n\geq
0}\frac{(2z)^{n}}{n!} \int_{\bb R^{dn}/\bb R^{d}} \sum_{H\in \cc G^{c}\cb{n}} w(H)\, dx.
\end{equation}
The right-hand side of \Cref{eq:SBP-Bal-2} is, up to a factor of
$\frac{1}{2}$, the pressure of the hard sphere gas at activity
$2z$. By \Cref{thm:BI} the pressure $\lim_{\Lambda\uparrow
\bb R^{d}}\abs{\Lambda}^{-1}\log Z^{b}_{\Lambda}(z)$ has a branched
polymer representation. It follows from what is presented in the next
section that this pressure is equal to $\lim_{\Lambda\uparrow
\bb R^{d}}\abs{\Lambda}^{-1} \log Z^{D}_{\Lambda}(z)$.
\subsubsection{Dimensional reduction for $Z_{\Lambda}^{u}(z)$}
\label{sec:SBP-Cn-DR}
The set of unbalanced graphs on $\cb{n}$ has a naturally associated
matroid $M_{D_{n}}$. The bases of the matroid are signed cycle rooted
spanning forests in which each cycle is unbalanced. The cycle may
consist of only two edges $(ij,+), (ij,-)$ for some $i\neq j$.
\begin{lemma}
\label{lem:SBP-Cn-Bicircular}
The matroid $M_{D_{n}}$ is the matroid associated to the hyperplane
arrangement $D_{n}$.
\end{lemma}
\begin{proof}
An edge $(ij,\pm)$ corresponds to the hyperplane defined by
$h^{\mp}_{ij}(x)=0$. Note that linear dependence of a set of normals
corresponding to a signed graph can only occur if some single component
is linearly dependent, so to establish linear dependence /
independence it suffices to consider each connected component separately.
In a base of the matroid each component has a single cycle. The only
possible non-trivial linear dependencies involve only edges in the
cycle: otherwise there is a vertex $i$ of degree $1$, and the
corresponding coordinate $x_{i}$ cannot have coefficient $0$ in any
linear combination. Attempting to determine a non-trivial linear
dependence thus has only one degree of freedom: once the coefficient
of a single edge is chosen, all other edges are determined. This
fact together with the fact that the cycle contains an odd number of
$-$ edges shows that any cycle is equivalent to a two-edge cycle
with edges $(ij,+)$, $(ij,-)$, and the corresponding normals are
linearly independent.
Similarly, given a spanning set that is not a base, there is a
component that contains two cycles. If there is a balanced cycle in
the component, then there is a linear dependence along this
cycle. If all cycles are unbalanced, then all cycles are in fact
edge disjoint. An argument as before shows that these can be
reduced to a pair of cycles $(ij,+),(ij,-)$, $(i'j',+),(i'j',-)$,
and a path connecting these cycles. The corresponding normals are
linearly dependent.
\end{proof}
\begin{theorem}
\label{thm:SBP-Cn-Result}
Let $Z_{D}^{BP,d}$ be the $\vec{\cc H}$-polymer partition function for
$\vec{\cc H} = (D_{1},D_{2},\dots)$ in $\bb R^{d}$. For all $z$ such
that the right-hand side converges and all $d\geq 1$,
\begin{equation}
\label{eq:SBP-Cn-Result}
\lim_{\Lambda\uparrow \bb R^{d}}
\frac{Z^{D}_{\Lambda}(z)}{Z_{\Lambda}^{b}(z)} =
Z^{BP,d+2}_{D}( - \frac{z}{2\pi}).
\end{equation}
\end{theorem}
\begin{proof}
By~\eqref{eq:Cn-HCG-4} the left-hand side is
$Z^{u}_{\Lambda}(z)$. Note that each graph that contributes is
unbalanced, which implies each connected component of the graph
contains an unbalanced cycle. Hence this is the generating function
of MMC associated to $M_{D_{n}}$, restricted to a finite volume
$\Lambda$. The theorem follows by applying
\Cref{thm:DR}. The infinite volume limit can be taken by the
monotone convergence theorem; that the terms are
monotonically increasing follows from \Cref{eq:DR-1b}.
\end{proof}
Thus unbalanced graphs express the lowest-order corrections to the
partition function of the type $D$ hard core gas compared to
$Z_{\Lambda}^{b}$, and these corrections can be written in terms of
$D_{n}$-polymers.
\subsubsection{Variations on the theme}
\label{sec:SBP-DR-Variations}
Similar arguments to those in
Sections~\ref{sec:SBP-Mayer}--~\ref{sec:SBP-Cn-DR} can be applied
to the arrangements with hyperplanes (in each case $i\neq j\in
\cb{n}$, $\ell\in \cb{n}$):
\begin{align*}
h_{ij}(x) &= x_{i} + x_{j}, \\
h_{ij}^{\pm}(x) &= x_{i}\pm x_{j}, \quad h_{\ell}(x)=0, \\
h^{m}_{ij}(x) &= x_{i} - \zeta^{m}x_{j},
\end{align*}
where in the last example $\zeta$ is a primitive $k$th root of unity
and $0\leq m\leq k-1$. The first example is the \emph{threshold
arrangement}, while the second is the \emph{type $B_{n}$ Coxeter
arrangement}. For a discussion of planar $B_{n}$-polymers
see~\cite[Section~6]{MP}. The class of examples in the third case are
the \emph{Dowling arrangements}; in general these are not complexified
arrangements and hence the results only apply to branched polymers in
$2d$-dimensional space. In all cases these arrangements enforce
certain symmetry constraints between the locations of spheres.
\bibliographystyle{ieeetr}
|
\section{Introduction}
Weak gravitational lensing has tremendous potential to inform our knowledge of the universe, on distance scales ranging from galactic to cosmological. In particular, {\it cosmic shear}, the statistical measurement of the distortion of observed galaxy shapes due to lensing by the distribution of matter along the line of sight, can provide information about the background cosmological model, via its sensitivity to the growth of structure and the strength of gravitational clustering on various scales. The first successful measurements of cosmic shear were carried out using patches of the sky with area of order one square degree~\citep{2000MNRAS.318..625B,2000astro.ph..3338K,2000Natur.405..143W,2000A&A...358...30V}. More recent measurements from the Stripe 82 region of the Sloan Digital Sky Survey~\citep{2012ApJ...761...15L,2014MNRAS.440.1322H}, the Canada France Hawaii Telescope Lensing Survey (e.g.~\citealt{2013MNRAS.432.2433H}), the Dark Energy Survey Science Verification run~\citep{2015arXiv150705552T}, and the Kilo-Degree Survey~\citep{2015MNRAS.454.3500K} have used $\mathcal{O}(100)$ square degrees and are beginning to provide useful low-redshift cosmological constraints that are complementary to the high-redshift information obtained from the cosmic microwave background. Indeed, there currently exist discrepancies between the two sets of constraints that have yet to be resolved (see e.g.~\citealt{2015MNRAS.451.2877M} for further discussion). Future surveys like the Wide Field InfraRed Survey Telescope (WFIRST)\footnote{\href{http://wfirst.gsfc.nasa.gov}{http://wfirst.gsfc.nasa.gov}}, the Large Synoptic Survey Telescope (LSST)\footnote{\href{http://www.lsst.org}{http://www.lsst.org}}, and Euclid\footnote{\href{http://sci.esa.int/euclid}{http://sci.esa.int/euclid}} will expand the sky coverage to $\mathcal{O}(10000)$ square degrees, and measure the shapes of orders of magnitude more galaxies than previous observational programs.
The observed shapes of galaxies are altered by gravitational lensing by only about 1\%, and it is the statistical correlations among these shapes that are the primary end product of cosmic shear observations. There are many possible sources of systematic errors that could contaminate the underlying cosmological signal, and each must be characterized in detail in order to obtain robust results. These include instrumental distortions of observed galaxy shapes caused by the telescope's point-spread function; biases in the methods used to measure shapes from raw images; uncertainties in the photometric redshifts obtained for each galaxy; ``intrinsic alignments" of galaxies with nearby structures, independently of the alignments caused by lensing by the line-of-sight distribution of matter; and uncertainties in the theoretical modeling of the distribution of matter on small scales.
This paper will focus on the last point, in particular on the effects of baryonic physics on the underlying two-point statistics of the distribution of matter. It is well-known~\citep{2004APh....22..211W,2004ApJ...616L..75Z,2006ApJ...640L.119J,2008ApJ...672...19R} that these effects can strongly influence cosmological constraints derived from cosmic shear measurements if not properly accounted for. For this reason, it is common to in cosmological analyses to simply exclude scales where these effects are important. Alternatively, a variety of schemes have been developed to allow inclusion of these scales by mitigating the effects of baryons on the desired information. These includes parametrizations of baryonic effects in the context of the halo model~\citep{2008PhRvD..77d3507Z,2011MNRAS.417.2020S,2013MNRAS.434..148S,2013PhRvD..87d3509Z,2014arXiv1410.6826M,2014MNRAS.445.3382M,2015MNRAS.454.1958M}, perturbation theory~\citep{2015JCAP...05..019L}, principal component analysis~\citep{2015MNRAS.454.2451E,2016MNRAS.459..971K}, or empirical fitting functions for the net effect~\citep{2015MNRAS.450.1212H} or individual baryonic components~\citep{2015JCAP...12..049S}, along with other techniques not based on direct parametrizations~\citep{2005PhRvD..72d3002H,2011MNRAS.416.1717K,2015JCAP...05..023B}.
However, one can also turn the question around, and treat these baryonic effects as {\it signal} to be constrained instead of nuisance to be removed or marginalized over. The advantage of using cosmic shear for these constraints is that it provides a relatively clean probe of the underlying matter distribution, free of the need to classify individual objects and explicitly relate them to dark matter halos or properties of their environments. Similarly, the effect of baryons on the total matter distribution can also be obtained relatively easily from a numerical simulation, by simply measuring the $N$-point matter statistics from snapshots of the simulation output. In recent years, such simulations have reached sufficient levels of detail that it is worthwhile to compare their predictions to observations of the real universe, and cosmic shear can act as nice complement to other observations that have already been exploited for this purpose (e.g.~\citealt{2010MNRAS.402.1536S,2015MNRAS.452.3529V}).
Rather than performing a separate analysis to constrain each implementation of baryonic physics one is interested in, it is possible to obtain generic, model-independent constraints via a principal component analysis of the deviations of the observed matter power spectrum from a fiducial ``dark matter-only" case. One can construct principal components of these deviations (in the form of linear combinations of deviations from the dark matter-only power spectrum at discrete points in the wavenumber-redshift plane) from a Fisher matrix corresponding to a specific survey: inverting and then diagonalizing this matrix not only gives the principal components corresponding to that survey, but also provides an automatic ranking of these principal components in terms of the expected constraints on their amplitudes. (For applications of this procedure to other problems in cosmology, see~\citealt{2003PhRvL..90c1301H,2009arXiv0901.0721A,2012PhRvD..86l3504S}.) One can then simply constrain these amplitudes in a likelihood analysis, and thereafter straightforwardly relate these constraints to any baryonic model of interest.
One may contrast this approach to that of~\citet{2015MNRAS.454.2451E}, which also applies principal component analysis to the issue of baryonic effects. The approach in that work is to find the principal components of the variations in the matter power spectrum between a range of specific baryonic models, and then remove these principal components from both the data and theory vectors in a likelihood analysis, in order to minimize the amount that cosmological constraints are contaminated by these effects. Our approach is rather to find the principal components of the matter power spectrum that will be best constrained by a cosmic shear measurements from a given survey, and then relate constraints on these principal components to constraints on a set of baryonic models as a second independent step.
In this paper, we carry out forecasts to assess the performance of this approach when it is applied to current and upcoming weak lensing surveys. Our forecasts marginalize over a variety of systematic effects relating both to calibration of the shear measurements and uncertainties in the theoretical modeling. For all surveys we consider, we find that 90\% of the constraining power contained in the shear correlation function is captured by at most nine principal components of the matter power spectrum. We use this fact to define a figure of merit: the reciprocal of the geometric mean on the expected one-sigma constraints of the amplitudes of the nine principal components. Comparing this figure of merit for different surveys, we find that the constraints on baryonic effects from Stage III surveys like the Dark Energy Survey or Hyper Suprime Cam Survey\footnote{\href{http://www.naoj.org/Projects/HSC/}{http://www.naoj.org/Projects/HSC/}} can improve upon those from currently available datasets by roughly an order of magnitude, while Stage IV surveys will only improve upon this by a factor of a few, with the constraining power being driven mainly by the number of galaxy shapes that can be measured by each survey.
We then apply this method to a representative set of models for baryonic effects on the matter power spectrum; specifically, we use nine models from the OverWhelmingly Large Simulations (OWLS;~\citealt{2010MNRAS.402.1536S,2011MNRAS.415.3649V}) suite. We find that the constraining power of future surveys is strongly dependent on the minimum angular scale at which measurements of the shear correlation functions can be used. In particular, measurements at scales of less than one arcminute would allow Stage III surveys to rule out the majority of the OWLS models at more than five-sigma confidence (or alternatively rule out the dark matter-only power spectrum if one of these models were correct). Furthermore, these measurements would also likely allow for the identification of a single ``best-fit" model among those that currently exist.
While we marginalize over a wide range of systematics, our conclusions rely on the existence of shear catalogs with well-understood properties (in terms of, for example, selection functions) at the appropriate scales. Provided that such an understanding can be attained (particularly for $\theta\lesssim 1'$), however, our results demonstrate that cosmic shear is capable of providing stringent tests of the results of current and future hydrodynamical simulations, opening an additional window on the physics of galaxy formation and evolution (or at least, its implementation in simulations).
The paper is organized as follows. In Sec.~\ref{sec:pca}, we describe the principal component method we use for parameterizing the matter power spectrum, while in Sec.~\ref{sec:mockdataandbaryons} we describe our assumptions about the surveys we provide forecasts for, the mock cosmic shear data and covariances we associate with these surveys, and the set of baryonic models we will use for our tests. In Sec.~\ref{sec:valopt}, we verify that most of the constraining power of each survey with respect to the matter power spectrum is contained in no more than the first nine principal components, and describe how we relate constraints on these principal components to a given baryonic model. In Sec.~\ref{sec:forecasts}, we compare the performance of various surveys via an appropriately-defined figure of merit, and also quantify this performance in terms of constraints on the baryonic models in our test set. We additionally examine the impact of varying our priors on the most important theoretical systematics. Finally, we provide additional comments and conclude in Sec.~\ref{sec:conclusions}.
\section{Method of Principal Components}
\label{sec:pca}
\begin{figure*}
\includegraphics[width=0.99\textwidth]{figures/des_y5_6modes_36x36.pdf}
\caption{\label{fig:example_modes}
The six principal components of $\Delta(k,z) \equiv P(k,z)/P_\text{DM-only}(k,z)-1$ that will be best constrained by measurements of the cosmic shear correlation functions $\xi_\pm$ from DES Y5, as determined by the procedure in Sec.~\ref{sec:pca}. (See Sec.~\ref{sec:mockdataandbaryons} for our assumptions about the details of these measurements.) The shapes of these principal components correspond to curves in the $(k,z)$ plane that contribute to multipoles at different $\ell$ values.
}
\end{figure*}
We begin by seeking a generalized parametrization of the matter power spectrum that allows for deviations from the DM-only case that are localized in wavenumber $k$ and redshift $z$. Such a parametrization will allow us to account for the fact that the strongest constraints on $P(k,z)$ from cosmic shear will be localized where the lensing kernels peak, and also for the fact that models for baryonic effects will have characteristic scale- and redshift-dependence that will vary from model to model.
One such parametrization simply uses the fractional deviation $\Delta$ between the full power spectrum and the DM-only one at discrete values of $k$ and $z$, taking the value of that deviation at each sample as a free parameter and smoothly interpolating between these samples at other values of $k$ and~$z$:
\begin{equation}
P(k_i,z_j) = P_\text{DM-only}(k_i,z_j) \left[ 1+\Delta_{ij} \right]\ .
\end{equation}
A version of this parametrization that assumes that $\Delta$ is $z$-independent was used in~\citet{2005APh....23..369H} and~\citet{2012JCAP...04..034H}. (For other works that do not rely on a $z$-independent parametrization, see~\citealt{2009ApJ...695..652B,2015JCAP...05..023B}.) For our purposes, it is important to relax this assumption, but this requires the introduction of at least several tens of new parameters $\{\Delta_{ij}\}$ into any analysis that would make use of this parametrization. Following the discussion above, some of these parameters will be much better constrained by a particular dataset than others. In addition, the nature of cosmic shear observables as projections of the matter power spectrum will lead to strong degeneracies among these parameters, since there are many ways to perturb the $\{\Delta_{ij}\}$ in a correlated way without significantly affecting the final signal.
Therefore, we further seek a convenient basis for these parameters that eliminates degeneracies as much as possible, and also identifies which basis elements are most relevant for a given set of observations. In fact, such a basis is straightforward to construct: given a covariance matrix for the $\{\Delta_{ij}\}$, its eigenvectors would specify a set of linear combinations of the $\{\Delta_{ij}\}$ that are maximally uncorrelated with one another, while the corresponding eigenvalues would indicate the expected variance on a measurement of each linear combination. Sorting these ``principal components" (PCs) by eigenvalue, we can then hope to identify a minimal subset of them that will suffice for a particular analysis (in our case, providing sufficient constraints on a range of baryonic models). Such a procedure has previously been applied, for example, to the reconstruction of a time-dependent equation of state for dark energy~\citep{2003PhRvL..90c1301H}, and of the primordial spectrum of curvature perturbations from inflation~\citep{2005PhRvD..72b3510K,2010PhRvD..82d3513D}.
For a given survey, the covariance matrix of the $\{\Delta_{ij}\}$ (along with other parameters related to the background cosmology and systematics) can be approximated by the inverse of the appropriate Fisher matrix. Recall that for a given vector of observed data points $\boldsymbol{d}$ with associated covariance~$\mathbf{Cov}$, along with a given model with parameters~$\{p_a\}$, the Fisher matrix is given by\footnote{The Fisher matrix will also contain terms that include the derivative of $\mathbf{Cov}$ with respect to the model parameters, but we will ignore those terms because they will generally be subdominant for a large-area survey (e.g.~\citealt{2009A&A...502..721E}).}
\begin{equation}
F_{ab} = \frac{{\partial}\boldsymbol{d}^{\rm T}}{{\partial} p_a} \mathbf{Cov}^{-1} \frac{{\partial}\boldsymbol{d}}{{\partial} p_b}
+ \frac{\delta^{ab}}{\sigma_a^2}\ ,
\end{equation}
where the last term incorporates optional Gaussian priors on each parameter~$p_a$. Upon inverting $F_{ab}$ and then diagonalizing the submatrix corresponding to the $\{\Delta_{ij}\}$, we obtain a new set of parameters $\alpha_a$, defined by
\begin{equation}
\alpha_a = \sum_\mu \beta_{a\mu} \Delta_\mu\ ,
\label{eq:alphaa_def}
\end{equation}
where the $(i,j)$ indices on $\Delta$ have been compressed into a single index $\mu$ for brevity. We use conventions where the $\{\boldsymbol{\beta}_{a}\}$ are orthonormal vectors, i.e.~$\sum_\mu \beta_{a\mu} \beta_{b\mu} = \delta_{ab}$. We will denote the eigenvalues (forecast variances of each $\alpha_a$) by $\sigma_a^2$.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/twosurveys_sigmai_over_sigma0.pdf}
\caption{\label{fig:example_sigmas}
Expected one-sigma uncertainties $\sigma_i$ on the amplitudes $\alpha_i$ of the most relevant PCs for DES Y5, normalized to the uncertainty on the best-constrained mode. The solid line corresponds to PCs constructed from a fine sampling of the $(k,z)$ plane (a $36\times36$ grid in $\log(k/[\,h\, {\rm Mpc}^{-1}\,])$ and $z$), while the dashed line corresponds to a coarser sampling (a $5\times5$ grid). The expected uncertainties increase rapidly with mode number, implying that a finite number of the lowest modes likely contain the bulk of a survey's information about the matter power spectrum. Furthermore, this increase is most rapid when the $(k,z)$ plane is sampled more coarsely, since each sample then acts as a wider ``bin" in the~$(k,z)$ plane and thus encapsulates more information. Forecasts for other surveys exhibit the same behavior.
}
\end{figure}
To visualize the results of this procedure, in Fig.~\ref{fig:example_modes} we show the 6 PCs whose amplitudes will be best constrained by (i.e.\ have the lowest variance when constrained using) measurements of the cosmic shear two-point function from DES Y5, where the assumed properties and treatment of systematics for DES Y5 will be described in Secs.~\ref{sec:mockdata}-\ref{sec:systematics}. These modes were constructed from a set of $\{\Delta_{ij}\}$ defined in a $36\times36$ grid in $\log(k/[\,h\, {\rm Mpc}^{-1}\,])$ and $z$, bounded by the edges of each panel in Fig.~\ref{fig:example_modes}. The lower bound for $k$, $0.1\,h\, {\rm Mpc}^{-1}\,$, has specifically been chosen to exclude linear scales, while the other boundaries have been chosen to encompass the range of wavenumbers and redshifts where DES is expected to be most informative (assuming measurements of $\xi_\pm$ to $\theta_{\rm min}=2'$). Similar considerations apply to our forecasts for other surveys.
The appearance of these modes in Fig.~\ref{fig:example_modes} reflects the sensitivity of the data to projections of $P(k,z)$ that contribute to different angular scales: specifically, a given multipole~$C_\ell$ will probe the matter power spectrum along the curve $k=\ell/\chi(z)$, where $\chi(z)$ is the comoving distance to redshift $z$ (see Eq.~\eqref{eq:cl_conv} in Sec.~\ref{sec:mockdata}). Measurements of $\xi_\pm$ in a given angular bin will be sensitive to a range of multipoles; it is also possible that sets of angular bins exist that are impacted equally by fractional variations in the power spectrum at a given level. Fig.~\ref{fig:example_modes} demonstrates that these angular bins (and their corresponding ranges of multipoles) are identified and combined together automatically by the PCA procedure.
In Fig.~\ref{fig:example_sigmas}, we show the projected constraints on the amplitudes of the most important PCs for DES Y5, normalized to the constraint on the PC with lowest expected variance. The solid line in this figure uses PCs obtained from the same $36\times36$ grid in $\log(k/[\,h\, {\rm Mpc}^{-1}\,])$ and $z$ for $\Delta_{ij}$ that was used to produce Fig.~\ref{fig:example_modes}, while the dashed line arises from a much coarser $5\times5$ grid in $\log(k/[\,h\, {\rm Mpc}^{-1}\,])$ and $z$. (The four best-constrained PCs corresponding to this coarser sampling are shown in Fig.~\ref{fig:example_modes_coarse}.)
\begin{figure}
\begin{centering}
\includegraphics[width=\columnwidth]{figures/des_y5_6modes_5x5.pdf}
\caption{\label{fig:example_modes_coarse}
As Fig.~\ref{fig:example_modes}, but displaying PCs constructed from a set of $\{\Delta_{ij}\}$ that sample the $(k,z)$ plane a factor of $\sim$5 more coarsely in both $k$ and $z$. With this coarser sampling, measurements at a greater range of angular scales are important for each PC, with the result that these four PCs contain a much larger fraction of the information to be found in DES than the first four PCs shown in Fig.~\ref{fig:example_modes}. This type of sampling of the $(k,z)$ plane is representative of what we will use in the remainder of this work.
}
\end{centering}
\end{figure}
The steep dropoff in constraining power (generically also seen in forecasts for other surveys, not shown in Fig.~\ref{fig:example_sigmas}) indicates that a finite number of PCs will likely contain the majority of the information about the matter power spectrum that we could expect to extract from a given set of shear measurements. The difference between the dashed and solid lines further implies that a coarser sampling of the $(k,z)$ plane is more efficient in compressing the constraining power into a small set of PCs. Henceforth, in this work we will use a $5\times5$ grid to produce our results, since this is the smallest number of points that will allow us to use cubic splines to interpolate along each axis. Coarser grids paired with different interpolation methods may also be useful, but we will not pursue them in this work.
It is worth mentioning that in the original basis of~$\{\Delta_{ij}\}$ variables, some of these variables may be so poorly constrained that their corresponding elements in the Fisher matrix will cause the matrix to become ill-conditioned, impeding its inversion or diagonalization. This can be ameliorated by incorporating a large Gaussian prior for each variable, which will not affect the results for the most important modes. For our forecasts in this work, a prior of $\sigma(\Delta_{ij})=10^4$ was sufficient to eliminate these numerical issues.
\begin{table*}
\centering
\caption{Assumed properties of surveys considered in this work}
\label{tab:survey_info}
\begin{tabular}{lccccc}
\hline
& Sky coverage & Source density, & Number of & Source distribution, \\
& (deg$^2$) & $\bar{n}$ (gals/arcmin$^2$) & redshift bins & $n(z)$ \\
\hline
DES SV & 139 & 5.7 & 3
& From~\citep{2015arXiv150705598B} \\
DES Y5 & 5000 & 8 & 5
& Eq.~\eqref{eq:nz}: $z_{\rm max}=2.0$, $\alpha=1.75$, $z_0=0.25$, $\beta=1.0$ \\
HSC & 1400 & 20 & 10
& Eq.~\eqref{eq:nz}: $z_{\rm max}=3.0$, $\alpha=2.0$, $z_0=0.35$, $\beta=1.0$ \\
LSST & 18000 & 26 & 10
& Eq.~\eqref{eq:nz}: $z_{\rm max}=3.5$, $\alpha=1.25$, $z_0=0.50$, $\beta=1.0$ \\
Euclid & 15000 & 30 & 10
& Eq.~\eqref{eq:nz}: $z_{\rm max}=2.5$, $\alpha=1.25$, $z_0=0.35$, $\beta=1.0$ \\
WFIRST & 2200 & 45 & 10
& Eq.~\eqref{eq:nz}: $z_{\rm max}=4.0$, $\alpha=1.25$, $z_0=0.60$, $\beta=1.0$ \\
\hline
\end{tabular}
\end{table*}
\section{Mock Data and Baryonic Models}
\label{sec:mockdataandbaryons}
We will demonstrate the performance of the method described in Sec.~\ref{sec:pca} by carrying out forecasts in which it is applied to a number of mock data sets corresponding to current and upcoming weak lensing surveys, and comparing to what would be obtained from attempting to directly constrain a set of representative models for baryonic effects on the matter power spectrum. In this section, we describe how these mock data sets are constructed, our treatment of systematic effects that could be important for cosmic shear measurements on the relevant scales, and the baryonic models we use for our test set.
\subsection{Mock data}
\label{sec:mockdata}
For our forecasts, we consider a variety of upcoming surveys (along with the already-completed Dark Energy Survey Science Verification run, DES SV, which will serve as an indicator of the constraining power of currently available data). Table~\ref{tab:survey_info} displays the properties we will assume in this work. These properties are based mainly on~\citet{2016MNRAS.456..207K}, with following exceptions: for DES Y5, we have use a slightly more conservative value for the density of sources; for HSC, we have used the sky coverage and a more conservative value of the source density from~\citet{2015ApJ...806..186O}; and for Euclid, we have used the density of sources specified in~\citet{2011arXiv1110.3193L}. For all surveys, we will assume an intrinsic shape noise of $\sigma_\epsilon=0.27$ per component. Where appropriate, throughout the paper we will comment on the sensitivity of our forecasts to these assumptions.
For the data themselves, we will use the real-space two-point shear correlation functions $\xi_\pm(\theta)$, given by
\begin{equation}
\xi_{\pm}^{ij}(\theta) = \frac{1}{2\pi} \int d\ell \, \ell \, J_{0/4}(\ell\theta) \, C_\ell^{ij}\ ,
\end{equation}
where $J_{0/4}$ is the Bessel function of order 0 or 4, and $i,j$ denote the two tomographic redshift bins involved in the correlation. The multipoles of the convergence field, $C_\ell^{ij}$, are obtained from the matter power spectrum via
\begin{equation} \label{eq:cl_conv}
C_\ell^{ij} = \frac{9H_0^4 \Omega_{\rm m}^2}{4c^4} \int_0^{\chi_{\rm h}} d\chi
\frac{g^i(\chi) g^j(\chi)}{a^2(\chi)} \,P\!\left( \frac{\ell}{f_K(\chi)}, z(\chi) \right)\ ,
\end{equation}
where $\chi_{\rm h}$ is the comoving distance to the horizon and $f_K(\chi)$ is the comoving angular diameter distance (equal to $\chi$ in a flat universe, which we will assume in this work). The lens efficiency $g_i(\chi)$ is given by
\begin{equation} \label{eq:lens_eff}
g^i(\chi) = \int_\chi^{\chi_{\rm h}} d\chi' n_i(\chi') \frac{f_K(\chi'-\chi)}{f_K(\chi')}\ .
\end{equation}
Our base forecasts will be performed using 6 measurements of each of $\xi_+$ and $\xi_-$ per redshift bin pair, log-spaced between $\theta_{\rm min}=2'$ and $\theta_{\rm max}=300'$, although we will also investigate the consequences of other choices of $\theta_{\rm min}$.
For the distribution of sources $n_i(z)$ in each redshift bin, we will use the exact distributions from DES SV~\citep{2015arXiv150705598B} for the corresponding forecast. For DES Y5 and the other surveys, we employ the following parametrization with parameter values given in Table~\ref{tab:survey_info}, dividing the total $n(z)$ into bins with equal numbers of sources:
\begin{equation} \label{eq:nz}
n(z) \propto \Theta(z_{\rm max}-z) \, z^\alpha \exp\left[ -\left( \frac{z}{z_0} \right)^\beta \right]\ .
\end{equation}
We base the source distribution for HSC on~\citet{2011PhRvD..83b3008O}, while we follow~\citet{2016MNRAS.456..207K} for the other surveys. We make a conservative choice of 5 redshift bins for DES Y5, while we will assume that the other surveys will be capable of collecting sufficient statistics in 10 redshift bins.
For the covariance of the mock correlation function measurements, we will restrict ourselves to the Gaussian approximation, evaluated as in~\citet{2008A&A...477...43J}. We refer the reader to that paper for the derivation, and merely quote the final expression we use, which accounts for the difference in our conventions for $\sigma_\epsilon$:
\begin{widetext}
\begin{align} \nonumber
\text{Cov}_\text{G} \! \left[ \xi_\pm^{ij}(\theta_1), \xi_\pm^{kl}(\theta_2) \right]
&= \frac{1}{8\pi^2 f_{\rm sky}} \int d\ell \, \ell\, J_{0/4}(\ell\theta_1) J_{0/4}(\ell\theta_2)
\left[ \tilde{C}_{\ell_{1}}^{ik} \tilde{C}_{\ell_{1}}^{jl} + \tilde{C}_{\ell_{1}}^{il} \tilde{C}_{\ell_{1}}^{jk}
- (\delta^{ik} \delta^{jl} + \delta^{il}\delta^{jk}) \frac{\sigma_\epsilon^4}{n^2} \right] \\
&\quad + (\delta^{ik} \delta^{jl} + \delta^{il}\delta^{jk}) \delta_{\theta_1 \theta_2}
\delta_{+-}
\frac{\sigma_\epsilon^4}{8\pi^2 f_{\rm sky} \theta_1 \Delta\theta_1 n^2 }\ ,
\label{eq:xipm_cov}
\end{align}
\end{widetext}
\noindent where $\delta_{+-}=1$ for $\text{Cov}_\text{G} \! \left[ \xi_+, \xi_+ \right]$ or $\text{Cov}_\text{G} \! \left[ \xi_-, \xi_- \right]$ and 0 otherwise, and $n$ is the angular density of source galaxies in each redshift bin (in our forecasts, $n$ is the same for each bin in a given survey). In Eq.~\eqref{eq:xipm_cov}, we have defined
\begin{equation}
\label{eq:clplussn}
\tilde{C}_{\ell}^{ij} = C_{\ell}^{ij} + \delta^{ij} \frac{\sigma_\epsilon^2}{n}\ .
\end{equation}
For a subset of our forecasts, we have also implemented the leading non-Gaussian terms in the covariance (the trispectrum of the convergence, along with a halo sample variance term that accounts for the influence of modes beyond the survey window), using the halo model treatment described in~\citet{2015MNRAS.454.2451E}, to which we refer the reader for details. We find that the inclusion of these terms has negligible impact on our results, justifying our use of the Gaussian approximation for the remaining forecasts.
For our fiducial cosmology, we use the following best-fit parameters from~\citet{2015arXiv150201589P}: $\Omega_{\rm m}=0.316$, $\Omega_{\rm b}=0.0491$, $h_0=0.673$, $\sigma_8=0.83$, $n_{\rm s}=0.965$, $w=-1.0$, and $\tau=0.078$. In our forecasts, we allow these parameters to vary within the one-sigma priors determined from Planck. Previous work~\citep{2014PhRvD..90f3516N,2015MNRAS.450.1212H} has shown that it is important to consider the effect of massive neutrinos jointly with other effects on the small-scale matter power spectrum, and therefore we also marginalize over the sum of neutrino masses in our forecasts, assuming a fiducial value of $\Sigma m_\nu = 0.06$eV (for simplicity, we only assume a single massive neutrino species).
For each survey, we compute a mock data vector and mock covariance matrix with the properties and fiducial cosmology described above. We then use these to compute a Fisher matrix, where the set of varied parameters consists of the cosmological parameters in the previous paragraph, neutrino mass, the grid of 25 $P(k,z)$ values described in Sec.~\ref{sec:pca}, and the systematics parameters listed in the next section. The maximum redshift for the power spectrum grid is set roughly to the maximum redshift for each survey, while the maximum wavenumber is set based on the value $\theta_{\rm min}$.\footnote{Specifically, we use $k_{\rm max}=\{ 50, 100, 400 \} \,h\, {\rm Mpc}^{-1}\,$ for $\theta_{\rm min} = \{4',2',0.5'\}$. These $k_{\rm max}$ values we chosen to obtain convergence of the final forecast results with respect to small changes in $k_{\rm max}$. Therefore, the results are not strongly sensitive to the values of the power spectrum precisely at $k_{\rm max}$, but are only sensitive up to a wavenumber around a factor of a few smaller.} We then invert the Fisher matrix and diagonalize the submatrix corresponding to the $P(k,z)$ samples; the eigenvectors (PCs) and eigenvalues $\{\sigma_a^2\}$ are then used to obtain the results in the remainder of the paper.
We use CAMB~\citep{camb} to calculate the linear matter spectrum, and Halofit~\citep{halofit,halofit-update} as an approximation for the nonlinear corrections from gravitational evolution. Furthermore, we use the modified Halofit prescription from~\citet{2012MNRAS.420.2551B} to account for the effect of massive neutrinos. (In Sec.~\ref{sec:systematics}, we describe how we account for possible uncertainties in the modeling of the nonlinear power spectrum.) Most numerical computations are performed using CosmoSIS~\citep{2015A&C....12...45Z}, with the exception of the mock covariances, for which we use CosmoLike~\citep{2016arXiv160105779K}.
\subsection{Treatment of systematic errors}
\label{sec:systematics}
We have attempted to include a selection of possible systematic errors that will likely be relevant for cosmic shear measurements on the scales we are interested in.
Our implementation is as follows:
\begin{enumerate}
\item {\bf Shear calibration errors}:
Following the treatment of the DES Science Verification Data in~\citet{2015arXiv150705552T}, we assign a single free multiplicative parameter $m_i$ to each redshift bin, so that $C_\ell^{ij} \to (1+m_i)(1+m_j) C_\ell^{ij}$, and marginalize over each of these parameters, using a Gaussian prior of width 0.05 centered at zero. This choice is extensively discussed and motivated in~\citet{2015arXiv150705603J} in the context of DES SV, but the general features of that discussion are likely to apply to shear calibration issues in other surveys as well.
\item {\bf Photometric redshift uncertainties}:
For each redshift bin, we allow for a uniform translation $\delta z_i$ of the entire source distribution within that bin, such that $n_i(z) \to n_i(z-\delta z_i)$. We marginalize over each $\delta z_i$ with a prior of 0.05, which was found to reflect the uncertainty in the redshift distributions from DES SV~\citep{2015arXiv150705909B}; while this value is much larger than the uncertainty goal for photo-$z$'s from future (e.g.\ Stage III) observations, to be conservative, we use it for all surveys we consider.
\item {\bf Intrinsic alignments}:
We assume the non-linear tidal alignment model of~\citet{2007NJPh....9..444B}, with free overall amplitude $A_{\rm IA}$, assigned a wide prior of $\sigma(A_{\rm IA})=5$ and marginalized over in our analysis. We discuss this choice further in Sec.~\ref{sec:sys-priors}.
\item {\bf Uncertainty in power spectrum modeling}:
We treat this uncertainty by allowing for two constant fractional shifts in the value of the ``DM+neutrinos" power spectrum, above and below some transition wavenumber $k_{\rm tr}$, with exponentials smoothly connecting the two regimes:
\begin{align} \nonumber
&P(k,z) \to P(k,z) \left[ 1 + S_P^{\rm (low)} \exp({-k^2/k_{\rm tr}^2}) \right. \\
&\qquad\qquad\qquad\qquad\quad \left. +\,\, S_P^{\rm (high)} \left( 1-\exp[{-k^2/k_{\rm tr}^2}] \right) \right]\ .
\end{align}
We fix $k_{\rm tr}=0.5\,h\, {\rm Mpc}^{-1}\,$, and marginalize over the two amplitudes $S_P^{\rm (low)}$ and $S_P^{\rm (high)}$, with Gaussian priors of 0.02 and 0.05 respectively. These priors are roughly based on conservative estimates of the accuracy of currently-available fitting functions like Halofit or the CosmicEmu emulator~\citep{2014ApJ...780..111H}. In the future, the combination of improvements in perturbation theory~\citep[e.g.][]{2015arXiv150705326F} at large scales and more ambitious emulation programs or approaches to nonlinear modeling at smaller scales will likely decrease these errorbars dramatically. We explore the effect of varying the prior on $S_P^{\rm (high)}$ in Sec.~\ref{sec:sys-priors} (our results are insensitive to the prior on $S_P^{\rm (low)}$).
\item {\bf Other small-scale shear systematics}:
There exist several other systematics effects that could become relevant at high multipoles, or equivalently small ($\sim$arcminute) angular scales. Examples of these include higher-order terms in the expansion of the ``reduced shear" $\gamma/(1-\kappa)$, which is what is actually observed (rather than the pure shear $\gamma$ itself)~\citep{2009ApJ...696..775S,2010A&A...523A..28K}; {\it lensing bias}, the effect of lensing on observed properties of galaxies, which can lead to biased selection in regions of large magnification~\citep{2009ApJ...702..593S,2010A&A...523A..28K}; and environment-dependent selection biases, arising for example from excluding blended sources (which typically occur in regions of high convergence) from a shear catalogue~\citep{2011A&A...528A..51H,maccrann-thesis,maccrann-inprep}.
Fortunately, a simple power-law model, $\Delta C_\ell/C_\ell \propto \ell^p$ serves to mimic the impact of most of these effects quite closely, with $p \approx 0.5$ providing a good match to examples of explicit calculations found in the literature~\citep{2009ApJ...696..775S,2009ApJ...702..593S}. Therefore, we roll these effects into the following phenomenological model, letting $S_{\rm shear}$ be the overall amplitude of these effects at $\ell=10^4$:
\begin{equation}
C_\ell^{ij} \to C_\ell^{ij} \left[ 1 + S_{\rm shear} \left( \frac{\ell}{10^4}\right)^{0.5} \right]\ .
\end{equation}
Such a model was also used in~\citet{2006MNRAS.366..101H}, who pointed out that other additive systematics such as anisotropies in a telescope's point-spread function will likely have a similar impact on the angular power spectrum.
We marginalize over the amplitude $S_{\rm shear}$ with a prior of 0.05, which is roughly the level at which these effects are expected to contaminate near-term shear measurements~\citep{maccrann-thesis,maccrann-inprep}. We investigate how our results depend on the prior on $S_{\rm shear}$ in Sec.~\ref{sec:sys-priors}.
\end{enumerate}
In summary, we marginalize over $(2n_{\rm bins}+4)$ systematics parameters, along with the sum of neutrino masses.
\subsection{Models for baryonic effects}
\label{sec:baryon_models}
\begin{figure}
\includegraphics[width=\columnwidth]{figures/baryons_pkratio.pdf}
\caption{\label{fig:baryons_pkratio}
Ratio of total (baryons+DM) power spectra over DM-only power spectrum at $z=0$ for each baryonic model described in Sec.~\ref{sec:baryon_models}.
}
\end{figure}
To represent a range for the types of baryonic effects on the matter power spectrum that can be expected within reasonable scenarios, we consider the 9 unique models that compose the OWLS suite of hydrodynamic simulations~\citep{2010MNRAS.402.1536S,2011MNRAS.415.3649V}. A brief summary of the physical effects that are included in each model can be found in Table~1 of~\citet{2011MNRAS.415.3649V}.
These simulations were carried out using a modified version of Gadget-3, in periodic boxes with side length $100h^{-1} {\rm Mpc}$ using $512^3$ dark matter particles and an equal number of baryonic particles. Each simulation (including one that contains only dark matter) was started with the same initial conditions, so that the ratio $P_\text{baryons+DM}/P_\text{DM-only}$ can be obtained without the sample variance inherent in the initial perturbations\footnote{Of course, these ratios will likely have some dependence on the background cosmology, but since we are only using these models as representative of the possible range of baryonic effects on the matter power spectrum, we will ignore this cosmology-dependence in this work. Furthermore, the ratio $P_\text{baryons+DM}/P_\text{DM-only}$ will itself have nonzero sample variance, by virtue of the fact that it has been measured from a finite spatial volume. We will not attempt to quantify this sample variance in this work, but merely note that future precision comparisons with hydrodynamical simulations will eventually need to take this into account.}.
These ratios are shown in Fig.~\ref{fig:baryons_pkratio} at $z=0$ for each of the 9 models. These models can be phenomenologically classified into two groups:
\begin{enumerate}
\item {\it AGN, DBLIMFV1618, WDENS, WML1V848}: These include strong feedback processes that drive gas out of galaxies, smoothing the total density field and thereby suppressing the power spectrum for $k\lesssim 10\,h\, {\rm Mpc}^{-1}\,$.
\item {\it REF, WML4, NOSN\_NOZCOOL, NOZCOOL, NOSN}: These contain relatively weak feedback (or none at all), with the dominant effect on the power spectrum instead being an enhancement at $k\gtrsim 5\,h\, {\rm Mpc}^{-1}\,$ arising from gas cooling that increases the central density in halos.
\end{enumerate}
In Fig.~\ref{fig:baryons_contour}, we show contour plots of the ratio $P_\text{baryons+DM}/P_\text{DM-only}$ for the two models at the extremes of this classification: {\it AGN}, which includes feedback from active galactic nuclei in addition to supernovae, and {\it NOSN}, which does not include any feedback mechanisms but includes radiative cooling.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/owls_contour_AGN.pdf}
\includegraphics[width=\columnwidth]{figures/owls_contour_NOSN.pdf}
\caption{\label{fig:baryons_contour}
As Fig.~\ref{fig:baryons_pkratio}, but plotted as a function of wavenumber and redshift, and only showing the two most extreme models: one ({\it AGN}) in which the combined effects of AGN and supernova feedback lead to a strong suppression of the power spectrum at relatively large scales, and another ({\it NOSN}) that neglects all feedback processes but contains gas cooling that enhances the power spectrum at small scales.
}
\end{figure}
Of the OWLS models, {\it AGN} can be thought of as the most realistic, since the AGN feedback effects that it includes result in much better agreement with observations of low-redshift groups of galaxies than any of the other models (see~\citealt{2010MNRAS.402.1536S,2011MNRAS.415.3649V,2015MNRAS.452.3529V} for details). Some others, such as those that neglect radiative cooling processes known to be important in matching galaxy-scale observations, are decidedly unrealistic. Nevertheless we still find it useful test to test our principal-component method on the full range of scenarios.
\section{Validation of Method}
\label{sec:valopt}
The strategy of identifying the best-constrained principal components of $P(k,z)$, as described in Sec.~\ref{sec:pca}, will only be useful if these principal components can also provide a reasonable amount of information about the influence of baryonic effects on the power spectrum. In this section, we quantify the information content of the best-constrained PCs with respect to our chosen set of baryonic models, and demonstrate that a small number of these PCs provides sufficient information about the full set of models to lead to meaningful constraints.
We first define the following parametrization for each baryonic model separately, based around a free amplitude $\aowls$ that interpolates between the DM-only case ($\aowls=0$) and the exact DM+baryons power spectrum ($\aowls=1$):
\begin{align} \nonumber
P(k,z) &= P_\text{DM-only}(k,z) \\
&\quad\times \left[ 1 + \aowls \left( \frac{P_\text{DM+baryons}(k,z)}{P_\text{DM-only}(k,z)} - 1 \right) \right]\ .
\label{eq:aowls_def}
\end{align}
If we were only interested in a single baryonic model, we could imagine using Eq.~\eqref{eq:aowls_def} as our model for baryonic effects, and attempting to constrain $\aowls$ along with the other cosmological and systematics parameters. Alternatively, we can parametrize baryonic effects using the amplitudes~$\{\alpha_a\}$ of the $N$ most important PCs. To do so, we can solve Eq.~\eqref{eq:alphaa_def} for $\Delta_\mu$ at each $\mu\equiv(k_i,z_j)$, using only up to mode $N$ on the right hand side, and then smoothly interpolate between these discrete values of $\Delta_\mu$ to construct a function $\Delta_N(k,z)$, which will depend on the values $\{\alpha_1,\cdots,\alpha_N\}$. The power spectrum is then given by
\begin{equation}
P(k,z) = P_\text{DM-only}(k,z) \left[ 1 + \Delta_N(k,z) \right]\ .
\end{equation}
An intuitive way to determine how much information the PCs provide about different baryonic models is to relate constraints on the amplitudes $\{\alpha_a\}$ to a constraint on $\aowls$. If we treat each amplitude as providing an independent ``measurement" of $\aowls$ and apply inverse-variance weighting to these ``measurements," we obtain\footnote{The right-hand side of Eq.~\eqref{eq:sig2aowls} is also the square of the cumulative signal-to-noise of the first $N$ PCs, and therefore can alternatively be interpreted without the need of the $\aowls$ parametrization from Eq.~\eqref{eq:aowls_def}.}
\begin{equation}
\frac{1}{\sigma^2(\aowls)} = \sum_{a=1}^N \frac{1}{\sigma_a^2} \left( \frac{{\partial}\alpha_a}{{\partial}\aowls} \right)^2
= \sum_{a=1}^N \frac{ \left( \alpha_a^\text{(bar)} \right)^2 }{\sigma_a^2} \ ,
\label{eq:sig2aowls}
\end{equation}
where $\alpha_a^\text{(bar)}$ is the projection of the specific baryonic model onto the chosen PCs:
\begin{equation}
\alpha_a^\text{(bar)} = \sum_\mu \beta_{a\mu} \Delta_\mu^\text{(bar)}\ .
\end{equation}
(The second equality in Eq.~\eqref{eq:sig2aowls} follows from the fact that $\Delta_\mu^\text{(bar)}$, and therefore $\alpha_a^\text{(bar)}$, is linear in $\aowls$.)
\begin{figure}
\includegraphics[width=\columnwidth]{figures/sigaowls_ratio.pdf}
\caption{\label{fig:sigma_aowls_nmodes_surveys}
Ratio of the effective one-sigma constraint on $\aowls$ (defined in Eq.~\eqref{eq:aowls_def}), $\sigma(\aowls)$ from using all available PCs with that from using only the $N$ best-constrained PCs for the three surveys specified above each panel. For each set of modes, the effective constraint on $\aowls$ is calculated using Eq.~\eqref{eq:sig2aowls}. The black curve corresponds to the OWLS {\it AGN} model, while the blue band represents the range spanned by all OWLS models. This represent the fraction of the full survey's information that is retained by the first $N$ PCs of $P(k,z)$. Other Stage IV surveys give results comparable to LSST, and therefore we conclude that for all surveys we consider, the first $\sim$9 PCs contain 90\% of the total constraining power on our test set of baryonic models.
}
\end{figure}
When {\it all} available PCs are used in Eq.~\eqref{eq:sig2aowls} to determine $\sigma(\aowls)$, the result is very close to the output of forecasts in which we constrain $\aowls$ directly for each baryonic model in our testing set. However, it is sufficient to use a much smaller number of PCs to retain a comparable amount of constraining power. In Fig.~\ref{fig:sigma_aowls_nmodes_surveys}, we check how this constraining power scales with the number $N$ of the most important PCs we retain, by plotting the ratio of $\sigma(\aowls)$ obtained from all PCs and that from the first $N$ PCs. If we wish to retain 90\% of the constraining power of the full set of PCs, we may only keep the first 9 PCs in our analysis. This number is an upper bound across all surveys we have investigated, including those not shown in Fig.~\ref{fig:sigma_aowls_nmodes_surveys}, and can be considered as a validation of the usefulness of the PCA approach.
In the next section, we will perform forecasts for the constraining power of the first 9 PCs of various surveys. We use the same number of PCs for each survey in order to make meaning comparisons, but in principle, a given survey might require even fewer PCs to achieve meaningful constraints. For example, the top panel of Fig.~\ref{fig:sigma_aowls_nmodes_surveys} shows that for cosmic shear measurements comparable to those of DES SV, only the first 6 modes are required to capture 90\% of the full amount of information about the impact of baryons of the matter power spectrum. Optimization of the number of PCs in other cases is best left to future, survey-specific analyses.
\section{Forecasts}
\label{sec:forecasts}
\subsection{Performance of various surveys}
\label{sec:forecasts-fom}
In order to quantify the constraining power of different surveys with respect to generic baryonic effects on the matter power spectrum, we define a figure of merit $\mathcal{F}$ as the reciprocal of the geometric mean of the 68\% confidence intervals for the most important 9 PCs for each survey:
\begin{equation}
\label{eq:fom}
\mathcal{F} \equiv \left[ \prod_{a=1}^{9} \sigma_a \right]^{-1/9}\ .
\end{equation}
Since the PCs are uncorrelated by construction, the product $\prod_{a=1}^{9} \sigma_a$ is proportional to the volume of the hyperellipsoid corresponding to the joint 68\% confidence region of all PCs, but this product differs by many orders of magnitude for different surveys. On the other hand, the quantity in Eq.~\eqref{eq:fom} efficiently captures a kind of average of the constraints on the first 9 modes; from Eq.~\eqref{eq:sig2aowls}, we then find that the effective constraints on a generic model for baryonic effects (as parametrized by $\sigma(\aowls)$) will scale proportionally to~$\mathcal{F}$.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/fom.pdf}
\caption{\label{fig:fom}
Figure of merit $\mathcal{F}$ (defined in Eq.~\eqref{eq:fom}) for constraints on baryonic models, normalized to the value for DES SV, for forecasts performed with mock $\xi_\pm$ measurements down to $\theta_{\rm min}=2'$. We find that both HSC and the Y5 shear measurements of cosmic shear from DES will improve $\mathcal{F}$ by roughly an order of magnitude, implying a similar potential improvement in our information about baryonic effects on the matter power spectrum, while Stage IV surveys like LSST will improve on that by a further factor of a few.
}
\end{figure}
Fig.~\ref{fig:fom} presents the figure of merit for each survey we consider in this work, for forecasts using either $\xi_\pm$ measurements down to $\theta_{\rm min}=2'$, normalized to the figure of merit for DES SV to emphasize the improvement of future shear measurements over currently available data. Fig.~\ref{fig:fom} indicates that HSC and the Y5 shear measurements from DES will improve on current data by roughly an order of magnitude, while upcoming Stage IV surveys will only improve on that by a factor of a few. This can be understood intuitively by making use of the following rough scalings:
\begin{itemize}
\item $\mathcal{F} \propto f_{\rm sky}^{0.5}$, since the leading terms in the data covariance matrix scale like $f_{\rm sky}^{-1}$.
\item Because of shot noise in the shear correlation function, $\mathcal{F} \propto (n\sigma_\epsilon^{-2})^{0.6}$ (this power has been empirically determined by varying $n$ in forecasts for various surveys). If shot noise completely dominated the covariance matrix at all scales, we would expect $\mathcal{F} \propto (n\sigma_\epsilon^{-2})^1$, but the balance between the cosmic variance and shot noise terms in Eq.~\eqref{eq:clplussn} softens the dependence to what we quote here.
\item On average, the figure of merit scales with the number of redshift bins like $\mathcal{F} \propto n_{\rm bins}^{0.4}$. This power is steeper for surveys with broader redshift distributions, since in those cases, narrower redshift bins contain more information than for surveys that primarily probe lower redshifts.
\end{itemize}
Therefore, it is essentially the improvement in the expected number of galaxies with measured shapes (i.e.\ $f_{\rm sky} \times n$) that drives the large improvement of Stage III surveys over current data, as compared to the somewhat more modest improvement afforded by Stage IV surveys.
\begin{figure*}
\includegraphics[width=\textwidth]{figures/abs_constraints.pdf}
\caption{\label{fig:sigma_aowls_absolute}
{\it Top panels:} Expected one-sigma constraints on $\aowls$ (see Eq.~\eqref{eq:aowls_def}) for each survey and baryonic model we consider, when the first 9 PCs for each survey are used in the analysis. {\it Bottom panels:} Statistical significance at which each survey could rule out each model, again using the PCA approach with the amplitudes of the first 9 PCs included as free parameters in the analysis. Using measurements of $\xi_\pm^{ij}$ over the range $4'<\theta<300'$ ({\it left panels}), the constraints are strongest on models including prescriptions for feedback or other effects that affect the power spectrum at relatively large scales, e.g.~$k\lesssim1\,h\, {\rm Mpc}^{-1}\,$ at $z=0$ ({\it blue labels}). When $\theta_{\rm min}=2'$ is used instead ({\it right panels}), models with large deviations from the DM-only power spectrum on small scales (typically due to cooling effects on the centers of halos; {\it black labels}) are much better constrained due to the inclusion of the additional small-scale information.
}
\end{figure*}
\subsection{Expected constraints on specific baryonic models}
Beyond the model-independent approach of Sec.~\ref{sec:forecasts-fom}, we can also ask how our method can provide information about specific baryonic models when applied to different surveys. Specifically, we can use Eq.~\eqref{eq:sig2aowls} to translate the expected constraints on the 9 most important PCs into constraints on $\aowls$ for each of the models in Sec.~\ref{sec:baryon_models}. We can further translate these constraints into the statistical significance at which each model could be ruled out by a given survey (or, alternatively, the significance at which the DM-only power spectrum can be ruled out).
\begin{figure}
\includegraphics[width=\columnwidth]{figures/abs_constraints_xi_thmin0p5.pdf}
\caption{\label{fig:sigma_aowls_absolute_xi_thmin0p5}
Same as the bottom panels of Fig.~\ref{fig:sigma_aowls_absolute}, but assuming that measurements of $\xi_\pm$ can be used down to $\theta_{\rm min}=0.5'$. The gain in constraining power is significant (and also robust to the variations in systematics discussed in Sec.~\ref{sec:sys-priors}), with all other surveys we have considered (not shown) providing better than 20$\sigma$ constraints across the entire test set of baryonic models. Therefore, the coming generation of cosmic shear measurements could potentially be very informative with regards to the implementation of certain baryonic phenomena in simulations.
}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{figures/abs_constraints_xi_vs_cl.pdf}
\caption{\label{fig:xi_vs_cl}
Same as the bottom panels of Fig.~\ref{fig:sigma_aowls_absolute}, but comparing LSST forecasts using the real-space correlation function $\xi_\pm$ ({\it black solid lines}) or the angular power spectrum $C_\ell$ ({\it red dashed lines}) with different values of $\theta_{\rm min}$ and $\ell_{\rm max}$ as indicated in the figure. For each value of $\theta_{\rm min}$, we find that $\ell_{\rm max}$ can be chosen to provide very similar constraints on our test set of baryonic models, following the relationship $\theta_{\rm min} \simeq 1.3\pi/\ell_{\rm max}$.
}
\end{figure}
The top left panel of Fig.~\ref{fig:sigma_aowls_absolute} shows the resulting values for $\sigma(\aowls)$ for each survey and baryonic model, in the case where the shear correlation functions $\xi_\pm^{ij}$ have been measured down to $\theta_{\rm min}=4'$, while the bottom left panel shows the corresponding statistical significance of these constraints, given in terms of the ``number of sigmas" at which the model could be distinguished from the DM-only case. Unsurprisingly, the models that can be best constrained by the most powerful surveys are {\it AGN}, whose implementation of feedback has dramatic effects at large scales\footnote{Other more recent hydrodynamical simulations, such as Illustris~\citep{2014Natur.509..177V}, have even stronger effects on the power spectrum at large scales, and therefore can likely be constrained at least at the level of the {\it AGN} model we examine here.}, and {\it NOSN}, which lacks any feedback processes but implies a large enhancement at small scales due to cool gas altering the cores of halos. Other models with significant feedback or strong winds, such as {\it WDENS}, are the next-best constrained, in some cases even better than {\it NOSN} due to shape noise and other factors that reduce the capabilities of some surveys to probe the smallest scales.
When we consider measurements of $\xi_\pm$ down to $\theta_{\rm min}=2'$, shown in the right panels of Fig.~\ref{fig:sigma_aowls_absolute}, the corresponding increase in small-scale information serves to improve the constraining power by roughly a factor of two in many cases\footnote{We have also checked the case where $\theta_{\rm min}$ is set to $2'$ for $\xi_+$ and $\theta_{\rm min}=10'$ for $\xi_-$, since different scale cuts are sometimes used for $\xi_+$ and $\xi_-$ due to the increased vulnerability of $\xi_-$ to small-systematics as compared to the $\xi_+$. The results in this case are very similar to what we obtain using $\theta_{\rm min}=4'$ for both $\xi_+$ and $\xi_-$.}. The improvement is more modest for models where feedback reduces the impact of cooling on the clustering at the smallest scales, and in some cases, the slight shift in the angular coordinates of the measured $\xi_\pm$ data points even causes a very slight degradation in the constraints.
Given the importance of small scales for this analysis, we have also investigated the case where measurements of $\xi_\pm$ can be used down to $\theta_{\rm min}=0.5'$. The results for DES SV-like or Stage III datasets are shown in Fig.~\ref{fig:sigma_aowls_absolute_xi_thmin0p5}; for all other surveys we consider, the constraints on all baryonic models are better than 20$\sigma$. This conclusion is insensitive to the priors on systematic effects that we will return to in Sec.~\ref{sec:sys-priors}, but of course relies on sufficient control of all observational issues that could affect the usability of measurements at these scales. Nevertheless, it seems likely that data from Stage III surveys, or even partial data releases from these surveys, will be informative with respect to certain behaviors of baryons on the scales probed by cosmic shear measurements.
We have presented our results in the framework of distinguishing each model from the DM-only case, but with the high level of statistical significance that will be afforded by future surveys, it is likely that these measurements will be able to distinguish between different scenarios themselves. This can be accomplished by examining the joint posteriors on the amplitudes of each PC, and then comparing the projections of individual models onto these PCs. This procedure can in principle be applied to any model or simulation for which the total matter power spectrum is known, and may also be useful in constraining models containing their own continuous parameters.
Similar conclusions apply to cases where two-point statistics other than the correlation functions $\xi_\pm$ are used in the analysis. In Fig.~\ref{fig:xi_vs_cl}, we compare forecasts for LSST performed with either $\xi_\pm$ or the angular power spectrum $C_\ell$, with $\ell_{\rm max}$ chosen to reproduce as closely as possible the results from $\xi_\pm$ for each value of $\theta_{\rm min}$ we have previously considered. It is to be expected that $\theta_{\rm min}$ scales roughly as $\pi/\ell_{\rm max}$, and our explicit comparisons bear out this expectation, with $\theta_{\rm min} \simeq 1.3\pi/\ell_{\rm max}$. Specifically, we find that the $\xi_\pm$ forecasts with $\theta_{\rm min} = \{4',2',0.5'\}$ match quite closely with $C_\ell$ forecasts with $\ell_{\rm max}=\{3500,7000,30000\}$, respectively.
\subsection{Impact of systematics at small scales}
\label{sec:sys-priors}
\begin{figure*}
\includegraphics[width=1.0\textwidth]{figures/systematics_xi_tm4.pdf}
\caption{\label{fig:sys-xi-tm4}
For $\xi_\pm$ measurements with $\theta_{\rm min}=4'$, each panel shows the statistical significance (in ``number of sigmas", as in Figs.~\ref{fig:sigma_aowls_absolute}-\ref{fig:xi_vs_cl}) at which the specified baryonic model can be ruled out by the specified survey, as a function of the assumed level of two types of systematic uncertainty in the theoretical modeling. The parameter $S_\text{(shear)}$ denotes the uncertainty in the handling of shear-specific systematics (such as source-lens clustering) in the modeling of the shear correlation function, assumed to scale like $\Delta C_\ell \propto \ell^{0.5}$, while $S_P^\text{(high)}$ denotes the uncertainty in the DM-only power spectrum for $k<0.5\,h\, {\rm Mpc}^{-1}\,$ (see Sec.~\ref{sec:systematics} for details). Accuracy in both aspects of the modeling is more important for {\it AGN}-like models in which feedback affects the power spectrum on relatively large scales, while for cooling-dominated models. Meanwhile, Stage IV surveys will be sufficiently powerful that relatively large ($\sim$10\%) uncertainties in the modeling will not appreciably degrade the constraints.
}
\end{figure*}
Of the possible systematics we describe in Sec.~\ref{sec:systematics}, two of the more uncertain are the accuracy of the modeling of the DM-only power spectrum (including the effects of massive neutrinos) and the level of shear-specific systematics (e.g.\ lensing bias) that could be important when interpreting measurements at small scales. Our fiducial choice for the main results of the paper has been a 5\% uncertainty in the DM-only power spectrum for $k>0.5\,h\, {\rm Mpc}^{-1}\,$ (our $S_P^{\rm (high)}$ parameter) and a 5\% error budget for shear-specific systematics at $\ell=10^4$ (our $S_{\rm shear}$ parameter), with a profile scaling like $\Delta C_\ell \propto \ell^{0.5}$.
In Figs.~\ref{fig:sys-xi-tm4} and~\ref{fig:sys-xi-tm2}, we display representative variations in our forecasts when the priors on $S_P^{\rm (high)}$ and $S_{\rm shear}$ are varied away from their fiducial values, for the forecasts with $\theta_{\rm min}=4'$ and $2'$ respectively. For the {\it AGN} model and others where feedback affects the power spectrum on large scales, the accuracy in the modeling of the DM-only power spectrum is an important determinant of the overall constraining power, provided that the other shear systematics are controlled at a reasonable level. On the other hand, for models for which the constraining power is concentrated at smaller scales, the effects of baryons on the power spectrum at these scales are generally so strong that knowledge of the DM-only power spectrum to within 10\% will be sufficient to detect their presence from shear measurements alone.
Essentially the same conclusion holds for the prior on $S_{\rm shear}$: it has a relatively smaller impact on cooling-dominated models than on feedback-dominated scenarios. The increased precision of Stage IV surveys like LSST actually translates into {\it less} stringent requirements on the modeling of the DM-only power spectrum for {\it AGN}-like models, although the main science goals of these surveys (e.g.\ searches for new cosmological physics) will of course impose their own stricter requirements. In fact, for cooling-dominated models, we find that the constraints from Stage IV surveys are fairly insensitive to the systematics we vary in this section; measurements with $\theta_{\rm min} \lesssim 3'$ will be capable of distinguishing our entire test set of baryonic models from the DM-only case, even with $\sim$10\% uncertainties in the theoretical modeling on the relevant scales.
\section{Conclusions}
\label{sec:conclusions}
In this paper, we have assessed the ability of the two-point statistics of cosmic shear to act as a probe of the physics of galaxy formation, which is expected to impact the matter power spectrum. We have done so using a method that parameterizes the effects of baryons on the matter power spectrum in terms of a set of principal components (PCs). For a given survey, the method automatically classifies these PCs in terms of the expected constraints on each one, allowing for poorly-constrained PCs to be discarded from an analysis without a significant loss of information. These PCs are determined by the properties of a given survey, and are independent of the implementations of baryonic physics one might be interested in testing.
We have performed forecasts for the application of this method to a variety of surveys, ranging from those that have already been completed (the Dark Energy Survey Science Verification run) to Stage III (DES Y5, HSC) and Stage IV (WFIRST, LSST, Euclid) surveys. We have used the OWLS simulation suite, composed of nine different implementations of baryonic effects plus a reference dark matter-only run, to test how the (model-independent) output of the method translates into constraints on a range of specific models.
\begin{figure*}
\includegraphics[width=1.0\textwidth]{figures/systematics_xi_tm2.pdf}
\caption{\label{fig:sys-xi-tm2}
As Fig.~\ref{fig:sys-xi-tm4}, but corresponding to $\xi_\pm$ measurements with $\theta_{\rm min}=2'$ instead of $4'$.
}
\end{figure*}
We find that surveys operating now and planned in the near future have significant power to constrain the impact of galaxy formation on the matter power spectrum, with the potential to strongly constrain or rule out a majority of the models considered here. (Indeed, in~\citealt{2015MNRAS.450.1212H}, mild constraints were already obtained from CFHTLenS data, although using a different method than what we present here.) We emphasize that cosmic shear is a very clean signal to model: the connection between the observed correlations of galaxy shapes and the underlying matter clustering is completely determined by a weak-field calculation in general relativity, and does entail any of the astrophysical calibration or selection issues involved in many other probes of galaxy formation physics.
Our main results may be summarized as follows:
\begin{itemize}
\item For all surveys we consider, we find that 90\% of the constraining power with respect to baryonic effects on the matter power spectrum is encapsulated in no more than nine PCs, implying a likelihood analysis that makes use of this PCA method would need to add no more than nine additional parameters.
\item From these nine PCs, one can define a figure of merit as the reciprocal of the geometric mean of the expected one-sigma constraints on the corresponding amplitudes. By comparing this figure of merit for different surveys, we find that the constraints from Stage III surveys will improve on those from currently available shear measurements by roughly an order of magnitude, while Stage IV surveys will provide further improvements of a factor of a few. We argue that it is chiefly the number of galaxies with measured shapes that drives this improvement.
\item We find that the ultimate power of our method to constrain baryonic effects depends strongly on the minimum scale $\theta_{\rm min}$ at which measurements of the shear correlation functions $\xi_\pm$ can be used. In particular, if $\theta_{\rm min}$ can be pushed down to less than one arcminute, Stage III surveys should be able to rule out (at more than five-sigma confidence) the majority of the baryonic models we consider. We reach this conclusion after marginalizing over uncertainties in neutrino mass, photometric redshift distributions, shear calibration, and theoretical modeling of the power spectrum.
\end{itemize}
In our analysis, we have assumed that the background cosmology has already been mostly fixed by other measurements, such as temperature fluctuations in the cosmic microwave background. For our fiducial cosmology, we have chosen the latest best-fit model from Planck~\citep{2015arXiv150201589P}, marginalizing over their one-sigma uncertainties. In applying our method to data, care must be taken in choosing a fiducial cosmology that is consistent with the dataset being employed (accounting, for example, for the current tensions between cosmological constraints from weak lensing and the CMB), to avoid a wrong choice of cosmology biasing the constraints on baryonic effects. While we have not investigated the size of this possible bias in detail, it could easily be investigated in a future analysis simply by varying the choice of cosmology and examining the resulting change in the constraints on the PCs.
The primary product of a likelihood analysis making use of this method would be posterior constraints on the amplitudes of the most important PCs for that survey, with the PCs determined by a similar Fisher matrix procedure to what we have used here. The power spectrum from a given baryonic model can then be projected onto these PCs, and the resulting amplitudes can be compared with the constraints, without the need for an separate likelihood analysis corresponding to every model of interest. In this way, the results of a likelihood analysis including the PCs will act as a resource that can be used to assess the realism of currently-existing models or hydrodynamical simulations, and could also be used to guide the development of new models or simulations.
Furthermore, there currently exist models for baryonic effects on clustering that contain parameters connected to specific physical processes, such as halo expansion or gas ejection (e.g.~\citealt{2015MNRAS.454.1958M,2015JCAP...12..049S}). Constraints on the PCs from our method could also be mapped onto constraints on such parameters, and could therefore help to inform our knowledge about the related processes, in a more detailed manner than a wholesale acceptance or rejection of an individual model.
Finally, we note that even stronger constraints can likely be obtained by performing this kind of analysis on the combination of cosmic shear and other cosmological statistics. These include, for example, the cluster--mass correlation function, cluster Sunyaev-Zeldovich profiles, and galaxy--galaxy lensing. These may have significant additional power but introduce extra complexity in the modeling due for example to the need for a model for the cluster--halo connection in the first two cases and a bias model in the third case. Regardless, we have shown that, if sub-arcminute measurements can be robustly made, the connection between cosmic shear and the overall clustering of matter can be exploited to turn cosmic shear into an important probe of the physics of galaxy formation.
\section*{Acknowledgements}
This work received partial support from the U.S.\ Department of Energy under contract number DE-AC02-76SF00515.
We thank Tim Eifler, Andrew Hearin, Dragan Huterer, Elisabeth Krause, Niall MacCrann, and Michael Troxel for useful discussions and comments. We also thank Elisabeth Krause for providing code for computing the covariance matrices used in this work, and Stuart Marshall and the SLAC computing staff for essential computing support. This work has made extensive use of results from the OWLS simulation suite, and we thank the authors for making these data available from~\href{http://vd11.strw.leidenuniv.nl/}{http://vd11.strw.leidenuniv.nl/}.
This work also made use of the CosmoSIS software package, available from~\href{http://bitbucket.org/joezuntz/cosmosis}{http://bitbucket.org/joezuntz/cosmosis}. We thank Joe Zuntz and the CosmoSIS team for making this code available.
\bibliographystyle{mnras_sjf}
|
\section{}
\section{Introduction}
The reaction of H$_2$ + OH has emerged as a prototype reaction for four-atomic
systems. It contributes to fundamental processes in atmospheric chemistry,
astrochemistry, and combustion.\cite{pie09, rah03, gli15} The reaction being
surface catalyzed was shown to be one of the main routes of H$_2$O formation
in the interstellar medium.\cite{cup07,cup10,lam13a,lam14}
On surfaces the reaction was observed at the cryogenic temperature of 10~K
through quantum mechanical tunneling of atoms.\cite{oba12,mei16}
\new{For the gas phase reaction, a} number of studies on this reaction,
theoretical\cite{bha10, bha11, esp10, ngu11, fu10, fu15, cha04, man00} as well as experimental
\cite{rav81,tal96,kra04,ork06,lam13} down to 200~K has been performed. For an
overview of previous experimental and theoretical results, we refer to
reviews.\cite{cas02,smi02a}
In this article we present reaction rate constants of the title reaction down
to 100~K using instanton theory\cite{lan67,mil75,col77,cal77,gil77} and down
to 50~K using CVT/$\mu$OMT.\cite{liu93a,Truhlar_Faraday1994} Instanton
theory\new{\cite{aff81,col88,han90,
ben94,mes95,ric09,alt11,rom11,rom11b,ric16}} is a semiclassical theory
based on Feynman's path integrals.\cite{fey48} It takes multidimensional
tunneling into account while only the optimization of a tunneling path -- the
instanton -- is necessary.\cite{kae14} Instanton theory is meanwhile
frequently used to calculate reaction rates in different areas of
chemistry.\cite{cha75,mil94,mil95,mil97,sie99, sme03,qia07,and09,
gou10a,gou11,gou11b, rom11,gou10,jon10,mei11,gou11a,ein11,rom12,
kry12,kae13,alv14,kry14} Canonical variational theory \new{(CVT) minimizes
recrossing compared to transition state theory (TST). It} was used with
microcanonical optimized multidimensional tunneling
(CVT/$\mu$OMT)\cite{liu93a,Truhlar_Faraday1994} was used along with zero
curvature tunneling (ZCT),\cite{gar80,tru83} small curvature tunneling
(SCT),\cite{sko81} large curvature tunneling (LCT),\cite{gar83,gar85,fer01}
and microcanonical optimized multidimensional tunneling
($\mu$OMT)\cite{liu93a,Truhlar_Faraday1994} calculations down to 50~K. ZCT
assumes no deviation of the tunneling path from the classical minimum energy
path. Compared to that, SCT considers corner-cutting effects and LCT
approximates the tunneling path by a linear path from reactants' valley to
products' valley. The $\mu$OMT method takes into account that the tunneling
path depends on the energy by using the maximum of SCT and LCT tunneling
probabilities at each energy.\cite{fer06,fer07}
As the four-atomic system of H$_2$ + OH is of fundamental interest, a variety
of potential energy surfaces (PES) have been published.\cite{och98, wu00,
wal80,sch80,yan01,bet00} Recently, a global potential energy surface fitted
by a neural network to UCCSD(T)-F12a/AVTZ data was published (NN1
PES).\cite{che13} This PES was shown to give reliable results in, e.g.,
the study of the mode specificity of the H + HOD reaction. \cite{fu15} The NN1
PES was therefore applied here as well. Although several studies on thermal
rate constants for the title reaction appeared, \cite{ngu11,man00,mat98} for
instance, the semiclassical transition state theory (SCTST) calculations of
Nguyen et al. \cite{ngu10} -- even to investigate reaction rate constants of
all isotopologues \cite{ngu11} -- it seems that this is the first study which
provides rate calculations on the NN1 PES.
\new{The reaction profile consisting of the stationary points are shown in
\figref{reactionprofile}.}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=8cm]{figure1.eps}
\caption{Potential energy profile of the reaction H$_2$ + OH $\rightarrow$
H$_2$O + H. Relative to the separated reactants, the pre-reactive complex
has a potential energy of $-$2.1~kJ~mol$^{-1}$, the transition state
22.5~kJ~mol$^{-1}$ and the separated products $-$68.1~kJ~mol$^{-1}$.
\label{reactionprofile}
}
\end{center}
\end{figure}
For bimolecular reactions, it is in general possible that a (weakly) bound
van-der-Waals complex can lead to an increase of the bimolecular reaction rate
constant with decreasing temperature. This effect was studied experimentally
in the reaction of HBr + OH as well as in the reactions with nitric acid or
alcohols and OH radicals. \cite{ree15, bro01, sha13} In these cases, the
non-covalent interactions between the two reactants stem from the dipole
moments and polarizabilities of the reacting molecules.
\new{In contrast to these,
H$_2$ is less polarizable and has no permanent dipole moment.
Thus, the intermolecular interaction between H$_2$ and the OH radical
and the impact of the pre-reactive complex (PRC) [H$_2 \cdots$OH] are
expected to be small unless the temperature is much lower than
considered in this work.}
In this study we investigate the temperature dependence of the reaction rate
constant and compare it to published values. Furthermore, the temperature
dependence for the rate constants for all eight possible isotopologue
reactions and the resulting kinetic isotope effects (KIEs) have been studied.
\new{At low temperatures (below $T_\text{c}$) tunneling dominates} the reaction
rate. The nuclear mass has a high impact on the tunneling probability leading
to large kinetic isotope effects (KIEs).
\section{Methods}
In instanton theory, the instanton, the saddle point corresponding to the
transition state, is a closed Feynman path folded back onto itself which spans
the barrier region. At high temperature it is short and covers only the top of
the barrier while at low temperature it protrudes right into the reactant
state region. At the crossover temperature,
\begin{equation}
T_\text{c} = \frac{\hbar \omega_\text{TS}}{2 \pi k_\text{B}},
\label{eq:2}
\end{equation}
the instanton path generally collapses to a point and the theory becomes
inapplicable although extensions above the crossover temperature exist.\cite{zha14}
Here $\omega_\text{TS}$ is the absolute value of the imaginary
frequency at the transition state, $\hbar$ is the reduced Planck's constant,
and $k_\text{B}$ is Boltzmann's constant. $T_\text{c}$ is mass-dependent: for
the title reaction containing protium only is was found to be 276.2~K, while
204.2~K was found for the per-deuterated reaction. In many cases
$T_\text{c}$ can be used as a cheap and simple indication if atom tunneling is
important at the temperature of interest. Following \eqref{eq:2}, whenever
$\omega_\text{TS}$ is larger than 1300~cm$^{-1}$ atom tunneling is relevant at
room temperature.
The different H/D isotopologues are labeled as H$^1$H$^2$OH$^3$ such that the
reaction reads $\textnormal{H}^1\textnormal{H}^2+
\textnormal{O}\textnormal{H}^3 \rightarrow \textnormal{H}^1 +
\textnormal{H}^2\textnormal{O}\textnormal{H}^3$. DDOH therefore corresponds
to a reaction of OH with D$_2$ while HDOH corresponds to the reaction HD + OH
$\rightarrow $ H + DOH.
Vibrational modes were described by the harmonic approximation of the Feynman
path. The translational partition function was in all cases approximated by
the one of the ideal gas, which is identical to that of a quantum mechanical
particle in a box. The rotational partition function of the transition state
was obtained as the geometric mean value of the rotational partition functions
of all images along the instanton path treated as rigid quantum rotors. The
reactant molecules were, equivalently, treated as rigid rotors. The symmetry
number, the order of the rotational subgroup in the molecular point
group,\cite{fer07a} of the individual molecules was taken into account in the
rotational partition function, i.e. the one of H$_2$ and D$_2$ was divided by
two, while the one for HD is not.
The kinetic isotope effects are dominated by tunneling and by the zero-point
vibration. Neither the rotational nor the translational contribution have a
significant effect on the KIEs.
The NN1 PES\cite{che13} was interfaced with DL-FIND.\cite{kae09a}
Instantons were optimized starting from the classical transition state
or by starting from an already optimized instanton of similar temperature using the adapted
Newton--Raphson algorithm implemented in DL-FIND. \cite{rom11,rom11b}
The convergence criteria for the instanton optimization on the NN1 PES was $5\cdot 10^{-11}$
atomic units for the maximal component of the gradient. Note, that we use mass-weighted
coordinates and gradients with the masses in atomic units, i.e. relative to
the electron mass. This influences the convergence criterion.
The instanton is a closed Feynman path with images having pairwise identical
coordinates. The full path was represented by 512~images. Convergence with
respect to the number of images was tested at the most severe case with the
largest distances between adjacent images, the all-H reaction (HHOH) at 100~K.
In this case, the rate constant obtained with 4096 images for the full path
deviated by only 0.4~\% from the value obtained with 512~images \new{and is
only 2.4~\% higher compared to the value obtained with 194~images.} Smaller
deviations can be expected at higher temperature or for heavier
isotopologues. Thus, we consider the discretization to be converged with
respect to the number of images.
To test the quality of the NN1 PES we additionally calculated instanton rate
constants with on-the-fly energy calculations at the CCSD(T)-F12
level\cite{adl07,adl09} using the cc-pVDZ-F12 basis set.\cite{pet08} The
program package Molpro version 2012\cite{MOLPRO2012,wer11} interfaced to
DL-FIND\cite{kae09a} via ChemShell\cite{met14} was used for these
calculations. Due to the high computational demands of these calculations, 194
images were used and the
instanton optimizations were considered converged for all absolute gradient
components smaller than $10^{-8}$~a.u.
Below 100~K, the instanton path for HHOH stretches into the pre-reactive minimum with
parts of the path below the energy of the separated reactants, see
\figref{fig:Instanton}.
\new{Instanton rates are not valid for energies below the separated reactants, so
instanton rates are reported only down to 100~K for H-transfer.}
For the D-transfer the
whole instanton path remains above the reactants' energy for $T \textgreater
80$~K. Thus, instanton rates for D-transfer reactions are reported down to
80~K.
\new{CVT/$\mu$OMT was used down to 50~K.}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=16cm]{figure2.eps}
\caption{Potential energy along the instanton path at
130~K (left) and 80~K (right) relative to the energy of the
separated reactants. At 130~K the whole instanton path is above
the reactant's energy, at 80~K its ends are below that value.
Images of the corresponding instantons are inserted.
\label{fig:Instanton}
}
\end{center}
\end{figure}
The ZCT, SCT, LCT, and $\mu$OMT calculations on the NN1 PES\cite{che13} have
been performed using POLYRATE \new{2010}\cite{lu92,POLYRATE2010} based on canonical
variational transition state theory (CVT).\cite{fer06,fer07} For the LCT
calculations, the action integrals ($\theta$ integrals) and the sine of the
angle between the minimum energy path and the tunneling path were interpolated
to 2$^{nd}$ order.
\section{Results}
\subsection{Reaction Rate Constants}
The relevant stationary points on the potential energy surface of the title
reaction are depicted in \figref{reactionprofile}. Relative to the separated
reactants H$_2$ and OH, the potential energy on the NN1 PES is
$-2.11$~kJ~mol$^{-1}$ for the PRC, $22.50$~kJ~mol$^{-1}$ for the transition
state (TS) and $-68.08$~kJ~mol$^{-1}$ for the products (H + H$_2$O).
The relative energies of the corresponding stationary points
optimized on CCSD(T)-F12/cc-pVDZ-F12 level
are $-$1.77 kJ/mol for the PRC, 23.98 kJ/mol for the TS and $-$64.94 kJ/mol for
the products.
The imaginary harmonic frequency is 1206~i~cm$^{-1}$ on the NN1 PES and
1199~i~cm$^{-1}$ on CCSD(T)-F12/cc-pVDZ-F12 level.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=8cm]{figure3.eps}
\caption{Reaction rate constants for HHOH
compared to literature data. Experimental data:
``{\textcolor{red}{$\blacksquare$}}'' data from \cite{ork06},
``{\textcolor{green}{$\blacklozenge$}}'' data from \cite{tal96},
``{\textcolor{blue}{$\blacktriangle$}}'' data from \cite{rav81};
computational data:
``{\textcolor{red}{$\square$}}'' data from \cite{man00},
``{\textcolor{green}{$\bigcirc$}}'' data from \cite{ngu10},
``{\textcolor{blue}{$\Diamond$}}'' data from \cite{mat98}.
\label{fig:rates}
}
\end{center}
\end{figure}
The rate constant of the title reaction has been measured several times using
different techniques, see \figref{fig:rates}. An Arrhenius plot shows a
noticeable curvature already at 300~K and below,\cite{ork06,rav81} which is a
clear sign that the reaction is influenced by atom tunneling. Experimental
rate constants are available from 1000~K down to
200~K.\cite{tal96,ork06,lam13,rav81,kra04} The different sets agree quite
well, typically within 20--30\% of the rate constant.
Among the computational studies of this system, Matzkies and Manthe
\cite{mat98,man00} performed full-dimensional quantum dynamics calculations on
the Schatz--Elgersma PES \cite{wal80,sch80} using the multi-configuration
time-dependent Hartree \new{(MCTDH)} approach. At $T=300$ K, the lowest temperature
covered by close-coupling calculations employing a rigorously correct
statistical sampling scheme for the rotational degrees of freedom,\cite{man00}
their calculations overestimate the experimental rate constants by about a
factor of 2. In an earlier work,\cite{mat98} they calculated rate constants
down to 200 K, which are more than an order of magnitude higher than the
experimental value.\cite{ork06} Better agreement with the experimental values
was achieved by Nguyen et al.\cite{ngu10} by applying semiclassical
transition-state theory (SCTST) on high-level direct-dynamics energies. At
200~K they underestimate the experimental rate constant by a factor of 1.43,
which is comparable to the experimental error bar. They furthermore showed that
SCT gives significantly higher reaction rate constants at lower temperatures
compared to SCTST calculations. \cite{ngu10}
For the title reaction containing protium only (HHOH), the crossover
temperature is $T_\text{c}=276.2$ K. Using instanton theory, we calculated
bimolecular reaction rates using the NN1 PES at $T=270$ K and below as
it is only applicable below $T_\text{c}$. The results are
depicted in \figref{fig:rates}, numbers are given in the supporting
information.\cite{ThisSI} As expected,\cite{gou11a} instanton theory overestimates the
rate constant close to $T_\text{c}$. Agreement is improved at lower temperature. At
220~K our rate constant ($1.78\cdot10^{-16}$~cm$^3$~s$^{-1}$) is higher by a
factor of 1.95 than the results of flash photolysis resonance-fluorescence by
Orkin et al.;\cite{ork06} at 200~K, the lowest temperature at which comparison is
possible, it is still higher by a factor of 1.76.
A higher accuracy of
instanton theory can be expected at lower temperature due to the known
overestimation close to $T_\text{c}$.\cite{gou11a}
The deviation might reflect a deficiency of the rate theory or the
potential. To test the accuracy of the potential, we recalculated instantons and
rate constants on-the-fly on CCSD(T)-F12/cc-pVDZ-F12 level without fitting the
PES. The reaction rate constants obtained in this way agree better with the
experimental result,
overestimating it by a factor of 1.49 at 240~K and only by a factor of 1.12 at 200~K.
\new{
Thus, the NN1 PES leads to a slight overestimation of the rate constants.
However, we continue with the NN1 PES as
on-the-fly calculations for all isotopologues would be too costly.
}
While the absolute rate constants might be overestimated by a factor of
approximately 1.5 to 2.0, one can assume the KIEs incur smaller errors. We
assume roughly the same level of accuracy at low temperatures, where no other
experimental or computational data are available for comparison.
Given that instanton theory is expected to
be more accurate at lower temperatures, the rates using the NN1 PES, which
employs a better basis set, are more promising than the direct-dynamics
results. Still, the NN1 PES probably shows slight inaccuracies in the region
of the configurational space most relevant to tunneling, leading to a slight
overestimation of the reaction rate. However, we continue with the NN1
PES. While the absolute rate constants might be overestimated by a factor of
approximately 1.5 to 2.0, one can assume the KIEs incur smaller errors. We
assume roughly the same level of accuracy at low temperatures, where no other
experimental or computational data are available for comparison.
To test the accuracy of the rate theory, we performed CVT/ZCT, SCT, LCT,
and $\mu$OMT calculations on the NN1-PES. As in
this reaction the SCT rate constant is always higher than the one
obtained by LCT, the $\mu$OMT result is virtually indistinguishable from the results
obtained by SCT. Therefore no graph for SCT is shown in
\figref{fig:rates}. At temperatures below 300~K, CVT/LCT (and ZCT)
agrees well with instanton theory whereas SCT, and thus $\mu$OMT, give
significantly higher rate constants. At 200~K, $\mu$OMT overestimates
the reaction rate constants by a factor of 4.0, see
\tabref{tab:rates} and \figref{fig:rates}.
For comparison, \figref{fig:rates} includes the rate constant calculated
\new{by TST with all vibrations treated via quantum partition functions of
harmonic oscillators, i.e., accounting for the vibrational zero-point energy
but not for tunneling (Harmonic transition state theory, HTST)}. As
expected, it describes the rate constant very well at high temperatures (close
to and above 400 K) but deviates significantly below about 300 K.
\begin{table}[ht!]
\caption{
\label{tab:rates}
Reaction rate constants $k$ in cm$^3$~molecule$^{-1}$~s$^{-1}$ at 200~K obtained by different methods.
Experimental values are from reference \citenum{ork06}.
}
\begin{center}
\setlength{\tabcolsep}{1mm}
\begin{tabular}{ll}
\hline
Exp. & 4.3 $\cdot 10^{-16}$ \\
Instanton NN1 & 7.56$\cdot 10^{-16}$ \\
Instanton on-the-fly & 4.83$\cdot 10^{-16}$ \\
CVT/$\mu$OMT & 1.73$\cdot 10^{-15}$ \\
CVT/SCT & 1.73$\cdot 10^{-15}$ \\
CVT/LCT & 8.77$\cdot 10^{-16}$ \\
CVT/ZCT & 7.00$\cdot 10^{-16}$ \\
CVT & 2.76$\cdot 10^{-17}$ \\
\new{H}TST & 3.25$\cdot 10^{-17}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Kinetic Isotope Effects}
All eight possible isotopologues were investigated. The zero-point
energy (ZPE) corrected energies of PRC and TS, as well as $T_\text{c}$
are given in \tabref{tab:ZPE}, the rate constants are shown in
\figref{fig:iso}. Values of the KIEs at 160~K and 100~K (both instanton
and $\mu$OMT), and 50~K ($\mu$OMT) are given in \tabref{tab:KIE}.
For reactions with deuterium, the crossover temperature
is significantly reduced, see \tabref{tab:ZPE}.
Down to 120~K the curvature of the
resulting Arrhenius plot in \figref{fig:iso} is negligible for isotopologues
with D-transfer. Defazio et al. already mentioned that tunneling may not be
very important in the DDOH case.\cite{def03} This is certainly true in the
temperature range where experimental data is available, i.e. above 210~K.
\new{At lower temperature the reactions of all isotopologues are dominated by tunneling.}
A direct comparison between our instanton calculations and experimental data is
impossible for any of the deuterated cases, as no data is available below
$T_\text{c}$. Above 50 K a clear primary KIE, i.e., depending on the mass of
the atom to be transferred, is measurable.
\begin{table}[h!]
\caption{
\label{tab:ZPE}
ZPE corrected energies of the corresponding characteristic points of the
PES in~kJ~mol$^{-1}$ relative to the separated reactants. The crossover
temperature $T_\text{c}$ is given in K. $E_\text{a}$ refers to the
activation energy, the energy difference between TS and PRC.}
\begin{center}
\setlength{\tabcolsep}{1mm}
\begin{tabular}{crccccr}
\hline
& PRC & TS & $E_\text{a}$ &
$E_\text{a}$(ref \cite{ngu11})& $T_c$ \\
\hline
HHOH & 0.54 & 24.76 & 24.22 & 24.41 & 276.2 \\
HHOD & 0.41 & 23.50 & 23.09 & 23.19 & 276.1 \\
DHOH & 0.30 & 24.13 & 23.82 & 23.74 & 266.0 \\
DHOD & 0.17 & 22.84 & 22.67 & 22.48 & 265.8 \\
HDOH & 0.20 & 25.86 & 25.66 & 25.37 & 208.9 \\
HDOD & 0.04 & 24.57 & 24.53 & 24.12 & 208.8 \\
DDOH & $-$0.04 & 25.65 & 25.69 & 25.16 & 204.3 \\
DDOD & $-$0.22 & 24.31 & 24.53 & 23.86 & 204.2 \\
\hline
\end{tabular}
\end{center}
\end{table}
The KIE can stem from differences in zero-point energies or from
tunneling.\cite{per12a,sul13, per14} One may, of course, argue if the
harmonic approximation for zero-point energies is good enough to
estimate rate constants at such low temperatures. However, our
calculated vibrationally adiabatic barriers of the isotopologues agree
well (deviation $<0.7$~kJ~mol$^{-1}$) from the literature values
obtained by the more elaborate HEAT protocol,\cite{ngu11} see
\tabref{tab:ZPE}. It was shown previously \cite{ngu11} that including
anharmonicity changes the corresponding barrier height by less than
0.33~kJ~mol$^{-1}$.
Apart from the primary KIE, we observe that deuteration of the hydroxy
radical (OD) increases the reaction rate, leading to inverse KIEs.
Depending on the deuteration of the other sites, OD increases the
rates by factors of 1--3, see \tabref{tab:KIE}. The main reason for
this effect is that the heavier deuterium atom lowers $E_\text{va}$ of
the transition state by reducing the zero-point energy of the
deformation modes of the two molecules with respect to each other.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=8cm]{figure4a.eps}
\includegraphics[width=8cm]{figure4b.eps}
\caption{Temperature dependence of the reaction rate constants of
all H/D isotopologues calculated with the instanton method (left) and
with CVT/$\mu$OMT (right).
\label{fig:iso}
}
\end{center}
\end{figure}
\begin{table}[ht!]
\caption{
\label{tab:KIE}
Kinetic isotope effects at \new{160~K}, 100~K, and 50~K with respect to HHOH.
}
\begin{center}
\begin{tabular}{l| D{.}{.}{3}D{.}{.}{3}| c D{.}{.}{3}D{.}{.}{3}D{.}{.}{3}}
\hline
&\multicolumn{2}{c|}{Instanton} &&\multicolumn{3}{c}{$\mu$OMT}\\
Isotopes &\multicolumn{1}{c}{160 K} &\multicolumn{1}{c|}{100 K}
&\,&\multicolumn{1}{c}{160 K} &\multicolumn{1}{c}{100 K} &\multicolumn{1}{c}{50 K} \\
\hline
HHOH & 1.00 & 1.00 & & 1.00 & 1.00 & 1.00 \\
DHOH & 3.19 & 3.95 & & 3.37 & 4.99 & 6.89 \\
HHOD & 0.649 & 0.561 & & 0.828 & 0.732 & 0.610 \\
DHOD & 1.99 & 2.04 & & 2.58 & 3.19 & 3.42 \\
HDOH & 41.5 & 224 & & 49.1 & 176 & 382 \\
DDOH & 30.2 & 225 & & 32.4 & 172 & 558 \\
HDOD & 23.9 & 108 & & 35.4 & 109 & 216 \\
DDOD & 17.1 & 104 & & 22.3 & 95.3 & 229 \\
\hline
\end{tabular}
\end{center}
\end{table}
The reaction rate constants obtained with $\mu$OMT are higher than the ones
obtained with instanton theory by a factor of 4.2 for HHOH and 5.4 for HDOH at
100~K. It is obvious from \figref{fig:rates} and \tabref{tab:rates} that
CVT/$\mu$OMT generally overestimates the reaction rate constants for this
reaction, see also Fig. S1\new{, because in $\mu$OMT the tunneling path is not
optimized}. Apart from that, the rate constants seem to follow the same
trends, in particular the KIEs obtained by both methods agree reasonably well,
see \tabref{tab:KIE}.
Instanton theory provides a dominant tunneling path for each specific
temperature. At low temperature, that path is almost
temperature-independent. The atoms contribute quite differently to that
tunneling path. Geometries and the energy along the instanton path are depicted in
\figref{fig:Instanton}. In the low-temperature limit for HHOH, the hydrogen
atom to be transferred is delocalized over 1.34~\AA{}, the one that remains as
isolated hydrogen atom after the reaction over 0.80~\AA{}. Both oxygen and
hydrogen of OH contribute to the tunneling much weaker, they are delocalized
over 0.14 and 0.21~\AA{}, respectively. Deuteration changes these
contributions: for HDOH, the transferred deuterium is delocalized over only
1.25~\AA{} while the other tunneling path length remain almost unchanged
(0.77, 0.15, and 0.21~\AA).
We found primary H/D-KIEs of $>200$ at 50~K using CVT/$\mu$OMT. At even lower
temperature than reported here, the KIE can be expected to be at least
similarly strong. Consequently we expect a significant influence of this
reaction and its KIE on the deuterium fractionation of molecules in the
interstellar medium.
\section{Summary}
We calculated reaction rate constants of H$_2$ + OH $\rightarrow$ H + H$_2$O
down to 100~K using instanton theory and down to 50~K using CVT/$\mu$OMT on
the NN1 PES \cite{che13} for all H/D isotopologues. Atom tunneling sets in at
about 250~K for H-transfer and at about 200~K for D-transfer. A significant
primary H/D KIE of about 200 is found at 100~K and of 300--600 at 50~K.
At 80--50~K the reaction rate constants of the H-transfer reaction become
almost temperature-independent due to atom tunneling. Our results clearly
indicate that the title reaction may well be relevant for processes in the
interstellar medium at even lower temperature, even including deuterium.
\begin{acknowledgments}
Alexander Denzel is acknowledged for performing the initial calculations.
This work was financially supported by the German Research Foundation (DFG)
within the Cluster of Excellence in Simulation Technology (EXC 310/2) at the
University of Stuttgart.
\end{acknowledgments}
|
\section{Statistical Background}
\label{sec:statistical_background}
Asymptotic analysis plays an important role in statistics. It considers the extreme cases where the number of observations is increased with no bounds. In such scenarios, discussions on different notions of convergence of the sequence of random variables naturally arise. Generally speaking, there are three major types of stochastic convergence: \emph{convergence in probability, convergence in law (distribution)} and \emph{convergence with high probability (almost surely)}.
Here, we focus on the two former modes of convergence, discuss two fundamental results based on them and formalize our notations regarding parameter estimators. Further details of the following definitions and results can be found in any standard statistical textbook such as~\citet{lehmann1998theory}.
\subsection{Convergence of Random Variables}
Throughout this section, $\{\bz_1,\bz_2,...,\bz_n,...\}$, denoted simply by $\{\btheta_n\}$, is a sequence of multivariate random variables lying in $\Omega\subseteq\mathbb{R}^d$. Also suppose that $\bz_0$ is a constant vector and $\tilde{\bz}$ is another random variable in the same space $\Omega$.
\begin{definition}
\label{defn:convergence_in_prob}
We say that the sequence $\{\bz_n\}$ \emph{converges in probability} to $\bz_0$ and write $\bz_n\overset{P}{\rightarrow}\bz_0$, iff for every $\varepsilon>0$ we have:
\begin{equation}
P(|\theta_{ni}-\theta_{0i}|>\varepsilon) \enskip \to \enskip 0,\quad \mbox{for all } i=1,...,d.
\end{equation}
\end{definition}
Convergence in probability is invariant with respect to any continuous mapping:
\begin{proposition}[\citet{brockwell1991time}, Proposition 6.1.4]
\label{prop:convergent_continuous_mappings}
If $\bz_n\overset{P}{\to}\bz_0$ and $g:\Omega\to\mathbb{R}$ is a continuous function at $\bz=\bz_0$, then $g(\bz_n)\overset{P}{\to}g(\bz_0)$.
\end{proposition}
\begin{definition}
\label{defn:convergence_in_law}
We say that a sequence $\{\bz_n\}$ \emph{converges in law (in distribution)} to the random variable $\tilde{\bz}$ and write $\bz_n\overset{L}{\rightarrow}\tilde{\bz}$, iff the sequence of their joint CDFs, $F_n$, point-wise converges to the joint CDF of $\tilde{\bz}$:
\begin{equation}
\label{eq:convergence_in_law}
F_n(\mathbf{a}) = P(\theta_{n1}\leq a_1,...,\theta_{nd}\leq a_d) \enskip \rightarrow \enskip F(\mathbf{a}) = P(\tilde{\theta}_{1}\leq a_1,...,\tilde{\theta}_{d}\leq a_d) \quad \forall\mathbf{a}\in\calc_{F}\subseteq \mathbb{R}^d,
\end{equation}
where $\calc_{F}$ is the set of continuity points of the CDF $F$.
\end{definition}
Equation~(\ref{eq:convergence_in_law}) means that for large values of $n$, the distribution of $\bz_n$ can be well approximated by the distribution of $\tilde{\bz}$. Note that throughout this paper, for simplicity, we say that a random sequence $\{\bz_n\}$ converges to a distribution with density function $p(\bz)$, or write $\bz_n\overset{L}{\to} p(\bz)$, instead of fully saying that $\{\bz_n\}$ converges in law to a random variable with that distribution.
Note that $\bz_n \overset{P}{\to}\bz_0$ suggests that $\bz_n-\bz_0 \overset{L}{\to}\delta(\bz)$ where $\delta$ is the Kronecker delta function, which can be viewed as the density function of a degenerate distribution at $\bz=\mathbf{0}$. This, however, does not give any information about the speed with which $\bz_n$ converges to $\bz_0$. In order to take the speed into account, we consider the convergent distribution of the sequence $\{a_n\cdot(\bz_n - \bz_0)\}$, where $a_n$ is any sequence of positive integers and $a_n\to\infty (n\to\infty)$. In practice $a_n$ is usually considered to have the form $n^r$ with $r>0$.
\begin{definition}
\label{defn:convergence_rate}
Assume $\bz_n\overset{P}{\to}\bz_0$. We say that the sequence $\{\bz_n\}$ converges to $\bz_0$ with \emph{rate of convergence} $r>0$, iff $n^r(\bz_n-\bz_0)$ converges in law to a random variable with non-degenerate distribution. Furthermore, the non-degenerate distribution is the \emph{asymptotic distribution} of $\bz_n$.
\end{definition}
\noindent Next, we discuss some of the classic results in asymptotic statistics:
\begin{theorem}[\bfseries Law of Large Numbers, \citet{brockwell1991time}]
\label{thm:LLN}
Let $\btheta_1,...,\btheta_n$ be a set of independent and identically distributed (i.i.d) samples. If $\mathbb{E}[\btheta_i] = \bmu$, then
\begin{equation}
\label{eq:LLN}
\bar{\btheta}_n \equals \frac{1}{n}\sum_{i=1}^n\btheta_i \overset{P}{\rightarrow}\bmu.
\end{equation}
\end{theorem}
\begin{theorem}[\bfseries Central Limit Theorem, \citet{lehmann1998theory}]
\label{thm:CLT}
Let $\btheta_1,...,\btheta_n$ be a set of i.i.d samples with mean $\mathbb{E}[\btheta_i]=\bmu$ and covariance $\mbox{\normalfont Cov}[\btheta_i] =\bSigma$ (with a symmetric and positive semi-definite matrix $\bSigma$), then the sequence of sample averages $\left\{\bar{\btheta}_n\right\}$ with $\bar{\bz}_n=\frac{1}{n}\sum_{i=1}^n\btheta_i$ converges to the true mean with convergence rate $1/2$. Moreover, its asymptotic distribution is a zero-mean Gaussian distribution with covariance matrix $\bSigma$, that is:
\begin{equation}
\label{eq:CLT}
\sqrt{n}\cdot(\bar{\btheta}_n - \bmu) \enskip\overset{L}{\rightarrow}\enskip \caln(\mathbf{0},\bSigma).
\end{equation}
\end{theorem}
\noindent The following results are very useful when deriving the asymptotic distribution of a random sequence under a continuous mapping:
\begin{theorem}
\label{thm:multivariate_delta_method}
{\normalfont\bfseries (Multivariate Delta Method, first order, \citet{lehmann1998theory})} Let $\{\bz_n\}$ be a sequence of random variables such that it converges to $\bz_0$ with rate of convergence $1/2$ and a normal asymptotic distribution, that is $\sqrt{n}\cdot(\bz_n-\bz_0)\overset{L}{\to}\caln(\mathbf{0},\bSigma)$. If $g:\mathbb{R}^d\to\mathbb{R}$ is a continuously differentiable mapping and $\nabla_{\bz}g(\bz_0)\neq\mathbf{0}$, then
\begin{equation}
\sqrt{n}\cdot \bigg[g(\bz_n) - g(\bz_0)\bigg]\enskip \overset{L}{\to} \enskip \caln\left(0,\nabla_{\bz}^\top g(\bz_0)\bSigma\nabla_{\bz} g(\bz_0)\right).
\end{equation}
\end{theorem}
\begin{theorem}[\bfseries Multivariate Delta Method, second order]
\label{thm:delta_method_zeroder}
Let $\{\bz_n\}$ be a sequence of random variables such that it converges to $\bz_0$ with rate of convergence $1/2$ and a normal asymptotic distribution, that is $\sqrt{n}\cdot(\bz_n-\bz_0)\overset{L}{\to}\caln(\mathbf{0},\bSigma)$. If $g:\mathbb{R}^d\to\mathbb{R}$ is a continuously differentiable mapping where $\nabla_{\bz}g(\bz_0)=\mathbf{0}$ and $\nabla_{\bz}^2g(\bz_0)$ is non-singular in a neighborhood of $\bz_0$, then the sequence $\{g(\bz_n) - g(\bz_0)\}$ converges in law to a mixture of random variables with first-degree Chi-square distributions, and the rate of convergence is one. More specifically,
\begin{equation}
\label{eq:delta_method_zeroder}
n\cdot \bigg[g(\bz_n) - g(\bz_0)\bigg]\enskip \overset{L}{\to} \enskip \sum_{i=1}^d\lambda_{i}\chi_1^2,
\end{equation}
where $\lambda_{i}$'s are eigenvalues of $\bSigma^{1/2}\nabla_{\bz}g(\bz_0)\bSigma^{1/2}$.
Moreover, variance of this asymptotic distribution can be written as
\begin{equation}
\label{eq:var_delta_method_zeroder}
\frac{1}{2}\left\|\bSigma^{1/2}\nabla_{\bx}^2g(\bx_0)\bSigma^{1/2} \right\|_F^2,
\end{equation}
where $\|\cdot\|_F$ is the Frobenius norm.
\end{theorem}
\begin{proof}
For proof see Appendix~\ref{app:proof_2nd_Delta_Method}.
\end{proof}
\subsection{Parameter Estimation}
\label{subsec:parameter_estimation}
\noindent Now suppose that the set of independent and identically distributed (i.i.d) set of samples $\bx_1,...\bx_n$ are generated from an underlying distribution that belongs to a parametric family, for which the density function $p(\bx|\btheta)$ can be represented by a multivariate parameter vector $\btheta$. Assume the true parameter is $\btheta_0$, that is $\{\bx_i\}\sim p(\bx|\btheta_0),i=1,...,n$. An \emph{estimator} $\btheta_n=\btheta(\bx_1,...,\bx_n)$ is a function that maps the observed random variables to a point in the parameter space $\Omega$. The subscript $n$ in $\btheta_n$ indicates its dependence on the sample size. Since the observations are generated randomly, the estimators are also random and thus $\{\btheta_n\}$ can be viewed as a sequence of random variables. There are some reserved terms for such a sequence, which we introduce in the remaining of this section:
\begin{definition}[\bfseries Consistency]
We say that an estimator $\btheta_n$ is \emph{consistent} iff $\btheta_n\overset{P}{\to}\btheta_0$.
\end{definition}
\noindent Based on Theorem~\ref{thm:LLN}, sample average of the observation set is a consistent estimator of the true mean of the samples. Another important characteristic of estimators is based on the following bound over their covariance matrices:
\begin{theorem}[\bfseries Cram\'{e}r-Rao, \citet{lehmann1998theory}]
\label{thm:CR_bound}
Let $\bx_1,...,\bx_n\sim p(\bx|\btheta_0)$ and $\btheta_n=\btheta(\bx_1,...,\bx_n)$ be an estimator. If the first moment of $\btheta_n$ is differentiable with respect to the parameter vector and its second moment is finite, then the following inequality holds for every $\btheta\in\Omega$:
\begin{equation}
\label{eq:CR_bound}
\mbox{\normalfont{Cov}}[\btheta_n]\enskip \succeq \enskip -\left(\nabla_{\btheta}\mathbb{E}[\btheta_n] \right)^\top \bI(\btheta)^{-1} \nabla_{\btheta}\mathbb{E}[\btheta_n].
\end{equation}
\end{theorem}
\noindent The right-hand-side of~(\ref{eq:CR_bound}) is called the \emph{Cramer-Rao bound} of the estimator, where the middle term is the inverse of the \emph{Fisher information matrix} of the parametric distribution $p(\bx|\btheta)$:
\begin{equation*}
\bI(\btheta) \equals \mathbb{E}\left[\nabla_{\btheta}\log p(\bx|\btheta)\cdot\nabla_{\btheta}^\top\log p(\bx|\btheta)\right]
\end{equation*}
Theorem~\ref{thm:CR_bound} suggests that for an unbiased estimator $\btheta_n$, the inequality over the covariance matrix becomes: $\mbox{Cov}[\btheta_n]\succeq\bI(\btheta)^{-1}, \forall \btheta\in\Omega$.
\begin{definition}[\bfseries Efficiency]
We say that an estimator $\btheta_n$ is \emph{efficient}, iff it attains the Cramer-Rao bound, that is $\mbox{\normalfont Cov}[\btheta_n]$ achieves the lower-bound in~(\ref{eq:CR_bound}) for every $n=1,2,...$~. Furthermore, we say that $\btheta_n$ is \emph{asymptotically efficient}, iff the lower bound is attained asymptotically (when $n\to\infty$).
\end{definition}
\section{Proof of Second-order Multivariate Delta Method}
\label{app:proof_2nd_Delta_Method}
In order to prove this theorem, we have to formulate the statistical Taylor expansion. This, in turn, needs a brief introduction of stochastic order notations.
\subsection{Stochastic Order Notations}
\label{subapp:stochastic_order_notations}
The stochastic order notations are denoted by $o_p$ and $O_p$, where the former is equivalent to convergence in probability (Definition~\ref{defn:convergence_in_prob}) and the latter implies \emph{boundedness in probability}. In what follows, if otherwise stated, $\{\bz_n\}$ is a sequence of multivariate random variables lying in $\Omega\subseteq\mathbb{R}^d$ and $\{a_n\}$ is a sequence of strictly positive real numbers. The skipped proofs can be found in many textbooks on asymptotic theory, such as \citet[Chapter 6]{brockwell1991time}.
\begin{definition}
\label{eq:stochastic_small_o}
We write $\bz_n=o_p(a_n)$ iff
\begin{equation}
\frac{\theta_{in}}{a_n} \equals o_p(1), \quad \mbox{for all } i=1,...,d
\end{equation}
\end{definition}
\begin{definition}
We write $\bz_n=O_p(a_n)$ iff the sequence $\left\{\frac{\theta_{in}}{a_n} \right\}$ is bounded in probability for every $i=1,...,d$, that is, for every $\epsilon>0$ there exists $\delta_\epsilon$ such that
\begin{equation}
P\left(\left|\frac{\theta_{in}}{a_n}\right| > \delta_\epsilon \right) \enskip < \enskip \epsilon, \quad n=1,2,...
\end{equation}
\end{definition}
We also need the following propositions:
\begin{proposition}[\citet{brockwell1991time}]
\label{prop:convergence_types_relationship}
Let $\{\theta_n\}$ and $\{\eta_n\}$ be two sequences of scalar random variables, and $\{a_n\}$ and $\{b_n\}$ be two sequences of positive real numbers. If $\theta_n=O_p(a_n)$ and $\eta_n=o_p(b_n)$, then
\begin{enumerate}[label=(\roman*)]
\item $\theta_n^2 = O_p(a_n^2)$
\item $\theta_n\eta_n=o_p(a_nb_n)$
\end{enumerate}
\end{proposition}
\begin{proposition}
\label{prop:vector_norm_convergence}
The followings are true\footnote{Unless subscripted, $\|\cdot\|$ denotes the $L_2$ norm in all the equations.}:
\begin{enumerate}[label=(\roman*)]
\item\label{item:norm_convergence_op} $\bz_n=o_p(a_n) \enskip \Leftrightarrow \enskip \|\bz_n\|=o_p(a_n)$.
\item\label{item:norm_convergence_Op} $\bz_n=O_p(a_n) \enskip \Leftrightarrow \enskip \|\bz_n\|=O_p(a_n)$.
\end{enumerate}
\end{proposition}
\begin{proof}
The proof of part~\ref{item:norm_convergence_op} can be found in~\citet[Proposition 6.1.2]{brockwell1991time}. Here, we only prove part~\ref{item:norm_convergence_Op}:
\vspace{.25cm}
\noindent $(ii, \Rightarrow)$ : Since $\bz_n=O_p(a_n)$, for every $\varepsilon>0$ and for every $i=1,...,d$, there exists a coefficient $\delta_i>0$ such that
\begin{equation}
P\left(|\theta_{ni}|>a_n\cdot\delta_i\right) \enskip < \enskip \frac{\varepsilon}{d} \quad,n=1,2,....
\end{equation}
Define $\delta_{\mbox{\footnotesize max}}=\max\{\delta_1,...,\delta_d\}$ and note that we can write
\begin{align}
\left\{\bz_n:\enskip \sum_{i=1}^d |\theta_{ni}|^2 > \left(d\cdot a_n\cdot\delta_{\mbox{\footnotesize max}}\right)^2 \right\} &\enskip \subseteq \enskip \left[\bigcap_{i=1}^d \left\{\bz_n:\enskip |\theta_{ni}|\leq a_n\cdot\delta_{\mbox{\footnotesize max}}\right\}\right]^c \nonumber\\
&\equals \bigcup_{i=1}^d \left\{\bz_n:\enskip |\theta_{ni}|>a_n\cdot\delta_{\mbox{\footnotesize max}}\right\}
\end{align}
implying that
\begin{align}
P\left(\|\bz_n\|^2 > \left(d\cdot a_n\cdot\delta_{\mbox{\footnotesize max}}\right)^2 \right) \enskip &\leq \enskip P\left( \bigcup_{i=1}^d \left\{\bz_n:\enskip |\theta_{ni}|>a_n\cdot\delta_{\mbox{\footnotesize max}}\right\} \right) \nonumber\\
&\leq \enskip \sum_{i=1}^dP(|\theta_{ni}|>a_n\cdot\delta_{\mbox{\footnotesize max}})
\end{align}
Furthermore, for every $i=1,...,d$ we have $\delta_{\mbox{\footnotesize max}}\geq\delta_i$, consequently the interval $(a_n\delta_{\mbox{\footnotesize max}},\infty)$ is a subset of $(a_n\delta_i,\infty)$ and $P(|\theta_{ni}|>a_n\delta_{\mbox{\footnotesize max}}) \leq P(|\theta_{ni}|>a_n\delta_i)$. This implies that
\begin{equation}
P\left(\|\bz_n\|^2 > \left(d\cdot a_n\cdot\delta_{\mbox{\footnotesize max}}\right)^2 \right) \enskip \leq \enskip \sum_{i=1}^d P(|\theta_{ni}|>a_n\cdot\delta_i) \enskip < \enskip \varepsilon.
\end{equation}
Therefore, for every $\varepsilon>0$, we can choose $\delta_\varepsilon = d\cdot\delta_{\mbox{\footnotesize max}}$ such that $P\left(\frac{\|\bz_n\|}{a_n}>\delta_\varepsilon\right)<\varepsilon$ for every $n=1,2,...$, that is $\|\bz_n\|=O_p(a_n)$.
\noindent $(ii, \Leftarrow)$ : Suppose $\|\bz_n\|=O_p(a_n)$, that is for every $\varepsilon>0$ we can find $\delta_\varepsilon>0$ such that
\begin{equation}
P(\|\bz_n\|>a_n\cdot\delta_\varepsilon) \enskip < \varepsilon \quad, n=1,2,...
\end{equation}
It is clear that for any given $i\in\{1,...,d\}$ we have
\begin{equation}
\label{eq:component_norm_sets}
\left\{\bz_n: |\theta_{ni}|>a_n\cdot\delta_\varepsilon \right\} \enskip \subseteq \enskip \left\{\bz_n: \|\bz_n\|>a_n\cdot\delta_\varepsilon \right\}
\end{equation}
hence
\begin{equation}
P(|\theta_{ni}|>a_n\cdot\delta_\varepsilon)\enskip \leq \enskip P(\|\bz_n\|>a_n\cdot\delta_\varepsilon) \enskip < \epsilon \quad, n=1,2,...
\end{equation}
meaning that $\theta_{ni} = O_p(a_n), i=1,...,d$ or equivalently $\bz_n = O_p(a_n)$.
\end{proof}
\begin{proposition}
\label{prop:Op_convergent_op}
If $\bz_n=O_p(a_n)$ and $a_n\to0 (n\to\infty)$, then $\bz_n=o_p(1)$.
\end{proposition}
\begin{proof}
The goal is to show $\bz_n=o_p(1)$ or equivalently $\|\bz_n\|=o_p(1)$ by proving that $P(\|\bz_n\|>\varepsilon)\to0 (n\to\infty)$ for every $\varepsilon>0$. Fix $\varepsilon$ to a positive real number. In order to have the sequence of probability numbers $\{P(\|\bz_n\|>\varepsilon)\}$ converging to zero, for every $\varepsilon_0>0$ there should exist a positive integer $N>0$ such that
\begin{equation}
\label{eq:convergent_to_zero}
P(\|\bz_n\|>\varepsilon)<\varepsilon_0 \quad \forall n>N.
\end{equation}
Because of the assumption of being bounded by $a_n$, that is $\bz_n=O_p(a_n)$ or equivalently $\|\bz_n\|=O_p(a_n)$, we can choose a real number $\delta_0>0$ such that
\begin{equation}
\label{eq:bounded_by_an}
P(\|\bz_n\|>a_n\delta_0)\enskip < \enskip \varepsilon_0 \quad n=1,2,...
\end{equation}
On the other hand, since $a_n\to0(n\to\infty)$, there exists a large enough number $N_0>0$ such that $0<a_n<\frac{\varepsilon}{\delta_0}$ for all $n>N_0$. Therefore we get:
\begin{equation}
[0,a_n\delta_0] \enskip \subseteq \enskip [0,\varepsilon] \quad \forall n>N_0
\end{equation}
implying that
\begin{equation}
\label{eq:N0_inequality}
P(\|\bz_n\|\leq a_n\delta_0) \enskip \leq \enskip P(\|\bz_n\|\leq\varepsilon) \quad \forall n>N_0
\end{equation}
From inequalities~(\ref{eq:bounded_by_an}) and~(\ref{eq:N0_inequality}), and noticing that the latter holds for all $n$ whereas the former is satisfied when $n>N_0$, one can write:
\begin{equation}
P(\|\bz_n\|>\varepsilon) \enskip \leq \enskip P(\|\bz_n\|>a_n\delta_0)\enskip<\enskip\varepsilon_0 \quad \forall n>N_0
\end{equation}
Therefore, for every $\varepsilon_0>0$, equation~(\ref{eq:convergent_to_zero}) is guaranteed if $N$ is chosen to be equal to $N_0$ so that inequality~(\ref{eq:N0_inequality}) is satisfied. Similarly, this can be written for every $\varepsilon>0$, thus the proof is complete.
\end{proof}
\begin{proposition}[\citet{serfling2009approximation}, Chapter 1]
\label{prop:boundedness_convergence_in_law}
Let $\{\bz_n\}$ be a sequence of random variables. If there exists a random variable $\bz_0$ such that $\bz_n\overset{L}{\to}\bz_0$, then $\bz_n=O_p(1)$.
\end{proposition}
\subsection{Second-order Statistical Taylor Expansion}
\label{subapp:2nd_statisticaly_Taylor}
Now we are ready to establish the second-order statistical Taylor expansion.
\begin{theorem}
\label{thm:taylor_in_probability_2nd}
Let $\{\bz_n\}$ be a sequence of random vectors in a convex and compact set $\Omega\subseteq \mathbb{R}^d$ and $\bz_0\in \Omega$ be a constant vector such that $\bz_n - \bz_0 = O_p(a_n)$ where $a_n\to0 (n\to\infty)$. If $g:\Omega\to\mathbb{R}$ is a $\calc^3$ function , then
\begin{equation}
g(\bz_n) \equals g(\bz_0) \enplus \nabla_{\bz}^\top g(\bz_0)(\bz_n-\bz_0) \enplus \frac{1}{2}(\bz_n-\bz_0)^\top\nabla_{\bz}^2g(\bz_0)(\bz_n-\bz_0) \enplus o_p(a_n^2).
\end{equation}
\end{theorem}
\begin{proof}
Since $g$ is twice continuously differentiable in a neighborhood of $\bz_0$, it can be written in terms of the Taylor expansion as
\begin{equation}
g(\bz) \equals g(\bz_0) \enplus (\bz-\bz_0)^\top\nabla_{\bz}g(\bz_0) \enplus \frac{1}{2}(\bz-\bz_0)^\top\nabla_{\bz}^2g(\bz_0)(\bz-\bz_0) \enplus r_2(\bz,\bz_0)
\end{equation}
where $r_2(\bz,\bz_0)$ is the Lagrange remainder of second order. Based on Taylor's polynomial theorem for multivariate functions, there exists a number $t\in[0,1]$ such that $\bz^*=t\bz + (1-t)\bz_0\in\Omega$ (due to convexity of $\Omega$) and
\begin{equation}
r_2(\bz,\bz_0) \equals \frac{1}{6}\sum_{1\leq i,j,k\leq d} \frac{\partial^3 g(\bz^*)}{\partial \theta_i\partial \theta_j\partial \theta_k}(\theta_i-\theta_{0i})(\theta_j-\theta_{0j})(\theta_k-\theta_{0k})
\end{equation}
But since $\Omega$ is compact and $g\in\calc^3$, the third derivative of $g$ is bounded\footnote{This is because of the following Theorem in real analysis:
\begin{theorem}
Let $X$ and $Y$ be two vector spaces. If $g:X\to Y$ is continuous and $X$ is compact, then $f(X)$ is compact in $Y$.
\end{theorem}
In special case of this theorem, when $Y=\mathbb{R}$, compactness of $f(X)$ is equivalent to boundedness and closedness.} and therefore there exists $M>0$ such that
\begin{equation}
\left|\frac{\partial^3 g(\bz)}{\partial \theta_i\partial \theta_j\partial \theta_k}\right| \enskip \leq \enskip M \quad,\forall \bz\in \Omega\enskip,\forall i,j,k\in\{1,...,d\}
\end{equation}
Hence the Lagrange remainder can be bounded by
\begin{align}
\label{eq:bounded_lagrange_remainder}
|r_2(\bz,\bz_0)| &\enskip \leq \enskip \frac{M}{6}\sum_{1\leq i,j,k\leq3} |\theta_i-\theta_{0i}|\cdot|\theta_j-\theta_{0j}|\cdot|\theta_k-\theta_{0k}| \nonumber\\
&\equals \frac{M}{6}\|\bz-\bz_0\|_1^3\nonumber\\
&\enskip\leq\enskip \frac{M'}{6}\|\bz-\bz_0\|^3
\end{align}
where $M'=c_uM$ and $c_u$ is obtained from the equivalence of norms in $\mathbb{R}^d$ \footnote{Two norm functions $\|\cdot\|_{(1)}$ and $\|\cdot\|_{(2)}$, in a vector space $\Omega$, are called \emph{equivalent} iff there exist constants $c_u\geq c_d>0$ such that
\begin{equation}
c_d\|\bz\|_{(2)} \enskip \leq \enskip \|\bz\|_{(1)} \enskip \leq \enskip c_u\|\bz\|_{(2)} \quad,\forall \bz\in \Omega.
\end{equation}}. Now define the function $h:\Omega\to\mathbb{R}$ as below
\begin{equation}
\label{eq:definition_hz}
h(\bz) \enskip := \enskip \left\{\begin{array}{ll}
\displaystyle \frac{r_2(\bz,\bz_0)}{\|\bz-\bz_0\|^2/2} & ,\bz\neq\bz_0 \\[.25cm]
0 & ,\bz=\bz_0
\end{array}\right.
\end{equation}
Note that $h(\bz)$ is continuous at $\bz=\bz_0$: due to boundedness of $r_2(\bz,\bz_0)$, $h(\bz)$ is also bounded by
\begin{equation}
|h(\bz)| \enskip \leq\enskip \frac{M'}{3}\|\bz-\bz_0\|.
\end{equation}
Hence, for every $\varepsilon>0$, we can select $\delta_\varepsilon=\frac{3\varepsilon}{M'}$ such that the following continuity condition holds
\begin{equation}
\|\bz-\bz_0\|<\delta_\varepsilon \enskip \Rightarrow\enskip |h(\bz)| \enskip \leq \enskip \varepsilon.
\end{equation}
Continuity of $h(\bz)$ at $\bz=\bz_0$ implies $\lim_{\bz\to\bz_0}h(\bz)=h(\bz_0)=0$. Furthermore, since $\bz_n-\bz_0=O_p(a_n)$ and $a_n\to0(n\to\infty)$, Proposition~\ref{prop:Op_convergent_op} suggests that $\bz_n-\bz_0=o_p(1)$. These two enable us to use Proposition~\ref{prop:convergent_continuous_mappings} and write
\begin{equation}
h(\bz_n) - h(\bz_0) \equals h(\bz_n) \equals o_p(1).
\end{equation}
Finally, from equation~(\ref{eq:definition_hz}) and Propositions~\ref{prop:convergence_types_relationship},~\ref{prop:vector_norm_convergence} and~\ref{prop:Op_convergent_op}, we can write that
\begin{equation}
r_2(\bz_n,\bz_0) \equals h(\bz_n)\cdot\frac{\|\bz_n-\bz_0\|^2}{2}\equals o_p(1)\cdot O_p(a_n^2)\equals o_p(a_n^2)
\end{equation}
\end{proof}
\subsection{Second-order Multivariate Delta Method}
\label{subapp:2nd_multivariate_delta_method}
Finally, here is the proof of second-order multivariate Delta method (Theorem~\ref{thm:delta_method_zeroder}):
\begin{proof}
From assumption of the Theorem, $\sqrt{n}(\bz_n-\bz_0)\overset{L}{\to}\caln(\mathbf{0},\bSigma)$, and Proposition~\ref{prop:boundedness_convergence_in_law}, one conclude that $\sqrt{n}(\bz_n-\bz_0)=O_p(1)$ and therefore $\bz_n-\bz_0 = O_p\left(\frac{1}{\sqrt{n}}\right)$. Thus we can use Theorem~\ref{thm:taylor_in_probability_2nd} with $a_n=\frac{1}{\sqrt{n}}$ to write:
\begin{equation}
g(\bz) \equals g(\bz_0) \enplus (\bz-\bz_0)^\top\nabla_{\bz}g(\bz_0) \enplus \frac{1}{2}(\bz-\bz_0)^\top\nabla_{\bz}^2g(\bz_0)(\bz-\bz_0) \enplus o_p\left(\frac{1}{n}\right),
\end{equation}
hence
\begin{align}
n\bigg[g(\bz) - g(\bz_0)\bigg] &\equals \frac{1}{2}\left[\sqrt{n}\cdot(\bz-\bz_0)\right]^\top\nabla_{\bz}^2g(\bz_0)\left[\sqrt{n}\cdot(\bz-\bz_0)\right] \enplus o_p(1) \nonumber\\
&\enskip \overset{L}{\to} \enskip \frac{1}{2}\caln(\mathbf{0},\bSigma)^\top \nabla_{\bz}^2g(\bz_0) \caln(\mathbf{0},\bSigma)\nonumber\\
&\equals \frac{1}{2}\caln(\mathbf{0},\mathbb{I}_d)^\top \left[\bSigma^{1/2}\nabla_{\bz}^2g(\bz_0)\bSigma^{1/2} \right]\caln(\mathbf{0},\mathbb{I}_d)
\end{align}
Define $\boldsymbol{\Gamma} := \bSigma^{1/2}\nabla_{\bz}^2g(\bz_0)\bSigma^{1/2}$ and rewrite the right-hand-side element-wise as
\begin{align}
\frac{1}{2}\caln(\mathbf{0},\mathbb{I}_d)^\top \boldsymbol{\Gamma} \caln(\mathbf{0},\mathbb{I}_d) \equals \frac{1}{2}\sum_{i=1}^d \lambda_{i} \caln(0,1)^2 \equals \frac{1}{2}\sum_{i=1}^d\lambda_{i}\chi_1^2,
\end{align}
where $\lambda_i$'s are eigenvalues of $\boldsymbol{\Gamma}$. Finally, noting that the terms in the Chi-square mixture are independent, variance of the convergent random variable can be easily computed as
\begin{align}
\label{eq:asymptotic_variance_zeroder}
\mbox{Var}\left[\frac{1}{2}\sum_{i=1}^d \lambda_{i} \chi_1^2 \right] &\equals \frac{1}{4}\sum_{i=1}^d\lambda_{i}^2\cdot\mbox{Var}\left[\chi_1^2\right]\nonumber\\
&\equals \frac{1}{2}\sum_{i=1}^d\lambda_{i}^2 \nonumber\\
&\equals \frac{1}{2}\left\|\bSigma^{1/2}\nabla_{\bx}^2g(\bx_0)\bSigma^{1/2} \right\|_F^2,
\end{align}
\end{proof}
\section{Proof of Lemma~\ref{lemma:equivalent_set_optimization}}
\label{app:proof_Lemma}
We first substitute the score function of the classifier
\begin{equation}
\nabla_{\btheta} \log p(y|\bx,\btheta) \equals \frac{\nabla_{\btheta} p(y|\bx,\btheta)}{p(y|\bx,\btheta)}\nonumber\\
\end{equation}
into formulation Monte-Carlo approximation of $\bI_q$ to get:
\begin{align}
\hat{\bI}(\btheta;X_q) &\equals \frac{1}{|X_q|}\sum_{\bx\in X_q}\sum_{y=1}^cp(y|\bx,\btheta)\cdot\frac{\nabla_{\btheta} p(y|\bx,\btheta)\nabla_{\btheta}^\top p(y|\bx,\btheta)}{p(y|\bx,\btheta)^2} \enplus \delta\mathbb{I}_d \\
&\equals \frac{1}{|X_q|}\sum_{\bx\in X_q} \sum_{y=1}^c\frac{\nabla_{\btheta}p(y|\bx,\btheta).\nabla_{\btheta}^\top p(y|\bx,\btheta)}{p(y|\bx,\btheta)} \enplus \delta\mathbb{I}_d
\end{align}
Define the vector $\bv_{\btheta}(\bx,y):=\nabla_{\btheta}p(y|\bx,\btheta)\bigg/\sqrt{p(y|\bx,\btheta)}$ and rewrite $\hat{\bI}(\btheta;X_q)$ as:
\begin{equation}
\label{eq:Iq_v}
\hat{\bI}(\btheta;X_q) \equals \frac{1}{|X_q|}\sum_{\bx\in X_q} \sum_{y=1}^c \bv_{\btheta}(\bx,y).\bv_{\btheta}(\bx,y)^\top +\delta\cdot\mathbb{I}_d.
\end{equation}
On the other hand, since $X_q\subset X_p$ we can write $\hat{\bI}(\btheta;X_p)$ in terms of $\hat{\bI}(\btheta;X_q)$ by breaking the summation over $X_p$ into summations over $X_q$ and $X_p-X_q$ as follows:
\begin{align}
\hat{\bI}(\btheta;X_p) &\equals \frac{|X_q|}{|X_p|}\left[\frac{1}{|X_q|}\sum_{\bx\in X_q} \sum_{y=1}^c \bv_{\btheta}(\bx,y).\bv_{\btheta}(\bx,y)^\top +\delta\cdot\mathbb{I}_d \right] \nonumber\\
&\phantom{\equals} \enplus \frac{1}{|X_p|}\sum_{\bx\in X_p-X_q} \sum_{y=1}^c \bv_{\btheta}(\bx,y).\bv_{\btheta}(\bx,y)^\top \enplus \delta\left(\frac{|X_p|-|X_q|}{|X_p|}\right)\cdot \mathbb{I}_d \nonumber\\
&\equals \left(\frac{|X_q|}{|X_p|}\right)\cdot\hat{\bI}(\btheta;X_q) \enplus \frac{1}{|X_p|}\sum_{\bx\in X_p-X_q} \sum_{y=1}^c \bv_{\btheta}(\bx,y).\bv_{\btheta}(\bx,y)^\top\nonumber\\
&\phantom{\equals} \enplus \delta\left(\frac{|X_p|-|X_q|}{|X_p|}\right)\cdot\mathbb{I}_d
\end{align}
Now that we related the Fisher information matrices to each other, we can compute the product of $\hat{\bI}(\btheta;X_p)$ and $\hat{\bI}(\btheta;X_q)^{-1}$:
\begin{align}
\hat{\bI}(\btheta;X_q)^{-1}\hat{\bI}(\btheta;X_p) &\equals \left(\frac{|X_q|}{|X_p|}\right)\cdot\mathbb{I}_d \enplus \frac{\hat{\bI}(\btheta;X_q)^{-1}}{|X_p|}\left[\sum_{\bx\in X_p-X_q} \sum_{y=1}^c \bv_{\btheta}(\bx,y)\cdot\bv_{\btheta}(\bx,y)^\top \right] \nonumber\\
&\phantom{\equals} \enplus \delta\left(\frac{|X_p|-|X_q|}{|X_p|}\right)\cdot\hat{\bI}(\btheta;X_q)^{-1}
\end{align}
Applying the trace function to both sides of the equation will result:
\begin{align}
\label{eq:objective_wtrace}
\mbox{tr}\left[\hat{\bI}(\btheta;X_q)^{-1}\hat{\bI}(\btheta;X_p)\right] &\equals \frac{|X_q|\cdot d}{|X_p|} + \frac{1}{|X_p|}\sum_{\bx\in X_p-X_q} \sum_{y=1}^c \mbox{tr}\left[\hat{\bI}(\btheta;X_q)^{-1} \bv_{\btheta}(\bx,y)\cdot\bv_{\btheta}(\bx,y)^\top\right]\nonumber\\
&\phantom{\equals} \enplus \delta\left(\frac{|X_p|-|X_q|}{|X_p|}\right)\cdot\mbox{tr}\left[\hat{\bI}(\btheta;X_q)^{-1}\right] \nonumber\\
&\enskip \approx\enskip \frac{|X_q|\cdot d}{|X_p|} \enplus \frac{1}{|X_p|}\sum_{\bx\in X_p-X_q} \sum_{y=1}^c \bv_{\btheta}(\bx,y)^\top\hat{\bI}(\btheta;X_q)^{-1} \bv_{\btheta}(\bx,y),
\end{align}
where the last term is dropped since the overloading constant, $\delta$, is assumed to be small. Furthermore, the term including $\hat{\bI}(\btheta;X_q)^{-1}$ can be approximated by replacing the weighted harmonic mean of the eigenvalues of $\hat{\bI}(\btheta;X_q)$ by their weighted arithmetic mean~\citep{Hoi2006}:
\begin{equation}
\label{eq:inv_Iq_approx}
\bv_{\btheta}(\bx,y)^\top\hat{\bI}(\btheta;X_q)^{-1} \bv_{\btheta}(\bx,y) \enskip \approx \enskip \frac{\|\bv_{\btheta}(\bx,y)\|^4}{\bv_{\btheta}(\bx,y)^\top\hat{\bI}(\btheta;X_q)\bv_{\btheta}(\bx,y)}.
\end{equation}
Note that this approximation becomes exact when the condition number of $\hat{\bI}(\btheta;X_q)$ is one. Substituting $\hat{\bI}(\btheta;X_q)$ from equation~(\ref{eq:Iq_v}) into the denominator of the approximation above yields:
\begin{equation}
\bv_{\btheta}(\bx,y)^\top\hat{\bI}(\btheta;X_q)\bv_{\btheta}(\bx,y) \equals \frac{1}{|X_q|}\sum_{\bx'\in X_q}\sum_{y'=1}^c\left[\bv_{\btheta}(\bx,y)^\top \bv_{\btheta}(\bx',y') \right]^2 \enplus \delta\|\bv_{\btheta}(\bx,y)\|^2
\end{equation}
Integrating this approximation with equation~(\ref{eq:objective_wtrace}), and assuming that the value of $\btheta$ is not located at the stationary point of the conditional density $p(y|\bx,\btheta)$ (hence $\bv_{\btheta}(\bx,y)$ is not the zero vector), results:
\begin{align}
\label{eq:objective_wtrace_simp}
\mbox{tr}&\left[\hat{\bI}(\btheta;X_q)^{-1}\hat{\bI}(\btheta;X_p)\right] \enskip\approx\enskip \frac{|X_q|\cdot d}{|X_p|} \nonumber\\
&+\enskip \frac{1}{|X_p|}\sum_{\bx\in X_p-X_q}\sum_{y=1}^c\frac{1}{\delta\cdot\|\bv_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')}
\end{align}
where
\begin{equation}
\label{eq:defn_g}
g_{\btheta}(\bx,y,\bx') \enskip := \enskip \frac{1}{|X_q|}\sum_{y'=1}^c \left[\frac{\bv_{\btheta}(\bx,y)^\top\bv_{\btheta}(\bx',y')}{\|\bv_{\btheta}(\bx,y)\|^2}\right]^2
\end{equation}
Finally in~(\ref{eq:objective_wtrace_simp}), removing the constants we get
\begin{align}
\label{eq:set_maximization}
\argmin_{X_q\subset X_p} \enskip &\mbox{tr}\left[\hat{\bI}(\btheta;X_q)^{-1}\hat{\bI}(\btheta;X_p)\right] \nonumber\\
&\approx \enskip \operatorname*{arg\,max}_{X_q\subset X_p} \sum_{\bx\in X_p-X_q}\sum_{y=1}^c\frac{-1}{\delta\cdot\|\bv_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')}
\end{align}
\section{Proof of Theorem~\ref{thm:submodularity}}
\label{app:proof_submodularity}
Proof of this Theorem is a generalization of the discussion by~\citet{Hoi2006}, with clarification of all the assumptions and approximations made.
First, note that the function $f_{\btheta}$ can be broken into simpler terms $f_{\btheta}(X_q)=\sum_{y=1}^cf_{\btheta}(X_q;y)$, where
\begin{equation}
f_{\btheta}(X_q;y) \equals \sum_{\bx\in X_p-X_q}\frac{-1}{\delta\cdot\|v_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')},\quad \forall X_q\subseteq X_p.
\end{equation}
Therefore, in order to prove submodularity and monotonicity of $f_{\btheta}$, it suffices to prove these properties for $f_{\btheta}(\cdot;y)$ for all $y\in\{1,...,c\}$.
Fix $y$ and take any subset $X_q\subseteq X_p$ and $\bxi\in X_p-X_q$. Then, we can write:
\begin{eqnarray}
f_{\btheta}(X_q\cup\{\bxi\};y) &\equals &\sum_{\bx\in X_p-(X_q\cup\{\bxi\})}\frac{-1}{\delta\cdot\|v_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')} \\
&\equals &\sum_{\bx\in X_p-X_q}\frac{-1}{\delta\cdot\|v_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')} \nonumber\\
& &+\enskip \frac{1}{\delta\cdot\|v_{\btheta}(\bxi,y)\|^{-2} + \sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')}. \nonumber
\end{eqnarray}
We then form the discrete derivative of $f_{\btheta}(\cdot;y)$ at $X_q$ to get:
\begin{align}
\label{eq:discrete_derivative_fy}
&\rho_{f_{\btheta}(\cdot;y)}(X_q;\bxi) \equals f_{\btheta}(X_q\cup\{\bxi\};y) \enskip-\enskip f_{\btheta}(X_q;y) \nonumber\\
&\equals \sum_{\bx\in X_p-X_q}\left[ \frac{-1}{\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}} + \sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')} + \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}} + \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')} \right] \nonumber\\
&\phantom{\equals} \enplus \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}} + \sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')}.
\end{align}
The right-hand-side can be rewritten as
\begin{align}
\label{eq:discrete_derivative_fy_re}
&\sum_{\bx\in X_p-X_q}\left[ \frac{g_{\btheta}(\bx,y,\bxi)}{\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')\right)\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')\right)} \right]\nonumber\\
&\phantom{\equals} \enplus \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')}.
\end{align}
Since by definition $g_{\btheta}(\bx,y,\bx')\geq0,\forall\bx,y,\bx'$, all of the terms in~(\ref{eq:discrete_derivative_fy_re}) are non-negative and therefore $\rho_{f_{\btheta}(\cdot;y)}(X_q;\bxi)\geq0$. This is true for any $X_q\subseteq X_p$ hence monotonicity of $f_{\btheta}(\cdot;y)$ is obtained.
Now let us take any superset $X_{q'}$ such that $X_{q}\subseteq X_{q'}\subseteq X_p$ and $\bxi\in X_p-X_{q'}$, and form the difference between their corresponding discrete derivatives. From~(\ref{eq:discrete_derivative_fy_re}) we will have:
\begin{align}
\label{eq:discrete_derivative_diff}
& \rho_{f_{\btheta}(\cdot;y)}(X_q;\bxi) - \rho_{f_{\btheta}(\cdot;y)}(X_{q'};\bxi) \nonumber\\
&= \sum_{\bx\in X_p-X_q}\left[ \frac{g_{\btheta}(\bx,y,\bxi)}{\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')\right)\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')\right)} \right] \nonumber\\
& + \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bxi,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')} \nonumber \\
&- \sum_{\bx\in X_p-X_{q'}}\left[ \frac{g_{\btheta}(\bx,y,\bxi)}{\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_{q'}\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')\right)\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_{q'}} g_{\btheta}(\bx,y,\bx')\right)} \right] \nonumber\\
& - \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bxi,y)\|^{2}}+\sum_{\bx'\in X_{q'}\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')}.
\end{align}
From non-negativity of $g_{\btheta}$ and that $X_q\subseteq X_{q'}$, we can conclude that for any $\bx\in X$ and $y\in\{1,...,c\}$:
\begin{align}
\label{eq:inequality_q}
\sum_{\bx'\in X_{q'}} g_{\btheta}(\bx,y,\bx') \enskip &\geq \enskip \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx') \nonumber \\
\Leftrightarrow \quad \left[\sum_{\bx'\in X_{q'}} g_{\btheta}(\bx,y,\bx')+\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}\right]^{-1} \enskip &\leq \enskip \left[\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')+\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}\right]^{-1} \nonumber\\
\Leftrightarrow \quad -\left[\sum_{\bx'\in X_{q'}} g_{\btheta}(\bx,y,\bx')+\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}\right]^{-1} \enskip &\geq \enskip -\left[\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')+\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}\right]^{-1}
\end{align}
Similarly, since $X_q\cup\{\bxi\} \subseteq X_{q'}\cup\{\bxi\}$ we will get:
\begin{equation}
\label{eq:inequality_q_bxi}
\phantom{\Rightarrow \quad }-\left[\sum_{\bx'\in X_{q'}\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')+\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}\right]^{-1} \geq -\left[\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')+\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}\right]^{-1}
\end{equation}
Applying the inequalities~(\ref{eq:inequality_q}) and~(\ref{eq:inequality_q_bxi}) into euqation~(\ref{eq:discrete_derivative_diff}) results:
\begin{align}
\label{eq:ineq_delta_rho}
& \rho_{f_{\btheta}(\cdot;y)}(X_q;\bxi) - \rho_{f_{\btheta}(\cdot;y)}(X_{q'};\bxi)\nonumber \\
&\geq \sum_{\bx\in X_p-X_q}\left[ \frac{g_{\btheta}(\bx,y,\bxi)}{\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')\right)\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')\right)} \right] \nonumber \\
&- \sum_{\bx\in X_p-X_{q'}}\left[ \frac{g_{\btheta}(\bx,y,\bxi)}{\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')\right)\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')\right)} \right] \nonumber \\
& + \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bxi,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')} \enskip-\enskip \frac{1}{\frac{\delta}{\|\bv_{\btheta}(\bxi,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bxi,y,\bx')}
\end{align}
which yields\footnote{The inequality in~(\ref{eq:ineq_delta_rho}) is obtained by the fact that, for every four positive real numbers $a$, $a_0$, $b$ and $b_0$, if we have $-a\geq-a_0$ and $-b\geq-b_0$ (similar to~(\ref{eq:inequality_q}) and~(\ref{eq:inequality_q_bxi})), then $$-a\cdot b\equals(-a)\cdot b\enskip\geq\enskip(-a_0)\cdot b\equals a_0\cdot(-b)\enskip\geq\enskip a_0\cdot(-b_0)\equals-a_0\cdot b_0.$$}
\begin{equation}
\label{eq:submodular_inequality}
\sum_{\bx\in X_{q'}-X_q} \frac{g_{\btheta}(\bx,y,\bxi)}{\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q\cup\{\bxi\}} g_{\btheta}(\bx,y,\bx')\right)\left(\frac{\delta}{\|\bv_{\btheta}(\bx,y)\|^{2}}+\sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')\right)}\enskip \geq \enskip 0.
\end{equation}
Inequality~(\ref{eq:submodular_inequality}) holds for any $X_q\subseteq X_p$; hence submodularity of $f_{\btheta}(\cdot;y)$ stands for all $y\in\{1,...,c\}$ and $\btheta\in\Omega$.
\section{The Framework and Assumptions}
\label{sec:framework_assumptions}
In this paper, we deal with classification problems, where each covariate, represented by a feature vector $\bx$ in vector space $X$, is associated with a numerical class label $y$. Assuming that there are $1<c<\infty$ classes, $y$ can take any integer among the set $\{1,...,c\}$. Suppose that the pairs $(\bx,y)$ are distributed according to a parametric joint distribution $p(\bx,y|\btheta)$, with the parameter space denoted by $\Omega\subseteq\mathbb{R}^d$. Using a set of observed pairs as the training data, $\call_n:=\{(\bx_1,y_1),...,(\bx_n,y_n)\}$, we can estimate $\btheta$ and predict the class labels of the unseen test samples, e.g. by maximizing $p(y|\bx,\btheta)$. In active learning, the algorithm is permitted to take part in designing $\call_n$ by choosing a set of data points $\{\bx_1,...,\bx_n\}$, for which the class labels are then generated using an external oracle.
In addition to the framework described in the last section (see subproblems (i) to (iii)), we make the following assumptions regarding the oracle, our classification model and the underlying data distribution:
\begin{enumerate}
\setcounter{enumi}{-1}
\item\label{item:parameter_dependence} The dependence of the joint distribution to the parameter $\btheta$ comes only from the class-conditional distribution and the marginal distribution does not depend on $\btheta$, that is:
\begin{equation}
\label{eq:type2_distribution}
p(\bx,y|\btheta) \equals p(y|\bx,\btheta) p(\bx).
\end{equation}
\citet{Zhang2000} referred to joint distributions with such parameter dependence as type-II models, as opposed to type-I models which have parameter dependence in both class conditionals and marginal. They argue that active learning is more suitable for type-II models. Moreover, maximizing the joint with respect to the parameter vector in this model, becomes equivalent to maximizing the posterior $p(y|\bx,\btheta)$ (inference step in sub-problem (ii)).
\item\label{item:identifiability} \textit{(Identifiability):} The joint distribution $P_{\btheta}$ (whose density is given by $p(\bx,y|\btheta)$) is identifiable for different parameters. Meaning that for every distinct parameter vectors $\btheta_1$ and $\btheta_2$ in $\Omega$, $P_{\btheta_1}$ and $P_{\btheta_2}$ are also distinct. That is
$$\forall \btheta_1\neq\btheta_2\in\Omega \enskip \exists A\subseteq X\times\{1,...,c\} \quad \mbox{s.t.} \quad P_{\btheta_1}(A)\neq P_{\btheta_2}(A).$$
\item\label{item:support} The joint distribution $P_{\btheta}$ has common support for all $\btheta\in\Omega$.
\item\label{item:oracle_distribution} \textit{(Model Faithfulness): } For any $\bx\in X$, we have access to an oracle that generates a label $y$ according to the conditional $p(y|\bx,\btheta_0)$. That is, the posterior parametric model matches the oracle distribution. We call $\btheta_0$ the true model parameter.
\item\label{item:iid} \textit{(Training joint): } The set of observations in $\call_n:=\{(\bx_1,y_1),...,(\bx_n,y_n)\}$ are drawn independently from the \emph{training}/\emph{proposal}/\emph{query} joint distribution of the form $p(y|\bx,\btheta_0)q(\bx)$ where $q$ is the training marginal with no dependence on the parameter.
\item\label{item:test_iid} \textit{(Test joint): } The unseen test pairs are distributed according to the \emph{test}/\emph{true} joint distribution of the form $p(y|\bx,\btheta_0)p(\bx)$ where $p$ is the test marginal with no dependence on the parameter.
\item\label{item:differentiability} \textit{(Differentiability): } The log-conditional $\log p(y|\bx,\btheta)$ is of class $\calc^3(\Omega)$ as a function of $\btheta$ and for all $(\bx,y)\in X\times\{1,...,c\}$ \footnote{We say that a function $f:X\to Y$ is of $\mathcal{C}^p(X)$, for an integer $p>0$, if its derivatives up to $p$-th order exist and are continuous at all points of $X$.}.
\item\label{item:openness}
The parameter space $\Omega$ is compact and there exists an open ball around the true parameter of the model $\btheta_0\in\Omega$.
\item\label{item:FisherInformation} \textit{(Invertibility): } The Fisher information matrix (reviewed in section~\ref{subsec:fisher_information}) of the joint distribution is positive definite and therefore invertible for all $\btheta\in\Omega$, and for any type of marginal that is used under assumption \ref{item:parameter_dependence}.
\end{enumerate}
\noindent Regarding assumptions \ref{item:iid} and \ref{item:test_iid}, note that the training and test marginals are not necessarily equal. The test marginal is usually not known beforehand and $q$ cannot be set equal to $p$ in practice, hence $q$ can be viewed as a proposal distribution. Such inconsistency is what \citet{shimodaira2000improving} called \emph{covariate shift in distribution}. In the remaining sections of the report, we use subscripts $p$ and $q$ for the statistical operators that consider $p(\bx)$ and $q(\bx)$ as the marginal in the joint distribution, respectively. We explicitly mention $\bx$ as the input argument in order to refer to marginal operators. For instance, $\mathbb{E}_q$ denotes the joint expectation with respect to $q(\bx)p(y|\btheta,\bx)$, whereas $\mathbb{E}_{q(\bx)}$ denotes the marginal expectation with respect to $q(\bx)$.
\section{Background}
\label{sec:background}
Here, we provide a short review of maximum likelihood estimation (MLE) as our inference method, and briefly introduce Fisher information of a parametric distribution. These two basic concepts enable us to explain some of the key properties of MLE, upon which our further analysis of FIR objective relies. Note that our focus in this section is on sub-problem~\ref{subproblem:passive_learning} with the assumptions listed above.
\subsection{Maximum Likelihood Estimation}
\label{subsec:MLE}
In this section, we review maximum likelihood estimation in the context of classification problem.
Given a training data set $\call_n=\{(\bx_1,y_1),...,(\bx_n,y_n)\}$, a maximum likelihood estimate (MLE) is obtained by maximizing the log-likelihood function over all pairs inside $\call_n$, with respect to the parameter $\btheta$:
\begin{align}
\label{eq:MLE}
\hat{\btheta}_n \equals \operatorname*{arg\,max}_{\btheta} \enskip \log p\left(\call_n|\btheta \right),
\end{align}
Under the assumptions \ref{item:parameter_dependence} and \ref{item:iid}, the optimization in~(\ref{eq:MLE}) can be written as
\begin{align}
\label{eq:MLE_conditional}
\hat{\btheta}_n \equals \operatorname*{arg\,max}_{\btheta} \enskip \sum_{i=1}^n \log p(y_i|\bx_i,\btheta),
\end{align}
Equation~(\ref{eq:MLE_conditional}) shows that MLE does not depend on the marginal when using type-II model. Hence, in our analysis we focus on the conditional log-likelihood as the classification objective, and simply call it the log-likelihood function when viewed as a function of the parameter vector $\btheta$, for any given pair $(\bx,y)\in X\times\{1,...,c\}$:
\begin{equation}
\ell(\btheta;\bx,y) \enskip := \enskip \log p(y|\bx,\btheta).
\end{equation}
Moreover, for any set of pairs independently generated from the joint distribution of the training data, such as $\call_n$ mentioned in~\ref{item:iid}, the log-likelihood function will be:
\begin{equation}
\label{eq:loglikelihood_function}
\ell(\btheta;\call_n) \equals \sum_{i=1}^n \ell(\btheta;\bx_i,y_i)\equals \sum_{i=1}^n \log p(y_i|\bx_i,\btheta).
\end{equation}
hence the MLE can be rewritten as
\begin{equation}
\label{eq:MLE_ell}
\hat{\btheta}_n \equals \operatorname*{arg\,max}_{\btheta} \enskip \sum_{i=1}^n \ell(\btheta;\bx_i,y_i).
\end{equation}
Doing this maximization usually involves the computation of the stationary points of the log-likelihood, which requires calculating $\nabla_{\btheta}\ell(\btheta;\call_n)=\sum_{i=1}^n\nabla_{\btheta}\ell(\btheta;\bx_i,y_i)$.
For models assumed in~\ref{item:parameter_dependence}, each of the derivations in the summation is equal to the \emph{score function} defined as the gradient of the joint log-likelihood:
\begin{equation}
\label{eq:score_loglikelihood}
\nabla_{\btheta}\ell(\btheta;\bx,y) \equals \nabla_{\btheta}\log p(y|\bx,\btheta) \equals \nabla_{\btheta}\log p(\bx,y|\btheta) ,
\end{equation}
Equation~(\ref{eq:score_loglikelihood}) implies that the score will be the same no matter whether we choose the training or test distribution as our marginal. Furthermore, under regularity conditions~\ref{item:differentiability}, the score is always a zero-mean random variable\footnote{Score function is actually zero-mean even under weaker regularity conditions.}.
Finally, using MLE to estimate $\hat{\btheta}_n$, class label of a test sample $\bx$ will be predicted as the class with the highest log-likelihood value:
\begin{equation}
\hat{y}(\bx) \equals \operatorname*{arg\,max}_{y} \ell(\hat{\btheta}_n;\bx,y).
\end{equation}
\subsection{Fisher Information}
\label{subsec:fisher_information}
Here we give a very short introduction to Fisher information. More detailed descriptions about this well-known criterion can be found in various textbooks, such as~\citet{lehmann1998theory}.
Fisher information of a parametric distribution is a measure of information that the samples generated from that distribution provide regarding the parameter. It owes part of its importance to the Cram\'{e}r-Rao Theorem (see Appendix~\ref{subsec:parameter_estimation}, Theorem~\ref{thm:CR_bound}), which guarantees a lower-bound for the covariance of the parameter estimators.
Fisher information, denoted by $\bI(\btheta)$, is defined as the expected value of the outer-product of the score function with itself, evaluated at some $\btheta\in\Omega$. In our classification context, taking the expectation with respect to the training or test distributions gives us the training or test Fisher information criteria, respectively:
\begin{align}
\label{eq:FI_training_test}
\bI_q(\btheta) \enskip &:= \enskip \mathbb{E}_q\left[ \nabla_{\btheta}\log p(\bx,y|\btheta)\cdot\nabla_{\btheta}^\top\log p(\bx,y|\btheta) \right] \nonumber\\
\bI_p(\btheta) \enskip &:= \enskip \mathbb{E}_p\left[ \nabla_{\btheta}\log p(\bx,y|\btheta)\cdot\nabla_{\btheta}^\top\log p(\bx,y|\btheta) \right]
\end{align}
Here, we focus on $\bI_q$ to further explain Fisher information criterion. Our descriptions here can be directly generalized to $\bI_p$ as well. First, note that from equation~(\ref{eq:score_loglikelihood}) and that the score function is always zero-mean, one can reformulate the definition as:
\begin{align}
\label{eq:fisher_information_loglikelihood}
\bI_q(\btheta) &\equals \mathbb{E}_q\left[ \nabla_{\btheta}\ell(\btheta;\bx,y)\cdot \nabla_{\btheta}^\top\ell(\btheta;\bx,y) \right] \nonumber\\
&\equals \mbox{Cov}_q\left[\nabla_{\btheta}\ell(\btheta;\bx,y) \right]
\end{align}
Under the differentiability conditions~\ref{item:differentiability}, it is easy to show that we can also write the Fisher information in terms of the Hessian matrix of the log-likelihood:
\begin{equation}
\label{eq:fisher_hessian}
\bI_q(\btheta) \equals -\mathbb{E}_q\left[\nabla_{\btheta}^2\ell(\btheta;\bx,y) \right]
\end{equation}
Recall that the subscript $q$ in equations~(\ref{eq:fisher_information_loglikelihood}) and~(\ref{eq:fisher_hessian}) indicates that the expectations are taken with respect to the joint distribution that uses $q(\bx)$ as the marginal, that is $p(\bx,y|\btheta)=q(\bx)p(y|\bx,\btheta)$. Expansion of the expectation in~(\ref{eq:fisher_hessian}) results
\begin{align}
\label{eq:fisher_information_integration}
\bI_q(\btheta) &\equals -\mathbb{E}_{q(\bx)}\left[ \mathbb{E}_{y|\bx,\btheta}\left[ \nabla_{\btheta}^2\ell(\btheta;\bx,y) | \bx,\btheta \right] \right] \nonumber\\
&\equals -\int_{\bx\in X} q(\bx) \left[ \sum_{y=1}^c p(y|\bx,\btheta) \cdot \nabla^2_{\btheta}\ell(\btheta;\bx,y) \right]d\bx
\end{align}
\subsection{Some Properties of MLE}
\label{subsec:properties_MLE}
In this section, we formalize some of the key properties of MLE, which make this estimator popular in various fields.
They are also very useful in the theoretical analysis of FIR, provided in the next section. More detailed descriptions of these properties, together with the proofs that are skipped here, can be found in different sources, such as~\citet{wasserman2004all} and~\citet{lehmann1998theory}.
Note that a full understanding of the properties described in this section requires the knowledge of different modes of statistical convergence, specifically, convergence in probability ($\overset{P}{\to}$), and convergence in law ($\overset{L}{\to}$). A brief overview of these concepts are given in Appendix~\ref{sec:statistical_background}.
\begin{theorem}[\citet{lehmann1998theory}, Theorem 5.1]
\label{thm:MLE_consistency}
If the assumptions~\ref{item:parameter_dependence} to~\ref{item:openness} hold, then there exists a sequence of solutions $\left\{\hat{\btheta}^*_n\right\}_{n=1}^\infty$ to $\nabla_{\btheta}\ell(\btheta;\call_n)=0$ that converges to the true parameter $\btheta_0$ in probability.
\end{theorem}
Note that Theorem~\ref{thm:MLE_consistency} does not imply that convergence holds for \emph{any} sequence of MLEs. Hence, if there are multiple solutions to equation $\nabla_{\btheta}\ell(\btheta;\call_n)=0$ (the equation to solve for finding the stationary points) for every $n$, it is not obvious which root to select as $\hat{\btheta}_n^*$ to sustain the convergence. Therefore, while consistency of the MLE is guaranteed for models with a unique root of the score function evaluated at $\call_n$, it is not trivial how to build a consistent sequence when multiple roots exist. Here, in order to remove this ambiguity, we assume that either the roots are unique, become asymptotically unique, or we have access to an external procedure guiding us to select the proper roots so that $\hat{\btheta}_n\overset{P}{\to}\btheta_0$. We will denote the selected roots the same as $\hat{\btheta}_n$ from now on.
\begin{theorem}[\citet{lehmann1998theory}, Theorem 5.1]
\label{thm:asymptotic_efficiency_MLE}
Let $\hat{\btheta}_n$ be the maximum likelihood estimator based on the training data set $\call_n$. If the assumptions~\ref{item:parameter_dependence} to~\ref{item:FisherInformation} hold, then the MLE $\hat{\btheta}_n$ has a zero-mean normal asymptotic distribution with the covariance equal to the inverse Fisher information matrix, and with the convergence rate of $1/2$:
\begin{equation}
\label{eq:MLE_convergence_in_law}
\sqrt{n}(\hat{\btheta}_n-\btheta_0)\enskip \overset{L}{\rightarrow} \enskip \caln\left( \mathbf{0},\bI_{q}(\btheta_0)^{-1} \right)
\end{equation}
\end{theorem}
Theorems~\ref{thm:asymptotic_efficiency_MLE} and Cram\'{e}r-Rao bound (see Appendix~\ref{sec:statistical_background}), together with the consistency assumption, i.e. $\hat{\btheta}_n\overset{P}{\to}\btheta_0$, imply that MLE is an asymptotically efficient estimator with the efficiency equal to the training Fisher information. One can rewrite~(\ref{eq:MLE_convergence_in_law}) as
\begin{equation}
\sqrt{n}\cdot\bI_q(\btheta_0)^{1/2}(\hat{\btheta}_n-\btheta_0) \enskip \overset{L}{\to}\enskip \caln(\mathbf{0},\mathbb{I}_d)
\end{equation}
In the following corollary, we see that if we substitute $\bI_q(\btheta_0)$ with $\bI_q(\hat{\btheta}_n)$, the new sequence still converges to a normal distribution:
\begin{corollary}[\citet{wasserman2004all}, Theorem 9.18]
\label{cor:MLE_convergence_in_law_thetahat}
Under the assumptions of Theorem~\ref{thm:asymptotic_efficiency_MLE}, we get
\begin{equation}
\label{eq:MLE_convergence_in_law_thetahat}
\sqrt{n}\cdot\bI_{q}(\hat{\btheta}_n)^{1/2}(\hat{\btheta}_n-\btheta_0)\enskip \overset{L}{\rightarrow} \enskip \caln\left( \mathbf{0},\mathbb{I}_d \right)
\end{equation}
\end{corollary}
\subsection{Practical Issues}
\label{subsec:practical_issues}
The main difficulties consist of (1) having unknown variables in the objective, such as the test marginal, $p(\bx)$, and the true parameter, $\btheta_0$, and (2) lack of closed form for Fisher information matrices for most cases. In the next two sections, we review different hacks and solutions that have been proposed to resolve these issues.
\subsubsection{Replacing $\btheta_0$ by $\hat{\btheta}_{n'}$}
\label{subsubsec:replacing_theta}
Since $\btheta_0$ is not known, the simplest idea is to replace it by the current parameter estimate, that is $\hat{\btheta}_{n'}$~\citep{fukumizu2000statistical,settles2008analysis,Hoi2006,hoi2009batch,chaudhuri2015convergence}. Clearly, as the algorithm keeps running the iterations ($n'$ increases), the approximate objective (which contains $\btheta_{n'}$ instead of $\btheta_0$) gets closer to the original objective. This is due to the regularity and invertibility conditions assumed for the log-likelihood function and Fisher information matrices, respectively. Moreover, \citet{chaudhuri2015convergence} analyzed how this approximation effects the querying performance in finite-sample case.
Their analysis is done only for pool-based active learning, and when the test marginal $p(\bx)$ is a \emph{uniform distribution} $U(\bx)$ over the pool $X_p$. It is also assumed that the Hessian $\frac{\partial^2 \ell(\btheta;\bx,y)}{\partial \btheta^2}$ is independent of the class labels $y$, and therefore can be viewed as the conditional Fisher information $\bI(\btheta,\bx)$ (that is $\bI_p(\btheta)=\mathbb{E}_{p(\bx)}[\bI(\btheta,\bx)]$). Furthermore, there assumed to exist four positive constants $L_1,L_2,L_3, L_4\geq0$ such that the following four inequalities hold for all $\bx\in X_p$, $y\in\{1,...,c\}$ and $\btheta\in\Theta$:
\begin{align}
\label{eq:Chaudhuri_assumptions}
\nabla\ell(\btheta_0;\bx,y)^\top\bI_p(\btheta_0)^{-1}\nabla\ell(\btheta_0;\bx,y) \enskip&\leq\enskip L_1 \\
\left\|\bI_p(\btheta_0)^{-1/2}\bI(\btheta_0,\bx)\bI_p(\btheta_0)^{-1/2} \right\| \enskip &\leq \enskip L_2 \nonumber\\
\left\|\bI_p(\btheta_0)^{-1/2}(\bI(\btheta',\bx)-\bI(\btheta'',\bx))\bI_p(\btheta_0)^{-1/2} \right\| \enskip&\leq\enskip L_3(\btheta'-\btheta'')^\top\bI_p(\btheta_0)(\btheta'-\btheta'') \nonumber\\
-L_4\|\btheta - \btheta_0\|_2\bI(\btheta_0,\bx) \enskip\preceq\enskip \bI(\btheta,\bx)-\bI(\btheta_0,\bx) \enskip&\preceq\enskip L_4\|\btheta-\btheta_0\|_2\bI(\btheta_0,\bx) \nonumber
\end{align}
where $\btheta'$ and $\btheta''$ are any two parameters in a fixed neighborhood of $\btheta_0$. Then, provided that $n'$ is large enough, the following remark can be shown regarding the relationship between the FIRs computed at $\btheta_0$ and an estimate $\hat{\btheta}_{n'}$:
\begin{remark}
Let the assumptions~(\ref{item:parameter_dependence}) to~(\ref{item:FisherInformation}) and those in~(\ref{eq:Chaudhuri_assumptions}) hold. Moreover, assume that the Hessian is independent of the class labels. If $n'$ is large enough, then the following inequality holds for any $\beta\geq10$ with high probability:
\begin{equation}
\label{eq:replacement_inequality}
\mbox{\normalfont tr}\left[\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)\right] \enskip \leq \enskip \frac{\beta+1}{\beta-1}\cdot\mbox{\normalfont tr}\left[\bI_q(\hat{\btheta}_{n'})^{-1}\bI_p(\hat{\btheta}_{n'})\right]
\end{equation}
The minimum value for $n'$ that is necessary for having this inequality with probability $1-\delta$, increases quadratically with $\beta$ and reciprocally with $\delta$~\citep[Lemma 2]{chaudhuri2015convergence}.
\end{remark}
\begin{proof}
It is shown in the proof of Lemma 2 in \citet{chaudhuri2015arxiv} that under assumptions mentioned in the statement, the following inequalities hold with probability $1-\delta$:
\begin{equation}
\frac{\beta-1}{\beta}\bI(\bx,\btheta_0) \enskip\preceq\enskip \bI(\bx,\hat{\btheta}_{n'}) \enskip\preceq\enskip \frac{\beta+1}{\beta}\bI(\bx,\btheta_0).
\end{equation}
Taking expectation with respect to $p(\bx)$ and $q(\bx)$ result:
\begin{subequations}
\begin{align}
\frac{\beta-1}{\beta}\bI_p(\btheta_0) \enskip\preceq\enskip \bI_p(\hat{\btheta}_{n'}) \enskip\preceq\enskip \frac{\beta+1}{\beta}\bI_p(\btheta_0) \label{ineq:p_expectation}\\
\frac{\beta-1}{\beta}\bI_q(\btheta_0) \enskip\preceq\enskip \bI_q(\hat{\btheta}_{n'}) \enskip\preceq\enskip \frac{\beta+1}{\beta}\bI_q(\btheta_0) \label{ineq:q_expectation}
\end{align}
\end{subequations}
Since $\bI_q(\btheta_0)$ and $\bI_q(\hat{\btheta}_{n'})$ are assumed to be positive definite, we can write~(\ref{ineq:q_expectation}) in terms of inverted matrices\footnote{For any two positive definite matrices $\mathbf{A}$ and $\mathbf{B}$, we have that $\mathbf{A}\succeq\mathbf{B} \enskip\Rightarrow\enskip \mathbf{A}^{-1}\preceq\mathbf{B}^{-1}$.}:
\begin{equation}
\label{ineq:q_expectation_inv}
\frac{\beta}{\beta+1}\bI_q^{-1}(\btheta_0) \enskip\preceq\enskip \bI_q^{-1}(\hat{\btheta}_{n'}) \enskip\preceq\enskip \frac{\beta}{\beta-1}\bI_q^{-1}(\btheta_0)
\end{equation}
Now considering the first inequalities of~(\ref{ineq:p_expectation}) and~(\ref{ineq:q_expectation_inv}), multiplying both sides and taking the trace result~(\ref{eq:replacement_inequality}).
\end{proof}
Inequality~(\ref{eq:replacement_inequality}) implies that minimizing $\mbox{\normalfont tr}\left[\bI_q(\hat{\btheta}_{n'})^{-1}\bI_p(\hat{\btheta}_{n'})\right]$ (or an approximation of it) with respect to $q$ in each iteration of FIR-based querying algorithms, namely through the operation $\cala(\call_{n'},\hat{\btheta}_{n'})$ (line~\ref{line:generate_queries} of Algorithm~\ref{alg:AL}), is equivalent to upper bound minimization of the original cost function, i.e. left-hand-side of~(\ref{eq:FIR_variance_upper_bound}).
\subsubsection{Monte-Carlo Approximation}
Computation of Fisher information matrices is intractable unless when the marginal distributions are very simple or when they are restricted to be PMFs over finite number of samples. The latter is widely used in pool-based active learning, when the samples in the pool are assumed to be generated from $p(\bx)$. In such cases, one can simply utilize a Monte-Carlo approximation to compute $\bI_p(\tilde{\btheta})$. More specifically, denote the set of observed instances in the pool by $X_p$. Then the test Fisher information at any $\btheta\in\Omega$ can be approximated by
\begin{equation}
\label{eq:MC_approx_Ip}
\bI_p(\btheta)\approx\hat{\bI}(\btheta;X_p) \enskip := \enskip \frac{1}{|X_p|}\sum_{\bx\in X_p}\sum_{y=1}^c p(y|\bx,\btheta)\nabla_{\btheta}\ell(\btheta;\bx,y)\nabla_{\btheta}^\top\ell(\btheta;\bx,y) \enplus \delta \cdot\mathbb{I}_d
\end{equation}
where $\delta$ is a small positive number and the weighted identity matrix is added to ensure positive definiteness. It is important to remark that when using equation~\ref{eq:MC_approx_Ip}, we are actually utilizing some of the test samples in the training process, hence we cannot use those in $X_p$ in order to evaluate the performance of the trained classifier.
Similarly, $\bI_q(\hat{\btheta}_{n'})$ can be estimated based on a candidate query set $X_q$. Let $X_q$ be the set of samples drawn independently from $q(\bx)$. Then we can have the approximation $\bI_q(\hat{\btheta}_{n'})\approx\hat{\bI}(\hat{\btheta}_{n'};X_q)$. Putting everything together, the best query set $X_q\subseteq X_p$ is chosen to be the one that minimizes the approximate FIR querying objective:
\begin{equation}
\label{eq:hat_Xq_Xp_objective}
\mbox{tr}\left[\hat{\bI}(\hat{\btheta}_{n'};X_q)^{-1}\hat{\bI}(\hat{\btheta}_{n'};X_p)\right].
\end{equation}
Note that this objective is directly written in terms of $X_q$, and therefore the queries can be deterministically determined by fixing all the rest (including the current parameter estimate $\hat{\btheta}_{n'}$) and optimizing with respect to $X_q$. Therefore, such settings are usually called \emph{deterministic} active learning, as opposed to the probabilistic nature of~(\ref{eq:FIRAL}).
\subsubsection{Bound Optimization}
\label{subsubsec:bound_opt}
There are other types of approximation methods occurring in the optimization side. These methods are able to remove part of the unknown variables by doing upper-bound minimization or lower-bound maximization. Recall that in active learning, the querying objective is to be optimized with respect to $q$ (or $X_q$ in pool-based scenario). In a very simple example, when $d=1$, note that the $\bI_p(\btheta_0)$ is a constant scalar in~(\ref{eq:q_optimization}) and hence can be ignored. Hence, in the scalar case, we can simply focus on maximizing the training Fisher information. In the multivariate case, though it is not clear what measure of $\bI_q(\hat{\btheta}_n)$ to optimize, one may choose the objective to be $|\bI_q(\hat{\btheta}_n)|$ (where $|\cdot|$ is the determinant function),\footnote{Similar to \emph{D-optimality} in Optimal Experiment Design~\citep{fedorov1972theory}.} or $\mbox{tr}[\bI_q(\hat{\btheta}_n)]$.\footnote{Similar to \emph{A-optimality} in Optimal Experiment Design~\citep{fedorov1972theory}.} The latter is worth paying more attention due to the following inequality~\citep{yang2000matrix}:
\begin{equation}
\label{eq:trace_inequality}
\mbox{tr}[\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)] \enskip \leq \enskip \mbox{tr}[\bI_q(\btheta_0)^{-1}]\cdot\mbox{tr}[\bI_p(\btheta_0)].
\end{equation}
Since $\mbox{tr}[\bI_p(\btheta_0)]$ is a constant with respect to $q$, minimizing the right-hand-side of inequality~(\ref{eq:FIR_variance_upper_bound}) can itself be approximated by another upper-bound minimization:
\begin{equation}
\label{eq:upper_bound_minimization}
\argmin_{q}\enskip\mbox{tr}[\bI_q(\btheta_0)^{-1}]
\end{equation}
This helps removing the dependence of the objective to the test distribution $p$. A lower bound can also be established for the FIR. Using the inequality between arithmetic and geometric means of the eigenvalues of $\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)$, one can see that $d\cdot|\bI_q(\btheta_0)^{-1}|\cdot|\bI_p(\btheta_0)|\leq\mbox{tr}[\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)]$. Hence, when minimizing the upper-bound by minimizing the trace of $\bI_q(\btheta_0)^{-1}$, one should be careful about the determinant of this matrix as a term influencing the lower-bound of the objective.
In practice, of course, the minimization in~(\ref{eq:upper_bound_minimization}) can be difficult due to matrix inversion. Thus, sometimes it is further approximated by
\begin{equation}
\label{eq:maximization_trace_Iq}
\argmax_q \enskip\mbox{tr}[\bI_q(\btheta_0)].
\end{equation}
Hence, algorithms that aim to maximize $\mbox{tr}[\bI_q(\btheta_0)]$, indeed introduce three layers of objective approximations through equations~(\ref{eq:trace_inequality}) to~(\ref{eq:maximization_trace_Iq}). As discussed before, the dependence of the objectives in all the layers (in either~(\ref{eq:upper_bound_minimization}) or~(\ref{eq:maximization_trace_Iq})), can be removed by replacing it with the current estimate $\btheta_{n'}$.
\subsection{Some Existing Algorithms}
\label{subsec:algorithms}
In this section, we discuss several existing algorithms for implementing the query selection task based on minimization of FIR. We will analyze these algorithms, sorted according to date of their publication, in the context of our unifying framework.
Besides the categorizations that have already been described in previous sections, it is also useful to divide the querying algorithms into two categories based on the size of $X_q$: \emph{sequential active learning} where a single sample is queried at each iteration, i.e. $|X_q|=1$; and \emph{batch active learning} where the size of the query set is larger than one. The non-singleton query batches are usually generated greedily, with the batch size $|X_q|$ fixed to a constant value.
Table~\ref{tab:alg_summary} lists the algorithms that we reviewed in the following sections together with a summary of their properties and the approximate objective that they optimize for querying. Note that among these algorithms, the one by~\citet{chaudhuri2015convergence} makes extra assumptions as is described in Section~\ref{subsubsec:replacing_theta}.
\begin{table}[t!]
\caption{Reviewed FIR-based active learning algorithms for discriminative classifiers}
\label{tab:alg_summary}
\centering
\begin{tabular}{|| c | c | c || c c || c c || c c ||}
\hline
\multicolumn{2}{||c|}{\textbf{Algorithm}} & \textbf{Obj.} & \textbf{Prob.} & \textbf{Det.} & \textbf{Pool} & \textbf{Syn.} & \textbf{Seq.} & \textbf{Batch} \\ [0.5ex]
\hline\hline
1 & \citet{fukumizu2000statistical} & (\ref{eq:maximization_trace_Iq}) & \ding{51} & & & \ding{51} & \ding{51} & \ding{51} \\
\hline
2 & \citet{Zhang2000} & (\ref{eq:maximization_trace_Iq}) & & \ding{51} & \ding{51} & \ding{51} & \ding{51} & \\
\hline
3 & \small \citet{settles2008analysis} & (\ref{eq:hat_Xq_Xp_objective}) & & \ding{51} & \ding{51} & & \ding{51} & \ding{51} \\
\hline
4 & \citet{Hoi2006,hoi2009batch} & (\ref{eq:hat_Xq_Xp_objective}) & & \ding{51} & \ding{51} & & \ding{51} & \ding{51} \\
\hline
5 & \citet{chaudhuri2015convergence} & (\ref{eq:replacement_inequality}) & \ding{51} & & \ding{51} & & \ding{51} & \ding{51} \\
\hline
\end{tabular}
\end{table}
\vspace{.25cm}
\noindent\textbf{Algorithm 1 } \textit{(\citet{fukumizu2000statistical})}
\vspace{.25cm}
This algorithm is the classification version of the \emph{probabilistic active learning} proposed by~\citet{fukumizu2000statistical} for regression problem. The assumption is that the proposal belongs to a parametric family and is of the form $q(\bx;\balpha)$, where $\balpha$ is the parameter vector of the family. In this \emph{parametric} active learning, the best set of parameters $\hat{\balpha}$ is selected using the current parameter estimate and the query set is sampled from the resulting proposal distribution $X_q\sim q(\bx;\hat{\balpha})$.
This algorithm makes no use of the test samples and optimizes the simplified objective in~(\ref{eq:maximization_trace_Iq}) to obtain the query distribution $q(\bx)$. Denote the covariates of the current training data set $\call_{n'}$ by $X_{\call}$. As is described in section~(\ref{subsec:practical_issues}), the new trace objective can be approximated by Monte-Carlo formulation using the old queried samples $X_{\call}$ as well as the candidate queries $X_q$ to be selected in this iteration:
\begin{equation}
\label{eq:MC_approx_alg1}
\mbox{tr}\left[\hat{\bI}(\hat{\btheta}_{n'};X_{\call}\cup X_q)\right],
\end{equation}
\begin{algorithm}[!t]
\caption{\citet{fukumizu2000statistical}}
\label{alg:parametric_AL}
\textbf{Inputs:} Current estimation of the parameter $\hat{\btheta}_{n'}$, size of the query set $|X_q|$\\
\textbf{Outputs:} The query set $X_q$\\[-5pt]
\noindent\rule{.5\columnwidth}{.5pt}\\
\SetNlSkip{1em}
\DontPrintSemicolon
\tcc{\footnotesize Parameter Optimization}
\nl$\hat{\balpha} \equals \argmax_{\balpha} \mathbb{E}_{q(\bx;\balpha)}\left[\sum_{y=1}^cp(y|\bx,\hat{\btheta}_{n'})\nabla_{\btheta}^\top\ell(\hat{\btheta}_{n'};\bx,y)\nabla_{\btheta}\ell(\hat{\btheta}_{n'};\bx,y) \right]$\;
\tcc{\footnotesize Sampling from the parametric proposal}
\nl$\bx_i \enskip\sim\enskip q(\bx;\hat{\balpha}) \qquad,i=1,...,|X_q|$\;
\nl\textbf{return} $X_q=\left\{\bx_1,...,\bx_{|X_q|}\right\}$\;
\end{algorithm}
More specifically, the new parameter vector is obtained by maximizing the expected contribution of the queries $X_q$ generated from $q(\bx;\balpha)$ to this objective. Taking expectation of~(\ref{eq:MC_approx_alg1}) with respect to $q(\bx;\balpha)$ yields:
\begin{equation}
\label{eq:parametric_objective}
\mathbb{E}_{q(\bx;\balpha)}\left[\mbox{tr}\left[\hat{\bI}(\hat{\btheta}_{n'};X_{\call}\cup X_q)\right]\right] \equals
\mbox{tr}\left[\frac{n'}{n' + |X_q|} \hat{\bI}(\hat{\btheta}_{n'};X_{\call}) + \frac{1}{n'+|X_q|}\mathbb{E}_{q(\bx;\balpha)}\left[\hat{\bI}(\hat{\btheta}_{n'};X_q)\right] \right].
\end{equation}
Recall that $n'$ is the size of the current training data set $\call_{n'}$. The first term in~(\ref{eq:parametric_objective}) is independent of the query set $X_q$ (assuming that the size $|X_q|$ is fixed to a constant), hence we focus only on the second term in our optimization. Noting that the queries are generated independently, we can rewrite this term as:
\begin{align}
\label{eq:expected_contribution}
\mathbb{E}_{q(\bx;\balpha)}\bigg[\mbox{tr}[\hat{\bI}(\hat{\btheta}_{n'};X_q)]\bigg] &\equals
\mathbb{E}_{q(\bx;\balpha)}\left[\frac{1}{|X_q|}\sum_{\bx\in X_q}\mbox{tr}\left[\hat{\bI}(\hat{\btheta}_{n'};\bx)\right]\right] \enskip -\enskip (|X_q|-1)\delta\cdot\mathbb{I}_d \nonumber\\
&\equals \mathbb{E}_{q(\bx;\balpha)}\bigg[\mbox{tr}\left[\hat{\bI}(\hat{\btheta}_{n'};\bx)\right]\bigg] \enskip -\enskip (|X_q|-1)\delta\cdot\mathbb{I}_d
\end{align}
From equation~(\ref{eq:expected_contribution}), if we are to select a single query, the parameter vector $\balpha$ can be obtained by maximizing the expected contribution of that single query to the trace objective, that is:
\begin{align}
\label{eq:alphan_optimization}
\hat{\balpha} &\equals \argmax_{\balpha} \enskip \mathbb{E}_{q(\bx;\balpha)}\bigg[\mbox{tr}\left[\hat{\bI}(\hat{\btheta}_{n'};\bx)\right]\bigg] \nonumber\\
&\equals \argmax_{\balpha} \enskip \mathbb{E}_{q(\bx;\balpha)}\left[\sum_{y=1}^cp(y|\bx,\hat{\btheta}_{n'})\nabla_{\btheta}^\top\ell(\hat{\btheta}_{n'};\bx,y)\nabla_{\btheta}\ell(\hat{\btheta}_{n'};\bx,y) \right]
\end{align}
The optimization~(\ref{eq:alphan_optimization}) does not depend on $X_{\call}$, and therefore we do not need to explicitly feed this algorithm with $\call$. All it needs is an estimation of the parameter $\hat{\btheta}_{n'}$. The two-step procedure of generating queries from parametric query distribution is shown in Algorithm~\ref{alg:parametric_AL}. This algorithm can be used in both sequential and batch modes by changing the number of samples drawn from $q(\bx;{\alpha})$.
We emphasize that Algorithm~\ref{alg:parametric_AL} is probabilistic, meaning that with any fixed parameter estimate $\hat{\btheta}_{n'}$, the next set of queries are \emph{not} deterministically selected. The optimization is performed with respect to the parameters of the proposal distribution, which are then used to sample $X_q$. \citet{fukumizu2000statistical} claims that introducing such randomness into active learning, which increases exploration against exploitation, may prevent the algorithm from falling into local optima. Also note that this algorithm is not pool-based, meaning that it does not select the queries from a pool of observed instances, although could be constrained to do so.
\vspace{.25cm}
\noindent\textbf{Algorithm 2} \textit{(\citet{Zhang2000})}
\vspace{.25cm}
\citet{Zhang2000} started from optimization problem~(\ref{eq:maximization_trace_Iq}), and introduced even additional simplifications to it, specifically considering the use of a binary logistic regression classifier. Here, we discuss their formulation using a general discriminate framework.
In their algorithm, a single query is selected at each iteration. Denote it by $X_q = \{\bx_q\}$.
Similar to the previous section, the Fisher information matrix $\bI_q$ can be approximated by Monte-Carlo approximation. \citet{Zhang2000} discarded the expectation with respect to the proposal distribution in~(\ref{eq:alphan_optimization}) or equivalently consider $q$ to be a uniform distribution. Therefore, the optimization with respect to parameters turned into a direct optimization with respect to the single query $\bx_q$:
\begin{align}
\label{eq:trace_Iq_expanded}
\bx_q \equals \argmax_{\bx\in X} \enskip \sum_{y=1}^cp(y|\bx,\hat{\btheta}_{n'})\nabla_{\btheta}^\top\ell(\hat{\btheta}_{n'};\bx,y)\nabla_{\btheta}\ell(\hat{\btheta}_{n'};\bx,y)
\end{align}
This single-step deterministic approach, shown in Algorithm~\ref{alg:deterministic_AL_Zhang}, is very similar to the probabilistic approach described above, except that there is no intermediate parameter optimization step.
\begin{algorithm}[!t]
\caption{\citet{Zhang2000}}
\label{alg:deterministic_AL_Zhang}
\textbf{Inputs:} Current estimation of the parameter $\hat{\btheta}_{n'}$\\
\textbf{Outputs:} The query singleton set $X_q=\{\bx_q\}$\\[-5pt]
\noindent\rule{.5\columnwidth}{.5pt}\\
\SetNlSkip{1em}
\DontPrintSemicolon
\nl$\bx_q \enskip\gets\enskip \argmax_{\bx} \enskip \sum_{y=1}^cp(y|\bx,\hat{\btheta}_{n'})\nabla_{\btheta}^\top\ell(\hat{\btheta}_{n'};\bx,y)\nabla_{\btheta}\ell(\hat{\btheta}_{n'};\bx,y)$\;
\nl\textbf{return} $X_q=\{\bx_q\}$\;
\end{algorithm}
It is important to note that Algorithm~\ref{alg:deterministic_AL_Zhang} can be used in pool-based active learning as well. This can be done by constraining $\bx_q$ to be a member of a pool of samples, in which case it can even be extended to batch querying by sorting the unlabeled samples based on their objective values and taking the highest ones. However, such iterative optimization is not efficient, because the resulting queries will most probably be close to each other and therefore contain redundant information.
\newpage
\vspace{.25cm}
\noindent\textbf{Algorithm 3} \textit{(\citet{settles2008analysis})}
\vspace{.25cm}
Inspired by \citet{Zhang2000}, \citet{settles2008analysis} employed Fisher information ratio to develop a pool-based active learning, which can be used in either sequential or batch querying. The pool that is used here is the set of unlabeled samples, $X_p$, which are assumed to be drawn from the test marginal $p(\bx)$. Queries are chosen from $X_p$, that is $X_q\subseteq X_p$. The test Fisher information matrix can be approximated by Monte-Carlo simulation over the samples in $X_p$, meaning $\hat{\bI}\left(\hat{\btheta}_{n'};X_p\right)$.
Similar to Algorithm~\ref{alg:parametric_AL}, the updated training Fisher information matrix after querying a set $X_q$ can be approximated by $\hat{\bI}\left(\hat{\btheta}_{n'};X_{\call}\cup X_q \right)$. Thus, since we do have an approximation of both Fisher information matrices, the objective to minimize is chosen to be in the form of~(\ref{eq:hat_Xq_Xp_objective}).
Similar to the \citet{Zhang2000} algorithm, the proposal distribution $q$ is ignored in the objective (or equivalently considered as being uniform). An additional assumption \citet{settles2008analysis} made to simplify the optimization task is:
\begin{align}
\label{eq:settles_simplifying_assumption}
\argmin_{X_q\subset X_p} \enskip \mbox{tr}&\left[\hat{\bI}\left(\hat{\btheta}_{n'};X_{\call}\cup X_q\right)^{-1}\hat{\bI}\left(\hat{\btheta}_{n'};X_p\right)\right] \enskip\approx\enski
\argmin_{X_q\subset X_p} \enskip \mbox{tr}\left[\hat{\bI}\left(\hat{\btheta}_{n'};X_q\right)^{-1}\hat{\bI}\left(\hat{\btheta}_{n'};X_p\right)\right].
\end{align}
This simplified optimization is easy to implement for sequential active learning. However, the combinatorial optimization required for batch active learning can easily become intractable. As shown in Algorithm~\ref{alg:deterministic_AL_Settles}, \citet{settles2008analysis} used a greedy approach to do this optimization (the inner loop).
\begin{algorithm}[!t]
\caption{\citet{settles2008analysis}}
\label{alg:deterministic_AL_Settles}
\textbf{Inputs:} Current estimation of the parameter $\hat{\btheta}_{n'}$, the set of unlabeled samples $X_p$, , size of the query set $|X_q|$\\
\textbf{Outputs:} The query set $X_q$\\[-5pt]
\noindent\rule{.5\columnwidth}{.5pt}\\
\SetNlSkip{1em}
\DontPrintSemicolon
\tcc{\footnotesize Initializing the query set for this iteration}
\nl$X_q \enskip \gets \enskip \varnothing$\;
\tcc{\footnotesize The loop for greedy batch querying}
\nl\For{$j=1\to |X_q|$}{
\tcc{\footnotesize Query optimization and adding the result into the query set}
\nl$X_q \enskip\gets\enskip X_q \enskip \cup \enskip \argmin_{\bx\in X_p} \mbox{tr}\left[\hat{\bI}\left(\hat{\btheta}_{n'};\bx \right)^{-1}\hat{\bI}\left(\hat{\btheta}_{n'};X_p\right)\right]$\label{line:settles_optimization}\;
\tcc{\footnotesize Removing the selected queries from the pool}
\nl$X_p \enskip \gets \enskip X_p - X_q$\;
}
\nl\textbf{return} $X_q$\;
\end{algorithm}
\vspace{.25cm}
\noindent \textbf{Algorithm 4} \textit{(\citet{Hoi2006} and~\citet{hoi2009batch})}
\vspace{.25cm}
The algorithms proposed by \citet{Hoi2006} and \citet{hoi2009batch} are very similar to the one developed by \citet{settles2008analysis}, described above, except that they use a more sophisticated optimization method. Their method shown in Algorithm~\ref{alg:deterministic_AL_Hoi}, is different from Algorithm~\ref{alg:deterministic_AL_Settles} mainly in the way that it greedily chooses the query at each inner loop iteration of the algorithm. While Algorithm~\ref{alg:deterministic_AL_Settles} exclusively considers the contribution of each $\bx\in X_q$, ignoring the samples selected in the previous iterations (hence $\hat{\bI}\left(\hat{\btheta}_{n'};\bx \right)$ in line~\ref{line:settles_optimization} of Algorithm~\ref{alg:deterministic_AL_Settles}), Algorithm~\ref{alg:deterministic_AL_Hoi} takes into account all the queries chosen so far (hence $\hat{\bI}\left(\hat{\btheta}_{n'};X_q\cup\{\bx\} \right)$ in line~\ref{line:hoi_optimization} in Algorithm~\ref{alg:deterministic_AL_Hoi}).
\citet{Hoi2006} and \citet{hoi2009batch} showed that when using \emph{binary logistic regression} classifier, their optimization~(\ref{eq:settles_simplifying_assumption}) can be done by maximizing a \emph{submodular} set function with respect to the query set $X_q$. This allowed them to use the well-known iterative algorithm proposed by~\citet{Nemhauser1978a}, which guarantees a tight lower-bound for maximization of submodular and monotone set functions.
In the rest of this section, we show that minimizing this objective obtained from the above-mentioned assumptions, can be efficiently approximated by a monotonically submodular maximizing under \emph{any} discriminative classifier. This is a generalization of the result derived by~\citet{Hoi2006} that is obtained in case of using logistic regression classifier. As a consequence, FIR can be efficiently optimized with guaranteed tight bounds~\citep{Nemhauser1978a,Nemhauser1978b}.
The following lemma shows that~(\ref{eq:settles_simplifying_assumption}) is approximately equivalent to maximizing a simplified set function, for any unlabeled sample pool $X_p$:
\begin{algorithm}[!t]
\caption{\citet{Hoi2006, hoi2009batch}}
\label{alg:deterministic_AL_Hoi}
\textbf{Inputs:} Current estimation of the parameter $\hat{\btheta}_{n'}$, the set of unlabeled samples $X_p$, size of the query set $|X_q|$\\
\textbf{Outputs:} The query set $X_q$\\[-5pt]
\noindent\rule{.5\columnwidth}{.5pt}\\
\SetNlSkip{1em}
\DontPrintSemicolon
\tcc{\footnotesize Initializing the query set}
\nl$X_q\gets \varnothing$\;
\tcc{\footnotesize The loop for greedy batch querying}
\nl\For{$j=1\to |X_q|$}{\label{line:Hoi_greedy_max_start}
\tcc{\footnotesize Query optimization}
\nl$\tilde{\bx} \equals \argmin_{\bx\in X_p} \mbox{tr}\left[\hat{\bI}\left(\hat{\btheta}_{n'};X_q\cup\{\bx\} \right)^{-1}\hat{\bI}\left(\hat{\btheta}_{n'};X_p\right)\right]$\label{line:hoi_optimization}\;
\tcc{\footnotesize Add the selected query into the query set}
\nl$X_q\gets X_q\cup\{\tilde{\bx}\}$\;
\tcc{\footnotesize Remove the selected instance from the pool}
\nl$X_p\gets X_p-\{\tilde{\bx}\}$\label{line:Hoi_greedy_max_finish}\;
}
\nl\textbf{return} $X_q$\;
\end{algorithm}
\begin{lemma}
\label{lemma:equivalent_set_optimization}
Let $X_p,X_q\subseteq X$ be two non-empty and finite subsets of samples randomly generated from $p(\bx)$ and its resample distribution $q(\bx)$, respectively, such that $X_q\subset X_p$, and the parameter $\delta\geq0$ in~(\ref{eq:MC_approx_Ip}) is a small constant. If assumptions~\ref{item:parameter_dependence}, \ref{item:iid}, \ref{item:differentiability} and \ref{item:FisherInformation} hold, then the following optimization problems are approximately equivalent for some function $g_{\btheta}:X\times\{1,...,c\}\times X\to\mathbb{R}^+$, $d$-dimensional non-zero vector $\bv_{\btheta}$ depending on $\bx$ and $y$, and for all $\btheta\in\Omega$ :
\begin{subequations}
\begin{align}
\label{eq:original_set_optimization}
(i)\quad \argmin_{X_q\subset X_p} \enskip &\mbox{\normalfont{tr}}\left[\hat{\bI}(\btheta;X_q)^{-1}\hat{\bI}(\btheta;X_p)\right] \\
\label{eq:equivalent_set_optimization}
(ii)\quad \argmax_{X_q\subset X_p} \enskip &\sum_{\bx\in X_p-X_q}\sum_{y=1}^c\frac{-1}{\delta\cdot\|\bv_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')}
\end{align}
\label{eq:set_optimizations}
\end{subequations}
The approximation is more accurate for smaller $\delta$ and well-conditioned Monte-Carlo approximation of proposal Fisher information matrix.
\end{lemma}
The proof can be found in Appendix~\ref{app:proof_Lemma}.
Note that Lemma~\ref{lemma:equivalent_set_optimization}, as stated above, does not depend on the size of $X_q$. However, just as before, in practice it is usually assumed that $|X_q|>0$ is fixed and therefore the optimizations in~(\ref{eq:set_optimizations}) should be considered with cardinality constraint. In general, combinatorial maximization problems can turn out to be intractable. Next, it is shown that the objective at hand is a monotonically submodular set function in terms of $X_q$ and therefore can be maximized efficiently with a greedy approach such as that shown in Algorithm~\ref{alg:deterministic_AL_Hoi}.
\begin{theorem}
\label{thm:submodularity}
Suppose $f_{\btheta}:2^{X_p}\to\mathbb{R}$ is defined as:
\begin{equation}
\label{eq:submodular_objective}
f_{\btheta}(X_q) \equals \sum_{\bx\in X_p-X_q}\sum_{y=1}^c\frac{-1}{\delta\cdot\|\bv_{\btheta}(\bx,y)\|^{-2} + \sum_{\bx'\in X_q} g_{\btheta}(\bx,y,\bx')},\quad \forall X_q\subseteq X_p
\end{equation}
with $\bv_{\btheta}$ a $d$-dimensional vector depending on $\bx$ and $y$, and $g_{\btheta}$ defined in~(\ref{eq:defn_g}).
Then $f_{\btheta}$ is a submodular and monotone (non-decreasing) set function for all $\btheta\in\Omega$.
\end{theorem}
The proof is in Appendix~\ref{app:proof_submodularity}.
The result above, together with Lemma~\ref{lemma:equivalent_set_optimization}, imply that the objective of~(\ref{eq:equivalent_set_optimization}) is a monotonically increasing set function with respect to $X_q$.
Below we present the main result that guarantees tight bounds for greedy maximization of monotonic submodular set functions. Details of this result, which is also shown to be the optimally efficient solution to submodular maximization, can be found in the seminal papers by~\citet{Nemhauser1978a} and~\citet{Nemhauser1978b}.
\begin{theorem}[\citet{Nemhauser1978a}]
\label{thm:greedy_bound}
Let $f_{\btheta}:2^{X_p}\to\mathbb{R}$ be any submodular and nondecreasing set function with $f(\varnothing)=0$ \footnote{This can always be assumed since maximizing a general set function $f(X_q)$ is equivalent to maximizing its adjusted version $g(X_q):=f(X_q) - f(\varnothing)$, which satisfies $g(\varnothing)=0$.}. If $X_q$ is the output of a greedy maximization algorithm, and $X_q^*$ is the optimal maximizer of $f_{\btheta}$ with a cardinality constraint (fixed $|X_q|$), then we have:
\begin{align}
\label{eq:deterministic_greedy_bound}
f_{\btheta}(X_q)\enskip \geq \enskip\left[1-\left(\frac{|X_q|-1}{|X_q|}\right)^{|X_q|} \right]f_{\btheta}(X_q^*)\enskip\geq\enskip \left(1-\frac{1}{e}\right)f_{\btheta}(X_q^*).
\end{align}
\end{theorem}
In Algorithm~\ref{alg:deterministic_AL_Hoi}, the inner loop (lines~\ref{line:Hoi_greedy_max_start} to~\ref{line:Hoi_greedy_max_finish}) implements the minimization in~(\ref{eq:settles_simplifying_assumption}) greedily. We have seen above that this set minimization is approximately equivalent to maximizing a submodular and monotone set maximization, which, in turn, is shown to be efficient.
\newpage
\vspace{.25cm}
\noindent \textbf{Algorithm 5} \textit{(\citet{chaudhuri2015convergence})}
\vspace{.25cm}
This algorithm uses FIR for doing a probabilistic pool-based active learning. It has extra assumptions in comparison to our general framework, which are briefly explained in Section~\ref{subsubsec:replacing_theta}. Note that these assumptions are to be made as well as those listed in Section~\ref{sec:framework_assumptions}. In such settings, \citet{chaudhuri2015convergence} gave a finite-sample theoretical analysis for FIR when applied to pool-based active learning.
More specifically, suppose $p(\bx)$ is a uniform PMF and $q(\bx)$ is a general PMF, both defined over the pool $X_p$. Using the notations in~(\ref{eq:Chaudhuri_assumptions}), the training Fisher information can be written as $\bI_q(\hat{\btheta}_{n'})=\sum_{\bx\in X_p}q(\bx)\bI(\hat{\btheta}_{n'},\bx)$. Now, assuming that $\bI_p(\hat{\btheta}_{n'})$ has a singular decomposition of the form $\sum_{j=1}^d\sigma_j\mathbf{u}_j\mathbf{u}_j^\top$, FIR can be written as:
\begin{align}
\label{eq:chaudhuri_objective}
\mbox{tr}\bigg[\bI_q(\hat{\btheta}_{n'})^{-1}\bI_p(\hat{\btheta}_{n'})\bigg] &\equals \sum_{j=1}^d\sigma_j\mbox{tr}\bigg[\bI_q(\hat{\btheta}_{n'})^{-1}\mathbf{u}_j\mathbf{u}_j^\top\bigg] \nonumber\\
&\equals \sum_{j=1}^d\sigma_j\mathbf{u}_j^\top\bI_q(\hat{\btheta}_{n'})^{-1}\mathbf{u}_j
\end{align}
Minimizing the last term in~(\ref{eq:chaudhuri_objective}) with respect to PMF $\{q(\bx) | \bx\in X_p\}$ is equivalent to a semidefinite programming after introducing a set of auxiliary variables $t_j, j=,1...,d$ and applying Schur complements~\citep{vandenberghe1996semidefinite}:
\begin{align}
\label{eq:chaudhuri_SDP}
\argmin_{q(\bx), \bx\in X_p} \quad &\sum_{j=1}^d \sigma_j t_j \\
\mbox{such that } \enskip
&\begin{bmatrix}
t_j & \mathbf{u}_j^\top \\
\mathbf{u}_j & \sum_{\bx\in X_p} q(\bx)\bI(\hat{\btheta}_{n'},\bx)
\end{bmatrix}
\succeq 0, \nonumber\\
& \sum_{\bx\in X_p}q(\bx)=1. \nonumber
\end{align}
The steps for this querying method is shown in Algorithm~\ref{alg:probabilistic_chaudhuri}. Note that the solution to~(\ref{eq:chaudhuri_SDP}) is slightly modified by mixing it with the uniform distribution over the pool. Such modification is mainly to establish their theoretical derivations. The mixing coefficient, $0\leq\lambda\leq1$ reciprocally depends on the number of queries. More specifically, \citet{chaudhuri2015convergence} made it equal to $1-\frac{1}{|X_q|^{1/6}}$. That is, as the number of queries increases, $\lambda$ shrinks and so does the modification. Furthermore, in their analysis, they assumed that sampling from $\tilde{q}(\bx)$ (line~\ref{line:sampline_qtilde} of Algorithm~\ref{alg:probabilistic_chaudhuri}) is done \emph{with replacement}. That is, label of a given sample might be queried multiple times.
\begin{algorithm}[!t]
\caption{\citet{chaudhuri2015convergence}}
\label{alg:probabilistic_chaudhuri}
\textbf{Inputs:} Current estimation of the parameter $\hat{\btheta}_{n'}$, the set of unlabeled samples $X_p$, size of the query set $|X_q|$\\
\textbf{Outputs:} The query set $X_q$\\[-5pt]
\noindent\rule{.5\columnwidth}{.5pt}\\
\SetNlSkip{1em}
\DontPrintSemicolon
\tcc{\footnotesize Solving the semidefinite programming}
\nl $q(\bx) \enskip\gets \enskip$ solution to~(\ref{eq:chaudhuri_SDP})\;
\tcc{\footnotesize Modification of the solution}
\nl $\tilde{q}(\bx) \enskip\gets \enskip \lambda q(\bx) + (1-\lambda)U(\bx)$ \;
\tcc{\footnotesize Sampling with replacement from the modified proposal}
\nl$\bx_i \enskip\sim\enskip \tilde{q}(\bx) \qquad, i=1,...,|X_q|$\label{line:sampline_qtilde}\;
\nl \textbf{return} $X_q=\left\{\bx_1,...\bx_{|X_q|}\right\}$\;
\end{algorithm}
\subsubsection{Comparison with Other Information-theoretic Objectives}
\label{subsubsec:comparison}
In the last part of this section, we compare FIR and two other common querying objectives from the field of information theory. Entropy of class labels and mutual information between labeled and unlabeled samples are two other common active learning objectives. Their goal is mainly to get the largest possible amount of information about \emph{class labels of unlabeled samples} from each querying iteration, hence naturally pool-based.
Entropy-based querying, also known as uncertainty sampling, directly measures the uncertainty with respect to class label of each unlabeled sample and query those with highest uncertainty. It has been hugely popular due to its simplicity and effectiveness especially in sequential active learning. However, it does not consider interaction between samples when selecting multiple queries, which can cause querying very similar samples (redundancy). Therefore, uncertainty sampling shows relatively poor performance in batch active learning. Mutual information, on the other hand, does not suffer from redundancy, however, it requires a much higher computational complexity.
These two objectives directly measure the amount of information each batch can have with respect to the class labels (hence prediction-based), as opposed to Fisher information as a measure of information regarding the distribution parameters (hence inference-based). However, there is no guarantee that by minimizing uncertainty of the class labels (or equivalently, choosing queries with highest amount of information about class labels), the prediction accuracy also increases. Whereas, as we showed earlier, FIR upper-bounds the expected asymptotic variance of a parameter inference loss function. From this point of view, FIR has a closer relationship with the performance of a classifier.
\setlength{\extrarowheight}{5pt}
\begin{table}[t!]
\caption{Computational complexity of different querying algorithms}
\label{tab:complexity}
\centering
\begin{tabular}{| c | c |}
\hline
\textbf{Algorithm} & \textbf{Complexity} \\ [0.5ex]
\hline\hline
Entropy & $O(|X_p|cd)$ \\[0.5ex]
\hline
Mutual Information & $O\left(|X_p|\cdot|X_q|\cdot c^{|X_q|+1}d\right)$ \\[0.5ex]
\hline\hline
\citet{Zhang2000} & $O(|X_p| cd)$ \\[0.5ex]
\hline
\small \citet{settles2008analysis} & $O\left(|X_q|\cdot|X_p|\cdot(cd+d^3)\right)$ \\[0.5ex]
\hline
\citet{Hoi2006,hoi2009batch} & $O\left(|X_q|\cdot|X_p|\cdot(cd+cd|X_q|+d^3)\right)$ \\[0.5ex]
\hline
\citet{chaudhuri2015convergence} & $O\left( d^3|X_p|^2 + d^4|X_p| + d^5\right)$ \\[0.5ex]
\hline
\end{tabular}
\end{table}
Table~\ref{tab:complexity} shows computational complexity of the querying objectives. The algorithm by \citet{fukumizu2000statistical} is excluded from this table since it cannot be used in pool-based sampling. Also note that the complexity reported for mutual information is for the case when it is optimized greedily. Nevertheless, it still contains an exponential term in its complexity. Entropy-based and \citet{Zhang2000} have the lowest complexity, but in the expense of introducing redundancy into the batch of queries. Algorithms by \citet{settles2008analysis}, \citet{Hoi2006,hoi2009batch} and \citet{chaudhuri2015convergence} become very expensive when $d$ is large, whereas mutual information can easily get intractable for selecting batches of higher size (large $|X_q|$). Observe that algorithm by \citet{Hoi2006, hoi2009batch} is more expensive than \citet{settles2008analysis}. Recall that despite similarities in appearance, the former guarantees tight bound for its greedy optimization, whereas the latter does not.
The complexity for the algorithm by \citet{chaudhuri2015convergence} is computed assuming that a barrier method (following path) is used as its numerical optimization~\citep{boyd2004convex}. From Table~\ref{tab:complexity}, this algorithm is the only one whose complexity increases quadratically with size of the pool $|X_p|$, and therefore can get significantly slow for huge pools. Furthermore, it does not depend on $|X_q|$ since the optimization in~(\ref{eq:chaudhuri_SDP}) as its main source of computation, only depends on $|X_p|$ and $d$ (computing $\bI(\hat{\btheta}_{n'},\bx)$ is assumed to cost $O(1)$ for each $\bx\in X_p$ as it is taken to be independent of $y$).
\subsection{Asymptotic Distribution of MLE-based Classifier}
\label{subsec:asymptotic_distribution}
Recall that the estimated parameter $\hat{\btheta}_n$ is obtained from a given proposal distribution $q(\bx)$. The log-likelihood ratio function, at a given pair $(\bx,y)$, is defined as:
\begin{equation}
\label{eq:likelihood_ratio}
\ell(\hat{\btheta}_n;\bx,y) \enskip - \enskip \ell(\btheta_0;\bx,y).
\end{equation}
This ratio can be viewed as an example of the classification loss function whose expectation with respect to the test joint distribution of $\bx$ and $y$, results in the \emph{discrepancy} between the true conditional $p(y|\bx,\btheta_0)$ and MLE conditional $p(y|\bx,\hat{\btheta}_n)$~\citep{murata1994network}. Here, we analyze this measure asymptotically as $(n\to\infty)$. Primarily, note that based on continuity of the log-likelihood function \ref{item:differentiability} and consistency of MLE (Theorem~\ref{thm:MLE_consistency}), equation~(\ref{eq:likelihood_ratio}) converges in probability to zero for any $(\bx,y)$.
Furthermore, equation~(\ref{eq:likelihood_ratio}) is dependent on both the true marginal $p(\bx)$ (through the test pairs, where it should be evaluated) and the proposal marginal $q(\bx)$ (through the MLE $\hat{\btheta}_n$). In the classification context, \citet{Zhang2000} claimed that the expected value of this ratio with respect to both marginals converges to $\mbox{tr}[\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)]$ with the convergence rate equal to unity. In the scalar case, $\mbox{tr}[\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)]$ is equal to the ratio of the Fisher information of the true and proposal distributions, the reason why it is sometimes referred to as the \emph{Fisher information ratio}~\citep{settles2008analysis}. This objective have been widely studied in linear and non-linear regression problems~\citep{fedorov1972theory,mackay1992information,murata1994network,cohn1996neural,fukumizu2000statistical}. However, it is not as fully analyzed in classification.
\citet{Zhang2000} and many papers following them \citep{Hoi2006,settles2008analysis,hoi2009batch}, used this function as an \emph{asymptotic} objective in active learning to be optimized with respect to the proposal $q$. Here, we show that this objective can also be viewed as an \emph{upper bound} for the expected variance of the asymptotic distribution of~(\ref{eq:likelihood_ratio}).
First, we investigate the asymptotic distribution of the log-likelihood ratio in two different cases:
\begin{theorem}
\label{thm:asymptotic_distribution}
If the assumptions~\ref{item:parameter_dependence} to~\ref{item:FisherInformation} hold, then, at any given $(\bx,y)\in X\times\{1,...,c\}$:
\begin{enumerate}[label=(\Roman*)]
\item In case $\nabla_{\btheta}\ell(\btheta_0;\bx,y)\neq\mathbf{0}$, the log-likelihood ratio follows an asymptotic normality with convergence rate equal to $1/2$. More specifically
\begin{equation}
\label{eq:asymptotic_distribution}
\sqrt{n}\cdot\bigg(\ell(\hat{\btheta}_n;\bx,y) - \ell(\btheta_0;\bx,y)\bigg) \enskip \overset{L}{\to} \enskip \caln\bigg(0, \mbox{\normalfont{tr}}\big[\nabla_{\btheta}\ell(\btheta_0;\bx,y)\cdot\nabla_{\btheta}^\top\ell(\btheta_0;\bx,y)\cdot\bI_q(\btheta_0)^{-1}\big]\bigg).
\end{equation}
\item In case $\nabla_{\btheta}\ell(\btheta_0;\bx,y)=\mathbf{0}$ and $\nabla_{\btheta}^2\ell(\btheta_0;\bx,y)$ is non-singular, the asymptotic distribution of the log-likelihood ratio is a mixture of first-order Chi-square distributions, and the convergence rate is one. More specifically:
\begin{equation}
\label{eq:asymptotic_distribution_zeroder}
n\cdot \bigg(\ell(\hat{\btheta}_n;\bx,y) \enskip - \enskip \ell(\btheta_0;\bx,y)\bigg) \enskip \overset{L}{\to} \enskip \sum_{i=1}^d \lambda_{i}\cdot\chi_1^2
\end{equation}
where $\lambda_i$'s are eigenvalues of $\bI_q(\btheta_0)^{-1/2}\nabla_{\btheta}^2\ell(\btheta_0;\bx,y)\bI_q(\btheta_0)^{-1/2}$.
\end{enumerate}
\end{theorem}
\begin{proof}
Due to assumptions~\ref{item:parameter_dependence} to~\ref{item:openness}, Theorem~\ref{thm:asymptotic_efficiency_MLE} holds and therefore we have $\sqrt{n}\cdot(\hat{\btheta}_n - \btheta_0)\overset{L}{\to}\caln(\mathbf{0},\bI_q(\btheta_0)^{-1})$. The rest of the proof is based on the Delta method in the two modes described in Appendix~\ref{sec:statistical_background} (Theorems~\ref{thm:multivariate_delta_method} and~\ref{thm:delta_method_zeroder}):
\begin{enumerate}[label=(\Roman*)]
\item $\nabla_{\btheta}\ell(\btheta_0;\bx,y)\neq\mathbf{0}$ :
Since the expected log-likelihood function, evaluated at a given pair $(\bx,y)$, is assumed to be continuously differentiable \ref{item:differentiability} and that $\nabla_{\btheta}\ell(\btheta_0;\bx,y)\neq\mathbf{0}$, we can apply Theorem~\ref{thm:multivariate_delta_method} to $\ell(\hat{\btheta}_n;\bx,y) - \ell(\btheta_0;\bx,y)$ to write:
\begin{equation}
\sqrt{n}\cdot \bigg(\ell(\hat{\btheta}_n;\bx,y) \enskip - \enskip \ell(\btheta_0;\bx,y) \bigg) \enskip\overset{L}{\to}\enskip \caln\bigg(0 \enskip,\enskip \nabla_{\btheta}^\top\ell(\btheta_0;\bx,y)\cdot\bI_q(\btheta_0)^{-1}\cdot \nabla_{\btheta}\ell(\btheta_0;\bx,y) \bigg),
\end{equation}
where the scalar variance can also be written in a trace format.
\item $\nabla_{\btheta}\ell(\btheta_0;\bx,y)=\mathbf{0}$ and $\nabla_{\btheta}^2\ell(\btheta_0;\bx,y)$ non-singular :
In this case, the conditions in Theorem~\ref{thm:delta_method_zeroder} are satisfied (with $\bSigma=\bI_q(\btheta_0)^{-1}$ and $g(\btheta)=\ell(\btheta;\bx,y)$), and therefore we can directly write~(\ref{eq:asymptotic_distribution_zeroder}) from equations~(\ref{eq:delta_method_zeroder}).
\end{enumerate}
\end{proof}
Theorem~\ref{thm:asymptotic_distribution} regards the log-likelihood ratio~(\ref{eq:likelihood_ratio}) evaluated at any arbitrary pair $(\bx,y)$. Note that if we consider the training pairs in $\call_n$, which are used to obtain $\hat{\btheta}_n$, it is known that the ratio evaluated at the training set converges to a single first-degree Chi-square distribution, that is
\begin{equation}
\label{eq:asymptotic_distribution_training}
\ell(\hat{\btheta}_n;\call_n) \enskip - \enskip \ell(\btheta_0;\call_n) \enskip \overset{L}{\to} \enskip \frac{1}{2} \chi_1^2
\end{equation}
Theorem~\ref{thm:asymptotic_distribution} implies that variance of the asymptotic distribution of the log-likelihood ratio in case (I) is $\mbox{\normalfont{tr}}\big[\nabla_{\btheta}\ell(\btheta_0;\bx,y)\cdot\nabla_{\btheta}^\top\ell(\btheta_0;\bx,y)\cdot\bI_q(\btheta_0)^{-1}\big]$, whereas in case (II), from Theorem~\ref{thm:delta_method_zeroder} (see Appendix~\ref{sec:statistical_background}), the variance is $\frac{1}{2}\left\|\bI_q(\btheta_0)^{-1/2}\nabla_{\btheta}^2\ell(\btheta_0;\bx,y)\bI_q(\hat{\btheta}_n)^{-1/2} \right\|_F^2$. Therefore, it is evident that the variance of the log-likelihood ratio at any $(\bx,y)$ is reciprocally dependent on the training Fisher information. From this point of view, one can set the training distribution such that it leads to a Fisher information that minimizes this variance. Unless the parameter and hence the Fisher information is univariate, it is not clear what objective to optimize with respect to $q$ such that the resulting Fisher information minimizes the variance.
\subsection{Establishing the Upper Bound}
\label{subsec:upper_bound}
In the next theorem, we show that the Fisher information ratio
, $\mbox{tr}\left[ \bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)\right]$, is a reasonable candidate objective to minimize in order to get a training distribution $q$ for the multivariate case:
\begin{theorem}
\label{thm:FIR_variance_upper_bound}
If the assumptions~\ref{item:parameter_dependence} to~\ref{item:FisherInformation} hold, then:
\begin{equation}
\label{eq:FIR_variance_upper_bound}
\mathbb{E}_p\left[\mbox{\normalfont Var}_q\left(\lim_{n\to\infty}\sqrt{n}\cdot[\ell(\hat{\btheta}_n;\bx,y)-\ell(\btheta_0;\bx,y)]\right) \right] \enskip \leq \enskip \mbox{\normalfont tr}\bigg[\bI_q(\btheta_0)^{-1}\bI_p(\btheta_0)\bigg].
\end{equation}
The equality holds when the set of pairs $(\bx,y)$ where we have zero score function at $\btheta_0$, i.e. $\nabla_{\btheta}\ell(\btheta_0;\bx,y)=\mathbf{0}$, has measure zero under the true joint distribution $P_{\btheta_0}$ in $X\times\{1,...,c\}$.
\end{theorem}
\begin{proof}
Note that, from Theorem~\ref{thm:asymptotic_distribution}, when $\nabla_{\btheta}\ell(\btheta_0;\bx,y)=0$ the convergence rate of the log-likelihood ratio is one and therefore it is of $O_p\left(\frac{1}{n}\right)$. Thus, in this case we have $\sqrt{n}\cdot[\ell(\hat{\btheta}_n;\bx,y)-\ell(\btheta_0;\bx,y)]=O_p\left(\frac{1}{\sqrt{n}}\right)$ and it converges to zero in probability (and in law). Now, define the region $R_0\subseteq X\times\{1,...,c\}$ by
\begin{equation}
R_0 \enskip := \enskip \left\{(\bx,y)| \nabla_{\btheta}\ell(\btheta_0;\bx,y)=0 \right\}
\end{equation}
Variance of the asymptotic distribution of $\sqrt{n}\cdot[\ell(\hat{\btheta}_n;\bx,y)-\ell(\btheta_0;\bx,y)]$, considering both cases $\nabla_{\btheta}\ell(\btheta_0;\bx,y)=0$ (with probability $P_{\btheta_0}(R_0)$) and $\nabla_{\btheta}\ell(\btheta_0;\bx,y)\neq0$ (with probability $1-P_{\btheta_0}(R_0)$), can be written as:
\begin{align}
\label{eq:FIR_variance_upper_bound_noexp}
\mbox{Var}&\left(\lim_{n\to\infty}\sqrt{n}\cdot[\ell(\hat{\btheta}_n;\bx,y)-\ell(\btheta_0;\bx,y)]\right) \nonumber \\
&\equals [1-P_{\btheta_0}(R_0)]\cdot \mbox{\normalfont tr}\big[\nabla_{\btheta}\ell(\btheta_0;\bx,y)\cdot\nabla_{\btheta}^\top\ell(\btheta_0;\bx,y)\cdot\bI_q(\btheta_0)^{-1}\big] \enplus P_{\btheta_0}(R_0)\cdot0 \nonumber\\
&\leq \enskip \mbox{\normalfont tr}\big[\nabla_{\btheta}\ell(\btheta_0;\bx,y)\cdot\nabla_{\btheta}^\top\ell(\btheta_0;\bx,y)\cdot\bI_q(\btheta_0)^{-1}\big]
\end{align}
Taking the expectation of both sides with respect to the true joint, gives the inequality~(\ref{eq:FIR_variance_upper_bound}). If the set of pairs $(\bx,y)$ where $\nabla_{\btheta}\ell(\btheta_0;\bx,y)=\mathbf{0}$ form a zero measure set under $P_{\btheta_0}$, then $P_{\btheta_0}(R_0)=0$ and we get equality in~(\ref{eq:FIR_variance_upper_bound_noexp}) and hence an equality in~(\ref{eq:FIR_variance_upper_bound}).
\end{proof}
Theorem~\ref{thm:FIR_variance_upper_bound} implies that minimizing the Fisher information ratio with respect to $q$, is indeed the upper-bound minimization of the expected variance of the asymptotic distribution of the log-likelihood ratio.
\section{Introduction}
\label{sec:intro}
\input{Intro}
Before going through the main discussion in section~\ref{sec:theoretical_analysis}, we formalize our classification model assumptions, set the notations and review the basics and some of the key properties of our inference method, maximum likelihood estimation, in sections~\ref{sec:framework_assumptions} and~\ref{sec:background}. The statistical background required to follow the remaining sections is given in Appendix~\ref{sec:statistical_background}.
\input{ClassificationModel}
\section{Fisher Information Ratio as an Upper Bound}
\label{sec:theoretical_analysis}
\input{TheoreticalAnalysis}
\section{Fisher Information Ratio in Practice}
\label{subsec:inpractice}
\input{InPractice}
\section{Conclusion}
\label{sec:conclusion}
In this paper, we focused on active learning algorithms in classification problems whose objectives are based on Fisher information criterion. As the primary result, we showed the dependency of the variance of the asymptotic distribution of log-likelihood ratio on the Fisher information of the training distribution. Then, we used this dependency to derive our novel theoretical contribution by establishing the Fisher information ratio (FIR) as the upper bound of such asymptotic variance. We discussed that several layers of approximations can be employed in practice to simplify FIR; simplifications, that can usually be avoided in pool-based active learning. Finally Monte-Carlo simulations and greedy algorithms can be used to evaluate and optimize the (simplified) FIR objective, respectively. Using this framework, we can distinguish the main differences between some of the FIR-based querying methods in the classification context. Such comparative analysis, not only shed light on the assumptions and simplifications of the existing algorithms, it can also be helpful for finding suitable directions in developing novel active learning algorithms based on the Fisher information criterion.
\begin{appendices}
|
\section{Introduction}\label{sec:intro}
Suppose that the observed data consists of $\mathbf{z}_i = (y_i,\mathbf{x}_i)$, $i = 1,\dotsc,n$, where $y_i \in \cY
\subseteq \mathbb{R}$ and $\mathbf{x}_i \in \cX \subseteq \mathbb{R}^d$.
Suppose further that $\mathbf{z}_1,\dotsc,\mathbf{z}_n \sim \rho$ are iid from some probability distribution $\rho$
on $\cY \times \cX$.
Let $\rho(\cdot\mid\mathbf{x})$ denote the conditional distribution of $y_i$ given $\mathbf{x}_i = \mathbf{x} \in \cX$ and
let $\marginal$ denote the marginal distribution of $\mathbf{x}_i$.
Our goal is to use the available data to estimate the regression function of $y$ on $\mathbf{x}$,
\[
\reg(\mathbf{x}) \ = \ \int_{\cY} y \ \dif\rho(y\mid \mathbf{x})
\,,
\]
which minimizes the mean-squared prediction error
\[
\int_{\cY\times \cX}
\Parens{ y - f(\mathbf{x}) }^2 \ \dif\rho(y,\mathbf{x})
\]
over $\marginal$-measurable functions $f \colon \cX \to \mathbb{R}$.
More specifically, for an estimator $\est$ define the risk
\begin{align}\label{risk}
\risk(\est)
& \ = \ \mathbb{E}\Brackets{ \int \Parens{ \reg(\mathbf{x}) - \est(\mathbf{x}) }^2 \ \dif\marginal(\mathbf{x}) }
\ = \ \mathbb{E}\Brackets{ \, \norm{\reg - \est}^2_{\marginal} } \,,
\end{align}
where the expectation is computed over $\mathbf{z}_1,\dotsc,\mathbf{z}_n$, and ${\norm{\cdot}_{\marginal}}$ denotes
the norm on $L^2(\marginal)$; we seek estimators $\est$ which minimize $\risk(\est)$.
This is a version of the random design nonparametric regression problem.
There is a vast literature on nonparametric regression, along with a huge variety of corresponding
methods~\citep[e.g.,][]{gyorfi2002distribution,wasserman2006all}.
In this paper, we focus on regularization and kernel methods for estimating $\reg$.\footnote{Here, we mean
``kernel'' as in \emph{reproducing kernel Hilbert space}, rather than
\emph{kernel-smoothing}, which is another popular approach to nonparametric regression.}
Most of our results apply to general regularization operators.
However, our motivating examples are two well-known regularization techniques: Kernel ridge
regression (which we refer to as ``KRR''; KRR is also known as Tikhonov regularization) and kernel
principal component regression (``KPCR''; also known as spectral cut-off regularization).
Our main theorem is a new upper bound on the risk of a general class of kernel-regularization
methods, which includes both KRR and KPCR (Theorem~\ref{thm:UB}).
The theorem substantially generalizes previously published bounds (see Section \ref{sec:related} for
a discussion of related work) and illustrates the dependence of the risk on three important
features: (i) the structure of the ambient reproducing kernel Hilbert
space (RKHS), (ii) the specific regularization technique
employed, and (iii) the regularity (often interpreted as smoothness) of the function to be
estimated.
One
consequence of the theorem is that the regularization methods studied in this paper
(including KRR and KPCR) achieve the minimax rate for estimating $\reg$ in a variety of settings.
A second consequence is that certain regularization methods (including
KPCR, but not KRR) may adapt to
favorable regularity of $\reg$ to attain even faster convergence rates, while others (notably KRR)
are limited in this regard due to a well-known \emph{saturation}
effect~\citep{neubauer1997converse,mathe2004saturation,bauer2007regularization}.
This illustrates a striking advantage that KPCR may have over KRR in these settings.
\section{Related work}\label{sec:related}
Kernel ridge regression has been studied extensively in the literature.
Indeed, bounds for KRR that are similar to our Theorem \ref{thm:UB} have been derived by
\citet{caponnetto2007optimal}.
Moreover, it is well-known that KRR is minimax in many of the settings considered in this paper,
such as those described in Corollaries \ref{cor:poly}--\ref{cor:gauss}~\citep{caponnetto2007optimal, zhang2005learning, zhang2013divide}.
However, these cited results apply only to KRR, while the results in this paper apply to a
substantially larger class of regularization operators (including, for example, KPCR).
Beyond KRR, there has also been significant research into more general regularization methods, like
those considered in this paper.
However, our bounds are sharper than previously published results on general regularization
operators.
For instance, unlike our Theorem \ref{thm:UB}, the bounds of \citet{bauer2007regularization} do not
illustrate the dependence of the risk on the ambient Hilbert space.
Thus, while our approach immediately implies that many of the regularization methods under
consideration are minimax optimal, it seems difficult (if not impossible) to draw this conclusion
using the approach of \citeauthor{bauer2007regularization}.
General regularization operators are studied by~\citet{caponnetto2010cross}, but their results
require a semi-supervised setting where an additional pool of unlabeled data is available.
One of the major practical implications of this paper is that KPCR may have significant advantages
over KRR in some settings.
This has been observed previously by other researchers; others have even noted that Tikhonov
regularization (KRR) saturates, while spectral cut-off regularization (KPCR) does not
\citep{mathe2004saturation, bauer2007regularization, logerfo2008spectral}.
Our results (Theorem \ref{thm:UB} and Corollaries \ref{cor:poly}--\ref{cor:gauss}) sharpen these observations by precisely
quantifying the advantages of unsaturated regularization operators in terms of adaptability and
minimaxity.
In other related work, \citet{dhillon2013risk} have illustrated the potential advantages of KPCR
over KRR in finite-dimensional problems with linear kernels; though their work is not framed in
terms of saturation and general regularization operators, it relies on similar concepts.
We recently became aware of simultaneous independent work of \citet{blanchard2016optimal} that
proves upper bounds on the risk for a comparable class of regularization methods (which includes KRR
and KPCR), as well as minimax lower bounds that match the upper bounds; their results are
specialized to RKHS's where the covariance operator has polynomially decaying eigenvalues.
Compared to that work, our results require a weaker moment condition on the noise for risk bounds, apply to a much broader class of RKHS's, and also consider target functions that live in
finite-dimensional subspaces (see Proposition~\ref{prop:finite_adapt}). Another
distinguishing feature of the present work is that it contains numerical
experiments that illustrate the implications of our theoretical
bounds in some practical settings (Section \ref{sec:experiments}).
The main engine behind the technical results in this paper is a collection of large-deviation
results for Hilbert-Schmidt operators.
The required machinery is developed in the appendix.
These results build on straightforward extensions of results of \citet{tropp2015intro} and
\citet{minsker2011bernstein}.
Our most precise results for KPCR and achieving parametric rates for estimation over
finite-dimensional subspaces (Proposition~\ref{prop:finite_adapt}) rely on slightly different
arguments, which are based on well-known eigenvalue perturbation results that have been adapted to
handle Hilbert-Schmidt operators (e.g., the Davis-Kahan $\sin \Theta$ theorem
\citep{davis1970rotation}).
\section{Statistical setting and assumptions}
Our basic assumption on the distribution of $\mathbf{z} = (y,\mathbf{x}) \sim \rho$ is that the residual variance is
bounded; more specifically, we assume that there exists a constant $\sigma^2 > 0$ such that
\begin{equation}\label{resid}
\int_{\cY} \left[ y - \reg(\mathbf{x}) \right]^2 \ d\rho(y\mid\mathbf{x}) \ \leq \ \sigma^2
\end{equation}
for almost all $\mathbf{x} \in \cX$.
\citet{zhang2013divide} also assume~\eqref{resid}; this assumption is slightly weaker than the
analogous assumption of \citet{bauer2007regularization} (Equation (1) in their paper).
Note that~\eqref{resid} holds if $y$ is bounded almost surely or if $y
= f^{\dagger}(\mathbf{x}) + \epsilon$, where $\epsilon$ is independent of $\mathbf{x}$, $\mathbb{E}(\epsilon) = 0$, and $\mathrm{Var}(\epsilon) = \sigma^2$.
Let $K \colon \cX \times \cX \to \mathbb{R}$ be a symmetric positive-definite kernel function.
We assume that $K$ is bounded---i.e., that there exists $\kappa^2 > 0$
such that $\sup_{\mathbf{x} \in \cX} K(\mathbf{x},\mathbf{x}) \ \leq \ \kappa^2$.
Additionally, we assume that there is a countable basis of eigenfunctions $\braces{\psi_j}_{j =
1}^{\infty} \subseteq L^2(\marginal)$ and a sequence of corresponding eigenvalues $t_1^2 \geq t_2^2
\geq \dotsb \geq 0$ such that
\begin{equation}\label{mercer}
K(\mathbf{x},\tilde\mathbf{x}) \ = \ \sum_{j = 1}^{\infty} t_j^2\psi_j(\mathbf{x})\psi_j(\tilde\mathbf{x}) \,, \quad \mathbf{x},\tilde\mathbf{x} \in \cX
\end{equation}
and the convergence is absolute.
Mercer's theorem and various generalizations give conditions under which representations
like~\eqref{mercer} are known to hold \citep{carmeli2006vector}; one of the simplest examples is
when $\cX$ is a compact Hausdorff space, $\marginal$ is a probability measure on the Borel sets of
$\cX$, and $K$ is continuous.
Observe that
\begin{align*}
\sum_{j = 1}^{\infty} t_j^2
& \ = \ \sum_{j = 1}^{\infty} t_j^2 \int_{\cX} \psi_j(\mathbf{x})^2 \ \dif\marginal(\mathbf{x})
\ = \ \int_{\cX} K(\mathbf{x},\mathbf{x}) \ \dif\marginal(\mathbf{x})
\ \leq \ \kappa^2 \,;
\end{align*}
in particular, $\braces{t_j^2} \in \ell^1(\mathbb{N})$.
Let $\mathcal{H} \subseteq L^2(\marginal)$ be the RKHS corresponding to
$K$ \citep{aronszajn1950theory} and let $\phi_j \ = \ t_j\psi_j$, $j = 1,2,\ldots$.
It follows from basic facts about RKHSs that $\braces{\phi_j}_{j = 1}^{\infty}$ is an orthonormal
basis for $\mathcal{H}$ (if $t_J^2 > t_{J+1}^2 = 0$, then $\Braces{\phi_j}_{j=1}^J$ is an orthonormal basis for
$\mathcal{H}$).
Furthermore, $\mathcal{H}$ is characterized by
\[
\mathcal{H} \ = \ \Braces{
f = \sum_{j = 1}^{\infty} \theta_j\psi_j \in L^2(\marginal);
\ \sum_{j = 1}^{\infty} \frac{\theta_j^2}{t_j^2} < \infty
}
\]
and the inner product
\[
\dotp{ f, \tilde f }_{\mathcal{H}}
\ = \
\Dotp{
\sum_{j = 1}^{\infty} \theta_j\psi_j,\sum_{j = 1}^{\infty} \tilde{\theta}_j\psi_j
}_{\mathcal{H}}
\ = \
\sum_{j = 1}^{\infty} \frac{\theta_j\tilde{\theta}_j}{t_j^2}
\,
\]
(the corresponding norm is denoted by $\norm{\cdot}_{\mathcal{H}}$).
Our main assumption on the relationship between $y$, $\mathbf{x}$, and the kernel $K$ is that
\begin{equation}\label{reg}
\reg \ \in \ \mathcal{H}
\,.
\end{equation}
This is a regularity or smoothness assumption on $\reg$. Many of the results in this paper can be modified, so that they apply to settings where $\reg \notin
\mathcal{H}$, by replacing $\reg$ with an appropriate projection of $\reg$ onto $\mathcal{H}$ and including an
approximation error term in the corresponding bounds.
This approach leads to the study of {\it oracle inequalities} \citep{zhang2005learning,zhang2013divide,
koltchinskii2006local, steinwart2009optimal, hsu2014random}, which we do not pursue
in detail here.
However, investigating oracle inequalities for general regularization operators may be of interest
for future research, as most existing work focuses on ridge regularization.
Another interpretation of condition \eqref{reg} is that it is a minimal regularity condition on
$\reg$ for ensuring that $\reg$ can be estimated consistently using the kernel methods considered
below. One key aspect of the upper bounds in Section \ref{sec:main} is that they
show $\reg$ can be estimated more efficiently, if it satisfies
stronger regularity conditions (and if the regularization method used
is sufficiently adaptable).
A convenient way to formulate a collection of regularity conditions
with varying strengths, which will be useful in the sequel, is as follows.
For $\zeta \geq 0$, define the Hilbert space
\begin{equation}\label{Hzeta}
\mathcal{H}_\zeta
\ = \
\Braces{
f = \sum_{j = 1}^{\infty} \theta_j\psi_j \in L^2(\marginal);
\ \sum_{j = 1}^{\infty} \frac{\theta_j^2}{t_j^{2(1+\zeta)}} < \infty
}
\,.
\end{equation}
Then $\mathcal{H} = \mathcal{H}_0$ and $\mathcal{H}_{\zeta_2} \subseteq \mathcal{H}_{\zeta_1}$ whenever $\zeta_2 \geq \zeta_1 \geq 0$.
The norm on $\mathcal{H}_{\zeta}$ is defined by $\norm{ f }_{\mathcal{H}_{\zeta}}^2
\ = \
\sum_{j = 1}^{\infty} \frac{\theta_j^2}{t_j^{2(1+\zeta)}}$.
Additionally, positive integers $J$ define the finite-rank subspace
\begin{equation}\label{HJ}
\mathcal{H}_J^\circ
\ = \
\Braces{
f = \sum_{j = 1}^{J} \theta_j\psi_j \in L^2(\marginal);
\ \theta_1,\ldots,\theta_J \in \mathbb{R}
}
\,.
\end{equation}
We have the inclusion $\mathcal{H}_J^\circ \subseteq \mathcal{H}_{J+1}^\circ$ and, if $t_1^2,\ldots,t_J^2 > 0$, then
$\mathcal{H}_J^\circ \subseteq \mathcal{H}_\zeta$ for any $\zeta \geq 0$.
In particular, it is clear from \eqref{Hzeta}--\eqref{HJ} that $\reg \in \mathcal{H}_{\zeta}$ is a stronger regularity condition than $\reg \in \mathcal{H}$ and that $\reg \in \mathcal{H}_J^{\circ}$ is an even stronger condition (provided the $t_j^2$ are
strictly positive).
Conditions such as $\reg \in \mathcal{H}_{\zeta}$ and $\reg \in \mathcal{H}_J^{\circ}$ are known as {\em source
conditions} elsewhere in the literature \citep[e.g.,][]{bauer2007regularization,caponnetto2010cross}.
\section{Regularization}
As discussed in Section~\ref{sec:intro}, our goal is find estimators $\est$ that minimize the
risk~\eqref{risk}.
In this paper, we focus on regularization-based estimators for $\reg$.
In order to precisely describe these estimators, we require some additional notation for various
operators that will be of interest, and some basic definitions from regularization theory.
\subsection{Finite-rank operators of interest}
For $\mathbf{x} \in \cX$, define $K_{\mathbf{x}} \in \mathcal{H}$ by $K_{\mathbf{x}}(\tilde{\mathbf{x}}) = K(\mathbf{x},\tilde{\mathbf{x}})$, $\tilde{\mathbf{x}} \in
\cX$. Let $X = (\mathbf{x}_1,\dotsc,\mathbf{x}_n)^\top \in \mathbb{R}^{n \times d}$ and $\mathbf{y} = (y_1,\dotsc,y_n)^\top \in
\mathbb{R}^n$.
Additionally, define the finite-rank linear operators $S_X: \mathcal{H} \to \mathbb{R}^n$ and $T_X: \mathcal{H} \to \mathcal{H}$ (both
depending on $X$) by
\begin{align*}
S_X\phi
& \ = \
(\dotp{ \phi,K_{\mathbf{x}_1} }_{\mathcal{H}}, \dotsc, \dotp{ \phi,K_{\mathbf{x}_n} }_{\mathcal{H}})^{\top}
\ = \ (\phi(\mathbf{x}_1), \dotsc, \phi(\mathbf{x}_n))^{\top} \,, \\
T_X\phi
& \ = \
\frac{1}{n}\sum_{i = 1}^n \dotp{ \phi,K_{\mathbf{x}_i} }_{\mathcal{H}} K_{\mathbf{x}_i}
\ = \ \frac{1}{n} \sum_{i = 1}^n \phi(\mathbf{x}_i) K_{\mathbf{x}_i}
\,,
\end{align*}
where $\phi \in \mathcal{H}$.
Let $\dotp{\cdot,\cdot}_{\mathbb{R}^n}$ denote the normalized inner-product on $\mathbb{R}^n$, defined by $\dotp{
\bfv,\tilde\bfv }_{\mathbb{R}^n} = n^{-1} \bfv^\top\tilde\bfv$ for $\bfv = (v_1,\dotsc,v_n)^\top, \
\tilde\bfv = (\tilde{v}_1,\dotsc,\tilde{v}_n)^\top \in \mathbb{R}^n$.
Then the adjoint of
$S_X$ with respect to $\dotp{\cdot,\cdot}_{\mathcal{H}}$ and
$\dotp{\cdot,\cdot}_{\mathbb{R}^n}$, $S_X^*: \mathbb{R}^n \to \mathcal{H}$, is given by
$S_X^*\bfv
\ = \
\frac{1}{n}\sum_{i = 1}^n v_iK_{\mathbf{x}_i}$.
Additionally, we have $T_X = S_X^*S_X$.
Finally, observe that $S_XS_X^* \colon \mathbb{R}^n \to \mathbb{R}^n$ is given by the $n \times n$ matrix $S_XS_X^*
= n^{-1}\mathbf{K}$, where $\mathbf{K} = (K(\mathbf{x}_i,\mathbf{x}_j))_{1 \leq i,j\leq n}$; $\mathbf{K}$ is the {\it kernel matrix}, which
is ubiquitous in kernel methods and enables finite computation.
\subsection{Basic definitions}
A family of functions $\linv \colon \intco{0,\infty} \to \intco{0,\infty}$ indexed by $\lambda > 0$ is
called a \emph{regularization family} if it satisfies the following three conditions:\footnote{%
This definition follows \citet{engl1996regularization} and \citet{bauer2007regularization}, but is
slightly more restrictive.%
}
\begin{description}
\item[R1] $\sup_{0 < t \leq \kappa^2} |t \linv(t)| < 1$.
\item[R2] $\sup_{0 < t \leq \kappa^2} |1 - t \linv(t)| \leq 1$.
\item[R3] $\sup_{0 < t \leq \kappa^2} | \linv(t)| < \lambda^{-1}$.
\end{description}
The main idea behind a regularization family is that it ``looks'' similar to $t \mapsto 1/t$, but is
better-behaved near $t = 0$, i.e., it is bounded by $\lambda^{-1}$.
An important quantity that is related to the adaptability of a regularization
family is the {\em qualification} of the regularization.
The qualification of the regularization family $\braces{\linv}_{\lambda > 0}$ is defined to be the
maximal $\xi \geq 0$ such that
\[
\sup_{0< t \leq \kappa^2}
\abs{ 1 - t \linv(t) }
t^{\xi}
\ \leq \
\lambda^{\xi}
\,.
\]
If a regularization family has qualification $\xi$, we say that it ``saturates at $\xi$.''
Two regularization families that are the major motivation for the results in this paper are ridge
(Tikhonov) regularization, where $\linv(t) = \ \frac1{t+\lambda}$
and principal component (spectral cut-off) regularization, where
\begin{equation}\label{PCA}
\linv(t) \ = \ \sinv(t) \ = \ \frac1t \ind{ t \geq \lambda }
\,.
\end{equation}
Observe that ridge regularization has qualification 1 and principal
component regularization has qualification $\infty$.
Another example of a regularization family with qualification $\infty$
is the Landweber iteration, which can be viewed as a special case of
gradient descent~\citep[see,
e.g.,][]{rosasco2005spectral,bauer2007regularization,logerfo2008spectral}.
\subsection{Estimators}
\label{sec:est}
Given a regularization family $\braces{\linv}_{\lambda > 0}$, we define the $\linv$-regularized estimator
for $\reg$,
\begin{equation}\label{regest}
\estl
\ = \
\linv(T_X) S_X^* \mathbf{y}
\,.
\end{equation}
Here, $\linv$ acts on the spectrum (eigenvalues) of the finite-rank operator $T_X$ (which is the
same as the spectrum of the kernel matrix $\mathbf{K}$, up to scaling).
Therefore, a finitely-computable representation is $\estl = \sum_{i=1}^n \hat\gamma_i
K_{\mathbf{x}_i}$, where $(\hat\gamma_1,\dotsc,\hat\gamma_n)^\top = \linv(\mathbf{K}/n) \mathbf{y} / n$; computing the
$\hat\gamma_i$ involves an eigenvalue decomposition of the matrix $\mathbf{K}$.
The dependence of $\estl$ on the regularization family is implicit; our results hold for any
regularization family except where explicitly stated otherwise.
The estimators $\estl$ are the main focus of this paper.
\section{Main results}\label{sec:main}
\subsection{General bound on the risk}
\begin{theorem}\label{thm:UB}
Let $\estl$ be the estimator defined in \eqref{regest} with regularization family
$\braces{\linv}_{\lambda>0}$.
Let $0 \leq \delta \leq 1$ and assume that there is some $\kappa_{\delta}^2 > 0$ such that
\begin{equation}\label{psibd}
\sup_{\mathbf{x} \in \cX}
\sum_{j = 1}^{\infty} t_j^{2(1 - \delta)} \psi_j(\mathbf{x})^2
\ \leq \
\kappa_{\delta}^2
\ < \
\infty
\end{equation}
Assume that the source condition $\reg \in \mathcal{H}_{\zeta}$ holds for some $\zeta \geq 0$, and that $\linv$
has qualification at least $\max\braces{(\zeta+1)/2,1}$.
Define the effective dimension $\dl =
\sum_{j=1}^\infty
\frac{t_j^2}{t_j^2 + \lambda}$.
Finally, assume that $(8/3+2\sqrt{5/3}) \kappa_{\delta}^2 / n \leq \lambda^{1-\delta} \leq \kappa_{\delta}^2$.
The following risk bound holds:
\begin{align}\label{UB}
\risk(\estl)
& \ \leq \
2^{\zeta+3}\norm{\reg}_{\mathcal{H}_\zeta}^2 \lambda^{\zeta+1}
+ \frac{4\dl\sigma^2}{n}
\\ \nonumber
& \qquad
+ 4\dl \Parens{
\norm{\reg}_{\mathcal{H}}^2 t_1^2
+ \frac{\kappa^2\sigma^2}{\lambda n}
}
\exp\Biggl(-\frac{3\lambda^{1-\delta}n}{28\kappa_{\delta}^2}\Biggr)
\\ \nonumber
& \qquad
+ \ind{\zeta>1} \cdot 16\zeta^2(3/2)^{\zeta-1}
\cdot \norm{\reg}_{\mathcal{H}_\zeta}^2
(t_1^2+\lambda)^{\zeta-1} \kappa^4
\Biggl(
\frac{34}{n} + \frac{15}{n^2}
\Biggr)
\,.
\end{align}
\end{theorem}
Theorem \ref{thm:UB} is proved in Appendix~\ref{sec:proof-UB}. The
first two terms in the upper bound \eqref{UB} are typically the dominant terms.
In the upper bound \eqref{UB}, the interaction between the kernel $K$ and the distribution $\marginal$ is reflected in the
\emph{effective dimension} $\dl$~\citep[see, e.g.,][]{zhang2005learning,caponnetto2007optimal}.
The regularity of $\reg$ enters through norm of $\reg$ (both the $\mathcal{H}$-
and $\mathcal{H}_{\zeta}$-norms) and the
exponent on $\lambda$.
The
condition \eqref{psibd} in Theorem \ref{thm:UB} is always satisfied by taking $\delta = 0$ and
$\kappa_0^2 = \kappa^2$. Requiring \eqref{psibd} with $\delta > 0$
imposes additional conditions on the RKHS $\mathcal{H}$. For Corollary
\ref{cor:poly} below (which applies when the eigenvalues $\{t_j^2\}$
have polynomial-decay), we take $\delta = 0$ and
$\kappa_0^2 = \kappa^2$. The stronger condition with $\delta > 0$ is required
to obtain obtain minimax rates for
kernels where the eigenvalues $\{t_j^2\}$ have exponential or
Gaussian-type decay (see Corollaries \ref{cor:exp}--\ref{cor:gauss}).
Risk bounds on general regularization
estimators similar to $\estl$ were previously obtained by
\citet{bauer2007regularization}.
However, their bounds \citep[e.g., Theorem 10 in][]{bauer2007regularization} are independent of the
ambient RKHS $\mathcal{H}$, i.e., they do not depend on the eigenvalues $\braces{t_j^2}$.
Our bounds are tighter than those of \citet{bauer2007regularization} because we take advantage of
the structure of $\mathcal{H}$.
In contrast with our Theorem \ref{thm:UB}, the results of \citet{bauer2007regularization} do not give
minimax bounds (not easily, at least), because minimax rates must
depend on the $t_j^2$.
\subsection{Implications for kernels characterized by their eigenvalues' rate of
decay}
We now state consequences of Theorem~\ref{thm:UB} that give explicit rates for estimating $\reg$ via
$\estl$, for any regularization family, under specific assumptions about the decay rate of the
eigenvalues $\{t_j^2\}$.
We first consider the case where the eigenvalues have polynomial decay.
\begin{corollary}\label{cor:poly}
Assume that $C \geq 0$ and $\nu > 1/2$ are constants such that $0 < t_j^2 \leq Cj^{-2\nu}$ for all
$j = 1,2,\ldots$.
Assume the source condition $\reg \in \mathcal{H}_{\zeta}$ for some $\zeta \geq 0$, and that $\linv$
has qualification at least $\max\braces{(\zeta+1)/2,1}$.
Finally, take $\lambda = C' n^{-\frac{2\nu}{2\nu(\zeta + 1) + 1}}$
for a suitable constant $C'>0$ so that the conditions on $\lambda$ from Theorem~\ref{thm:UB} are satisfied.
Then
\begin{align*}
\risk(\estl)
& \ = \
O\Parens{
\Braces{ \norm{\reg}_{\mathcal{H}_\zeta}^2 + \sigma^2 }
\cdot
n^{-\frac{2\nu(\zeta+1)}{2\nu(\zeta + 1) + 1}}
}
\,,
\end{align*}
where the constants implicit in the big-$O$ may depend on $\kappa^2$, $C$, $C'$, $\nu$, and
$\zeta$, but nothing else.
\end{corollary}
\begin{remark}
Observe that if $\linv$ has qualification at least $\max\braces{(\zeta+1)/2,1}$ (and the other
conditions of Corollary \ref{cor:poly} are met), then $\hat{f}_{\lambda}$ obtains the minimax rate for
estimating functions over $\mathcal{H}_{\zeta}$ \citep{pinsker1980optimal}.
Thus, if $g_{\lambda}$ has higher qualification, then $\hat{f}_{\lambda}$
can effectively adapt to a broader range of subspaces
$\mathcal{H}_{\zeta} \subseteq \mathcal{H} = \mathcal{H}_0$. In particular, KPCR (with infinite qualification) can
adapt to source conditions with arbitrary $\zeta \geq
0$; on the other hand, KRR satisfies the conditions of Corollary \ref{cor:poly} only when $\zeta
\leq 1$, because KRR has qualification 1.
\end{remark}
\begin{remark}
As mentioned earlier, a very similar result for polynomial decay eigenvalues was independently and
simultaneously obtained by \citet{blanchard2016optimal} for essentially the same class of
regularization operators. Our Theorem \ref{thm:UB}, from
which the corollary follows, applies to a broader
class of kernels than results of \citeauthor{blanchard2016optimal}.
\end{remark}
When the eigenvalues $\braces{t_j^2}$ have exponential or Gaussian-type decay, the rates are nearly
the same as in finite dimensions.
\begin{corollary}\label{cor:exp}
Assume that $C, \alpha \geq 0$ are constants such that $0 < t_j^2 \leq Ce^{-\alpha j}$ for all
$j = 1,2,\ldots$.
Assume that $\linv$
has qualification at least $1$ and that \eqref{psibd} holds for any $0 < \delta \leq 1$.
Finally, take $\lambda = C'n^{-1}\log(n)$ for a suitable constant $C'>0$ so that the conditions on $\lambda$
from Theorem~\ref{thm:UB} are satisfied.
Then
\begin{align*}
\risk(\estl)
& \ = \
O\Parens{
\Braces{ \norm{\reg}_{\mathcal{H}}^2 + \sigma^2 }
\cdot
\frac{\log(n)}{n}
}
\,,
\end{align*}
where the constants implicit in the big-$O$ may depend on $\kappa^2$, $C$, $C'$, $\alpha$, $\delta$,
and $\kappa_{\delta}^2$, but nothing else.
\end{corollary}
\begin{corollary}\label{cor:gauss}
Assume that $C, \alpha \geq 0$ are constants such that $0 < t_j^2 \leq Ce^{-\alpha j^2}$ for all
$j = 1,2,\ldots$.
Assume that $\linv$
has qualification at least $1$ and that \eqref{psibd} holds for any $0 < \delta \leq 1$.
Finally, take $\lambda = C'n^{-1}\sqrt{\log(n)}$ for a suitable constant $C'>0$ so that the conditions
on $\lambda$ from Theorem~\ref{thm:UB} are satisfied.
Then
\begin{align*}
\risk(\estl)
& \ = \
O\Parens{
\Braces{ \norm{\reg}_{\mathcal{H}}^2 + \sigma^2 }
\cdot
\frac{\sqrt{\log(n)}}{n}
}
\,,
\end{align*}
where the constants implicit in the big-$O$ may depend on $\kappa^2$, $C$, $C'$, $\alpha$, $\delta$,
and $\kappa_{\delta}^2$, but nothing else.
\end{corollary}
\begin{remark}
In Corollaries \ref{cor:exp} --\ref{cor:gauss}, we get minimax
estimation over $\mathcal{H} = \mathcal{H}_0$ \citep{pinsker1980optimal}.
However, our bounds are not refined enough to pick up any potential improvements which may be had
if a stronger source condition is satisfied (e.g., $\reg \in \mathcal{H}_{\zeta}$ for $\zeta > 0$) .
This is typical in settings like this because the minimax rate is already quite fast, i.e., within
a log-factor of the parametric rate $n^{-1}$.
\end{remark}
\subsection{Parametric rates for finite-dimensional kernels and subspaces}
\label{sec:finite}
If the kernel has finite rank (i.e., $t_j^2 = 0$ for $j$ sufficiently large), then it follows directly from Theorem
\ref{thm:UB} that
$\risk(\estl) = O\parens{
\braces{ \norm{\reg}_{\mathcal{H}}^2 + \sigma^2 } / n
}$.
If the kernel has infinite rank, but $\reg $ is contained in the finite-dimensional subspace
$\mathcal{H}_J^{\circ}$ for some
$J < \infty$, then Theorem \ref{thm:UB}
can still be applied, provided $\linv$ has high qualification. Indeed, if $\linv$
has infinite qualification and $\reg \in
\mathcal{H}_J^{\circ}$, then it follows that $\reg \in \mathcal{H}_{\zeta}$ for all $\zeta
\geq 0$ and Theorem \ref{thm:UB} implies that for any $0 < \alpha \leq 1$,
$\risk(\estl)
=
O\parens{
\braces{ \norm{\reg}_{\mathcal{H}}^2 + \sigma^2 } / n^{1-\alpha}
}$ for appropriately chosen $\lambda$.
In fact, we can improve on this rate for KPCR.
The next proposition implies that the risk of KPCR matches the
parametric rate $n^{-1}$ for $\reg \in \mathcal{H}_J^{\circ}$; the proof requires a different
argument, based on eigenvalue perturbation theory, which we give in
Appendix~\ref{sec:proof-finite_adapt}.
\begin{proposition}\label{prop:finite_adapt}
Let $\kpcr$ be the KPCR estimator, with principal component regularization
\eqref{PCA}, and assume that $\reg \in \mathcal{H}_J^{\circ}$.
Let $0 < r < 1$ be a constant and let $\lambda = (1-r)t_J^2$.
If $rt_J^2 \geq \kappa^2/n^{1/2} + \kappa^2/(3n)$, then
\begin{align*}
\risk(\kpcr)
& \ \leq \
\frac{1}{n}\cdot
\Braces{
\frac{34\kappa^6}{r^2t_J^4}
\norm{\reg}^2_{\mathcal{H}} + \frac{3\kappa^2}{(1-r)t_J^2}\sigma^2
}
+ \kappa^2 \norm{\reg}{\mathcal{H}}^2
\Braces{
\frac{15\kappa^4}{n^2r^2t_J^4}
+ 4\exp\Parens{
-\frac{nr^2t_J^4}{2\kappa^4 + 2\kappa^2rt_J^2/3}
}
}
\,.
\end{align*}
\end{proposition}
Proposition~\ref{prop:finite_adapt} implies that KPCR may reach the parametric rate for estimating $\reg
\in\mathcal{H}_J^{\circ}$. On the other hand, it is known that KRR may perform dramatically worse than KPCR in these
settings due
to the saturation effect \citep[see, e.g.,][]{caponnetto2007optimal,
dicker2016ridge, dhillon2013risk}.
\section{Numerical experiments}\label{sec:experiments}
\subsection{Simulated data}
This simulation study shows how KPCR is able to adapt to
highly structured signals, while KRR requires more favorable structure
from the ambient RKHS.
For this experiment, we take $\cX = \{1,2,\ldots,2^{13}\}$. The data distribution
$\rho$ on $\cY \times \cX$ is specified as follows.
The marginal distribution on $\cX$ is $\marginal(x) \propto x^{-1/2}$, the
regression function $\reg$ is given by $\reg(x) = \sum_{j=1}^5 \ind{x=j}$, and $\rho(\cdot\mid x)$
is normal with mean $\reg(x)$ and variance $1/4$. To compute
$\hat{f}_{\lambda}$, we use the discrete kernel $K(x,\tilde x) = \ind{x = \tilde x}$
Using an iid~sample of size $n=2^{13}$, we compute $\estl$ (either KRR or KPCR) for $\lambda$ in a discrete
grid of $2^{10}$ values uniformly spaced between $10^{-5}$ and $0.02$, and then choose the value of
$\lambda$ for which $\estl$ has smallest \emph{validation mean-squared error}
$n^{-1} \sum_{i=1}^n (y_i^v-\estl(x_i^v))^2$,
computed using a separate iid~sample $\braces{ (x_i^v,y_i^v)
}_{i=1}^n$ of size $n=2^{13}$.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.47\textwidth]{expt6_validmse.pdf} &
\includegraphics[width=0.47\textwidth]{expt6_mse.pdf}
\\
(a) & (b)
\\
\includegraphics[width=0.49\textwidth]{expt6_fhat.pdf} &
\includegraphics[width=0.47\textwidth]{alphas.pdf}
\\
(c) & (d)
\end{tabular}
\end{center}
\caption{\label{fig:mse}
(a) Validation-MSE of $\krr$ and $\kpcr$
for $\marginal(x) \propto x^{-1/2}$ as $\lambda$ varies;
(b) $\marginal$-MSE of $\krr$ and $\kpcr$
for $\marginal(x) \propto x^{-1/2}$ as $\lambda$ varies;
(c) estimated functions $\est_{\KRR,\hat\lambda_{\KRR}}$ and
$\est_{\KPCR,\hat\lambda_{\KPCR}}$ for $\marginal(x) \propto x^{-1/2}$;
(d) $\marginal$-MSE of $\est_{\KRR,\hat\lambda_{\KRR}}$ and
$\est_{\KPCR,\hat\lambda_{\KPCR}}$ for $\marginal(x) \propto x^{-\alpha}$ as $\alpha$
varies.
}
\end{figure}
Figure~\ref{fig:mse}(a) shows the validation-MSE of each $\estl$; the plot of \emph{$\marginal$-MSE}
$\norm{\reg-\estl}_{L^2(\marginal)}^2$ has roughly the same shape, just shifted down by $\sigma^2=1/4$
(Figure~\ref{fig:mse}(b)).
The selected $\lambda$ is $\hat\lambda_{\KRR} = 0.001534$ for KRR, and $\hat\lambda_{\KPCR} = 0.001573$ for KPCR.
These choices of $\lambda$ yield the final estimators, $\est_{\KRR,\hat\lambda_{\KRR}}$ and
$\est_{\KPCR,\hat\lambda_{\KPCR}}$; the $\marginal$-MSE is $0.0034$ for KRR, and $0.0003$ for KPCR.
In Figure~\ref{fig:mse}(c),
we plot the functions $\est_{\KRR,\hat\lambda_{\KRR}}$ and
$\est_{\KPCR,\hat\lambda_{\KPCR}}$; the KRR function is non-zero for much of the domain, while the KPCR
function is zero for nearly all of the domain (like $\reg$).
We repeat the above simulation for different marginal distributions $\marginal(x) \propto x^{-\alpha}$,
for $1/2 \leq \alpha \leq 2$, which imply different eigenvalue sequences $\{t_j^2\}$.
The mean and standard deviation of the $\marginal$-MSE's over $10$ repetitions are shown in
Figure~\ref{fig:mse}(b).
This confirms KPCR's to adapt to the regularity of $\reg$, regardless of the ambient RKHS;
KRR requires more structure to achieve similar results.
\subsection{Real data}
We also compared KRR and KPCR using three ``weighted degree'' kernels designed for recognizing
splice sites in genetic sequences~\citep{sonnenburg08machine}.\footnote{We use the first three
kernels for the data obtained from \url{http://mldata.org/repository/data/viewslug/mkl-splice/}.}
The $3300$ samples are divided into a training set ($1000$), validation set ($1100$), and testing set
($1200$).
For each kernel, we use the training data to compute $\estl$ for $\lambda$ in a discrete grid of $2^{10}$
equally-spaced values between $10^{-5}$ and $0.4$, and select the value of $\lambda$ on which the MSE of
$\estl$ on the validation set is smallest.
The MSE on the testing set and the intrinsic dimension $d_{\hat\lambda}$ for the selected
$\hat\lambda$ (on the training data) are as follows:
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
& Kernel 1 & Kernel 2 & Kernel 3 \\
\hline
MSE of $\est_{\KRR,\hat\lambda_{\KRR}}$
& \textbf{0.181452}
& 0.172223
& 0.167745 \\
\hline
MSE of $\est_{\KPCR,\hat\lambda_{\KPCR}}$
& 0.187059
& \textbf{0.168067}
& \textbf{0.164159} \\
\hline
$d_{\hat\lambda_{\KRR}}$
& 175.6879
& 373.0029
& 738.0712 \\
\hline
$d_{\hat\lambda_{\KPCR}}$
& 170.9016
& 275.3560
& 581.1381 \\
\hline
\end{tabular}
\end{center}
KRR outperforms KPCR with Kernel 1, where the intrinsic dimension of the kernel is low, while the
reverse happens with Kernels 2 and 3, where the intrinsic dimensions
are high. This resonates with our theoretical results, which suggest that
KRR requires low intrinsic dimension to perform most effectively.
\section{Discussion}
Our unified analysis for a general class of regularization families in
nonparametric regression highlights two important statistical
properties.
First, the results show minimax optimality for this general class in
several commonly studied settings, which was only previously
established for specific regularization methods.
Second, the results demonstrate the adaptivity of certain
regularization families to subspaces of the RKHS, showing that
these techniques may take advantage of additional smoothness
properties that the signal may possess.
It is notable that the most well-studied family, KRR/Tikhonov
regularization, does not possess this adaptability property.
\bibliographystyle{plainnat}
|
\section{Introduction}
\label{section:Intro}
Hyperspectral images consist of very high-dimensional pixel observations that allow reconstruction of the spectral profiles of objects imaged thanks to the acquisition of several hundred narrow spectral bands. The supervised classification of these pixels is a challenging task, which commonly arises in remote sensing imaging \cite{camps2014advances,gualtieri1999support,Fauvel2,fauvel2013advances}. Structure of the hyperspectral imagery is seldom studied in a comprehensive manner, with most approaches focusing either on spatial information building on tools available for normal imagery or with a focus on spectral information without a principled way to make use of both. We propose a novel approach to classification based on kernel embeddings of distributions which utilizes both the spatial and spectral information in the data. While aimed at hyperspectral imaging, the method we propose is general and can be applied to other types of data.
Kernel methods and support vector machines have been employed in the hyperspectral imaging in \cite{fang2015classification,li2013generalized} the pixel data is lifted into a potentially infinite-dimensional feature space, called reproducing kernel Hilbert space (RKHS), where linear separating hyperplanes are sought. However, spectral information contained in pixels is often not sufficient for such task, and as we will see, including the local / spatial information available in the imagery is key to obtain good classification accuracy. Our approach is to encode the spatial neighbourhood of each pixel as a random sample from a distribution associated to that pixel and to treat such distribution as an additional feature for classification. In order to add consistent spatial information, we also use the Hadamard multiplication of two kernels. Where one kernel is the kernel embeddings of distributions, and the other one is the linear kernel of spatial information similarly to \cite{campsValls1,fauvel2007spectral}.
In Section \ref{sec:related_work}, related work is reviewed. Section \ref{sec:background} provides the background on kernel embeddings of distributions, random features for fast approximations to kernel methods, and on mathematical morphology, which allow us to analyse and understand the geometrical structures of images. Section \ref{sec:theory} studies the consistency and convergence rate of the proposed method and experiments are given in Section \ref{sec:experiments}.
\section{Related work}
\label{sec:related_work}
Many techniques aim to include the spatial information in the classification process. Of particular interest are those combining feature space representations describing the spatial information with those describing the pixels. Morphological feature spaces have been considered in several publications, with impressive results
\cite{Dalla_Mura1,Fauvel2,Pesaresi,lefevre}. On the other hand, kernel methods have also been studied extensively, and more particularly the compositions of kernels \cite{li2013generalized,li2015multiple,campsValls1,Fauvel1}, which allow building new feature space representations.
We marry these approaches with a framework of \cite{Smola2007, muandet2012learning,szabo2014two,oliva2015deep}, where instead of the usual feature map, sending each data point to the feature space, a whole distribution can be represented in the RKHS. This yields a framework for learning on distributions via their representations in this RKHS. In our approach, each pixel is associated to a distribution of its neighbours -- effectively, a hyperspectral image is treated as a set of such distributions. This is similar to the approach to regression applied in \cite{szabo2014two} to the multispectral imaging data. However, the authors of \cite{szabo2014two} partition a multispectral image and classify the partitions - with a goal to obtain responses at the level of the \emph{groups of neighbouring pixels}, which suffices when the goal is to predict an averaged quantity of an image area (e.g. aerosol concentration as studied in \cite{szabo2014two}) and the pixel-level classification is not considered. Another related line of work is that of \cite{Volpi}, where they used the mean map on hyperspectral to perform a dimensionality reduction.
\section{Background}
\label{sec:background}
\subsection{Mean Map Kernel}
Let $ k : {\mathcal X} \times {\mathcal X} \rightarrow \mathbb{R}$ be a positive definite kernel. By Moore-Aronszajn theorem \cite{BerTho04}, there is a unique RKHS $\mathcal{H}$ of real-valued functions on ${\mathcal X}$ where $\langle g,k(\cdot,x)\rangle_\mathcal{H}=g(x)$, for all $g\in\mathcal H, x\in{\mathcal X}$, implying that $k$ corresponds to an inner product between features and, in particular, $k(x,x')=\langle k(\cdot,x),k(\cdot,x')\rangle_\mathcal{H}$. This means that $k(\cdot,x)$ can be viewed as a feature of $x\in \mathcal X$. For many typical choices of kernels $k$, the RKHS $\mathcal{H}$ is infinite-dimensional. Now, let $X$ denote a random variable following a distribution $\mathcal{P}$. The \emph{mean map} or the \emph{kernel embedding} \cite{Smola2007,SriGreFukLanSch10} of $\mathcal{P}$ is defined as:
\begin{eqnarray}
\mu_{\mathcal{P}} := \mathbb{E}_X [k(\cdot, X) ] = \int_{\mathcal X} k(\cdot,x) \ \mathrm{d}\mathcal{P}(x),
\end{eqnarray}
where the expectation is over $\mathcal{H}$. For \emph{characteristic} kernels \cite{Sriperumbudur2011}, which include Gaussian RBF, Matern family and many others, this embedding is injective on the space of all probability distributions (i.e. captures information on all moments, akin to a characteristic function).
Further, if we are given two random variables, $X$ following the distribution $\mathcal{P}$, and $Y$ following the distribution $\mathcal{Q}$, the inner product between the corresponding embeddings is given as
\small
\begin{eqnarray}
\langle \mu_{\mathcal{P}},\mu_{\mathcal{Q}} \rangle_\mathcal{H} = \mathbb{E}_{X,Y} [k(X, Y) ],
\end{eqnarray}
\normalsize
which is sometimes referred to as a \emph{mean map kernel}.
For a random sample $\{ x_1,\ldots,x_n \}$, drawn independently and identically distributed from $\mathcal{P}$, we can define the empirical mean map:
\small
\begin{eqnarray}
\widehat{\mu}_{\mathcal{P}} = \frac{1}{n} \sum_{i=1}^n k(\cdot,x_i),
\end{eqnarray}
\normalsize
and for random samples $\{ x_1,\ldots,x_n \}$ from $\mathcal{P}$ and $\{ y_1,\ldots,y_m \}$ from $\mathcal{Q}$, we obtain the empirical mean map kernel:
\small
\begin{eqnarray}
\langle\widehat{\mu}_{\mathcal{P}},\widehat{\mu}_{\mathcal{Q}}\rangle_\mathcal{H} = \frac{1}{nm} \sum_{i=1}^n \sum_{j=1}^m k(x_i ,y_j).
\end{eqnarray}
\normalsize
\subsection{Random features for kernels}
The computational and storage requirements for kernel methods on large datasets can be prohibitive in practice due to the need to compute and store the kernel matrix. If we consider a dataset of $n$ $D$-dimensional observations, the storage requirements are $O(n^2)$ and the calculation takes $O(Dn^2)$ operations. A remedy developed by \cite{Rahimi2007} is to approximate translation-invariant kernels in an unbiased way using a random feature representation. Namely, any translation-invariant positive definite kernel $k$, such that $ \forall(x,y) \in \mathcal{X}^2$, $k(x,y) = \kappa(x-y)$ can be written as
$k(x,y) = \textbf{E}_{\omega\sim\Lambda}\left[\cos(\omega^\top x)\cos(\omega^\top y)\right.$ $+\left.\sin(\omega^\top x)\sin(\omega^\top y)\right]$,
where $\omega \in \mathbb{R}^D$ follows some distribution $\Lambda$ (spectral measure of the kernel). Thus, by sampling i.i.d. vectors $\omega_1,\ldots,\omega_N$ from $\Lambda$, we can approximate kernel $k$ by $\hat{k}$ defined by:
$\hat{k}(x,y) = \frac{1}{N}\sum_{j=1}^N \left(\cos(\omega_j^\top x)\cos(\omega_j^\top y)\right.$ $+\left. \sin(\omega_j^\top x)\sin(\omega_j^\top y)\right)$,
so that the original feature map $k(\cdot, x)$, potentially living in an infinite-dimensional space, is approximated by an explicit $2N-$dimensional feature vector:
\footnotesize
\begin{eqnarray}
\hat{Z}(x) = \sqrt{\frac{1}{N}}\begin{bmatrix}\cos(\omega_1^\top x), \ldots , \cos(\omega_N^\top x), \sin(\omega_1^\top x),\ldots, \sin(\omega_N^\top x)\end{bmatrix}^T.
\end{eqnarray}
\normalsize
Thus, the mean map and the mean map kernel can be estimated using these finite-dimensional representations. In this contribution, we will focus on Gaussian RBF kernels for which the spectral measure $\Lambda$ is also Gaussian.
\subsection{Random features mean map on hyperspectral images}
Let us now turn our attention to a hyperspectral image $h$. Around each pixel location $x_i$, we consider a square patch $\mathcal{P}_{x_i}^{(s)}$ of size $s$ where we will treat the pixels as a random sample from a distribution $\mathcal{P}_i$ specific to the location $x_i$. Instead of calculating the kernel between individual data points, we will calculate kernel between these distributions.
An empirical mean map kernel is thus given simply by:
\begin{eqnarray}
K_{mm}(x_i,x_j)&= &\langle \hat \mu_{P_i}, \hat \mu_{P_j} \rangle_\mathcal H \\\nonumber
&=&\frac{1}{s^2} \sum_{l_1 \in \mathcal{P}_{x_i}} \sum_{l_2 \in \mathcal{P}_{x_j}} k(h(x_{l_1}),h(x_{l_2}))\\\nonumber
&\approx&\frac{1}{s^2} \sum_{l_1 \in \mathcal{P}_{x_i}} \sum_{l_2 \in \mathcal{P}_{x_j}} \hat{Z}(h(x_{l_1}))^\top\hat{Z}(h(x_{l_2})),
\end{eqnarray}
where $h(x)$ denotes the measurement vector at location $x$ and in the last line we employ a random feature approximation of $k$.
It should be noted that there may be outliers in a patch, which can damage the estimation of the mean. Similarly to the work of \cite{flaxman}, we proposed to use a weighted mean map, where the weights depend on spatial information. The kernels we obtain, called convolutional kernels, have also been used in \cite{mairal2014}. In contrast to \cite{mairal2014}, however, we will use random feature expansions to explicitly represent the feature space.
The convolutional kernel is defined as:
\small
\begin{eqnarray*}
\widehat{KCN}(x_i,x_j)= \sum_{l1 \in \mathcal{P}_{x_i}} \sum_{l2 \in \mathcal{P}_{x_j}} \| h(x_{l1}) \|_2 \| h(x_{l2})\|_2 \\ e^{-\frac{1}{2\beta^2} \| x_{l1}-x_{l2} \|_2 } e^{-\frac{1}{2\sigma^2}\| \tilde{h}(x_{l1})-\tilde{h}(x_{l2}) \|_2 },
\end{eqnarray*}
\normalsize
where $\tilde{h}$ represents a normalised version of $h$, such that for all $i \in [1,n]$ =$\| \tilde h(x_i)\|_2 =1$. So we do a product of a kernel on the positions, another on the magnitudes, and a third one is an RBF kernel between spectra.
This formula can be interpreted as a weighted mean map which is defined by:
\scriptsize
\begin{eqnarray}\label{KCN1}
\widehat{KCN}(x_i,x_j)= \langle\widehat{\mu}_{P(\mathcal{P}_{x_i})},\widehat{\mu}_{P(\mathcal{P}_{x_j})}\rangle \\
\widehat{KCN}(x_i,x_j)= \frac{1}{s^2} \sum_{l1 \in \mathcal{P}_{x_i}} \sum_{l2 \in \mathcal{P}_{x_j}} \| h(x_{l1}) \|_2 . \| h(x_{l2})\|_2 \hat{\mathbf{k}}(x_{l1},x_{l2}),
\end{eqnarray}
\normalsize
where $\hat{\mathbf{k}}$ is a positive definite kernel arising from the random feature space expansion.
\section{Theoretical Analysis}
\label{sec:theory}
Let us consider that the data are partitioned into sets following the same distribution, then the structure of our data is given by $ \{(\{x_{i,n} \}_{n=1}^{N_i} ,y_i)\}_{i=1}^l $ with $x_{i,1}, \ldots, x_{i,N_i} \overset{i.i.d.}{\sim} x_i$, where $(x_i,y_i)$ are drawn from a joint meta distribution $\mathcal{M}$. We follow the notation of \cite{szabo2014two}. Let us denote $\Phi$ a loss function. Let us write the following expected risk function of the data for the SVM problem:
\small
\begin{eqnarray}
\mathcal{R}(f)= \inf_{f \in \mathcal{H}} \mathbb{E}_{(x,y)\sim\mathcal{M}} \left( \Phi(f(x)y)\right)
\end{eqnarray}
\normalsize
We can modify it to mean map embedding classification problem :
\small
\begin{eqnarray}
\mathcal{R}_{\mu}(f)= \inf_{f \in \mathcal{H}} \mathbb{E}_{(x,y)\sim\mathcal{M}} \left( \Phi(f(\mu_x)y)\right)
\end{eqnarray}
\normalsize
We can also write the empirical risk function, for mean map embedding classification problem :
\small
\begin{eqnarray}
\hat{\mathcal{R}}_{\mu}(f)= \inf_{f \in \mathcal{H}} \frac{1}{n}\sum_{i=1}^n \left( \Phi(f(\mu_{x_i})y_i) \right)
\end{eqnarray}
\normalsize
Finally we can also write the empirical risk function, for the empirical mean map embedding classification problem :
\small
\begin{eqnarray}
\hat{\mathcal{R}}_{\hat{\mu}}(f)= \inf_{f\in \mathcal{H}} \frac{1}{n}\sum_{i=1}^n \left( \Phi(f(\hat{\mu}_{x_i})y_i) \right)
\end{eqnarray}
\normalsize
Then we would like to obtain an inequality between $\mathcal{R}_{\mu}(f)$ and $ \hat{\mathcal{R}}_{\hat{\mu}}(f)$. To do that, inspired by \cite{muandet2012learning}, we derive a inequality $\mathcal{R}_{\mu}(f)$ and $ \mathcal{R}(f)$ : \\
\begin{theorem}
Given that $x \sim P$ an arbitrary probability distribution with variance $\sigma^2$, a Lipschitz continuous function $f : \mathbb{R} \rightarrow \mathbb{R}$ with constant $C_f$, an arbitrary loss function $\Phi: \mathbb{R} \rightarrow \mathbb{R}$ that is Lipschitz continuous in the second argument with constant $C_l$ , it follows that :
\begin{eqnarray}
\mathcal{R}_{\mu}(f) - \mathcal{R}(f) \leq C_l C_f^2 \mathbb{E}_{(x)} \| x-\mu_{x} \|^2 \mathbb{E}_{(y)} ( y^2)
\end{eqnarray}
\end{theorem}
The proof of this theorem can be found on the supplementary materials.
Then we might use \cite{cortes2013learning} where we have an inequality between $\mathcal{R}_{\mu}(f)$ and $ \hat{\mathcal{R}}_{\mu}(f)$ :
\begin{theorem}
Let $ \mathcal{G} = \Phi(\mathcal{H}, .)$ denote the loss class, let $ \mathcal{R}_n(\mathcal{G} )$ denote the Rademacher complexity. Let $ \Sigma(\mathcal{G})^2 = \sup_{g \in \mathcal{G}} \mathbb{E}(g^2)$ be a bound on the variance of the functions in
$\mathcal{G}$. If the trace of the kernel is bounded, the loss function $\Phi: \mathbb{R} \rightarrow \mathbb{R}$ that is Lipschitz continuous, for any $\delta>0$ , the following bound holds with probability at least $1-\delta$
\scriptsize
\begin{eqnarray*}
\hat{\mathcal{R}}_{\mu}(f)- \mathcal{R}_{\mu}(f) \leq 8 \mathcal{R}_n(\mathcal{G} ) + \Sigma(\mathcal{G}) \sqrt{\frac{8 \log(2/\delta)}{n}}+ \frac{3 \log(2/\delta)}{n}
\end{eqnarray*}
\normalsize
\end{theorem}
\begin{theorem}
Given that $f : \mathbb{R} \rightarrow \mathbb{R}$ is a Lipschitz continuous function with constant $C_f$, an arbitrary loss function $\Phi: \mathbb{R} \rightarrow \mathbb{R}$ that is Lipschitz continuous with constant $C_{l2}$ , it follows that :
\small
\begin{eqnarray*}
\hat{\mathcal{R}}_{\hat{\mu}}(f) -\hat{\mathcal{R}}_{\mu}(f) \leq \frac{1}{n}C_{l}C_f^2 \hat{\mathbb{E}}\left( \| \mu_x -\hat{\mu}_x \|^2\right)\hat{\mathbb{E}}\left( (y)^2\right)
\end{eqnarray*}
\normalsize
\end{theorem}
The proof of this theorem can be found on the supplementary materials.
We also need the following theorem proved in \cite{Smola2007}
\begin{theorem}
Assume that $\|g\|_{\infty} \leq R$ for all $g \in \mathcal{H}$
with $\|g\|_{\mathcal{H}} \leq 1$, and that $k$ is an universal kernel. Then with probability at least $1 - \delta $ :
\begin{eqnarray*}
|\mu[P] - \mu[X]| \leq 2\mathcal{R}_n(\mathcal{H}, P) + R\sqrt{\log\left(1/\delta\right)/n}
\end{eqnarray*}
\normalsize
where $\mathcal{R}_n(\mathcal{H}, P)$ denotes the Rademacher average associated with $P$ and
$\mathcal{H}$.
\end{theorem}
Then by combining the previous theorems we easily have the following theorem.
\begin{theorem}\label{theorem1}
Given the conditions of the previous theorems. Then with probability at least $1 - \delta $ :
\scriptsize
\begin{eqnarray*}
\hat{\mathcal{R}}_{\hat{\mu}}(f) - \mathcal{R}(f) \leq C_l C_f^2 \left[ \mathbb{E}_{(x)} \| x-\mu_{x} \|^2 \mathbb{E}_{(y)} ( y^2)\right. \\
\left. + \left( 2\mathcal{R}_n(\mathcal{H}, P) + R\sqrt{\log\left(1/\delta\right)/n} \right)\hat{\mathbb{E}}\left( (y)^2\right) \right] \\
+8 \mathcal{R}_n(\mathcal{G} ) + \Sigma(\mathcal{G}) \sqrt{\frac{8 \log(2/\delta)}{n}}+ \frac{3 \log(2/\delta)}{n}
\end{eqnarray*}
\normalsize
\end{theorem}
\section{Experiments}
\label{sec:experiments}
We evaluate the classification accuracy of the proposed approach using two standard datasets: the AVIRIS Indian Pines, and the ROSIS
University of Pavia. The first data set is an image of dimension $145 \times 145$ pixels, with $D = 224$ spectral bands and its geometrical
resolution is of 3.7 m. The training set is composed of $80$ pixels, and the image is composed of $16$ classes. The dimensions of the second data set are $ 610 \times 340$ pixels, with $D = 103$ spectral bands and its geometrical resolution is of 1.3 m. The training set is composed of $3921$ pixels and a testing set of $42776$ pixels, and the image is composed of $9$ classes. There is a commonly used testing set for Pavia data, and we report performance on this testing set. In the first data set there is no testing set so we generate 20 Monte-Carlo simulations, selecting randomly 5 pixels per class, then aggregate the result of the classification. We used the Morphological Profile (MP) feature \cite{Pesaresi, Dalla_Mura1} space which is commonly used in pixel classifiation and is described in the supplementary material. We also use the product of the two kernels where one is the MP kernel and the other is the KMM kernels. This kind of techinique has been previously explored in \cite{li2013generalized,li2015multiple,campsValls1}. In contrast to the previous work, which approximate this product of kernels thanks to addition of kernels, we can do the real multiplication since we work with finite dimension Hilbert spaces. The results of classification are reported in Table \ref{table1} (Indian Pine), and Table \ref{table2} (Pavia). The classification algorithm used is the C-SVM \cite{CC01a} were the parameter $C$ was selected with $5-$fold cross-validation on a grid $C=2^i$,with $i \in [-15,15]$. The results on Table \ref{table1} and \ref{table2} show us that kernel mean map can perform as well as state of art results on these images. The size of scale $s$ seems to be important, this is linked with the theorem \ref{theorem1}, where we see that increasing the size of the scale increase in a way the size of the training set. An explanation of the parameters of evaluation can be found in \cite{fauvel2007spectral} ( pages 166-167) and also in the supplementary materials.
\begin{table}[width=0.35\columnwidth]
\begin{center}
\tiny
\begin{tabular}{l l l l l}
kernel & \bf parameters & \bf OA & \bf kappa statistic & \bf AA \\
\hline \hline
linear kernel & ~ & $54.6 \pm 3.3$ & $49.5 \pm 3.5$ & $58.2 \pm 2.7$\\
\hline
random rbf & ~ & $53.9 \pm 2.7$ & $48.6 \pm 3.0$ & $58.0 \pm 2.8$\\
\hline
$\hat{Kmm}$ & s=3 &$57.6 \pm 4.2$ & $48.1 \pm 4.3$ & $59.1 \pm 3.8$\\
\hline
$\hat{Kmm}$ & s=10 &$66.5 \pm 3.2$ & $62.0 \pm 3.3$ & $65.4 \pm 3.8$\\
\hline
$\hat{Kmm}$ & s=15 &$70.0 \pm 4.1$ & $66.4 \pm 3.9$ & $68.6 \pm 2.3$\\
\hline
$\hat{Kmm}$ & s=20 &$70.15 \pm 3.5$ & $66.7 \pm 3.8$ & $64.1 \pm 2.3$\\
\hline
$K_{MP}$ & ~ &$62.9 \pm 4.6$ & $58.5 \pm 5.5$ & $66.5 \pm 2.3$\\
\hline
$K_{MP} \times \hat{Kmm}$ & s=15 & $73.0 \pm 3.7$ & $69.7 \pm 3.7$ & $76.3 \pm 2.3$\\
\hline
$\widehat{KCN}$ & s=7 & $\bf 77.9 \pm 3.4$ & $ \bf 76.4 \pm 3.7$ & $\bf 74.8 \pm 2.4$\\
\hline
$\widehat{KCN}$ & s=15 & $73.0 \pm 3.7$ & $69.7 \pm 3.7$ & $76.3 \pm 2.3$\\
\hline
\hline
\end{tabular}
\end{center}
\normalsize \caption{Overall accuracy, kappa statistic, and Average accuracy obtained for different kernels, applied on the AVIRIS Indian Pines hyperspectral data set. I have run 20 Monte Carlo simulations. I have selected on the training set just 5 samples per class.} \label{table1}
\end{table}
\begin{table}[width=0.35\columnwidth]
\begin{center}
\tiny
\begin{tabular}{l l l l l}
kernel & \bf parameters & \bf OA & \bf kappa statistic & \bf AA \\
\hline \hline
linear kernel & ~ & $73.2$ & $66.6$ & $78.5$\\
\hline
random rbf & ~ & $78.1 \pm 2.5$ & $74.5 \pm 2.5$ & $80.4 \pm 1.8$\\
\hline
$\hat{Kmm}$ & s=3 &$90.0 \pm 2.1$ & $87.4 \pm 2.5$ & $89.7 \pm 1.1$\\
\hline
$\hat{Kmm}$ & s=10 &$93.2 \pm 1.4$ & $91.1 \pm 1.7$ & $93.2 \pm 0.8$\\
\hline
$\hat{Kmm}$ & s=15 &$93.9 \pm 0.8$ & $91.9 \pm 1.0$ & $93.4 \pm 0.5$\\
\hline
$\hat{Kmm}$ & s=20 &$87.5 \pm 2.6$ & $83.0 \pm 2.5$ & $88.0 \pm 1.2$\\
\hline
$K_{MP}$ & ~ &$97.1$ & $96.2$ & $96.7$\\
\hline
$K_{MP} \times \hat{Kmm}$ & s=10 & $\bf97.4 \pm 0.6$ & $\bf 96.4 \pm 0.7$ & $\bf97.3 \pm 0.6$\\
\hline
$\widehat{KCN}$ & s=10 & $96.2 \pm 0.9$ & $94.9 \pm 1.2$ & $94.0 \pm 1.5$\\
\hline
$\widehat{KCN}$ & s=15 & $96.4 \pm 0.8$ & $95.1 \pm 1.2$ & $94.4 \pm 1.8$\\
\hline
\hline
\end{tabular}
\end{center}
\normalsize \caption{Overall accuracy, kappa statistic, and Average accuracy obtained for different kernels, applied on the University of Pavia hyperspectral data set. We used the classical training set.} \label{table2}
\end{table}
\normalsize
\begin{figure}[h]
\begin{center}
\begin{tabular}{ccc}
\includegraphics[width=0.23\columnwidth]{./ground_truth_I1.eps}&
\includegraphics[width=0.23\columnwidth]{./SVM_linear_I1.eps}&
\includegraphics[width=0.23\columnwidth]{./Krbf_estimated_I1.eps}\\
{\small (a)} & {\small (b)} & {\small (c)}\\
\includegraphics[width=0.23\columnwidth]{./KMM_S_15_I1.eps}&
\includegraphics[width=0.23\columnwidth]{./KMM_spatial_S_7_I1.eps}&
\includegraphics[width=0.23\columnwidth]{./KMM_spatial_S_15_I1.eps}\\
{\small (d)} & {\small (e)} & {\small (f)}
\end{tabular}
\end{center}
\caption{Classification maps for the Indian Pines hyperspectral image using different approaches, with just 5 points per class in the training set. In (a) ground truth, (b) the linear SVM, (c)the estimated RBF SVM, (d) kernel SVM with $KMM$ and $s=15$, (e) kernel SVM with $KCN$ and $s=7$, (f) kernel SVM with $KCN$ and $s=15$.} \label{fig:result_order}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tabular}{ccccc}
\includegraphics[width=0.16\columnwidth]{./ground_truth_I2.eps}&
\includegraphics[width=0.16\columnwidth]{./SVM_linear_I2.eps}&
\includegraphics[width=0.16\columnwidth]{./Krbf_estimated_I2.eps}&
\includegraphics[width=0.16\columnwidth]{./KMM_S_15_I2.eps}&
\includegraphics[width=0.16\columnwidth]{./KMM_spatial_S_13_I2.eps}\\
{\small (a)} & {\small (b)} & {\small (c)}&{\small (d)} & {\small (e)}
\end{tabular}
\end{center}
\caption{Classification maps for the Pavia hyperspectral image using different approaches, with just the classical training set. In (a) ground truth, (b) the linear SVM, (c)the estimated RBF SVM, (d) kernel SVM with $KMM$ and $s=15$, (e) kernel SVM with $KCN$ and $s=13$.} \label{fig:result_order}
\end{figure}
\section{conclusion}
\label{sec:conclu}
In this article we developed a new method for pixel classification in hyperspectral imaging. The method uses spatial information encoded through distributions of the neighbourhood around each pixel. Even with very simple kernel choices (Gaussian RBF applied directly to raw data) the obtained results are comparable with state-of-the-art. We establish convergence rates for, and prove that we are two-stage consistent. Further improvements are possible by using different feature spaces employing suitable representations of the individual pixels. We believe that we have established an interesting research direction where local distributions are treated as additional features for a supervised learning task, which is of particular interest in hyperspectral imaging where it is difficult to combine spatial and spectral information in a principled way and our approach can be viewed as a step in that direction.
\clearpage
\bibliographystyle{splncs}
|
\section{Introduction}
The electronic structure of the transition-metal dichalcogenides (TMDs) belonging to the orthorhombic and non-centrosymmetric $Pmn2_1$ space group, e.g. WTe$_2$, were recently recognized as candidates for possible topologically non-trivial electronic states. For instance, their monolayer electronic bands were proposed to be characterized by a non-trivial $Z_2 = 1$ topological invariant based on the parity of their valence bands, making their monolayers good candidates for a quantum spin Hall insulating ground-state \cite{TP_transition}. This state is characterized by helical edge states that are protected by time-reversal symmetry from both localization and elastic backscattering. Hence, these compounds could provide a platform for realizing low dissipation quantum electronics and spintronics \cite{TP_transition, MacDonald}.
However, the majority of gapped TMDs, such as semiconducting MoS$_2$ or WSe$_2$, crystallize either in a trigonal prismatic coordination or in a triclinic structure with octahedral coordination \cite{review1,review2} as is the case of ReS$_2$. Those crystallizing in the aforementioned orthorhombic phase, e.g. WTe$_2$, are semi-metals albeit displaying remarkable transport properties such as an enormous, non-saturating magnetoresistivity \cite{cava}. Strain is predicted to open a band gap \cite{TP_transition} in WTe$_2$, which might make it suitable for device development. In fact, simple exfoliation of its isostructural $\gamma-$MoTe$_2$ compound (where $\gamma$ refers to the orthorhombic semi-metallic phase) into thin atomic layers was claimed to induce a band gap \cite{MoTe2_MI_transition} in the absence of strain. Such a transition would contrast with band structure calculations finding that WTe$_2$ should remain semi-metallic when exfoliated down to a single atomic layer \cite{Lv}. The insulating behavior reported for a few atomic layers of WTe$_2$ was ascribed to an increase in disorder due to its chemical instability in the presence of humidity which would induce Anderson localization \cite{Morpurgo}, although more recently it was claimed to be intrinsic from transport measurements on encapsulated few-layered samples \cite{Cobden}.
Orthorhombic $\gamma-$MoTe$_2$ and its isostructural compound WTe$_2$ were also claimed, based on density functional theory calculations, to belong to a new class of Weyl semi-metals, called type-II, which is characterized by a linear touching between hole and electron Fermi surface pockets \cite{bernevig,felser,bernevig2, Hasan}. As for conventional Weyl points \cite{Weyl1, Weyl2}, these Weyl type-II points would also act as topological charges associated with singularities, i.e., sources and sinks, of Berry-phase pseudospin \cite{bernevig,felser,bernevig2,Hasan} which could lead to anomalous transport properties. A series of recent angle-resolved photoemission spectroscopy (ARPES) measurements \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang, ARPES_Liang, ARPES_Xu, ARPES_Tamai, ARPES_Belopolski, thirupathaiah} claim to observe a good overall agreement with these predictions. These studies observe the band crossings predicted to produce the Weyl type-II points, which would be located slightly above the Fermi-level, as well as the Fermi arcs projected on the surface of this compound \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang, ARPES_Liang, ARPES_Xu, ARPES_Tamai, ARPES_Belopolski, thirupathaiah}.
Here, motivated by the scientific relevance and the possible technological implications of the aforementioned theoretical predictions \cite{bernevig,felser,bernevig2,Hasan, TP_transition}, we evaluate, through electrical transport and torque magnetometry in bulk single-crystals, the electronic structure at the Fermi level and the topological character of orthorhombic $\gamma-$MoTe$_2$. Our goal is to contrast our experimental observations with the theoretical predictions and the reported ARPES results in order to validate their findings. This information could, for example, help us predict the electronic properties of heterostructures fabricated from single- or a few atomic layers of this compound. An agreement between the calculated geometry of the FS of $\gamma-$MoTe$_2$ with the one extracted from quantum oscillatory phenomena, would unambiguously support the existence of Weyl nodes in the bulk \cite{Weyl1,Weyl2} and, therefore, the existence of related non-trivial topological surface states or Fermi arcs \cite{Weyl2, Hasan, bernevig,felser,bernevig2,Hasan}. However, quantum oscillatory phenomena from $\gamma-$MoTe$_2$ single-crystals reveals a Fermi surface whose geometry is quite distinct from the one predicted by the DFT calculations based on its low temperature crystallographic structure. The extracted Berry-phase is found to be field-dependent. Still one does not obtain evidence for the topological character predicted for this compound when the Berry-phase is evaluated at low fields. Here, we show that shifts in the relative position of the electron and hole bands, implied by previous ARPES studies \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang, ARPES_Liang, ARPES_Xu, ARPES_Tamai, ARPES_Belopolski, thirupathaiah}, can replicate the angular dependence of the observed Fermi surface cross-sectional areas. However, these band shifts imply that the valence and electron bands would no longer cross and, therefore, that $\gamma-$MoTe$_2$ would not display the predicted Weyl type-II semi-metallic state.
\begin{figure*}[htp]
\begin{center}
\includegraphics[width=18 cm]{MoTe2_Weyl_fig1v2}
\caption{(a) Resistivity $\rho$, for currents flowing along the $a-$axis, as a function of the temperature $T$ for three representative single crystals displaying resistivity ratios $\rho(300 \text{ K})/\rho(2) \text{ K}$ between 380 and $\gtrsim 1000$. (b) $\rho$ as a function of $T$ for each single-crystal indicating that $T_c$ depends on sample quality: it increases from an average transition middle point value of $\sim 130$ mK for the sample displaying the lowest ratio to $\sim 435$ mK for the sample displaying the highest one. The apparent hysteresis is due to a non-ideal thermal coupling between the single-crystals, the heater and the thermometer. (c) $\rho$ as a function of the field $H$ applied along the \emph{c}-axis at a temperature $T = 1.7$ K for a fourth crystal characterized by a resistivity ratio of $\sim 207$. Notice i) the non-saturation of $\rho(H)$ and ii) that $\Delta \rho (\mu_0 H)/ \rho_0 = (\rho(\mu_0 H)- \rho_0)/ \rho_0$, where $\rho_0 = \rho(\mu_0 H=0 \text{ T}, T = 2 \text{ K})$ surpasses $1.4 \times 10^6$ \% at $\mu_0 H = 60$ T. (d) $\rho$ as a function of $\mu_0 H$ applied along the $b-$axis also at $T=1.7$ K and for the same single-crystal. (e) Shubnikov-de Haas signal superimposed onto the magnetoresistivity for $\mu_0 H \| c-$axis and for three temperatures, $T = 8$ K (blue line), 4.2 K (red line) and 1.7 K (black line), respectively. (f) Oscillatory signal (black line) superimposed onto the magnetic susceptibility $\Delta \chi = \partial (\tau/\mu_0 H)/\partial \mu_0 H$, where $\tau$ is the magnetic torque.}
\end{center}
\end{figure*}
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=14cm]{MoTe2_Weyl_fig3v2}
\caption{(a) Typical de Haas-van Alphen (red trace) and Shubnikov-de Haas (black trace) signals superimposed onto the magnetic torque and the magnetoresistivity respectively, for fields aligned along the \emph{c}-axis of $\gamma-$MoTe$_2$ single-crystals at $T \simeq 30$ mK. The same panel shows the FFT spectra for each signal revealing just two main frequencies or FS cross-sectional areas. (b) dHvA and SdH signals and corresponding FFT spectra obtained from the same single-crystals for $H$ aligned nearly along the \emph{a}-axis of each single-crystal. (c) and (d), Amplitude of the peaks observed in the FFT spectra for $H\|c$-axis and as a function of $T$ including the corresponding fits to the LK formula from which the effective masses are extracted. (e) and (f) Amplitude of two representative peaks observed in the FFT spectra for $H\|a$-axis and as a function of $T$ with the corresponding fits to the LK formula to extract their effective masses.}
\end{center}
\end{figure*}
\section{Methods and Experimental Results}
Very high quality single crystals of monoclinic $\beta-$MoTe$_2$ were synthesized through a Te flux method: Mo, 99.9999\%, and Te 99.9999 \% powders were placed in a quartz ampoule in a ratio of 1:25 heated up to 1050 $^{\circ}$C and held for 1 day. Then, the ampoule was slowly cooled down to 900 $^{\circ}$C and centrifuged. The ``as harvested" single-crystals were subsequently annealed for a few days at a temperature gradient to remove the excess Te. Magneto-transport measurements as a function of temperature were performed in a Physical Property Measurement System using a standard four-terminal configuration. Measurements of the Shubnikov-de Haas (SdH) and the de Haas-van Alphen (dHvA) effects were performed in dilution refrigerator coupled to a resistive Bitter magnet, with the samples immersed in the $^3$He-$^4$He mixture. Measurements of the dHvA-effect were performed \emph{via} a torque magnetometry technique, i.e. by measuring the deflection of a Cu-Be cantilever capacitively. Electrical transport measurements in pulsed magnetic fields were performed at the Dresden High Magnetic Field Laboratory using a 62 T magnet with a pulse duration of 150 ms. The sample temperature was controlled using a $^4$He bath cryostat (sample in He atmosphere) with an additional local heater for temperatures above 4.2 K. Synchrotron based X-ray measurements were performed in three single-crystals at the CHESS-A2 beam line using a combination of photon energies and cryogenic set-ups. The crystallographic data was reduced with XDS \cite{XDS}. The structures were solved with direct methods using SHELXS \cite{shelx}. Outlier rejection and absorption correction was done with SADABS. Least squares refinement on the intensities were performed with SHELXL \cite{shelx}. For additional detailed information on the experimental set-ups used, see Supplemental Information file \cite{supplemental}.
As illustrated by Fig. 1(a), the as synthesized single-crystals display resistivity ratios \emph{RRR} = $\rho(T = 300 \text{ K})/\rho(T = 2 \text{ K})$ ranging from $380$ to $> 2000$ which is one to two orders of magnitude higher than the $RRR$ values currently in the literature (see, for example, Ref. \onlinecite{MoTe2_SC}). Although not clearly visible in Fig. 1(a) due to its logarithmic scale, a hysteretic anomaly is observed in the resistivity around $ 240 $ K corresponding to the monoclinic to orthorhombic structural transition which stabilizes what we denominate as the orthorhombic $\gamma-$MoTe$_2$ phase. For a clearer exposure of this transition and related hysteresis, see Ref. \onlinecite{qiong}. These single-crystals were subsequently measured at much lower temperatures allowing us to determine their superconducting transition temperature $T_c$. Remarkably, and as seen in Fig. 1(b), we find that $T_c$ depends on sample quality, increasing considerably as the $RRR$ increases, suggesting that structural disorder suppresses $T_c$. For these measurements, particular care was taken to suppress the remnant field of the superconducting magnet since the upper critical fields are rather small (see Supplemental Fig. S1 \cite{supplemental}). The sample displaying the highest $RRR$ and concomitant $T_c$ was measured in absence of a remnant field. To verify that these differences in $T_c$ are not due to a poor thermal coupling between the sample and the thermometers, $T_c$ was measured twice by increasing and decreasing $T$ very slowly. The observed hysteresis is small relative to $T_c$ indicating that the measured $T_c$s are not an artifact. The values of the residual resistivities $\rho_0$ depend on a careful determination of the geometrical factors such as the size of the electrical contacts. Therefore, the $RRR$ provides a more accurate determination of the single-crystalline quality. In the past, the suppression of $T_c$ by impurities and structural defects was systematically taken as evidence for unconventional superconductivity \cite{andy, satoru1, satoru2}, e.g. triplet superconductivity \cite{Maeno} in Sr$_2$RuO$_4$. Nevertheless, the fittings of the upper-critical fields $H_{c2}$ to a conventional Ginzburg-Landau expression, shown in Fig. S1 \cite{supplemental}, points towards singlet pairing.
We have also evaluated the quality of our single crystals through Hall-effect \cite{qiong} and heat capacity measurements (see, Supplemental Fig. S2 \cite{supplemental}). Hall-effect reveals a sudden increase in the density of holes below $T = 40$ K, suggesting a possible temperature-induced Lifshitz-transition. While the heat capacity reveals a broad anomaly around $T = 66$ K, well-below its Debye temperature ($\Theta_D \simeq 120$ K), that would suggest that the structural degrees of freedom continue to evolve upon cooling below $T = 100$ K. Given that such structural evolution could affect the electronic band structure predicted for $\gamma-$MoTe$_2$ \cite{felser,bernevig2,Hasan}, we performed synchrotron X-ray scattering down to $ \sim 12$ K (see, Supplemental Fig. S3 \cite{supplemental}). We observe some variability in the lattice constants extracted among several single-crystals and a sizeable hysteresis in the range $125 \text{K} \leq T \leq 250 \text{K}$ associated with the structural transition observed at $T \simeq 250$ K, but no significant evolution in the crystallographic structure below 100 K. As we discuss below, there are negligible differences between the electronic bands calculated with the crystal structures collected at 100 K and at 12 K, respectively.
Figures 1(c) and 1(d) display the change in magnetoresistivity $\Delta \rho (\mu_0 H)/\rho_0 = (\rho(\mu_0 H)-\rho_0)/\rho_0$ as a function of the field $\mu_0 H$ for a crystal characterized by $RRR \sim 450$ when the electrical current flows along the crystalline \emph{a}-axis and the field is applied either along the \emph{c}- or the \emph{b}-axes, respectively. Similarly to WTe$_2$, for both orientations $\Delta \rho /\rho_0$ shows no sign of saturation under fields all the way up to 60 T while surpassing $1 \times 10^{6}$ \% for $\mu_0 H \| c$-axis \cite{cava}. For WTe$_2$ such anomalous magnetoresistivity was attributed to compensation between the density of electrons and holes \cite{cava,pletikosic,pippard}. Nevertheless, there are a number of subsequent observations \cite{Daniel} contradicting this simple scenario, such as i) a non-linear Hall response \cite{Joe}, ii) the suppression of the magnetoresistivity at a pressure where the Hall response vanishes \cite{WTe2_SC_1} (i.e. at perfect compensation), and iii) the observation of a pronounced magnetoresistivity in electrolyte gated samples with a considerably higher density of electrons with respect to that of holes \cite{Fuhrer}. It remains unclear if the proposed unconventional electronic structure \cite{bernevig,felser,bernevig2} would play a role on the giant magnetoresistivity of WTe$_2$, while its measured FS differs from the calculated one \cite{Daniel,behnia}. In contrast, we have previously shown that $\gamma$-MoTe$_2$ indeed is a well compensated semi-metal \cite{qiong}. The slightly smaller
magnetoresistivity of $\gamma-$MoTe$_2$ relative to WTe$_2$ is attributable to heavier effective effective masses, according to de Haas-van Alphen-effect discussed below, or concomitantly lower mobilities.
The best $\gamma-$MoTe$_2$ samples, i.e. those with $RRR \geq 2000$, display even more pronounced $\Delta \rho/\rho_0$ under just $\mu_0 H \simeq 10$ T. The oscillatory component superimposed on the magnetoresistivity corresponds to the Shubnikov-de Haas (SdH) effect resulting from the Landau quantization of the electronic orbits. Figure 1(e) shows the oscillatory, or the SdH signal as a function of inverse field $(\mu_0H)^{-1}$ for three temperatures. The SdH signal was obtained by fitting the background signal to a polynomial and subtracting it. Notice how for this sample and for $\mu_0 H \|$ \emph{c}-axis, the SdH signal is dominated by a single frequency. However for all subsequent measurements performed under continuous fields (discussed below) one observes the presence of two main frequencies very close in value, each associated to an extremal cross-sectional area $A$ of the FS through the Onsager relation $F = A(\hbar / 2\pi e)$ where $\hbar$ is the Planck constant and $e$ is the electron charge. To illustrate this point, we show in Fig. 1(f) the oscillatory signal extracted from the magnetic torque, i.e. $\mathbf{\tau} = \mathbf{M} \times \mu_0 \mathbf{H}$, or the de Haas-van Alphen effect (dHvA) collected from a $\gamma-$MoTe$_2$ single-crystal for fields aligned nearly along its \emph{c}-axis. Here $M = \chi \mu_0 H$ is the magnetization and $\chi(\mu_0 H, T)$ is its magnetic susceptibility. Figure 1(f) shows the oscillatory component of the magnetic susceptibility $\Delta \chi = \partial (\tau/\mu_0H)/ \partial(\mu_0H)$. The envelope of the oscillatory signal displays the characteristic ``beating" pattern between two close frequencies. This becomes clearer in the fast Fourier transform of the oscillatory signal shown below.
According to the Lifshitz-Onsager quantization condition \cite{kim,nagaosa}, the oscillatory component superimposed onto the susceptibility is given by:
\begin{eqnarray*}
\Delta \chi[(B)^{-1}] \propto \frac{T}{B^{5/2}}\sum_{l=1}^{\infty} \frac{\exp^{-l \alpha \mu T_D/ B}\cos(l g \mu \pi/2)}{l^{3/2}\sinh(\alpha \mu T/B)}
\end{eqnarray*}
\begin{eqnarray}
\times \cos\left\{ 2\pi \left[ \left( \frac{F}{B}-\frac{1}{2}+\phi_B \right)l + \delta \right] \right\}
\end{eqnarray}
where $F$ is the dHvA frequency, $l$ is the harmonic index, $\omega_c$ the cyclotron frequency, $g$ the Land\'{e} \emph{g}-factor, $\mu$ the effective mass in units of the free electron mass $m_0$, and $\alpha$ is a constant. $\delta$ is a phase shift determined by the dimensionality of the FS which acquires a value of either $\delta = 0$ or $ \pm 1/8$ for two- and three-dimensional FSs \cite{kim,nagaosa,kopelevich}, respectively. $\phi_B$ is the Berry phase which, for Dirac and Weyl systems, is predicted to acquire a value $\phi_B = \pi$ \cite{kim,nagaosa,kopelevich}. Finally, $T_D= \hbar /(2 \pi k_B \tau $) is the so-called Dingle temperature from which one extracts $\tau$ or the characteristic quasiparticle scattering time.
In Supplementary Figs. S4 and S5 \cite{supplemental}, we discuss the extraction of the Berry-phase of $\gamma-$MoTe$_2$ via fits to Eq. (1) of the oscillatory signal shown in Fig. 1(f). However, the geometry of the FS of $\gamma-$MoTe$_2$ evolves slightly as the field increases due to the Zeeman-effect, which precludes the extraction of its Berry phase. More importantly, one cannot consistently extract a value $\phi_B \simeq \pi$ when one limits the range in magnetic fields to smaller values in order to minimize the role of the Zeeman-effect. In other words, the dHvA-effect does not provide evidence for the topological character of $\gamma-$MoTe$_2$. Nevertheless, it does indicate that the Dingle temperature decreases as the field increases implying a field-induced increase in the quasiparticle lifetime. This effect should contribute to its large and non-saturating magnetoresistivity. We reported a similar effect for WTe$_2$ \cite{Daniel}.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width = 12 cm]{original_band_structure_and_FS}
\caption{(a) Electronic band-structures of $\gamma-$MoTe$_2$ calculated with and without the inclusion of spin-orbit coupling (SOC). These calculations are based on the crystallographic structure measured at $T=100$ K. (b) Comparison between the calculated electronic bands based upon the crystallographic structures measured at $T = 100$ and $T = 12$ K, respectively. (c) and (d) Respectively, side and top views of the calculated FS. (e), (f), (g) and (h) Fermi surface sheets resulting from electron bands. Notice the marked two-dimensional character of several of the electron-like FS sheets. (i), (k) and (l), Hole-like sheets around the $\Gamma-$point.}
\end{center}
\end{figure*}
\begin{figure*}[htb]
\begin{center}
\includegraphics[width = 10 cm]{MoTe2_FFT_DFT_angular_comparison_unshifted}
\caption{(a) Angular dependence of the FS cross-sectional areas or, through the Onsager relation, de Haas-van Alphen frequencies as with respect to the main crystallographic axes. (b) Experimentally observed dHvA spectra as a function of the frequency $F$ for several angles $\theta$ between the $c-$ and and the $b-$axis and for several values of the angle $-\phi$ between the $c-$ and the $a-$axes.}
\end{center}
\end{figure*}
Since the Berry-phase extracted from the dHvA-effect does not provide support for a topological semi-metallic state in $\gamma-$MoTe$_2$, it is pertinent to ask if the DFT calculations predict the correct electronic band-structure and related FS geometry for this compound, since both are the departing point for the predictions of Refs. \onlinecite{bernevig,felser,bernevig2}. To address this issue, we studied the dHvA-effect as a function of the orientation of the field with respect to the main crystallographic axes. Here, our goal is to compare the angular dependence of the cross-sectional areas determined experimentally, with those predicted by DFT.
Figures 2(a) and 2(b) display both the dHvA (red traces) and the SdH signals (black traces) measured in two distinct single crystals and for two field orientations, respectively along the $c-$ and the $a-$axes. As previously indicated, the dHvA and SdH signals were obtained after fitting a polynomial and subtracting it from the background magnetic torque and magnetoresistivity traces, respectively. The SdH signal was collected from a crystal displaying a $RRR \gtrsim 1000 $ at $T \simeq 25$ mK under fields up to 18 T, while the dHvA one was obtained from a crystal displaying $RRR \gtrsim 2000$ at $T \simeq 35$ mK under fields up to 35 T. Both panels also display the Fast Fourier transform (FFT) of the oscillatory signal. For fields along the $c-$axis, one observes two main peaks at $F_{\alpha} = 231$ T and at $F_{\beta} = 242$ T, as well as their first- and second harmonics and perhaps some rather small frequencies which could result from imperfect background subtraction. We obtain the same two dominant frequencies regardless of the interval in $H^{-1}$ used to extract the FFTs. Supplemental Fig. S6 \cite{supplemental} displays the dHvA signal for $H$ aligned nearly along the $b-$axis along with the corresponding FFT spectra which are again dominated by two prominent peaks. The observation of just two main frequencies for $\mu_0H \| c-$axis is rather surprising since, as we show below, DFT calculations, including the effect of the spin-orbit interaction, predict several pairs of electron-like corrugated cylindrical FSs along with pairs of smaller three-dimensional electron-like sheets in the First-Brillouin zone. Around the $\Gamma-$point, DFT predicts at least a pair of four-fold symmetric helix-like large hole sheets. This complex FS should lead to a rich oscillatory signal, contrary to what is observed. One might argue that the non-observation of all of the predicted FS sheets would be attributable to an experimental lack of sensitivity or to poor sample quality which would lead to low carrier mobility. Nevertheless, our analysis of the Hall-effect within a two-carrier model \cite{qiong}, yields electron- and hole-mobilities ranging between $10^4$ and $10^5$ cm$^2$/Vs at low $T$s which is consistent with both the small residual resistivities and the large resistivity ratios of our measured crystals. Given that the magnetic torque is particularly sensitive to the anisotropy of the FS, such high mobilities should have allowed us to detect most of the predicted FSs, particularly at the very low $T$s and very high fields used for our measurements. Hence, we conclude that the geometry of the FS ought to differ considerably from the one predicted by DFT.
In Figs. 2(c) and 2(d) we plot the amplitude of the main peaks observed in the FFT spectra for fields along the $c-$axis as a function of the temperature. Red lines are fits to the Lifshitz-Kosevich (LK) temperature damping factor, i.e. $x/\sinh x$ with $x = 14.69 \mu T/H$ and with $\mu$ being the effective mass in units of the free electron mass, from which we extract the masses associated with each frequency. As seen, for $H \| c-$axis one obtains $\mu_{\alpha} = 0.85$ $m_0$ and $\mu_{\beta}= 0.8$ $m_0$, which contrasts with the respective values obtained for $H \| a-$axis, namely $\mu_{\alpha, \beta} \simeq 1.5$ $m_0$, see Figs. 2(e) and 2(f). As previously mentioned for $\mu_0H \|b-$axis, we observe two main frequencies, but by reducing the $H^{-1}$ window to focus on the higher field region, we detect additional frequencies (See, Fig. S6 \cite{supplemental}) which are characterized by heavier effective masses, i.e. in the order of $2.5-2.9$ $m_0$. This indicates that $\gamma-$MoTe$_2$ displays a higher anisotropy in effective masses when compared to WTe$_2$ \cite{Daniel}, although these masses are consistent with its sizeable $\gamma_e$ coefficient. Supplemental Fig. S7 \cite{supplemental} displays several traces of the dHvA signal as functions of the inverse field for several angles between all three main crystallographic axes. These traces are used to plot the angular dependence of the FS cross-sectional areas in order to compare these with the DFT calculated ones.
\section{Comparison between experiments and the DFT calculations}
\begin{figure*}[htb]
\begin{center}
\includegraphics[width = 15 cm]{MoTe2_ARPES_vs_bands_fig}
\caption{(a) ARPES energy distribution map along $k_y$ for $k_x=0$ from Ref. \onlinecite{thirupathaiah}. (b) Derivative of the energy distribution map. Vertical blue lines indicate the diameter of the observed hole-pocket, with its area corresponding to a frequency $F \sim 0.33$ kT.
(c) Ribbons obtained from our DFT calculations showing the $k_z$-projection for all bulk bands which are plotted along the same direction as the ARPES energy distribution map in (a). Different colors are chosen to indicate distinct bands. Purple dotted line corresponds to the original position of the Fermi level $\epsilon_F$ according to DFT, while the white one corresponds to the position of $\epsilon_F$ according to ARPES. Notice that the ARPES bands are shifted by $\sim -50$ meV with respect to the DFT ones. (d) ARPES energy distribution map corresponding to a cut along the $k_x-$direction with $k_y=0$, from Ref. \onlinecite{ARPES_Tamai}. Vertical red dotted lines indicate the diameter of the observed electron pockets or $\sim 0.2$ \AA$^{-1}$. SS stands for ``surface-state". (e) DFT Ribbons showing the $k_z$-projection for all bulk bands plotted along the same direction as the ARPES EDM in (d). In both panels black lines depict the original position of $\epsilon_F$ while the yellow lines are guides to the eyes illustrating the difference in energy between the top of the deepest DFT calculated hole-band and its equivalent according to ARPES. }
\end{center}
\end{figure*}
\begin{figure*}[htb]
\begin{center}
\includegraphics[width=17 cm]{MoTe2_F_shifted_bands_v2}
\caption{(a) Experimentally observed dHvA spectra as a function of $F$ for several angles $\theta$ and $-\phi$, where the magenta and the blue lines act
as guides to the eyes and as identifiers of respectively, electron- and hole-like orbits according to the shifted band structure. Clear blue line depicts a possible hole-orbit associated with very small peaks in the FFT spectra. (b) Angular dependence of the dHvA orbits, or frequencies on the FS resulting from the shifted bands in absence of spin-orbit coupling, where magenta and blue markers depict electron- and hole-like orbits on the FS, respectively. Notice the qualitative and near quantitative agreement between the calculations and the experimental observations. (d) Angular dependence of the dHvA frequencies for the shifted electron and hole-bands in the presence of spin-orbit coupling. Here, electron-orbits are depicted by magenta and orange markers while the hole ones are indicated by blue and clear blue markers. The experimental data are better described by the non spin-orbit split bands.}
\end{center}
\end{figure*}
\begin{figure*}[htb]
\begin{center}
\includegraphics[width = 14 cm]{MoTe2_shifted_bands_and_FS_fig_v2}
\caption{(a) Electronic band structure calculated with the inclusion of SOC after the hole-like bands have been shifted by $-50$ meV and the
electron ones by $+35$ meV with the goal of reproducing the observed dHvA frequencies and their angular dependence. Notice that these shifts suppress the crossings between the electron- and hole-bands and therefore the Weyl type-II points of the original band structure. (b) Fermi surface resulting from these band-shifts. (c) FS top view. (d) and (e) Electron-like FS sheets. (f) and (g) Hole-like sheets. }
\end{center}
\end{figure*}
Several recent angle-resolved photoemission spectroscopy (ARPES) studies \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang, ARPES_Liang, ARPES_Xu, ARPES_Tamai,ARPES_Belopolski,thirupathaiah}
claim to find a broad agreement between the band structure calculations, the predicted geometry of the Fermi surface, the concomitant existence
of Weyl type-II points \cite{bernevig,felser,bernevig2}, and the related Fermi arcs on the surface states of $\gamma-$MoTe$_2$.
Several of these experimental and theoretical studies claim that the electronic structure of this compound is particularly sensitive to its precise crystallographic structure.
Inter-growth of the $2H-$phase or the temperature used to collect to X-ray diffraction data, typically around 100 to 230 K, are claimed to have a considerable effect on the calculations \cite{bernevig2, ARPES_Tamai}. Given the few frequencies observed by us, it is pertinent to ask if the mild evolution of the crystallographic structure as a function of the temperature shown in Fig. S3 \cite{supplemental} would affect the geometry of the FS of $\gamma-$MoTe$_2$. To address this question we performed a detailed angular-dependent study of the frequencies extracted from both the SdH and the dHvA effects in $\gamma-$MoTe$_2$ in order to compare these with the angular dependence of the FS cross-sectional areas predicted by the calculations.
In the subsequent discussion we compare the angular dependence of our dHvA frequencies with calculations performed with the Quantum Espresso \cite{QE} implementation of the density functional theory in the GGA framework including spin-orbit coupling (SOC). The Perdew-Burke-Ernzerhof (PBE) exchange correlation functional \cite{PBE} was used with fully relativistic norm-conserving pseudopotentials generated using the optimized norm-conserving Vanderbilt pseudopotentials as described in Ref. ~\onlinecite{ONCVPP}.
The 4\emph{s}, 4\emph{p}, 4\emph{d} and 5\emph{s} electrons of Mo and the 4\emph{d}, 5\emph{s} and 5\emph{p} electrons of Te were treated as valence electrons.
After careful convergence tests, the plane-wave energy cutoff was taken to be 50 Ry and a $k-$point mesh of $20\times 12\times 6$ was used to sample the
reducible Brillouin Zone (BZ) used for the self-consistent calculation. The Fermi surfaces were generated using a more refined $k-$point mesh
of $45\times 25\times 14$. FS sheets were visualized using the XCrysden software \cite{xcrysden}. The related angular dependence of the quantum oscillation frequencies was calculated using the skeaf code \cite{skeaf}. As shown in Fig. 3 the results are very close to those obtained by using the VASP and the Wien2K implementations of DFT (see, Fig. S8 in SI \cite{supplemental}), and also to those reported by Refs. \onlinecite{felser,bernevig2,ARPES_Huang,thirupathaiah}.
Figure 3(a) displays the electronic band structure of $\gamma-$MoTe$_2$, based on its structure determined at $T = 100$ K, with and without the inclusion of SOC. As previously reported \cite{felser,bernevig2}, electron- and hole-bands intersect along the $\Gamma-X$ direction at energies slightly above $\varepsilon_F$ creating a pair of Weyl type-II points. Figure 3(b) shows a comparison between band structures based on the crystallographic lattices determined at 12 K and at 100 K, respectively. Both sets of electronic bands are nearly identical and display the aforementioned crossings between hole- and electron-bands thus indicating that the electronic structure remains nearly constant below 100 K. Figures 3(c) and 3(d) provide a side perspective and a top view of the overall resulting FS within BZ, respectively. The main features of the DFT calculations are the presence of two-dimensional electron pockets, labeled $e_1$ and $e_2$ in Figs.~3(e) and 3(f) and of large ``star-shaped'' hole-pockets near the $\Gamma-$point, labeled as the $h_2$ and the $h_3$ sheets in Figs.~3(k) and 3(l). These electron and hole pockets nearly ``touch''. Due to the broken inversion symmetry, these bands are not Kramer's degenerate,
and hence the spin-orbit split partners of the corresponding electron and hole pockets are located inside the corresponding bigger sheets.
The $h_1$ hole pocket and the $e_3$ and $e_4$ electron pockets are very sensitive to the position of $\varepsilon_F$ disappearing when $\varepsilon_F$ is moved
by only $\pm 15$ meV.
Figures 4(a) and 4(b) present the angular dependence of the calculated and of the measured FFT spectra of the oscillatory signal (raw data in Fig.~S7 \cite{supplemental}), respectively. In this plot the Onsager relation was used to convert the theoretical FS cross-sectional areas into oscillatory frequencies.
In Fig.~4(b) $\theta$ refers to angles between the $c-$ and the $a-$axis, where $\theta=0^{\circ}$ corresponds to $H\parallel c-$axis,
while $\phi$ corresponds to angles between the $c-$ and the $b-$axis, again relative to the $c-$axis.
As seen, there are striking differences between both data sets with the calculations predicting far more frequencies than the measured ones.
More importantly, for fields oriented from the \emph{c}-axis towards either the $a-$ or the $b-$axis, one observes the complete absence of experimental frequencies around $\sim 1$ kT which, according to the calculations, would correspond to the cross-sectional areas of the hole-pockets $h_2$ and $h_3$. In addition, while many of the predicted electron orbits show a marked two-dimensional character, diverging as the field is oriented towards the $a-$ or the $b-$axis, the experimentally observed frequencies show finite values for fields along either axis. This indicates that these orbits are three-dimensional in character, despite displaying frequencies close to those predicted for the $e_1$ and the $e_2$ pockets for fields along the $c-$axis. These observations, coupled to the non-detection of all of the predicted orbits, in particular the large hole $h_2$ and $h_3$ Fermi surfaces, indicate unambiguously that the actual geometry of the FS of $\gamma-$MoTe$_2$ is different from the calculated one. Notice that frequencies inferior to $F = 100$ T, which correspond to the smaller electron- and hole-pockets and which are particularly sensitive to the position of $\varepsilon_F$ as previously mentioned, were not included in Fig. 4(a) for the sake of clarity.
The calculation shows a significant difference between the SOC-split theoretical bands, which is highlighted by the absence of a frequency around 0.5 kT associated with $e_2$ pocket, along with its presence in association with the $e_1$ pocket. This contrast between both orbits is due to the presence of a ``handle-like'' structure (see Fig. 3(e)) in $e_1$ which gives a maximum cross-section close to the BZ edge. However, at this position there is no maximum cross-section within the BZ for the $e_2$ pocket since its ``handle'' is missing (see Fig.~3(f)). This marked difference in topology between the FSs of both spin-orbit split partners indicates that the strength of the SOC provided by the DFT calculations tends to be considerably larger than the one implied by our experiments. In fact, from the twin peaks observed in the experimental FFT spectra having frequencies around 250 T for fields along the \emph{c}-axis, which are likely to correspond to SOC-split bands due to their similar angular dependence, we can infer that the actual SO-splitting is far less significant than the value predicted by the calculations. We have investigated the possibility of an overestimation of the strength of the SOC within our calculations
which is the mechanism driving the DFT prediction of a large number of dHvA frequencies displaying remarkably different angular dependencies.
For instance, we calculated the angular dependence of the FSs without the inclusion of SOC. This leads to just one, instead of a pair of distinct SOC-split bands, which in fact display angular dependencies very similar to those of orbits $h_3$, $e_1$ and $e_3$ in Fig.~4(a). Notice that part of the discrepancy is attributable to the inter-planar coupling which is not well captured by the DFT calculations \cite{Son}. DFT suggests that this compound is van der Waals like by predicting several two-dimensional (i.e. cylindrical like) FS sheets, when the experiments indicate that the overall FS displays a marked three-dimensional character. This indicates that the inter-planar coupling is stronger than implied by DFT. In any case, from Figs.~4(a) and 4(b) and the above discussion, it is clear that there are significant discrepancies between the calculated and the measured FSs.
Given that the proposed Weyl type-II scenario \cite{bernevig,felser,bernevig2} hinges on a possible touching between electron- and hole-pockets, it is critical to understand their exact geometry, or the reason for the disagreement between predictions and experiments, before one can make any assertion on the existence of the Weyl type-II points in $\gamma-$MoTe$_2$.
To understand the source of the disagreement between calculations and our measurements, we now focus on a detailed comparison between our DFT calculations and a selection of ARPES studies. Figure 5(a) corresponds to data from Ref. \onlinecite{thirupathaiah} depicting an ARPES energy distribution map (EDM) along $k_y$ while keeping $k_x =0$. Figure 5(b) plots its derivative. In both figures the $\Gamma-$point corresponds to $k_y=0$. According to the calculations, this EDM should reveal two valence bands intersecting $\varepsilon_F$ around the $\Gamma-$point; the first leading to two small hole-pockets, or the $h_1$ sheets at either side of $\Gamma$, with the second SOC-split band producing the larger $h_2$ and $h_3$ sheets. Instead, ARPES observes just one band intersecting $\varepsilon_F$ which leads to a single FS sheet of cross-sectional area $S_{\text{FS}} \sim \pi (0.1 \text{ \AA}^{-1})^2$ as indicated by the vertical blue lines in Fig.~5(b). This observation by ARPES questions the existence of the band (or of its intersection with $\varepsilon_F$) responsible for the large $h_2$ and $h_3$ hole-pockets with this band being the one previously reported to touch the electron band that produces the $e_1$ pocket and creating in this way the Weyl type-II points \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang,ARPES_Tamai,ARPES_Belopolski}. Notice that our dHvA measurements do not reveal any evidence for the original $h_2$ and $h_3$ sheets, thus, being in agreement with this ARPES observation. Furthermore, the Onsager relation $F= S(\hbar/2\pi e)$ yields a frequency of $\sim 330$ T for $S_{\text{FS}}$ which, in contrast, is close to the frequencies observed by us for $\mu_0H \| c-$axis. Figure 5(c) displays the band structure calculated ``ribbons'' obtained by projecting the $k_z$ dependence of the bands onto the $k_x-k_y$ plane. This representation of the band-structure provides a better comparison with the ARPES EDMs.
As seen, there is a good overall agreement between the ARPES and the DFT bands, as previously claimed \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang,ARPES_Tamai,ARPES_Belopolski}, except for the exact position of $\varepsilon_F$. The purple line depicts the position of $\varepsilon_F$ according to the DFT calculations while the white line depicts the position of $\varepsilon_F$ according to ARPES. A nearly perfect agreement between DFT and ARPES is achievable by shifting the DFT valence bands by $\sim - 50$ meV, which is what ends suppressing the $h_2$ and $h_3$ FS hole sheets from the measured ARPES EDMs. As shown through Figs.~5(d) and 5(e), this disagreement between the ARPES and the DFT bands is observed in the different ARPES studies\cite{ARPES_Tamai}. Figure 5(d) corresponds to an EDM along the $k_x$ $(k_y=0)$ direction of the BZ. As shown in Fig.~5(e), DFT reproduces this EDM quite well. Nevertheless, as indicated by the yellow lines in both figures, which are positioned at the top of the deepest valence band observed by ARPES, the ARPES bands are displaced by $\sim -45$ meV with respect to the DFT ones. Red dotted lines in Fig.~5(d) indicate the cross-sections of the observed electron pockets or $\sim \pi (0.1 \text{ \AA}^{-1})^2$. Therefore, to match our main dHvA frequencies, the electron bands would have to be independently and slightly displaced towards higher energies to decrease their cross-sectional area. To summarize, DFT and ARPES agree well on the overall dispersion of the bands of $\gamma-$MoTe$_2$, but not on their relative position with respect to $\varepsilon_F$.
Therefore, guided by ARPES, we shifted the overall valence bands of $\gamma-$MoTe$_2$, shown in Figs.~3(a) and 3(b), by -50 meV and the electron ones by +35 meV to recalculate the FS cross-sectional areas as a function of field orientation relative to the main crystallographic axes. The comparison between the measured dHvA cross-sectional areas and those resulting from the shifted DFT bands are shown in Fig.~6. Figure 6(a) displays the Fourier spectra, previously shown in Fig.~4(b), with superimposed colored lines identifying shifted electron (magenta) and hole (blue) orbits according to Fig.~6(b) which displays these frequencies as a function of field orientation for shifted non-SOC-split DFT bands. As seen, the qualitative and quantitative agreement is good, but not perfect. In contrast, Fig.~6(c) displays these orbits/frequencies as a function of field orientation for SOC-split DFT bands. Clearly, and as previously discussed, the approach used to evaluate the effect of the SOC in $\gamma-$MoTe$_2$ seems to overestimate it for reasons that remain to be clarified.
Concerning the Weyl physics in $\gamma-$MoTe$_2$, the displacement of the bands, introduced here to explain our observations based on the guidance provided by previous ARPES studies, would eliminate the crossings between the electron- and the hole-bands as shown in Fig.~7(a). Finally, Figs.~7(b) to 7(g) display the geometry of the Fermi surface resulting from the shifted bands. Overall, the FS displays a distinctly more marked three-dimensional character, with the electrons and the-hole pockets remaining well-separated in \emph{k}-space.
This three-dimensionality is consistent with the observations of Ref. \onlinecite{ARPES_Jiang} which finds that the electronic bands do disperse along the $k_z-$direction implying that $\gamma-$MoTe$_2$ cannot be considered a van der Waals coupled solid. Notice that DFT tends to underestimate the inter-planar coupling in weakly coupled compounds which is at the heart of the disagreement between the calculations and our observations.
\section{Conclusions}
In conclusion, quantum oscillatory phenomena reveal that the geometry of the Fermi surface of $\gamma-$MoTe$_2$ is quite distinct from the one predicted by previous electronic band-structure calculations. Our low-temperature structural analysis \emph{via} synchrotron X-ray diffraction measurements indicates the absence of an additional structural transition below the monoclinic to orthorhombic one that would explain this disagreement, while heat-capacity measurements provide no evidence for an electronic phase-transition upon cooling. In contrast, a direct comparison between DFT calculations and the band-structure reported by angle resolved photoemission spectroscopy reveals a disagreement on the position of the valence bands relative to the Fermi-level, with the experimental valence bands shifted by $\sim -50$ meV relative to the DFT ones. Therefore, one should be careful concerning the claims of a broad agreement between the calculations and the electronic bands revealed by ARPES measurements \cite{ARPES_Huang, ARPES_Deng, ARPES_Jiang, ARPES_Liang, ARPES_Xu, ARPES_Tamai}.
Here, we show that it is possible to describe the angular-dependence of the observed de Haas-van Alphen Fermi surface cross-sectional areas by shifting the position of the DFT bands relative to the Fermi level as indicated by ARPES. However, with this adjustment, the Weyl points, which result from band-crossings that are particularly sensitive to small changes in the lattice constants, are no longer present in the band-structure of $\gamma-$MoTe$_2$. Although our approach of modifying the band structure in order to obtain an agreement with both ARPES and de Haas-van Alphen experiments has only a phenomenological basis, our findings do shed a significant doubt on the existence of the Weyl points in the electronic band structure of $\gamma-$MoTe$_2$.
Finally, this study combined with the ARPES results in Ref. \onlinecite{daniel2}, indicate that there ought to be a Lifshitz-transition \cite{wu} upon W doping in the $\gamma-$Mo$_{1-x}$W$_x$Te$_2$ series, leading to the disappearance of the central hole pockets in $\gamma-$MoTe$_2$ in favor of the emergence of hole-pockets at either side of the
$\Gamma-$point in $\gamma-$Mo$_{1-x}$W$_x$Te$_2$.
\begin{acknowledgments}
We acknowledge helpful discussions with R. M. Osgood and A. N. Pasupathy.
J.Y.C. is supported by NSF-DMR-1360863.
L.~B. is supported by DOE-BES through award DE-SC0002613
for experiments under high magnetic fields and at very low temperatures,
and by the U.S. Army Research Office MURI Grant W911NF-11-1-0362 for the synthesis
and physical characterization of two-dimensional materials and their heterostructures.
We acknowledge the support of the HLD-HZDR, member of the European Magnetic Field Laboratory (EMFL).
Research conducted at the Cornell High Energy Synchrotron Source (CHESS) is supported by the NSF \& NIH/NIGMS via NSF award DMR-1332208.
The NHMFL is supported by NSF through NSF-DMR-1157490 and the
State of Florida.
\end{acknowledgments}
|
\section{INTRODUCTION}
A large body of observational evidence has been gathered
in support of an accelerated expansion of the universe at the present time.
This includes, in particular, measurements of the luminosity distance of type Ia supernovae
\cite{Riess:1998cb},
the anisotropy of the cosmic microwave background
\cite{Spergel:2006hy,Ade:2013sjv},
weak lensing \cite{Jain:2003tba},
baryon acoustic oscillations \cite{Eisenstein:2005su},
and the large scale structure of the universe \cite{Tegmark:2003ud}.
To explain the current accelerated expansion of the Universe
is one of the most challenging problems of modern cosmology.
In the current cosmological standard model, the $\Lambda$CDM model,
a small positive cosmological constant is included in the
Einstein field equations to model this acceleration.
It is therefore clearly of considerable interest, to study the influence
of a cosmological constant on further solutions of the Einstein equations,
such as black hole spacetimes, in particular.
For a deeper understanding of the gravitational field of massive objects
and in order to accurately predict observational
effects (such as light deflection, gravitational time delay, perihelion
shift and Lense-Thirring effect),
it is mandatory to have very good knowledge of the motion
of test particles and light rays in the spacetimes of interest.
But only analytical methods allow for arbitrarily high accuracy
of the prediction of this motion and the associated observables.
In the Schwarzschild spacetime
the equations of motion of test particles and light
rays were solved analytically in terms of elliptic functions by Hagihara in
1931 \cite{Y. Hagihara}.
The geodesic equations in the
Reissner-Nordstr\"om, Kerr, and Kerr-Newman space-times have the same
mathematical structure \cite{Chandrasekhar:1985kt} and can be solved analogously.
This analytical method was recently further advanced and applied to the
hyperelliptical case, where the analytical solution of the equations
of motion in the 4-dimensional Schwarzschild-(A)dS,
Reissner-Nordstr\"om-(A)dS and Kerr-(A)dS spacetimes was
presented \cite{Hackmann:2008zza,Hackmann:2008tu,Hackmann:2009nh,Hackmann:2010zz,Grunau:2010gd,Enolski:2010if}.
These mathematical tools were also
applied to the geodesic motion in Taub-NUT
and wormhole spacetimes \cite{Kagramanova:2010bk,Kagramanova:2013mwv},
and to higher dimensional spacetimes, including static black hole spacetimes
and Myers-Perry spacetimes
\cite{Hackmann:2008tu,Kagramanova:2012hw,Diemer:2014lba,Diemer:2013fza},
while in five dimensional black ring spacetimes, the
equations of motion could be solved analytically in special cases
\cite{Grunau:2012ai,Grunau:2012ri}.
Moreover, the motion of test particles was studied in various black
string spacetimes including field theoretical cosmic string
spacetimes and black holes pierced by a black string
\cite{Aliev:1988wv,Galtsov:1989ct,Chakraborty:1991mb,Ozdemir:2003km,Ozdemir:2004ne,Grunau:2013oca,Hackmann:2009rp,Hackmann:2010ir}.
In addition, the geodesic equations were solved analytically in a static black hole
spacetime of $f(R)$ gravity \cite{Soroushfar:2015wqa},
for Ho\v{r}ava-Lifshitz black holes \cite{Enolski:2011id},
and for BTZ and GMGHS black holes
\cite{Soroushfar:2015dfz,Soroushfar:2016yea}.
On the other hand, the large body of current cosmological data
could also be taken to indicate that
General Relativity itself should be extended.
The latter would also be supported by
the necessity for dark matter as revealed from astrophysical and
cosmological observations
(see e.g.~\cite{Bertone:2004pz})
and moreover by theoretical arguments
at the ultraviolet scale (e.g., quantum gravity, initial singularities).
Consequently, numerous theoretical attempts
to model the evolution of the Universe
are based on the modification of gravity
(see e.g.~the reviews
\cite{Capozziello:2011et,Clifton:2011jh,Joyce:2014kja,Jain:2010ka,Koyama:2015vza}).
Popular suggestions to modify gravity include theories with
higher powers of the Riemann and Ricci tensors as well as the curvature scalar $R$.
Lovelock theory \cite{Lovelock:1971yv}
and $f(R)$ gravity \cite{Sotiriou:2008rp,DeFelice:2010aj,Capozziello:2011et}
are such examples, where the Einstein-Hilbert action is
generalized accordingly.
A change of the action has influence on the dynamics
of the Universe, but it may also affect the dynamics at the galactic or solar system scales.
Clearly, any modifications of the action must retain the well tested sector of
General Relativity, like its description of the solar system.
However, modified theories may yield different answers from General Relativity in the
strong field regime.
It is therefore essential to inquire about the existence of black holes and about
their properties in modified theories of gravity
(see e.g.~\cite{Berti:2015itd}).
In general,
the study of black holes in these theories may reveal interesting features
not present in General Relativity.
Focussing on black holes in $f(R)$ theories
\cite{Brevik:2004sd,Cognola:2005de,Saffari:2007zt,delaCruzDombriz:2009et, Capozziello:2009jg, Sebastiani:2010kv, Moon:2011hq,Larranaga:2011fv, Cembranos:2011sr, delaCruzDombriz:2012xy, Hendi:2014mba},
we note that a particular class of solutions
is obtained when the curvature scalar is constant, $R=R_0$.
Taking the trace of the field equations then specifies this constant in terms of the
function $f(R)$ and its derivative at $R_0$. A comparison with General Relativity and its
black hole solutions reveals, that the finite curvature scalar
acts basically like a cosmological constant. In vacuum therefore the
Schwarzschild-(A)dS
and Kerr-(A)dS solutions are recovered, when certain rescalings are performed.
When adding charge to the solutions by including an electromagnetic field,
the Reissner-Nordstr\"om-(A)dS and the Kerr-Newman--(A)dS solutions can be
recovered after certain rescalings,
because the trace of the energy momentum tensor vanishes.
In this paper we study the geodesic motion in the spacetime of the static and
rotating charged black hole (Kerr-Newman-(A)dS spacetime).
Since after certain rescalings the Kerr-Newman-(A)dS black hole of
General Relativity also describes a rotating charged black hole
in $f(R)$ gravity, the current analysis can also be applied
to this black hole solution in $f(R)$ gravity.
Let us mention, however, that the stability of this $f(R)$ black hole
can only be established, after the function $f(R)$ is specified.
We here analyze the possible
orbit types using effective potential techniques and parametric
diagrams. Furthermore, we present the analytical solutions of the
equations of motion for test particles and light. The equations of
motion are of elliptic and hyperelliptic type and the solutions are
given in terms of the Weierstrass $\wp$, $\zeta$ and $\sigma$
functions as well as the Kleinian $\sigma$ functions.
Complete integrability of the geodesic equations is guaranteed
by the presence of four integrals of motion of the Kerr-Newman-(A)dS
spacetime, with the forth one being the Carter constant.
Similar analyses of geodesic motion and solutions of the equations
of motion were presented in the Kerr-(A)dS spacetime \cite{Hackmann:2010zz}
and in the Kerr-Newman spacetime \cite{Hackmann:2013pva},
but in the Kerr-Newman-(A)dS spacetime the geodesic motion
has not been analyzed analytically before in great detail, although some aspects were studied in \cite{Heisnam:2014,Stuchlik:1997gk}.
The analytical solution of the geodesic equation of light and a study of the gravitational lensing and frame dragging of light were presented in \cite{Kraniotis:2014paa}. However, neither an analysis of all possible orbits for particles and light was done nor the full
set of analytic solutions to the equations of motions was found.
Our paper is organized as follows: In Sec.~(\ref{field})
we give a brief review of the field equations
and the metric of the rotating black hole in $f(R)$ gravity
and its connection to General Relativity.
In Sec.~(\ref{rotatinggeodesic}) we present the equations of motion
in the Kerr-Newman-(A)dS spacetime. We analyze the geodesic motion
in Sec.~(\ref{sec:analysis}) and give a list of all possible orbit types.
The analysis is given separately for the static case
(Reissner-Nordstr\"om-(A)dS) and the rotating case (Kerr-Newman-(A)dS).
In Sec.~(\ref{analytical solutions}) we present the full set of
analytical solutions of the geodesic equations in the general rotating case
of the Kerr-Newman-(A)dS black hole. Some example orbits in the static
and the rotating case are shown in Sec.~(\ref{sec:orbits}).
We conclude in Sec.~(\ref{conclusions}).
\boldmath
\section{FIELD EQUATIONS IN $f(R)$ MODIFIED GRAVITY AND RESCALINGS}\label{field}
\unboldmath
In this section we give a brief review of the field equations and the metric of the rotating black hole in $f(R)$, which represents the Kerr-Newman-(A)dS spacetime after certain rescalings.
In four dimensions the action of $f(R)$ gravity with a Maxwell field is given by
\begin{align}\label{action}
S=S_{g}+S_{M},
\end{align}
where $S_{g}$ and $S_{M}$ are the gravitational and the electromagnetic actions
\begin{equation}
S_{g}=\dfrac{1}{16 \pi G} \int d^{4} x \sqrt{ -g }(R+f(R)),
\end{equation}
\begin{equation}\label{action2}
S_{M}=\dfrac{-1}{16 \pi} \int d^{4} x \sqrt{-g }[F_{\mu\nu}F^{\mu\nu}],
\end{equation}
where $G$ is the gravitational constant, which we will set to one,
$g$ is the determinant of the metric, $R$ is the curvature scalar, and $R + f(R)$ is the
function defining the modified gravity theory under consideration.
From the above action, the Maxwell
equations take the form
\begin{equation}
\nabla_{\mu}F^{\mu\nu}=0 ,
\end{equation}
while the field equations in the metric formalism are
\begin{align}\label{rmiyo}
R_{\mu\nu}\big(1+ f'(R)\big)-\frac{1}{2}\big(R+f(R)\big)g_{\mu\nu}
+\big(g_{\mu\nu}\nabla^{2}-\nabla_{\mu}\nabla_{\nu}\big)f'(R)=2T_{\mu\nu},
\end{align}
where $R_{\mu\nu}$ is the Ricci tensor, $\nabla$ denotes the usual
covariant derivative, and the stress-energy tensor of the
electromagnetic field is given by
\begin{equation}
T_{\mu\nu}=F_{\mu\rho}F_{\nu}^{\rho}-\dfrac{g_{\mu\nu}}{4}F_{\rho\sigma}F^{\rho\sigma},
\end{equation}
and has vanishing trace
\begin{equation}
T^{\mu} _{\,\,\mu}=0 .
\label{tracezero}
\end{equation}
Taking the trace of Eq.~(\ref{rmiyo}) under the assumption, that
the curvature scalar is constant, $R=R_{0}$, leads to
\begin{equation}
R_{0}\big(1+f'(R_{0})\big)-2\big(R_{0}+f(R_{0})\big)=0 .
\label{R_0}
\end{equation}
This is the same equation as in the vacuum case, because
the matter field has vanishing trace, Eq.~(\ref{tracezero}),
and it determines the constant value of the curvature scalar
\begin{equation}
R_{0}=\dfrac{2f(R_{0})}{f'(R_{0})-1} .
\end{equation}
Using this relation in Eq.~(\ref{rmiyo}) gives the field equations
\begin{align}
R_{\mu\nu} - \dfrac{1}{2} \dfrac{f(R_{0})}{f'(R_{0})-1} g_{\mu\nu}
=\dfrac{2}{1+f'(R_{0})}T_{\mu\nu}.
\end{align}
Comparison with the Einstein equations in the presence of a cosmological constant
$\Lambda$ then indicates an equivalence of the two sets of equations
for $R_0=4 \Lambda$, when we further rescale the left hand side
of the equations
adequately.
Consequently, up to rescalings, the Kerr-Newman-(A)dS solution of
General Relativity is also a solution of the field equations in $f(R)$ gravity
\cite{Larranaga:2011fv, Cembranos:2011sr, delaCruzDombriz:2012xy}.
Thus, the stationary black hole solution can be obtained in Boyer-Lindquist like
coordinates $(t,r,\theta,\varphi)$ as follows \cite{Larranaga:2011fv}
\begin{align}\label{rotatingmetric}
ds^{2}=-\dfrac{\Delta_{r}}{\rho^{2}} \left[dt-\dfrac{a \sin^{2}\theta
d\varphi}{\Xi}\right]^{2}
+\dfrac{\rho^{2}}{\Delta_{r}}dr^{2}+\dfrac{\rho^{2}}{\Delta_{\theta}}d\theta^{2}
+\dfrac{\Delta_{\theta} \sin^{2}\theta}{\rho^{2}}\left[adt-\dfrac{r^{2}+a^{2}}{\Xi}d\varphi
\right]^{2} ,
\end{align}
with
\begin{align}
\Delta_{r}=(r^{2}+a^{2}) \left(1-\dfrac{R_{0}}{12}r^{2} \right)-2Mr+\dfrac{Q^{2}}{(1+f'(R_{0}))},
\end{align}
\begin{align}
\Xi=1+\dfrac{R_{0}}{12}a^{2}, \qquad
\rho^{2}=r^{2}+a^{2} \cos^{2}\theta, \qquad
\Delta_{\theta}=1+\dfrac{R_{0}}{12}a^{2} \cos^{2}\theta ,
\end{align}
where $ Q $ is the electric charge, $ a $ is the angular momentum per mass of the black hole,
and $R_{0}$ enters like a cosmological constant ($R_{0}=4\Lambda $),
yielding a non-asymptotically flat de Sitter or anti-de Sitter spacetime, when $R_0$ is finite.
Note, that the electric charge enters with a scaling factor in the metric.
As in the Kerr spacetime there is a ringlike singularity defined by $\rho^2=0$,
and the horizons are given by $\Delta_r=0$.
\section{THE GEODESIC EQUATIONS}\label{rotatinggeodesic}
In this section we derive the equations of motion for a rotating charged black hole Eq.(\ref{rotatingmetric}),
using the Hamilton-Jacobi formalism, and later introduce effective potentials for the $r$- and $\theta$-motion.
The Hamilton-Jacobi equation
\begin{equation}\label{Hamilton}
\dfrac{\partial S}{\partial\tau} +\frac{1}{2} \ g^{ij}\dfrac{\partial S}{\partial x^{i}}\dfrac{\partial S}{\partial x^{j}}=0
\end{equation}
can be solved with an ansatz for the action
\begin{equation}
\label{S}
S=\frac{1}{2}\varepsilon \tau - Et+L_{z}\phi +S_{\theta}(\theta) + S_{r} (r).
\end{equation}
The constants of motion are the energy $E$ and the angular momentum $L$ which are given
by the generalized momenta $P_{t}$ and $P_{\phi}$
\begin{equation}\label{constants of motion}
P_{t}=g_{tt}\dot{t}+g_{t \varphi}\dot{\varphi}=-E, \qquad P_{\phi}=g_{\varphi \varphi}\dot\varphi +g_{t \varphi}\dot{t} =L.
\end{equation}
Using Eqs.(\ref{Hamilton})--(\ref{constants of motion}) we get
\begin{align}\label{ds/dr.ds/dtheta}
\Delta_{\theta}\left(\dfrac{\partial S}{\partial\theta}\right)^{2}+\varepsilon a^{2} \cos^{2}\theta -\dfrac{2aEL\Xi - E^{2}a^{2}\sin^{2}\theta}{\Delta_{\theta}}+\dfrac{L^{2}\Xi^{2}}{
\Delta_{\theta}\sin^{2}\theta}=-\Delta_{r}(\dfrac{\partial S}{\partial r})^{2}-\nonumber\\ \varepsilon r^{2}+\dfrac{(a^{2}+r^{2})^{2}E^{2}+a^{2}L^{2} \Xi^{2}-2aEL\Xi (r^{2}+a^{2})}{\Delta_{r}} ,
\end{align}
where each side depends on $r$ or $\theta$ only.
With the separation ansatz Eq.(\ref{S}) and with the help of the Carter constant \cite{Carter:1968rr},
we derive the equations of motion:
\begin{align}\label{Joda}
\rho^{4}\left(\dfrac{dr}{d\tau}\right)^{2}=-\Delta_{r}(K+\varepsilon r^{2})+\big[(a^{2}+r^{2})E-aL\Xi \big]^{2}=R(r),
\end{align}
\begin{align}\label{thetatho}
\rho^{4}\left(\dfrac{d\theta}{d\tau}\right)^{2}=\Delta_{\theta}(K-\varepsilon a^{2}\cos^{2}\theta)-\dfrac{1}{\sin^{2}\theta}\big(aE \sin^{2}\theta -L\Xi \big)^{2} =\Theta(\theta),
\end{align}
\begin{align}
\rho^{2}\left(\dfrac{d\varphi}{d\tau}\right)=\dfrac{aE\Xi (a^{2}+r^{2})-a^{2}\Xi^{2}L}{\Delta_{r}}-\dfrac{1}{\Delta_{\theta}\sin^{2}\theta}(a\Xi E \sin^{2}\theta -\Xi^{2}L),
\end{align}
\begin{align}\label{ttho}
\rho^{2}\left(\dfrac{dt}{d\tau}\right)=\dfrac{E(r^{2}+a^{2})^{2}-aL\Xi(r^{2}+a^{2})}{\Delta_{r}}-\dfrac{\sin^{2}\theta}{\Delta_{\theta}}\left(E a^{2}-\dfrac{L\Xi a}{\sin^{2}\theta}\right).
\end{align}
In the following we will explicitly solve these equations.
Eq.(\ref{Joda}) suggests the introduction of an effective potential $V_{{\rm eff},r}$ such that $V_{{\rm eff},r}=E$
corresponds to $\left(\dfrac{dr}{d\tau}\right)^{2}=0$
\begin{equation}
V_{{\rm eff},r}=\dfrac{L\Xi a\pm \sqrt{\Delta_{r}(K+\varepsilon r^{2})}}{a^{2}+r^{2}},
\end{equation}
where $\left(\dfrac{dr}{d\tau}\right)^{2}\geq 0$ for $E\leq V_{{\rm eff},r}^{-}$ and $E\geq V_{{\rm eff},r}^{+}$.
In the same way an effective potential
corresponding to Eq.(\ref{thetatho}) can be introduced
\begin{align}
V_{{\rm eff},\theta}=\dfrac{L\Xi \pm \sqrt{\Delta_{\theta}\sin^{2}\theta (K-\varepsilon a^{2}\cos^{2}\theta)}}{a \sin^{2}\theta}.
\end{align}
but here $\left(\dfrac{d\theta}{d\tau}\right)^{2}\geq 0$ for $V_{\rm eff,\theta}^{-}\leq E \leq V_{\rm eff,\theta}^{+}$.
\\
Introducing the Mino time $\lambda$ \cite{Mino:2003yg} connected to the proper time
$\tau$ by $\dfrac{d\tau}{d\lambda}=\rho^{2}$, the equations of motions read
\begin{align}\label{d}
\left(\dfrac{dr}{d\lambda}\right)^{2}=-\Delta_{r}(K+\varepsilon r^{2})+\big[(a^{2}+r^{2})E-aL\Xi \big]^{2}=R(r),
\end{align}
\begin{align}
\left(\dfrac{d\theta}{d\lambda}\right)^{2}=\Delta_{\theta}(K-\varepsilon a^{2}\cos^{2}\theta)-\dfrac{1}{\sin^{2}\theta}\big(aE \sin^{2}\theta -L\Xi \big)^{2} =\Theta(\theta),
\end{align}
\begin{align}
\left(\dfrac{d\varphi}{d\lambda}\right)=\dfrac{aE\Xi (a^{2}+r^{2})-a^{2}\Xi^{2}L}{\Delta_{r}}-\dfrac{1}{\Delta_{\theta}\sin^{2}\theta}(a\Xi E \sin^{2}\theta -\Xi^{2}L),
\end{align}
\begin{align}\label{t}
\left(\dfrac{dt}{d\lambda}\right)=\dfrac{E(r^{2}+a^{2})^{2}-aL\Xi(r^{2}+a^{2})}{\Delta_{r}}-\dfrac{\sin^{2}\theta}{\Delta_{\theta}}(E a^{2}-\dfrac{L\Xi a}{\sin^{2}\theta}).
\end{align}
we introduce dimensionless quantities to rescale the parameters
\begin{align}
\tilde{r}=\dfrac{r}{M} , \qquad \tilde{a}=\dfrac{a}{M} , \qquad \tilde{t}=\dfrac{t}{M} , \qquad \tilde{L}=\dfrac{L}{M} , \qquad \tilde{K}=\dfrac{K}{M^{2}} , \nonumber\\ \tilde{R_{0}}=R_{0}M^{2}, \qquad \tilde{Q}=\dfrac{Q}{M},\qquad \gamma = M\lambda.
\end{align}
Then the equations (\ref{d}) - (\ref{t}) can be rewritten as
\begin{align}\label{drd}
\left(\dfrac{d\tilde{r}}{d\gamma}\right)^{2}=-\Delta_{\tilde{r}}(\tilde{K}+\varepsilon
\tilde{r}^{2})+ \left[ (\tilde{a}^{2}+\tilde{r}^{2})E
-\tilde{a}\tilde{L}\Xi \right]^{2}=\tilde{R}(\tilde{r}),
\end{align}
\begin{align}\label{dthetad}
\left(\dfrac{d\theta}{d\gamma}\right)^{2}=\Delta_{\theta}(\tilde{K}
-\varepsilon \tilde{a}^{2}\cos^{2}\theta)
-\dfrac{1}{\sin^{2}\theta}\left(\tilde{a}E \sin^{2}\theta
-\tilde{L}\Xi \right)^{2}=\tilde{\Theta}(\theta),
\end{align}
\begin{align}\label{dphi}
\left(\dfrac{d\varphi}{d\gamma}\right)=\dfrac{ \tilde{a}E \Xi (\tilde{a}^{2}+\tilde{r}^{2})-\tilde{a}^{2}\Xi^{2}\tilde{L}}{\Delta_{\tilde{r}}}
-\dfrac{1}{\Delta_{\theta}\sin^{2}\theta} \big(\tilde{a}\Xi E
\sin^{2}\theta -\Xi^{2}\tilde{L} \big),
\end{align}
\begin{align}\label{dtd}
\left(\dfrac{d\tilde{t}}{d\gamma}\right)=\dfrac{E(\tilde{r}^{2}+\tilde{a}^{2})^{2} - \tilde{a}\tilde{L}\Xi
(\tilde{r}^{2}+\tilde{a}^{2})}
{\Delta_{\tilde{r}}}-\dfrac{\sin^{2}\theta}{\Delta_{\theta}} \left(E
\tilde{a}^{2}-\dfrac{\tilde{L}\Xi \tilde{a}}{\sin^{2}\theta} \right).
\end{align}
\section{ANALYSIS OF THE GEODESIC EQUATIONS}\label{sec:analysis}
In this section we will analyze the geodesic equations and give a list of all possible orbits. First we will study the special case of a static charged black hole (Reissner-Nordstr\"om-(A)dS) and then we will give a full analysis of the general rotating charged black hole solution (Kerr-Newman-(A)dS).
\subsection{The static case}
In this section, we investigate the possible orbit types in the static case with the help of the analytical solutions which are described in previous sections, parameter diagrams (see Fig.\ref{pic:staticparametric-diagrams}) and the effective potential (see Fig.\ref{pic:potentials}).
In the static case $a=0$ the motion is confined to a plane and therefore the geodesic equations reduce to
\begin{equation}\label{dr/dphi}
(\frac{dr}{d\varphi})^2=\frac{r^4}{L^2}(E^2-(1-\frac{2M}{r}+\frac{q^{2}}{r^{2}}-\frac{1}{12}R_{0}{r}^{2})(\varepsilon+\frac{L^2}{r^2}))=:R(r),
\end{equation}
\begin{equation}\label{dr/dt}
(\frac{dr}{dt})^2=\frac{1}{E^2}(1-\frac{2M}{r}+\frac{q^{2}}{r^{2}}-\frac{1}{12}R_{0}{r}^{2})^2(E^2-(1-\frac{2M}{r}+\frac{q^{2}}{r^{2}}-\frac{1}{12}R_{0}{r}^{2})(\varepsilon+\frac{L^2}{r^2})),
\end{equation}
where we introduced $ q^{2}=\frac{Q^{2}}{(1+f^{'}(R_{o}))} $. An effective potential can be defined as
\begin{equation}
V_{eff}=(1-\frac{2m}{r}+\frac{q^{2}}{r^{2}}-\frac{1}{12}R_{0}{r}^{2})(\varepsilon+\frac{L^2}{r^2}).
\label{eqn:veff}
\end{equation}
The shape of an orbit depends on the energy $E$ and the angular momentum $L$
of the test particle or light ray, as well as the charge $ q $ and the cosmological constant $ \Lambda $.
The mass can be absorbed through a rescaling of the radial coordinate and the parameters
\begin{equation}
\tilde{r}=\frac{r}{M},\qquad
\tilde{q}=\frac{q}{M},\qquad \mathcal{L}=\frac{M^{2}}{L^{2}},\qquad \tilde{R_{0}}=\frac{1}{12}R_{0} M^{2}.
\end{equation}
Thus, Eq.~(\ref{dr/dphi}) can be written as
\begin{equation}\label{dr/dphin}
(\frac{d\tilde{r}}{d\varphi})^2=\varepsilon\tilde{R_{0}}\mathcal{L}\tilde{r}^{6}+((E^{2}-\varepsilon)\mathcal{L}+\tilde{R_{0}})\tilde{r}^{4}+(2\varepsilon\mathcal{L})\tilde{r}^{3}-(1+\varepsilon\mathcal{L}\tilde{q}^{2} )\tilde{r}^{2}+2\tilde{r}-\tilde{q}^{2}=R(\tilde{r}).
\end{equation}
In following we give a list of the possible orbits types. Let $\tilde{r}_-$ be the inner horizon and $\tilde{r}_+$ be the outer event horizon.
\begin{enumerate}
\item \textit{Escape orbit} (EO) with range $\tilde{r} \in [r_1, \infty)$ where $r_1>\tilde{r}_+$.
\item \textit{Two-world escape orbit} (TEO) with range $[r_1, \infty)$ where $0<r_1 < r_-$.
\item \textit{Bound orbit} (BO) with range $\tilde{r} \in [r_1, r_2]$ with $r_1, r_2 > r_+$.
\item \textit{Many-world bound orbit} (MBO) with range $\tilde{r} \in [r_1, r_2]$ where $0<r_1 \leq r_-$ and $r_2 \geq r_+$.
\item \textit{Terminating orbit} (TO) with ranges
\begin{enumerate}
\item either $\tilde{r} \in [0, \infty)$ (\textit{Terminating escape orbit} -- TEO)
\item or $\tilde{r} \in [0, r_1]$ with $r_1\geq \tilde{r}_+$ (\textit{Terminating bound orbit} -- TBO).
\end{enumerate}
TOs only occur for $q\neq 0$, otherwise the charge will provide a potential barrier preventing the geodesic from reaching the singularity at $\tilde{r}=0$.
\end{enumerate}
These five regular types of geodesic motion correspond to different arrangements
of the real and positive zeros of $ R(r)$ defining the borders of $ R(r) \geq 0 $ or, equivalently,
$ E^2 \geq V_{eff} $.
Eq.~(\ref{dr/dphin}) implies that $ R(\tilde{r}) \geq 0 $ is a necessary condition for the existence of a
geodesic and, thus, that the positive zeros of $ R(\tilde{r}) $ are the turning points of the orbits.
If for a given set of parameters
$ \tilde{R_{0}}, \tilde{q}, \varepsilon, E^{2}, \mathcal{L} $
the polynomial $ R(\tilde{r}) $ has $n$ positive
zeros, then for varying $ E^{2} $ and $L$
this number can only change if two zeros merge to one.
Solving $ R(\tilde{r})=0, \frac{d R(\tilde{r}) }{d\tilde{r}}=0 $ for $ E^{2} $ and $ \mathcal{L} $, for $ \varepsilon=1 $, yields
\begin{equation}\label{LE^2T}
E^{2}=\frac{(\tilde{r}(\tilde{r}-2)+\tilde{q}^{2}-\tilde{R_{0}}\tilde{r}^{4})^{2}}{\tilde{r}^{2}(\tilde{r}^{2}-3\tilde{r}+2\tilde{q}^{2})},\qquad\quad
\mathcal{L}=-\frac{\tilde{r}^{2}-3\tilde{r}+2\tilde{q}^{2}}{\tilde{r}^{2}(\tilde{R_{0}}\tilde{r}^{4}+\tilde{q}^{2}-\tilde{r})}
\end{equation}
and for $ \varepsilon=0 $, yields
\begin{equation}\label{LE^2N}
\mathcal{L}=\frac{1}{E^{2}}
(\frac{2(1+\sqrt{9-8\tilde{q}^{2}})}{(3+\sqrt{9-8\tilde{q}^{2}})^{3}}-
\tilde{R_{0}}).
\end{equation}
In Fig.~\ref{pic:staticparametric-diagrams},
the results of this analysis are shown for both test particles ($ \varepsilon=1 $)
and light rays ($ \varepsilon=0 $).
\begin{figure}[h]
\centering
\subfigure[$\varepsilon=1$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.25$]{
\includegraphics[width=0.36\textwidth]{static-parameter-eps1.eps}
}
\subfigure[$\varepsilon=0$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.25$]{
\includegraphics[width=0.36\textwidth]{static-parameter-eps0.eps}
}
\caption{Parametric $\mathcal{L}$-$E^2$-diagrams of the $\tilde{r}$-motion.
For $\tilde{R}_0>0$ the polynomial $R(\tilde{r})$ has a single positive zero in region I, three positive zeros in region II
and five positive zeros in region III. If $\varepsilon =0$ then region III vanishes, implying that
stable bound orbits for light do not exist outside the horizons. If $\tilde{R}_0<0$ then there are two positive zeros in the regions I,II, and four positive zeros in region III.}
\label{pic:staticparametric-diagrams}
\end{figure}
In the Parametric $\mathcal{L}$-$E^2$-diagrams three regions of geodesic motion with different numbers of zeros can be identified (in the following $r_{i}<r_{i+1}$ is assumed):
\begin{enumerate}
\item Region I:
\begin{enumerate}
\item $\tilde{R}_0>0$: $R (\tilde{r})$ has a single positive real zero $r_1$ and $ R(\tilde{r})\geq 0 $ for $\tilde{r} \geq r_1$. The only possible orbit type is EO.
\item $\tilde{R}_0<0$: $R (\tilde{r})$ has two positive zeros $r_1, r_2$ and $ R(\tilde{r})\geq 0 $ for $ {r_{1}} \leq \tilde{r} \leq {r_{2}} $. The only possible orbit type is MBO.
\end{enumerate}
\item Region II:
\begin{enumerate}
\item $\tilde{R}_0>0$: $R (\tilde{r})$ has three positive zeros $r_1, r_2, r_3$ with $ R(\tilde{r})\geq 0 $ for $ {r_{1}} \leq \tilde{r} \leq {r_{2}} $ and $ r_{3} \leq\tilde{r} $. Possible orbit types are MBO and EO.
\item $\tilde{R}_0<0$: $R (\tilde{r})$ has two positive zeros $r_1, r_2$ and $ R(\tilde{r})\geq 0 $ for $ {r_{1}} \leq \tilde{r} \leq {r_{2}} $. The only possible orbit type is MBO.
\end{enumerate}
\item Region III:
\begin{enumerate}
\item $\tilde{R}_0>0$: $R (\tilde{r})$ has five positive zeros $r_1, r_2, r_3, r_4, r_5$ with $ R(\tilde{r})\geq 0 $ for $ {r_{1}} \leq \tilde{r} \leq {r_{2}} $,
$ {r_{3}} \leq \tilde{r} \leq {r_{4}} $ and $ r_{5} \leq\tilde{r} $. Possible orbit types are MBO, BO and EO.
\item $\tilde{R}_0<0$: $R (\tilde{r})$ has four positive zeros $r_1, r_2, r_3, r_4$ with $ R(\tilde{r})\geq 0 $ for $ {r_{1}} \leq \tilde{r} \leq {r_{2}} $ and $ {r_{3}} \leq \tilde{r} \leq {r_{4}} $. Possible orbit types are MBO and BO.
\end{enumerate}
\end{enumerate}
Terminating orbits are possible in all three regions iff the black hole is uncharged $\tilde{q}=0$. For light rays only regions I and II appear and therefore stable bound orbits do not exist for $\varepsilon =0$. A summary of possible orbit types for $\tilde{R}_0>0$ and $\tilde{R}_0<0$
can be found in Tables~\ref{tab:staticorbit-types} and~\ref{tab:staticorbit-types2} respectively.
\begin{table}[h]
\begin{center}
\begin{tabular}{|cccc|}\hline
zeros & region & range of $\tilde{r}$ & orbit \\
\hline\hline
1 & I &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-}(-2,0)(4,0)
\end{pspicture}
& TEO
\\ \hline
3 & II &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-*}(-2,0)(0,0)
\psline[linewidth=1.2pt]{*-}(1,0)(4,0)
\end{pspicture}
& MBO, EO
\\
& &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-*}(-1.5,0)(-0.5,0)
\psline[linewidth=1.2pt]{*-}(1,0)(4,0)
\end{pspicture}
& MBO, EO
\\ \hline
5 & III &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-*}(-2,0)(0,0)
\psline[linewidth=1.2pt]{*-*}(1,0)(2,0)
\psline[linewidth=1.2pt]{*-}(3,0)(4,0)
\end{pspicture}
& MBO, BO, EO
\\ \hline\hline
\end{tabular}
\caption{Types of orbits in the spacetime of a static charged black hole for $\tilde{q}\neq 0$ and a positive cosmological constant $\tilde{R}_0>0$. The range of the orbits is represented by thick lines. The dots show the turning points of the orbits. The positions of the horizons are marked by vertical double lines. The single vertical line indicates $\tilde{r}=0$. Terminating orbits exist in all three regions only if $\tilde{q}= 0$.}
\label{tab:staticorbit-types}
\end{center}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{|cccc|}\hline
zeros & region & range of $\tilde{r}$ & orbit \\
\hline\hline
2 & I, II &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-*}(-2,0)(0,0)
\end{pspicture}
& MBO
\\
& II &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-*}(-1.5,0)(-0.5,0)
\end{pspicture}
& MBO
\\ \hline
4 & III &
\begin{pspicture}(-2.5,-0.2)(4,0.2
\psline[linewidth=0.5pt]{->}(-2.5,0)(4,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-1.5,-0.2)(-1.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=1.2pt]{*-*}(-2,0)(0,0)
\psline[linewidth=1.2pt]{*-*}(1,0)(2,0)
\end{pspicture}
& MBO, BO
\\ \hline\hline
\end{tabular}
\caption{Types of orbits in the spacetime of a static charged black hole for $\tilde{q}\neq 0$ and a negative cosmological constant $\tilde{R}_0<0$. The range of the orbits is represented by thick lines. The dots show the turning points of the orbits. The positions of the horizons are marked by vertical double lines. The single vertical line indicates $\tilde{r}=0$. Terminating orbits exist in all three regions only if $\tilde{q}= 0$.}
\label{tab:staticorbit-types2}
\end{center}
\end{table}
For $\tilde{R}_0>0$ a plot of the effective potential introduced in Eq.\eqref{eqn:veff} with example energies corresponding to the different regions is shown in Fig~\ref{pic:potentials}. Here the possible orbit types can be identified.
\begin{figure}[h]
\centering
\subfigure[Effective potential in the range $\tilde{r}\in {[0,250 ]} $ ]{
\includegraphics[width=0.38\textwidth]{static-potential2.eps}
}
\subfigure[Closeup of figure (a)]{
\includegraphics[width=0.38\textwidth]{static-potential1.eps}
}
\caption{Effective potential (blue) with parameters $\varepsilon=1$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.75$ and $\mathcal{L}=0.076$. The horizons are depicted as vertical dashed lines. Example energies of region I, II and III (compare figure \ref{pic:parametric-diagrams} and table \ref{tab:orbit-types}) are given as red horizontal lines.}
\label{pic:potentials}
\end{figure}
\subsection{The rotating case}
In this section we analyse the equations of motion in the rotating case (Kerr-Newman-(A)dS spacetime) and investigate the possible orbit types.
\subsubsection{Types of latitudinal motion}
In this subsection and the next subsection, we use the function $ \tilde{\Theta}(\theta) $ in Eq.~(\ref{dthetad}) and the polynomial $\tilde{R}(\tilde{r}) $ in Eq.~(\ref{drd}), to determine the possible orbits of light and test particles.
First we substitute $\upsilon = \cos^{2}\theta$ with $ \theta \in [0, 1]$ in the function $\tilde{\Theta}(\theta)$:
\begin{align}
\tilde{\Theta}(\upsilon)=(1+\dfrac{\tilde{R}_{0}}{12}\tilde{a}^{2}\upsilon)(\tilde{K}-\varepsilon \tilde{a}^{2} \upsilon)- \left(\tilde{a}^{2}E^{2}(1-\upsilon)-2\tilde{L}\Xi \tilde{a}E+\dfrac{\tilde{L}^{2}\Xi^{2}}{(1-\upsilon)} \right) ,
\end{align}
Geodesic motion is possible if $\tilde{\Theta}(\theta)\geqslant 0$, then real values of the coordinate $ \theta $ are obtained. This condition also implies that $\tilde{K}>0$ for all geodesics with $\tilde{R}_0>-\frac{12}{\tilde{a}^2}$, or $\Lambda >-\frac{3}{\tilde{a}^2}$. From the observational side $\Lambda >-\frac{3}{\tilde{a}^2}$ is always true, since the cosmological constant acquires a very small positive value.
The number of zeros of $ \tilde{\Theta}(\theta) $, which are the turning points of the latitudinal motion, only changes if a zero crosses
$0$ or $1$, or if a double zero occurs. $\upsilon=0$ is a zero of $\tilde{\Theta}$ if
\begin{align}
\tilde{\Theta}(\upsilon =0)=\tilde{K}- \big(\tilde{a}^{2}E^{2}-2\tilde{L}\Xi \tilde{a}E+\tilde{L}^{2}\Xi^{2} \big) ,
\end{align}
and therefore
\begin{align}
\tilde{L}=\dfrac{E\tilde{a}\pm \sqrt{\tilde{K}}}{\Xi}.
\end{align}
Since $\upsilon=1$ is a pole of $\tilde{\Theta}(\upsilon)$ for $\tilde{L}\neq 0$, it is only possible that $\upsilon=1$ is a zero of $\tilde{\Theta}(\upsilon)$ if $\tilde{L}=0$,
\begin{align}
\tilde{\Theta}(\upsilon =1, \tilde{L}=0)=\left(1+\dfrac{\tilde{R}_{0}}{12}\tilde{a}^{2}\right)(\tilde{K}-\varepsilon \tilde{a}^{2}).
\end{align}
To remove the pole of $\tilde{\Theta}(\upsilon)$ at $\upsilon = 1$ we consider
\begin{align}\label{thetanoo}
\tilde{\Theta}^{'}({\upsilon})=(1-\upsilon)\left(1+\dfrac{\tilde{R}_{0}}{12}\tilde{a}^{2}\upsilon\right)(\tilde{K}-\varepsilon \tilde{a}^{2}\upsilon)-\big( \tilde{a}E(1-\upsilon)-\tilde{L}\Xi \big)^{2},
\end{align}
where $ \tilde{\Theta}({\upsilon}) =\frac{1}{1-\upsilon}\tilde{\Theta}^{'}({\upsilon}) $. Then double zeros fulfill the conditions
\begin{align}\label{x}
\tilde{\Theta}^{'}({\upsilon})=0 \qquad \text{and} \qquad \frac{ d\tilde{\Theta}^{'}({\upsilon})}{d\upsilon}=0,
\end{align}
which yields
\begin{align}
\tilde{L}=\dfrac{\left( 6E \pm \sqrt{36 E^{2}+3 \tilde{K}\tilde{R}_{0}} \right) (\tilde{R}_{0}\tilde{a}^{2}+12)}{12\tilde{R}_{0}\tilde{a}\Xi}.
\end{align}
From the condition of $\upsilon=0$ being a zero and the condition of double zeros, we can plot parametric $\tilde{L}$-$E^2$-diagrams, see figure \ref{fig:parameterplot-theta}. These reveal two regions in which geodesic motion is possible. The function $\tilde{\Theta}$ has a single zero $\upsilon_0$ in region a, therefore the geodesics will cross the equatorial plane ($\tilde{K}>(E\tilde{a}-\tilde{L}\Xi)^2$). In region b, the function $\tilde{\Theta}$ has two zeros $\upsilon_1$, $\upsilon_2$, which corresponds to motion above or below the equatorial plane ($\tilde{K}<(E\tilde{a}-\tilde{L}\Xi)^2$). If $\tilde{K}=(E\tilde{a}-\tilde{L}\Xi)^2$, the geodesics will reside in the equatorial plane.
\begin{figure}[h]
\centering
\includegraphics[width=0.38\textwidth]{parameterplot-theta.eps}
\caption{ \label{fig:parameterplot-theta} $\varepsilon =1$, $\tilde{a}=0.4$, $\tilde{K}=0.2$, $\tilde{R_0}=4\cdot 10^{-5}$: Parametric $\tilde{L}$-$E^2$-diagram for the function $\tilde{\Theta}$. $\tilde{\Theta}$ possesses one zero in region a and two zeros in region b. In the grey areas geodesic motion is not possible.}
\end{figure}
\subsubsection{Types of radial motion}
The zeros of the polynomial $\tilde{R}$ are the turning points of orbits of
light and test particles, and therefore $\tilde{R}$ determines the possible
types of orbits
\begin{align}
\tilde{R}(\tilde{r})=-\Delta_{\tilde{r}}(\varepsilon \tilde{r}^{2}
+\tilde{K})+ \big[(\tilde{a}^{2}
+\tilde{r}^{2})E-\tilde{a}\tilde{L}\Xi \big]^{2},
\end{align}
with
\begin{align}
\Delta_{\tilde{r}}=(\tilde{r}^{2} +\tilde{a}^{2})\left(1+\dfrac{\tilde{R_{0}}}{12}\tilde{r}^{2}\right)
-2\tilde{r}+q^2
\end{align}
where we introduced
\begin{align}
q^2=\frac{\tilde{Q}^{2}}{(1+f'(R_{0}))} \ .
\end{align}
There are \emph{bound orbits}, where test particles move back and forth between two turning points, and \emph{escape orbits}, where the black hole is approached, but the test particles turn around at a certain point to escape towards infinity. \emph{Terminating orbits} end in the singularity, if they reach simultaneously $\tilde{r}=0$ and $\vartheta=\frac{\pi}{2}$, such that $\rho^2=0$.
If a test particle crosses the black hole horizons twice or even multiple times,
it can enter another universe. These orbits are called \emph{two-world orbits} or \emph{many-world orbits}. Due to the ring singularity it is possible that a geodesic crosses $\tilde{r}=0$,
which is then called \emph{transit orbit} or \emph{crossover orbit} \cite{Hackmann:2010zz}.
Below we give a list of all possible orbits. Let $\tilde{r}_+$ be the outer event horizon and $\tilde{r}_-$ be the inner horizon of the black hole:
\begin{enumerate}
\item \textit{Transit orbit} (TrO) with range $\tilde{r} \in (-\infty, \infty)$.
\item \textit{Escape orbit} (EO) with range $\tilde{r} \in [r_1, \infty)$ with $r_1>\tilde{r}_+$, or with range $\tilde{r} \in (-\infty, r_1]$ with $r_1<0$.
\item \textit{Two-world escape orbit} (TEO) with range $[r_1, \infty)$ where $0<r_1 < r_-$.
\item \textit{Crossover two-world escape orbit} (CTEO) with range $[r_1, \infty)$ where $r_1 < 0$.
\item \textit{Bound orbit} (BO) with range $\tilde{r} \in [r_1, r_2]$ with
\begin{enumerate}
\item $r_1, r_2 > r_+$ or
\item $ 0 < r_1, r_2 < r_-$
\end{enumerate}
\item \textit{Many-world bound orbit} (MBO) with range $\tilde{r} \in [r_1, r_2]$ where $0<r_1 \leq r_-$ and $r_2 \geq r_+$.
\item \textit{Terminating orbit} (TO) with ranges either $\tilde{r} \in [0, \infty)$ or $\tilde{r} \in [0, r_1]$ with
\begin{enumerate}
\item $r_1\geq \tilde{r}_+$ or
\item $0<r_1<\tilde{r}_-$
\end{enumerate}
\end{enumerate}
The type of an orbit is determined by the number of real zeros of the polynomial $\tilde{R}$. This number changes if a double zero occurs
\begin{equation}
\tilde{R}(\tilde{r})=0 \qquad \text{and} \qquad \frac{d \tilde{R}(\tilde{r})}{d\tilde{r}}=0 \, .
\label{eqn:doublezero}
\end{equation}
Additionally the distinction between positive ($r$) and negative ($r$) zeros is interesting,
since the geodesics can cross $\tilde{r}=0$. The number of positive and negative zeros changes if $\tilde{R}(\tilde{r}=0)=0$. Taking both these conditions into account, we can plot parametric $\tilde{L}$-$E^2$-diagrams, which show five regions with different numbers of zeros. In region I $\tilde{R}$ has no zeros. In region II there are two negative zeros. A negative and a positive zero are possible in region III. Region IV has three positive and a single negative zero. Five positive zeros and a negative zero appear in region V. In figure \ref{pic:parametric-diagrams} examples of the parametric $\tilde{L}$-$E^2$-diagrams of the $\tilde{r}$-motion can be seen. We combined them with the parametric diagrams of the $\theta$-motion.
\begin{figure}[h]
\centering
\subfigure[$\varepsilon=1$, $\tilde{a}=0.7$, $\tilde{K}=12$, $R_0=4\cdot 10^{-5}$, $q=0.7$]{
\includegraphics[width=0.3\textwidth]{parameterplot1.eps}
}
\subfigure[$\varepsilon=1$, $\tilde{a}=0.7$, $\tilde{K}=12$, $R_0=4\cdot 10^{-5}$, $q=0.7$]{
\includegraphics[width=0.3\textwidth]{parameterplot2.eps}
}
\subfigure[$\varepsilon=1$, $\tilde{a}=0.9$, $\tilde{K}=0.3$, $R_0=4\cdot 10^{-5}$, $q=0.2$]{
\includegraphics[width=0.3\textwidth]{parameterplot3.eps}
}
\caption{\label{pic:parametric-diagrams} Combined $\tilde{L}$-$E^2$-diagrams of the $\tilde{r}$-motion (green lines) and $\theta$-motion (blue lines). The dashed green lines show, where $\tilde{R}(\tilde{r}=0)=0$. The polynomial $\tilde{R}$ has no zero in region I, 2 negative zeros in region II, 1 negative and 1 positive zeros in region III, 3 positive and 1 negative zeros in region IV, 5 positive and 1 negative zeros in region V. Inside the grey areas the $\theta$-equation does not allow geodesic motion. In regions marked with the letter a, the orbits cross $\theta=\frac{\pi}{2}$, but not $\tilde{r}=0$. Whereas in regions marked with the letter b, $\tilde{r}=0$ can be crossed but $\theta=\frac{\pi}{2}$ is never crossed.}
\end{figure}
The regions I and II intersect only with region b so here the orbits will not cross the equatorial plane. The regions IV and V only intersect with region a, therefore the orbits will cross the equatorial plane. Region III intersects both with the regions a and b.
In the regions I and II the geodesics can cross $\tilde{r}=0$, but in the regions III, IV and V $\tilde{r}=0$ cannot be crossed.
We conclude that the only way for a geodesic to reach the singularity (Terminating Orbit) is $\tilde{R}(\tilde{r}=0)=0$ and $\tilde{\Theta}(\theta=\frac{\pi}{2})=0$. This is the case if $\tilde{K}=(E\tilde{a}-\tilde{L}\Xi)^2$ and additionally $q=0$.\\
We use the parametric $\tilde{L}$-$E^2$-diagrams and the effective potential of the $\tilde{r}$-equation (see figure \ref{pic:potential}) to determine all possible orbit types. Table \ref{tab:orbit-types} shows an overview. If null geodesics ($\varepsilon =0$) are considered, the region V vanishes from the parametric $\tilde{L}$-$E^2$-diagrams, therefore orbits of type F and G (see table \ref{tab:orbit-types}) are not possible for light. This implies that bound orbits outside the horizons ($\tilde{r}>\tilde{r}_+$) are only possible for particles but not for light.\\
\begin{figure}[h]
\centering
\subfigure[$\varepsilon=1$, $\tilde{a}=0.7$, $\tilde{K}=12$, $R_0=4\cdot 10^{-5}$, $q=0.7$, $\tilde{L}=0.5$]{
\includegraphics[width=0.3\textwidth]{potential1a.eps}
}
\subfigure[closeup of figure (a)]{
\includegraphics[width=0.3\textwidth]{potential1b.eps}
}
\subfigure[$\varepsilon=1$, $\tilde{a}=0.7$, $\tilde{K}=1$, $R_0=4\cdot 10^{-5}$, $q=0.7$, $\tilde{L}=0.5$]{
\includegraphics[width=0.3\textwidth]{potential2.eps}
}
\subfigure[$\varepsilon=1$, $\tilde{a}=0.9$, $\tilde{K}=0.3$, $R_0=4\cdot 10^{-5}$, $q=0.3$, $\tilde{L}=1.25$]{
\includegraphics[width=0.3\textwidth]{potential3a.eps}
}
\subfigure[closeup of figure (d)]{
\includegraphics[width=0.3\textwidth]{potential3b.eps}
}
\subfigure[$\varepsilon=1$, $\tilde{a}=0.9$, $\tilde{K}=0.3$, $R_0=4\cdot 10^{-5}$, $q=0.2$, $\tilde{L}=1.45$]{
\includegraphics[width=0.3\textwidth]{potential4.eps}
}
\caption{ \label{pic:potential} Plots of the effective potential together with examples of energies for the different orbit types of table \ref{tab:orbit-types}. The blue and green curves represent the two branches of the effective potential. In the grey area the red dashed lines correspond to energies. The red dots mark the zeros of the polynomial $R$, which are the turning points of the orbits. In the grey area no motion is possible since $\tilde{R} <0$. In the dashed area the $\theta$-equation does not allow geodesic motion ($\tilde{\Theta} <0$). The vertical black dashed lines show the position of the horizons.}
\end{figure}
The preceding analysis was done for $\tilde{R}_0 >0$ which implies a positive cosmological constant. Since the motion of test particles and light in this spacetime is similar to the Kerr-(A)dS spacetime we refer to \cite{Hackmann:2010zz} for an analysis concerning a negative cosmogical constant. In comparison with the parametric diagrams of \cite{Hackmann:2010zz} it is obvious that the overall behaviour and the possible orbit types are the same, still, there are some differences to the Kerr-(A)dS spacetime caused by the parameter $q$. In the spacetime of a rotating black hole, there are orbits that do not cross $\tilde{r}=0$ and also do not cross the equatorial plane ($\theta=\frac{\pi}{2}$). This occurs in the region IIIb (see figure \ref{pic:parametric-diagrams} and table \ref{tab:orbit-types}), which is not present for the Kerr-(A)dS case $q=0$. In the Kerr-(A)dS spacetime the green dashed line in figure \ref{pic:parametric-diagrams} will coincide with the blue lines from the $\theta$ parametric plot, so that region III is not splitted into an a and a b part. Therefore, in the Kerr-(A)dS case an orbit that crosses $\tilde{r}=0$ will not cross $\theta=\frac{\pi}{2}$ and an orbit that crosses $\theta=\frac{\pi}{2}$ will not cross $\tilde{r}=0$.
A further difference to the Kerr case is that terminating orbits do not exist for $q\neq 0$. Only orbits with $\tilde{K}=(E\tilde{a}-\tilde{L}\Xi)^2$ and simultaneously $q=0$ end in the singularity.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|lccll|}\hline
type & zeros & region & range of $\tilde{r}$ & orbit \\
\hline\hline
A & 0 & Ib &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-}(-4,0)(3.5,0)
\end{pspicture}
& TrO
\\ \hline
B & 2 & IIb &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3.5,0)
\psline[linewidth=1.2pt]{*-}(-3,0)(3.5,0)
\end{pspicture}
& EO, CTEO
\\ \hline
C & 2 & IIIa,b &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-}(-1,0)(3.5,0)
\end{pspicture}
& EO, TEO
\\
C$_-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-}(-0.5,0)(3.5,0)
\end{pspicture}
& EO, TEO
\\
C$_0$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3.5,0)
\psline[linewidth=1.2pt]{*-}(-2.5,0)(3.5,0)
\end{pspicture}
& EO, TO/TEO
\\ \hline
D & 4 & IVa &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-1,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, MBO, EO
\\
D$_-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-0.5,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, MBO, EO
\\
D$_+$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-1,0)(1,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, MBO, EO
\\
D$_\pm$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-0.5,0)(1,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, MBO, EO
\\
D$_0$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2.5,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, TO/MBO, EO
\\
D$_0+$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2.5,0)(1.0,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, TO/MBO, EO
\\ \hline
E & 4 & IVa &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-}(-1,0)(3.5,0)
\end{pspicture}
& EO, BO, TEO
\\
E$_-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-}(-0.5,0)(3.5,0)
\end{pspicture}
& EO, BO, TEO
\\
E$_0$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2.5,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-}(-1,0)(3.5,0)
\end{pspicture}
& EO, TO/BO, TEO
\\
E$_0-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2.5,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-}(-0.5,0)(3.5,0)
\end{pspicture}
& EO, TO/BO, TEO
\\ \hline
F & 6 & Va &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-1,0)(1.5,0)
\psline[linewidth=1.2pt]{*-*}(2,0)(2.5,0)
\psline[linewidth=1.2pt]{*-}(3,0)(3.5,0)
\end{pspicture}
& EO, MBO, BO, EO
\\
F$_-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-0.5,0)(1.5,0)
\psline[linewidth=1.2pt]{*-*}(2,0)(2.5,0)
\psline[linewidth=1.2pt]{*-}(3,0)(3.5,0)
\end{pspicture}
& EO, MBO, BO, EO
\\
F$_+$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-1,0)(1.0,0)
\psline[linewidth=1.2pt]{*-*}(2,0)(2.5,0)
\psline[linewidth=1.2pt]{*-}(3,0)(3.5,0)
\end{pspicture}
& EO, MBO, BO, EO
\\ \hline
G & 6 & Va &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-*}(-1,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, BO, MBO, EO
\\
G$_-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-*}(-0.5,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, BO, MBO, EO
\\
G$_0$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2.5,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-*}(-1,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, TO/BO, MBO, EO
\\
G$_0-$ & & &
\begin{pspicture}(-4,-0.2)(3.5,0.2
\psline[linewidth=0.5pt]{->}(-4,0)(3.5,0)
\psline[linewidth=0.5pt](-2.5,-0.2)(-2.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](-0.5,-0.2)(-0.5,0.2)
\psline[linewidth=0.5pt,doubleline=true](1,-0.2)(1,0.2)
\psline[linewidth=1.2pt]{-*}(-4,0)(-3,0)
\psline[linewidth=1.2pt]{*-*}(-2.5,0)(-1.5,0)
\psline[linewidth=1.2pt]{*-*}(-0.5,0)(1.5,0)
\psline[linewidth=1.2pt]{*-}(2,0)(3.5,0)
\end{pspicture}
& EO, TO/BO, MBO, EO
\\ \hline\hline
\end{tabular}
\caption{\label{tab:orbit-types} Types of orbits in the Kerr-Newman-dS-spacetime ($\tilde{R}_0>0$). The range of the orbits is represented by thick lines. The dots show the turning points of the orbits. The positions of the event horizon and the Cauchy horizon are marked by a vertical double line. The cosmological horizon is not displayed here since it is not relevant for the orbits. The single vertical line indicates $\tilde{r}=0$.}
\end{center}
\end{table}
\section{ANALYTICAL SOLUTION OF THE GEODESIC EQUATIONS}\label{analytical solutions}
In this section, we will present the analytical solutions of the geodesic equations (\ref{drd})--(\ref{dtd}) in the Kerr-Newman-(A)dS spacetime. We will treat each equation separately and give the solutions in terms of the Weierstrass $\wp$, $\zeta$ and $\sigma$ functions as well as the Kleinian $\sigma$ function.
\subsection{$\theta$ motion}\label{thetam}
We start with the differential equation (\ref{dthetad}) describing the $\theta$ motion
\begin{align}
\left(\dfrac{d\theta}{d\gamma} \right)^{2}=
\tilde{\Theta}(\theta)=\Delta_{\theta}(\tilde{K}-\varepsilon
\tilde{a}^{2}\cos^{2}\theta)-\dfrac{1}{\sin^{2}\theta} \left(
\tilde{a}E \sin^{2}\theta -\tilde{L}\Xi \right)^{2},
\end{align}
and substitute $\upsilon=cos^{2}\theta$ to simplify the equation
\begin{align}\label{doo}
\left(\dfrac{d\upsilon}{d\gamma}\right)^{2}=4\upsilon \tilde{\Theta}^{'}({\upsilon})=4\upsilon(1-\upsilon)\left(1+\dfrac{\tilde{R}_{0}}{12}\tilde{a}^{2}\upsilon\right)(\tilde{K}-\varepsilon \tilde{a}^{2}\upsilon)-4\upsilon\big( \tilde{a}E(1-\upsilon)-\tilde{L}\Xi \big)^{2}.
\end{align}
\subsubsection{Timelike geodesics}
The differential equation (\ref{doo}) is of elliptic type, since $4\upsilon \tilde{\Theta}^{'}({\upsilon})$ is in general a polynomial of order four. Here we consider the case $ \varepsilon = 1 $. Assuming that $\tilde{\Theta}^{'}({\upsilon})$
has only simple zeros, equation (\ref{doo}) can be solved in terms of the
Weierstrass elliptic $\wp$ function. To get the solution we
transform $4\upsilon \tilde{\Theta}^{'}({\upsilon})$ into the Weierstrass
form $(4y^{3}-g_{2}y-g_{3})$ with the constants $g_{2}$ and $g_{3}$.
First, we apply the substitution $\upsilon=\xi^{-1}$ yielding
\begin{align}\label{dkhi}
\left(\dfrac{d\xi}{d\gamma}\right)^{2}=\tilde{\Theta}_{\xi},
\end{align}
where
\begin{align}\label{khiii}
\tilde{\Theta}_{\xi}=:4\xi^{3} \left[ \tilde{K}-(E\tilde{a}
-\tilde{L}\Xi)^{2} \right] +4\xi^{2} \left[
\tilde{a}^{2}(-\varepsilon+\frac{1}{12}\tilde{K}\tilde{R}_{0}
+2E^{2})-(\tilde{K}+2E\tilde{L}\Xi \tilde{a}) \right] \nonumber\\
+4\xi \left[ \tilde{a}^{2}(-\frac{1}{12}\tilde{K}\tilde{R}_{0}
+\varepsilon -E^{2}-\frac{1}{12}\tilde{R}_{0}\tilde{a}^{2}
\varepsilon) \right] +\frac{1}{3}\tilde{R}_{0}\tilde{a}^{4}\varepsilon
=:\sum_{i=1}^3 a_{i}\xi^{i}
\end{align}
is a now a polynomial of order three. Second, we substitute $\xi=\frac{1}{a_{3}}(4y-\frac{a_{2}}{3})$, which gives
\begin{align}\label{dy}
(\dfrac{dy}{d\gamma})^{2}=4y^{3}-g_{2}y-g_{3},
\end{align}
where the Weierstrass invariants are
\begin{align}
g_{2}=\frac{1}{16} \left( \frac{4}{3} a_{2}^{2}-4a_{1}a_{3} \right),
\end{align}
\begin{align}\label{g3}
g_{3}=\frac{1}{16} \left( \frac{1}{3}a_{1}a_{2}a_{3}
-\frac{2}{27}a_{2}^{3}-a_{0}a_{3}^{2} \right).
\end{align}
The differential equation (\ref{dy}) represents an elliptic intregal of the first kind, which can be solved by \cite{Hackmann:2008zza,M.Abramowitz, E. T. Whittaker}
\begin{align}\label{PW}
y(\gamma)=\wp \left(\gamma -\gamma_{\theta , in};g_{2},g_{3} \right).
\end{align}
Finally, the solution of Eq.(\ref{dthetad}) is given by
\begin{align}
\theta \big(\gamma \big)=\arccos \big( \pm \sqrt{\dfrac{a_{3}}{4\wp
(\gamma -\gamma_{\theta ,in}; g_{2},g_{3})-\frac{a_{2}}{3}}} \big),
\end{align}
where $\gamma_{\theta ,in}=\gamma_{0}+
\int_{y_{0}}^{\infty}\dfrac{dy'} {\sqrt{4y'^{3}-g_{2}y'-g_{3}}}$
and $y_{0}=\dfrac{a_{3}}{4 cos^{2}(\theta_{0})}+\dfrac{a_{2}}{12}$
depends only on the initial values $\gamma_{0}$ and $\theta_{0}$.
Since the $\theta$ motion is symmetric with respect to the equatorial plane
$\theta=\frac{\pi}{2}$, the sign of the square root can be chosen so that $\theta(\gamma)$
is either in $(0,\frac{\pi}{2})$ (positive sign) or in
$(\frac{\pi}{2},\pi)$ (negative sign).
\subsubsection{Null geodesics}
The differential
equation (\ref{doo}) for $ \varepsilon = 0 $ is already a
polynomial of degree three and, thus, with the standard substitution
$\upsilon =\frac{1}{b_{3}} \big(4y-\frac{b_{2}}{3} \big)$ where
$4\upsilon \tilde{\Theta}^{'}({\upsilon})=\sum_{i=1}^3 b_{i}\upsilon^{i}$
transforms the problem to the form Eq.~(\ref{dy}). The solution is
then given by
\begin{align}
\theta \left( \gamma \right)=\arccos \left( \pm \sqrt{\frac{4}{b_{3}}\wp
(\gamma - \gamma_{\theta ,in}; g_{2},g_{3})-\frac{b_{2}}{3 b_{3}}}
\right)
\end{align}
where $\gamma_{\theta ,in }$,$g_{2}$, and $g_{3}$ are as above with
$a_{i}$ replaced by $b_{i}$.
\subsection{r motion}\label{rmotion}
The dynamics of $r$ are described by the differential equation (\ref{drd})
\begin{align}
\left(\dfrac{d\tilde{r}}{d\gamma}\right)^{2}
=\tilde{R}(\tilde{r})=\Delta_{\tilde{r}}(-\varepsilon
\tilde{r}^{2}-\tilde{K})+ \big[ (\tilde{a}^{2}+\tilde{r}^{2})E
-\tilde{a}\tilde{L}\Xi \big]^{2} \, .
\end{align}
Here the solution procedure is more complicated because $\tilde{R}$ is in general a polynomial of order six. However, for null geodesics the order of the polynomial is reduced to four. In the following we will consider timelike and null geodesics separately.
\subsubsection{Null geodesics}
Considering light, i.e. $\varepsilon=0$,
$\tilde{R}$ is simplified to a polynomial of degree four and therefore the differential
equation (\ref{drd}) is of elliptic type. Then we can solve it using the method of section \ref{thetam}.
By substituting first $\tilde{r}=\xi^{-1}+\tilde{r}_{\tilde{R}}$,
where $\tilde{r}_{\tilde{R}}$ is a zero of $\tilde{R}$, and then
$\xi=\frac{1}{b_{3}}(4y- \frac{b_{2}}{3})$, where
$b_{i}=\frac{1}{(4-i)!}\frac{d^{(4-i)}\tilde{R}}{d\tilde{r}^{(4-i)}}(\tilde{r}_{\tilde{R}})$,
Eq.~(\ref{drd}) aquires the form of Eq.~(\ref{dy}). Again, this can be solved with the help of the Weierstrass elliptic $\wp$ function, so that the result is
\begin{align}\label{rr}
\tilde{r}(\gamma)=\dfrac{b_{3}}{4\wp(\gamma
-\gamma_{\tilde{r},in};g_{2},g_{3})-\frac{b_{2}}{3}}+\tilde{r}_{\tilde{R}},
\end{align}
where
$\gamma_{\tilde{r},in}=\gamma_{0}+\int_{y_{0}}^{\infty}\dfrac{dy'}
{\sqrt{4y'^{3}-g_{2,r}y'-g_{3,r}}}$ and
$y_{0}=\frac{b_{3}}{4(\tilde{r}_{0}-r_{R})}+\frac{b_{2}}{12}$
depends only on the initial values $\gamma_{0}$ and $\tilde{r}_{0}$
and $g_{2}$, $g_{3}$ are defined as in Eq.(\ref{g3}) with
$a_{i}=b_{i}$.
\subsubsection{Timelike geodesics}
Considering particles, i.e.
$\varepsilon=1$, and assuming that $\tilde{R}$ has only simple zeros,
the differential equation (\ref{drd}) is of hyperelliptic type. As presented in \cite{Hackmann:2010zz}, this equation can be solved in terms of derivatives of the Kleinian $\sigma$ function. To begin with, Eq.~(\ref{drd}) is transformed into the standard form with the substitution
$\tilde{r}=\pm \frac{1}{u}+\tilde{r}_{\tilde{R}}$ where $\tilde{r}_{\tilde{R}}$ is a zero of $\tilde{R}$.
Then we get
\begin{align}\label{udu}
\left( u\dfrac{du}{d\gamma} \right)^{2}=\tilde{R}_{u},
\end{align}
where
\begin{align}\label{Ru}
\tilde{R}_{u}=\sum_{i=0}^{5} c_{i} u^{i}, \qquad
c_{i}=\dfrac{\big( \pm 1
\big)^{i}}{(6-i)!}\dfrac{d^{(6-i)}\tilde{R}}{d
u^{(6-i)}}(\tilde{r}_{\tilde{R}}).
\end{align}
A separation of variables leads to
\begin{equation}\label{Phi}
\gamma - \gamma_{0}
=\int_{u_{0}}^{u}\frac{udu}{\sqrt{\tilde{R}_{u}}},
\end{equation}
where $u_{0}=u(\gamma_{0})$. Considering the solution of the
integral (\ref{Phi}), we have to address two points. First, due to the two branches of the square root the integrand is not well defined in the complex
plane. Second, the solution $ u(\gamma) $ should not depend on the chosen path of
integration \cite{Hackmann:2008zza}. Let $ \zeta $ be a closed
integration path and
\begin{equation}\label{Omega}
\omega=\oint_{\zeta}\frac{udu}{\sqrt{\tilde{R}_{u}}},
\end{equation}
then also
\begin{equation}\label{PhiOmega}
\gamma - \gamma_{0}-\omega
=\int_{u_{0}}^{u}\frac{udu}{\sqrt{\tilde{R}_{u}}},
\end{equation}
should be true. Hence, the solution
$ u(\gamma) $
of our problem has to fulfill
\begin{equation}\label{PO}
u(\gamma)=u(\gamma - \omega)
\end{equation}
for every $ \omega\neq{0} $ obtained from Eq.~(\ref{Omega}). These
two issues can be solved if we consider Eq.~(\ref{Phi}) to be
defined on the Riemann surface $ y^{2}=\tilde{R}_{u}(x) $ of genus $
g=2 $ and introduce a basis of canonical holomorphic and meromorphic
differentials $ dz_{i} $ and $ dr_{i} $, respectively,
\begin{equation}\label{dz}
dz_{1}=\frac{dx}{\sqrt{\tilde{R}_{x}}}, \qquad\qquad\qquad
dz_{2}=\frac{xdx}{\sqrt{\tilde{R}_{x}}},
\end{equation}
\begin{equation}\label{dr}
dr_{1}=\frac{a_{3}x+2a_{4}x^{2}+3a_{5}x^{3}}
{4\sqrt{\tilde{R}_{x}}}dx, \qquad\qquad\qquad
dr_{2}=\frac{x^{2}dx}{4\sqrt{\tilde{R}_{x}}},
\end{equation}
and real $2\omega_{ij}$, $2\eta_{ij}$ and imaginary
$2\omega^{\prime}_{ij}$, $2\eta^{\prime}_{ij}$ period matrices
\begin{equation}\label{2w}
2\omega_{ij}=\oint_{a_{j}} {dz_{i}}, \qquad\qquad\qquad
2\omega^{\prime}_{ij}=\oint_{b_{j}} {dz_{i}},
\end{equation}
\begin{equation}\label{2n}
2\eta_{ij}=\oint_{a_{j}} {dr_{i}}, \qquad\qquad\qquad
2\eta^{\prime}_{ij}=\oint_{b_{j}} {dr_{i}}.
\end{equation}
The equation (\ref{Phi}) is a hyperelliptic integral of the first kind and can be
solved by \cite{Hackmann:2010zz,Enolski:2010if}
\begin{align}\label{ugamma}
u(\gamma)=-\frac{\sigma_{1}}{\sigma_{2}}(\gamma_{\sigma}),
\end{align}
where $ \sigma_{i} $ is the $i$-th derivative of the Kleinian $\sigma$
function
\begin{equation}
\sigma(z)=Ce^{z^{t}kz} \theta[K_{\infty}](2\omega^{-1}z;\tau),
\end{equation}
which is given by the Riemann $\theta $-function with characteristic
$K_{\infty} $,
\begin{equation}
\theta(z;\tau)=\sum_{m\in \mathbb{Z}^{2}}e^{i{\pi}m^{t}({\tau}m+2z)}.
\end{equation}
A number of parameters enter here: the symmetric Riemann matrix
$\tau=(\omega^{-1}\omega^{\prime}) $, the period-matrix $(2\omega,
2\omega^{\prime}) $, the period-matrix of the second kind $ (2\eta,
2\eta^{\prime}) $, the matrix $ k=\eta(2\omega)^{-1} $, and the
vector of Riemann constants with base point at infinity $
2K_{\infty} = (0, 1)^{t} + (1, 1)^{t}\tau$. The constant $C$ can be
given explicitly, see e.g.~\cite{V. M. Buchstaber}, but is not
important here. In Eq.(\ref{ugamma}) the argument $\gamma_{\sigma}$ is
an element of the one-dimensional $\sigma$ divisor: $ \gamma_{\sigma} =
(f(\gamma -\gamma_{\tilde{r},in}),\gamma
-\gamma_{\tilde{r},in})^{t} $ with
$\gamma_{\tilde{r},in}=\sqrt{c_{5}}\gamma_{0}+\int_{u_{0}}^{\infty}\dfrac{u'
du'}{\sqrt{\tilde{R}_{u'}}}$ and $u_{0}=\pm
(\tilde{r}_{0}-\tilde{r}_{\tilde{R}})^{-1}$ depends only on the
initial values $\gamma_{0}$ and $\tilde{r}_{0}$, and the function $
f$ can be found from the vanishing condition
$\sigma((f(x),x)^{t})=0$ \cite{Hackmann:2010zz}, so it describes the $\theta$-divisor. Then the solution of the
$\tilde{r}$ equation is given by
\begin{align}
\tilde{r}(\gamma)=\mp \frac{\sigma_{2}}{\sigma_{1}}(\gamma_{\sigma}).
\end{align}
Here the sign depends on the sign that was chosen in the substitution
$\tilde{r}=\pm \frac{1}{u}+\tilde{r}_{\tilde{R}}$. The functions $ \sigma_{1}
$ and $ \sigma_{2} $ depend on $\gamma_{\sigma},\omega, \eta, \tau $ and also on the polynomial $ \tilde{R}_{u}$
according to Eqs.~($ \ref{Phi}-\ref{2n} $), which contains all the
parameter-dependence of the modified gravity solution. The solution of $
\tilde{r} $ is the analytic solution of the equation of motion of a
test particle in the Kerr-Newman-(A)dS spacetime. This solution is valid in all regions of
this spacetime.
\subsection{$\varphi$ motion}\label{fii}
The $\varphi$-equation (\ref{dphi}) can be rewritten using the
$\tilde{r}$- and $\theta$-equations, (\ref{drd}) and (\ref{dthetad})
\begin{align}
d\varphi = \dfrac{\tilde{a}E \Xi
(\tilde{a}^{2}+\tilde{r}^{2})-\tilde{a}^{2}\Xi^{2}\tilde{L}}{\Delta_{\tilde{r}}\sqrt{\tilde{R}}}d\tilde{r}
- \dfrac{\tilde{a}\Xi E
\sin^{2}\theta-\Xi^{2}\tilde{L}}{\Delta_{\theta}\sin^{2}\theta
\sqrt{\tilde{\Theta}(\theta)}}d\theta .
\end{align}
Intergrating this equation gives an $\tilde{r}$-dependent integral
$I_{r}$ and a $\theta$-dependent integral $I_{\theta}$ which can be
solved separately
\begin{align}\label{phi}
\varphi - \varphi_{0}= \int_{\tilde{r}_{0}}^{\tilde{r}}
\dfrac{\tilde{a}E \Xi
(\tilde{a}^{2}+\tilde{r}^{2})-\tilde{a}^{2}\Xi^{2}\tilde{L}}{\Delta_{\tilde{r}}\sqrt{\tilde{R}}}d\tilde{r}
- \int_{\theta_{0}}^{\theta}\dfrac{\tilde{a}\Xi E
\sin^{2}\theta-\Xi^{2}\tilde{L}}{\Delta_{\theta}\sin^{2}\theta
\sqrt{\tilde{\Theta}(\theta)}}d\theta = I_{r}-I_{\theta} .
\end{align}
\subsubsection{The $\theta$-dependent integral}
Let us first consider the $\theta$-dependent integral
\begin{align}
I_{\theta}=\int_{\theta_{0}}^{\theta}\dfrac{(\tilde{a}E\Xi
\sin^{2}\theta - \Xi^{2}\tilde{L})d\theta}
{\Delta_{\theta}\sin^{2}\theta \sqrt{\tilde{\Theta}(\theta)}},
\end{align}
which can be simplified by the substitution $\upsilon=\cos^{2}\theta$
\begin{align}
I_{\theta}=\mp\int_{\upsilon_{0}}^{\upsilon}\dfrac{\tilde{a}E \Xi
(1-\upsilon)-\Xi^{2}\tilde{L}}{\Delta_{\upsilon}(1-\upsilon)\sqrt{4\upsilon
\tilde{\Theta}^{'}({\upsilon})}}d\upsilon '\, ,
\end{align}
where the polynomial $\Theta^{'}({\upsilon})$ is given in Eq.~(\ref{thetanoo}) and
$\Delta_{\upsilon}=1+\frac{\tilde{R}_{0}}{12}\tilde{a}^{2}\upsilon$.
Assuming $4\upsilon\Theta^{'}({\upsilon})$ has only simple zeros and is a polynomial of order four, then $I_{\theta}$ is an elliptic integral of the third kind. In this case the solution to $I_{\theta}$ is given by~\cite{Hackmann:2010zz}
\begin{align}\label{thetafi}
I_{\theta}=\frac{|a_{3}|}{a_{3}} \bigg\lbrace(\tilde{a}\Xi
E-\Xi^{2}\tilde{L})(\upsilon - \upsilon_{0})-
\sum_{i=1}^{4}\dfrac{a_{3}}{4\wp'(\upsilon_{i})}
\bigg(\zeta(\upsilon_{i})(\upsilon -\upsilon_{0})+\log
\dfrac{\sigma(\upsilon
-\upsilon_{i})}{\sigma(\upsilon_{0}-\upsilon_{i})}+2\pi i k_{i}
\bigg) \nonumber\\ \left(\tilde{a}^{3}\frac{\tilde{R_{0}}}{12}(E\Xi
-\tilde{a}\frac{\tilde{R_{0}}}{12}\Xi^{2}
\tilde{L})(\delta_{i1}+\delta_{i2})+\Xi^{2}
\tilde{L}(\delta_{i3}+\delta_{i4}) \right) \bigg\rbrace ,
\end{align}
where the coefficients $a_{i}$ of the polynomial $4\upsilon\Theta^{'}({\upsilon})$ are given in subsection \ref{thetam} and
\begin{align}
\wp(\upsilon_{1})=\frac{a_{2}}{12}-\frac{1}{48}\tilde{a}^{2}
\tilde{R_{0}}a_{3}=\wp(\upsilon_{2}),
\nonumber\\
\wp(\upsilon_{3})=\frac{a_{2}}{12}+\frac{1}{4}a_{3}=\wp(\upsilon_{4}) \, .
\end{align}
Also we have $\upsilon=\upsilon(\gamma)=\gamma-\gamma_{\theta,in}$, where
$\gamma_{\theta,in}$ is defined in Eq.~(\ref{PW}) and
$\upsilon_{0}=\upsilon(\gamma_{0})$. The different branches of the logarithm are represented by the integers $k_{i}$. The details of the computation can be
found in ref.~\cite{Hackmann:2010zz}.
\subsubsection{The $r$-dependent integral}
Next we will solve the
$\tilde{r}$-dependent integral
\begin{align}
I_{r}=\int_{\tilde{r}_{0}}^{\tilde{r}}\dfrac{\tilde{a}E
\Xi (\tilde{a}^{2}+\tilde{r}^{2})-\tilde{a}^{2}\Xi^{2}\tilde{L}}
{\Delta_{\tilde{r}}\sqrt{\tilde{R}}}d\tilde{r}.
\end{align}
Here we will distinguish between timelike and null geodesics as the equation simplifies in the latter case.\\
\paragraph{Null geodesics}
Considering light, i.e.~$\varepsilon=0$,
the polynomial $\tilde{R}$ is of order four, and therefore $I_{r}$ is an elliptic integral of the third kind and can be solved
analogously to $I_{\theta}$. We apply the same substitutions
$\tilde{r}=\frac{1}{\xi}+\tilde{r}_{\tilde{R}}$ and
$\xi=\frac{1}{b_{3}}(4y-\frac{b_{2}}{3})$, as in subsection
(\ref{rmotion}) for the case $\varepsilon=0$, then perform a partial
fraction decomposition, and finally substitute $y=\wp
(\upsilon)$. Then we get
\begin{align}
\frac{b_{3}}{\mid
b_{3}\mid}I_{r}=\sum_{i=1}^{4}C_{i}\int_{\upsilon_{0}}^{\upsilon}\dfrac{d\upsilon}{\wp
(\upsilon)-y_{i}}-\dfrac{\tilde{a}E\Xi
(\tilde{a}^{2}+\tilde{r}^{2}_{\tilde{R}})-\tilde{a}^{2}\Xi^{2}\tilde{L}}{\Delta_{\tilde{r}=
\tilde{r}_{\tilde{R}}}}\int_{\upsilon_{0}}^{\upsilon}d\upsilon \, ,
\end{align}
where $b_{3}$ is given in Eq.~(\ref{g3}), and the $y_{i}$ are the four zeros of
$\Delta_{y(\tilde{r})}$. The constants $C_{i}$ arise from the partial
fraction decomposition and depend on the parameters of the test particle and the metric.
The integrand $(\wp
(\upsilon)-y_{i})^{-1}$ has simple poles $\upsilon_{i1}$,
$\upsilon_{i2}$, where $\wp (\upsilon_{i1})=y_{i}=\wp (\upsilon_{i2})$.
$I_{r}$ can be integrated according to ref.~\cite{Hackmann:2010zz}, and the solution is
\begin{align}
I_{r}=\frac{|b_{3}|}{b_{3}} \bigg\lbrace\sum_{i=1}^{4}
\sum_{j=1}^{2}\dfrac{C_{i}}{\wp ' (\upsilon_{ij})} \big[ \xi
(\upsilon_{ij})(\upsilon-\upsilon_{0})+ \log \sigma
(\upsilon-\upsilon_{ij}) \nonumber\\ - \log
\sigma(\upsilon_{0}-\upsilon_{ij}) \big] -
\dfrac{\tilde{a}E\Xi
(\tilde{a}^{2}+\tilde{r}^{2}_{\tilde{R}})-\tilde{a}^{2}\Xi^{2}\tilde{L}}{\Delta_{\tilde{r}
=\tilde{r}_{\tilde{R}}}}(\upsilon-\upsilon_{0})\bigg\rbrace,
\end{align}
with $\upsilon=\upsilon(\gamma)=\gamma-\gamma_{\tilde{r},in}$ and $
\upsilon_{0}=\upsilon (\gamma_{0})$, where $\gamma_{\tilde{r},in}$ is given
in Eq.~(\ref{rr}).\\
\paragraph{Timelike geodesics}
Considering particles, i.e.~$\varepsilon=1$,
and assuming that the polynomial $\tilde{R}$ has only simple zeros,
$I_{r}$ is a hyperelliptic integral of the third kind.
The first step in the solution procedure is to transform $\tilde{R}$ to the standard form by the substitution $\tilde{r}=\pm 1/u+\tilde{r}_{\tilde{R}}$, where $\tilde{r}_{\tilde{R}}$ is a zero of $\tilde{R}$ (see section \ref{rmotion}). Next we apply a partial fraction decomposition to the integrand, so that the solution method of ref.~\cite{Hackmann:2010zz} can be used. The solution of $I_{r}$ is
\begin{align}\label{akh}
I_{r}=\mp\dfrac{\tilde{a}u_{0}}{\mid u_{0} \mid} \big
\lbrace C_{1}(\omega -\omega_{0})+ C_{0}(f(\omega)-f(\omega_{0}))
\nonumber\\ +\sum_{i=1}^{4}\dfrac{C_{2,i}}{\sqrt{\tilde{R}_{u_{i}}}}
[\frac{1}{2} \log
\frac{\sigma(W^{+}(\omega))}{\sigma(W^{-}(\omega))}-\frac{1}{2}\log
\frac{\sigma(W^{+}(\omega_{0}))}{\sigma(W^{-}(\omega_{0}))}
\nonumber\\ -(f(\omega)-f(\omega_{0}), \omega - \omega_{0}) \big(
\int_{u_{i}^{-}}^{u_{i}^{+}} d\vec{r} \big) ] \big \rbrace .
\end{align}
with $\omega =\omega (\gamma)=\gamma
-\gamma_{\tilde{r},in}$ and $\omega_{0}=\omega (\gamma_{0})$. Again the constants $C_{i}$ arise from the partial fraction decomposition.
$\tilde{R}_{u}$ is defined in Eq.~(\ref{udu}), and the $u_{i}$ are the four zeros of
$\Delta_{\tilde{r}=\pm1/u+\tilde{r}_{\tilde{R}}}, u_{0}=\pm
(\tilde{r}-\tilde{r}_{\tilde{R}})^{-1}$.
The functions $W^{\pm}$ are
given by $W^{\pm}(\omega):=
(f(\omega),\omega)^{t}-2\int_{\infty}^{u_{i}^{\pm}} d\vec{z}$ with $u_{i}^{\pm}=(u_{i}\pm \sqrt{\tilde{R}_{u_{i}}})$ (compare~\cite{Hackmann:2010zz}).
\subsection{t motion}
The $\tilde{t}$-equation (\ref{dtd}) can be rewritten using the
$\tilde{r}$- and $\theta$-equations, (\ref{drd}) and (\ref{dphi})
\begin{align}
d\tilde{t}=\dfrac{E(\tilde{r}^{2}+\tilde{a}^{2})^{2}-\tilde{a}\tilde{L}\Xi(\tilde{r}^{2}
+\tilde{a}^{2})}
{\Delta_{\tilde{r}}}\frac{dr}{\sqrt{\tilde{R}}}
-\frac{\sin^{2}\theta}{\Delta_{\theta}}(E
\tilde{a}^{2}-\frac{\tilde{L}\Xi
\tilde{a}}{\sin^{2}\theta})\dfrac{d\theta}{\sqrt{\tilde{\Theta}(\theta)}},
\end{align}
and has the same structure as the equation for the $\varphi$ motion. Integrating the $\tilde{t}$-equation yields
\begin{align}
\tilde{t}-\tilde{t}_{0}=\big[ \int_{r_{0}}^{r}\dfrac{E(\tilde{r}^{2}
+\tilde{a}^{2})^{2}-\tilde{a}
\tilde{L}\Xi(\tilde{r}^{2}+\tilde{a}^{2})}{\Delta_{\tilde{r}}} \frac{dr}{\sqrt{\tilde{R}}}
- \int_{\theta_{0}}^{\theta} \frac{\sin^{2}\theta}{\Delta_{\theta}}(E
\tilde{a}^{2}-\frac{\tilde{L}\Xi
\tilde{a}}{\sin^{2}\theta})\dfrac{d\theta}{\sqrt{\tilde{\Theta}(\theta)}}
\big] = \tilde{I}_{r}-\tilde{I}_{\theta}.
\end{align}
The solutions can be found in the same way as in section \ref{fii}. For the $\theta$-dependent part we have \cite{Hackmann:2010zz}
\begin{align}
\tilde{I}_{\theta}= a_{3}(\upsilon -\upsilon_{0}) -
\sum_{i=1}^{2}\dfrac{a_{3}\tilde{a}^{2}R_{0}}{4\wp' (\upsilon_{i})}
\bigg[\zeta(\upsilon_{i})(\upsilon - \upsilon_{0}) +\log \sigma
(\upsilon - \upsilon_{i})-\log \sigma (\upsilon_{0}-\upsilon_{i}) \bigg],
\end{align}
where $a_{2}$ and $a_{3}$ are given in
Eq.~(\ref{khiii}), $\wp(\upsilon_{1})=\frac{a_{2}}{12}+\frac{1}{4}\tilde{a}^{2}R_{0}a_{3}=\wp(\upsilon_{2})$,
and $\upsilon=\upsilon(\gamma)=2\gamma -\gamma_{\theta,in}$ with the initial value
$\upsilon_{0}=\upsilon(\gamma_{0})$.
Considering light, i.e.~$\varepsilon=0$,
the solution for the $\tilde{r}$-dependent part is very simple and
given by \cite{Hackmann:2010zz}
\begin{align}
\tilde{I}_{r}=\frac{|b_{3}|}{b_{3}} \bigg\lbrace\sum_{i=1}^{4}
\sum_{j=1}^{2}\dfrac{\tilde{C}_{i}}{\wp ' (\upsilon_{ij})} \big[ \xi
(\upsilon_{ij})(\upsilon-\upsilon_{0})+ \log \sigma
(\upsilon-\upsilon_{ij}) \nonumber\\ - \log
\sigma(\upsilon_{0}-\upsilon_{ij}) \big] -
\dfrac{\tilde{a}\tilde{L}\Xi (\tilde{r}_{\tilde{R}}^{2}
+\tilde{a}^{2})-E(\tilde{r}_{\tilde{R}}^{2}+\tilde{a}^{2})^{2}}{\Delta_{\tilde{r}=\tilde{r}_{\tilde{R}}}}
(\upsilon-\upsilon_{0})\bigg\rbrace ,
\end{align}
where $b_{3}$ is given in Eq.~(\ref{rr}), the $\tilde{C}_{i}$ arise from the partial fraction decomposition, and
$\wp(\upsilon_{i1})=y_{i}=\wp(\upsilon_{i2})$, where
$y_{i}$ are the four zeros of $\Delta_{y(\tilde{r})}$. The variable $\upsilon=\upsilon(\gamma)=\gamma -\gamma_{\tilde{r},in}$ has the initial value
$\upsilon_{0}=\upsilon(\gamma_{0})$.\\
In the case of timelike geodesics, i.e.~$\varepsilon=1$, the solution of the now hyperelliptic $\tilde{r}$-dependent part is
given by \cite{Hackmann:2010zz}
\begin{align}
\tilde{I}_{r}=\dfrac{u_{0}}{\sqrt{c_{5}} \mid u_{0} \mid} \big
\lbrace \tilde{C}_{1}(\omega -\omega_{0})+
\tilde{C}_{0}(f(\omega)-f(\omega_{0})) \nonumber\\
+\sum_{i=1}^{4}\dfrac{\tilde{C}_{2,i}}{\sqrt{\tilde{\tilde{R}}_{u_{i}}}}
[\frac{1}{2} \log
\frac{\sigma(W^{+}(\omega))}{\sigma(W^{-}(\omega))}-\frac{1}{2}\log
\frac{\sigma(W^{+}(\omega_{0}))}{\sigma(W^{-}(\omega_{0}))}
\nonumber\\ -(f(\omega)-f(\omega_{0}), \omega - \omega_{0}) \big(
\int_{u_{i}^{-}}^{u_{i}^{+}} d\vec{r} \big) ] \big \rbrace .
\end{align}
For the notation see Eq.~(\ref{akh}). The constants
$\tilde{C}_{0},\tilde{C}_{1},\tilde{C}_{2,i}$ result from the partial fraction decomposition.
\section{THE ORBITS}\label{sec:orbits}
The analytical solutions can be used to plot the orbits of test particles and light rays. We present example of the orbits around the static charged (A)dS black hole (Reissner-Nordstr\"om-(A)dS) and the rotating charged (A)dS black hole (Kerr-Newman-(A)dS).
\subsection{The static case}
Some examples of timelike and null geodesics in the static case can be found in Figs.~\ref{pic:orbits1} and \ref{pic:orbits2}.
In Fig.~\ref{pic:orbits1}, two bound orbits of test particles are shown: a bound orbit outside the horizons (Fig.~\ref{pic:orbits1}(a)) and a many-world bound orbit (Fig.~\ref{pic:orbits1}(b)). On the many-world bound orbit both horizons are crossed several times and each time the test particles emerge into
another universe. Note that the test particle is reflected at the potential barrier behind the horizons arising from the charge.
An escape orbit and a two-world escape orbit are depicted in Figs.~\ref{pic:orbits2}(a) and~\ref{pic:orbits2}(b) respectively.
The two-world escape orbit crosses both horizons twice and escapes to another universe. Also the reflection at the potential barrier is visible.
\begin{figure}[h!]
\centering
\subfigure[Bound orbit with parameters $\varepsilon=1$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.75$, $\mathcal{L}=0.076$, $E=\sqrt{0.918}$.]{
\includegraphics[width=0.45\textwidth]{static-bo-particles.eps}
}
\subfigure[Many-world bound orbit with parameters $\varepsilon=1$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.995$, $\mathcal{L}=0.8$, $E=\sqrt{0.2}$. The particle is reflected at the potential barrier arising from the charge.]{
\includegraphics[width=0.45\textwidth]{static-mbo-particles.eps}
}
\caption{Two examples of particle orbits in the Reissner-Nordstr\"om-(A)dS spacetime.
The blue curves depict the orbits and the black dashed circle indicte the positions of the horizons.}
\label{pic:orbits1}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[Escape orbit with parameters $\varepsilon=0$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.75$, $\mathcal{L}=0.1$, $E=\sqrt{0.46}$.]{
\includegraphics[width=0.45\textwidth]{static-eo-light.eps}
}
\subfigure[Two-world escape orbit with parameters $\varepsilon=0$, $\tilde{R}_0=\frac{1}{3}\cdot 10^{-5}$, $\tilde{q}=0.95$, $\mathcal{L}=5$, $E=\sqrt{0.8}$. Light is reflected at the potential barrier arising from the charge.]{
\includegraphics[width=0.45\textwidth]{static-teo-light2.eps}
}
\caption{Two examples of light orbits in the Reissner-Nordstr\"om-(A)dS spacetime.
The blue curves depict the orbits and the black dashed circle indicte the positions of the horizons.}
\label{pic:orbits2}
\end{figure}
\clearpage
\subsection{The rotating case}
Here we show some orbits in the Kerr-Newman-(A)dS spacetime. Figure \ref{pic:bo-eo}, shows two example plots of a bound orbit for particles and an escape orbit for light. A transit orbit crossing $r=0$ can be seen in Figure \ref{pic:tro-teo}(a). A two-world escape orbit which crosses both horizons twice and escapes to another universe is depicted in \ref{pic:tro-teo}(b). In Figure \ref{pic:innerbo-mbo}(a), a bound orbit hidden behind the inner horizon is shown. Figure \ref{pic:innerbo-mbo}(b), shows a many-world bound orbit, where both horizons are crossed several times.
\begin{figure}[h!]
\centering
\subfigure[Bound orbit with parameters $\varepsilon=1$, $a=0.7$, $K=2$, $q=0.7$, $R_0=4\cdot 10^{-5}$, $L=1.9$, $E=0.84$.]{
\includegraphics[width=0.45\textwidth]{bo-particles.eps}
}
\subfigure[Escape orbit with parameters $\varepsilon=0$, $a=0.7$, $K=2$, $q=0.7$, $R_0=4\cdot 10^{-5}$, $L=1.9$, $E=0.75$.]{
\includegraphics[width=0.45\textwidth]{eo-light.eps}
}
\caption{Two examples of possible orbits in the Kerr-Newman-(A)dS spacetime. The blue lines show the path of the orbits and the sphere represents the event horizon.}
\label{pic:bo-eo}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[Transit orbit with parameters $\varepsilon=0$, $a=0.7$, $K=2$, $q=0.5$, $R_0=4\cdot 10^{-5}$, $L=0.5$, $E=3.75$.]{
\includegraphics[width=0.45\textwidth]{tro-light.eps}
}
\subfigure[Two-world escape orbit with parameters $\varepsilon=0$, $a=0.7$, $K=1$, $q=0.7$, $R_0=4\cdot 10^{-5}$, $L=0.5$, $E=1.5$.]{
\includegraphics[width=0.45\textwidth]{teo-light.eps}
}
\caption{Two examples of possible orbits in the Kerr-Newman-(A)dS spacetime. The blue lines show the path of the orbits and the spheres represent the inner and outer horizon.}
\label{pic:tro-teo}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[Bound orbit behind the inner horizon with parameters $\varepsilon=1$, $a=0.9$, $K=0.3$, $q=0.2$, $R_0=4\cdot 10^{-5}$, $L=1.45$, $E=1.02$.]{
\includegraphics[width=0.45\textwidth]{innerbo-particles.eps}
}
\subfigure[Many-world bound orbit with parameters $\varepsilon=0$, $a=0.7$, $K=1$, $q=0.7$, $R_0=4\cdot 10^{-5}$, $L=0.5$, $E=0.05$.]{
\includegraphics[width=0.45\textwidth]{mbo-light.eps}
}
\caption{Two examples of possible orbits in the Kerr-Newman-(A)dS spacetime. The blue lines show the path of the orbits and the spheres represent the inner and outer horizon.}
\label{pic:innerbo-mbo}
\end{figure}
\section{conclusions}\label{conclusions}
In this paper we discussed the motion of test particles and light
rays in the spacetime of the static and rotating charged black hole (Kerr-Newman-(A)dS spacetime).
After reviewing the spacetime and presenting the corresponding
equations of motion, we classified the possible types of geodesic motion
by an analysis of the zeros of the polynomials underlying the
$\theta$- and $r$-motion. The geodesic equations were solved in
terms of Weierstrass elliptic functions and derivatives of Kleinian
$\sigma$ functions.
Using effective potential techniques and parametric diagrams, the
possible types of orbits were derived.
Finally, a number of orbits were illustrated.
The techniques employed in this
paper, present a useful tool to calculate the exact orbits,
and the results obtained should prove valuable in order to analyze their properties,
including observables like the periastron shift of bound
orbits, the light deflection of flyby orbits, the deflection angle
and the Lense-Thirring effect. For the calculation of these
observables analogous formulas to those given in~\cite{Hackmann:2008zza,Rindler:2007zz,
Bhattacharya:2009rv,Bhattacharya:2010xh,Kraniotis:2003ig,Kraniotis:2004cz} may be used.
The analytical solutions of the equations of motion are also useful in the context of AdS/CFT, since geodesics in an AdS spacetime can be related to CFT propagators (see e.g. \cite{Balasubramanian:1999zv}).
\begin{acknowledgements}
S.G.~and J.K.~would like to acknowledge support by the DFG
Research Training Group \emph{Models of Gravity}.
\end{acknowledgements}
\bibliographystyle{amsplain}
|
\subsection*{\hbox{}\hfill{\normalsize\sl #1}\hfill\hbox{}}}
\textheight 23truecm
\textwidth 15truecm
\addtolength{\oddsidemargin}{-1truecm}
\addtolength{\topmargin}{-1truecm}
\makeatletter \def\l@section{\@dottedtocline{1}{0em}{1.2em}} \makeatother
\begin{document}
\centerline{\Large\bf Harder-Narasimhan stacks}
\centerline{\Large\bf for principal bundles in higher dimensions}
\bigskip
\centerline{\bf Sudarshan Gurjar and Nitin Nitsure}
\begin{abstract}
Let $G$ be a connected split reductive group over a
field $k$ of arbitrary characteristic, chosen suitably.
Let $X\to S$
be a smooth projective morphism of locally noetherian $k$-schemes,
with geometrically connected fibers.
We show that for each Harder-Narasimhan type $\tau$ for principal
$G$-bundles,
all pairs consisting of a principal $G$-bundle on a fiber of $X\to S$ together
with a given canonical reduction of HN-type $\tau$ form an
Artin algebraic stack $Bun_{X/S}^{\tau}(G)$ over $S$. Moreover,
the forgetful $1$-morphism
$Bun_{X/S}^{\tau}(G) \to Bun_{X/S}(G)$ to the stack of
all principal $G$-bundles on fibers of $X\to S$ is
a schematic morphism, which is of finite type, separated and injective
on points.
The notion of a relative canonical reduction that we use was defined earlier
in arXiv:1505.02236, where we showed that a
stronger result holds in characteristic zero, namely, the $1$-morphisms
$Bun_{X/S}^{\tau}(G) \to Bun_{X/S}(G)$ are locally closed imbeddings
which stratify $Bun_{X/S}(G)$ as $\tau$ varies.
\end{abstract}
\bigskip
\section{Introduction}
Let $G$ be a split reductive group over a field $k$, such that
the following hypothesis $(*)$ is satisfied
(we will say more about this hypothesis at the end of the Introduction).
{\it $(*)$ Preservation of canonical reductions under field extensions}:
If $L/K/k$ are extension fields of $k$, if
$H = P/R_u(P)$ where $P$ is a standard parabolic in $G$
and if $E$ a semistable principal $H$-bundle
on a geometrically irreducible smooth projective curve $X$ over $K$, then
the base change $E_L$ is a semistable
principal $H$-bundle on $X_L$.
Let $S$ be a locally noetherian scheme over $k$ and let $X\to S$ be a
smooth projective morphism with geometrically connected fibers.
Let there be chosen a split maximal torus in $G$ and a Borel containing it,
and let $\tau$ be an element of the resulting closed positive Weyl chamber.
For any $S$-scheme $T$ and a principal $G$-bundle
$E$ on $X_T$, we defined in [G-N-2] the notion of a
relative canonical reduction $[L,\phi]$
of $E$ of Harder-Narasimhan type $\tau$ (which is recalled later).
We define an $S$-groupoid
$Bun_{X/S}^{\tau}(G)$ which attaches to $T$ the groupoid
whose objects are pairs consisting of a principal $G$-bundle $E$ on
$X_T$ and a relative canonical reduction
of $E$ of HN-type $\tau$. We denote by $Bun_{X/S}(G)$
the algebraic stack of all $G$-bundles on $X/S$.
The main result of this note is the following.
\begin{theorem}\label{stack version of main theorem}
The $S$-groupoid $Bun_{X/S}^{\tau}(G)$ is an
algebraic stack over $S$. The natural forgetful
$1$-morphism $Bun_{X/S}^{\tau}(G) \to Bun_{X/S}(G)$
is a schematic morphism, which is of finite type, separated and injective
on points.
\end{theorem}
The above theorem can be equivalently re-formulated as follows.
\begin{theorem}\label{family version of main theorem}
Let $E$ be a principal $G$-bundle on $X$.
There exists a scheme $S^{\tau}(E)$ over $S$
which has the universal property that for any $S$-scheme $T$, the set
of all relative canonical reductions of type $\tau$ of the pullback $E_T/X_T/T$
is in a natural bijection with the set of all $S$-morphisms from
$T$ to $S^{\tau}(E)$. Moreover, the
morphism $S^{\tau}(E) \to S$ is of finite type, separated and
injective.
\end{theorem}
If $k$ is of characteristic zero, it can be shown
(see [G-N-2] Theorem 1.1 and Proposition 7.4)
that each $S^{\tau}(E)$ is a locally closed subscheme of $S$,
and as $\tau$ varies over the closed
positive Weyl chamber $\ov{C}$, these subschemes stratify $S$.
Correspondingly in the Theorem \ref{stack version of main theorem}
above, $Bun_{X/S}^{\tau}(G) \to Bun_{X/S}(G)$ is a locally closed substack,
and these stratify $Bun_{X/S}(G)$ as $\tau$ varies
(see [G-N-2] Theorem 7.7).
The stronger results in characteristic zero are
made possible by the uniqueness and the infinitesimal uniqueness
of a canonical reduction in characteristic zero.
It is known (see [He])
that the property of infinitesimal uniqueness does
not necessarily hold in the finite characteristic case (failure of the
Behrend conjecture [Be], which in the context of principal bundles
says that the canonical reduction of a principal
$G$-bundle over a curve has no infinitesimal deformations).
Instead, one has the (weaker) results of this note.
\pagestyle{myheadings}
\markright{Gurjar and Nitsure: Harder-Narasimhan
stacks in positive characteristic.}
The importance of the hypothesis ${\bf (*)}$ is that it allows the
definition of a moduli functor for $G$-bundles of a given HN-type
on curves (and also on higher dimensional projective varieties).
In fact (see [He]), $(*)$ is a consequence of the Behrend conjecture.
It is known that if $G = GL_{n,k}$ or $SL_{n,k}$,
then the Behrend conjecture is satisfied for all $k$, and
if $char(k) =0$, then it is satisfied for all $G$.
Moreover, the conjecture always holds for
classical groups, and holds for exceptional simple groups
whenever $char(k)$ is large enough
(see Theorem 1 in [He]).
{\it Question}: More generally,
if $S$ is a quasi-finite, flat scheme over $\mathop{\rm Spec}\nolimits \mathbb Z$
and if $G$ is obtained by base change from a reductive group
scheme ${\mathfrak G}$ defined over $S$,
then one may ask whether
there exists a nonempty open subscheme $S'\subset S$
such that the Behrend conjecture (or at least the hypothesis ${\bf (*)}$)
holds whenever $\mathop{\rm Spec}\nolimits k$ factors via $S'$.
\section{Preliminaries}
Let $G$
be a reductive group over a field $k$ of arbitrary characteristic,
such that $G$ is split over $k$, together with a chosen split
maximal torus and a Borel containing it.
Let $K$ be an extension field over $k$, and
$X$ a smooth irreducible projective variety over $K$ with
a very ample line bundle ${\mathcal O}_X(1)$. Let $E$ be a principal $G$-bundle defined
on $X$ (or defined on a big open subscheme $U$ of $X$,
where bigness of $U$ signifies that $X-U$ is of codimension
$\ge 2$).
Recall that $E$ is said to be {\it semistable} w.r.t. the choice
of ${\mathcal O}_X(1)$ if for any standard
parabolic $P\subset G$, any section $\sigma: W\to E/P$ defined on
a big open subscheme $W$ of $U$, and any dominant character
$\chi : P \to {\mathbb G}_{m,K}$, we have
$$\deg(\chi_*\sigma^*E) \le 0$$
where $\sigma^*E$ is the principal $P$-bundle on $W$
defined by the reduction $\sigma$,
and $\chi_*\sigma^*E$ is the ${\mathbb G}_m$-bundle obtained by
extending its structure group via $\chi: P \to {\mathbb G}_m$,
which is equivalent to a line bundle on $W$. This line
bundle extends uniquely (up to a unique isomorphism) to a line bundle
on $X$, denoted again by $\chi_*\sigma^*E$, and
$\deg(\chi_*\sigma^*F)$ is its degree w.r.t. ${\mathcal O}_X(1)$.
A {\it rational reduction} of the structure group to a standard parabolic
$P$ is a section $\sigma: U \to E/P$
of $E/P \to X$ over a big open subscheme $U\subset X$.
Recall that a {\it canonical reduction} of $E$ is
a rational reduction
of structure group of $E$ to a standard parabolic $P\subset G$
for which
the following two conditions hold:
(1) If $\rho: P \to L = P/R_u(P)$ is the Levi quotient of $P$
(where $R_u(P)$ is the unipotent radical of $P$) then the principal $L$-bundle
$\rho_*\sigma^*E$ is a semistable principal $L$-bundle defined
on the big open subscheme $U$ on which $\sigma$ is defined.
(2) For any non-trivial character $\chi: P \to {\mathbb G}_m$
whose restriction to the chosen maximal torus $T\subset B \subset P$
is a linear combination $\sum n_i\alpha_i$
of simple roots $\alpha_i \in \Delta$ where $n_i \ge 0$, and at least
one $n_i \ne 0$, we have $\deg(\chi_*\sigma^*E) > 0$.
To any such reduction, one associates a Harder-Narasimhan type
$\tau \in \ov{C}$ (see [G-N-2], section 4 for an exposition).
We recall the following well known fact
(originally proved by Behrend [Be] for curves).
\begin{proposition}\label{uniqueness of reduction}
Let $G$ be a reductive group over a field $k$ of arbitrary characteristic,
such that $G$ is split over $k$, together with a chosen a split
maximal torus and a Borel containing it.
Let $K$ be an extension field over $k$, and
$X$ a smooth irreducible projective variety over $K$ with
a very ample line bundle. Let $E$ be a principal $G$-bundle on
$X$. Then $E$ admits a unique canonical reduction.
\end{proposition}
Let $T\subset B\subset G$ be the chosen split torus and Borel.
Given a standard parabolic $P\supset B$, let $\lambda_P \in X^*(T)$
be a chosen dominant weight such that $\lambda_P$ is a character on $P$
which lies in the negative
ample cone for $G/P$.
Let $V_{\lambda_P}$ be a chosen irreducible representation of $G$ with
highest weight $\lambda_P$, and let $0\ne v\in V_{\lambda_P}$
be a chosen highest weight vector. Then for the action of $G$
on the projective space ${\bf P}(V_{\lambda_P})$ of lines in
$V_{\lambda_P}$, the isotropy
subgroup scheme at the point $[v] \in {\bf P}( V_{\lambda_P})$ is $P$,
and we get a closed $G$-equivariant
embedding $G/P \hra {\bf P}(V_{\lambda_P})$ under which $eP\mapsto [v]$.
With the above notation, we recall the
definition of a relative canonical reduction made in
[G-N-2]. Let $X\to S$ be a
smooth projective morphism with geometrically connected fibers,
where $S$ is a noetherian scheme over $k$, with
a given relatively very ample line bundle ${\mathcal O}_{X/S}(1)$ on $X$.
For any principal $G$-bundle $E$ on $X$,
let $E(V_{\lambda_P})$ denote
the associated vector bundle on $X$ corresponding to
the representation $V_{\lambda_P}$.
As defined in [G-N-2], a
{\it relative rational reduction} of structure group
of $E$ from $G$ to $P$ is an equivalence class $[L,f]$ of pairs $(L,f)$,
where $L$ is a line bundle on $X$
and $f: L\to E(V_{\lambda_P})$ is an injective ${\mathcal O}_X$-linear
homomorphism of sheaves,
such that \\
(i) the open subscheme $U = \{ x\in X\,|\,\mathop{\rm rank}\nolimits(f_x) =1\} \subset X$
is relatively big over $S$, that is, for each $s\in S$
the fiber $U_s$ has complementary codimension
$\ge 2$ in the fiber $X_s$, and \\
(ii) the section $U\to {\bf P}(E(V_{\lambda_P}))$
defined by $f$ factors via the natural closed embedding
$E/P \hra {\bf P}(E(V_{\lambda_P}))$.\\
Two such pairs $(L,f)$ and $(L',f')$ are {\it equivalent}
if there exists an isomorphism
$\phi :L \to L'$ such that $f = f'\circ \phi$.
In the special case where $S = \mathop{\rm Spec}\nolimits K$ for a field $K$,
the above definition is equivalent to the usual definition
of a rational reduction to $P$ that we recalled earlier
([G-N-2] Proposition 3.2). Finally, we say that a pair
$(L,f)$ as above defines a
{\it canonical reduction $[L,f]$ of type $\tau$} if its restriction to each
fiber $X_s$ of $X\to S$ is a canonical reduction of $E_s$
of constant type $\tau$ (this is well-defined).
The following remark shows that
if $S$ has a Zariski open cover $(W_i)$ and we have relative rational
$P$-reductions $[L_i,f_i]$
of $E_{W_i}/X_{W_i}/W_i$ which are represented by pairs $(L_i,f_i)$
which are equivalent over each $W_i\cap W_j$,
then there exists a unique relative rational reduction $[L,f]$ of $E/X/S$
which restricts to these.
\refstepcounter{theorem}\paragraph{Remark \thetheorem} (Sheaf property.) Let $Y$ be a scheme, ${\mathcal E}$ be a sheaf of ${\mathcal O}_Y$-modules,
and $(U_i)$ be an open cover of $Y$. Let for each $i$ there be given a line
bundle $L_i$ on $U_i$ together with an injective ${\mathcal O}_{U_i}$-linear
homomorphism of sheaves $f_i: L_i \to {\mathcal E}|_{U_i}$. Suppose that
for each $U_{ij} = U_i\cap U_j$, there exists an element
$g_{ij} \in {\mathbb G}_m(U_{ij})$ (that is, a nowhere vanishing regular
function on $U_{ij}$) such that $f_i = g_{ij}\cdot f_j$
(we do not assume any cocycle condition on the $g_{ij}$'s).
Then there exists a line bundle $L$ on $Y$ and an injective
${\mathcal O}_Y$-linear homomorphism $f: L \to {\mathcal E}$, such that
for any $i$, there exists an isomorphism $h_i : L|_{U_i} \to L_i$
with $f|_{U_i} = f_i\circ h_i$.
Moreover, if $(L',f')$ is another such pair,
then there exists a unique ${\mathcal O}_Y$-linear isomorphism $\phi: L'\to L$
such that $f' = f\circ \phi$.
For, the image subsheaves $\im(f_i)
\subset {\mathcal E}|_{U_i}$ coincide over $U_i\cap U_j$, so they
glue together to define a global subsheaf $L\subset {\mathcal E}$.
Take $f: L\hra {\mathcal E}$ to be the inclusion. Then the pair $(L,f)$ has the
desired property. Given any other such
$(L',f')$, the image of the homomorphism $f'$ is the subsheaf
$L\subset {\mathcal E}$, so $f'$ factors through
$L$ to give rise to a homomorphism $\phi: L'\to L$ with the desired property.
\section{Proofs}
If $T$ is a scheme and ${\mathcal E}$ is a sheaf of ${\mathcal O}_T$-modules, we will denote by
$ \Gamma(T, {\mathcal E})^{\times} \subset \Gamma(T, {\mathcal E})$ the subset which consists
of all nowhere vanishing global sections of ${\mathcal E}$. In particular,
$\Gamma(T, {\mathcal O}_T)^{\times} = {\mathbb G}_m(T)$ is the group of all
invertible regular functions on $T$. Note that the group $ {\mathbb G}_m(T)$
acts on the set $ \Gamma(T, {\mathcal E})^{\times}$ by scalar multiplication.
The following lemma is a projective
version of the result of Grothendieck on the
representability by a linear scheme for sections of direct images
(see [EGA III 7.7.8, 7.7.9], and [Ni-2, 5.8] for an exposition).
We expect this lemma, though elementary, to be of independent interest.
\begin{lemma}\label{projective representation}
Let $X\to S$ be a proper morphism of noetherian schemes,
and let ${\mathcal F}$ be a coherent ${\mathcal O}_X$-module that is flat over $S$.
Consider the contravariant functor
$\Psi' :(Schemes/S)^{op} \to Sets$ which associates to any $S$-scheme $T$
the quotient set
$$%
\Psi'(T) =
\frac{\Gamma(T,{\pi_T}_*{\mathcal F}_T)^{\times}}
{{\mathbb G}_m(T)}$$
where $\Gamma(T,{\pi_T}_*{\mathcal F}_T)^{\times}$ is the set of all nowhere
vanishing sections of ${\pi_T}_*{\mathcal F}_T$, on which ${\mathbb G}_m(T)$
acts by scalar multiplication.
Then the sheafification $\Psi$ of $\Psi'$
in the big Zariski site over $S$
is representable by the $S$-scheme
$${\mathbb P}({\mathcal Q} ) = Proj_S \, Sym_S ^{\bullet}({\mathcal Q}),$$
where ${\mathcal Q}$ denotes the Grothendieck $Q$-sheaf of ${\mathcal F}/X/S$ (locally over $S$,
we can take ${\mathcal Q}$ to be the cokernel of the transpose of the $0$th
differential of a Grothendieck semicontinuity complex for ${\mathcal F}/X/S$).
\end{lemma}
{\it Proof.} If ${\mathcal E}$ is any coherent sheaf $S$,
then ${\mathbb P}({\mathcal E} ) = Proj_{S}Sym_S^{\bullet}({\mathcal E})$ represents the functor
$\varphi : (Schemes/S)^{op} \to Sets$ which is the
sheafification in the big Zariski site over $S$ of
the functor that associates to any $S$-scheme $T$ the
quotient set
$$\frac{Hom({\mathcal E}_T, {\mathcal O}_T)^{sur}}{{\mathbb G}_m(T)}$$
where $Hom({\mathcal E}_T, {\mathcal O}_T)^{sur} \subset Hom({\mathcal E}_T, {\mathcal O}_T)$ consists
of all surjective homomorphisms ${\mathcal E}_T \to {\mathcal O}_T$, and
${\mathbb G}_m(T)$ acts on it by scalar multiplication (which is the
restriction of the action of ${\mathbb G}_m(T)$ on $Hom({\mathcal E}_T, {\mathcal O}_T)$).
The Grothendieck sheaf ${\mathcal Q}$ has the universal property
(see [EGA III 7.7.8, 7.7.9]) that we have a natural bijection
$$\Gamma(T,{\pi_T}_*{\mathcal F}_T) \cong Hom({\mathcal Q}_T,{\mathcal O}_T).$$
This bijection is equivariant under the action of ${\mathbb G}_m(T)$ on both sides by
scalar multiplication. As the bijection is functorial in $T$,
by pull-back to any point of $T$ it
follows that nowhere vanishing sections of ${\pi_T}_*{\mathcal F}_T$ exactly
correspond to
surjective homomorphisms ${\mathcal Q}_T \to {\mathcal O}_T$, and hence we get a
natural ${\mathbb G}_m(T)$-equivariant bijection
$$\Gamma(T,{\pi_T}_*{\mathcal F}_T)^{\times} \cong Hom({\mathcal Q}_T,{\mathcal O}_T)^{sur}.$$
The lemma follows on passage to the quotient sets under ${\mathbb G}_m(T)$.
\hfill$\square$
\refstepcounter{theorem}\paragraph{Remark \thetheorem} Unlike the functor $T \mapsto \Gamma(T,{\pi_T}_*{\mathcal F}_T)$,
which is representable in the projective case
if and only if ${\mathcal F}$ is flat over $S$ (see [Ni-1]),
the functor $\Psi$ may or may not be representable when ${\mathcal F}$
is not flat. For example, let $X = S = \mathop{\rm Spec}\nolimits k[t]$ for a field $k$, and let
${\mathcal F}_1 = (k[t]/(t))^{\sim}$ and ${\mathcal F}_2 = {\mathcal F}_1 \oplus {\mathcal O}_X$,
which are coherent sheaves on $X =S$ which are not flat over $S$.
We leave it to the reader to verify that
the corresponding functor $\Psi$ is representable for ${\mathcal F}_1$,
and it is not representable for ${\mathcal F}_2$.
{\iffalse
{\footnotesize Solution: We can assume $k= \ov{k}$.
For ${\mathcal F}_1$, the scheme $\mathop{\rm Spec}\nolimits k[t]/(t)
\hra \mathop{\rm Spec}\nolimits k[t]$ represents $\Psi$. For ${\mathcal F}_2$, suppose $\Psi$ is
represented by a scheme $p: R\to \mathop{\rm Spec}\nolimits k[t]$. The fiber of
$R$ over the origin
$x_0 = \mathop{\rm Spec}\nolimits k[t]/(t)$ must be the projective line $p^{-1}(x_0) = P^1_k$,
while over $U = \mathop{\rm Spec}\nolimits k[t,t^{-1}]$ the restriction $R_u \to U$ is
an isomorphism. Let $V \subset R$ be an affine open
subscheme with $V \cap p^{-1}(x_0) \ne \emptyset$. Then as
$k= \ov{k}$, we will have at least two distinct $k$-valued points $y_1$
and $y_2$ in $V \cap p^{-1}(x_0)$. From the definition of
${\mathcal F}_2$, it can be seen that there will exist global sections
$s_1,s_2 \in \Gamma(S,R)$ with $s_i(x_0) = y_i$. But then
$s_1|_U = s_2|_U$, which contradicts the separatedness of
$V \to S$.}
\fi}
\bigskip
{\bf Proof of Theorem \ref{family version of main theorem}.}
An exposition of the basic facts about
the relative Picard scheme $Pic_{X/S}$ that we need can be found in [K].
We first treat the case where $\pi: X\to S$ admits
a global section $\sigma : S\to X$. Under this assumption,
there exists a Poincar\'e line bundle
on $X\times_S Pic_{X/S}$, fixed up to (a non-unique)
isomorphism by the requirement that its pullback to $Pic_{X/S}$
under the section
$(\sigma, \id_{Pic_{X/S}}): Pic_{X/S} \to X\times_S Pic_{X/S}$
is a trivial line bundle.
The choice of $\tau \in \ov{C}$ determines the Hilbert polynomial
$h \in \mathbb Q[t]$ of any line bundle
$L$ which occurs in a canonical reduction $[L, f: L\to E_s(V_{\lambda_P})]$
of type $\tau$ of the principal $G$-bundle $E_s = E|_{X_s}$ on a fiber $X_s$.
Let $J = Pic_{X/S}^h \subset Pic_{X/S}$ be the open and closed subscheme where
the Hilbert polynomial of the line bundle is $h$. Let $\LL$
denote the restriction of the Poincar\'e line bundle
to $X\times_S J\subset X\times_S Pic_{X/S}$.
Let ${\mathcal F}$ be the coherent ${\mathcal O}$-module on $X\times_S J$
defined by
$${\mathcal F} = \un{Hom}(\LL, E_J(V_{\lambda_P}))$$
and let ${\mathcal Q}$ denote the coherent ${\mathcal O}_J$-module
which is the Grothendieck $Q$-sheaf for ${\mathcal F}$. Let
$Y = {\mathbb P}({\mathcal Q})$ be the corresponding projective scheme over $J$,
which has the universal property given by
Lemma \ref{projective representation}.
Over $Y$, we have a universal element
$f \in \Psi(Y)$ in the notation of Lemma \ref{projective representation},
which can be represented by a Zariski open
cover $(U_i)$ of $Y$ together with a family of homomorphism
$f_i : \LL_{U_i} \to E_{U_i}(V_{\lambda_P})$.
Note that on $U_i\cap U_j$, the homomorphisms $f_i$ and $f_j$
differ by scalar multiplication by an element of ${\mathbb G}_m( U_i\cap U_j)$.
Let $Y_1 \subset Y$ be the union of the open subschemes of $U_i$ where
$f_i : \LL_{U_i} \to E_{U_i}(V_{\lambda_P})$ is fiberwise injective
in a relatively large open subscheme of $X_{U_i}$.
Let $Y_2\subset Y$ be the closed subscheme which is the union of
the closed subschemes of $U_i$ where the homomorphism
$f_i : \LL_{U_i} \to E_{U_i}(V_{\lambda_P})$ factors via the cone
$\wh{E/P}\subset E_{U_i}(V_{\lambda_P})$ over $E/P \subset
{\bf P}(E_{U_i}(V_{\lambda_P}))$.
Let $Y_3 = Y_1\cap Y_2 \subset Y$, which is the locally closed
subscheme, where the $f_i : \LL_{U_i} \to E_{U_i}(V_{\lambda_P})$ define
a rational reduction of structure group to $P\subset G$.
Let $Y_4\subset Y_3$ be the open and closed subscheme of $Y_3$ where the
topological type of the reduction is given by $\tau$. Finally, let
$Y_5 \subset Y_4$ be the open subscheme of $Y_4$ where the extension under
the Levi quotient $P \to P/R_u(P)$ is semistable.
By Lemma \ref{projective representation},
is immediate from its construction that the $S$-scheme $Y_5$
represents the functor $T\mapsto \Phi_{E/X/S}^{\tau}(T)$ which is the set of all
relative canonical reductions of type $\tau$ of the pullback $E_T/X_T/T$.
Now we come to a general case, where $X$ may not necessarily admit a
global section over $S$. As $X\to S$ is by assumption smooth,
there exists a surjective separated \'etale morphism $p : S'\to S$
such that the base change $X'= X_{S'}$ admits a global section
$S' \to X'$. Let $E' = E_{S'}$.
Hence by the above special case, there exists a scheme
$Y'\to S'$ which represents the functor
$\Phi_{E'/X'/S'}^{\tau} : (Schemes/S')^{opp} \to Sets$.
Let $S'' = S'\times_S S'$, and let
$p_1,p_2 : S'' \stackrel{\to}{\scriptstyle\to} S'$ be the
two projections. We write $\pi = p\circ p_1 = p\circ p_2$,
and $X'' = \pi^*X$ and $E'' = (\id \times \pi)^*E$,
so that we have natural identifications
$$p_1^*X' = X'' = p_2^*X'$$
and
$$(\id \times p_1)^* E' = E'' = (\id \times p_2)^*E'.$$
Note that $p_1^* Y'$ and $p_2^*Y'$ respectively represent
the two functors $\Phi_{(\id \times p_1)^* E'/X''/S''}^{\tau}$
and $\Phi_{(\id \times p_2)^* E'/X''/S''}^{\tau}$, and both
these functors have a natural isomorphism with the functor
$\Phi_{E''/X''/S''}^{\tau}$ which comes from the above
identifications $(\id \times p_1)^* E' = E'' = (\id \times p_2)^*E'$.
Hence we get an isomorphism of the representing $S''$-schemes
$$g: p_1^* Y' \stackrel{\sim}{\to} p_2^*Y'.$$
As the bundle $E$ is defined over the base $S$,
the functor $\Phi_{E/X/S}^{\tau}$ is defined over $S$-schemes, and
the other functors ($\Phi_{(\id \times p_1)^* E'/X''/S''}^{\tau}$ etc.)
are obtained from it by base changes. It follows that
the above isomorphism $g$ between the representative schemes
for these functors satisfies the cocycle condition when pulled back to
$S'\times _S S'\times_S S'$.
Hence $Y'$ descends to an algebraic space over $S$. By its construction,
$Y$ represents $\Phi_{E/X/S}^{\tau}$.
As the square
$$%
\begin{array}{ccc}
Y' & \to & Y \\
\da & & \da \\
S' & \to & S
\end{array}$$
is cartesian, as $S'\to S$ is an \'etale cover
and as $Y'\to S'$ is of finite type,
it follows that $Y\to S$ is of finite type.
By the same reasoning, as $Y'\to S'$ is separated by its construction,
it follows that $Y\to S$ is separated.
Next, we note that the Proposition \ref{uniqueness of reduction}
implies that $Y'\to S'$ is
injective at the level of underlying sets. By the above
reasoning, this implies that $Y\to S$ too is injective
at the level of underlying sets. In particular, $Y\to S$ is quasi-finite.
Given the above properties of $Y\to S$,
the Proposition \ref{SP representability by scheme} below
implies that $Y$ is a scheme. This is the desired scheme $S^{\tau}(E)$ by
its construction.
\hfill$\square$
\begin{proposition}\label{SP representability by scheme}
{\rm ([Stacks Project] Tag 03XX, Proposition 55.47.2.)}
Let $S$ be a scheme. Let $f:X \to T$ be a morphism of
algebraic spaces over $S$. Assume that
$T$ is representable,
$f$ is locally quasi-finite, and
$f$ is separated.
Then $X$ is representable.
\end{proposition}
{\bf Proof of Theorem \ref{stack version of main theorem}.}
We begin by recalling that the
stack $Bun_{X/S}(G)$ of $G$-bundles on fibers of $X/S$ is algebraic.
To see this, choose a closed embedding $G \hra GL_{n,k}$
as group schemes over $k$, and consider
the induced $1$-morphism of stacks $Bun_{X/S}(G)\to Bun_{X/S}(GL_{n,k})$.
The stack $Bun_{X/S}(GL_{n,k})$ is just the
stack of rank $n$ vector bundles on fibers of $X/S$, so it is an
algebraic stack (see [L-MB]). Given any $GL_{n,k}$-bundle $E$ on $X$,
the reductions of its structure group to $G$ are the sections of
$E/G \to X$, so they are parameterized by a suitable open subscheme of
the Hilbert scheme $Hilb_{(E/G)/S}$ (see for example [Ni-2] section 5.6.2
for an exposition).
This shows the $1$-morphism $Bun_{X/S}(G)\to Bun_{X/S}(GL_{n,k})$ is schematic,
which implies that the stack $Bun_{X/S}(G)$ is algebraic.
Next, given $\tau \in \ov{C}$, consider the $1$-morphism from the stack
$Bun_{X/S}^{\tau}(G)$ of Corollary 1.2 to the stack $Bun_{X/S}(G)$.
The Theorem \ref{family version of main theorem}
shows that this $1$-morphism is schematic and has the desired properties.
\hfill$\square$
{\bf Acknowledgement} S. Gurjar will like to thank the Tata Institute of
Fundamental Research for its generous hospitality during the
preparation of this article.
\bigskip
\bigskip
{\small
\parskip=2pt
{\large \bf References}
\bigskip
[Be] Behrend, K. : Semi-stability of reductive group schemes over curves.
Math. Ann. 301 (1995), 281-305.
[EGA] Grothendieck, A. and Dieudonn\'e, J. :
{\it \'El\'ements de g\'eom\'etrie alg\'ebrique},
Publ. Math. IHES., vols. 4, 8, 11, 17, 20, 24, 28, 32 (1960-1967).
[Gu] Gurjar, S. :
Restriction theorems for principal bundles in arbitrary characteristic.
J. Algebra 426 (2015), 79-91.
[G-N-1] Gurjar, S. and Nitsure, N. : Schematic Harder-Narasimhan
stratification for families of principal bundles and $\Lambda$-modules.
Proc. Indian Acad. Sci. (Math. Sci.) 124 (2014), 315-332.
[G-N-2] Gurjar, S. and Nitsure, N. :
Schematic Harder-Narasimhan stratification for families of
principal bundles in higher dimensions. arXiv:1505.02236
[He] Heinloth, J. : Bounds for Behrend's conjecture on the canonical
reduction. Int. Math. Res. Not. IMRN Vol. 2008, rnn045.
[La-MB] Laumon, G. and Moret-Bailly, L. : {\it Champs alg\'ebriques},
Springer (2000).
[K] Kleiman, S. : The Picard scheme. Part 5
of {\it Fundamental Algebraic Geometry -- Grothendieck's FGA
Explained}, Fantechi et al, Math. Surveys and Monographs Vol.
123, American Math. Soc. (2005).
[M-R] Mehta, V. B. and Ramanathan, A. : Semistable sheaves on
projective varieties and their restriction to curves.
Math. Ann. 258 (1981/82), 213-224.
[Ni-1] Nitsure, N. : Representability of Hom implies flatness.
Proc. Indian Acad. Sci. Math. Sci. {\bf 114} (2004)
[Ni-2] Nitsure, N. : Construction of Hilbert and Quot schemes. Part 2
of {\it Fundamental Algebraic Geometry -- Grothendieck's FGA
Explained}, Fantechi et al, Math. Surveys and Monographs Vol.
123, American Math. Soc. (2005).
[Ni-3] Nitsure, N. :
Schematic Harder-Narasimhan stratification.
Internat. J. Math. 22 (2011), 1365-1373.
[Stacks Project] : http://stacks.math.columbia.edu/
\bigskip
}
{\footnotesize
Sudarshan Gurjar \hfill Nitin Nitsure\\
Department of Mathematics \hfill School of Mathematics\\
Indian Institute for Technology, Bombay
\hfill Tata Institute of Fundamental Research\\
Powai \hfill Homi Bhabha Road\\
Mumbai 400 076 \hfill Mumbai 400 005\\
India \hfill India\\
{\tt <EMAIL>} \hfill {\tt <EMAIL>}
\bigskip
\bigskip
}
\end{document}
|
\section*{Appendix}
In the first section of Appendix \ref{app:revDMC} we provide a short review of the DMC method,
followed by a description of the DMC algorithm, the problem of the divergences in proximity of the nodal surface,
the instabilities in DMC simulations and the size-consistency issue met when DMC is stabilized by slightly modifying the algorithm. All this is used to contextualize the methodological improvements of this work.
Appendices \ref{app:MW}, \ref{app:big} and \ref{app:ice} provide further details on the three examples shown in the paper.
\section{Review of DMC }\label{app:revDMC}
DMC energy evaluations are mostly concerned with the
{\em mixed estimator}, defined as:
\begin{equation}\label{eq:Emix}
E_\textit{mix} = \frac{ \left< \phi \right| \hat H \left| \psi_G \right> }{ \left< \phi \right| \left. \psi_G \right> }
\end{equation}
where
$\psi_G$ is the {\em guiding function}
(a parametrized wave function optimized within VMC schemes in order to be as close as possible to the ground state) and
$\phi$ is the {\em exact ground state} wave function of the Hamiltonian $\hat H$.
As long as $\psi_G$ has a non-zero overlap with $\phi$, $E_\textit{mix}$ is equivalent to the {\em pure estimator}
$\frac{ \left< \phi \right| \hat H \left| \phi \right> }{ \left< \phi \right| \left. \phi \right>}$.
The exact wave function $\phi({\bf R})$ can be obtained from the solution
$\Phi({\bf R},t)$ of the imaginary time Schr\" odinger equation
\begin{equation}\label{eqn:itse}
- \frac{\partial \Phi({\bf R},t)}{\partial t} = -\frac{1}{2} \nabla^2 \Phi({\bf R},t) - (E_T - V_P({\bf R})) \Phi({\bf R},t)
\end{equation}
where $t$ is the time, ${\bf R}=({\bf r}_1, \ldots, {\bf r}_N)$ specifies the coordinates of the $N$ electrons,
$V_P$ is the potential energy and $E_T$ is an energy offset.
Given the boundary condition $\Phi({\bf R},0) = \psi_G({\bf R})$, for time $t \to \infty$
the imaginary time solution converges to the ground state:
$$
\lim_{t\to \infty} \Phi({\bf R},t) = \phi({\bf R}).
$$
It is often convenient to write the time evolution of $\Phi$ in terms of the Green function $G({\bf R \leftarrow R'};t)$:
\begin{equation}\label{eqn:gphi}
\Phi({\bf R},t_0+t) = \int G({\bf R \leftarrow R'};t) \Phi({\bf R'},t_0) d{\bf R'} .
\end{equation}
The Green function $G({\bf R \leftarrow R'};t)$, which satisfies an equation analogous to that of $\Phi$, prescribes how to propagate further in time the distribution $\Phi$.
Formally, we can write:
\begin{equation}\label{eqn:formalGbare}
G({\bf R \leftarrow R'};t) = \left< {\bf R} \right| e^{-t(\hat H - E_T)} \left| {\bf R'} \right> \,.
\end{equation}
Unfortunately, $G({\bf R \leftarrow R'};t)$ is not exactly known for realistic systems.
However, by considering that the time interval $t$ can be divided in $n$ smaller intervals of time $\tau = t/n$,
and iteratively using Eq.~\ref{eqn:gphi} to write
$\Phi({\bf R},t_{i})$ in terms of $\Phi({\bf R},t_{i-1})$,
with
$i=1,\ldots,n$ and
$t_i=t_0+i \tau$,
we obtain the following expression for the Green function:
\begin{widetext}
\begin{equation}\label{eqn:expandGbare}
G({\bf R \leftarrow R'};t) = \int G({\bf R \leftarrow R}_1;\tau) \ldots G({\bf R}_{n-1} \leftarrow {\bf R'},\tau) d{\bf R}_1 \ldots d{\bf R}_{n-1} \,.
\end{equation}
\end{widetext}
For a small enough time step $\tau$, the Green function can be approximated using the Trotter-Suzuki formula, which results in:
\begin{equation}\label{eqn:appG1}
G({\bf R \leftarrow R'};\tau) \approx G_b({\bf R \leftarrow R'};\tau) G_d({\bf R \leftarrow R'};\tau)
\end{equation}
where
$$
G_d({\bf R \leftarrow R'};\tau) = (2 \pi \tau)^{-\frac{3}{2} N } \exp\left[ - \frac{ ( {\bf R - R'} )^2 }{ 2 \tau } \right]
$$
is a diffusion term, and
$$
G_b({\bf R \leftarrow R'};\tau) = \exp\left[ \tau \frac{ 2 E_T - V_P({\bf R}) - V_P({\bf R'}) }{ 2 } \right]
$$
is a branching term.
The DMC algorithm is a stochastic realization of Eq.~\ref{eqn:gphi}, in which a series of {\it walkers} initially distributed as some $\Phi({\bf R},0)$
is propagated ahead in time with the short time approximation to the Green function in Eq.~\ref{eqn:appG1}. In the long time limit the walkers become distributed as $\phi({\bf R})$.
The method works perfectly well for bosons, as the ground state of the Hamiltonian is node-less. However, the fermionic ground state is generally difficult to calculate, because it is an excited state of the Hamiltonian. The difficulty comes from the fact that in the time evolution of Eq.~\ref{eqn:itse} the weight of the ground state becomes exponentially
dominant compared to excited states, and so the fermionic signal is quickly lost into noise.
The common solution is to embrace the {\em fixed node approximation}: $\Phi({\bf R},t)$ in constrained to have the same nodal surface of some guiding function $\psi_G({\bf R})$. The constraint makes DMC only approximate, and the variational principle then implies that the fixed-node DMC energy is an upper bound of the true fermionic ground state energy. If the nodal surface of the guiding function is exact then also the fixed-node DMC energy is exact.
The fixed-node constraint is conveniently implemented by introducing the mixed distribution $f({\bf R},t) = \psi_G({\bf R}) \Phi({\bf R},t)$,
which satisfies the equation:
\begin{equation}\label{eqn:mix}
- \frac{\partial f({\bf R},t)}{\partial t} = -\frac{1}{2} \nabla^2 f({\bf R},t) +\nabla \cdot [{\bf V}({\bf R}) f({\bf R},t)] - S({\bf R}) f({\bf R},t)
\end{equation}
(see Eq.~\ref{eqn:mixed}), where
${\bf V(R)} \equiv \nabla \log \left| \psi_G({\bf R}) \right|$ is the {\em drift velocity}, or local gradient, and
$S({\bf R}) \equiv E_T - E_L({\bf R})$ is the {\em branching} term, with $E_L({\bf R}) = \psi_G({\bf R})^{-1} \hat H \psi_G({\bf R})$ the {\em local energy}.
Note that in Eq.~\ref{eqn:mix} there is an additional drift term that was not present in the original imaginary time Schr\" odinger equation for $\Phi$.
The mixed distribution $f$ has the border condition
$
f({\bf R},0) = \psi_G({\bf R})^2
$
and, in the limit of large time $t$:
$$
\lim_{t\to \infty} f({\bf R},t) = \psi_G({\bf R}) \phi({\bf R}).
$$
Thus, the mixed estimator can be written as:
\begin{equation}\label{eq:Emix2}
E_\textit{mix} = \lim_{t\to \infty} \frac{ \int E_L({\bf R}) f({\bf R},t) d{\bf R} }{ \int f({\bf R},t) d{\bf R} }.
\end{equation}
It is convenient to write the time evolution of $f$ in terms of the Green function $\tilde G({\bf R \leftarrow R'};t)$, which prescribes how to propagate further in time the distribution $f$:
\begin{equation}\label{eqn:gf}
f({\bf R},t_0+t) = \int \tilde G({\bf R \leftarrow R'};t) f({\bf R'},t_0) d{\bf R'} ,
\end{equation}
where $\tilde G({\bf R \leftarrow R'};t)$ satisfies an equation analogous to that of $f$, and formally can be written as:
\begin{equation}\label{eqn:formalG}
\tilde G({\bf R \leftarrow R'};t) = \frac{\psi_G({\bf R})}{\psi_G({\bf R'})} \left< {\bf R} \right| e^{-t(\hat H - E_T)} \left| {\bf R'} \right> \,.
\end{equation}
Again, $\tilde G({\bf R \leftarrow R'};t)$ is not exactly known for realistic systems, but we can use the same trick of splitting $t$ in $n$ time steps of length $\tau=t/n$.
We obtain the following expression for the Green function:
\begin{widetext}
\begin{equation}\label{eqn:expandG}
\tilde G({\bf R \leftarrow R'};t) = \int \tilde G({\bf R \leftarrow R}_{n-1};\tau) \ldots \tilde G({\bf R}_{1} \leftarrow {\bf R'},\tau) d{\bf R}_1 \ldots d{\bf R}_{n-1} \,.
\end{equation}
For a small enough time step $\tau$,
$\tilde G({\bf R}_i, {\bf R}_{i+1};\tau)$
is approximated by the Green functions for purely drift, diffusion and branching processes.
This leads to:
\begin{equation}\label{eqn:appG}
\tilde G({\bf R \leftarrow R'};\tau) \approx \tilde G_b({\bf R \leftarrow R'};\tau) \tilde G_d({\bf R \leftarrow R'};\tau)
\end{equation}
where
$$
\tilde G_d({\bf R \leftarrow R'};\tau) = (2 \pi \tau)^{-\frac{3}{2} N } \exp\left[ - \frac{ ( {\bf R - R' - \tau V(R')} )^2 }{ 2 \tau } \right]
$$
is the drift-diffusion term,
and
$$
\tilde G_b({\bf R \leftarrow R'};\tau) = \exp\left[ \tau \frac{ S({\bf R}) + S({\bf R'}) }{ 2 } \right]
$$
is the branching term.
\end{widetext}
Eq.~\ref{eqn:mix} also introduces {\em importance sampling}.
Beside concentrating the sampling in the important part of the phase space, an additional
advantage of importance sampling over simple sampling is that the branching term depends on the local energy $E_L({\bf R})$, and not on the potential energy $V_P({\bf R})$. Since $E_L({\bf R})$ is much smother than $V_P({\bf R})$, and it is constant in the limit of $\psi_G \sim \phi$, the stability of the DMC simulation is greatly enhanced.
The error on this approximate expression for $\tilde G({\bf R}_i, {\bf R}_{i+1};\tau)$ can be evaluated using the
Zassenhaus formula~\cite{Suzuki:1977bw},
and the leading correction is of order ${\cal O}(\tau^2)$.
This translates into an error of order ${\cal O}(\tau)$ on $\tilde G({\bf R, R'};t)$ (see Eq.~\ref{eqn:expandG}).
In the limit of $ \tau \rightarrow 0$ the error on the Green function is zero, but the computational cost is $\propto 1/\tau$ because $\tilde G_b({\bf R \leftarrow R'};t)$ is split in $n=t/\tau$ terms.
\subsection{ DMC algorithm }
We discuss here how the DMC algorithm actually works.
At each time $t$ the distribution $f({\bf R},t)$ can be represented by a discrete set $\{ {\bf R}^\alpha(t),w^\alpha(t) \}_{\alpha=1,\ldots,n_w(t)}$ of walkers ({\em i.e.} sampling points ${\bf R}^\alpha$ with a weight $w^\alpha$),
such that
$f({\bf R},t) \sim { \sum_\alpha w^\alpha \delta({\bf R} - {\bf R}^\alpha) / \sum_\alpha w^\alpha }$.
By using the Metropolis algorithm we can easily generate an ensemble of configurations
$\{ {\bf R}^\alpha \}_{\alpha=1,\ldots,n_w}$ ({\em i.e.}, a set of walkers with unit weight) that correspond to the initial distribution $f({\bf R},0)= \psi_G({\bf R}_n)^2$.
In DMC we need to project forward in time the walkers in order to calculated the mixed distribution for $f({\bf R},t\to \infty)$.
If in Eq.~\ref{eq:Emix2} we
express the mixed distribution $f({\bf R},t)$ as in Eq.~\ref{eqn:gf} (with initial distribution $f({\bf R},0) = \psi_G({\bf R})^2$),
and we expand the Green function as in Eq.~\ref{eqn:expandG} (with $t=n\tau$),
we obtain that the mixed estimator is rewritten in the following way:
\begin{widetext}
\begin{equation}\label{eqn:Emix3}
E_\textit{mix} = \lim_{n\to \infty} \frac{
\int E_L({\bf R}_n)
\tilde G({\bf R}_n \leftarrow {\bf R}_{n-1};\tau) \ldots
\tilde G({\bf R}_{1} \leftarrow {\bf R}_0,\tau) \psi_G({\bf R}_n)^2
d{\bf R}_0 \ldots d{\bf R}_n
}{
\int
\tilde G({\bf R}_n \leftarrow {\bf R}_{n-1};\tau) \ldots
\tilde G({\bf R}_{1} \leftarrow {\bf R}_0,\tau) \psi_G({\bf R}_n)^2
d{\bf R}_0 \ldots d{\bf R}_n
} \, ,
\end{equation}
and using the approximation in Eq.~\ref{eqn:appG} for the Green function with small $\tau$ we have:
\begin{equation}\label{eqn:Emix4}
E_\textit{mix} \simeq \lim_{n\to \infty} \frac{
(2 \pi \tau)^{-\frac{3}{2} n N }
\int
E_L({\bf R}_n)
\prod_{i=0}^{n-1} \left\{
\exp\left[ - \frac{ ( {\bf R}_{i+1} - {\bf R}_i - \tau {\bf V}({\bf R}_i) )^2 }{ 2 \tau } \right]
\exp\left[ \tau \frac{ S({\bf R}_{i+1}) + S({\bf R}_i) }{ 2 } \right]
\right\}
\psi_G({\bf R}_n)^2
d{\bf R}_1 \ldots d{\bf R}_n
}{
(2 \pi \tau)^{-\frac{3}{2} n N }
\int
\prod_{i=0}^{n-1} \left\{
\exp\left[ - \frac{ ( {\bf R}_{i+1} - {\bf R}_i - \tau {\bf V}({\bf R}_i) )^2 }{ 2 \tau } \right]
\exp\left[ \tau \frac{ S({\bf R}_{i+1}) + S({\bf R}_i) }{ 2 } \right]
\right\}
\psi_G({\bf R}_n)^2
d{\bf R}_1 \ldots d{\bf R}_n
} \, .
\end{equation}
\end{widetext}
Thus, according to the RHS of Eq.~\ref{eqn:Emix4}, each walker
evolves in time according to a branching-drift-diffusion process:
given the configuration ${\bf R}^\alpha_i$ and weight $w^\alpha_i$ at time $t=i*\tau$, the walker drift-diffuse as follows:
\begin{equation}\label{eq:driftdiff}
{\bf R}^\alpha_{i} \to
{\bf R}^\alpha_{i+1} = {\bf R}^\alpha_i + \tau {\bf V}({\bf R}^\alpha_i) + \sqrt{\tau} { \eta} \,,
\end{equation}
where ${ \eta}$ is a $3N$-dimensional random vector generated from a normal distribution with zero mean and unit variance, and the walker weight evolves as:
\begin{equation}\label{eq:weight}
w^\alpha_{i} \to
w^\alpha_{i+1} = w^\alpha_i * \exp\left[ \tau \frac{ S({\bf R}^\alpha_{i+1}) + S({\bf R}^\alpha_i) }{ 2 } \right] \,.
\end{equation}
The evolution of the weight is efficiently realized by using a branching (birth/death) algorithm, where walkers with small weight are killed and walkers with high weight are replicated \cite{foulkes01}.
Moreover, a Metropolis acceptance/rejection move is usually introduced after the drift-diffusion step\cite{Reynolds:1982,umrigar93}, in order to satisfy the detailed balance and reduce the time-step error, and with that an efficient time-step $\tau_\textrm{\small eff}$, which rescales the nominal time-step $\tau$ taking into account the acceptance probability, is used in Eq.~\ref{eq:weight} in place of $\tau$.
Finally, given the chosen time-step $\tau$ and a sufficiently large number $n$ of DMC steps, the mixed energy is calculated as:
\begin{equation}\label{eqn:Emix5}
E_\textit{mix}^\tau =
\frac{
\left< E_L({\bf R}_{n}^\alpha) w_{n}^\alpha \right>_\alpha
}{ \left< w_{n}^\alpha \right>_\alpha } \,,
\end{equation}
where $\left< \cdot \right>_\alpha$ is the average over all the walkers.
Clearly, this evaluation is affected by a stochastic error inversely proportional to the square root of the number $n_w$ of walkers.
In order to increase the precision of the evaluations it is not necessary to use a huge number of walkers; it is much more efficient, because of the equilibration time, to propagate further in time the walkers and to use the following expression to evaluate the mixed energy:
\begin{equation}\label{eqn:actualEmix}
E_\textit{mix}^\tau =
{1 \over M} \sum_{m=1}^M
\frac{
\left< E_L({\bf R}_{n+m}^\alpha) w_{n+m}^\alpha \right>_\alpha
}{ \left< w_{n+m}^\alpha \right>_\alpha } \,.
\end{equation}
Notice that in Eq.~\ref{eqn:actualEmix} the walkers provide almost independent evaluations, but the local energies are instead serially correlated, with a correlation time proportional to $\tau$. Thus, in evaluating the stochastic error for the mixed energy it is important to get rid of the serial correlation, for instance by using the ``blocking method'' \cite{flyvbjerg_error_1989}.
Sometimes the estimator actually used can be slightly different from Eq.~\ref{eqn:actualEmix} -- for instance some corrections are sometimes introduced in order to correct for the finite population bias ({\em i.e.}, having a finite number of walkers can introduce a bias) -- but the size-consistency issue here addressed is unaffected by these corrections.
\subsection{Divergences in proximity of the nodal surface}
Close to the nodal surface $\Sigma_G$ of the guiding function $\psi_G$ the approximation in Eq.~\ref{eqn:appG} is problematic, because a configuration ${\bf R}$ at a distance $\delta$ from $\Sigma_G$ has both the local gradient $\bf V(R)$ and the local energy $E_L({\bf R})$ (and consequently the branching term $S({\bf R})$) diverging in modulus as $1/\delta$, leading to instabilities and big finite time step errors.
This problem has been tackled both by \citet{depasquale88} and \citet{umrigar93}, who proposed modifications for
$\bf V(R)$ and for $S({\bf R})$ for $\bf R$ close to $\Sigma_G$ to eliminate these divergences.
These modifications are strictly related to the size-inconsistency issue addressed in this work.
\subsection{DMC instabilities}\label{app:instab}
DMC instabilities are uncontrolled walker population fluctuations ({\em i.e.}, weights $w^\alpha_i$ experiencing huge changes in a single step $i\to i+1$, see Eq.~\ref{eq:weight}), which jeopardize the DMC energy evaluations and makes the simulation unfeasible. They are mainly due to walkers reaching regions of diverging local energy (because of the pseudo-potential or proximity to the nodal surface), and in particular for $E_L({\bf R}) \to -\infty$ the branching term $S({\bf R})$ leads to proliferation of walkers from just one problematic configuration.
Instabilities are strictly related with time step $\tau$: with small $\tau$ instabilities are usually under control, but as larger and larger values of $\tau$ are considered instabilities are more often observed.
The reason is that the diffusion step is random and proportional to $\sqrt{\tau}$, see Eq.~\ref{eq:driftdiff}, and if the time-step is too large there is some chance to fall into the problematic regions, because the drift step
is unable to keep electrons away for the divergences.
A small enough $\tau$ allows the drift step to recover from a ``bad'' diffusion step.
As a matter of fact, DMC simulations with no modifications to the drift and branching terms are stable only for tiny values of $\tau$, making schemes as those proposed by \citet{depasquale88} or \citet{umrigar93} necessary in actual calculations, the latter being much more stable than the former.
The new limiting scheme proposed in this work
(which is the same of \citet{umrigar93} for the drift, Eq.~3 of the letter, and the one in Eq.~5 of the letter for the branching) appears as effective as the limiting scheme of \citet{umrigar93} (see Eqs.~3 and 4 of the letter), if not better, in keeping the DMC simulation stable.
A pragmatic way to recover from a diverging population count (population explosion) is to back-track the simulation to a region far from the instability, run the random number generator idle for a number of cycles, and resume the DMC simulation. Often this procedure sends the simulation to a different region of phase space, avoiding the instability. However, if the instabilities are too frequent, the simulation becomes impractical or even impossible. To highlight the improvement in the stability of the calculations using the new limiting procedure, consider for example the CH$_4$ - H$_2$O dimer in the bound configuration. Using the UNR limiting procedure and $\tau= 0.05$ a.u. we encountered 32 population explosions in $\sim 26,000$ steps (population size: 20,480 walkers). No simulations were possible with any larger value of time step. By contrast, using the new limiting procedure we observed no instabilities in $\sim 176,000$ steps at $\tau= 0.05$ a.u., and also no instabilities in $\sim 250,000$ steps at $\tau= 0.1$ a.u..
\subsection{ Size-consistency in DMC }\label{sec:sizecons}
As discussed in the letter, a method is size-consistent if the energy $E_{AB}$ of any system $AB$ constituted by the two non-interacting subsystems $A$ and $B$, is equal to the sum $E_A+E_B$ of the energies of individual subsystems.
As in the letter, we assume here to deal with systems that are size-consistent when described with a single Slater determinant (so, also with a Jastrow correlated single Slater determinant).
In this section we show that the fixed-node DMC with importance sampling ({\em i.e.}, Eq.~\ref{eqn:Emix5}) is size-consistent for any $\tau$, but if the modifications to the branching proposed by \citet{umrigar93} are used DMC is size-consistent only in the limit of $\tau\to 0$.
Clearly, any configuration ${\bf R}^{[AB]}$ of the systems $AB$ is given by the configurations ${\bf R}^{[A]} $ and ${\bf R}^{[B]} $ of the subsystems $A$ and $B$, because any electron in AB belongs either to the subsystem $A$ or to $B$.
Mathematically, this means that the vectorial space where the configurations ${\bf R}^{[AB]}$ live is the direct sum of the two vectorial spaces where ${\bf R}^{[A]} $ and ${\bf R}^{[B]} $ live, and we can write (with a little abuse of notation):
\begin{equation}\label{eq:Rab}
{\bf R}^{[AB]} = {\bf R}^{[A]} \oplus {\bf R}^{[B]} \,.
\end{equation}
As discussed in the letter, the guiding wave function factorizes, {\em i.e.}:
\begin{equation}\label{eq:psiab}
\psi_G^{[AB]}({\bf R}^{[AB]}) = \psi_G^{[A]}({\bf R}^{[A]}) \otimes \psi_G^{[B]}({\bf R}^{[B]})
\end{equation}
whenever $A$ and $B$ are far away.
From the properties of the hamiltonian operator it follows that the local energy is additive:
\begin{equation}\label{eq:elocadd}
E_L^{[AB]}({\bf R}^{[AB]}) = E_L^{[A]}({\bf R}^{[A]}) + E_L^{[B]}({\bf R}^{[B]}) \,,
\end{equation}
which proves that VMC is size-consistent.
Moreover, considering that the drift velocity is the local gradient, it is easy to show that:
\begin{equation}\label{eq:driftadd}
{\bf V}^{[AB]}({\bf R}^{[AB]}) = {\bf V}^{[A]}({\bf R}^{[A]}) \oplus {\bf V}^{[B]}({\bf R}^{[B]}) \,,
\end{equation}
where the symbol $\oplus$ is used in the same way as in Eq.~\ref{eq:Rab}.
In order to address the properties of the DMC mixed energy $E_\textit{mix}^\tau$ evaluated for a finite value $\tau$ of the time-step, we can consider Eq.~\ref{eqn:Emix5}.
According to Eq.~\ref{eq:weight}, the weight is
$w_{n}^\alpha = \exp{\left[ \tau \sum_{i}^n S({\bf R}_{i}^\alpha) \right]}$ (here, for simplicity, we have slightly simplified the expression, neglecting that the first and last step have a weight that is 1/2)
and including that the branching term $S({\bf R}_{i}^\alpha) = E_T - E_L({\bf R}_{i}^\alpha)$, it is straightforward to see that:
\begin{equation}
\label{eqn:Emix6}
E_\textit{mix}^\tau =
\frac{
\left< E_L({\bf R}_{n}^\alpha) e^{ - \tau \sum_{i}^n E_L({\bf R}_{i}^\alpha) } \right>_\alpha
}{ \left< e^{ - \tau \sum_{i}^n E_L({\bf R}_{i}^\alpha) } \right>_\alpha } \,.
\end{equation}
By using Eq.~\ref{eqn:Emix6}, the additivity of the local energy (Eq.~\ref{eq:elocadd}) and of the drift velocity (Eq.~\ref{eq:driftadd}), and some algebra, it is easy to prove that:
\begin{widetext}
\begin{equation}\label{eqn:SCdmc}
\textrm{DMC with no modifications:} \qquad
{E_\textit{mix}^\tau}^{[AB]} =
{E_\textit{mix}^\tau}^{[A]} + {E_\textit{mix}^\tau}^{[B]}
\,,
\end{equation}
for any value of the time-step $\tau$, and of course also for $\tau\to 0$.
The main point of the proof is that the additivity of the local energy imply the factorization of the weight, {\em i.e.}:
\begin{equation}
e^{ - \tau \sum_{i}^n E_L^{[AB]}({{\bf R}_{i}^\alpha}^{[AB]})} =
e^{ - \tau \sum_{i}^n E_L^{[A]}({{\bf R}_{i}^\alpha}^{[A]})} *
e^{ - \tau \sum_{i}^n E_L^{[B]}({{\bf R}_{i}^\alpha}^{[B]})}
\qquad \textrm{for any walker } \alpha.
\end{equation}
In principle, it could be explicitly tested that DMC with no modifications satisfy the size-consistency for any finite time-step, but in practice it can be done only for very small values of $\tau$ because of the instabilities discussed in Section~\ref{app:instab}.
The UNR modification to the drift, as reported in Eq.~3 of the letter, does not affect the additivity of the drift (because the correction is performed independently for each electron), and we have that:
\begin{equation}\label{eq:UNRdriftadd}
\textrm{DMC with UNR modifications:} \qquad
{\bf \bar V}^{[AB]}({\bf R}^{[AB]}) = {\bf \bar V}^{[A]}({\bf R}^{[A]}) \oplus {\bf \bar V}^{[B]}({\bf R}^{[B]}) \,.
\end{equation}
which clearly does not affect the size-consistency of the method.
The source of the size-inconsistency is instead the UNR modification to the branching term, see Eq.~4 in the letter, because we have that:
\begin{equation}\label{eq:UNRbranchadd}
\textrm{DMC with UNR modifications:} \qquad
{\bar S}^{[AB]}({\bf R}^{[AB]}) \neq {\bar S}^{[A]}({\bf R}^{[A]}) + {\bar S}^{[B]}({\bf R}^{[B]}) \,.
\end{equation}
because of the term ${\bar V / V}$ appearing in the expression of $\bar S$.
This imply that the weight of a DMC realization does not factorize any more, that is:
\begin{equation}\label{eqn:UNRweight}
\textrm{DMC with UNR modifications:} \qquad
e^{ \tau \sum_{i}^n \bar S^{[AB]}({{\bf R}_{i}^\alpha}^{[AB]})} \neq
e^{ \tau \sum_{i}^n \bar S^{[A]}({{\bf R}_{i}^\alpha}^{[A]})} *
e^{ \tau \sum_{i}^n \bar S^{[B]}({{\bf R}_{i}^\alpha}^{[B]})}
\,.
\end{equation}
However, in the limit of $\tau \to 0$ we have that $\bar V \to V$ and $\bar S \to S$, thus UNR approaches asymptotically the case of no modifications, where size-consistency is proven.
The scheme proposed in this work (named here ZSGMA, from authors' names), see Eqs.~5 and 6 in the letter, is exactly size-consistent for $E_\textrm{cut}\to \infty$ (namely, for $\alpha\to\infty$ or $\tau\to0$), because the branching $\bar S$ becomes equivalent to $S$, which factorizes exactly, so we recover the unmodified DMC algorithm.
The method is only approximated for finite $E_\textrm{cut}$;
the modified branching term is not exactly additive, {\em i.e.}
${\bar S}^{[AB]}({\bf R}^{[AB]}) \neq \sum_{i}^n {\bar S}^{[A]}({\bf R}^{[A]}) + \sum_{i}^n {\bar S}^{[B]}({\bf R}^{[B]})$,
but what we approximatively satisfy is that:
\begin{equation}\label{eq:ZSGMAbranch}
\textrm{DMC with ZSGMA modifications:} \qquad
\sum_{i}^n {\bar S}^{[AB]}({\bf R}^{[AB]}) \sim
\sum_{i}^n {\bar S}^{[A]}({\bf R}^{[A]}) + \sum_{i}^n {\bar S}^{[B]}({\bf R}^{[B]}) \,,
\end{equation}
at least when $E_\textrm{cut}$ is large enough.
This happens because, assuming that $E_T$ is properly set,
we have that $\bar S$ can be seen as a random variable of zero mean and a variance proportional to $\sqrt{N}$.
In order to satisfy Eq.~\ref{eq:ZSGMAbranch}, at least approximatively, we require that the number of times we perform a cut on $\bar S$ is independent on the size of the system and with a random sign. This implyes a value of $E_\textrm{cut} \propto \sqrt{\textrm{VAR}(\bar S)}$.
\end{widetext}
\newpage
\section{Water-Methane dimer}\label{app:MW}
\begin{figure}[ht]
\includegraphics[width=3.5in]{plot_absene_MW.eps}
\caption{\label{fig:Fig2} Energies of the
CH$_4$ (top panel),
H$_2$O (middle),
and unbounded CH$_4$-H$_2$O (bottom)
systems as function of time step $\tau$, calculated using the UNR and present work prescriptions for the limits on the branching factor. Error bars are smaller than the size of the symbols. }
\end{figure}
In Fig.~\ref{fig:Fig2} we display the energy of the dimer, $E_{\rm CH_4-H_2O(\rm shifted)}$ as well and the energies of the monomers, $E_{\rm CH_4}$ and $E_{\rm H_2O}$, computed in independent calculations performed with simulation cells containing either the CH$_4$-H$_2$O(\rm shifted) dimer or the isolated CH$_4$ and H$_2$O monomers, respectively.
Single particle wavefunctions were obtained using a plane-wave cutoff of 300 Ry, and re-expanded in terms of B-splines with the natural grid spacing $a=\pi/G{\rm max}$, where $G_{\rm max}$ is the magnitude of the largest plane wave in the expansion.
The Jastrow factor used in the trial wavefunction of the system included a two-body electron-electron (e-e) term; three different two-body electron-nucleus (e-n) terms for C, O and H, respectively; and three different three-body electron-electron-nucleus (e-e-n) terms, for C, O and H. Of course, for the isolated CH$_4$ and H$_2$O systems we only included the e-n and the e-e-n terms for C, H and O, H, respectively, but a part form this difference the Jastrow factors were exactly the same in all systems. The cutoff radii of the e-e, e-n, and e-e-n terms were all lower than 3.5~\AA, and the large distance between the two molecules guarantees that the overlap between their respective orbitals is effectively zero. Therefore the trial wavefunction of the dimer $\psi_{\rm CH_4-H_2O(shifted)}$, is effectively the appropriately antisymmetrised product of the trial wavefunctions $\psi_{\rm CH_4}$ and $\psi_{\rm H_2O}$ of the CH$_4$ and the H$_2$O sub-systems, respectively: $\psi_{\rm CH_4-H_2O(shifted)} = \psi_{\rm CH_4} \otimes\psi_{\rm H_2O}$. The variances of the local energy with the variational Monte Carlo (VMC) distributions were $\sim 0.72, 0.26$ and $0.45$ Ha$^2$ for the CH$_4$-H$_2$O, CH$_4$ and H$_2$O systems, respectively.
As seen in the paper, the finite time-step error in the binding energy, whenever the $E_b$ evaluation is used, is mostly due to the size consistency error.
The speedup obtained by using present work prescriptions for the branching factor in comparison with UNR branching factor is of two orders of magnitude, as it is shown in Fig.~\ref{fig:accuracy}(left). In this system there is the possibility to use $E_{bs}$ and to alleviate the size-consistency issue of the UNR prescription for the branching factror. However, when big clusters or molecular crystals are considered, $E_{bs}$ could be an unfeasible choice.
\section{The C$_{60}$-C$_{60}$H$_{28}$ complex}\label{app:big}
As for the water-methane dimer, single particle wavefunctions were obtained using a plane-wave cutoff of 300 Ry, and re-expanded in terms of B-splines with the natural grid spacing $a=\pi/G{\rm max}$.
The Jastrow factor (e-e), (e-n) and (e-e-n) terms, and was constructed with the same procedure as in the water-methane system, i.e. by ensuring that it is the same in all systems. The variances of the VMC local energies were $\sim 11, 5.4$ and $5.8$ Ha$^2$ for the C$_{60}$-C$_{60}$H$_{28}$, C$_{60}$ and C$_{60}$H$_{28}$ systems, respectively.
\begin{figure}
\includegraphics[width=3.5in]{plot_absene_big.png}
\caption{\label{fig:Fig7} Energies of
C$_{60}$ (top panel), C$_{60}$H$_{28}$(middle), and
unbonded C$_{60}$-C$_{60}$H$_{28}$ (bottom) systems as function of time step $\tau$, calculated using the UNR and present work prescriptions for the limits on the branching factor. Error bars are smaller than the size of the symbols.}
\end{figure}
In Fig.~\ref{fig:Fig7} we display the energy of the supramolecular system, $E_{\rm C_{60}-C_{60}H_{28}}$ as well as the energies of the monomers, $E_{\rm C_{60}}$ and $E_{\rm C_{60}H_{28}}$, computed in independent calculations performed with simulation cells containing either the isolated C$_{60}$ and C$_{60}$H$_{28}$ molecules, respectively.
The improved accuracy of present work prescriptions for the branching factor in comparison with the UNR branching factor can be appreciated in Fig.~\ref{fig:accuracy}(center).
\section{Two dimensional square ice}\label{app:ice}
We considered a monolayer of flat square ice of water, that is a system with 2-dimensional periodicity that is attaining considerable attention~\cite{Chen:2016gz,AlgaraSiller:2015cb}.
The unit cell include four water molecules, and here we considered a $4 \times 4$ supercell, for a total of 64 waters in the system.
The cohesive energy is obtained by subtracting the energy of the relevant number of isolated water molecules.
Single particle wavefunctions were obtained using a plane-wave cutoff of 600 Ry, and re-expanded in terms of B-splines with the natural grid spacing $a=\pi/G{\rm max}$.
The larger plane-wave cutoff used for these calculations resulted in a lower variance of the VMC local energies, which was $\sim 0.28$~Ha$^2$ for the isolated molecule, and $\sim 19.8$~Ha$^2$ for the square ice (corresponding to $\sim 0.31$~Ha$^2$ per water molecule).
At the VMC level of theory the evaluated cohesive energy is -0.108(4)~eV, that is severely underestimated (by a factor 4) with respect to the DMC evaluations.
\begin{figure}
\includegraphics[width=3.5in]{plot_absene_2dice_v2.png}
\caption{\label{fig:ice}
Energies of an isolated water molecule (top), and of a water molecule in a periodic two dimensional squale ice (bottom) systems as function of time step $\tau$, calculated using the UNR and present work prescriptions for the limits on the branching factor. Error bars are smaller than the size of the symbols. }
\end{figure}
In Fig.~\ref{fig:ice} we display the energy of the isolated water molecule, as well as the energy per water in the square lattice 2-dimensional system.
A comparison with Fig.~\ref{fig:Fig2} shows that the higher quality of the trial wavefunctions for this system results in a lower time step error.
The speedup obtained with present work prescriptions for the branching factor in comparison with the UNR branching factor can be appreciated in Fig.~\ref{fig:accuracy}(left).
\include{paper.bbl}
\end{document}
|
\section*{Abstract}
Spatial big data have the ``velocity,'' ``volume,'' and ``variety'' of big data sources and \edit{contain additional geographic information}. Digital data sources, such as medical claims, mobile phone call data records, and geo-tagged tweets, have entered infectious disease epidemiology as novel sources of data to complement traditional infectious disease surveillance. In this work, we provide examples of how spatial big data have been used thus far in epidemiological analyses and describe opportunities for these sources to improve disease mitigation strategies and public health coordination. In addition, we consider the technical, practical, and ethical challenges with the use of spatial big data in infectious disease surveillance and inference. Finally, we discuss the implications of the rising use of spatial big data in epidemiology to health risk communication, and public health policy recommendations and \edit{coordination across scales}. \\
\noindent
\textbf{Keywords: spatial big data, spatial epidemiology, disease mapping, infectious disease, digital epidemiology, statistical bias}\\
\noindent
Abstract word count: 134; Text word count: 2771
\end{titlepage}
\doublespacing
\section*{Introduction}
During one of epidemiology's formative moments, John Snow mapped London households with cholera and succeeded in highlighting the risk associated with the Broad Street pump. Since then, spatial investigations have played a critical role in improving our understanding of the associations between risks and disease outcomes. In infectious disease epidemiology, we ask: \emph{Which populations are at higher risk for disease? Where did this outbreak originate? Where can we expect future disease outbreaks to arise?} \textemdash fundamentally, these are spatial questions that rely on spatial data for answers.
Traditional infectious disease epidemiology is built on the foundation of \edit{relatively} high quality and high accuracy data on disease (e.g., serological diagnostics) and behavior (e.g., \edit{vaccination} surveys). These data are usually characterized by small size, but they benefit from \edit{control groups or designed observational samples from known underlying populations, thus rendering} it possible to make population-level inferences. On the other hand, digital infectious disease epidemiology \edit{typically} uses existing digital traces, re-purposing them to identify patterns in health-related processes. Digital data are electronic in form, and can often be characterized as ``big data'' when they are produced in large volumes (i.e. a large number of subjects or a large number of measurements per subject), with high velocity (i.e. data created in near real-time), and have variety in sources and organizational structures \cite{Laney2001}. When big data are characterized by fine spatial granularity, identifying point or areal locations, we refer to them \edit{here} as \edit{\textit{spatial big data}}. Big data provide opportunities for infectious disease epidemiology and public health because they increase accessibility to populations over space and time; data on personal beliefs, behaviors, and health outcomes are now available at unprecedented breadth and depth. The trade-off to this tremendous access is \edit{the potential for} loss of quality and accuracy. Streams of digital data relevant to public health \edit{may} serve as proxies for a desired measure, \edit{but these datasets may not meet the assumptions for} standard \edit{methods of epidemiological comparison (e.g., self-reported symptoms on social media and serological diagnoses both serve as proxies for ``true cases,'' but they have different biases and collection procedures, and represent different populations)}.
The trade-off between access and accuracy, and the task of separating true signal from large and varied noise characterizes the challenge and opportunity of big data for infectious disease epidemiology \cite{Khoury2014}. In this piece, we focus specifically on spatial big data and its applications to the field of spatial epidemiology. We highlight the opportunities for spatial big data to improve spatial modeling and data coverage, and describe ongoing challenges as spatial big data become more pervasive in informing disease surveillance, disease control, and public health policy.
\section*{Spatial big data open new doors in epidemiology}
True to the promise of \emph{variety} in big data streams, several familiar technologies produce spatial big data that can be used for infectious disease surveillance and modeling. Social media sites like Facebook and Twitter allow users to tag individual posts with specific locations, linking \edit{geography} to specific health behaviors. Mobile phones send signals with GPS locations and their call data records are spatially referenced through cell tower locations, both of which enable the measurement of human activity and mobility \edit{\cite{Wesolowski2012a, Wesolowski2015a}}. Web search data may capture user location through internet IP addresses, and online encyclopedia (Wikipedia) access logs may identify locations based on the search language \cite{Ginsberg2009, Generous2014}. Administrative medical claims and pharmacy transactions indicate the location of healthcare facilities and drugstores where patients seek care and medications \edit{\cite{Viboud2014, Pivette2014}}. Restaurant reservation cancellations on sites like OpenTable may provide insight into disease incidence in specific cities \cite{Nsoesie2014}.
Infectious disease epidemiology has already witnessed an impact from spatial big data, and the development of new methodologies and improvements to computational efficiency will only increase the potential of these data sources. Satellite imaging to infer climate, land use, and population density information has contributed to a better understanding of the spatial distribution of critical mosquito disease vectors and the seasonal epidemic dynamics of measles \citep[e.g.,][]{Kraemer2015, Bharti2011}. HealthMap, an automated, online news and outbreak reporting aggregator, has enabled the assimilation of disparate sources of disease occurrence data and has been used to examine spatial dynamics of cholera \cite{tuite2011}. Mobile phone call data records have provided insights into human mobility that have informed risk maps, importation potential, and spatial dynamics of dengue and malaria \citep[e.g.,][]{Wesolowski2015a}. Medical claims data have been used to examine spatial heterogeneity in influenza epidemic timing and severity \cite{Gog2014, Lee2015}, while geo-referenced Twitter data has been used to identify spatial anti-vaccination sentiment \cite{Salathe2011}.
While these studies with spatial big data have leveraged the fine spatial resolution to develop a detailed understanding of disease risk, there remain untapped opportunities with real-time surveillance, large-scale ecological inference, and adaptive disease mitigation strategies. Harnessing disease data from digital sources may enable epidemiological analyses to be performed at finer spatial scales in areas with poor coverage from traditional public health surveillance, and traditional and digital sources of spatial big data may be combined to \edit{account for the bias and gaps in each \citep[e.g.,][]{Shaman2013}}. The assimilation of multiple spatial big data sources through flexible \edit{statistical} modeling methods, and the continuous nature of data streams could enable near real-time dynamic disease mapping and risk mapping in the near future. For example, Bayesian statistical approaches have emerged as tools for merging multi-scale big data sources, incorporating explicit spatial dependencies into maps and models, and providing a framework for joining disease surveillance data across spatial scales while explicitly capturing the variation \edit{in} measurement bias across locations \citep[e.g.,][]{Corberan-Vallet2014}. \edit{Finally, access to multiple spatial scales of data allows one scale with missing observations to ``borrow information'' from a different scale through the addition of contextual effects in modeling inference \cite{Aregay2015b}}.
\section*{Spatial big data presents technical challenges}
While big data offer significant opportunities for epidemiological modeling and analysis, they also present a variety of technical and practical challenges. The measurement of incomplete and unrepresented populations, the lack of consistency and reliability in data over time, and \edit{the need for data and model validation} are broad challenges with big data and statistical analysis that are discussed elsewhere \edit{\citep[e.g.,][]{Lazer2014, Althouse2015}}. Here, we discuss a narrower set of challenges that arise specifically from the spatial nature of big data.
\subsection*{Spatial coverage and representation}
Spatial big data may provide precise spatial information, but careful users should question the validity of available data. For one, we know that sources of spatial big data have biases in usership rates and demographics by location (Figure \ref{fig:fig1}A) \edit{\cite{Hecht2014}}. Medical claims record data only from insured and care-seeking populations, which may vary systematically according to socioeconomic and demographic characteristics. Social media sites where users volunteer spatial data tend to have more users and higher quality information per capita in urban areas compared to rural ones \cite{Hecht2014}. Mobile phone \edit{ownership} varies by gender and literacy, and phone sharing between multiple individuals and SIM card switching complicate comparisons of these data across locations \edit{\cite{Wesolowski2012a, Wesolowski2015a}}. \edit{As we cannot often measure the heterogeneities in user populations, these heterogeneities can translate into poor choices in sampling design (e.g., how to stratify samples to get a representative population). Beyond heterogeneities in user populations, the populations captured by big data (e.g., Twitter users) are not usually relevant to epidemiology; even if we could generate an unbiased sample of the population, it may not provide information important to public health. All of these issues complicate} analyses that seek to compare different locations. Ultimately, issues with spatial coverage and representation cause problems for statistical inference, which often depend on assumptions of independent random variation and representative sampling for validity. Future research should compare analyses of spatial big data and analyses of designed observational data in order to demonstrate the \edit{validity} of spatial big data samples and to understand which features of a big data sample can produce \edit{robust} statistical inference.
\subsection*{Spatial uncertainty and noise}
Each source of big data provides a different type of spatial insight, despite their shared feature of high spatial resolution. Users of social media volunteer their geographical locations in their profiles or \edit{posts}, while internet search engines can log spatial information automatically every time a web search is performed. Sometimes the data are tied to a static location, as in the case of medical claims and health care facilities, but the cell towers associated with call data records and the locations of geo-tagged tweets vary dynamically over time (Figure \ref{fig:fig1}B). Across the combinations of features \textemdash self-reported or automated, static or dynamic \textemdash among these data sources, there are additional layers of uncertainty to consider in the context of epidemiology. \edit{For one, when spatial information in big data is not clearly specified, systematic biases in the results may be generated from the data cleaning process itself (e.g., addresses may be less likely to be geolocated in rural areas) \citep[e.g.,][]{Skelly2002}}. Second, locations of potential transmission events will often differ from locations where disease is reported. While these components are explicitly differentiated in medical claims data (i.e., transmitted in the community, reported at health care facilities), social media posts affiliated with dynamic movements could provide undifferentiated information about both transmission and reporting event locations. Big data provides information at unprecedented levels of spatial precision, but the spatial information fundamental to infectious disease epidemiology (e.g., location and conditions that caused a disease transmission event) continues to remain obscured. As big data becomes more prevalent in epidemiological analysis, public health officials should take care not to conflate spatial precision with spatial accuracy in statistical inference for disease transmission and control.
\subsection*{Spatial scales and misalignment}
When spatial big data are available at the level of individuals or precise spatial coordinates, practitioners may need to choose the scale of analysis and aggregate data accordingly. Analyzing data at the individual scale is prone to overfitting and the ``atomistic fallacy,'' in which we may make incorrect inferences at the group or population-level based on relationships observed in individual-level data \edit{\cite{Lawson2006}}. For example, if we observe an association between body mass index (BMI) and hospitalization for influenza \edit{among individuals}, it may be incorrect to assume that populations with a high average BMI would have higher rates of influenza-associated hospitalization. On the other hand, analyzing data at aggregated scales is prone to the ``ecological fallacy,'' where inferences about individuals are derived falsely from population-level observations \edit{\cite{Gotway2002, Lawson2006}}. \edit{As an example}, if we observed a negative association between average income and cholera prevalence at a national scale, it would be erroneous to assume that poor individuals have higher risk of cholera than wealthy individuals. Similarly, statistical relationships between predictors and disease outcomes may change when analyses are performed at different spatial aggregations. For instance, Google Flu Trends \edit{attempted} to estimate influenza activity across different regions of the United States by \edit{modeling the relationship between Google search terms and visits for influenza-like illness (ILI)}, as reported in traditional flu surveillance systems \cite{Ginsberg2009}. However, the set of search terms identified as \edit{``most predictive'' of ILI activity were tuned to specific spatial scales}, which means that \edit{terms that fit well at the region-level} may not apply to finer resolution data \edit{\cite{Ginsberg2009, Olson2013}}. Additionally, spatial questions often require the use of multiple data sources, and spatial misalignment arises when data are collected at different spatial scales and need to be incorporated into a single analysis. For instance, we may seek to understand the spatial distribution of cases at the state level when data were collected at the parish or county level (switching between two areal scales), or translate case data associated with household coordinates to cases at the county level (switching between point and areal scales) (Figure \ref{fig:fig1}C). Spatial big data has expanded the types of spatial information available for data aggregation \textemdash posts geo-tagged on social media might provide information at the level of countries, cities, neighborhoods, landmarks, and latitude-longitude coordinates \textemdash potentially engaging statistical change of support problems, even for one individual in a single day \cite{Gotway2002}. The multiplicity of highly resolved spatial scales also poses concerns for standard data checks, since traditional public health data will not necessarily be available at scales appropriate for validating comparisons to spatial big data \cite{Viboud2014, Shaman2013}. Finally, choices about how to deal with spatial misalignment have consequences for modeling results. For instance, recent studies have asked whether Zika virus-associated microcephaly was occurring at unusually high rates in different Brazilian states. Birth rate data might be collected at one spatial scale according to regular demographic surveys, but data systems tracking microcephalic live births would likely have finer spatial detail. Depending on the choice of spatial scale, the combination of these two data sources creates the potential for both over and under estimation of microcephaly rates.
\subsection*{Spatial confidentiality and ethics}
The practice of collecting data without seeking appropriate ethical approval presents some risk for digital infectious disease epidemiology, and the access to fine-grain spatial information further deepens this concern. \edit{Safeguards currently implemented for collecting and sharing spatial big data have focused on the obfuscation and aggregation of shared data to protect privacy, and the anonymization and de-identification of individuals.} \edit{Many research institutions have standardized practices to protect individual privacy that follow the guidance of institutional review boards (IRBs), disclosure review boards for public use data, and federal laws (e.g., the Health Insurance Portability and Accountability Act of 1996 (HIPAA) in the United States), but these organizations do not often recognize high resolution spatial data as a source that should be covered under human subjects protection policies \cite{NRCPanel2007}.} Several studies have provided examples where seemingly anonymized data could be mined (or linked with other databases) for de-anonymization: de Montjoye et al. \cite{DeMontjoye2013a} showed that four spatio-temporal position points from mobile phone records can be sufficient to uniquely identify 95\% of individuals in a large de-identified dataset; and Homer et al. \cite{Homer2008} showed that the sheer quantity of data collected could be sufficient to re-identify individuals in a genetic database. These issues already push the limits of existing ethical review mechanisms and our understanding of de-anonymization. In the future, guidelines to protect privacy and confidentiality may require: the masking of individual-level records through the aggregation of data to coarser spatial resolutions (Figure \ref{fig:fig1}D), the provision of synthetic datasets that attempt to mimic underlying distributions \cite{Kinney2011}, or the distillation of spatial big data to parameters commonly used in epidemiological models. Investigations may consider the optimal choice of spatial scale in the context of trade-offs between the accurate representation of process heterogeneity and the protection of privacy \cite{NRCPanel2007} and the improvement of computational efficiency \cite{Deeth2016}. \edit{Nevertheless, public data become increasingly vulnerable to breaches of privacy as additional data are released and data mining techniques improve over time}.
\section*{Implications for public health communication and policy}
The promise of high spatial and temporal resolutions in spatial big data opens opportunities for change in the standard practice of public health. In circumstances where adjacent or subordinate administrative units issue separate public health recommendations (e.g., United States federal, state, and local governments may issue independent flu vaccination recommendations), spatial big data may enable these entities to derive their policies from analyses of a common dataset and encourage coordination of preparedness activities \edit{across scales} \citep[e.g.,][]{Lee2015}. \edit{There is a growing panoply of adaptive, behavioral, and health economic modeling methods aimed at identifying the most effective interventions for human and livestock diseases.} As these methods begin to find use during ongoing outbreaks, the combination of spatial big data and adaptive models could enable the real-time adaptive management of infectious diseases and the coordination of disease control efforts across spatial scales.
In the long term, \edit{some sources of} big data \edit{may} become more readily available at finer spatial resolutions than the administrative regions at which policy decisions are made, \edit{even to the level} of the individual. Spatial big data has already changed consumer marketing strategies; rather than targeting geographic areas with certain socioeconomic and behavioral characteristics, marketers can now target individual users based on the behaviors demonstrated in their digital traces \cite{Dalton2015}. Should epidemiological modeling and design reflect these cultural changes to public health data? Perhaps an analogous scenario would see individual epidemiological data being used to inform optimal intervention strategies, \edit{ignoring the administrative boundaries that typically constrain decision making.} It is difficult to imagine how such a public health infrastructure could operate \textemdash resources must still be coordinated and expended by administrative units, and policy decisions must still apply to populations (rather than individuals) to maintain feasibility. Nevertheless, epidemiological analyses with spatial big data expand the possibilities for multi-scale coordination of infectious disease surveillance, response, and forecasting.
The real-time high volume nature of spatial big data makes more epidemiological information readily available to policymakers, but it also creates challenges for the communication of public health information. Spatial big data enables small area analyses, which are simultaneously highly precise to spatial locations and highly uncertain in modeling results about risk of disease. Similarly, the rise of epidemic forecasting technologies based on spatial big data might present predictions about risk and epidemic outcomes in precise locations even though the forecasts themselves are subject to uncertainty \cite{Shaman2013}. Consumers of analyses derived from spatial big data \textemdash clinicians, public health officials, epidemiologists, modelers \textemdash should develop conscientious practices for communicating uncertainty about spatial results to the public.
\clearpage
\begin{figure}[ht!]
\caption{\textbf{A)} Spatial big data have spatial biases in the populations they represent. For instance, \edit{as reported by the 2013 American Community Survey,} there is spatial variation in home internet access across the United States, which might affect the populations \edit{generating search query data in Google Trends}. \textbf{B)} With static spatial data (depicted left), individuals (represented with different colors) report case events (points) at fixed locations. For instance, the \edit{two} individuals each visited the same doctor's office with symptoms multiple times (points along the `time' axis), so their events are recorded at the same position along the `space' axis \edit{(see overlapping trajectories in the lower part of the `space' axis)}, while the \edit{teal} individual visited a different doctor's office with symptoms three times in a similar time period. Events from the same individual are connected with a dashed line. With dynamic spatial data (depicted right), events are recorded as individuals move through space. For example, the \edit{dark blue} individual \edit{(see trajectory that begins earliest on the `time' axis)} recorded four events when they tweeted about symptoms at work, at the grocery store, at the pharmacy, and at home, so their case events occur at four different positions along the `space axis. Events occur in time dynamically (as show in this figure), but events may also be aggregated to regular time intervals (e.g., weekly). \textbf{C)} Data at different spatial scales may have different magnitudes and variability in time after adjusting for population size, even if they are derived from the same data source. For instance, we observe time-varying fluctuations and variation in epidemic peak timing and magnitude in the county-level disease data (grey) that are lost in the state-level data (black). \textbf{D)} One possible method to protect privacy is to mask individual-level data by aggregating collected data to larger spatial resolutions. In reality, individuals (black circles) may be connected to other individuals through mobile phone calls (black lines). The publicly released data may be aggregated to the level of neighborhoods (green circles), and the number of calls between individuals from different neighborhoods (green lines) would be represented with different weights (here, depicted with varying thickness according to number of individual calls).}
\centering
\includegraphics[width=0.75\textwidth]{Fig1.pdf}
\label{fig:fig1}
\end{figure}
\clearpage
\section*{Footnote page}
\subsection*{Funding}
This work was supported by the Jayne Koskinas Ted Giovanis Foundation for Health and Policy (JKTG) [dissertation support grant to ECL]; the National Cancer Institute at the National Institutes of Health [grant number R01CA172805 to ABL]; the Fogarty International Center at the National Institutes of Health; and the Research and Policy for Infectious Disease Dynamics program of the Science and Technology Directorate at the Department of Homeland Security. The opinions, findings, and conclusions or recommendations expressed in this material are those of the author and not necessarily those of JKTG, its directors, officers, or staff.
\subsection*{Acknowledgments}
The authors thank Shashank Khandelwal and two anonymous reviewers for their careful comments on earlier drafts of this work.
\subsection*{Conflicts of interest}
The authors declare no conflicts of interest.
\subsection*{Corresponding author}
Shweta Bansal\\
Assistant Professor, Department of Biology\\
Georgetown University\\
<EMAIL>; 202-687-9256\\
Elizabeth C. Lee (alternate)\\
Georgetown University\\
<EMAIL>
\clearpage
\renewcommand\refname{References}
\bibliographystyle{vancouver}
|
\section{Introduction}
Image restoration is one of the most fundamental and important problems in low-level image processing, which is the operation to recover (as good as possible) the clean image $u:\Omega\rightarrow \mathbb{R}$, $\Omega\subset\mathbb{R}^2$, from a contaminated image $f:\Omega\rightarrow \mathbb{R}$ as
\begin{eqnarray*}
f=Ku+\eta,
\end{eqnarray*}
where $K$ is linear degraded operator (blur operator) and $\eta$ is an additive noise. An ideal restored model is expected to enhance image by reducing degradations in the image acquisition process and preserving edges as much as possible. However, it is often difficult to simultaneously remove the noise and enhance edges because both the noise and edges are high frequency signals.
During the past several decades, the models based on variational partial differential equation (PDE) have been attracted much attention such as TV-based models \cite{4, 5,10} and nonlocal-based models \cite{26,27} and also obtained some satisfactory results \cite{1,2,3}. Different to aforementioned models, which are developed based on the image domain, the authors in \cite{28,29,38} proposed to consider the image as an embedded surface $\mathcal{M}\in\mathbb{R}^3$ denoted by
\begin{eqnarray*}
\Omega\rightarrow \mathcal{M}:\mathrm{x} \rightarrow\mathcal{U}(\mathrm{x}),
\end{eqnarray*}
where $\mathrm{x}:=(x_1,x_2)$ denotes the local coordinates of the surface and $\mathcal{U}(x_1,x_2):=\big(x_1,x_2,u(x_1,x_2)\big)$. Note that $\Omega$ and $\mathcal{M}$ are viewed as Riemannian manifold equipped with suitable metrics. By introducing metrics $\mathrm{d}^2s\alpha\mathrm{d}^2x_1+\alpha\mathrm{d}^2x_2$ on $\Omega$ and $\mathrm{d}^2\tilde{s}=\alpha\mathrm{d}^2x_1+\alpha\mathrm{d}^2x_1+\mathrm{d}^2u$ on $\mathcal{M}$,
we can obtain
\begin{eqnarray}\label{11}
\mathrm{d^2\tilde{s}}=(\mathrm{d}x_1,\mathrm{d}x_2)
\left[\begin{array}{cc}\alpha+u_{x_1}^2 & u_{x_1}u_{x_2}\\u_{x_1}u_{x_2}&\alpha+u_{x_2}^2
\end{array}\right]
\left[\begin{array}{cc}\mathrm{d}x_1\\ \mathrm{d}x_2
\end{array}\right],
\end{eqnarray}
where $\alpha>0$ is a shrinkaging parameter for the local coordinates $(\mathrm{d}x_1,\mathrm{d}x_2)$ and $\mathrm{d^2(\cdot)}$ denotes $\big(\mathrm{d(\cdot)}\big)^2$ with the convention. In order to obtain a restored approximation $u$ from $f$, we need to search for $\mathcal{M}$ with the minimal area. In this way, singularities are smoothed. Let $g$ denote the determinant of the second-order square matrix in (\ref{11}). We consider to minimize $u$ as follows
\begin{eqnarray}\label{12}
\min_{u}~~\mathcal{J}_{\alpha}(\nabla u):=\int_{\Omega}\sqrt{g}\mathrm{dx}=\sqrt{\alpha}\int_{\Omega}\sqrt{\alpha+|\nabla u|^2}\mathrm{dx}.
\end{eqnarray}
The Euler-Lagrange equation of (\ref{12}) gives
\begin{eqnarray}\label{13}
E_{\alpha}(u)=0,
\end{eqnarray}
where $$E_{\alpha}(u):=\mathrm{div}\left(\frac{\nabla u}{\sqrt{|\nabla
u|^2+\alpha}}\right).$$
It is clear that the mean curvature of $\mathcal{M}$ is zero when $\alpha=1$. Surfaces of zero mean curvature are known as minimal surfaces. Thus, we can solve (\ref{13}) by embedding it into the following dynamical scheme
\begin{eqnarray*}
\frac{\mathrm{d\mathcal{X}}}{\mathrm{dt}}(t)=E_{\alpha}(u),
\end{eqnarray*}
where $\mathcal{X}(t)=(x_1,x_2,u(t,x_1,x_2))$. However, this scheme only considers how to regularize the image while ignores to preserve the image features. Therefore, we propose a novel model by introducing a data fitting term as follows
\begin{eqnarray}\label{14}
\min_{u}~~\frac{\lambda}{2}\left\|Ku-f\right\|_2^2+\int_{\Omega}\sqrt{\alpha+|\nabla u|^2}\mathrm{dx},
\end{eqnarray}
where $\lambda>0$ is a positive parameter and $\|\cdot\|_2$ denotes the $L^2$-norm.
The proposed model (\ref{14}) closely relates to the classic ROF model proposed by Rudin, Osher and Fatemi (ROF model) \cite{4} when $\alpha=0$. On the other hand, when $\alpha>0$ as in our model (\ref{14}), extra smoothness is introduced to the Total Variation (TV). We can employ similar numerical methods to solve the smoothing ROF method (\ref{14}) as the classical ROF model including the time marching scheme \cite{4} and the fixed point iteration scheme \cite{8}. These methods are usually restricted to the Courant-Friedrichs-Lewy (CFL) condition and the data scale of the operator inversion. In this paper, we propose a primal-dual method to solve the model (\ref{14}). To the best of our knowledge, this method has not been used to solve the model (\ref{14}) although it has been verified on some non-smoothing models in the field of image processing and machine learning. Moreover, we firstly employ the Legendre-Fenchel transformation to reformulate the minimization problem (\ref{14}) as a saddle-point problem and use the alternative updating scheme to solve the primal and dual variables. The proposed primal-dual algorithm can help to avoid the difficulties when working solely with the primal variable or dual variable \cite{4,7}. We use numerical experiments to demonstrate the proposed primal-dual algorithm can achieve the solution in a reasonable time.
The contents of the paper are arranged as follows. In section 2, we give some preliminaries of the primal-dual method and use it to solve the proposed model. Some numerical comparisons are done between the proposed method and other classic numerical methods in section 3. We give the concluding remarks in section 4.
\section{The basic results}
Firstly, we define the convex conjugation as \cite{15} in the following way
\begin{eqnarray}\label{21}
H^*(s)=\sup_{t}\left\{\langle s,t\rangle-H(t)\right\}
\end{eqnarray}
for a function $H:\mathcal{Z}\rightarrow\mathds{R}\cup\{\infty\}$ in order to obtain the saddle-point problem from the model (\ref{14}). Here, $\mathcal{Z}$ denotes a finite-dimensional Hilbert space. Throughout the rest of the paper, we assume the images are matrices with the size of $N\times N$ and with the periodic boundary condition. Let us define the Euclidean space $X=R^{N\times N}$ and $Y=X\times X$. The usual scalar products can be denoted as $\langle \textbf{v},\textbf{w}\rangle_X:=\displaystyle\sum^n_{i=1}\sum^n_{j=1}\textbf{v}_{i,j}\textbf{w}_{i,j}$ with the norm $\|\textbf{v}\|_X=\sqrt{\langle \textbf{v},\textbf{v}\rangle_X}$ for $\textbf{v},\textbf{w}\in X$ and $\langle \textbf{p},\textbf{q}\rangle_Y:=\displaystyle\sum^n_{i=1}\sum^n_{j=1}\sum^2_{\iota=1}\textbf{p}^{\iota}_{i,j}
\textbf{w}^{\iota}_{i,j}$ for $\textbf{p},\textbf{q}\in Y$. In the following, we define the discrete gradient $\nabla u=(D^+_xu,D^+_yu)$ with the forward difference operators
\begin{eqnarray*}
D^+_xu_{i,j}=\left\{
\begin{array}{ll}
u_{i+1,j}-u_{i,j}\hspace{3pt}&\mbox{if}\hspace{3pt}1\leq i<n, 1\leq j\leq n,\\
u_{1,j}-u_{i,j}\hspace{3pt}&\mbox{if}\hspace{3pt}i=n, 1\leq j\leq n,
\end{array}
\right.\\
D^+_yu_{i,j}=\left\{
\begin{array}{ll}
u_{i,j+1}-u_{i,j}\hspace{3pt}&\mbox{if}\hspace{3pt}1\leq i\leq n, 1\leq j<n,\\
u_{i,1}-u_{i,j}\hspace{3pt}&\mbox{if}\hspace{3pt}1\leq i\leq n, j=n.
\end{array}
\right.
\end{eqnarray*}
We also define the backward difference operators as
\begin{eqnarray*}
D^-_xp^1_{i,j}=\left\{
\begin{array}{ll}
p^1_{i,j}-p^1_{i-1,j}\hspace{3pt}&\mbox{if}\hspace{3pt}1<i\leq n, 1\leq j\leq n,\\
p^1_{i,j}-p^1_{n,j}\hspace{3pt}&\mbox{if}\hspace{3pt}i=1, 1\leq j\leq n,
\end{array}
\right.\\
D^-_yp^2_{i,j}=\left\{
\begin{array}{ll}
p^2_{i,j}-p^2_{i,j-1}\hspace{3pt}&\mbox{if}\hspace{3pt}1\leq i\leq n, 1<j\leq n,\\
p^2_{i,j}-p^2_{i,n}\hspace{3pt}&\mbox{if}\hspace{3pt}1\leq i\leq n, j=1.
\end{array}
\right.
\end{eqnarray*}
Based on the relation $\langle u, \mathrm{div}\textbf{p}\rangle_X=-\langle\nabla u, \textbf{p}\rangle_Y$ in \cite{6}, we can obtain the divergence operator as
\begin{eqnarray*}
\mbox{div}\textbf{p}=\nabla_x^-p^1+\nabla_y^-p^2
\end{eqnarray*}
for $\textbf{p}\in Y$.
Therefore, the discrete equivalent of (\ref{14}) can be denoted by
\begin{eqnarray}\label{22}
\min_u\left\|Ku-f\right\|_X^2+\left\|\sqrt{\alpha+|\nabla u|^2}\right\|_X,
\end{eqnarray}
where $|\nabla u|^2=\left(D_x^+u\right)^2+\left(D_y^+u\right)^2$. Following from the convex conjugate (\ref{21}) (See Exam 8.5 in \cite{11}), there is
\begin{eqnarray*}
g(t)=\sqrt{\alpha+t^2}\Longleftrightarrow g(t)=\sup_{|s|\leq1}\left\{\langle t, s\rangle+\sqrt{\alpha\left(1-s^2\right)}\right\}.
\end{eqnarray*}
Thus, we can rewrite the minimization problem (\ref{22}) as a min-max problem
\begin{eqnarray}\label{23}
\min_u\max_{\|\textbf{p}\|_{\infty}\leq1}~\left\{\frac{\lambda}{2}\|Ku-f\|_X^2-\langle u,\mathrm{div}\textbf{p}\rangle_X+\left\|\sqrt{\alpha\left(1-|\textbf{p}|^2\right)}\right\|_X\right\},
\end{eqnarray}
for $\textbf{p}\in Y$.
Since the subjective function (\ref{23}) is proper convex, we can interchange the order of min and max and solve the problem using the primal-dual scheme \cite{7}. We separate (\ref{23}) into the following two subproblems.
\begin{itemize}
\item[$\bullet$]For the primal variable $u$ in (\ref{23}): By ignoring the unrelated term to $u$, we can obtain
\begin{eqnarray*}
\min_u~~\left\{\frac{\lambda}{2}\|Ku-f\|_X^2-\langle u,\mathrm{div}\textbf{p}\rangle_X\right\},
\end{eqnarray*}
with its optimality condition as
\begin{eqnarray*}
\lambda K^T(Ku-f)-\mathrm{div}\textbf{p}=0,
\end{eqnarray*}
where $T$ denotes the matrix transpose.
In general, the blurring operator matrix $K$ is ill-posed, we can use the gradient method to compute $u$
\begin{eqnarray*}
u^k-u^{k+1}=\tau\left(\lambda K^T(Ku^{k+1}-f)-\mathrm{div}\textbf{p}\right).
\end{eqnarray*}
Due to the assumption of the periodic boundary condition, we can use the Fast Fourier Transform (FFT) to solve $u$ as
\begin{eqnarray}\label{24}
u^{k+1}_{\textbf{p}}:=u^{k+1}=\mathcal{F}^{-1}\left(\frac{\mathcal{F}(u^k)+\lambda\tau \mathcal{F}(K^Tf)+\tau \mathcal{F}(\mathrm{div}\textbf{p})}{\mathcal{F}(I)+\lambda\tau\mathcal{F}(K^TK)}\right),
\end{eqnarray}
where $\mathcal{F}^{-1}$ denotes the inverse transform of $\mathcal{F}$ and $I$ is the identity matrix.
\item[$\bullet$]For the dual variable $\textbf{p}$ in (\ref{23}): By ignoring the unrelated term to $\textbf{p}$ and introducing an indicator function
\begin{eqnarray*}
\chi_{\mathcal{K}}\left(\textbf{p}\right)=\left\{
\begin{array}{ll}
0,&\mbox{if}\hspace{5pt}\textbf{p}\in\mathcal{K},\\
+\infty,&\mbox{if}\hspace{5pt}\textbf{p}\not\in\mathcal{K},
\end{array}\right.
\end{eqnarray*}
where $\mathcal{K}:=\left\{|p|_{\infty}\leq1\right\}$. Then we have
\begin{eqnarray*}
\max_{\textbf{p}}~~\left\{\langle\nabla u,\textbf{p}\rangle_Y
+\left\|\sqrt{\alpha\left(1-|\textbf{p}|^2\right)}\right\|_X+\chi_\mathcal{K}(\textbf{p})\right\}.
\end{eqnarray*}
The optimality condition of the above maximization problem is
\begin{eqnarray*}
\left(\nabla u+\partial\chi_\mathcal{K}(\textbf{p})\right)\sqrt{1-|\textbf{p}|^2}-\sqrt{\alpha}\textbf{p}=0.
\end{eqnarray*}
Note that $\partial\chi_\mathcal{K}(\textbf{p})=0$ because indicator function is a constant function. Here, $\partial$ denotes the sub-gradient defined by $\partial\hbar(\bar{y}):=\{v|\hbar(\bar{x})-\hbar(y)\geq(v,\bar{x}-\bar{y})\}$ at the point $\bar{y}$ for a function $\hbar$. Therefore, using the projection gradient method, we have
\begin{eqnarray}\label{25}
\textbf{p}^{k+1}_u:=\textbf{p}^{k+1}=\frac{\textbf{p}^k+\sigma\left(\nabla u
\sqrt{1-\left|\textbf{p}^k\right|^2}-\sqrt{\alpha}\textbf{p}^k\right)}{\max\left\{1,~
\left|\textbf{p}^k+\sigma\left(\nabla u\sqrt{1-\left|\textbf{p}^k\right|^2}-\sqrt{\alpha}\textbf{p}^k\right)\right|\right\}}.
\end{eqnarray}
\end{itemize}
\begin{theorem}\label{th21}
Assume that $\tau\sigma<1/8$ and choosing $u^1=f$ and $\overline{\textbf{p}}^k=\textbf{0}$,
then the sequence $\left\{\left(u^{k+1},\textbf{p}^{k+1}\right)\right\}$ generated by \begin{eqnarray}\label{26}
\left\{
\begin{array}{ll}
u^{k+1}=\mathcal{F}^{-1}\left(\frac{\mathcal{F}(u^k)+\lambda\tau \mathcal{F}(K^Tf)+\tau \mathcal{F}(\mathrm{div}\overline{\textbf{p}}^k)}{\mathcal{F}(I)+\lambda\tau\mathcal{F}(K^TK)}\right);\\
\overline{u}^k \hspace{7pt}= 2u^{k+1}-u^k;\\
\textbf{p}^{k+1}=\frac{\textbf{p}^k+\sigma\left(\nabla\overline{u}^k
\sqrt{1-\left|\textbf{p}^k\right|^2}-\sqrt{\alpha}\textbf{p}^k\right)}{\max\left\{1,~
\left|\textbf{p}^k+\sigma\left(\nabla \overline{u}^k\sqrt{1-\left|\textbf{p}^k\right|^2}-\sqrt{\alpha}\textbf{p}^k\right)\right|\right\}};\\
\overline{\textbf{p}}^k\hspace{5pt}=\textbf{p}^{k+1};
\end{array}\right.
\end{eqnarray}
converges to the saddle point $\left(u^*,\textbf{p}^*\right)$ of the problem (\ref{23}). Furthermore, $u^*$ is the solution of the problem (\ref{22}).
\end{theorem}
\begin{remark}
For Theorem \ref{th21}, if we set $F(u)=\frac{\lambda}{2}\|Ku-f\|_X^2$ and $G(\nabla u)=\left\|\sqrt{|\nabla u|^2+\alpha}\right\|_X$, the iteration scheme (\ref{26}) is the exact Algorithm 1 used in \cite{7}. Furthermore, the convergence can be kept since $\|\nabla\|_Y\leq 1/8$ (See Theorem 1 in \cite{7}). Note that the operator norm $\|\cdot\|_Y$ is defined as $\|A\|_Y=\max \big\{\|A\textbf{z}\|_Y,\hspace{3pt}\textbf{z}\in Y~ \mbox{with}~ \|\textbf{z}\|_Y\leq1 \big\}$.
\end{remark}
\section{Numerical implementations}
In numerical implementations, we consider to use the proposed model (\ref{14}) for the basic image restoration problems, i.e., image denoising and deblurring. In fact, the model (\ref{14}) has many other applications, for example the CT or MRI medical image reconstruction problems and image inpainting problems with different operators $K$, etc.. In order to demonstrate the advantage of the proposed Primal-Dual Method (PDM), we compare it with another two classic numerical methods, which are the Time Marching Method (TMM) \cite{4} and the Fixed Point Method (FPM) \cite{5}. All the algorithms will stop when $\max\left\{\frac{\left\|u^{k+1}-u^{k}\right\|_X}{\left\|u^k\right\|_X}, \frac{\left|E(u^{k+1})-E(u^{k})\right|}{\left|E(u^k)\right|}\right\}\leq10^{-5}$ or the iteration arrives to 500. The simulations are preformed in Matlab 7.14(R20014a) on a PC with an Intel Core i5 M520 at 2.40 GHz and 4 GB of memory.
We use the ``Lena'' image of different sizes, i.e., $128\times128$, $256\times256$, $512\times512$ and another four images of size $256\times256$ in the numerical experiments, which are shown in Figure \ref{fig1}. To standardize the discussions, we first normalize the pixel values of the test image $\bar{f}$ to [0,255] by using the linear-stretch formula as $f=255\times\left. (\bar{f}-min(\bar{f}))\middle/(max(\bar{f})-min(\bar{f}))\right.$, where $max$ and $min$ represent the maximum and minimum of $\bar{f}$, respectively.
\begin{figure*}[ht!]
\centering
\begin{minipage}[htbp]{0.18\linewidth}
\centering
\includegraphics[width=1.2in]{lena.eps}
{(a) Lena}
\end{minipage}~
\begin{minipage}[htbp]{0.18\linewidth}
\centering
\includegraphics[width=1.2in]{hou.eps}
{(b) House}
\end{minipage}~
\begin{minipage}[htbp]{0.18\linewidth}
\centering
\includegraphics[width=1.2in]{cam.eps}
{(c) Cameraman}
\end{minipage}~
\begin{minipage}[htbp]{0.18\linewidth}
\centering
\includegraphics[width=1.2in]{pep.eps}
{(d) Peppers}
\end{minipage}~
\begin{minipage}[htbp]{0.18\linewidth}
\centering
\includegraphics[width=1.2in]{ba1.eps}
{(e) Barchettas}
\end{minipage}\\
\caption{\label{fig1}Test images in numerical implementations.}
\end{figure*}
For the sake of simplicity, we use $\sigma$ to denote standard deviation of the white Gaussian noise and $G(hsize,\sigma)$ denotes the symmetric Gaussian low-pass filter of size $hsize$ with standard deviation $\sigma$. In order to compare the visual perception and quality metric point of view, the performance of each method is evaluated in terms of signal to noise ratio ($SNR$) and structural similarity index ($SSIM$): the higher $SNR$ and $SSIM$ the better the restoration results. In addition, we explicitly give the update scheme of the TMM \cite{4} and the FPM \cite{11,8} as follows
\begin{eqnarray}
&&\frac{u^{k+1}-u^k}{dt}=\left(\lambda K^T(Ku^k-f)-E_{\alpha}(u^k)\right)\hspace{10pt}(\rightarrow TMM)\\
&&(\lambda K^TK+E_{\alpha}(u^k))u^{k+1}=\lambda K^Tf \hspace{45pt}(\rightarrow FPM)
\end{eqnarray}
by choosing a suitable original value $u^0$.
We set $\alpha=10^{-2}$ for all of numerical implementations. Note the matrix operator in the left of the FPM is symmetric and positive definite. Therefore, we employ the conjugate gradient method to solve it as \cite{11,8}.
We first analyze the performance of our PDM by comparing it with the TMM and FPM for restoration of Lena images with different sizes. Here, the blur and noisy images are corrupted by additive white Gaussian noise with $\sigma=10$ and the Gaussian blur with $G(21,0.6)$. Table \ref{tab21} illustrates the values of $SNR$, $SSIM$ and CPU time when the numerical algorithms stop. We can observe that $SNR$ and CPU time increases while $SSIM$ decreases when the image size increases. Besides, large parameter $\lambda$ is required to penalize the data fitting term when we increase the size of the test image without increasing the level of noises. From the comparisons of $SNR$, $SSIM$, $Time$ and $Ite$, our PDM is shown better than both the TMM and FPM, especially the CPU time. In fact, the TMM needs more iterations to obtain the steady solution and the FPM needs to solve the linear equation by the numerical methods such as the conjugated gradient method with inner iterations. All of these mean that our proposed PDM is more suitable to deal with the large scale image than the other two methods.
\begin{table}[htbp]
\centering
\footnotesize{\begin{tabular}{|c|c|c|c||c|c|c||c|c|c|}
\hline {Image}&\multicolumn{9}{c|} {Lena image contaminated by noise with $\sigma=10$.} \\
\hline {(S,$\lambda$)}&\multicolumn{3}{c||} {$(128\times128,0.14)$}&\multicolumn{3}{c||}{$(256\times256,0.12)$} &\multicolumn{3}{c|}{$(512\times512,0.11)$} \\
\hline{Method}&TMM&FPM&PDM&TMM&FPM&PDM&TMM&FPM&PDM\\
\hline SNR&17.8059&17.7039&\textbf{18.1492} &18.8162&18.7139&\textbf{19.2730}&19.9109&19.8497&\textbf{20.3225}\\
\hline SSIM&0.8229&0.8397&\textbf{0.8421} &0.7271&0.7272&\textbf{0.7464} &0.6230&0.5907&\textbf{0.6277}\\
\hline Time(s)&1.6224&4.3056&\textbf{0.6864} &8.1433&23.7746&\textbf{3.0108} &68.6716&126.0488&\textbf{14.8981}\\
\hline Ite&59&\textbf{16}&32 &79&\textbf{18}&36 &138&\textbf{19}&38\\
\hline
\hline {Image}&\multicolumn{9}{c|} {Lena image contaminated by noise with $\sigma=10$ and Gaussian blur with $G(21.0.6).$}\\
\hline {(S,$\lambda$)}&\multicolumn{3}{c||} {$(128\times128, 0.3)$}&\multicolumn{3}{c||}{$(256\times256,0.25)$} &\multicolumn{3}{c|}{$(512\times512, 0.19)$} \\
\hline {Method}&TMM&FPM&PDM&TMM&FPM&PDM&TMM&FPM&PDM\\
\hline {SNR} &14.8131&15.3534&\textbf{15.3787}&15.6258&\textbf{16.8751} &16.2984&18.2138&18.7663&\textbf{18.7726}\\
\hline {SSIM}&0.7568 &0.7919 &\textbf{0.7902} &0.6160 &\textbf{0.6903} &0.6514 &0.5595&0.5663&\textbf{0.5778}\\
\hline {Time(s)}&1.7784 &23.6654&\textbf{2.6520} &4.6020 &101.1822&\textbf{2.6988} &56.6596&403.1534&\textbf{19.9057}\\
\hline {Ite}&51&\textbf{29}&91&31&\textbf{30}&21&102&\textbf{24}&37\\
\hline
\end{tabular}}
\caption{\label{tab21}The related data by restoring contaminated ``Lena'' image of different sizes. Here, $\lambda$ and "Ite" denote the regularization parameter and iteration number, respectively, and "Time" denotes the CPU time described by second. }
\end{table}
Next, we test our PDM on other degraded images, where each test image has its own specialty, i.e., ``House" has much more sharp edges, ``Cameraman" has more affine regions, and ``Peppers" is a relatively smooth image. To generate the degraded images, we add the additive white Gaussian noise and apply the Guassian convolution to the images. Similar results are obtained on these three test images, as shown in Table \ref{tab22}. It is observed that our PDM is more efficient than the other two methods.
\begin{table}[htbp]
\centering
\footnotesize{\begin{tabular}{|c|c|c|c||c|c|c||c|c|c|}
\hline {Description }&\multicolumn{9}{c|} {Restore noisy images generated by the second row of Figure \ref{fig1}} \\
\hline {Image}&\multicolumn{3}{c||}{House } &\multicolumn{3}{c||} {Cameraman }&\multicolumn{3}{c|}{Peppers } \\
\hline {$(\sigma,\lambda)$}&\multicolumn{3}{c||} {$(15,0.06)$}&\multicolumn{3}{c||}{$(18,0.06)$} &\multicolumn{3}{c|}{$(20,0.06)$} \\
\hline{Method}&TMM&FPM&PDM&TMM&FPM&PDM&TMM&FPM&PDM\\
\hline SNR&18.0276&17.5888&\textbf{18.1996} &16.4785&16.5229&\textbf{17.1333} &15.4266&15.7958&\textbf{15.7982}\\
\hline SSIM&\textbf{0.4280}&0.3641&0.4117 &0.4268&0.4091&\textbf{0.4383} &0.6083&0.6440&\textbf{0.6333}\\
\hline Time(s)&52.5255&42.4635&\textbf{6.0060} &50.0919&42.5103&\textbf{5.8812} &40.8723&38.7350&\textbf{5.6160}\\
\hline Ite&500&\textbf{21}&66 &500&\textbf{22}&68 &415&\textbf{21}&66\\
\hline
\hline {Description}&\multicolumn{9}{c|} {Restore noisy and blur images generated by the second row of Figure \ref{fig1}}\\
\hline {Image}&\multicolumn{3}{c||}{House }&\multicolumn{3}{c||} {Cameraman }&\multicolumn{3}{c|}{Peppers} \\
\hline {$(\sigma,G,\lambda)$}&\multicolumn{3}{c||} {$(15,G(21,0.6),0.07)$}&\multicolumn{3}{c||}{$(18,G(21,0.6),0.08)$} &\multicolumn{3}{c|}{$(20,G(21,0.6),0.07)$} \\
\hline {Method}&TMM&FPM&PDM&TMM&FPM&PDM&TMM&FPM&PDM\\
\hline {SNR} &\textbf{16.9536}&16.5206&16.7436 &14.4791& 14.2598& \textbf{14.5334} &18.2138&18.7663&\textbf{18.7726}\\
\hline {SSIM}&\textbf{0.3831}&0.3309&0.3624 &0.3732&0.3597&\textbf{0.3760} &0.5595&0.5663&\textbf{0.5778}\\
\hline {Time(s)}&50.9811&93.7410&\textbf{7.1136} &43.8831&97.4850&\textbf{5.7876} &56.6596&403.1534&\textbf{19.9057}\\
\hline {Ite}&461&\textbf{26}&66 &389&\textbf{26}&54 &102&\textbf{24}&37\\
\hline
\end{tabular}}
\caption{\label{tab22}The related data by restoring contaminated image ``House'', ``Cameraman'' and ``Peppers''. }
\end{table}
Finally, we test our PDM with different smoothing parameters $\alpha$ to validate the effect of $\alpha$ in (\ref{23}). We use the image ``Barchetta'', Figure \ref{fig1} (e), and generate the degraded images by the white Gaussian noise with $\sigma=10$ and the Gaussian blurring with $G(21,0.6)$. As shown in Table \ref{tab23}, we can obtain the overall best numerical results when $\alpha=0.01$. Furthermore, it is worthy to point out that both the TM and FPM can not be used when the model (\ref{23}) degenerates to the classic ROF model. We can use the PDM to solve (\ref{23}) as did in \cite{7} when $\alpha=0$ since the PDM does not depend on the smoothing of the model. By testing our PDM with different $\alpha$, we find that $\alpha=0.01$ is the best choice, which gives comparable or better results than the ROF model.
\begin{table}[htbp]
\centering
\footnotesize{\begin{tabular}{|c|c||c|c|c||c|c|c||c|c|c|}
\hline { Description}&\multicolumn{10}{c|} {Barchetta image contaminated by noise with $\sigma=10$} \\
\hline {$(\alpha,\lambda)$}&\multicolumn{1}{c||}{(0,0.19) }&\multicolumn{3}{c||}{(0.001,0.18) } &\multicolumn{3}{c||} {(0.01,0.16)}&\multicolumn{3}{c|}{(0.1,0.20) } \\
\hline{Method}&PDM&TMM&FPM&PDM&TMM&FPM&PDM&TMM&FPM&PDM\\
\hline SNR&18.6239&17.9886&18.5400&\textbf{18.6234} &18.0055&18.3863&18.6264 &18.0459&18.6002&18.6216\\
\hline SSIM&0.7537&0.7309&0.7540&\textbf{0.7540} &0.7352&0.7503&0.7537 &0.7340&0.7543&0.7537\\
\hline Time(s)&3.1563&29.2344&28.7031&\textbf{3.1719} &20&19.7031&2.4688&8.2188&14.4688&63.6406\\
\hline Ite&39&232&17&\textbf{34} &135&17&28 &52&15&500\\
\hline
\hline {Description}&\multicolumn{10}{c|} {Barchetta image contaminated by noise with $\sigma=10$ and Gaussian blurring with $G(21,0.6)$.}\\
\hline {$(\alpha,\lambda)$}&\multicolumn{1}{c||}{(0,0.33) }&\multicolumn{3}{c||}{(0.001,0.33) } &\multicolumn{3}{c||} {(0.01,0.33)}&\multicolumn{3}{c|}{(0.1,0.31) } \\
\hline {Method}&PDM&TMM&FPM&PDM&TMM&FPM&PDM&TMM&FPM&PDM\\
\hline {SNR} &15.7181&14.3768&15.7041&\textbf{15.7188} &15.2739&15.7061&15.7186 &15.3533&15.7340&15.7039\\
\hline {SSIM}&0.6722&0.6126&0.6728&\textbf{0.6723} &0.6520&0.6730&0.6715 &0.6588&0.6730&0.6732\\
\hline {Time(s)}&11.5938&6.4688&108.1719&\textbf{11.9063} &15.6250&67.4688&8.9219 &19.0313&9.1719&58.6406\\
\hline {Ite}&103&28&24&\textbf{104} &86&23&75 &117&6&500\\
\hline
\end{tabular}}
\caption{\label{tab23}The related data by restoring contaminated ``Barchettas'' images with different values of $\alpha$. }
\end{table}
\section{Conclusions}
We presented an image restored model based on the minimized surface regularization, which closely relates to the smoothing ROF model \cite{4}. By using the property of conjugate function, we first reformulate the proposed model as a min-max problem and use the primal-dual method \cite{7} to solve the optimization problem. Theoretical convexity conditions guarantee the proposed algorithm converges to a unique global minimizer. Numerical experiments demonstrate that the proposed method holds the potential for efficient and stable computation by compared to the classic time marching method (TMM) \cite{4} and the lagged diffusivity fixed point method (FPM) \cite{11,8}, especially for the large-scale image. In the future, we would like to extend the proposed method to other image processing problem such as image inpainting, reconstruction, registration and also for vector value images, etc.
\section*{Acknowledgements}
The authors acknowledge the financial support by the NSF of China (Nos.U1304601,11401170), Foundation of Henan Educational Committee of China (No.14A110018) and the National Basic Research Program of China (973 Program)(No.2015CB856003).
\bibliographystyle{elsarticle-num}
|
\section{INTRODUCTION}
The strong Goldbach's conjecture refers to Christian Goldbach's
theory which was represented by him with a letter to the great
Swiss mathematician Leonard Euler in 1742. He asked him for
solving this problem. He spent much time trying to prove it, but
never succeeded. In this theory, he claims that "every even
integer greater than two could be a sum of two prime numbers". The
word "strong" is opposition to word "weak" where refers to "every
odd number greater than five could be a sum of three prime
numbers". Since two hundreds years ago, several relevant proofs on
this subject have been conducted. Using the sieve method, a lot of
mathematicians as Brun [1] in 1920 have verified various results
who expressed � every even integer is a sum of product at most 9
primes and product at most another 9 primes". Wang [2] expressed
in 1962, "every even integer is a sum of 1 prime and product at
most 4 primes". Pan [3] obtained in 1962 that every even integer
is a sum of 1 prime and product at most 5 primes. Richert [4]
obtained same result in 1969, but product at most 3 primes instead
at most 5 primes. Chen [5]in 1973 proved same for 1 prime and
product at most 2 primes. R'enyi [6] in 1947 represented that
every even integer is a sum of 1 prime and product at most �c�
primes. Montgomery and Vaughan found an exceptional set in
Goldbach's problem by the circle method [7]. We presented [8-9]
by the two papers in 2009 and 2010 stating 5 new Functions,11
new Lemmas, a novel explicit formula for computing $\pi(x)$,
several methods, a new conjecture and some new experimental
computations to prove Goldbach�s conjecture experimentaly. C.D.
Pan and C.B. Pan in [10] made the proof of some Theorems in
analytic number Theory so that we make use of some of them here. Most of these
Theorems are about the proof of the Prime Number Theorem.
\setcounter{equation}{0}
\section{ELEMENTARY DEFINITIONS}
The following new function is defined as $n$ be a natural number
and $N$ an even positive integer,then
\begin{equation}
\eta(n)=\left\{
\begin{array}{ll}
0~~~ & \mbox{if~~~~$N\equiv n~mod(p)$~~$and~~~p\leq
\sqrt{N}$}\\\\1~~~ & \mbox{if~~~~$otherwise$}\end{array} \right.\
\end{equation}\\
where $p$ denotes a prime number and $gcd(p,N)=1$.From the
function $\eta(n)$ is recognized properties of sieve function.
Just, based on the above function and known functions as
$\zeta(s)$,$\theta(x)$,$\pi(x)$ and $\psi(x)$ one is able to
define the several new following functions (2-2) to (6-2).
\begin{equation}
\zeta(s,\eta)=\sum_{n=1}^{\infty}\frac{\eta(n)}{n^{s}}\
\end{equation}
\begin{equation}
\theta(x,\eta)=\sum_{p\leq x}\eta(p)log(p)\
\end{equation}
\begin{equation}
\pi(x,\eta)=\sum_{p\leq x}\eta(p)\
\end{equation}
\begin{equation}
\psi(x,\eta)=\sum_{n\leq x}\eta(n)\Lambda(n)\
\end{equation}
\begin{equation}
\Psi(s,\eta)=\sum_{n=1}^{\infty}\frac{\eta(n)\Lambda(n)}{n^{s}}\
\end{equation}
where
\begin{equation}
\Lambda(n)=\left\{
\begin{array}{ll}
log(p)~~~ & \mbox{if~~~~$log(n)=mlog(p)$}\\\\0 &
\mbox{if~~~$otherwise$}\end{array} \right.\
\end{equation}
\begin{equation}
\theta(x)=\sum_{p\leq x}log(p)\
\end{equation}
\begin{equation}
\pi(x)=\sum_{p\leq x}1\
\end{equation}
\begin{equation}
\psi(x)=\sum_{n\leq x}\Lambda(n)\
\end{equation}
The main scope of this paper is to show that $\pi(N,\eta)\geq 1$ where
$N$ denotes a sufficient large even integer.This proves
the strong Goldbach's conjecture for sufficiently even
integers.\\\\
\textbf{Lemma 2.1}\\\\ \textit{If $\pi(N,\eta)\geq 1$, then the strong Goldbach's conjecture holds}\\\\
\textit{\textbf{Proof:}}\\\\ let
\begin{equation} \pi(N,\eta)=\sum_{p\leq N}\eta(q)=\sum_{p\leq N}\left\{
\begin{array}{ll}
0 & \mbox{if~~~~$N\equiv q~mod(p)$~~$and~~~p\leq \sqrt{N}$}\\\\1 &
\mbox{if~~~~$otherwise$}\end{array} \right.\
\end{equation}\\
Consider a contradiction as:$\pi(N,\eta)=0$\\
~~~~If $\pi(N,\eta)=0$ then consider a typical prime as $2<q<N$,
therefore based on the above relation, one should have for any q:
$N-q<\sqrt{N}$ or $N-q>\sqrt{N}$.If $N-q<\sqrt{N}$ then $N-q=kp$
and based on the equation (1-2), one concludes that $\eta(q)=0$.If
$N-q>\sqrt{N}$, one has two cases:the first $N-q$ is prime and
based upon the equation (1-2), $\eta(q)=1$. In the second
case,$N-q$ is not a prime but at least one of its factors will be
under $\sqrt{N}$(every integer as $m$ has at least a prime factor
less than $sqrt(m)$, because otherwise product of its prime
factors will be greater than itself i.e.$m$). Hence, based on the
equation (1-2), $\eta(q)=0$.From this argument, one could conclude
if $\pi(N,\eta)\geq 1$ then the first case will be happened. This
means that $N-q$ is a prime number and the strong Goldbach's
conjecture holds. Using the Prime Number Theorem [10] could be
written:
\begin{equation}
\theta(x,\eta)=\pi(x,\eta)log(x)-\int_{2}^{x}\frac{\pi(u,\eta)}{u}du\
\end{equation}
\begin{equation}
\pi(x,\eta)=\frac{\theta(x,\eta)}{log(x)}+\int_{2}^{x}\frac{\theta(u,\eta)}{u(log(u))^{2}}du\
\end{equation}
\begin{equation}
\psi(x,\eta)=\theta(x,\eta)+O(x^{1/2})\
\end{equation} \setcounter{equation}{0}\
\section{LEMMAS AND THEOREMS TO APPROACH TO THE PROOF OF
$\pi(N,\eta)\geq 1$} In this section, we prove four Lemmas
using both Perron formula of Dirichlet�s series and the Residue
Theorem. These Lemmas are the base of a basic Theorem which will
be given in the next Section. Here, we give a theorem and
some lemmas along with their proofs.\\\\
\textbf{Lemma 3.1}\\\\ \textit{Let
$A(s)=\sum_{n=1}^{\infty}a(n)n^{-s}$,$\sigma_{a}<+\infty$ and
$H(u)$ and $B(u)$ be increasing functions so that $|a(n)|\leq
H(n)$,$n=1,2,...$ and $\sum_{n=1}^{\infty}a(n)n^{-\sigma}\leq
B(\sigma)$,$\sigma>\sigma_{a}$ for any $s_{0}=\sigma_{0}+it_{0}$
and $b_{0}\geq \sigma_{0}+b>\sigma_{a}$,$T\geq 1$. Also,$x\geq 1$,
then we have Perron formula as follows [10]:}\
\begin{eqnarray}
\sum_{n\leq
x}a(n)n^{-s_{0}}+\frac{1}{2}a(x)x^{-s_{0}}=\frac{1}{2\pi
i}\int_{b-iT}^{b+iT}A(s_{0}+s)\frac{x^{s}}{s}ds+O(\frac{x^{b}B(b+\sigma_{0})}{T})+\nonumber\\O(x^{1-\sigma_{0}}H(2x)min(1,\frac{log(x)}{T}))~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\
\end{eqnarray}
\textit{where "O" denotes a constant depending on $\sigma_{a}$ and
$b_{0}$. $\sigma_{a}$ denotes absolute convergence abscissa of
Dirichlet's series.} Refer to [10] in order to see the proof.\\\\
If $s_{0}=0$, then the relation (1-3) will be:\\\\
\begin{equation}
\sum_{n\leq x}a(n)+\frac{1}{2}a(x)=\frac{1}{2\pi
i}\int_{b-iT}^{b+iT}A(s)\frac{x^{s}}{s}ds+O(\frac{x^{b}B(b)}{T})+\\O(xH(2x)min(1,\frac{log(x)}{T}))\
\end{equation}
\textbf{Lemma 3.2}\\\\
\textit{Let $c_{1}$ be a positive constant, then $\zeta(s,\eta)$
has no zero point for $\sigma \geq
1-\frac{c_{1}}{log(|t|+2)}$ where $s=\sigma+it$.}\\\\
\textit{\textbf{Proof:}}\\\\ Let's define a new function as:
\begin{equation}
\zeta(s,\hat{\eta})=\sum_{n=1}^{\infty}\frac{(1-\eta(n))}{n^{s}}\
\end{equation}
If $\sigma \geq 1-\frac{c_{1}}{log(|t|+2)}$ then can be written:
\begin{equation}
|\zeta(s,\eta)|\leq |\zeta(s,\hat{\eta})|\
\end{equation}
Also,
\begin{equation}
|\zeta(s,\eta)+\zeta(s,\hat{\eta})|\leq |\zeta(s)|\
\end{equation}
Using Rouche's Theorem, the number of zeros of the function
$\zeta(s,\eta)$ is equal to the zeros of the function $\zeta(s)$.
Since $\zeta(s)$ has no zero in the set $\sigma \geq
1-\frac{c_{1}}{log(|t|+2)}$, therefore $\zeta(s,\eta)$ has also no
zero in the same set. This proves the Lemma.\\\\ \textbf{Lemma
3.3}\\\\ \textit{Let Dirichlet's series
$\Psi(s,\eta)=\sum_{n=1}^{\infty}\frac{\eta(n)\Lambda(n)}{n^{s}}$,
then, $\sigma_{c}= \sigma_{a}=1$ where $\sigma_{c}$ and
$\sigma_{a}$ denote the convergence and absolute convergence
abscissa series
respectively.}\\\\
\textit{\textbf{Proof:}}\\\\
Using the Prime Number Theorem for Arithmetic Progressions and
$gcd(k,N)=1$ one can write by referring to [11]:
\begin{equation}
\sum_{n\leq x,~n\equiv N
mod(k)}~\Lambda(n)=\frac{x}{\phi(k)}+O(\frac{x}{log^{H}x})\
\end{equation}
\begin{equation}
\lim_{x\rightarrow\infty}{\sum_{n\leq x,~n\equiv N
mod(k)}~\Lambda(n)}=\frac{x}{\phi(k)}\
\end{equation}
For sufficiently large number $N$:
\begin{equation}
\eta(n)=\prod_{p>\sqrt{N}, gcd(p,N)=1}(1-\frac{1}{\phi(p)})\
\end{equation}
Generally, one can write:
\begin{eqnarray}
\sum_{n\leq x}\eta(n)\Lambda(n)=\sum_{n\leq x}\prod_{p>\sqrt{N},
gcd(p,N)=1}(1-\frac{1}{\phi(p)})\Lambda(n)\
\end{eqnarray}\
and for sufficiently large number $N$ is;
\begin{eqnarray}
\sum_{n\leq x,~x\rightarrow\infty}\eta(n)\Lambda(n)=\sum_{n\leq
x,~x\rightarrow\infty}~~~\Lambda(n)\prod_{p>\sqrt{N},
gcd(p,N)=1}(1-\frac{1}{\phi(p)})=\nonumber\\ \lim_{x\rightarrow\infty}\psi(x)\left\{\prod_{p>\sqrt{N}~~,
gcd(p,N)=1}(1-\frac{1}{p-1})\right\}=~~~~~~~~~~~~~~~\nonumber\\
x\left\{\prod_{p>\sqrt{N}~~,
gcd(p,N)=1}(1-\frac{1}{p-1})\right\}~~~~~~~~~~~~~~~~~\
\end{eqnarray}\
Just,we give a conjecture namely the Sabihi's second
conjecture on the Goldbach's conjecture as below (to see our first
conjecture refer to [8]):
\\\\ \textbf{Sabihi's
Second Conjecture (SSC)}\\\\\textit{Let $\gamma$ be the Euler
constant,$p$ a prime number,$N$ a sufficiently large even
number,and Riemann Hypothesis holds,then:}
\begin{equation}
log(N)=\frac{4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))}{\prod_{p>\sqrt{N}, gcd(p,N)=1}(1-\frac{1}{p-1})}\
\end{equation}
Having the conjecture the relation (10-3) gives:
\begin{eqnarray}
\sum_{n\leq x}\eta(n)\Lambda(n)=4e^{-\gamma}\times~~~~~~~~~~~~~~~~~~~~~~\nonumber\\
\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{x}{log(N)}\
\end{eqnarray}
Assume we hold $N$ to a constant value and $x\rightarrow \infty$
then could be written:
\begin{equation}
\sigma_{c}=\lim_{x\rightarrow \infty}\frac{log|\sum_{n\leq
x}\eta(n)\Lambda(n)|}{log(x)}=1\
\end{equation}\
Therefore, $\sigma_{c}=\sigma_{a}=1$ and the Lemma is
proven.\\\\
\textbf{Lemma 3.4}\\\\ \textit{Let Dirichlet's series in
Lemma 3.3 have a pole at $s=1$,then it has a residue of the following form at same pole:}
\begin{equation}
4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{1}{log(N)}\
\end{equation}\
\textit{\textbf{Proof:}}\\\\
It is well-known
\begin{equation}
\lim_{s\rightarrow
1}(1-s)\sum_{n=1}^{\infty}\frac{\Lambda(n)}{n^{s}}=1\
\end{equation}\
On the other hand
\begin{equation}
\lim_{x\rightarrow\infty}\sum_{n\leq x,~n\equiv N
mod(k)}\frac{\Lambda(n)}{n^{s}}=\frac{1}{\phi(k)}\sum_{n=1}^{\infty}\frac{\Lambda(n)}{n^{s}}
\end{equation}\
Consequently
\begin{equation}
\lim_{x\rightarrow\infty}\sum_{n\leq x,~n\equiv N
mod(k)}\frac{\eta(n)\Lambda(n)}{n^{s}}=\sum_{n=1}^{\infty}~~\prod_{p>\sqrt{N},~gcd(p,N)=1}(1-\frac{1}{p-1})\sum_{n=1}^{\infty}\frac{\Lambda(n)}{n^{s}}
\end{equation}\
and this is equal to
\begin{equation}
4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{1}{log(N)}\sum_{n=1}^{\infty}\frac{\Lambda(n)}{n^{s}}\
\end{equation}\
Just, one could obtain the residue of the Dirichlet's series as
below:
\begin{eqnarray}
\lim_{s\rightarrow
1}(1-s)\sum_{n=1}^{\infty}\frac{\eta(n)\Lambda(n)}{n^{s}}=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\times \nonumber\\
\frac{1}{log(N)}(\lim_{s\rightarrow
1}(1-s))\sum_{n=1}^{\infty}\frac{\Lambda(n)}{n^{s}}=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\times ~~~~~~~~~~~~~~~~\nonumber\\
\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{1}{log(N)}~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\end{eqnarray}
\section{PROOF OF INEQUALITY $\pi(N,\eta)\geq 1$ FOR SUFFICIENTLY LARGE EVEN INTEGER $N$}
In this section, we prove one main theorem. On
the basis of this theorem,the inequality $\pi(N,\eta)\geq 1$ under two
conditions will be proved.The first codition is to be assumed
trueness of Riemann Hypothesis (RH) and the second is to be
assumed to hold true Sabihi's second conjecture (SSC).\\\\
\textbf{Sabihi's Theorem}\\\\
\textit{Let $N$ be a sufficiently large even integer.Let
both Sabihi's Second Conjecture (SSC) and Riemann Hypothesis (RH)
hold then:}
\begin{eqnarray}
\psi(N,\eta)=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{N}{log(N)}+\nonumber\\O(N\{H(2N)min(1,\frac{log(N)}{T})-e^{-c\sqrt{log(N)}}\})~~~~~~~~~~~~~~~~~~~~~~~\
\end{eqnarray}\
\begin{eqnarray}
\pi(N,\eta)=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{N}{log^{2}N}+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\\int_{2}^{N}\frac{du}{log^{2}u}+O(\frac{xH(2N)min(1,\frac{log(N)}{T})-\frac{Ne^{-c\sqrt{log(N)}}}{log(N)}-\frac{N^\frac{1}{2}}{log(N)}}{log(N)})+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\O(\int_{2}^{N}\frac{(e^{-c\sqrt{log(u)}}-u^\frac{-1}{2})du}{log^{2}u})~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\
\end{eqnarray}\
\textit{\textbf{Proof:}}\\\\
Let $a=1-c_{1}\frac{1}{log(T+2)},
b=1+\frac{1}{log(x)},logT=(log(x)^{\frac{1}{\alpha+1}}$ for
$0<\alpha<1$,and $H(u)\leq logu$, $B(u)\leq c_{2} log(x)$ where
$c_{2}$ is a positive constant.By applying the Lemma 3.1,the
relation (2-3)and assuming $A(s)=\sum_{n=1}^{\infty}a(n)n^{-s}$
from Lemma 3.1:
\begin{equation}
\sum_{n\leq x}a(n)=\sum_{n\leq x}\eta(n)\Lambda(n)=\psi(x,\eta)\
\end{equation}\
Let $C_{n}$ be a rectangle contour with vertices $a\pm iT$ and
$b\pm iT$ then if $T$ tends to infinity
\begin{equation}
I=\frac{1}{2\pi i}\int_{C_{n}}\Psi(s,\eta)\frac{x^{s}}{s}ds=J+K\
\end{equation}\
where $J$ denotes integral along the line joining $b-iT$ to $b+iT$
and K denotes the integral along the other three sides of
rectangle.Applying the relation (2-3)to the relation (23-4) gives
\begin{equation}
I=\frac{1}{2\pi
i}\int_{b-iT}^{b+iT}\Psi(s,\eta)\frac{x^{s}}{s}ds+O(f(x))+O(g(x))=
\end{equation}\
\begin{equation}
J=\frac{1}{2\pi i}\int_{b-iT}^{b+iT}\Psi(s,\eta)\frac{x^{s}}{s}ds\
\end{equation}\
and
\begin{equation}
K=\frac{1}{2\pi
i}\{\int_{b-iT}^{a-iT}+\int_{a-iT}^{a+iT}+\int_{a+iT}^{b+iT}\}\Psi(s,\eta)\frac{x^{s}}{s}ds\
\end{equation}\
\begin{equation}
O(f(x))=\frac{xlog(x)}{T}_{T\rightarrow \infty}=0\
\end{equation}\
\begin{eqnarray}
O(g(x))=(\int_{b-i\infty}^{a-i\infty}+\int_{a-i\infty}^{a+i\infty}+\int_{a+i\infty}^{b+i\infty})\Psi(s,\eta)\frac{x^{s}}{s}ds\leq \nonumber\\
(\int_{b-i\infty}^{a-i\infty}+\int_{a-i\infty}^{a+i\infty}+\int_{a+i\infty}^{b+i\infty})\sum_{n=1}^{\infty}\Lambda(n)n^{-s}\frac{x^{s}}{s}ds=\nonumber\\
(\int_{b-i\infty}^{a-i\infty}+\int_{a-i\infty}^{a+i\infty}+\int_{a+i\infty}^{b+i\infty})(Z(s))
\frac{x^{s}}{s}ds\nonumber\\
\ll(\int_{b-i\infty}^{a-i\infty}+\int_{a-i\infty}^{a+i\infty}+\int_{a+i\infty}^{b+i\infty})log^{2}|t|\frac{x^{s}}{s}ds\
\end{eqnarray}\
Referring to [10],one can conclude that $z(s)\ll log^{2}|t|$ since
$\sigma>1-c_{1}log^{2}|t|$. By manipulating the last right hand
term of inequality(28-4) we obtain as below:
\begin{equation}
O(g(x))=xe^{-c\sqrt{log(x)}}\
\end{equation}\
Just, residue theorem expresses that
\begin{equation}
I=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{x}{log(N)}\
\end{equation}\
The relation (2-3) gives:
\begin{equation}
\sum_{n\leq
x}a(n)=\psi(x,\eta)=I-O(f(x))-O(g(x))+O(\frac{x^{b}B(b)}{T})+O(xH(2x)min(1,\frac{log(x)}{T}))\
\end{equation}\
Consequently
\begin{eqnarray}
\psi(x,\eta)=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{x}{log(N)}+\nonumber\\O(x\{H(2x)min(1,\frac{log(x)}{T})-e^{-c\sqrt{log(x)}}\})~~~~~~~~~~~~~~~~~~~~~~~\
\end{eqnarray}\
If we apply $x=N$ for a sufficiently large even integer, then the
first formula of the theorem is proven. By applying and combining
the relations(12-2)to(14-2)and(32-4)could be written:
\begin{eqnarray}
\pi(x,\eta)=\frac{\psi(x,\eta)-O(x^\frac{1}{2})}{log(x)}+\int_{2}^{x}\frac{\psi(u,\eta)-O(u^\frac{1}{2})}{ulog^{2}(u)}du=4e^{-\gamma}\prod_{p>2}(1-\frac{1}{(p-1)^2})\times~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\
\prod_{p>2~,p\mid
N}\frac{p-1}{p-2}(1+O(\frac{1}{log(N)}))\frac{x}{log(N)log(x)}+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\O(\frac{xH(2x)min(1,\frac{log(x)}{T})}{log(x)})-O(\frac{xe^{-c\sqrt{log(x)}}}{log(x)})+\int_{2}^{x}\frac{du}{log^{2}u}+O(\int_{2}^{x}\frac{e^{-c\sqrt{log(u)}}du}{log^{2}u})-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\O(\frac{x^\frac{1}{2}}{log(x)})-O(\int_{2}^{x}\frac{du}{u^\frac{1}{2}log^{2}u})~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\
\end{eqnarray}\
Again, applying $x=N$ to the relation (33-4) for a sufficiently
large even integer, the second formula of the theorem is also
proven. In the above relation, the following relation and
inequality have been applied.Applying the Prime Number Theorem:
\begin{equation}
\psi(x)=x+O(xe^{-c\sqrt{log(x)}})\
\end{equation}
and
\begin{equation}
\int_{2}^{x}\frac{\psi(u,\eta)du}{ulog^{2}u}\leq
\int_{2}^{x}\frac{\psi(u)du}{ulog^{2}u}\
\end{equation}
From the relation (33-4) and replacing $x$ by $N$ and tending $N$ to infinity, we easily see
that $\pi(N,\eta)>1$ or $\pi(N,\eta)\neq 0$ and the theorem is
proven.This proves the strong Goldbach's conjecture for
sufficiently large even integers.
|
\section{Introduction}
There is a common belief that the color confinement and the related mass gap problem
in quantum chromodynamics (QCD) need a consistent non-perturbative quantum theory
for their resolution \cite{colorconft}.
A so-called proton spin crisis \cite{spincr1,spincr2,spincr3} represents another puzzle
which is closely related to non-perturbative dynamics of constituent quarks and gluons.
Besides, recent studies reveal deep conceptual problems in definitions of the
momentum, spin and orbital angular momentum operators of quarks and gluons
(see \cite{review1} for a review and references there in).
In this respect, an important step towards a strict and
non-perturbative theory of QCD is to describe the dynamical content of classical
non-perturbative solutions in a pure $SU(2)$ Yang-Mills theory and establish
their relationship to fundamental observable quantities in QCD
such as vacuum gluon condensate, glueball spectrum and others.
Non-linear structure and rich topology of the Yang-Mills theory
lead to a wide class of various exact solutions \cite{actor}.
Monopoles and instantons represent the most well-known
topological solutions which have numerous physical implications
(see, for ex., \cite{coleman85,rajaraman}).
Much less is known about non-linear wave type solutions
and especially about their physical meaning.
Non-linear transverse plane wave solutions representing analogues of the electro-magnetic
plane waves were found in 80s by Coleman \cite{coleman}.
The existence of another type of non-linear plane wave solutions with a mass scale parameter
\cite{mat1} is an important manifestation of the conformal symmetry in the
Yang-Mills theory. Various representations of the non-linear plane waves
and some special non-linear spherical wave solutions were considered
in \cite{lahno95,tura,smilga,frasca09,tsap}.
There is a hope that the knowledge of full structure of non-linear wave solutions
in the Yang-Mills theory can provide a novel approach towards
non-perturbative description of quantum chromodynamics.
In the present paper we undertake an attempt to describe a special class
of stationary non-linear wave solutions in a pure $SU(2)$ Yang-Mills theory (or QCD)
which admit intrinsic mass scale parameter. Among such solutions
there are non-linear stationary plane wave solutions \cite{mat1}
and non-stationary spherically symmetric solutions \cite{p14,p15,p16,p17,p18} which
resemble kink type solitons in two-dimensional coordinate plane $(t,r)$.
It is known that in a pure Yang-Mills theory a stable solitonic
solutions do not exist due to the Coleman theorem \cite{coleman2}.
This implies that any wave packet solution with a finite total energy
must decay by radiating its energy to space infinity. It is surprising that a regular stationary monopole like solution
with a finite energy density everywhere does exist even in a pure $SU(2)$ QCD
without introducing any additional matter fields.
We demonstrate the existence of a class of such regular stationary
solutions which represent a system of the Wu-Yang monopole
interacting to the off-diagonal gluons in a special gauge.
All considered non-linear propagating solutions and the new proposed stationary monopole like
solutions possess intriguing properties, namely, they admit a mass scale parameter,
non-vanishing longitudinal projections of the color fields and vanished classical spin density.
This indicates to existence of massive spinless states in the quantum theory.
Such an idea, that stationary solutions may describe particles (or pseudo-particles)
and even might be related to the vacuum structure in QCD,
was sounded long time ago \cite{derr,jackiwRMP,jackiw77}. Certainly, to obtain rigorous description
of particle spectrum based on standard QCD as a fundamental theory of strong interaction one has
to construct a consistent non-perturbative quantization scheme for QCD which remains
an unresolved problem so far.
The existence of the non-linear wave type solutions
with a non-zero mass and vanished spin manifests inconsistence of the perturbative QCD
which has massless vector gluons in the initial Fock space of physical states.
So even at classical level one should formulate a strict notion of the gluon spin density
operator on the class of massive non-linear wave solutions.
A consistent definition of the gluon spin and angular momentum operators
in QCD represents unresolved problem and there is still controversy between different
approaches to this problem \cite{review1}. In particular, there is no unique gauge invariant
and at the same time explicitly Lorentz frame independent definition for the gluon spin and orbital momentum operators.
In the present paper we show that for the stationary non-linear wave solutions with the mass parameter
it is possible to construct a unique gauge-invariant and Lorentz frame independent gluon spin density operator
at classical level.
\section{Non-linear propagating solutions with a mass scale parameter}
In this section we overview briefly the known non-linear propagating wave type solutions which
possess a mass scale parameter and non-vanishing longitudinal projections of
color fields \cite{mat1,p14,p15,p16,p17,p18}.
\subsection{Non-linear spinless plane waves}
We consider a class of non-linear wave solutions
admitting mass scale parameters in the
case of a pure $SU(2)$ Yang-Mills theory.
The respective Lagrangian and equations of motion
are as follows
\begin{eqnarray}
{\cal L}&=&-\displaystyle\frac{1}{4}F_{a\mu\nu}F^{a\mu\nu}, \nonumber \\
\big(D^\mu F_{\mu\nu}\big)^a &\equiv&
\partial^\mu F_{\mu\nu}^a+g\epsilon^{abc}A^{b\mu} F^c_{\mu\nu} =0,
\label{Lagr0}
\end{eqnarray}
where the field strength is given by
$$
F^a_{\mu\nu}
=\partial_\mu A_\nu^a-\partial_\nu A_\mu^a
+g\epsilon^{abc}A^b_\mu A^c_\nu
$$
with the group structure constants $\epsilon^{abc}$ and the coupling
constant $g$.
The known Coleman non-linear transverse plane wave solutions \cite{coleman}
form a family with six arbitrary functions depending
on light-cone coordinates which correspond to six transverse
propagating massless gluon modes in agreement with perturbative
description of gluons in QCD.
An important feature of the non-Abelian gauge theory is that
it admits a three-parametric family of non-linear massive wave solutions
with a field strength having non-vanishing longitudinal projections
along the wave vector \cite{lahno95,tura,smilga,frasca09}.
Such plane waves can be interpreted as non-perturbative
longitudinal modes of gluon. The solutions can be reproduced
by using a simple ansatz
\begin{eqnarray}
A_i^a(x)&=&\delta_i^a \phi_i(u),~~~~~A_0^3(x)=\beta \phi_3(u), \label{ansf1234}
\end{eqnarray}
where $u\equiv k_0 t+ k_3 z$ and $\beta=v/c$ is a kinematic
variable proportional to wave velocity ``$v$'' in units of the
speed of light
``$c$''.
The non-vanishing field strength components are the following:
\renewcommand{\arraystretch}{1.4}
\begin{equation}
\begin{array}{rcl}
F^1_{10}&=&-\partial_0 \phi_1,~~~~~~~~F^1_{20}=g\beta \phi_2\phi_3, \\
F^2_{10}&=&-g\beta \phi_1\phi_3,~~~~F^2_{20}=-\partial_0 \phi_2, \\
F^3_{30}&=&-(1-\beta^2)\partial_0\phi_3, \\
F^1_{13}&=& -\beta \partial_0 \phi_1,~~~~~~F^2_{13}=-g\phi_1 \phi_3, \\
F^1_{23}&=&g \phi_2 \phi_3 ,~~~~~~~~F^2_{23}=-\beta\partial_0 \phi_2, \\
F^3_{12}&=& g\phi_1 \phi_2.
\end{array}
\label{fieldstr}
\end{equation}
\noindent
One can verify that ansatz (\ref{ansf1234}) leads to
non-vanishing electric and magnetic longitudinal projections.
By direct substitution of the ansatz into the Yang-Mills equations
and imposing a constraint $k_3=\beta k_0$ one can
reduce all equations of motion to three ordinary differential equations (ODE)
\begin{eqnarray}
&& k^2 \displaystyle\frac{{\rm d}^2\phi_1}{{\rm d}u^2}+g^2 \phi_1
\Big(\phi_2^2+(1-\beta^2)\phi_3^2\Big)=0, \nonumber \\
&& k^2\displaystyle\frac{{\rm d}^2\phi_2}{{\rm d}u^2}
+g^2 \phi_2 \Big(\phi_1^2+(1-\beta^2)\phi_3^2\Big)=0, \label{eqbeta} \\
&& k^2\displaystyle\frac{{\rm d}^2\phi_3}{{\rm d}u^2} +g^2 \phi_3
\Big(\phi_1^2+\phi_2^2\Big)=0, \nonumber
\end{eqnarray}
where $k^2 \equiv k_0^2-k_3^2$. Note that
the constraint $k_3=\beta k_0$ provides
a Lorenz gauge condition $\partial^\mu A_\mu^a=0$ for the
gauge potential.
In the rest frame, $\beta=0$,
the equations (\ref{eqbeta}) describe
a classical mechanical system of three anharmonic oscillators
with the following Hamiltonian
\begin{eqnarray}
H&=&\displaystyle\frac{1}{2}\big(\dot \phi_1^2+\dot \phi_2^2+\dot \phi_3^2\big)
+\displaystyle\frac{g^2}{2} \big(\phi_1^2\phi_2^2+\phi_2^2\phi_3^2+\phi_3^2\phi_1^2\big).
\quad \label{Ham}
\end{eqnarray}
In general, such a system represents a
non-integrable problem \cite{mat-savv}.
It has been proved as well that
the corresponding quantum mechanical system possesses a
pure discrete energy spectrum despite the presence of zero energy valleys
in the Hamiltonian \cite{simon83}.
Let us consider two special cases when the system of equations
(\ref{eqbeta}) becomes integrable.
The first case, (I), corresponds to a constraint
$\phi_1=\phi_2\equiv \phi$, $\phi_3=0$, and, the second case, (II),
is determined by setting $\phi_1=\phi_2=\phi_3\equiv \phi$ and imposing an additional
condition $\beta=0$.
Corresponding equations are the following
\renewcommand{\arraystretch}{2.5}
\begin{equation}
\begin{array}{rcl}
{\rm (I)}:&~&\displaystyle
k^2 \frac{{\rm d}^2\phi(u)}{{\rm d}u^2}+g^2 \phi^3(u)=0, \\
{\rm (II)}:&~&\displaystyle
k^2 \frac{{\rm d}^2\phi(u)}{{\rm d}u^2}+2 g^2
\phi^3(u)=0.
\end{array}
\label{eq1}
\end{equation}
\noindent
Various representations of these equations
have been obtained in $SU(2)$ Yang-Mills theory
by using different methods \cite{tura,lahno95,smilga,frasca09,mat-savv}.
Solutions to the equation (\ref{eq1}) are given by the Jacobi elliptic
function
\renewcommand{\arraystretch}{2.0}
\begin{equation}
\begin{array}{rcl}
\phi_{\rm I}(u)&=&\displaystyle\frac{\sqrt 2 \mu_1}{\sqrt g}\, {\rm sn}
\Big[\displaystyle\frac{\mu_1}{k} \sqrt g (u+u_{01}),-1\Big], \\
\phi_{\rm II}(u)&=&\displaystyle\frac{\mu_2}{\sqrt {g}}\, {\rm sn}
\Big[\displaystyle\frac{\mu_2}{k} \sqrt g (u+u_{02}),-1\Big],
\end{array}
\label{phi0}
\end{equation}
\noindent
where $\mu_i, u_{0i}$ are integration constants. One can set $u_{0i}=0$ since $u_{0i}$
is the parameter corresponding to translation invariance.
The argument of the elliptic function can be re-written
in a Lorentz invariant form as $p^\mu x_\mu$
which implies a dispersion relation
$p^2=\mu^2 g$.
One should stress that two solutions (\ref{phi0}) are gauge non-equivalent
and have different implementations in the mathematical structure of the Yang-Mills theory.
\subsection{Spherically symmetric non-stationary finite energy solutions}
In this subsection we overview briefly known spherically-symmetric non-stationary solutions
\cite{p14,p15,p16,p17,p18}. We consider a restricted class
of such solutions which have a spherically symmetric
initial shape in the rest frame, i.e., the total linear
momentum is supposed to be zero.
A class of finite energy solutions is described by the following ansatz
for the gauge potential in spherical coordinates
$(r,\theta,\varphi)$
with one unconstrained function $\psi(t, r)$ \cite{p14,p15,p16,p17,p18}
\begin{eqnarray}
A_m^a= -\epsilon^{abc}\hat n^b \partial_m \hat n^c (1+\psi(t,r)), \label{wuyang}
\end{eqnarray}
where $\hat n=\vec r/r$. It is clear, that the ansatz (\ref{wuyang})
describes a generalized time dependent Wu-Yang monopole configuration.
A known static Wu-Yang monopole corresponds to a special case $\psi(t,r)=0$,
and a trivial pure gauge vacuum configuration is described by $\psi(t,r)= \pm 1$.
Direct substitution of the ansatz into the equations of motion
leads to one non-trivial independent partial differential equation
\begin{eqnarray}
\partial_{t}^2\psi-\partial_{r}^2\psi+\displaystyle\frac{1}{r^2}\psi(g^2\psi^2-1)=0. \label{eqspher}
\end{eqnarray}
The equation (\ref{eqspher}) represents a second order hyperbolic equation
which admits a wide class of wave solutions determined uniquely
by given initial conditions. We give one example of finite energy solutions
which looks like a kink in two-dimensional coordinate plane $(t,r)$
(others can be found in \cite{p14,p15,p16,p17,p18}).
To find a numeric solution one chooses an initial profile function $\psi(t=0,r)$
in such a manner that the magnetic field vanishes at space
infinity $\psi(t=0,r)=1-r^2 {\rm e}^{-r^2}$.
This profile function describes a monopole like configuration
with a maximal magnetic charge at a finite distance from
the center, and a vanishing total magnetic charge at large distances.
A corresponding numeric solution demonstrates
a soliton structure in the effective 1+1 space-time $(t, r)$;
the solution $\psi(t, r)$ represents a lump
in the coordinate plane $(t, r)$, FIG. \ref{Fig5}, which moves
in radial direction to space infinity with the light speed.
\begin{figure}[htp]
\centering
\includegraphics[width=70mm,height=55mm, bb=0 0 541 356]{kink2.jpg}
\caption[fig5]{A kink type solution
with initial conditions $\psi(0, r)=1-r^2 {\rm e}^{-r^2},
~\partial_t \psi(0, r)=0$.}\label{Fig5}
\end{figure}
Note that one finds a similar behavior of the magnetic field for the
system of localized monopoles and antimonopoles in the Weinberg-Salam
model \cite{pakIJMP15}.
We conclude that a pure QCD admits a wide class
of non-linear wave solutions.
All solutions carry zero spin and the non-linear plane waves
has the dispersion relation with the mass parameter.
Taking into account that a general solution has
additional three parameters corresponding to orientation
in the internal space of the group $SU(2)$,
one can interpret the non-linear pane waves as
three longitudinal dynamic degrees of freedom
in addition to six transverse dynamic degrees
of freedom represented by Coleman non-Abelian
plane waves \cite{coleman}.
Note that transverse non-linear waves remain massless,
so the mass appears only from the solutions containing
longitudinal projections of the field strength.
This is contrary to other approaches like the models of massive gluodynamics
where the gluon mass is introduced either through spontaneous symmetry breaking
or by adding explicit mass terms \cite{math}.
A principal advantage of our treatment is that the color gauge symmetry
remains unbroken, so the theory is still renormalizable in a standard perturbative
quantum field theory.
\section{Wu-Yang monopole dressed in off-diagonal gluon field}
Let us rewrite the ansatz (\ref{wuyang})
by performing an appropriate $SU(2)$ gauge transformation
in a so-called Abelian gauge \cite{choprd80}
as follows
\begin{eqnarray}
A_\theta^2&=&\psi (t, r), \nonumber \\
A_\varphi^1&=& -\psi(t, r) \sin \theta, \label{spherwav} \\
A_\varphi^3&=&\displaystyle\frac{1}{g} \cos \theta. \nonumber
\end{eqnarray}
All other components of the gauge potential are identically zero.
The expression for the Abelian component $A_\varphi^3$ contains coordinate singularity
along the $z$-axis, so such a gauge represents a singular gauge.
One should stress, that final gauge-invariant quantities (like the Lagrangian and energy density)
are regular and do not depend on a chosen gauge.
We prefer the ansatz written in the Abelian gauge \cite{choprd80},
(\ref{spherwav}), since such notation is more suitable for description of stationary
monopole solutions in $SU(N)$ Yang-Mills theory and multimonopole
configurations. Besides, due to gauge invariant decomposition of the gauge potential \cite{choprd80}
one can treat the Abelian gauge potential $A_\varphi^3$ as a static Wu-Yang monopole field
and the function $\psi(t,r)$ as a dynamic degree of freedom describing the off-diagonal gluon.
We are interested in regular stationary wave solutions
to equation (\ref{eqspher}). Let us write down
the energy functional corresponding to the ansatz (\ref{spherwav}),
\begin{eqnarray}
E&=&\int\!{\rm d}r\, {\rm d}\theta\, {\rm d}\varphi\,
\sin \theta\Big ((\partial_t \psi)^2+(\partial_r \psi)^2
\nonumber
\\
&&\hspace*{27mm}
+\displaystyle\frac{1}{2g^2r^2}(g^2 \psi^2-1)^2 \Big ) \nonumber \\
&\equiv& 4 \pi \int\! {\rm d}r \, {\cal E}(t, r), \label{entotspher}
\end{eqnarray}
where ${\cal E}(t, r)$ is an effective energy density
defined on (1+1)-dimensional half-plane $(r\geq 0, \ 0<t<\infty)$.
The effective energy density can be treated as
a Hamiltonian of two-dimensional $\lambda \phi^4$ theory
with a radial dependent coupling $\lambda \equiv 1/(2g^2r^2)$.
The field strength components
contain the following non-vanishing projections
of the color magnetic and electric field
\begin{eqnarray}
F_{r\theta}^2&=&\partial_r \psi,~~~~~~~F_{r\varphi}^1=-\partial_r \psi \sin\theta, \nonumber \\
F_{\theta\varphi}^3&=&\displaystyle\frac{1}{g}(g^2\psi^2-1)\sin \theta, \label{Ftf} \\
F_{t\theta}^2&=&\partial_t \psi, ~~~~~~~F_{t\varphi}^1=-\partial_t\psi\sin\theta. \nonumber
\end{eqnarray}
The radial magnetic field component $F_{\theta\varphi}^3$
generates a non-zero magnetic flux through a sphere with a center
at the origin, $r=0$.
So that, the color magnetic charge of the monopole depends on time and distance from the center.
Note that various static generalized Wu-Yang monopoles have been considered before,
however, all of them have singularities in agreement with the
Derrick's theorem \cite{derr}.
The equation (\ref{eqspher}) admits a local non-static solution near the origin
which removes the singularity of the monopole at the center
\begin{eqnarray}
\psi(t,r)&=&\displaystyle\frac{1}{g}+\sum_{n=1} c_{2 n}(t) r^{2 n}, \nonumber \\
c_4(t)&=& \displaystyle\frac{1}{10} \big (3 g c_2^2(t) +c_2''(t)\big ), \nonumber \\
c_6(t)&=&\frac{1}{28}\big(c_4''(t)+6gc_2(t)c_4(t)+g^2c_2^3(t)\big), \nonumber \\
&\vdots& \label{locsol}
\end{eqnarray}
where the coefficient functions $c_{2n}(t)$ ($n\geq 2$) are determined in
terms of one arbitrary function $c_2(t)$.
In asymptotic region, $r\simeq \infty$,
the non-linear
equation (\ref{eqspher}) reduces to a free D'Alembert
equation which has a standing spherical wave solution
\begin{eqnarray}
\psi(t, r) &\simeq& a_0+A_0 \cos (M r) \sin(M t)
+{\mathcal O}\Big(\displaystyle\frac{1}{r}\Big), \label{asymsol}
\end{eqnarray}
where $a_0, A_0$ are integration constants,
the mass scale $M$ appears due to scaling invariance
in the theory under dilatations $r \rightarrow Mr, t\rightarrow Mt $.
One should stress, that the asymptotic solution represents a standing spherical wave only
in the leading order and our solution can not be treated as a superposition
of out-going and in-going spherical waves due to absence of superposition principle in the
non-linear theory. Besides, as we will see below, the parameters $a_0, ~A_0$ are not independent
free parameters. Moreover, all known before non-linear spherical wave solutions in the Yang-Mills theory
are singular. Note that the series expansion for the local solution
starts from the factor $1/g^2$ which reflects interrelationship of the solution
with the non-perturbative topological origin of the Wu-Yang monopole. In particular,
the presence of such a term cancels the singularity of the Wu-Yang monopole.
The second term in the series expansion, $c_2(t) r^2$, contains quadratic dependence on the radial
coordinate. This provides finiteness of the energy density at the origin.
The equation (\ref{eqspher}) represents a hyperbolic partial differential equation
and admits a correct Cauchy problem setup with arbitrary
initial conditions for the function $\psi(t,r)$ at the initial time $t=0$.
For instance, one can choose any initial profile function periodic along the
radial direction. However, such a solution will not be stationary in general.
To find a stationary solution we will solve a Cauchy problem imposing
initial conditions with a periodic in time initial conditions at the origin $r=0$.
To solve numerically such problem one chooses a rectangular numeric
domain $(L_0 \leq r \leq L,~ 0\leq t \leq L)$. Since at $r=0$ one has coordinate
singularity which implies the stiffness numeric problem we introduce a small number
$L_0$ and then check convergence of the numeric solution in the limit $L_0 \rightarrow 0$.
We use the local solution (\ref{locsol})
to impose initial Dirichlet conditions along the boundary $r=L_0$.
The initial profile function $c_2(t)$ can be chosen arbitrarily as any regular
periodic function, we set it for simplicity in terms of ordinary sine function
\begin{equation}
c_2(t)=c_0+c_1 \sin (M t), \label{locsol10}
\end{equation}
where $c_0, \, c_1$ are numeric constants.
Note that only one of two parameters in the local solution (\ref{locsol10})
is free, the other is fixed by the requirement that a
numeric solution matches the asymptotic solution (\ref{asymsol}).
Dimensional analysis implies that the energy of the solutions
is proportional to $M$ so that the energy vanishes in the limit $M\rightarrow 0$.
This might cause some doubts on existence of the solution.
However, one should stress that standard arguments
on existence of static solitonic solutions based on the Derrick's theorem
\cite{derr} are not applicable to the case of stationary solutions
which satisfy the extremum principle of the classical action,
not the energy functional. In the case of Yang-Mills theory the action is conformal invariant
and its first variational derivative with respect to the scale parameter $M$ vanishes
identically. So the parameter $M$ determines the scale of the space-time coordinates
and can be set to one without loss of generality.
With this one can solve numerically the equation (\ref{eqspher}), and
the corresponding solution is depicted in Fig.\ref{Fig1}.
\begin{figure}[h!]
\centering
\includegraphics[width=70mm,height=52mm,bb=0 0 562 377]{gr1nov17c.jpg}
\caption[fig1]{Stationary monopole solution;
$(0\leq r,t \leq L)$, ~$L=8 \pi$, $c_0=-0.041$,~$c_1=-0.523,~g=1$
.}\label{Fig1}
\end{figure}
Note that one has a stiffness numeric problem near the origin, so that
we have checked the regular structure and convergence of the numeric solution
in close vicinity of the origin up to $L_0=10^{-6}$ while keeping the
radial size of the numeric domain of order $L=64 \pi$.
A general stationary monopole
solution can be classified by two of three parameters $a_0, A_0, M$
characterizing the asymptotic behavior of the solution (\ref{asymsol}).
The mass scale parameter $M$ takes arbitrary values
whereas the mean value $a_0$ and amplitude $A_0$ parameters
are constrained. Numerical analysis implies the following dependence
of the amplitude $A_0$ of the monopole solution on its mean value parameter
$a_0$, Fig. 3. The dependence of the amplitude $A_0$ on the mean value $a_0$ of the monopole
solution is important in study of the quantum stability of the vacuum gluon condensation
based on using the classical stationary monopole solution \cite{p2}.
\begin{figure}[htp]
\centering
\includegraphics[width=70mm,height=55mm,bb=0 0 324 213]{Aug1A0dep-a0mod.jpg}
\caption[fig8]{Dependence of the amplitude $A_0$ of the monopole solution on the
mean value $a_0$,~ $M=1$.}\label{Fig8}
\end{figure}
Note that in the case of vanishing function $\psi(t, r)$ one has still a non-trivial solution
for the Wu-Yang monopole which has a singularity at the origin $r=0$.
The monopole is defined by the Abelian gluon field $A_\varphi^3=\cos \theta$
and the function $\psi(t, r)$ describes off-diagonal gluons.
The obtained solution can be treated as a Wu-Yang monopole dressed in a
spherical standing wave made of off-diagonal gluons. It is surprising that
the spherical standing wave regularizes the singularity of the Wu-Yang monopole
resulting in a finite energy density at the origin $r=0$. Another interesting feature
of the solution is that the standing wave does not screen completely the color monopole
charge at large distances. One can find that in the asymptotic region the function $\psi(t,r)$
oscillates around the value $b_0\simeq 0.65$.
So the radial component of the color magnetic field
$F_{\theta \varphi}^1$ has a non-zero averaged value
which provides a non-vanishing total color magnetic charge.
One can easily generalize the above consideration of $SU(2)$ stationary
monopole solutions to the case of $SU(N)$ Yang-Mills theory.
For the case of a pure $SU(3)$ QCD one has the following ansatz
for generalized Wu-Yang monopole solution corresponding to
color magnetic charge two
\begin{eqnarray}
&A_\theta^2=\psi_1 (t, r),~~~~~~~~~~~~~& A_\theta^5=\psi_2 (t, r), \nonumber \\
&A_\varphi^1= -\psi_1(t, r) \sin \theta, ~~~~~& A_\varphi^4= \psi_2(t, r) \sin \theta, \qquad\label{spherwav3} \\
&A_\varphi^3= \displaystyle\frac{1}{g} \cos \theta, ~~~~~~~~~~~~~~&
A_\varphi^8=-\displaystyle\frac{\sqrt 3}{g} \cos \theta, \nonumber
\end{eqnarray}
where the Abelian gauge potentials
$A_{\varphi}^{(3,8)}$, corresponding to generators of the Cartan subalgebra of $SU(3)$,
describe a static Wu-Yang monopole with a total color magnetic charge two,
$g_m^{\rm tot}=2$, \cite{choprl80}. The functions $\psi_{1}(t,r)$ and $\psi_{2}(t,r)$
correspond to dynamic degrees of freedom of the off-diagonal components of the gluon field.
One can verify that substitution of the ansatz (\ref{spherwav3}) into the equations of motion
of the pure $SU(3)$ QCD implies two independent partial differential equations for two functions
$\psi_{1,2}$
\begin{equation}
\begin{array}{rcl}
&&\partial^2_{t} \psi_1 -\partial^2_r \psi_1+\displaystyle\frac{g^2}{2 r^2} \psi_1
\Big(2 \psi_1^2-\psi_2^2-\displaystyle\frac{2}{g^2}\Big)=0, \\
&&\partial^2_{t} \psi_2 -\partial^2_r \psi_2+\displaystyle\frac{g^2}{2 r^2} \psi_2
\Big(2
\psi_2^2-\psi_1^2-\displaystyle\frac{2}{g^2}\Big)=0.
\end{array}
\label{eqsu3}
\end{equation}
In a special case, $\psi_1(t,r)=\psi_2(t,r)\equiv \psi(t,r)$, the equations (\ref{eqsu3})
reduce to one differential equation
\begin{eqnarray}
&&\partial^2_{t} \psi -\partial^2_r \psi+\displaystyle\frac{g^2}{2 r^2} \psi
\Big(\psi^2-\displaystyle\frac{2}{g^2}\Big)=0.
\end{eqnarray}
By simple rescaling $\psi(t,r) \rightarrow \sqrt 2 \psi(t,r)$
the last equation is transformed to the equation $(\ref{eqspher})$
for $SU(2)$ monopole solution.
Existence of stationary solutions with a magnetic charge and a finite energy density
is unexpected in a pure Yang-Mills theory since there is no such an analogue
in linear field models including the Maxwell type gauge field. Remind, that
a known 't Hooft-Polyakov monopole solution includes external Higgs scalar
fields which regularize the monopole singularity at the origin.
Besides, one can verify that the standard electroweak theory does not admit
such regular stationary monopole solutions.
So that, the quantum chromodynamics is a unique theory
among the currently known theoretical field models of fundamental interactions
which possesses such a surprising feature.
\section{Stability analysis}
Stability of static localized solitonic solutions can be verified by checking whether the second variation
of the energy with respect to small deformations becomes negative.
If solution possesses a specific space symmetry one should
take into account perturbations of most general field configuration as well.
We will show that the stationary spherically symmetric monopole solution
considered in the previous section
is unstable against small axially-symmetric perturbations $Q_\mu^a(r,\theta,t)$
around the classical solution determined by the ansatz
(\ref{spherwav}).
In the case of stationary solutions with infinite total energy one can apply a similar approach
to stability problem as in the case of solitons. We study an eigenvalue spectrum of the operator $\hat K_{\mu\nu}^{ab}$
defined by means of second variational derivatives of
the classical action $S[A]$
\begin{eqnarray}
&&\hat K_{\mu\nu}^{ab} Q_\nu^a=\lambda Q_\mu^a, \label{schreq} \\
&&\hat K_{\mu\nu}^{ab} \equiv \displaystyle\frac{\delta^2}{\delta Q_\mu^a \delta Q_\nu^b} {S[A_{cl}+Q]}. \label{operK}
\end{eqnarray}
It is instructive to show that despite on infinite total energy and asymptotic oscillating behavior
of the Lagrangian corresponding to the stationary monopole solution
the classical action is finite and well-defined.
One can check that non-linear plane wave and stationary monopole solutions
realize an absolute maximum of the classical action.
Let us verify this for the stationary monopole solution.
A solution for the stationary monopole can be represented in terms of Fourier series
as follows
\begin{eqnarray}
\psi(t,r)=C_0(r)+\sum_{n=1,2,3,...} P_n(r) \cos(n r) \sin (n t), \label{series}
\end{eqnarray}
where one has only cosine functions in the decomposition due to the structure of the local solution,
(\ref{locsol}), and the radial functions $C_0(r), P_n(r)$ satisfy
the boundary condition at the origin, $C_0(0)=1$, $P_n(0)=0$, and asymptotic solution (\ref{asymsol}), i.e.,
$P_n(r=\infty)=A_0$.
Let us consider for simplicity a leading mode approximation keeping only the first term in the series decomposition
(\ref{series}).
The functions $C_0(r), P_1(r)$ can be decomposed in series in degrees of orthogonal polynomials. For our purpose to
demonstrate that the variational method can be applied successfully to the classical action we choose a simple variational trial function for the monopole solution
\begin{eqnarray}
\psi(t,r)&=&1-\displaystyle\frac{(1-a_0) r^2}{1+b_0 r^2}\nonumber \\
&+&A_0 (1-e^{-d_0 r^2}) \cos (M r) \sin(M t), \label{interpolf}
\end{eqnarray}
where the conformal scale factor $M$ will be set to one in numeric calculations,
$a_0,b_0,A_0, d_0$ are fitting number parameters.
Note that in the case of the stationary monopole solution the convergence of the series (\ref{series})
is very fast and the expression (\ref{interpolf}) provides a good interpolation function
to the exact numeric solution.
Substituting the trial function (\ref{interpolf}) into the action one can perform integration over the time in the
interval $(0\leq t \leq 2 \pi)$,
integration over the spherical angles $(\theta, \varphi)$ is trivial and leads to a number factor $4\pi$.
The obtained effective action $A^{eff}$ is defined in one-dimensional space $(0\leq r \leq \infty)$
and contains a divergent term $A^{div}$ which originates
from the kinetic terms in the original Lagrangian
\begin{eqnarray}
A^{div}
&=& 4 \pi^2 \int dr A_0 \cos(2 r).
\end{eqnarray}
A simple regularization by introducing an exponential factor
$e^{-\varepsilon r}$ with an infinitesimally small number $\varepsilon$ leads to a vanished
value of the divergent term.
The remaining part of the effective action is well-defined and integration over the radial coordinate
can be easily performed numerically. With this, the action represents a regular function
of the trial parameters $(a_0, A_0, b_0, d_0)$. The first two parameters represent a mean value and amplitude
of the monopole solution in the asymptotic region, the parameters $b_0, d_0$ describe
matching profile of the local and asymptotic solutions along the radial direction.
One can fix one of two parameters $a_0, A_0$ since only one of them is independent, (see Fig. 3),
all other parameters are found by variational procedure which finds an extremum the action.
The numeric results for a given asymptotic amplitude $A_0=0.56$
show that the classical action has an absolute maximum, $S^{max}=-0.168...$,
with the following trial parameters values
\begin{eqnarray}
a_0&=&0.9982..., \nonumber \\
b_0&=&0.02588..., \nonumber \\
d_0&=&0.1786...\, .
\end{eqnarray}
The results are in a good qualitative agreement with the numeric results for the
stationary monopole solution presented in Figs. 2,3.
Now we consider the stability of the monopole solution under axially-symmetric
perturbations. We consider a small perturbation $Q(r,\theta,t)$
around the Abelian gauge potential $A_\varphi^3$ which describes a static Wu-Yang monopole
within the framework of the ansatz (\ref{spherwav})
\begin{eqnarray}
A_\varphi^3&=& \cos \theta+Q(r,\theta)\cos (M t). \label{flucfuncs}
\end{eqnarray}
We constrain our study by consideration of perturbations with time dependent factor $\cos (M t)$
with the same frequency $M$ as one in the monopole solution.
It is suitable to pass to standard notations in spherical coordinates defining
the perturbation field $\tilde Q(r,\theta)$ of mass dimension
\begin{eqnarray}
\tilde Q(r,\theta)&=& \displaystyle\frac{1}{r \sin\theta} Q(r,\theta).
\end{eqnarray}
For brevity of notations we denote the interpolation function $\psi(t,r)$, (\ref{interpolf}),
as follows
\begin{eqnarray}
\psi(t,r)&=&1+P_0(r) \cos (M t).
\end{eqnarray}
In the case of small perturbations it is enough to keep only terms quadratic in $Q(r,\theta)$
in the classical action. Substituting the perturbed potential $A_\varphi^3$, (\ref{flucfuncs}),
into the classical action and performing
integration over $(t,\varphi)$ one obtains the following
quadratic form for the operator $\hat K_{\mu\nu}^{ab}$, (\ref{operK}),
\begin{eqnarray}
v{\cal L}^{(2)}&=& 2 \pi^2 \Big [ -r^2 \sin^2 \theta (\partial_r \tilde Q)^2-\sin\theta (\partial_\theta {\tilde Q})^2 \nonumber \\
&-&
2 {\tilde Q}(\cos \theta \partial_\theta {\tilde Q}+r \sin \theta \partial_r {\tilde Q})- \Big (\csc\theta+\nonumber \\
&&(1-M^2 r^2 +\displaystyle\frac{3}{4} P_0^2\Big) \sin \theta {\tilde Q}^2
\Big ] ,
\end{eqnarray}
where $v\equiv r^2 \sin \theta$ is an integration volume in spherical coordinates.
Varying the last expression with respect to $\tilde Q$ one can write down explicitly the eigenvalue equation
(\ref{schreq}) for possible unstable modes
\begin{eqnarray}
&& -r^2 \sin\theta \Big [-\partial_r\partial_r-\displaystyle\frac{2}{r}\partial_r-\displaystyle\frac{1}{r}(\partial_\theta\partial_\theta +\cot \theta )
+V \Big ] {\tilde Q} \nonumber \\
&&
= \lambda {\tilde Q}, \label{eigeq} \\
&& V \equiv -M^2 + \displaystyle\frac{1}{r^2} \Big ( 1+\displaystyle\frac{1}{\sin^2\theta} +\displaystyle\frac{3}{4} P_0(r) \Big).
\end{eqnarray}
The equation resembles a Schrodinger type equation (up to opposite sign on the left hand side)
with a quantum mechanical potential $V$. Simple consideration shows that
due to presence of the factor $-M^2$ the potential $V$ is negative at large distance.
This may cause appearance of negative eigenvalues
for such a ``Schrodinger'' equation, or, equivalently, positive values for $\lambda$
in the original eigenvalue equation (\ref{eigeq}). Corresponding eigenfunctions
represent unstable perturbation modes which will increase the value of the non-perturbed action.
Therefore, the spherical monopole solution will be unstable under axially-symmetric field perturbations.
Note that in the case of spherically symmetric perturbations around the ansatz (\ref{spherwav})
a similar term $-M^2$ in the potential $V$ does not imply negative eigenvalues. A detailed
analysis shows that due to scaling invariance there are only zero perturbation modes corresponding
to zero eigenvalues.
To solve numerically the equation (\ref{eigeq}) we impose the following boundary conditions
for the perturbation field ${\tilde Q}$
A numeric solution corresponding to a positive eigenvalue closest to zero
is presented in Fig. 4a. The unstable mode ${\tilde Q}_1(r,\theta)$
has oscillating behavior with an amplitude decreasing in radial direction as $\displaystyle\frac{1}{r^2}$.
\begin{eqnarray}
{\tilde Q}(\infty,\theta)&=& 0 , ~~~~~~~{\tilde Q}(r, 0)= {\tilde Q}(r,\pi).
\end{eqnarray}
\begin{figure}[h!]
\centering
\subfigure[~]{\includegraphics[width=80mm,height=50mm]{u3fluctw=1finalF.jpg}}
\hfill
\subfigure[~]{\includegraphics[width=80mm,height=60mm]{u3Lagrfluct1w=1finalBFFF.jpg}}
\caption[fig4]{(a) Perturbation mode ${\tilde Q}_1(r,\theta)$ of even parity corresponding to the eigenvalue $\lambda_1=+1.498$;
(b) an integral density
$r^2 \sin\theta {\cal L}$ for the second variation of the classical action.}\label{Fig4}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[~]{\includegraphics[width=80mm,height=55mm]{u3fluct2w=1finalF.jpg}}
\hfill
\subfigure[~]{\includegraphics[width=80mm,height=55mm]{u3Endensfluct1w=1finalF.jpg}}
\caption[fig5]{(a) Perturbation mode ${\tilde Q}_2(r,\theta)$ of odd parity corresponding to the eigenvalue $\lambda_2=+2.320$;
(b) a respective energy density $\cal{E}$.}\label{Fig5}
\end{figure}
Note that since the eigenvalue equation (\ref{eigeq}) is linear
one can normalize the function ${\tilde Q}_1(r,\theta)$ to any small number. We keep non-renormalized
solutions ${\tilde Q}(r,\theta)$ in our numeric results.
Numeric estimate of the second variation of the classical action
confirms that it is positive. So, the perturbations increase the extremum value of the action
showing instability of the spherically symmetric monopole.
Finite energy perturbation modes can be classified into
two type field configurations with respect to reflection symmetry
$\theta \rightarrow \pi-\theta$. The first mode ${\tilde Q}_1$ keeps
its shape under the reflection and has even parity.
An unstable mode
corresponding to the next positive eigenvalue is depicted in Fig. 5a,
and it represents odd parity field configuration.
Unstable modes with larger positive eigenvalues correspond to presence of
higher Fourier modes in polar angle $(\theta)$.
All regular unstable modes with appropriate asymptotic behavior
lead to a finite total energy and finite second variation of the classical action.
In particular, the energy density ${\cal E}_1$ corresponding to the mode ${\tilde Q}_1$ has integrable singularities
at two points on $Z$-axis, $(r\simeq \pm 0.72, \theta=0,\pi)$, and the energy density
${\cal E}_2$ corresponding to the mode ${\tilde Q}_2$
has a finite maximum along two rings with centers located on the $Z$-axis
at finite distance from the origin, Fig. 5b.
Remind that ${\tilde Q}(r,\theta)$ is a perturbation around the Abelian gauge field $A_\varphi^3(r,\theta)=\cos\theta$ which describes a static Wu-Yang monopole.
The presence of two types of unstable modes with energy density maximums located near the $Z$-axis at finite
distance from the origin indicates to possible existence of two types of axially symmetric solutions
which might correspond to monopole-monopole and
monopole-antimonopole pairs.
In general, in non-linear theories
the stability of stationary solutions represents a non-trivial problem.
We perform analysis of the stability in linear approximation
for axially-symmetric perturbation around the Abelian gauge potential $A_\varphi^3$.
In this stage we may conclude
that spherically symmetric monopole is unstable under small axially-symmetric perturbations.
\section{Gluon spin density operator}
The classical non-linear solutions considered in the present paper have
common features: all of them have a mass scale parameter, longitudinal projections of color fields
and zero spin density. It is known that gluons in QCD gain a dynamical mass
due to non-perturbative quantum corrections.
Appearance of the rest mass in classical spinless non-linear wave solutions
implies possibility that corresponding spinless states may exist
in the quantum theory. The propagating non-linear waves described in Section II
can be treated as a longitudinal massive modes in addition to the
Coleman transverse non-linear waves. It is clear, that
such of a set of the non-linear transverse and longitudinal plane waves
provides an attractive possibility for finding a proper non-perturbative quantization
of QCD. This raises also a question of finding a consistent definition
of the gluon spin density operator corresponding to the propagating spinless waves.
A strict concept of the gluon spin operator includes several aspects,
among them, the gauge invariance and Lorentz frame independence are the most important
issues. Below we construct a unique gauge invariant and Lorentz frame independent definition
of the gluon spin density operator for non-linear spinless waves at classical level.
Let us recall the main problems related to gauge invariant
definitions of gluon spin and angular momentum operators in QCD.
The standard canonical decomposition of the nucleon
total angular momentum includes gauge non-invariant
terms corresponding to quark and gluon spin and orbital
momentum operators \cite{JM}
\begin{eqnarray}
J^{\rm can}_{\mu \nu}&=& \displaystyle\int{\rm d}^3x\,
\Big\{ {\bar \psi}
\gamma^0 \displaystyle\frac{\Sigma_{\mu\nu}}{2} \psi
-i {\bar \psi} \gamma^0 x_{[\mu} \partial_{\nu]} \psi
- A_{a[\mu} F^a_{\nu] 0}
\nonumber \\
&&
\hspace*{12mm}
- F_{0 \rho}^a x_{[\mu} \partial_{\nu]} A^\rho_{a}\Big\}.
\quad \label{canon}
\end{eqnarray}
The expressions for the canonical spin density $S_{\mu\nu}= -{A}_{a[\mu} F^a_{\nu] 0}$ and angular momentum operators
for quark and gluon are gauge non-invariant.
It was assumed that a gauge invariant
definition of the gluon spin operator in non-Abelian
theory did not exist \cite{JM, ji}.
A novel approach towards constructing a gauge invariant
nucleon spin decomposition has been proposed in \cite{chen1,chen2}.
The main idea in this approach is to separate physical degrees of freedom
of gluon from pure gauge degrees of freedom, i.e., one splits
the gauge potential into two parts
\begin{eqnarray}
A_\mu^a&=&{\cal A}_\mu^a+\hat A_\mu^a,
\end{eqnarray}
where
$\hat A_\mu^a$ is a physical gauge potential containing
only physical degrees of freedom, and
$ {\cal A}_\mu^a$ is a pure gauge potential satisfying the
pure gauge condition
\begin{equation}
{\cal F}^a_{\mu\nu} \mbox{\small $({\cal A})$}=0. \label{puregauge}
\end{equation}
Both potentials are constructed
in terms of the original unconstrained vector potential $A_\mu^a$.
A key point is that the physical vector potential $\hat A_\mu^a$
transforms under the gauge transformation
as a vector in adjoint representation of $SU(2)$ whereas the pure gauge potential ${\cal A}_\mu^a$
transforms as a gauge connection.
This allows to re-write the canonical
decomposition (\ref{canon}) in an explicit gauge invariant manner \cite{chen1,chen2}
\begin{eqnarray}
J^{\rm can}_{\mu \nu}&=& \displaystyle\int{\rm d}^3x\,
\Big\{ {\bar \psi}
\gamma^0 \displaystyle\frac{\Sigma_{\mu\nu}}{2} \psi
-i {\bar \psi} \gamma^0 x_{[\mu} {\cal D}_{\nu]} \psi
- \hat{A}_{a[\mu} F^a_{\nu] 0}
\nonumber\\
&&\hspace*{13mm}
- F_{0 \rho}^a x_{[\mu} \big({\cal D}_{\nu]} \hat{A}^\rho_{a}
-{\cal F}_{\nu ]}{}^\rho{}_a({\cal A})\big)\Big\},
\label{canoncov}
\end{eqnarray}
where $ {\cal D}_\mu \equiv \partial_\mu+{\cal A}_\mu$ is a covariant derivative containing
the pure gauge potential.
Each term in (\ref{canoncov}), including the gluon spin density operator
$\hat S_{\mu\nu}=- \hat{A}_{a[\mu} F^a_{\nu] 0}$, is explicitly gauge invariant due to
covariant transformation properties of the physical gauge potential $\hat{A}_\mu^a$.
It is worth to stress that an equation defining the physical
potential can be chosen by several ways, and the most
important issue in the definition of the physical potential
is a problem of uniqueness of gauge invariant and
Lorentz frame independent definition of spin and momentum
operators for quarks and gluon. Within the framework of the
formalism proposed by Chen et al \cite{chen1,chen2} a Lorentz frame independent
definition is proposed in \cite{cho1}, however, an explicit construction
of such a definition is unknown since a perturbative solution
of the equation for the physical gauge potential does not exist on the space
of classical free plane wave solutions for the gluon field
(see also the review \cite{review1} for current status of the problem).
Note that for massless particles a consistent concept of spin is given by the
notion of helicity. In the case of the Maxwell electrodynamics
the gauge potential must satisfy
two helicity gauge conditions,
\begin{eqnarray}
A_0&=&0, ~~~~A_3=0, \label{hel1}
\end{eqnarray}
to represent helicity eigenstates of the
operator $J_3$ of a little group $E(2)$ which provides
naturally the Lorentz invariance of the helicity operator \cite{yskim}.
The helicity conditions can be expressed in equivalent forms
using various combinations of a generalized axial gauge condition
\begin{eqnarray}
n^\mu A_\mu&=&0, \label{hel2}
\end{eqnarray}
where the constant vector $n^\mu$ specifies the temporal, axial or light-cone gauge
conditions.
The meaning of the helicity conditions is simple, they fix pure gauge degrees of freedom
while keeping two transverse dynamic degrees of freedom of the gluon.
Generalization of such description of the helicity operator to the case of
non-Abelian theory has been done in \cite{chen1, chen2,hatta,pakspin}.
Since the helicity operator is Lorentz frame independent, all definitions satisfying the
helicity conditions are consistent even though the
defining equations for the physical gauge potential
are not manifestly Lorentz invariant \cite{review1,pakspin}.
In the case of non-linear massive plane waves described by the ansatz
(\ref{ansf1234}) the helicity conditions $A^a_0=0, ~ \partial^i A^a_i=0$ are fulfilled
only in the rest frame, $\beta=0$. So that the definition of spin operator for such solutions
based on the helicity conditions becomes Lorentz frame dependent.
Solutions defined by the ansatz (\ref{wuyang},\ref{spherwav},\ref{spherwav3}) satisfy
the helicity conditions as well, $A_0^a=0, ~A_r^a=0$, but
break the Lorenz gauge condition. In addition, one can verify that
the canonical spin density for all these solutions vanishes identically.
This implies that for massive spinless non-linear waves
one should find a proper definition for gluon spin operator.
Note that the helicity conditions are consistent
with the Gauss law, which guarantees consistency with all equations of motion.
This is somewhat unexpected because the presence of helicity conditions
means that one has two transverse polarizations of gluon which
correspond to spin one particles. However, the non-linear wave type solutions
considered above admit projections of color fields along
the wave vector. This allows to interpret such non-linear waves as longitudinal modes,
raising a question about the number of dynamic degrees
of freedom for gluons.
It is clear, that the origin of such a subtlety comes from
the non-linearity of the gauge field strength expressed in terms
of the gauge potential. One possible way to revise
the notion of the gluon spin is to develope the old approach
based on constructing gauge invariant quantities from the field strength
\cite{weinberg64a,weinberg64b}.
Note that, so far the known gauge invariant definitions of the gluon spin operator
are conditioned by assumption that gluon has only two transverse dynamical
degrees of freedom per each color degree of freedom. Such an assumption is based on
the standard perturbative quantization of QCD.
The presence of propagating massive wave solutions, which are essentially non-linear,
implies that definition of the gluon spin should be revised to include
description of zero spin states.
Let us consider the definition for the physical gauge
potential based on the Lorenz gauge type equation
\begin{equation}
\big({\cal D}^\mu \hat A_\mu\big)^a=0. \label{lorconstr}
\end{equation}
One should stress that the last equation looks similar to Lorenz gauge fixing condition
in presence of a background field, however, it does not represent
any gauge fixing in a fact. An explicit expression for the physical gauge potential $\hat A_\mu^a$
is provided as a solution to equation (\ref{lorconstr}) in terms of a general
initial gauge potential $A_\mu^a$ in such a manner
that $\hat A_\mu^a$ still possesses full gauge freedom (it is transformed
as $SU(2)$ vector in adjoint representation, not as a gauge connection).
In practical use it is suitable to choose a real Lorenz gauge fixing condition
for the physical potential with a trivial pure gauge counterpart, ${\cal A}_\mu^a=0$.
This is a key idea of the approach towards resolving the problem of
gauge invariant definitions for gluon spin and orbital momentum operators \cite{chen1, chen2}.
The equation (\ref{lorconstr}) is unique (except the case of a generalized covariant Lorenz gauge type
equation for the physical gauge potential which will be discussed below) among all possible conditions
containing first derivatives and satisfying invariance under the Poincare group transformations.
The Lorenz gauge type condition (\ref{lorconstr})
was proposed in \cite{cho1} as a transversality condition.
In the case of Maxwell theory the definition of a gauge invariant photon spin operator
based on the solution of the Lorenz type equation (\ref{lorconstr}) encounters
a well-known problem of incompleteness of the Lorenz gauge \cite{pakspin}.
Namely, in the Lorenz gauge one has still a residual symmetry which
implies that equation of motion for the temporal component of the gauge potential
admits unphysical propagating modes. To fix such a symmetry one has to impose
an additional gauge condition.
We show that in the non-Abelian $SU(2)$ gauge theory the definition of a gauge invariant
spin operator based on the equation (\ref{lorconstr}) is possible
due to absence of the residual symmetry for the class of
non-linear wave solutions described by ansatz (\ref{ansf1234},\ref{wuyang},\ref{spherwav},\ref{spherwav3}).
To verify this we construct a formal series operator expansion
for such a solution using the perturbative method.
The solution for the physical potential $\hat A_\mu$ can be found
as a series
\begin{equation}
\hat A^a_\mu=\hat A_\mu^{a(0)}+g \hat A_\mu^{a(1)}+g^2 \hat{A}_\mu^{a(2)}+
\cdots.
\end{equation}
Expressing the pure gauge potential in terms of the
physical one, ${\cal A}^a_\mu=A^a_\mu-\hat A^a_\mu$,
one can find a solution to the equations (\ref{puregauge}) and (\ref{lorconstr})
in the leading order approximation
\begin{eqnarray}
\hat A^a_\mu&=&
P_{\bot}{}_{\mu}{}^{\nu} A_\nu^a \nonumber \\
&& +g\displaystyle\frac{1}{\Box} P_{\bot}{}_\mu{}^\nu
\epsilon^{abc}\Big(
\partial^\rho(A_\rho^b A^c_\nu)
+\hat A_\rho^{b(0)} \partial^\rho \hat A_\nu^{c(0)} \nonumber \\
&& \hspace*{24mm}-\partial^\rho(A^b_\rho\hat A_\nu^{c(0)})
+\partial^\rho A_\nu^b\hat A_\rho^{c(0)}
\Big) \nonumber \\
&&
+\hat A_\mu^{a\, \rm long}, \label{solphys} \\
\hat A_\mu^{a\, \rm long}&=&
-g\epsilon^{abc}
\Big(\displaystyle\frac{1}{\Box}\partial^\rho A^b_\rho\Big)
\Big(P_{\bot}{}_{\mu}{}^\sigma A^c_\sigma\Big), \nonumber \\
\hat A_\mu^{(0)}&=& P_{\bot}{}_{\mu}{}^{\nu} A_\nu^a, \nonumber
\end{eqnarray}
where we use the transverse projectional operator
$P_\bot{}_\mu{}^\nu=\delta_\mu^\nu-\frac{\partial_\mu \partial^\nu}{\Box}$.
Note that the solution (\ref{solphys}) for the physical gauge potential
includes an expression $\hat A_\mu^{a\, \rm long}$ containing the longitudinal
part, $\partial^\rho A_\rho^a$, of the unconstrained gauge potential.
The fact that a gauge invariant expression (\ref{solphys})
is given in terms of the gauge potential, not in terms of the field strength,
reflects the property of the non-Abelian gauge theory
where the field strength does not determine all gauge invariant quantities
as it occurs in the Abelian gauge theory.
One should stress that non-locality appearing in the solution for
the physical gauge potential is unphysical. Such a non-locality disappears
when we impose the real Lorenz gauge fixing condition on the physical potential;
in this case the pure gauge potential vanishes identically and
the physical gauge potential coincides with the initial general
gauge potential $A_\mu^a$. So that the non-local expression for the
gluon spin density operator reduces to the standard local expression
for the canonical spin density with $A_\mu^a$ satisfying the usual Lorenz gauge condition.
Let us verify that the Lorenz gauge condition is complete
and does not possess a residual symmetry on a space of the
non-linear wave type solutions. A small gauge variation of the
Lorenz gauge condition leads to the following equation
for the gauge parameter $\lambda^a$
\begin{eqnarray}
(\partial^\mu D_\mu \lambda)^a=0. \label{lorcond3}
\end{eqnarray}
For simplicity we consider a class of plane wave solutions provided by
the ansatz (\ref{ansf1234}).
One can verify that any solution to equation (\ref{lorcond3}) for
the parameters $\lambda^a$
does not belong to the same class of the non-linear
plane wave solutions determined by the ansatz (\ref{ansf1234}).
Let us consider a counter example, a generalized Lorenz gauge
$\big(D^\mu A_\mu\big)^a=0$
which contains a full covariant derivative $D^\mu$ including
the gauge potential.
Under small gauge variation it
leads to a covariant D'Alembert equation
\begin{eqnarray}
(D^\mu D_\mu \lambda)^a=0. \label{eqlambda}
\end{eqnarray}
Applying the ansatz (\ref{ansf1234})
for homogeneous solutions, one obtains
the following system of equations
\begin{eqnarray}
&& \ddot\lambda_1+g\lambda_1 (\phi_2^2+\phi_3^2)=0, \nonumber \\
&& \ddot\lambda_2+g\lambda_2 (\phi_3^2+\phi_1^2)=0, \\
&& \ddot\lambda_3+g\lambda_3 (\phi_3^2+\phi_1^2)=0. \nonumber
\end{eqnarray}
It is clear, that the above system of equations on the space of
the non-linear plane wave solutions $(\ref{phi0})$ admits solutions for $\lambda^a$
exactly of the same type. This implies appearance of a
residual symmetry at the level of a full nonlinear theory
and makes the covariant Lorenz gauge
incomplete. So that, a defining equation for the physical gauge potential
based on use of a generalized covariant Lorenz gauge type equation can not
provide a self-consistent definition of a gauge invariant and Lorentz covariant gluon spin operator.
Summarizing, we propose a new definition for the gluon spin operator
for a class of non-linear plane wave solutions which admit a non-vanishing mass and a total spin zero.
The definition for the physical gauge potential is based on the equation (\ref{lorconstr})
and we have provided an explicit construction for the physical gauge potential, (\ref{solphys}),
which leads to unique gauge invariant and Lorentz frame independent definition of the gluon spin operator
for massive spinless non-linear plane waves. Note, that equation (\ref{lorconstr}) was suggested
in \cite{cho1} as a possible definition of the gluon spin operator for massless gluons
with helicity $\pm1$. However, the equation (\ref{lorconstr}) does not admit solutions
due to presence of the residual symmetry. So that, a consistent
definition of the gluon spin operator for massless gluons
is possible only on the basis of the helicity equations (\ref{hel1}, \ref{hel2}).
\section{Conclusion}
We propose a new stationary generalized monopole solution which
can be treated (at least in the singular Abelian gauge)
as a system of the static Wu-Yang monopole and off-diagonal gluon.
An essential feature of that solution is
that it possesses a finite energy density everywhere in the whole space.
For a class of propagating non-linear waves described by the ansatz (\ref{ansf1234})
we construct a unique gauge invariant and Lorentz frame independent definition
for the gluon spin operator based on a Lorenz gauge type equation for
the physical gauge potential.
Note that in the case of the stationary monopole solution
defined by the ansatz (\ref{spherwav}) the canonical spin density
vanishes identically as well. However, the Lorenz gauge condition
is not fulfilled
\begin{eqnarray}
\partial^\mu A_\mu^a&=&\delta^{a,2} \displaystyle\frac{1}{r^2}\cot \theta \psi(r,t).
\end{eqnarray}
This is not surprising, since the stationary monopole solution does not represent
a stationary propagating solution like a plane wave. Passing to arbitrary Lorentz frame
such a solution will represent a moving lump with a maximal energy density around its center mass.
Due to this, the definition of the gluon spin operator for the stationary
generalized Wu-Yang monopole solutions and for non-stationary solutions corresponding to the ansatz (\ref{spherwav})
is based on a generalized axial gauge condition as for the massless gluon \cite{pakspin}.
One should note that, since the notion of spin
has a quantum mechanical origin, one has to perform properly the quantization
of the gluon field and construct the algebra of quantized
angular momentum and spin operators. Since for the case of non-linear solutions the
angular momentum and spin operators are non-linear
operator functions of the quantized gauge potential $A_\mu^a$,
one expects that such a quantum algebra can be realized as
a deformation or a non-linear representation of the standard Lie algebra of $SO(3)$ .
An important issue is to study possible physical implications of non-linear wave type solutions
in QCD. It has been shown that non-linear type I plane waves, (\ref{phi0}), provide
a simple estimate of glueball spectrum in qualitative agreement
with lattice calculation \cite{frasca09}.
Our stability analysis shows that the stationary spherically symmetric monopole is unstable under axially-symmetric perturbations. This indicates to existence of stationary monopole and monopole-antimonopole pair solutions. This issue will be considered in a subsequent paper \cite{p2}.
\acknowledgments
One of authors (DGP) acknowledges Prof. Mo-Lin Ge and Prof. C.M. Bai
for warm hospitality during his staying in Chern Institute of Mathematics,
and Dr. E. Tsoy for numerous discussions.
The work is supported by NRF of Korea, grant MSIP No.2014R1A2A1A01002306(ERND),
and by Brain Pool Program (KOFST), Sogang University;
by NSFC (Grants 11035006 and 11175215),
Rare Isotope Science Project of
Inst. for Basic Sci. funded by Ministry of Science, ICT and Future
Planning and NRF of Korea (2013M7A1A1075764), and by UzFFR (Grant F2-FA-F116).
|
\section{Introduction}
In this paper, we aim to investigate the error analysis via energy arguments of a semidiscrete Galerkin finite
element (FE) method for time-fractional diffusion problems of the form:
find $u=u(x,t)$ such that
\begin{subequations}\label{a}
\begin{alignat}{2}\label{a1}
&u'(x,t) + \partial_t^{1-\alpha} {\cal{L}} u(x,t)
=f(x,t) &&\quad\mbox{ in }\Omega\times (0,T],
\\ \label{a2}
&u(x,t)= 0 &&\quad\mbox{ on }\partial\Omega\times (0,T],
\\ \label{a3}
&u(x,0)=u_0(x) &&\quad\mbox{ in }\Omega,
\end{alignat}
\end{subequations}
where ${\cal{L}} u= -{\rm div}({\bf a}(x)\nabla u)$, $\Omega$ is a bounded, convex polygonal domain in $\mathbb{R}^2$ with
boundary $\partial \Omega$, $f$, $a$ and $u_0$ are given functions defined on their
respective domains. Here, $u'$ is the partial derivative of
$u$ with respect to time and $\partial_t^{1-\alpha}:={^R}{\rm D}^{1-\alpha}$ is the
Riemann--Liouville time-fractional
derivative defined by: for $0<\alpha<1$,
\begin{equation} \label{Ba}
\partial_t^{1-\alpha} \varphi(t):=\frac{\partial }{\partial t}\mathcal{I}^{\alpha}\varphi(t):=\frac{\partial }{\partial t}\int_0^t\omega_{\alpha}(t-s)\varphi(s)\,ds\quad\text{with} \quad
\omega_{\alpha}(t):=\frac{t^{\alpha-1}}{\Gamma(\alpha)},
\end{equation}
($\mathcal{I}^{\alpha}$ is the Riemann--Liouville time-fractional integral). As $\alpha \rightarrow 1^-$, $\partial_t^{1-\alpha}\;( {\cal{L}}u)$ converges
to $ {\cal{L}}u$, and thus, problem \eqref{a} reduces to the classical diffusion equation.
Throughout the paper, we assume that the source term $f$ and the diffusivity coefficient function ${\bf a}$ are sufficiently regular and
\begin{equation}\label{eq: A positive}
0<a_{\min} \le {\bf a}(x) \le a_{\max}<\infty\quad { \rm on
} ~~\overline \Omega.\end{equation}
Several numerical techniques for the problem \eqref{a} (with constant diffusivity coefficient) in one and several space variables
have been proposed with various types of spatial discretizations including finite difference, FE
or spectral element methods, see \cite{CockburnMustapha2015, KMP2015, JLPZ2015, MustaphaMcLean2011}. For the time discretization,
different time-stepping schemes (implicit and explicit) have been investigated including finite difference, convolution quadrature,
and discontinuous Galerkin methods, see \cite{ChenLiuAnhTurner2012, CuestaLubichPalencia2006, Cui2012, Mustapha2011, MustaphaMcLean2011, RZ2013, ZS2011}. The error analyses in most studies in the existing
literature typically assume that the solution $u$ of \eqref{a} is sufficiently regular
including at $t = 0$, which is not practically the case. Indeed, assuming high
regularity on $u$ imposes additional compatibility conditions on the given data,
which are not reasonable in many cases.
Though the numerical approximation of the solution $u$ of \eqref{a} was considered by many authors
over the last decade,
the optimality of the estimates with respect to the solution smoothness expressed through the problem data,
$f$ and $u_0$, was considered in a few papers for the case of constant diffusivity and quasi-uniform FE meshes. Obtaining sharp error bounds under reasonable regularity
assumptions on $u$ has proved challenging. McLean and Thom\'ee \cite{MT2010}
established the first optimal $L^2(\Omega)$-error estimates for the Galerkin FE solution of \eqref{a}
with respect to the regularity of initial data. More precisely, for $t\in(0,T]$, convergence rates of order $t^{\alpha(\delta-2)/2} h^2$ ($h$
denoting the maximum diameter of the spatial mesh elements) were proved assuming that the initial data $u_0 \in \dot H^\delta(\Omega)$ for $\delta=0,\,2$ (see, Section \ref{sec:WRT} for the definition of these spaces). The proof was based on some refined estimates of the Laplace transform in time for the error. In \cite{McLeanThomee2010} and by using a similar approach, the same authors derived $O(t^{-\alpha(2-\delta)/2}h^2|\log h|^2)$ convergence rates in the stronger $L^\infty(\Omega)$-norm. For $\delta=0$, $u_0$ assumed to be in $L^\infty(\Omega)$, while $u_0 \in C^2(\Omega) \cap C_0(\overline \Omega)$ for $\delta=2$.
Recently, McLean and Mustapha \cite{McLeanMustapha2015} studied the error analysis of a first order
semidiscrete time-stepping scheme for problem \eqref{a} with $f\equiv 0$ allowing
nonsmooth $u_0,$ using discrete Laplace transform technique. Since standard energy
arguments are used heavily in the error analysis of Galerkin FEs for classical diffusion equations, it is more
pertinent to extend the analysis to these time-fractional order diffusion problems with a variable diffusivity. Since $t^m$ and $\partial_t^{1-\alpha}{\cal{L}} $ do not commute, extending these arguments to problem \eqref{a} is not a straightforward task, especially in the case of nonsmooth $u_0$.
The main motivation of this work is to derive optimal error estimates of the semidiscrete Galerkin FE
method for the problem \eqref{a} for both smooth and nonsmooth initial data using
energy arguments. Since the solution $u$ has limited smoothing property \cite{Mclean2010}, a repeated use of the integral operator like $\mathcal{I}\phi(t):=\int_{0}^t \phi(s)\;ds$ (see \cite{GP2011,GPY2014}) along with $t^m$ type weights is an essential
tool to provide optimal error
estimates. Earlier, for smooth initial data ($u_0 \in \dot H^2(\Omega)$), error analysis of the Galerkin FE method for the
fractional diffusion equation \eqref{a} was considered in \cite{MustaphaMcLean2011} using energy arguments
and they have derived quasi-optimal error estimates of order $O(h^2|\log h|)$ in $L^{\infty} (L^2)$-norm.
Below, we briefly summarize our main results obtained in this article: for $t\in (0,T]$ and when $u_0 \in \dot H^\delta(\Omega)$ with $0\le \delta\le 2$,
\begin{itemize}
\item {\it a priori} optimal error estimate in $L^2(\Omega)$-norm of order $t^{-\alpha(2-\delta)/2}h^2$ is established, see Theorem \ref{thm: smooth and nonsmooth}. Consequently, for a quasi-uniform FE mesh, an $O(t^{-\alpha(2-\delta)/2}h)$ error estimate in $H^1(\Omega)$-norm is obtained, see Remark \ref{rem: 1}. By dropping the quasi-uniformity mesh assumption, we showed that this error bound remains valid for $0\le \delta\le 1,$ see Theorem \ref{thm: H1 bound}. However, for $1<\delta\le 2,$ we derived an $O(t^{-\alpha(1-\delta)/2}\max\{1,(h/t)^{-\alpha(1-\delta)/2}\})$ error estimate, which is reduced to $t^{-\alpha(2-\delta)/2}h^2$ for $t \ge Ch$.
\item For $\Omega \subset \mathbb{R}^2$, and assuming that $u_0 \in L^\infty(\Omega)$ and the FE mesh is quasi-uniform, a quasi-optimal error estimate of order $t^{-\alpha(3-\delta)/2}|\ln h|^{5/2}\, h^2$ in the stronger $L^\infty(\Omega)$-norm is proved, see Theorem \ref{thm: smooth and nonsmooth-5}.
\end{itemize}
The proposed technique has several attractive features. Some of these are: (1) allowing variable coefficients, and smooth and nonsmooth initial data in the error analysis, (2) the quasi-uniform FE mesh assumption is not required to show the convergence results in $H^m$-norm for $m=0,1$, and (3) the proposed technique can be applied to other fractional model problems with
smooth and nonsmooth initial data. For instance, we discussed in Section \ref{sec: Caputo} the extension of the achieved error bounds in Theorems \ref{thm: smooth and nonsmooth}, \ref{thm: H1 bound} and \ref{thm: smooth and nonsmooth-5} to the time-fractional diffusion equation: for $0<\alpha<1$,
\begin{equation}\label{D}
{^C}\partial_t^{\alpha} u(x,t)
+ {\cal{L}} u(x,t)= f(x,t),
\end{equation}
where ${^C}\partial_t^{\alpha} v(t):=\mathcal{I}^{1-\alpha}v'(t)$ is the time-fractional Caputo derivative. The error estimate in Theorem \ref{thm: smooth and nonsmooth} provides an improvement of the result obtained by Jin {\it et al.} \cite[Theorem 3.7]{JLZ2013}, where the error analysis of the lumped mass FE method was also considered. For constant diffusivity and under the assumption that the mesh is quasi-uniform, the FE error bound in \cite{JLZ2013} involves a logarithmic factor which was derived using a semigroup approach. The error analysis approach in this
work can also be used to investigate the error estimates for the FE method applied
to the time-fractional Rayleigh-Stokes problem (described by the time-fractional differential equation) presented in the recent work \cite{EJLZ2016}, which is close to the one in \cite{MT2010}.
Outline of the paper. In Section \ref{sec:WRT}, we recall some smoothness properties of the solution $u$, we also state and derive some technical results.
In Section \ref{sec:Semi-discrete FE}, we introduce our semidiscrete FE scheme and recall some FE error results. We claim that a direct application of energy arguments to problem \eqref{a} does not lead to optimal convergence rates even when the initial data $u_0 \in \dot H^2(\Omega)$. In Section \ref{sec: LinftyL2}, for both smooth and nonsmooth initial data, we derive error estimates for the FE problem in the $L^2(\Omega)$-norm, see Theorem \ref{thm: smooth and nonsmooth}. The generalized Leibniz formula is an essential ingredient in our error analysis.
In Section \ref{sec: LinftyH1}, a superconvergence result in the $H^1$-norm
is obtained, see Theorem \ref{sup-conv-1}, and as a consequence, an optimal gradient FE error estimate in the $L^2(\Omega)$-norm is derived in Theorem \ref{thm: H1 bound}, and a quasi-optimal FE error bound in the $L^\infty(\Omega)$-norm is achieved for $\Omega \subset \mathbb{R}^2$, see Theorem \ref{thm: smooth and nonsmooth-5}.
Particularly relevant
to this {\it a priori} error analysis is the appropriate use of several properties of the
time-fractional integral and derivative operators.
In Section \ref{sec: Caputo}, we show that the achieved error estimates for problem \eqref{a} are valid for the FE discretization of the fractional diffusion model \eqref{D}. Numerical tests are presented in Section \ref{sec: Numerical} to
confirm some of our theoretical findings.
Throughout the paper, $C$ is a generic positive constant that may depend on $\alpha$ and $T$, but is independent of the spatial mesh element size $h$.
\section {Regularity and technical results}\label{sec:WRT} Smoothness properties of the solution $u$ of the fractional diffusion problem \eqref{a} play a key role in the error analysis of the Galerkin FE method,
particularly, since $u$ has singularity near $t=0$, even for smooth given data. Below, we state the required regularity results for problem \eqref{a} in terms of the initial data $u_0$ and the source term $f$. Over the convex domain $\Omega$, by combining the results of Theorems 4.1, 4.2 and 5.6 in \cite{Mclean2010}, for $0\le r,\,\mu \le 2$, we obtain
\begin{equation}\label{eq: regularity property}
\begin{aligned}
t^\ell\|u^{(\ell)}(t)\|_{r+\mu}&\le Ct^{-\alpha\mu/2}\Big[\|u_0\|_r
+(1+t^{\alpha\mu/2})\sum_{m=0}^{\ell+1}\int_0^t s^m\|f^{(m)}(s)\|_r\,ds\Big]\\
&\le C(1+T^{\alpha\mu/2}) t^{-\alpha\mu/2}d_r(u_0,f),
\end{aligned}
\end{equation}
for $\ell\in \{0,1\}$,
where
$d_r(u_0,f)=\|u_0\|_{r}
+\sum_{m=0}^2\int_0^T s^m\|f^{(m)}(s)\|_{r}\,ds.$ Here, $\|\cdot\|_r$ denotes the norm on the Hilbert space $\dot H^r(\Omega)$ defined by
\[\|v\|_r^2 =\|A^{r/2}v\|^2 =\sum_{j=1}^\infty \lambda_j^r (v,\phi_j)^2,\]
where $\{\lambda_j\}_{j=1}^\infty$ (with $0<\lambda_1< \lambda_2< \ldots$) are the eigenvalues of the
operator elliptic operator ${\cal{L}}$ (subject homogeneous Dirichlet boundary conditions)
and $\{\phi_j\}_{j=1}^\infty$ are the associated orthonormal eigenfunctions.
Noting that, $\dot H^r(\Omega)=H^r(\Omega)$ for $0\le r<1/2,$ and for convex polygonal domains, $\dot H^r(\Omega)=H^r(\Omega)\cap H^1_0(\Omega)$ for $1/2<r < 5/2,$ where $H^r(\Omega)$ (with $H^0(\Omega)=L^2(\Omega)$) is the standard Sobolev space.
Next, we state the positivity properties of the fractional operators $\mathcal{I}^{\alpha}$ and $\partial_t^{1-\alpha}$, and derive some technical results that will be used in the subsequent sections. By \cite[Lemma 3.1(ii)]{MustaphaSchoetzau2014} and since the bilinear form $A(\cdot, \cdot)$ associated with the operator ${\cal{L}}$ (that is, $A(v,w)=({\bf a}\nabla v, \nabla w)$) is symmetric
positive definite on the Sobolev space $H_0^1(\Omega)$, it follows that for piecewise time continuous functions $\varphi:[0,T] \to H_0^1(\Omega),$
\begin{equation}\label{eq: positive of Ia}
\int_0^TA(\mathcal{I}^\alpha\varphi,\varphi)\,dt\ge \cos(\alpha \pi/2)\int_0^T\|\sqrt{ a}\,\nabla \mathcal{I}^{\alpha/2}\varphi\|^2\,dt \ge 0~~{\rm for}~~0<\alpha<1,
\end{equation}
where $\|\varphi\|:=\sqrt{(\varphi,\varphi)}$ denotes the $L^2$-norm.
By \cite[Lemma A.1]{McLean2012} and again since the bilinear form $A(\cdot, \cdot)$ is symmetric
positive definite, the following holds: for
$W^{1,1}(0,T;H_0^1(\Omega))$,
\begin{equation}\label{eq: positive of Ba}
\int_0^TA(\partial_t^{1-\alpha}\varphi(t),\varphi(t))\,dt \ge \frac{1}{2}\sin(\alpha\pi/2) T^{\alpha-1}
\int_0^T\| \sqrt{{\bf a}} \nabla \varphi(t)\|^2\,dt\,.
\end{equation}
The next lemma will be used frequently in our convergence analysis. In the proof, we use the following integral inequality:
if for any $\tau\in(0,t)$, $ |\phi(\tau)|^2\le |\phi(0)|^2+2\,\int_{0}^\tau|\phi(s)|\,|\psi(s)|\,ds, $ then $|\phi(t)|\le |\phi(0)|+\int_0^t\,|\psi(s)|\,ds.$
\begin{lemma}\label{lem: reg use} Let $\kappa \in \{0,1\}$ and let ${\mathcal B}^\alpha=\partial_t^{1-\alpha}$ or ${\mathcal B}^\alpha=\mathcal{I}^\alpha.$ Assume that
\begin{equation} \label{eq: reg 1}
\kappa(v(t),\chi)+(1-\kappa)(v'(t),\chi)+A({\mathcal B}^\alpha v(t),\chi)=
(w(t),\chi),\quad \forall~ \chi \in V_h,
\end{equation}
for $t \in (0,T]$.
Then
\[
\kappa\int_0^t \|v\|^2\,ds+(1-\kappa)\|v(t)\|^2\le
(1-\kappa)\Big(\|v(0)\|+\int_0^t \|w\|\,ds\Big)^2+\kappa\int_0^t \|w\|^2\,ds.\]
\end{lemma}
\begin{proof}
Choose $\chi=v$ in \eqref{eq: reg 1}, and then, integrate over the interval $(0,t)$ to obtain
\[
2\kappa\int_0^t \|v\|^2\,ds+(1-\kappa)[\|v(t)\|^2-\|v(0)\|^2]+2\int_0^t A({\mathcal B}^\alpha v,v)ds=
2\int_0^t (w,v)\,ds.\]
Now, use the positivity properties of $\mathcal{I}^\alpha$ in \eqref{eq: positive of Ia} and of $\partial_t^{1-\alpha}$
in \eqref{eq: positive of Ba} to find that
\[2\kappa\int_0^t \|v\|^2\,ds+(1-\kappa)\|v(t)\|^2\le
(1-\kappa)\|v(0)\|^2+2\int_0^t \|w\| \,\|v\|\,ds.
\
For $\kappa=0$, an application of the integral inequality (stated above) yields the desired inequality. However, for $\kappa=1$, we use the inequality $2\|w\| \,\|v\|\le \|w\|^2+ \|v\|^2$ and the desired inequality follows after simplifying.
\end{proof}
\section {Semi-discrete FE method} \label{sec:Semi-discrete FE}
This section focuses on a semidiscrete Galerkin FE scheme for problem \eqref{a}. To define the scheme, let $\mathcal{T}_h$ be a
family of regular triangulations (made of simplexes $K$) of
the domain $\overline{\Omega}$ and let $h=\max_{K\in \mathcal{T}_h}(\mbox{diam}K),$ where $h_{K}$ denotes
the diameter of the element $K.$ The FE space
$V_h$ on $\mathcal{T}_h$ is given by
$$V_h=\{v_h\in C^0(\overline {\Omega})\;:\;v_h|_{K}\;\mbox{is linear for all}~
K\in \mathcal{T}_h\; \mbox{and} \; v_h|_{\partial \Omega}=0\}.$$
The weak formulation for problem \eqref{a} is
to find $u:( 0,T]\longrightarrow H^1_0$ such that
\begin{equation} \label{weak}
(u',v )+ A(\partial_t^{1-\alpha} u,v )= (f,v )\quad
\forall v\in H^1_0
\end{equation}
with given $ u(0)=u_0.$ Thus, the standard semidiscrete FE formulation for \eqref{a} is
to seek $u_h:(0,T]\longrightarrow V_h$ such that
\begin{equation} \label{semi}
(u_h',v_h)+ A(\partial_t^{1-\alpha} u_h,v_h)= (f,v_h)\quad
\forall v_h\in V_h
\end{equation}
with given $u_h(0)\in V_h$ to be defined later.
To derive {\it a priori} error estimates for the FE scheme (\ref{semi}), we split the error
$$e:=(u- R_h u)-( u_h-R_hu)=:\rho-\theta,$$
where the Ritz projection $R_h : H_0^1(\Omega) \rightarrow V_h $ is defined by the following relation:
$A(R_h v-v, \chi)= 0$ for all $\chi\in V_h.$
For $t\in (0,T]$, the projection error $\rho(t)$ satisfies the following estimates
\cite{thomee1997}: for $\ell=0,1,$
\begin{equation}\label{rho-estimate}
\|\rho^{(\ell)}(t)\|_j\leq C h^{m-j} \|u^{(\ell)}(t)\|_m,\qquad~~j=0,1,~~m=1,2.
\end{equation}
and hence, by using the regularity property in \eqref{eq: regularity property}, we observe
\begin{equation}\label{eq: rho}
\|\rho^{(\ell)}(t)\|
\le Ch^m t^{-\ell-\max\{0,\alpha(m-\delta)/2\}}d_\delta(u_0,f),\quad {\rm for}~~0\le \delta\le 2.
\end{equation}
Next, we show that a direct application of energy arguments to problem (\ref{a}) does not yield
satisfactory results due to the low regularity of the continuous solution.
From (\ref{weak}) and (\ref{semi}), the decomposition $e=\rho-\theta$, and the property of the
elliptic projection, we obtain the equation in $\theta$ as
\begin{equation} \label{sup-2}
(\theta',\chi)+A(\partial_t^{1-\alpha}\theta,\chi)=
(\rho',\chi)\quad \forall~\chi \in V_h.
\end{equation}
Then, the following result holds, whose proof can be found in \cite{MT2010, MustaphaMcLean2011}.
\begin{theorem}\label{th1} For $t\in (0,T],$ we have
\begin{equation}\label{e1}
\|u(t)-u_h(t)\|\leq
\|\theta_0\|+\int_0^t \|\rho'(s)\|\,ds+\|\rho(t)\|.
\end{equation}
\end{theorem}
Noting that, the solution $u$ of problem \eqref{a} has singularity near $t=0$. For instance, if $f\equiv 0$ and
$u_0 \in \dot H^2(\Omega)$, then by (\ref{rho-estimate}) and the
regularity property $\|u'(t)\|_m\le Ct^{\alpha(1-m/2)-1}\|u_0\|_2$ for $m=1,2,$
\begin{align*}
\int_0^t \|\rho'(s)\|\,ds
&\le Ch\int_0^\epsilon \|u'(s)\|_1\,ds+Ch^2\int_\epsilon^t \|u'(s)\|_2\,ds\\
&\le C\Big(h\int_0^\epsilon s^{\alpha/2-1}\,ds+h^2\int_\epsilon^t s^{-1}\,ds\Big)\|u_0\|_2\\
&\le C\Big(h \epsilon^{\alpha/2} + h^2|\log \epsilon |)\|u_0\|_2 \le Ch^2|\log h|\|u_0\|_2\quad{\rm for}~~\epsilon=h^{2/\alpha}.
\end{align*}
This leads to a quasi-optimal $O(h^2|\log h|)$ convergence for the spatial discretization by linear FEs. To achieve an optimal $O(h^2)$ convergence, a stronger
regularity assumption on $u$ is required and that, in turn imposes severe restrictions on initial data. Thus, the upper bound in Theorem \ref{th1} is not sharp even for the case of smooth initial data, that is, $H^2$-regularity on the initial data $u_0$ is not sufficient to get an optimal $O(h^2)$ convergence rate. Furthermore, it is clear that this upper bound is not suitable for the case of nonsmooth initial data. Therefore, we propose in the next section an approach via delicate energy arguments that provides valid error bounds for the problem with both smooth and nonsmooth initial data.
\section {$L^2(\Omega)$-error estimates} \label{sec: LinftyL2}
For convenience, we introduce the notations:
\[\Theta_i(t):=t^i\theta(t)\quad{\rm and}\quad \dot \Theta_i(t):=t^i\theta'(t)~~{\rm for}~~i=1,\,2.\]
In the next lemma, based on the generalized Leibniz formula for fractional derivatives, we state and show some identities
for our subsequent use.
\begin{lemma}\label{Ia}
For $0<\alpha<1$, the followings hold: \\
\mbox{(a)} $t\partial_t^{1-\alpha}\theta= \partial_t^{1-\alpha}\Theta_1-(1-\alpha)\mathcal{I}^\alpha \theta$,\\
\mbox{(b)} $t\mathcal{I}^\alpha\theta = \mathcal{I}^\alpha\Theta_1+\alpha\mathcal{I}^{1+\alpha}\theta$.
\end{lemma}
\begin{proof} The first identity follows from the fractional Leibniz formula. To show the second identity, noting first that $\partial_t^{1-\alpha}\Theta_1=\mathcal{I}^\alpha \Theta'_1=\mathcal{I}^\alpha\theta+\mathcal{I}^\alpha \dot \Theta_1.$ Hence, by (a),
\begin{equation}\label{k3a}
t\partial_t^{1-\alpha} \theta(t)= \mathcal{I}^\alpha \dot \Theta_1(t)+\alpha\mathcal{I}^\alpha\theta(t).
\end{equation}
Now, we replace $\theta$ by $\mathcal{I}\theta$ in (\ref{k3a}) to obtain the second identity in the lemma.
\end{proof}
Next, we derive an upper bound of $\Theta_1$. To do so, we let $u_h(0)=P_h u_0,$ where $P_h :L^2(\Omega)\rightarrow V_h$ denotes the $L^2$-projection defined by $(P_h v-v, \chi)= 0$ for all $\chi\in V_h.$
\begin{lemma}\label{lem: estimate of Theta in l2 norm}
Let $u_h(0)=P_h u_0.$ Then, we have
\[
\int_0^t\|\Theta_1\|^2\,ds
\le 3\int_0^t \Big( s^2\|\rho\|^2+ 2\big(\mathcal{I}\|\rho\|\big)^2\Big)\,ds.
\]
\end{lemma}
\begin{proof}
We integrate
(\ref{sup-2}) over the time interval $(0,t)$ and obtain
\begin{equation} \label{sup-3}
(\theta,\chi)+A(\mathcal{I}^\alpha\theta,\chi)=
(\rho+e(0),\chi)\quad \forall~\chi \in V_h.
\end{equation}
But $(e(0),\chi)= 0$ because $u_h(0)=P_h u_0.$ Therefore,
\begin{equation} \label{sup-3 ph}
(\theta,\chi)+A(\mathcal{I}^\alpha\theta,\chi)=(\rho,\chi)\quad \forall~\chi \in V_h.
\end{equation}
Multiply by $t$ and use $t\mathcal{I}^\alpha\theta= \mathcal{I}^\alpha\Theta_1+\alpha\mathcal{I}^{1+\alpha}\theta$
by Lemma \ref{Ia}\,(b) to find that
\[
(\Theta_1,\chi)+A(\mathcal{I}^\alpha\Theta_1,\chi)=
t(\rho,\chi)-\alpha A(\mathcal{I}^{1+\alpha}\theta,\chi)\quad \forall~\chi \in V_h.
\]
However, from \eqref{sup-3 ph}, we get
\begin{equation}\label{eq: I alpha+1}
A(\mathcal{I}^{\alpha+1}\theta,\chi)=
(\mathcal{I}(\rho-\theta),\chi)\quad \forall~\chi \in V_h,
\end{equation}
and thus,
\begin{equation} \label{sup-3-2}
(\Theta_1,\chi)+A(\mathcal{I}^\alpha\Theta_1,\chi)=
t(\rho,\chi)-\alpha (\mathcal{I}(\rho-\theta),\chi)\quad \forall~\chi \in V_h.
\end{equation}
Consequently, an application of Lemma \ref{lem: reg use} (with $\kappa=1$) yields
\begin{equation}\label{eq: second last step}
\begin{aligned}
\int_0^t\|\Theta_1\|^2\,ds
&\le \int_0^t \|s\rho-\alpha\mathcal{I}(\rho-\theta)\|^2\,ds
\le 3\int_0^t ( s^2\|\rho\|^2+ \|\mathcal{I}\rho\|^2+\|\mathcal{I}\theta\|^2)\,ds.
\end{aligned}
\end{equation}
To complete our proof, we rewrite \eqref{eq: I alpha+1} as
\[
(\mathcal{I}\theta,\chi)+A(\mathcal{I}^\alpha(\mathcal{I}\theta),\chi)=
(\mathcal{I}\rho,\chi)\quad \forall~\chi \in V_h,
\]
Again, an application of Lemma \ref{lem: reg use} (with $\kappa=1$) shows
\begin{align}\label{eq: theta less rho}
\int_0^t \|\mathcal{I}\theta\|^2\,ds
&\le \int_0^t \|\mathcal{I}\rho\|^2\,ds.
\end{align}
Substitute (\ref{eq: theta less rho}) in \eqref{eq: second last step} yields the desired bound.
\end{proof}
An upper bound of the term $\theta$ will be derived in the next lemma. Again, for convenience, we introduce the following notation
\begin{equation}\label{def: B1}
{\mathcal B}_1(t):=\int_0^t \Big( s^4\|\rho'(s)\|^2+s^2\|\rho(s)\|^2+ 2\big(\mathcal{I}\|\rho(s)\|\big)^2 \Big)\,ds\,.\end{equation}
For later use, by using the projection error estimates in \eqref{eq: rho} (with $\ell=0,1$ and $m=2$) for upper bounds of $\rho$ and $\rho'$, and then integrating, we find that for $t\in (0,T]$,
\begin{equation}\label{eq: bound of B1}
{\mathcal B}_1(t)\leq C\,h^4 t^{3-\alpha(2-\delta)}d_\delta^2(u_0,f),\quad {\rm for}~~0\le \delta\le 2\,.
\end{equation}
\begin{lemma}
\label{H1}
Let $u_h(0)=P_h u_0$. Then, the following estimate holds
\[
\|\theta(t)\|^2 \leq C\, t^{-3} {\mathcal B}_1(t),\quad {\rm for}~~t\in (0,T]\,.
\]
\end{lemma}
\begin{proof} We multiply (\ref{sup-2}) by $t^2$ so that
\begin{equation}\label{k-3}
(\dot\Theta_2,\chi)+A(t^2\partial_t^{1-\alpha}\theta,\chi)=
(t^2\rho', \chi),
\end{equation}
where $\dot\Theta_2= t^2 \theta'.$ From the fractional Leibniz formula, we have
\begin{align*}
t^2\partial_t^{1-\alpha}\theta&=\partial_t^{1-\alpha}\Theta_2- 2(1-\alpha)t\mathcal{I}^{\alpha} \theta+\alpha(1-\alpha)\mathcal{I}^{1+\alpha}\theta.
\end{align*}
Hence, we rearrange (\ref{k-3}) as
\begin{equation}\label{k2-1}
(\dot\Theta_2,\chi)+A(\partial_t^{1-\alpha}\Theta_2,\chi)=
(t^2\rho', \chi)+ (1-\alpha)\Big(2tA(\mathcal{I}^{\alpha} \theta,\chi)-\alpha A(\mathcal{I}^{1+\alpha}\theta,\chi)\Big),
\end{equation}
and then, by equations \eqref{sup-3 ph} and \eqref{eq: I alpha+1},
\begin{equation}\label{k3}
(\Theta_2',\chi)+A(\partial_t^{1-\alpha}\Theta_2,\chi)=
(t^2\rho'+ 2\alpha \Theta_1+ (1-\alpha)(2t\rho-\alpha\mathcal{I} (\rho-\theta)),\chi).
\end{equation}
Hence, by Lemma \ref{lem: reg use} (with $\kappa=0$), we obtain
\begin{align*
\|\Theta_2(t)\| &\leq \int_0^{ t} \Big(s^2\|\rho'(s)\|+ 2s\|\rho(s)\|
+2\|\Theta_1(s)\|+\|\mathcal{I}(\rho-\theta)\|\Big)\,ds\,,
\end{align*}
and thus, an application of the Cauchy-Schwarz inequality yields
\begin{align*
\|\Theta_2(t)\|^2 &\leq Ct\int_0^t \Big( s^4\|\rho'(s)\|^2+s^2\|\rho(s)\|^2+\|\Theta_1(s)\|^2+\|\mathcal{I}\rho\|^2+\|\mathcal{I}\theta\|^2\Big)\,ds\,.
\end{align*}
Therefore, by using the identity $\theta(t)=t^{-2}\Theta_2(t)$, the inequality in \eqref{eq: theta less rho} and Lemma \ref{lem: estimate of Theta in l2 norm} will complete the rest of the proof.
\end{proof}
In the next theorem, we derive optimal convergence results of the FE scheme \eqref{semi}
in the $L^2$-norm for both smooth and nonsmooth initial data $u_0$. For $u_0 \in \dot H^{\delta}(\Omega)$ with $0\le \delta\le 2,$ we show that the error is bounded by $C h^2 t^{-\alpha(2-\delta)/2}$ for each $t\in (0,T]$. Recall that, $\dot H^{\delta}(\Omega)=\{v\in H^{\delta}(\Omega):~v=0~{\rm on}~\partial \Omega\}$ for $1/2<\delta\le 2$, while $\dot H^{\delta}(\Omega)=H^{\delta}(\Omega)$ for $0\le \delta <1/2$. Noting that, in the limiting case $\alpha \rightarrow 1^{-}$, we recover the convergence rates for the parabolic equation $u'-{\cal{L}} u=f$.
\begin{theorem} \label{thm: smooth and nonsmooth}
Let $u$ and $u_h$ be the solutions of $(\ref{a})$ and $(\ref{semi})$,
respectively, with $u_h(0)=P_h u_0$. Then, for $u_0 \in \dot H^\delta(\Omega),$
$$
\|(u-u_h)(t)\| \leq
C h^2 t^{-\alpha(2-\delta)/2}d_\delta(u_0,f)\quad{\rm for}~~t \in (0,T]~~{\rm with}~~0\le \delta \le 2\,.
$$
\end{theorem}
\begin{proof} The desired result follows from the decomposition $u-u_h=\rho-\theta$, the estimate of $\theta$ in Lemma \ref{H1}, the bound in \eqref{eq: bound of B1}, and the estimate of $\rho$ in \eqref{eq: rho}.
\end{proof}
\begin{remark}\label{rem: 0}
In the proof of the above theorem, we used \eqref{eq: bound of B1} which follows from the projection estimate in \eqref{eq: rho} for $m=2.$ For $m=1$, we follow similar steps where $2-\delta$ will be replaced with $1-\delta$, to obtain ${\mathcal B}_1(t)\leq C\,h^2 t^{3-\alpha(1-\delta)}d_\delta^2(u_0,f)$ for $0\le \delta\le 1\,.$ Now, for $1<\delta\le 2$, we notice first that for $0\le q\le t$ with $t_h=\max\{t,h\}$,
$$\|\rho(q)\|\le \|\rho(t_h)-\rho(q)\|+\|\rho(t_h)\|\le \mathcal{I}(\|\rho'(t_h)\|)+\|\rho(t_h)\|\,.$$
Substitute this in the definition of ${\mathcal B}_1$ defined in \eqref{def: B1}, we observe
\[{\mathcal B}_1(t)\le Ct^2\int_0^t s^2\|\rho'(s)\|^2\,ds+Ct^3\Big(\mathcal{I}(\|\rho'(t_h)\|) +\|\rho(t_h)\|\Big)^2\,.\]
To estimate the first two terms, we use \eqref{rho-estimate} (with $\ell,m=1$) and the following regularity property (which follows from \cite[Theorems 4.2 and 5.6]{Mclean2010})
\begin{equation}\label{eq: regularity property-2}
\|u'(t)\|_1\le Ct^{\alpha (\delta-1)/2-1}\tilde d_\delta(u_0,f),\quad t\in(0,T],\quad{\rm for}~~1<\delta\le 2,
\end{equation}
where $\tilde d_\delta(u_0,f)=\|u_0\|_\delta +\sum_{m=0}^2 t^m\|f^{(m)}\|_{L^\infty(\dot H^1)},$ we arrive to
\[{\mathcal B}_1(t)\le Ct^3 h^2 t_h^{\alpha(\delta-1)}\tilde d_\delta(u_0,f) +Ct^3\|\rho(t_h)\|^2,\quad{\rm for}~~1<\delta\le 2\,.\]
However, by \eqref{eq: rho} and the inequality $t_h^{\alpha(\delta-1)/2}\le C t^{-\alpha(1-\delta)/2}(t_h/t)^{\alpha(\delta-1)/2}$ for $1<\delta\le2$, we find that
\[\|\rho(t_h)\| \le Ch^2 t_h^{-\alpha(2-\delta)/2}d_\delta(u_0,f)\le C h t^{-\alpha(1-\delta)/2}(t_h/t)^{\alpha(\delta-1)/2}d_\delta(u_0,f).\]
Therefore,
\[{\mathcal B}_1(t)\le C h^2 t^{3+\alpha(\delta-1)}(t_h/t)^{\alpha(\delta-1)}\tilde d^2_\delta(u_0,f),\quad{\rm for}~~1<\delta\le 2\,.\]
Consequently, by using the above bound of ${\mathcal B}_1$ in Theorem \ref{thm: smooth and nonsmooth}, we get the error estimate below that will be used in the forthcoming section to show the convergence of the gradient FE solution. For $t \in (0,T],$
\begin{equation} \label{thm: smooth and nonsmooth m=1}
\|e(t)\| \leq
C h t^{-\alpha(1-\delta)/2}D_{\delta,\alpha}(u_0,f,h/t),
\end{equation}
where
\[
D_{\delta,\alpha}(u_0,f,h/t)= \tilde d_\delta(u_0,f)\times\begin{cases} 1~~~&{\rm for}~~0\le \delta \le 1,\\
(t_h/t)^{\alpha(\delta-1)/2}~~~&{\rm for}~~1< \delta \le 2\,.\end{cases}
\]
\end{remark}
\begin{remark}\label{rem: 1}
Under the quasi-uniformity condition on $V_h$, for $t \in (0,T]$, from the decomposition $u-u_h=\rho-\theta$, the inverse inequality, the estimate of $\theta$ in Lemma \ref{H1}, and the estimate $\|\rho(t)\|_1 \le Ch\|u(t)\|_2\le Ct^{-\alpha(2-\delta)/2}d_\delta(u_0,f)$ (follows from the Ritz projection bound in \eqref{rho-estimate} (with $j=1$ and $m=2$) and the
regularity property (\ref{eq: regularity property}), we obtain the following optimal error estimate in the $H^1(\Omega)$-norm:
$$
\|\nabla(u-u_h)(t)\| \leq
C h t^{-\alpha(2-\delta)/2}d_\delta(u_0,f)\quad{\rm for}~~t \in (0,T]~~{\rm with}~~0\le \delta \le 2\,.
$$
By removing the quasi-uniformity mesh assumption, this error bound remains valid for $0\le \delta\le 1,$ see Theorem \ref{thm: H1 bound}.
\end{remark}
\begin{remark}\label{rem: 2} For smooth initial data $u_0\in {\dot H}^2(\Omega)$, one may choose $u_h(0)=R_hu_0$. An optimal convergence rate can be shown by following the proof of Theorem \ref{thm: smooth and nonsmooth} line-by-line, where the term $\rho$ in Lemma \ref{lem: estimate of Theta in l2 norm} should be replaced with $\tilde \rho:=\rho+e(0)$.\end{remark}
\section {$H^1(\Omega)$- and $L^\infty(\Omega)$-error estimates}\label{sec: LinftyH1}
In this section, for each $t \in (0,T]$, we show optimal convergence results for the gradient FE error in the $L^2(\Omega)$-norm, and quasi-optimal error bounds for the FE error in the $L^\infty(\Omega)$-norm, for both smooth and nonsmooth initial data $u_0.$ We start our analysis by deriving an upper bound of $\nabla \Theta_1$.
\begin{lemma}\label{theta}
For $0\le \delta \le 2$ and for $t\in(0,T]$, we have
\[
\int_0^{t} \|\nabla\Theta_1\|^2ds \leq C h^4 t^{3 -\alpha(3-\delta)}d^2_\delta(u_0,f)\,.\]
\end{lemma}
\begin{proof} Multiplying \eqref{sup-2} by $t$ and then using the identity $t\partial_t^{1-\alpha} \theta=\mathcal{I}^\alpha \dot \Theta_1+\alpha\mathcal{I}^\alpha\theta$ (Lemma \ref{Ia} (a)), we obtain
\begin{equation}\label{k2}
(\dot \Theta_1,\chi)+A(\partial_t^{1-\alpha}\Theta_1,\chi)=
(t\rho', \chi)+ (1-\alpha)A(\mathcal{I}^\alpha\theta,\chi).
\end{equation}
Then, a use of (\ref{sup-3 ph}) yields after simplifying
\begin{equation}\label{k3-1}
(\Theta_1',\chi)+A(\partial_t^{1-\alpha}\Theta_1,\chi)=
(t\rho',\chi)+\alpha(e, \chi)+(\rho, \chi)\,.
\end{equation}
Now, set $\chi=\Theta_1$ in \eqref{k3-1}, integrate the resulting equation over $(0,t)$, and use the positivity property of $\partial_t^{1-\alpha}$ in \eqref{eq: positive of Ba}, to find that
\begin{align*}\|\Theta_1(t)\|^2+\frac{1}{2}\sin(\alpha\pi/2) t^{\alpha-1}
\int_0^t\|\sqrt{{\bf a}} \nabla \Theta_1\|^2\,ds &\le
\int_0^t (s\|\rho'\|+\|e\|+\|\rho\|)\|\Theta_1\|\,ds.\end{align*}
By the integral inequality (stated before Lemma \ref{lem: reg use}), we observe
\begin{equation}\label{eq: bound theta}
\frac{1}{2}\sin(\alpha\pi/2) t^{\alpha-1}
\int_0^t\| \sqrt{{\bf a}}\nabla \Theta_1\|^2\,ds
\le \frac{1}{4} \Big(\int_0^t (s\|\rho'\|+\|e\|+\|\rho\|)ds\Big)^2.
\end{equation}
Therefore, the desired estimate follows from this bound, the error projection in \eqref{eq: rho} (for $\ell=0,1$ and with $m=2$), the achieved convergence results in Theorem \ref{thm: smooth and nonsmooth}, and the assumption on the diffusivity coefficient in \eqref{eq: A positive}.
\end{proof}
In the next theorem, we derive an error bound for $\nabla \theta(t)$ in the $L^2(\Omega)$-norm.
\begin{theorem}\label{sup-conv-1}
For $0\le \delta\le 2$, we have
$$
\|\nabla\theta(t)\|^2
\leq Ch^4 t^{-\alpha(3-\delta)}d_\delta^2(u_0,f),\quad {\rm for}~~ t\in(0,T]\,.
$$
\end{theorem}
\begin{proof}
Apply the operator $\mathcal{I}^{1-\alpha}$ \eqref{k2-1}, and use the identities $\mathcal{I}^{1-\alpha} \partial_t^{1-\alpha} \Theta_2=\Theta_2$ and $t\mathcal{I}^\alpha\theta = \mathcal{I}^\alpha\Theta_1+\alpha\mathcal{I}^{1+\alpha}\theta$ (by Lemma \ref{Ia} \mbox{(b)}) to get
\[
(\mathcal{I}^{1-\alpha} \dot\Theta_2,\chi)+A(\Theta_2,\chi)=
(\mathcal{I}^{1-\alpha}(t^2\rho'), \chi)+ (1-\alpha)\Big(2A(\mathcal{I} \Theta_1,\chi)+\alpha A(\mathcal{I}^2\theta,\chi)\Big).\]
Set $\chi=\dot \Theta_2$ follows by integrating the resulting equation from 0 to $t$ to obtain
\begin{multline*}
\int_0^t[(\mathcal{I}^{1-\alpha} \dot\Theta_2,\dot \Theta_2)+A(\Theta_2,\dot \Theta_2)]\,ds\\
\le
\int_0^t(\mathcal{I}^{1-\alpha}(s^2\rho'),\dot \Theta_2)\,ds+(1-\alpha)\int_0^tA(2\mathcal{I} \Theta_1+\alpha\mathcal{I}^2\theta,\dot \Theta_2)\,ds.\end{multline*}
However, by the continuity property of the operator $\mathcal{I}^{1-\alpha}$ (\cite[Lemma 3.1]{MustaphaSchoetzau2014}), we have
\[\Big|\int_0^t(\mathcal{I}^{1-\alpha}(s^2\rho'),\dot\Theta_2)\,ds\Big|\le C\int_0^t(\mathcal{I}^{1-\alpha}(s^2\rho'),s^2\rho')\,ds
+\int_0^t(\mathcal{I}^{1-\alpha}\dot\Theta_2,\dot\Theta_2)\,ds,\]
and so,
\[
\int_0^tA(\Theta_2,\dot \Theta_2)\,ds\le
C\int_0^t(\mathcal{I}^{1-\alpha}(s^2\rho'),s^2\rho')\,ds+(1-\alpha)\int_0^tA(2\mathcal{I} \Theta_1+\alpha\mathcal{I}^2\theta,\dot \Theta_2)\,ds.\]
Using the identity $2\mathcal{I} \Theta_1(t)=\Theta_2(t)-\mathcal{I} \dot \Theta_2$ and the inequality $\int_0^tA(\mathcal{I} \dot \Theta_2,\dot \Theta_2)\,ds\ge 0$, after some simplifications, we conclude that
\[
\alpha\int_0^tA(\Theta_2,\dot \Theta_2)\,ds\le
C\int_0^t\|\mathcal{I}^{1-\alpha}s^2\rho'\|\,\|s^2\rho'\|\,ds+\alpha(1-\alpha)\int_0^tA(\mathcal{I}^2\theta,\dot \Theta_2)\,ds.\]
Since
\[\int_0^t A(\Theta_2,\dot\Theta_2)\,ds=\frac{1}{2}\|\sqrt{{\bf a}}\nabla\Theta_2(t)\|^2
-2\int_0^t s\|\sqrt{{\bf a}}\nabla\Theta_1(s)\|^2\,ds, \]
we easily find that
\begin{multline}\label{sup-9-n}
\alpha\|\sqrt{{\bf a}}\nabla\Theta_2(t)\|^2
\le 4\alpha\int_0^t s\|\sqrt{{\bf a}}\nabla\Theta_1(s)\|^2\,ds
+C\int_0^t\|\mathcal{I}^{1-\alpha}s^2\rho'\|\,\|s^2\rho'\|\,ds\\
+2\alpha(1-\alpha)\int_0^tA(\mathcal{I}^2\theta,\dot \Theta_2)\,ds.
\end{multline}
Integrating the first bound in Lemma
\ref{theta} gives
\begin{equation}\label{kn1-n}
\int_0^t s\|\sqrt{{\bf a}}\nabla\Theta_1\|^2\,ds\leq Ct\int_0^t \|\nabla\Theta_1\|^2\,ds
\leq C h^4 t^{4-\alpha(3-\delta)}d^2_\delta(u_0,f)\,.
\end{equation}
To estimate the second term on the RHS of \eqref{sup-9-n}, we use the bound of $\rho'$ given in \eqref{eq: rho} (with $m=2$), the formula
\begin{equation}\label{eq: Inu}
\mathcal{I}^\nu (t^{\mu-1})= t^{\nu+\mu-1}\Gamma(\mu),\quad{\rm for}~~\nu,\,\mu >0,\end{equation}
and then integrate
\begin{equation}\label{kn4-n}
\begin{aligned}
\int_0^t(\mathcal{I}^{1-\alpha}(s^2\rho'),s^2\rho')\,ds& \le C \int_0^t s^{2-\alpha-\alpha(2-\delta)/2} s^{1-\alpha(2-\delta)/2}\,ds\, d^2_\delta(u_0,f)\\
&\le Ch^4 t^{4-\alpha-\alpha(2-\delta)}d^2_\delta(u_0,f)\,.
\end{aligned}
\end{equation}
For the last term on the RHS of \eqref{sup-9-n}, we apply
$\mathcal{I}^{2-\alpha}$ to (\ref{sup-3 ph}) to obtain
$A(\mathcal{I}^2 \theta,\chi)= (\mathcal{I}^{2-\alpha}e,\chi).$ Hence, integrating by parts, we find that
\[\begin{aligned}
\int_0^tA(\mathcal{I}^2\theta,\dot \Theta_2)\,ds
&=\int_0^t(s^2\mathcal{I}^{2-\alpha}e,\theta')\,ds\\
&=(\mathcal{I}^{2-\alpha}e(t),\Theta_2(t))-\int_0^t(2\mathcal{I}^{2-\alpha}e+s\mathcal{I}^{1-\alpha}e,\Theta_1)\,ds\,.\end{aligned}\]
Then, by using the estimate of $\theta$ in Lemma \ref{H1}, \eqref{eq: bound of B1}, and the estimate of $e$ in Theorem \ref{thm: smooth and nonsmooth}, we conclude after integrating and using the formula in \eqref{eq: Inu}, that
\[\begin{aligned}
\Big|\int_0^tA(\mathcal{I}^2\theta,\dot \Theta_2)\,ds\Big|&\le t^2\|\mathcal{I}^{2-\alpha}e(t)\|\,\|\theta(t)\|+2t\int_0^t\|\mathcal{I}^{2-\alpha}e(s)+s\mathcal{I}^{1-\alpha}e(s)\|\,\|\theta(s)\|\,ds\\
&\le Ch^4 t^{4-\alpha-\alpha(2-\delta)}d_\delta^2(u_0,f)\,.
\end{aligned}\]
A substitution of the estimates \eqref{kn1-n}, \eqref{kn4-n} and the above one in (\ref{sup-9-n}), follows by using
\eqref{eq: A positive} and the identity $\theta(t)=t^{-2}\Theta_2$ yield the desired estimate. \quad \end{proof}
Noting that, by using the estimates of $\rho$, $\rho'$ and $e$ from Remark \ref{rem: 0} in the inequality \eqref{eq: bound theta}, we observe
\[\int_0^{t} \|\nabla\Theta_1\|^2ds \leq C h^2 t^{3-\alpha(2-\delta)}D^2_{\delta,\alpha}(u_0,f,h/t)\,.\]
Hence, by following the steps in Theorem \ref{sup-conv-1}, and using the above bound instead of Lemma
\ref{theta}, and the bounds of $\rho'$ and $e$ achieved in Remark \ref{rem: 0}, we deduce that
\[
\|\nabla \theta(t)\|^2 \leq
C\, h^2\,t^{-\alpha(2-\delta)}D^2_{\delta,\alpha}(u_0,f,h/t)\,.
\]
Therefore, from the inequality $\|\nabla (u_h-u)(t)\|\le \|\nabla \theta(t)\| +\|\nabla \rho(t)\|$, the above bound, the bound $\eta$ in \eqref{rho-estimate} (with $j=1$ and $m=2$) and the regularity property (\ref{eq: regularity property}), we have the following result in term of a theorem.
\begin{theorem} \label{thm: H1 bound}
Let $u$ and $u_h$ be the solutions of $(\ref{a})$ and $(\ref{semi})$,
respectively, with $u_h(0)=P_h u_0$. For $u_0 \in \dot H^\delta(\Omega)$, for $t\in (0,T]$, we have
$$
\|\nabla(u-u_h)(t)\| \leq
C\, h\,t^{-\alpha(2-\delta)/2}\tilde d_\delta(u_0,f)\times \begin{cases} 1~~~&{\rm for}~~0\le \delta \le 1,\\
\max\{1,(h/t)^{\alpha(\delta-1)/2}\}~~~&{\rm for}~~1< \delta \le 2\,.\end{cases}$$
\end{theorem}
\begin{remark}\label{rem: 3}
The estimate in Theorem \ref{sup-conv-1} suggests that one can achieve a higher convergence rate for $\nabla(u_h-u)$ if an improved estimate of the error
$\nabla(R_hu-u)$ can be derived. This could be achieved using a superconvergent recovery procedure
of the gradient, which is possible on special meshes and for solutions in $H^3(\Omega)$ for each $t\in (0,T]$. Examples of special
meshes exhibiting superconvergence property are provided in \cite{MN-87}. K{\v r}\`i{\v z}ek and Neittaanm\"aki
\cite{MN-87} introduced an operator $G_h$ which postprocesses $\nabla R_h u(t)$ with the following properties:
\begin{itemize}
\item[(i)] If $u(t)\in H^3(\Omega)$, then $ \|\nabla u(t)-G_h(R_h u)(t)\|\leq Ch^2\|u(t)\|_{H^3(\Omega)}.$
\item[(ii)] For $\chi\in V_h$, we have $\|G_h(\chi)\|\leq C\|\nabla \chi\|.$
\end{itemize}
Now, if ${\cal T}_h$ is a triangulation of $\Omega$ such that these results are satisfied, then using
$$\|\nabla (u-u_h)(t)\|\le \|(\nabla u-G_h(R_h u))(t)\| +\|G_h(R_h u-u_h)(t)\|+ \|\nabla \theta(t)\|,$$ (i) and (ii), Theorem \ref{sup-conv-1}, and the inequality $\|u(t)\|_{H^3(\Omega)} \le Ct^{-\alpha(3-\delta)/2}d_\delta(u_0,f)$ for $1/2 < \delta\leq 2$, it is clear that the bound below holds for $t\in (0,T]$,
$$
\|\nabla(u-u_h)(t)\| \leq
C h^{2} t^{-\alpha(3-\delta)/2}d_\delta(u_0,f), \qquad 1/2< \delta\leq 2\,.
$$
\end{remark}
For $t\in (0,T]$, we show in the next theorem that the superconvergence result of $\nabla \theta$ in Theorem \ref{sup-conv-1} can be used to establish an optimal convergence rate (up to a logarithmic factor) in the stronger $L^\infty(\Omega)$-norm assuming that the initial data $u_0 \in \dot H^\delta(\Omega)$ for $0\le \delta \le 2.$ Without loss of generality, we assume that $f\equiv 0$ in problem \eqref{a}.
\begin{theorem}\label{thm: smooth and nonsmooth-5}
Let $u$ and $u_h$ be the solutions of $(\ref{a})$ and $(\ref{semi})$, respectively, with $u_h(0)=P_h u_0$. Assume that
$u_0 \in \dot H^\delta(\Omega)\cap L^\infty(\Omega)$ for $0\le \delta\le 2$. Under the quasi-uniformity condition on $V_h$, for $t \in (0,T]$, we have
$$
\|(u-u_h)(t)\|_{L^{\infty}(\Omega)} \leq
C|\ln h|^{\frac{5}{2}} h^2 t^{-\alpha(3-\delta)/2}\Big(\|u_0\|_\delta+\|u_0\|_{L^\infty(\Omega)}\Big)\quad {\rm for}~~0\le \delta\le 2.
$$
\end{theorem}
\begin{proof} By the Ritz projection error result (see \cite{thomee1997}) and the regularity estimate $\|u(t)\|_{W^{2,p}(\Omega)}\le Cp\|\mathcal L u(t)\|_{L^p(\Omega)}$, for $1<p<\infty$, we have
\begin{equation}\label{eq: infty 1}
\|\rho(t)\|_{L^\infty(\Omega)}\leq C |\ln h| h^{2-2/p}\|u(t)\|_{W^{2,p}(\Omega)} \leq C|\ln h| h^{2-2/p} p \|\mathcal L u(t)\|_{L^p(\Omega)}\,.\end{equation}
Applying the operator $\mathcal{I}^{1-\alpha}$ to both sides of \eqref{a} (with $f\equiv 0$), we arrive at
$\mathcal{I}^{1-\alpha} (u'(t))={\mathcal L} u(t)$. Insert this in \eqref{eq: infty 1}, then use the Sobolev embedding inequality ($\|v\|_{L^p(\Omega)}\leq C \sqrt{p} \|\nabla v\|$ for $v\in H^1_0(\Omega)$ with $2\le p<\infty$) and the regularity property \eqref{eq: regularity property-2}, yield
\begin{equation}\label{eq: infty 2}\begin{aligned}
\|\rho(t)\|_{L^\infty(\Omega)}&\le C|\ln h| h^{2-2/p} p^{3/2} \mathcal{I}^{1-\alpha} (\|u'(t)\|_1)\\
&\le C|\ln h| h^{2-2/p} p^{3/2} \Big(t^{-\alpha(3-\delta)/2}\|u_0\|_\delta\Big)\quad {\rm for}~~1< \delta\le 2.
\end{aligned}\end{equation}
To estimate $\|\rho(t)\|_{L^\infty(\Omega)}$ for $0\le \delta\le 1$, multiplying both sides of \eqref{a} (with $f\equiv 0$) by $t$, then applying the operator $\mathcal{I}^{1-\alpha}$ and using the identity in Lemma \ref{Ia} (a) (with $u$ in place of $\theta$), we find that
$\mathcal{I}^{1-\alpha} (tu'(t))={\mathcal L} (tu(t))-(1-\alpha)\mathcal{I} {\mathcal L} u(t)$. But, since $\mathcal{I} {\mathcal L} u(t)=\mathcal{I}^{2-\alpha} u'(t)=\mathcal{I}^{1-\alpha} [u(t)-u_0]$, it follows that
\[{\mathcal L} (tu(t))=\mathcal{I}^{1-\alpha} [(tu'(t))+(1-\alpha)(u(t)-u_0)].\]
Now, multiplying both side of \eqref{eq: infty 1} by $t$, then inserting the above bound and using again the Sobolev embedding inequality imply that
$$
t\|\rho(t)\|_{L^\infty(\Omega)}\leq C|\ln h| h^{2-2/p} p^{3/2} \Big(\mathcal{I}^{1-\alpha} [\|tu'(t)\|_1+\|u(t)\|_1]+t^{1-\alpha}\|u_0\|_{L^p(\Omega)}\Big),
$$
for any $2\le p<\infty.$ Hence, by \eqref{eq: regularity property-2} and the inequality $\|u(t)\|_1\le C t^{-\alpha(1-\delta)/2}\|u_0\|_\delta$ for $0\le \delta\le 1$ (follows from the regularity property \eqref{eq: regularity property} with $f\equiv 0$), we obtain
\begin{multline*}
t\|\rho(t)\|_{L^\infty(\Omega)}\le Ct|\ln h| h^{2-2/p} p^{3/2} \Big(t^{-\alpha(3-\delta)/2}\|u_0\|_\delta+t^{-\alpha}\|u_0\|_{L^p(\Omega)}\Big)\quad {\rm for}~~0\le \delta\le 1.
\end{multline*}
On the other hand, by the discrete Sobolev inequality and the achieved estimate in Theorem \ref{sup-conv-1}, we observe that
$$
\|\theta(t)\|_{L^\infty(\Omega)}\leq C|\ln h|\|\nabla \theta(t)\|\leq C|\ln h| h^2 t^{-\alpha(3-\delta)}\|u_0\|_\delta\quad {\rm for}~~0\le \delta\le 2.
$$
Finally, choose $p=|\ln h|$, and the desired convergence result follows then from $\|(u_h-u)(t)\|_{L^\infty(\Omega)}\leq \|\theta(t)\|_{L^\infty(\Omega)}
+\|\rho(t)\|_{L^\infty(\Omega)}$, \eqref{eq: infty 2}, and the above two bounds.
\end{proof}
\section{FE error analysis for the problem \eqref{D}}\label{sec: Caputo}
In this section, we justify that the achieved error estimates from the Galerkin FE semidiscrete solution of problem \eqref{a} are valid for the FE discretization of \eqref{D}. To see this, for each $t\in (0,T]$, we denote by $u_h(t)$ to be the Galerkin FE solution of \eqref{D}. Thus,
\begin{equation} \label{FE scheme caputo}
(\mathcal{I}^{1-\alpha} u_h',\chi)+A(u_h,\chi)=
(f,\chi)\quad \forall~\chi \in V_h\,.
\end{equation}
As before, we decompose the error as: $u-u_h:=\rho-\theta.$
One can check that $\theta$ satisfies
\begin{equation} \label{sup-2 caputo}
(\mathcal{I}^{1-\alpha} \theta',\chi)+A(\theta,\chi)=
(\mathcal{I}^{1-\alpha} \rho',\chi)\quad \forall~\chi \in V_h.
\end{equation}
Applying the operator $\mathcal{I}^\alpha$ to both sides and use the property $\mathcal{I}^\alpha\mathcal{I}^{1-\alpha}=\mathcal{I}$ yield
\[
(\mathcal{I} \theta',\chi)+A(\mathcal{I}^\alpha \theta,\chi)=
(\mathcal{I} \rho',\chi)\quad \forall~\chi \in V_h.
\]
A time differentiation of both sides shows
\begin{equation} \label{sup-2 caputo 1}
(\theta',\chi)+A(\partial_t^{1-\alpha} \theta,\chi)=
( \rho',\chi)\quad \forall~\chi \in V_h.
\end{equation}
Since the two equations \eqref{sup-2 caputo 1} and \eqref{sup-2} are identical, by following
the preceding error analysis in Sections \ref{sec: LinftyL2} and \ref{sec: LinftyH1}, one can show that the achieved error estimates in Theorems \ref{thm: smooth and nonsmooth}, \ref{thm: H1 bound} and \ref{thm: smooth and nonsmooth-5} remain valid.
\section{Numerical results}\label{sec: Numerical}
In this section, we focus on testing the achieved theoretical convergence results in Theorem \ref{thm: smooth and nonsmooth-5} on fractional model problems of the form \eqref{a} with different initial data. For the numerical illustration of the error bounds in Theorems \ref{thm: smooth and nonsmooth} and \ref{thm: H1 bound}, one can follow the convention in \cite[Section 6]{JLZ2013}.
We choose ${\cal{L}} = -\nabla^2$, $f\equiv 0$, $T=0.5$, and $\Omega=(0,1)\times (0,1)$ in problem \eqref{a}. The orthonormal eigenfunctions and
corresponding eigenvalues of~$-\nabla^2$
are
\[
\phi_{mn}(x,y)=2\sin(m \pi x)\sin(n \pi y)
\quad\text{and}\quad
\lambda_{mn}=(m^2+n^2)\pi^2\quad{\rm for}~~ m\,,n=1, 2, \ldots.
\] Separation of variables yields the series representation solution of problem \eqref{a}:
\begin{equation}\label{eq: u series}
u(x,y,t)=2\sum_{m,n=1}^\infty (u_0, \phi_{mn})
E_{\alpha}(-\lambda_{mn} t^{\alpha})\phi_{mn}(x,y),
\end{equation}
where $E_{\alpha}(t):=\sum_{p=0}^\infty\frac{t^p}{\Gamma(\alpha p+1)}$ is the
Mittag-Leffler function.
To compute the semidiscrete FE solution $u_h$, we discretize in time by the mean of generalized Crank-Nicolson scheme (see \cite{Mustapha2011}), this will then define the following fully-discrete scheme: for $1\le n\le N$
\begin{equation*
\tau_n^{-1}(u_h^n-u_h^{n-1},v_h)+ A(\mathcal{I}^\alpha \bar u_h(t_n)-\mathcal{I}^\alpha \bar u_h(t_{n-1}),v_h)= 0\quad
\forall v_h\in V_h,
\end{equation*}
where $N$ is the number of time mesh subintervals ($0=t_0<t_1<\ldots<t_N=T$), $\tau_n$ is the $n$th time step size. Here
$u_h^n \approx u_h(t_n)$ and $\bar u_h(s)=\frac{1}{2}(u_h^j+u_h^{j-1})$ when $s \in (t_{j-1},t_j)$ for $j\ge 2,$ while $\bar u_h(s)=u_h^1$ on the subinterval $(0,t_1).$ The modification on the first subinterval ensures that $\bar u_h$ does not depend on $u_h^0$ which is necessary for our numerical scheme in cases when $u_0$ is not sufficiently regular.
Following the convergence analysis in \cite{Mustapha2011}, we concentrate the time step near $t=0$ to compensate for the singular behaviour of the solution $u$ of problem \eqref{a}. So, we let $t_n=(n/N)^\gamma T$ for some fixed $\gamma\ge 1$ that will be chosen appropriately. For the spatial partition, let $\mathcal{T}_h$ be a family of uniform (right-angle) triangular mesh of the domain $\Omega$ with diameter ~$h=\sqrt{2}/M$, see Figure \ref{Fig: spatial mesh}. For measuring the error in our numerical solution at each time node $t_n$, we let ${\mathcal N}_h$ be the set of all triangular nodes of the mesh family $\mathcal{T}_{h_s}$ where the diameter $h_s$ is half the diameter of the finest mesh $\mathcal{T}_h$ in our spatial iterations, for instance, $h_s=\sqrt{2}/128$ in Tables \ref{table 1}--\ref{table 3} as well as in Figures \ref{fig1}--\ref{fig3}. To measure the errors, define the discrete-space maximum norm: $|\|v\||:=\max\{|v({\bf x})|,~{\bf x}\in {\mathcal N}_h\}\,.$ Thus, for large values of $M$, $|\|u_h^n-u(t_n)\||$ approximates the error $\|u_h^n-u(t_n)\|_{L^{\infty}(\Omega)}$ for each $1\le n\le N$.
In the three numerical examples below, we choose $\gamma=1.6$ and refine the time steps so that the spatial
errors are dominant. We evaluate the exact solution $u$ of problem \eqref{a} by truncating the Fourier series in \eqref{eq: u series} after $60$ terms.
\begin{figure}
\begin{center}
\caption{Triangular mesh $\mathcal{T}_h$ with $M=8$.}
\label{Fig: spatial mesh}
\includegraphics[width=6cm, height=6cm]{Figure4LebPres}
\end{center}
\end{figure}
{\bf Example 1.} Choose the initial data $u_0(x,y)=xy(1-x)(1-y)$, which has the Fourier sine coefficients
\[
(u_0,\phi_{mn})=8(1-(-1)^m)(1-(-1)^n)(mn\pi^2)^{-3},\quad {\rm for}~~m,n=1,\,2,\ldots.
\] The initial data $u_0 \in \dot H^{2+\epsilon}(\Omega)$ for $0\le \epsilon<1/2$ and $\notin \dot H^{2+\epsilon}(\Omega)$ for $\epsilon >1/2.$ Thus, by Theorem \ref{thm: smooth and nonsmooth-5} ($\delta=2$), for each time step $t_n$, we expect convergence of order $t_n^{-\alpha/2}|\ln h|^{5/2}h^2$ in the $L^\infty(\Omega)$-norm. Indeed, one can show the validity of the convergence results (up to some time logarithmic factor) in Theorem \ref{thm: smooth and nonsmooth-5} for $1<\delta<2.5^-$. Hence, the coefficient $t_n^{-\alpha/2}$ can be replaced with $t_n^{-\alpha/4}$. For $\alpha=0.75$, Figure \ref{fig1} shows how the error varies with $t$ for a sequence of solutions obtained by successively doubling the spatial mesh elements, using a log scale. (The same time mesh with $N=1000$ subintervals was used in all cases). In Table \ref{table 1}, we listed the time-space maximum error and its associated convergence rate ($CR$), where optimal convergence rates was observed (ignoring the logarithmic factors). Therefore, the coefficient $t_n^{-\alpha/4}=t_n^{-3/16}$ does not have much practical influence on the convergence rates. This is probably due to the fact that $u_0$ is also in $C^2(\Omega) \cap C_0(\overline \Omega),$ where an $O(h^2|\log h|^2)$ rate of convergence was proved in \cite[Theorem 4.2]{McLeanThomee2010}.
\begin{figure}
\begin{center}
\caption{The error $|\|u_h^n-u(t_n)\||$ as a function of~$t_n$ for Example 1.}
\label{fig1}
\includegraphics[width=10cm, height=6cm]{Figure1Case1.eps}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\caption{Behavior of the uniform error $\max_{n=1}^N |\|u_h^n-u(t_n)\||$ and the associated convergence rates as the
number of spatial mesh elements increases. In each case, we use $1000$ time subintervals.}
\label{table 1}
\begin{tabular}{|r|cc|}
\hline
$M$&
$\max_{n=1}^N |\|u_h^n-u(t_n)\||$& $CR$
\\
\hline
4& 1.2759e-02& \\
8& 3.3749e-03& 1.9186\\
16& 8.7940e-04& 1.9402\\
32& 2.2284e-04& 1.9805\\
64& 5.6414e-05& 1.9819\\
\hline
\end{tabular}
\end{center}
\end{table}
{\bf Example 2.} Choose $u_0(x,y)=xy\chi_{(0,1/2]\times(0,1/2]} +(1-x)y\chi_{(1/2,1)\times(0,1/2]}+
x(1-y)\chi_{(0,1/2]\times(1/2,1)}+(1-x)(1-y)\chi_{(1/2,1)\times(1/2,1)}$ which is less smooth, then the considered $u_0$ in the previous example, where $\chi_D$ denotes the characteristic function on the domain $D$. One can verify that $u_0$ has the Fourier sine coefficients
\[
(u_0,\phi_{mn})=2(1-(-1)^m)(1-(-1)^n)(mn\pi^2)^{-2}(-1)^{mn},\quad {\rm for}~~m,n=1,\,2,\ldots.
\]
The function $u_0 \in \dot H^{1+\epsilon}(\Omega)$ for $0\le \epsilon<1/2$. So, by Theorem \ref{thm: smooth and nonsmooth-5} ($\delta<1.5$), for each $t_n$, we expect $t_n^{-3\alpha/4}O(|\ln h|^{5/2}h^2)$ convergence rates in the $L^\infty(\Omega)$-norm from the spatial FE discretization. As in Figure \ref{fig1}, Figure \ref{fig2} shows how the error varies with $t$ for a sequence of solutions obtained by doubling the spatial mesh elements. (The time mesh with $N=1300$ subintervals was used in all cases). Table \ref{table 2} provides an alternative view of this data, listing the time-space maximum weighted error $E_\mu:=\max_{n=1}^N t_n^\mu |\|u_h^n-u(t_n)\||$ and its associated convergence rate $CR$. As expected, ignoring the logarithmic factors, the convergence rate is $2$ when $\mu \ge 3\alpha/4 \approx 0.56$, but the rate deteriorates for smaller values of $\mu$ (relatively far from $3\alpha/4$).
\begin{figure}
\begin{center}
\caption{The error $|\|u_h^n-u(t_n)\||$ as a function of~$t_n$ for Example 2.}
\label{fig2}
\includegraphics[width=10cm, height=6cm]{Figure1Case2.eps}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\caption{Behavior of the weighted error $E_\mu$, as the
number of spatial mesh elements increases, for different choices of the
power weight exponent $\mu$. In each case, we use $1300$ time subintervals.}
\label{table 2}
\begin{tabular}{|r|cc|cc|cc|cc|}
\hline
$M$&
$E_0$& $CR$&
$E_{0.25}$& $CR$&
$E_{0.5}$& $CR$&
$E_{0.75}$& $CR$\\
\hline
4& 3.008e-02& & 9.521e-03& & 3.610e-03& & 1.597e-03& \\%& 1.051e-03& \\
8& 1.054e-02& 1.513& 1.412e-03& 2.754& 5.342e-04& 2.757& 2.401e-04& 2.734\\%& 1.479e-04& 2.828\\
16& 5.441e-03& 9.536& 4.112e-04& 1.779& 1.279e-04& 2.062& 5.678e-05& 2.080\\%& 3.742e-05& 1.983\\
32& 1.876e-03& 1.536& 1.391e-04& 1.564& 3.344e-05& 1.936& 1.513e-05& 1.908\\%& 1.116e-05& 1.745\\
64& 8.667e-04& 1.114& 6.425e-05& 1.114& 8.598e-06& 1.959& 4.055e-06& 1.900\\%& 3.381e-06& 1.723\\
\hline
\end{tabular}
\end{center}
\end{table}
{\bf Example 3.} Choose $u_0(x,y)=1$, and so $u_0$ has the Fourier sine coefficients
\[
(u_0,\phi_{mn})=2(1-(-1)^m)(1-(-1)^n)(mn\pi^2)^{-1},\quad {\rm for}~~m,n=1,\,2,\ldots.
\]
The initial data function $u_0 \in \dot H^{\epsilon}(\Omega)\cap L^\infty(\Omega)$ for $0\le \epsilon<1/2$. As in the previous example, Figure \ref{fig3} shows a consistent decaying in the errors by doubling the number of spatial mesh elements. Another observation is the large impact of the very limited regularity of $u_0$ on the errors near $t=0$ in this example. For better justifications of this, see Table \ref{table 3} where the difference between the maximum error $E_0$ and the weighted error $E_1$ is very substantial, we also observed very good improvements in the convergence rates $CR$, but not yet optimal.
\begin{figure}
\begin{center}
\caption{The error $|\|u_h^n-u(t_n)\||$ as a function of~$t_n$ for Example 3.}
\label{fig3}
\includegraphics[width=10cm, height=6cm]{Figure1Case3.eps}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\caption{Behavior of the weighted error $E_\mu$, as the
number of spatial mesh elements increases, for different choices of the
power weight exponent $\mu$. In each case, we use $1300$ time subintervals.}
\label{table 3}
\begin{tabular}{|r|cc|cc|cc|cc|}
\hline
$M$&
$E_0$& $CR$&
$E_{0.5}$& $CR$&
$E_{0.75}$& $CR$&
$E_1$& $CR$\\
\hline
4& 1.0160e-00& & 3.453e-02& & 1.531e-02& & 9.898e-03& \\
8& 9.7501e-01& 0.0594& 8.545e-03& 2.015& 2.245e-03& 2.769& 1.525e-03& 2.699\\
16& 7.0054e-01& 0.4769& 3.852e-03& 1.150& 6.809e-04& 1.721& 4.783e-04& 1.672\\
32& 3.2311e-01& 1.1164& 1.776e-03& 1.117& 2.000e-04& 1.767& 1.442e-04& 1.730\\
64& 1.5301e-01& 1.0783& 8.409e-04& 1.078& 6.234e-05& 1.682& 4.945e-05& 1.544\\
\hline
\end{tabular}
\end{center}
\end{table}
|
\section{Introduction}
In the dueling bandits problem, we are faced with a collection of arms, and pull a pairs of arms while observing noisy binary feedback indicating which arm is better for each pulled pair. As in the classical multi-armed bandit problem, we wish to pull arms to quickly learn which arm is best and minimize the number of pulls to suboptimal arms.
Dueling bandits were introduced by \cite{yue2009interactively}, motivated by interactive optimization of web search and other information retrieval systems. The advantage of the dueling bandits formulation over the classical multi-armed bandits formulation in this application setting is that pairwise comparison results can be reliably inferred from implicit feedback, for example through interleaved rankings in \cite{radlinski2008does}, in contrast with cardinal evaluation obtained from explicit feedback, which is typically difficult to obtain, biased, and requires careful calibration \citep{joachims2007evaluating, yue2012k}.
Dueling bandits have been studied most frequently assuming strong regret, in which the regret is 0 if and only if both pulled arms are optimal. Several algorithms have been devised that assume the existence of a Condorcet winner, i.e., one that is preferred in comparison with each other arm. Algorithms with order-optimal strong regret, $O(N\log(T))$, in this setting include BTM \citep{yue2011beat}, RUCB \citep{zoghi2014relative} and RMED \citep{komiyama2015regret}. \cite{zoghi2015copeland} points out points out that a Condorcet winner does not necessarily exist, and that its probability of existence decreases dramatically with the number of arms. That work instead studies the dueling bandits assuming a Copeland winner, which is guaranteed to exist, and propose two algorithms, CCB and SCB, which achieve $O(N\log(T))$ strong regret in this more general setting.
The above papers on strong regret bound the binary strong regret, in which the regret is $1$ whenever it is strictly positive. \cite{ailon2014reducing} considered strong utility-based regret, in which each arm has a utility score from which preferences are derived, and the regret for failing to pull the maximum utility arm twice is a function of that maximal arm's utility and the utilities of the pulled arms.
Bandits have also been considered, though less frequently, in the weak regret setting, introduced by \cite{yue2012k}, in which regret is $0$ if either of the pulled arms is optimal. This setting is more appropriate for recommender systems, in which we offer the user a pair of items, and she selects the one that is preferred. $0$ regret is incurred as long as the best item is made available. While \cite{yue2012k} introduced weak regret, an algorithm with regret bounds first appeared in \cite{chen2017dueling}, which proposed the \textit{Winner Stays} (WS) algorithm that achieves $O(N\log(N))$ cumulative binary weak regret when arms have a total order and $O(N^2)$ in the Cordorcet winner setting. These bounds on binary weak regret have corresponding bounds on utility-based weak regret inflated by the difference in utility between the best and worst arms.
We consider utility-based weak regret, in the total order setting, when the total order is induced by a utility which is in turn a function of observable arm features, an unknown latent preference vector, and a known utility function.
This framework includes the commonly used logit or Bradley-Terry \citep{revelt1998mixed,yue2012k} and probit models \citep{franses2002econometric}.
We provide an algorithm, Comparing with the Best (CTB) that has expected cumulative utility-based weak regret that is constant in $T$, and that leverages the dependence between preferences over arms induced by the arm features and utility function to provide excellent empirical performance when prior information is available.
While our regret bound's dependence on $N$ is looser than \cite{chen2017dueling} (our dependence is $2^N$ in the worst case, and is $N^{2d}$ when the utility function is linear over a $d$-dimensional space of preferences and arm features), our algorithm is more flexible in its ability to problem structure induced by the feature vectors, and outperforms it empirically by a substantial margin when $N$ is small enough to allow computation that fully takes advantage of this problem structure.
Our exploitation of arm features is similar in spirit to work in the traditional (cardinal) multi-armed bandit setting on linear bandits \citep{rusmevichientong2010linearly, abbasi2011improved}.
The paper is structured as follows. In section~\ref{probForm}, we formulate our problem.
In section~\ref{Methods}, we introduce {\it Comparing The Best} (\CTB) which we show in section~\ref{results} has \CTB\ constant expected cumulative regret. In section~\ref{Imple}, we discuss a efficient implementation method for a specific class of prior information. In section~\ref{sec:Bayes}, we provide a Bayesian interpretation for \CTB. In section~\ref{sec:exp}, we compare \CTB\ with three benchmarks using simulated datasets, in which \CTB\ outperforms all benchmarks considered.
\newcommand{\prefdim}{d'}
\newcommand{\armdim}{d}
\section{Problem Formulation}
\label{probForm}
There are $N\geq 2$ arms, and each arm $i$ has an observable and distinct $\armdim$-dimensional feature vector $A_{i}$.
Preferences between pairs of arms $i,j$ are described by fixed but unknown probabilities $p_{i,j}$, where $p_{i,j} = 1 - p_{j,i}$ and $p_{i,j} \ne 0.5$ when $i \ne j$.
We denote $p=\min_{i<j} \max(p_{i,j},p_{j,i})$. By construction, $p>0.5$.
At each time t, we pull two arms $X_{t,0}$ and $X_{t,1}$ (this act is called a ``duel'') and we observe feedback $Y_{t}\in \{0,1\}$ indicating the winning arm: $Y_{t}=0$ indicates arm $X_{t,0}$ won and $Y_{t}=1$ indicates arm $X_{t,1}$ won.
Conditioned on the arms pulled and the history (the arms pulled and the identity of the winner at times $t'<t$), $Y_t$ is equal to $0$ with probability $p_{i,j}$.
We suppose that the arms have a total order, i.e., that there exists an ordering of the arms such that $p_{i,j} > 0.5$ if and only if arm $i$ is before arm $j$ in this order.
Moreover, we suppose this ordering is determined by a utility associated with each arm, $u(\theta, A_i)$, where $u$ is a known utility function and
$\theta \in \mathbb{R}^{\prefdim}$ is an unknown preference vector.
In particular, $p_{i,j} > 0.5$ if and only if $u(\theta, A_{i})>u(\theta,A_{j})$.
The assumption that the total order be determined by $u(\theta,A_i)$ is without loss of generality if we are willing to select $\prefdim$ to be sufficiently large and $u$ to allow sufficient flexibility, although one may also choose a smaller $\prefdim$ and a less flexible $u$ with the goal of obtaining smaller regret (described below) when these more restrictive modeling assumptions hold. We assume without loss of generality that the indices correspond to their ordering by utility, so $u(\theta, A_{1}) > u(\theta, A_{2}) > \cdots > u(\theta,A_{N})$.
Several commonly used discrete choice models fall within this framework. For example, our framework includes the logit or Bradley-Terry model \citep{revelt1998mixed,yue2012k}, in which $\prefdim=\armdim$, the utility function is $u(\theta,A_{i})=\theta\cdot A_{i}$ and
$p_{i,j} =\frac{\exp(u(\theta,A_{i}))}{\exp(u(\theta,A_{i})+u(\theta,A_{j}))}$.
Our framework also includes the probit model \citep{franses2002econometric} in which $\prefdim = \armdim$ and the utility function is the inner product as with the logit model, but $p_{i,j}=\Phi(u(\theta,A_{i})-u(\theta,A_{j}))$ where $\Phi(\cdot)$ is the standard normal cdf.
We define the utility-based weak regret $r(t)$ (henceforce referred to simply as the regret) at time $t$ as $r(t)=u(\theta,A_{1})-\max\{u(\theta,A_{X_{t,0}}), u(\theta,A_{X_{t,1}})\}$, which is the difference in utility between the best arm overall and the best arm available to the user from those offered.
The cumulative regret up to time $T$ is $R(T)=\sum_{t=1}^{T}r(t)$.
We measure the quality of an algorithm by its expected cumulative regret.
We now develop an algorithm \CTB, and show it has constant expected cumulative regret.
\section{The \textit{Comparing The Best} (CTB) Algorithm}
\label{Methods}
In this section we propose an algorithm {\it Comparing The Best} (\CTB) for this problem setting. This algorithm is based on the idea of ``cells'', which correspond to possible orderings of the arms by utility. It maintains a score for each cell, either explicitly or implicitly, which it initializes using optional prior information, and updates with the results from each duel.
We present a general version of \CTB\ in this section that admits any prior information and explicitly maintains a score for each cell. Because the number of cells is exponential in the number of arms, explicitly maintaining scores for each cell is computationally infeasible for large problems. Thus, after presenting our theoretical results for the general \CTB\ algorithm in section~\ref{results}, we present a computationally efficient implementation of our algorithm in section~\ref{Imple} that can be used when the prior information can be expressed in terms of an initial score for each pair of arms. Although we present our algorithm in a frequentist setting, we show in section~\ref{sec:Bayes} that the scores used for each cell correspond to a Bayesian posterior on the value of $\theta$, and \CTB\ has a natural Bayesian interpretation.
To define \CTB, we first define some terminology and notation: {\it winning spaces}, {\it cells}, a {\it score}, and the best arm corresponding to a cell. We begin with winning spaces.
\begin{definition}
Each pair of arms $i,j$ defines a {\it winning space} $H_{i,j} := \{X\in \mathbb{R}^{d}: u(X,A_{i}) \geq u(X, A_{j})\}$.
\end{definition}
When $\theta\in H_{i,j}$, arm $i$ is preferred over arm $j$.
We use the phrases ``arm $A_i$ wins over arm $A_j$ in a duel'', and ``winning space $H_{i,j}$ wins the duel'' interchangeably.
Each pair of arm determines two winning spaces and all winning spaces partition the space $\mathbb{R}^{d}$ into cells, where each cell is an intersection of winning spaces. To define notation to support working with cells, we first define $H_{i,j}(k)=H_{i,j}$ when $k=0$ and $H_{i,j}(k)=H_{j,i}$ when $k=1$. For a binary vector $V$, we let $V[k]$ denote the $k^{th}$ element of $V$. Then, we have the following definition.
\begin{definition}
The {\it cell} C corresponding to a length $\frac{N(N-1)}{2}$ binary vector V is
\begin{align*}
C(V):=\cap_{i<j}H_{i,j}\left(V\left[\frac12 (2N-i)(i-1)+j-i\right]\right).
\end{align*}
\end{definition}
We assign binary vectors indexing cells, all of length $\frac{N(N-1)}{2}$, to integers lexicographically.
Let $V_{k}$ denote the $k^{th}$ such binary vector, let $M=2^N$ denote the number of cells, and let $C_i = C(V_i)$.
With this definition,
$C_{1} = C(V_{1}) =C([0,0,\cdots,0])$ and thus $C_{1}=\cap_{i<j}H_{i,j}$ and $\theta \in C_{1}$.
Some cells $C_i$ may be empty. We call these {\it empty cells}.
Let $J_{k}=\{(i,j)|C_{k}\subseteq H_{i,j}\}$, which is the collection of indices of the winning spaces that contains $C_{k}$.
Figure~\ref{illustration} illustrates winning spaces and cells.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.5]{illustration.png}
\caption{Illustration of winning spaces and cells. The index of the cell and its corresponding binary vectors are: $C_{1}$ and $(0,0,0)$; $C_{2}$ and $(0,0,1)$; $C_{3}$ and $(0,1,0)$; $C_{4}$ and $(0,1,1)$; $C_{5}$ and $(1,0,0)$; $C_{7}$ and $(1,1,0)$; $C_8$ and $(1,1,1)$. In this case, cell $C_6$ is an empty cell since the intersection of $H_{2,1}$, $H_{1,3}$ and $H_{3,2}$ is empty.}
\label{illustration}
\end{figure}
\newcommand{\m}[1]{m_{#1}(0)}
\newcommand{\m{i}}{\m{i}}
We define a score $m_i(t)$ associated with each cell $C_i$ at time $t$. Later in section~\ref{sec:Bayes} we will interpret this score as a monotone transformation of the posterior probability that $\theta$ is in this cell. This score will be initialized to some value $\m{i}$, discussed below, and then will be incremented each time a winning space containing $C_i$ wins a duel. That is,
\begin{align}
m_{i}(t)&=\m{i} +\sum_{k=1}^{t}\mathbbm{1}\{C_{i}\subseteq H_{X_{k,1},X_{k,2}}(Y_k)\}.
\label{update}
\end{align}
Each cell $C_{i}$ assigns a preference order to the arms. Let $B(i)$ be the arm that would be best if $\theta$ were in $C_i$. More formally, $B(i)$ is the unique $j$ such that $C_{i}\subseteq H_{j,k}$, $\forall k\neq j$. Since $\theta \in C_{1}$, we know $B(1)=1$.
With this notation, we now define the {\it Comparing The Best} (\CTB) algorithm in Algorithm~\ref{algo1}. \CTB\ pulls the arm that is best according to the cell with the highest score $m_i(t)$, and the arm that is best according to the cell with the highest score among those that have different best arm from the first arm chosen.
If we interpret $m_i(t)$ as being a monotone transformation of the posterior probability that $\theta \in C_i$, then we are selecting arms by selecting two cells that have different best arms, and are together most likely to contain $\theta$.
\begin{algorithm}[H]
\For{t $\leq$ T}{
Step 1: Pick $X_{t,0}=B(\mathop{\mathrm{argmax}}_{i}m_{i}(t))$, breaking ties arbitrarily\\
Step 2: Pick $X_{t,1}=B\left(\mathop{\mathrm{argmax}}_{i:B(i)\neq X_{t,0}}m_{i}(t)\right)$, breaking ties arbitrarily\\
Step 3: Observe the noisy feedback $Y_{t}$ and update $m_{i}(t)$ using Equation~\eqref{update}\\
Step 4: t=t+1
}
\caption{{\it Comparing The Best} (\CTB)}
\label{algo1}
\end{algorithm}
\paragraph{Choice of $\m{i}$:}
Here we offer guidance on the choice of $\m{i}$, which is left general in the description of \CTB\ to allow the user the flexibility to influence the arms pulled with prior information about the value of $\theta$, and to trade off regret against CTB's computational performance. In doing so, there are four considerations:
First, by setting $\m{i}$ larger for those cells that the user believes are more likely to contain $\theta$, the user encourages \CTB\ to select those cells more often. If the user correctly sets $\m{i}$ larger for the cell that contains $\theta$, this tends to pull the best arm more often and decrease regret. We show in section~\ref{sec:Bayes} that $\m{i}$ can be interpreted in terms of the prior probability that $\theta \in C_i$, and one can leverage this relationship to convert prior information on $\theta$ into values for $\m{i}$.
Second, by setting $\m{i}$ to be $-\infty$ for those cells that user is certain do not contain $\theta$, she can lead \CTB\ to never select those cells. One may safely do this for empty cells, in which model assumptions imply $\theta$ cannot reside. Doing this for other cells is dangerous, as setting cell $\m{1}$ to $-\infty$ can cause \CTB\ to have linear regret.
Third, in the absence of prior information, one may simply set $\m{i}=0$ for all cells that may contain $\theta$. We show in the next section show that as long as $\m{1}>-\infty$, the expected cumulative regret is finite.
Fourth, there is a computational aspect to setting $\m{i}$. We show below in section~\ref{Imple} that if each $\m{i}$ can be written as a sum across pairs of arms of a score associated with each pair, then we can implement \CTB\ in a computationally efficient manner that scales to many arms. In contrast, if one sets $\m{i}$ without enforcing structure, the computation required to implement Algorithm~\ref{algo1} grows exponentially with the number of arms.
With these considerations in mind, we propose 3 specific ways to set $\m{i}$, and evaluate them in numerical experiments:
\begin{itemize}
\item For situations with loose computational requirements or few arms, and no prior information, we recommend setting $m_i = 0$ for all non-empty cells and $m_i=-\infty$ for all empty cells. We call this \CTBone.
\item For situations with strict computational requirements and no prior information, we recommend setting $m_i = 0$ for all cells. Then \CTB\ can be implemented using the efficient method described in section~\ref{Imple}. We call this \CTBtwo.
\item For situations with loose computational requirements or few arms, and strong prior information, we recommend setting $m_i$ from the prior according to the method described in section~\ref{sec:Bayes}. We call this CTB$-3$.
\end{itemize}
\section{Theoretical Results}
\label{results}
In this section, we prove the expected cumulative regret of \CTB\ is bounded by a constant.
The main idea behind our proof is to show that for each cell $C_{i}$ with $B(i)\neq 1$, $\mathbb{E}[\sum_{t=0}^{\infty}1\{m_{i}(t)\geq m_{1}(t)\}]$ is bounded by a constant. We show this in turn by relating $m_{1}(t)-m_{i}(t)$ to a random walk with a larger probability of increasing than of decreasing. The following lemma, whose proof is in the supplement, allows us to bound the number of times this stochastic process takes values less a constant.
\begin{lemma}
\label{basic}
Let $p \in (0.5,1]$.
Suppose $Z(t)$ is a stochastic process with filtration $\mathcal{F}_t$, $Z(0)=0$
and $P(Z(t+1)=Z(t)+1|\mathcal{F}_t)\geq p$,
then we have $\mathbb{E}\left[\sum_{t=0}^{\infty} \mathbbm{1} \{ Z(t)\leq S\}\right] \leq \frac{p+S(2p-1)}{(2p-1)^{2}}$ for $S\in \mathbb{N}$.
\end{lemma}
We now proceed with the larger proof by defining
\begin{align}
q_{i,j}(t)&=\sum_{k=1}^{t}\mathbbm{1}\{X_{k,0}=i, X_{k,1}=j, Y_{k}=0\} +\sum_{k=1}^{t}\mathbbm{1}\{X_{k,0}=j,X_{k,1}=i,Y_{k}=1\}, \label{update2}
\end{align}
which is the number of times up to time $t$ that arm $i$ beats arm $j$ in a duel.
Then we can rewrite $m_{i}(t)$ in terms of $q_{i,j}(t)$ as,
\begin{align}
m_{k}(t)=\m{k}+\sum_{(i,j)\in J_{k}}q_{i,j}(t).
\label{reconstruct}
\end{align}
The definition of $C_{1}$ implies $J_{1}=\{(i,j),\forall i<j\}$ and $m_{1}(t) = \m{1}+\sum_{i<j}q_{i,j}(t)$. Let $N_{i,j}(t) = q_{i,j}(t) + q_{j,i}(t)$ denote the number of times we have pulled arms $i$ and $j$.
The next lemma shows $\mathbb{E}[N_{i,j}(t)]$ is bounded by a constant for $1<i<j$.
\begin{lemma}
\label{mainLemma}
For $1<i<j$, if $\m{1} > -\infty$, we have $\mathbb{E}[N_{i,j}(t)]\leq M'\frac{p-\Delta(2p-1)}{(2p-1)^{2}}$, where $M'$ is the number of cells $i$ with $\m{i} > -\infty$, and $\Delta=\min_{s=1,\cdots M}\{\m{1}-\m{s}\}\leq 0$.
\end{lemma}
\begin{proof}
Let $1<i<j$.
Let $D_{i,j}(t)$ be an indicator function equal to $1$ if and only if we pull arms $i$ and $j$ at time $t$.
Given that we pull arm $i$, we can only also pull arm $j$ when there is a cell $C_{s}$ under which $j$ is the best arm and for which $m_{s}(t) \ge m_{1}(t)$.
Moreover, under the assumption that $\m{1}>-\infty$, $m_{s}(t) \ge m_{1}(t)$ is only possible if $\m{s}>-\infty$.
Thus, $D_{i,j}(t) = 1$ implies
$\max_{s:B(s)=j,\m{s}>-\infty} m_{s}(t)\geq m_{1}(t)$.
Adopting the convention here and in the rest of the proof that maxima and sums over sets of cells are taken only over those cells with $\m{s} > -\infty$, we have
\begin{align}
D_{i,j}(t)
&= D_{i,j}(t)\cdot \mathbbm{1}\left\{\max_{s:B(s)=j}m_{s}(t)\geq m_{1}(t)\right\} \nonumber \\
&\leq D_{i,j}(t)\sum_{s:B(s)=j}\mathbbm{1}\{m_{s}(t)\geq m_{1}(t)\} \nonumber \\
&= D_{i,j}(t)\sum_{s:B(s)=j}\mathbbm{1}\left\{\sum_{(i^{'},j^{'})\in J_{s}}q_{i^{'},j^{'}}(t)+\m{s}\geq \sum_{(i^{'},j^{'})\in J_{1}}q_{i^{'},j^{'}}(t)+\m{1}\right\} \nonumber \\
&=D_{i,j}(t)\sum_{s:B(s)=j}\mathbbm{1}\left\{\sum_{(i^{'},j^{'})\in J_{s}\setminus J_{1}}q_{i^{'},j^{'}}(t)+\m{s}\geq \sum_{(i^{'},j^{'})\in J_{1}\setminus J_{s}}q_{i^{'},j^{'}}(t)+\m{1}\right\} \nonumber \\
&=D_{i,j}(t)\sum_{s:B(s)=j}\mathbbm{1}\left\{\sum_{(i^{'},j^{'})\in J_{s}\setminus J_{1}} q_{i^{'},j^{'}}(t)-q_{j^{'},i^{'}}(t) \geq \m{1}-\m{s}\right\} \nonumber \\
&\leq D_{i,j}(t)\sum_{s:B(s)=j}\mathbbm{1}\left\{\sum_{(i^{'},j^{'})\in J_{s}\setminus J_{1}} q_{i^{'},j^{'}}(t)-q_{j^{'},i^{'}}(t) \geq \Delta\right\}, \nonumber
\end{align}
where the fourth equation holds because $J_s$ has the property that $(i',j') \in J_s \iff (j',i') \notin J_s$, and similarly for $J_1$. Thus,
$(i^{'},j^{'})\in J_{s}\setminus J_{1}
\iff i',j' \in J_s\text{ and } i',j' \notin J_1
\iff j',i' \notin J_s\text{ and } j',i' \in J_1
\iff (j^{'},i^{'})\in J_{1}\setminus J_{s}$.
Thus, we have
\begin{align}
&N_{i,j}(t) =\sum_{k=1}^{t}D_{i,j}(k) \leq
D_{i,j}(k)
\sum_{s:B(s)=j}\sum_{k=1}^{t}\mathbbm{1}\left\{\sum_{(i^{'},j^{'})\in J_{s}\setminus J_{1}} q_{i^{'},j^{'}}(k)-q_{j^{'},i^{'}}(k) \geq \Delta \right\}.\nonumber
\end{align}
Fix an $s$ with $B(s)=j$ and
let $Z(k)=\sum_{(i^{'},j^{'})\in J_{s}\setminus J_{1}} q_{i^{'},j^{'}}(k)-q_{j^{'},i^{'}}(k)$, so that
\begin{equation*}
N_{i,j}(t) \leq
\sum_{s:B(s)=j}
\sum_{k=1}^{t} D_{i,j}(k) \cdot \mathbbm{1}\left\{Z(k) \geq \Delta \right\}.
\end{equation*}
We observe that $Z(k)$ is like a random walk, except that changes in only some time periods. We now describe the conditional distribution of $Z(k+1)$ given the history up to time $k$. Later, we will refer to the $\sigma$-algebra generated by this history as $mathcal{H}_k$.
\begin{itemize}
\item
If the arms $X_{k,0}$, $X_{k,1}$ that we pull satisfy
$(X_{k,0},X_{k,1})\in J_{s}\setminus J_{1}$,
then $Z(k+1) \in \{Z(k)-1,Z(k)+1\}$
and the conditional probability that
$Z(k+1)=Z(k)-1$ is $p_{X_{k,1},X_{k,0}}\geq p$.
This lower bound holds because $(X_{k,0},X_{k,1})\notin J_{1}$ implies
$X_{k,1} < X_{k,0}$.
\item
Similarly, if
$(X_{k,1},X_{k,0})\in J_{s}\setminus J_{1}$,
then $Z(k+1) \in \{Z(k)-1,Z(k)+1\}$ as before,
and the conditional probability that
$Z(k+1)=Z(k)-1$ is $p_{X_{k,0},X_{k,1}}\geq p$,
because
$(X_{k,1},X_{k,0})\notin J_{1}$ implies
$X_{k,0} < X_{k,1}$.
\item Otherwise, if neither
$(X_{k,0},X_{k,1})$
nor
$(X_{k,1},X_{k,0})$ is in $J_{s}\setminus J_{1}$,
then $Z(k+1)=Z(k)$.
\item The definition of $J_1$ prevents having both $(X_{k,0},X_{k,1})$ and $(X_{k,1},X_{k,0})$ in $J_{s}\setminus J_{1}$.
\end{itemize}
When $D_{i,j}(k) = 1$, so that we pull arms $i$ and $j$ (either $X_{t,0}=i$ and $X_{t,1}=j$ or vice versa) we will be in one of the first two cases,
because $B(s)=j$ implies cell $s$ considers $j$ to be the best arm, and so $(j,i) \in J_s$, and $i<j$ implies $(j,i) \notin J_1$. Thus, $D_{i,j}(k) = 1$ implies $Z(k+1) \ne Z(k)$, and we have
\begin{equation*}
N_{i,j}(t) \leq
\sum_{s:B(s)=j}
\sum_{k=1}^{t} \mathbbm{1}\left\{Z(k+1) \ne Z(k), Z(k) \geq \Delta \right\}.
\end{equation*}
We will perform a random time change to study the dynamics over only those time periods where $Z(k)$ changes. Define $\tau_0 = 0$, $\tau_{m}=\min_{k}\{k>\tau_{m-1},Z(k)\neq Z(k+1)\}$.
Because the event $Z(k) \neq Z(k+1)$ is measurable given the history at time $k$, $\mathcal{H}_k$, as described in the dynamics of $Z(\cdot)$ above, each $\tau_m$ is a stopping time.
Define $\zeta=\inf\{m:\tau_{m}=\infty\}$,
which is the lifetime of the random change of time.
We have,
\begin{equation}
N_{i,j}(t) \leq
\sum_{s:B(s)=j}
\sum_{m=1}^{\zeta-1} \mathbbm{1}\left\{Z(\tau_m) \geq \Delta \right\}.
\label{eq:proof1}
\end{equation}
We let $W(m)=Z(\tau_m)$ for $m < \zeta$ (i.e., $m$ with $\tau_m < \infty$), and $W(m) = W(m-1) + \epsilon_m$ for $m\ge \zeta$, where $\epsilon_m$ are iid random variables taking value $-1$ with probability $p$ and value $1$ with probability $1-p$.
Observe that $\zeta$ is measurable with respect to $\mathcal{H}_\infty$, so that the event $m < \zeta$ is measurable with respect to $\mathcal{H}_{\tau_m}$. We define an augmented filtration, letting $\mathcal{F}_m$ to be the $\sigma$-algebra generated by
$\mathcal{H}_{\tau_{\min(m,\zeta)}}$ and $(\epsilon_{m'} : m' \le m)$.
With this construction, $W(m+1) - W(m) \in \{-1,+1\}$ and
$P\left(W(m+1) = W(m)-1 | \mathcal{F}_m \right) \ge p$.
Thus, by Lemma~1,
\begin{equation*}
\sum_{m=1}^{\zeta} \mathbbm{1}\left\{Z(\tau_m) \geq \Delta \right\}
= \sum_{m=1}^{\zeta} \mathbbm{1}\left\{W(m) \geq \Delta \right\}
\le \sum_{m=1}^{\infty} \mathbbm{1}\left\{W(m) \geq \Delta \right\}
\le \frac{p-\Delta(2p-1)}{(2p-1)^{2}}.
\end{equation*}
Combining this with \eqref{eq:proof1} and using the fact that the number of cells with $\m{s} > -\infty$, $M'$, bounds the sum over $s$, we obtain our result.\qedhere
\end{proof}
Based on Lemma~\ref{mainLemma} and a union bound, we obtain our main theorem:
\begin{theorem}
\label{thm:1}
Let $\Lambda=u(\theta,A_{1})-u(\theta,A_{N})$.
If $\m{1}>-\infty$, \CTB's expected cumulative regret is bounded by $\frac{(N-1)(N-2)}{2}M'\frac{p-\Delta(2p-1)}{(2p-1)^{2}}\Lambda$.
\end{theorem}
In general, $M'$ can be as large as $2^N$. However, as discussed above, we may set $\m{i}=-\infty$ for all the empty cells and assign finite $\m{i}$ to empty cells (\CTBone). In this setting, since each cell assigns a ranking over arms and different cells give different rankings, we can bound $M'$ by the number of permutations of $N$ arms, $N!$. Moreover, when the utility function is linear and $\prefdim=\armdim$, results in \cite{jamieson2011active} show $M'$ is $O(N^{2\prefdim})$.
\section{Computation for Decomposable $m_i$}
\label{Imple}
\CTB\ achieves a constant expected cumulative regret. However, a naive implementation of Algorithm~\ref{algo1} requires a great deal of memory to store $m_{i}(t)$ for each cell, which makes it computationally challenging for problems with many arms. In this section, we consider a special case of \CTB\ where $m_i(0)$ can be expressed in terms of an initial score for each pair of arms. Specifically, we suppose that there exists a $r_{i,j}$ such that
\begin{equation}
m_k(0)=\sum_{(i,j)\in J_{k}}r_{i,j} \quad \forall k.
\label{eq:decompose}
\end{equation}
Here $r_{i,j}$ can be interpreted as a prior indicating the extent to which we believe that arm $i$ is preferred over arm $j$. In this special case, we describe an efficient computation method that scales to problems with many arms.
Instead of storing $m_{i}(t)$, this method
stores $r_{i,j}$ and $q_{i,j}(t)$ and uses them to reconstruct $m_{i}(t)$ with Equation~\ref{reconstruct}.
Then, Steps 1 and 2 in Algorithm~\ref{algo1} are written as optimization problems in which $m_i(t)$ is replaced by this expression in terms of $q_{i,j}(t)$ and $r_{i,j}$. Toward this end, let $e_{i,j}$ denote a binary variable that will take value $e_{i,j}=1$ if we are to select a cell in $H_{i,j}$ and $0$ otherwise. Then, based on Equation~\ref{reconstruct}, maximizing $m_{i}(t)$ is equivalent to maximizing $\sum_{i,j:i\neq j}e_{i,j}\times (q_{i,j}(t)+r_{i,j})$.
To find the best arm suggested by $\mathop{\mathrm{argmax}}_{i}m_{i}(t)$ in Step~1, and suggested by a similar argmax in Step~2, it is sufficient to find $\max_{i:B(i)=k} m_{i}(t)$ for each arm $k$. This is the cell with largest $m_{i}(t)$ among those that believe $k$ is best. This problem is:
\begin{align}
\label{IP}
\begin{split}
\text{maximize } \displaystyle &\sum_{i,j:i\neq j} e_{i,j}\times (q_{i,j}(t)+r_{i,j}) \\
\text{subject to } \displaystyle
&e_{k,j}=1,\ \forall j\neq k\\
&e_{i,j}+e_{j,i}=1,\ i,j=1 ,..., N, i\neq j\\
&e_{i,j} \in \{0,1\},\ \forall i\neq j
\end{split}
\end{align}
There are three conditions in Equation~\ref{IP}. The first condition is $e_{k,j}=1$ $\forall j\neq k$, which means cell $C_{\ell}$ that satisfies the first condition must lie in the winning space $H_{k,j}$, $\forall j\neq k$. In other words, $C_{\ell}$ ranks arm $A_{k}$ better than any others and thus $B(\ell)=k$. The second and third condition together guarantee that cell $C_{\ell}$ either belongs to $H_{i,j}$ or $H_{j,i}$.
Though Equation~\ref{IP} is an integer linear programming problem, which are usually computationally challenging, it is in fact easy to solve: the maximum value of this problem is reached when $e_{i,j}=1$ if $r_{i,j}+q_{i,j}(t) > q_{j,i}(t)+r_{j,i}$ for all $i\neq j$, $e_{i,j}=0$ if this strict inequality is reversed, and breaking ties arbitrarily between the solutions $(e_{i,j}=1, e_{i,j}=0)$ and
$(e_{i,j}=0, e_{i,j}=1)$ for those $i,j$ with equality.
Denote the maximum value of this problem at time $t$ as $f(k,t)$. After knowing $f(k,t)=\max_{B(i)=k} m_{i}(t)$, finding the arm with largest $m_{i}(t)$ in Step 1 is equivalent to finding $\mathop{\mathrm{argmax}}_{k}f(k,t)$. Finding the arm with large $m_{i}(t)$ among those with a different best arm than $X_{t,0}$ in Step~2 is equivalent to finding $\mathop{\mathrm{argmax}}_{k \ne X_{t,0}}f(k,t)$.
For general values of $m_i(0)$ that do not satisfy \eqref{eq:decompose}, finding the largest $m_i(t)$ is computationally challenging. However, in applications, instead of setting $m_i(0)$ directly, we may have some prior information about the probability that the user prefers arm $i$ over arm $j$. This information can be used to construct $r_{i,j}$ since \CTB\ guarantees constant regret regardless of the values that $m_i(0)$ take.
\section{Bayesian Interpretation}
\label{sec:Bayes}
Although our problem is formulated in a frequentist setting, we show here that \CTB\ has a Bayesian interpretation. In this section, we construct a Bayesian posterior on $\theta$ given a prior and given an assumption that $p_{i,j}=q>0.5$ for all $i<j$, where $q$ may be the same or different from $p$, and $p_{i,j}$ may or may not be constant across $i,j$ in reality.
We put a prior distribution $p_0$ on $\theta$, which induces a prior on the identity of the cell containing $\theta$. The prior probability that $\theta$ is in cell $i$ is written $p_0(C_i)$, and is obtained by integrating $p_0$ over $C_i$. Let $p_t(C_i)$ indicate the posterior probability that $\theta$ is in cell $C_i$, at time $t$, given $p_{i,j}=q$ for all $i<j$. The following pair of lemmas give recursive and non-recursive expressions for $p_t$.
\begin{lemma}
\label{lemma:bayes}
For compactness of notation, let $i=X_{t,0}$ and $j=X_{t,1}$.
Then the posterior distribution $p_{t+1}$ is,
\[
p_{t+1}(x) =
\begin{cases}
\frac{p_{t}(x)q}{p_{t}(H_{i,j}(Y_k))q+(1-p_{t}(H_{i,j}(Y_k)))(1-q)} & \quad \text{if } x\in H_{i,j}(Y_t)\\
\frac{p_{t}(x)(1-q)}{p_{t}(H_{i,j}(Y_k))q+(1-p_{t}(H_{i,j}(Y_k)))(1-q)} & \quad \text{if } x\notin H_{i,j}(Y_t)\\
\end{cases}
\]
\end{lemma}
Based on this lemma, we can rewrite the posterior distribution in terms of $m_{i}(t)-m_{i}(0)$.
\begin{lemma}
\label{posterior}
For each cell $C_{i}$, the posterior distribution after t comparison is
\begin{align}
p_{t}(C_{i}) \propto p_{0}(C_{i})q^{m_{i}(t)-m_{i}(0)}(1-q)^{t-m_{i}(t)+m_{i}(0)}. \nonumber
\end{align}
\end{lemma}
We leave the proof of both Lemmas to the appendix.
Lemma~\ref{posterior} allows us to rewrite $p_{t}(C_{i})$ as
\begin{align}
p_{t}(C_{i}) \propto p_{0}(C_{i})q^{m_{i}(t)-m_{i}(0)}(1-q)^{t-m_{i}(t)+m_{i}(0)} \nonumber \\
\propto p_{0}(C_{i})(\frac{q}{1-q})^{m_{i}(t)-m_{i}(0)}. \nonumber
\end{align}
Thus, choosing the cell to maximize the posterior probability is equivalent to choosing the cell to maximize
$\log(p_{0}(C_{i}))+(m_{i}(t)-m_{i}(0))\log\left(\frac{q}{1-q}\right).$
Thus, if
\begin{equation}
\label{eq:mi0}
m_i(0)=\log(p_{0}(C_{i})) \big/ \log\left(\frac{q}{1-q}\right),
\end{equation}
then maximizing the posterior probability that $\theta$ is in $C_i$ is equivalent to maximizing $m_i(t)$,
the first cell selected by \CTB\ is the cell with the largest posterior probability of containing $\theta$, and the second cell selected is the largest among those with a different best arm from the first.
Thus, if one has prior information about the location of $\theta$ and an estimate $q$ of a typical value of $p_{ij}$, then a natural way to set $m_i(0)$ is via \eqref{eq:mi0}.
In addition, since $p_{0}(C_{i})=0$ for empty cells, following \eqref{eq:mi0} also sets $m_i(0)=-\infty$ for these cells as discussed before.
\section{Numerical Experiments}
\label{sec:exp}
In this section, we compare the three variants of \CTB\ described in section~\ref{Methods}, CTB-1, CTB-2, and CTB-3, with three benchmarks: Thompson Sampling, Relative Upper Confidence Bound (RUCB) and Winner-Stays (WS).
\begin{itemize}
\item Thompson sampling uses a posterior distribution over $\theta$ computed by beginning with a prior distribution on the location of $\theta$, and updating it using Bayes rule and knowledge of $p_{i,j}$. At time t, it generates $\theta_{t}$ from this posterior distribution $p_{t}$ and pulls the two arms that $\theta_{t}$ ranks as best and second best. In our implementation, we track the prior/posterior explicitly by storing a probability for each cell. We emphasize that Thompson sampling as we consider it here requires knowledge of $p_{i,j}$ which is not typically not available.
\item RUCB is as described in \cite{zoghi2014relative}. We choose it as our benchmark over other algorithms designed for strong regret from the literature because it works well relative to other algorithms designed for strong regret in previous literature when a Condorcet winner exists, and existence of a Condorcet winner is a consequence of our total order assumption. Though there are algorithms that outperform RUCB in some settings such as CCB and SCB \citep{zoghi2015copeland}, they typically work better when a Condorcet winner does not exist.
\item WS is as described in \cite{chen2017dueling}, and is selectetd because it is designed for the weak regret setting. In our plots, WS-W is the variant of WS designed specifically for weak regret.
\end{itemize}
We consider two experimental settings described below, with results pictured in Figure~\ref{fig:result2}.
\begin{figure*}[!h]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{binary-plot-wrucb.png}
\caption{Binary Regret and Constant $p_{i,j}$}
\label{fig:binary}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=1\textwidth]{utility-plot-wrucb.png}
\caption{Bradley-Terry Regret and $p_{i,j}$}
\label{fig:utility}
\end{subfigure}
\caption{
Performance comparison of the three \CTB\ variants from section~\ref{Methods} against benchmarks WS-W, RUCB and Thompson Sampling (THOM) using simulated datasets. CTB$-3$\ and Thompson sampling use prior information, and in this group CTB$-3$\ performs best. Among the four algorithms that do not use prior information, \CTBone\ performs best. \CTBtwo\ under-performs WS-W in the binary regret setting and for $t=100,200$ in the Bradley-Terry setting, and outperforms WS-W when $t=300,400,500$ in the Bradley-Terry setting.}
\label{fig:result2}
\end{figure*}
Since RUCB performs poorly in both experiments compared with other algorithms, we set the $y$-axis to emphasize the relative performance of the other algorithms. We include a plot over a wider $y$-axis showing RUCB's performance in the supplement.
\subsection{Binary Regret and Constant $p_{i,j}$}
\label{sec:binary}
In this experimental setting, we set $p_{i,j}=0.8$ for all $i<j$. We have $N=20$ arms uniformly generated from the $2$-dimensional unit circle. The preference vector $\theta$ is generated uniformly at random from the $2$-dimensional unit circle. We set regret to $1$ if both of the pulled arms are not optimal, i.e. $u(\theta,A_1)=1$ and $u(\theta,A_i)=0$ for $i\neq 1$. To satisfy our previous assumption that $u(\theta,A_i)$ be distinct across $i$, we may equivalently set $u(\theta,A_i)=i\cdot\epsilon$, and take $\epsilon$ small.
Figure~\ref{fig:binary} shows that \CTBone\ and CTB$-3$\ perform comparably and both outperform WS-W and Thompson Sampling. \CTBtwo\ does not perform as well as WS-W and Thompson Sampling.
Both Thompson sampling and CTB$-3$\ have access to the correct prior and use the true value of $p$ to perform updating.
\subsection{Bradley-Terry Regret and $p_{i,j}$}
\label{sec:utility}
In this experimental setting, we set utility using the Bradley-Terry model described in section~\ref{probForm}. As in the first experimental setting, we have $N=20$ arms on the $2$-dimensional unit circle. Among these arms, $19$ are uniformly generated from $\{x<0,y<0,x^2+y^2=1\}$ and $1$ arm is uniformly generated from $\{x>0, y>0, x^2+y^2=1\}$. The user's preference $\theta$ is also uniformly generated from $\{x>0, y>0, x^2+y^2=1\}$, but the Bayesian algorithms (CTB$-3$ and Thompson sampling) use another less information prior: that $\theta$ is uniform on the unit circle. Thompson sampling performs its update using the true $p_{i,j}$, while CTB$-3$\ uses a rough approximation of $q=0.6$ to set $m_i(0)$ to model the fact that we would not know $p$ or $p_{i,j}$ in practice.
Figure~\ref{fig:utility} shows that both CTB$-3$\ and Thompson Sampling takes advantage of the prior information and the dependence among arms. CTB$-3$\ uses this information more efficiently and significantly outperforms Thompson Sampling. Among the four algorithms (\CTBone, \CTBtwo, RUCB and WS) that do not use prior information, \CTBone\ performs best. Though \CTBtwo\ does not perform as well as WS at $t=100, 200$, it outperforms WS when $t=300,400,500$.
\section{Conclusion}
In this paper, we consider dueling bandits for weak regret, with application to recommender systems and online content recommendation. We formulate a new setting which differs from the traditional dueling bandits in which arms are dependent. We propose an algorithm \CTB, and show it has constant expected cumulative regret and strong empirical performance.
\newpage
|
\section*{Introduction}
Radios are ubiquitous in modern society. Between cellular devices, wearable devices, computing devices, medical devices, and other devices we carry and operate regularly through each day, radio frequency communications have become the most pervasive and convenient way we communicate information in our daily lives. Unfortunately, our spectrum resources are limited and our data needs are growing in a seemingly unbounded manner. Despite this wireless spectrum crunch, our methods for allocation, adaptation and optimization of spectrum use remains very much in the dark ages today. Spectrum is still allocated in a static fashion, and devices are oblivious and unaware of the use of or the availability of resources in their direct vicinity. The field of cognitive radio and dynamic spectrum access have attempted to address this through the introduction of expert systems which attempt to perform spectrum sensing and some degree of characterization of their environment, but their impact has been heavily limited by their inability to generalize to new regions, protocols, emitters, and radio propagation environments.
Generalized policy learning has and continues to be an open challenge in CS and AI for many years, however in recent years advances in reinforcement learning have made massive strides towards the advancement of this field. Recent Work by Minh \cite{mnih2013playing}, Silver \cite{nature-go}, Sutton \cite{sutten_textbook}, and others has begun to demonstrate the ability to learn exceedingly complex and varied tasks using deep neural network based policy function approximation to implement Q-Learning.
To address this problem of learning to rapidly understand the surrounding radio environment, we introduce a radio signal search environment for the recently released Gym RL framework from OpenAI in which to begin evaluating and scoring different approaches.
We also implement a general purpose open source Deep Neural network based Q-Learning function approximation learner for Gym using Keras primitives to learn a policy for rapidly exploring this environment through its set of discrete actions and observations \cite{keras}.
\section{Reinforcement Learning Policy}
We introduce KeRLym \cite{kerlym}, an open source deep reinforcement learning agent collection written in python using Keras to implement GPU optimized deep neural networks on top of Theano \cite{theano} and TensorFlow \cite{tensorflow}. OpenAI recently released Gym \cite{duan2016benchmarking}, a collection of reinforcement learning benchmark environments, an API to easily use them, and a web based high-score board for algorithm comparison. We leverage this API in our reinforcement learner to provide a standard agent interface and to rapidly provide a wide range of tasks we can test its performance and tuning against.
\subsection{Policy Learning}
Since Google Deepmind's Nature paper on Deep-Q Networks \cite{mnih2013playing}, there has been a surge of interest in the capabilities of reinforcement learning use deep neural network policy network approximation. This is an exciting and growing area with much potential improvement for learning algorithms yet to come. For the scope of this work we implement a parametric version of the Deep Q-Learning algorithm along with a Double Q-Learning \cite{van2015deep} implementation in the KeRLym toolbox. We implement a variety of function approximation networks which can be used inside them including dense fully connected networks, convolutional networks similar to those used in the Atari paper, and recurrent networks leveraging LSTM which may improve sequence learning in POMDPs as discussed in \cite{hausknecht2015deep}.
The approaches in are similar for both algorithms, a value function $Q(s,a;\theta)$ is updated using a form of stochastic gradient descnt, SGD, in the form of:
\begin{equation}
\theta_{t+1} = \theta_t + \alpha \left( Y_t, - Q\left( S_t, A_t; \theta_t \right) \right) \nabla_\theta Q(S_t, A_t; \theta_t)
\end{equation}
However, in single Q-learning we directly compute $Y_t$ in a greedy manner using our latest $\theta$ as:
\begin{equation}
Y_t = R_{t+1} + \gamma \underset{a}{ \mathrm{max} } Q\left(S_{t+1}, a; \theta_t\right)
\end{equation}
Whereas in double Q-learning we maintain two sets of weights $\theta$ and $\theta'$ which we alternate between using for decision making and greedy policy update purposes:
\begin{equation}
Y_t = R_{t+1} + \gamma Q \left( S_{t+1}, \underset{a}{ \mathrm{argmax} } Q\left(S_{t+1}, a; \theta_t\right) ; \theta' \right)
\end{equation}
This helps to reduce overestimation value bias and imrpoves policy learning rate and stability for many tasks.
We implement $\epsilon$-greedy learning with a default constant value of 0.1, to choose the greedy policy 90\% of the time, simply to avoid the tuning required with $epsilon$ decay schedules for stability of comparison of this work.
We also implement experience replay, keeping around a memory of 1,000,000 previous actions to draw training samples from in addition to the new experience gained each time-step. We use a learning rate of 0.001 in a Keras Adam \cite{adam} solver, and a discount rate of $\gamma = 0.99$ in our experiments.
Within the KeRLym toolbox we hope to extend the number of solvers available
\subsection{Deep-Q Network Implementation}
Our Q function $Q(s,a,\theta)$ is a Deep Neural Network with random initial parameters $\theta$ implemented in the Keras framework on top of Theano, running on an Nvidia Titan X. We zero the output regression layer weights to reduce initial error in value function output. We start with a similar architecture to the convolutional network used by Mnih et al in \cite{mnih2013playing}, but make changes which show improvement in our domain and account for the input information form.
Since we are passing both scalar stored variables containing sensor information, and contiguous frequency domain values into the value function as the current state, we treat each input configuration value as an independent discrete input with fully connected logic, while we reduce the parameter space and allow frequency domain filters to form and be used shift-invariantly on the power spectrum by using a set of convolutional layers, similar to our approach on raw time-domain samples in \cite{convmodrec}.
Ultimately we concatenate the activations from both of these paths into dense fully connected layers to perform the output regression task for output action-value estimates.
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\textwidth]{net2.jpg}
\caption{Action-Value Network Architecture}
\label{fig:net}
\end{figure}
\section{Radio Search Environment}
\subsection{Environment Overview}
Typical electronic devices such as cellular phones contain at this point highly flexible Radio Frequency Integrated Circuits (RFIC) which allow the frequency tuning and digitization of relatively large arbitrary bands of interest. Typically they are programmed in an exceedingly simple way by a carrier to brute force through a small list of carrier-assigned channels and bandwidth, however they are in fact capable of tuning to relatively arbitrary center frequencies between 100 MHz and 6GHz and providing often powers of two decimations of a 10-20 MHz wide bandwidth.
Instead of brute force search for signals on several carrier centric bands, we propose instead to allow machine learning to derive a general search policy to identify signals providing useful connectivity while optimizing for minimal search time, battery consumption and power usage possible. To do this we boil the search task down into a relatively small set of possible discrete actions which may be taken towards the end-goal.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\textwidth]{actions.png}
\caption{Initial Radio Discrete Action Set}
\label{fig:actionset}
\end{figure}
\subsection{Environment Implementation}
We begin by building an environment for the Gym Reinforcement Learning environment to attempt to mirror our problem statement, and a reasonable set of assumptions for what a real system could do and sense, but at a relatively small scale of complexity for initial work.
We simulate a single radio receiver sampling at a bandwidth of 20 MHz, which can be decimated and re-tuned using the set of discrete actions in \ref{fig:actionset}. The discrete actions we allow are as follows, where we refer to the variables: center frequency (fc), bandwidth (bw), maximum bandwidth (bwmax), minimum bandwidth (bwmin), maximum center frequency (fcmax), and minimum center frequency (fcmin).
\begin{itemize}
\item Freq Down: $fc -= max(bw/2, fc_min)$
\item Freq Up: $fc += min(bw/2, fc_max)$
\item BW Down Left: $bw = max(bw/2, bw_min); fc -= bw/2$
\item BW Down Left: $bw = max(bw/2, bw_min); fc += bw/2$
\item BW Max: $bw = bw_max$
\item Detect: Assert that a signal is in the current window.
\item Finished: Assert that all signals in band have been detected.
\end{itemize}
The environment chooses a random frequency within the band of interest (100MHz to 200MHz in this work) to place a single sinusoidal tone. For each agent observation, it returns a small band-limited window into the environment tuned to the chosen center frequency and bandwidth. The Detect action asserts that there is a signal within the current band either correctly or falsely, Finish assets that we have correctly found the signal and our search path is complete, and bandwidth and frequency actions change our receiver configuration according to the list above. A single optimal path to a solution through the environment might look something like shown in figure \ref{fig:search}. In this case each look window for a time-step is represented by a red bar above the wideband power spectrum plot.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.75\textwidth]{search_plot.png}
\caption{Environment Search Scenario}
\label{fig:search}
\end{figure}
\section{Training Considerations}
There are numerous ways to define penalties and rewards for this search process within the environment which pose a number of different considerations for the training process, we propose 3 potential rewards schemes below.
\begin{table}[!htb]
\centering
\begin{tabular}{||c c c c||}
\hline
Scheme & A & B & C \\ [0.5ex]
\hline\hline
Detect(True) & 1 & 1 & 0 \\
\hline
Detect(False) & 0 & -1 & 0 \\
\hline
BW-(True) & 1 & 1 & 0 \\
\hline
BW-(False) & 0 & 0 & 0 \\
\hline
Finish(True) & 1 & 1 & nfound*depth \\ [1ex]
\hline
Finish(False) & 0 & -1 & 0 \\ [1ex]
\hline
\end{tabular}\\
\caption{Environment Reward and Penalty Schemes}
\end{table}
Oour agent's goal at run-time is to detect the signal present somewhere in the band and localize the signal using BW-L and BW-R actions to zoom in on it, these rewards and penalties are designed to reflect that.
Scheme A results in perhaps the fastest training rate and simples approach towards directly rewarding good actions, Scheme B provides a strong disincentive for false positive actions, but slows down learning, and Scheme C provides a simple final score which requires a more delayed-reward style of learning.
\section{Conclusions and Future Work}
We can plot a number of statistics during the training process which give us insight into how the training is going. Shown in figure \ref{fig:plots} we have the training statistics under Scheme B with early exiting (on Finish(False)). From this graph it is clear that we are learning a relatively clear separation between good and bad action values, as can be seen from the separation in the 3rd plot, and that our reward is growing and our finishing time is growing long enough to succeed some of the time.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\textwidth]{plots_v21.png}
\caption{Plots During Network Training}
\label{fig:plots}
\end{figure}
'
In future work we hope to provide a more comprehensive trade between the trade offs described above, learn a policy which performs at a more satisfying reward level, and and compare the impact of reward/penalty schemes on traditional receiver operating characteristics, ROC, curves for performance. We are excited about the potential in this area and positive this approach will be fruitful.
\noindent
\printbibliography
\nocite{*}
\end{document}
|
\section{Introduction}
In the last 5 years or so, randomized coordinate descent (RCD) methods \cite{ShalevTewari11, Nesterov:2010RCDM, UCDC, PCDM} have become immensely popular in a variety of machine learning tasks, with supervised learning being a prime example. The main reasons behind the rise of RCD-type methods is that they can be easily implemented, have intuitive appeal, and enjoy superior theoretical and practical behaviour when compared to classical methods such as SGD \cite{robbins1951}, especially in high dimensions, and in situations when solutions of medium to high accuracy are needed. One of the most important success stories of RCD is in the domain of training linear predictors via regularized empirical risk minimization (ERM).
The highly popular SDCA algorithm \cite{SDCA} arises as the application of RCD \cite{UCDC} to the {\em dual problem} associated with the (primal) ERM problem\footnote{Indeed, the analysis of SDCA in \cite{SDCA} proceeds by applying the complexity result from \cite{UCDC} to the {\em dual problem}, and then arguing that the same rate applies to the primal suboptimality as well.}. In practice, SDCA is most effective in situations where the number of examples ($n$) exceeds the number of features ($d$). Since the dual of ERM is an $n$ dimensional problem, it makes intuitive sense to apply RCD to the dual. Indeed, RCD can be seen as a randomized decomposition strategy, reducing the $n$ dimensional problem to a sequence of (randomly generated) one-dimensional problems.
However, if the number of features exceeds the number of examples, and especially when the difference is very large, RCD methods \cite{PCDM} have been found very attractive for solving the {\em primal problem} (i.e., the ERM problem) directly. For instance, distributed variants of RCD, such as Hydra \cite{Hydra} and its accelerated cousin Hydra$^2$ \cite{Hydra2} have been successfully applied to solving problems with billions of features.
Recently, a variety of novel primal methods for ERM have been designed, including SAG \cite{SAG}, SVRG \cite{SVRG}, S2GD \cite{S2GD}, proxSVRG \cite{proxSVRG}, mS2GD \cite{mS2GD}, SAGA \cite{SAGA}, MISO \cite{MISO} and S2CD~\cite{S2CD}. As SDCA, all these methods improve dramatically on SGD \cite{robbins1951} as a benchmark, which they achieve by employing one of a number of variance-reduction strategies. However, these methods have essentially identical identical theoretical behavior to SDCA, including the property that these methods thrive in the data-laden domain (i.e., $n\gg d$). In this sense, in our comparison of primal vs dual RCD, these methods should be viewed as ``dual methods''.
\subsection{Contributions}
In this paper we provide the first joint study of these two approaches---applying RCD to the primal vs dual problems---and we do so in the context of L2-regularized ERM. First, we show through a rigorous theoretical analysis that for dense data, the intuition that the primal approach is better than the dual approach when $n\geq d$, and vice versa, is precisely correct. However, we show that for sparse data, this does not need to be the case: primal RCD can significantly outperform dual RCD even if $d\ll n$, and vice versa, dual RCD can be much faster than primal RCD even if $n\ll d$. In particular, we identify that the face-off between primal and dual RCD boils down to the comparison of as single quantity associated with the data matrix and its transpose. Moreover, we show that, surprisingly, a single sampling strategy minimizes both the (bound on the) number of iterations and the overall expected complexity of RCD. Note that the latter complexity measure takes into account also the average cost of the iterations, which depends on the structure and sparsity of the data, and on the sampling strategy employed. We confirm our theoretical findings using extensive experiments with both synthetic and real data sets.
\section{Primal and Dual Formulations of ERM}
Let $\mathbf{X}\in \mathbb{R}^{d\times n}$ be a data matrix, with $n$ referring to the number of examples and $d$ to the number of features. With each example $\mathbf{X}_{:j}\in \mathbb{R}^d$ we associate a loss function $\phi_j:\mathbb{R}\to \mathbb{R}$, and pick a regularization constant $\lambda>0$. The key problem of this paper is the L2-regularized ERM problem
\begin{equation}
\label{def:Primal}
\min_{w \in \mathbb{R}^d} \left[ P(w) ~{:=}~ \frac{1}{n}\sum_{j=1}^n \phi_j(\langle \mathbf{X}_{:j}, w \rangle) + \frac{\lambda}{2}\|w\|_2^2 \right],
\end{equation}
where $\langle\cdot, \cdot\rangle$ denotes the standard Euclidean inner product and $\|w\|_2~{:=}~\sqrt{\langle w,w\rangle}$. We refer to \eqref{def:Primal} as the {\em primal problem}. We assume throughout that the functions $\{\phi_j\}$ are convex and $\beta$-smooth: \begin{equation} \label{eq:phi_smooth}
\phi_j(s) + \phi_j'(s)t \leq \phi_j(s + t) \leq \phi_j(s) + \phi_j'(s)t + \frac{\beta}{2}t^2, \qquad \text{for all} \qquad s,t\in \mathbb{R}.
\end{equation}
The {\em dual problem} of \eqref{def:Primal} is
\begin{equation}\label{def:Dual}
\max_{\alpha \in \mathbb{R}^n} \left[ D(\alpha) ~{:=}~ - \frac{1}{2\lambda n^2}\left\| \mathbf{X}\alpha\right\|_2^2 - \frac{1}{n}\sum_{j=1}^n \phi_j^*(-\alpha_j) \right],
\end{equation}
where $\phi_j^*:\mathbb{R}\to \mathbb{R}$ is the convex conjugate of $\phi_j$, defined by $
\phi_j^*(s) ~{:=}~ \sup \{ st - \phi_j(t) \;:\; t \in \mathbb{R}\}.$
It is well known that \cite{SDCA, Quartz} that $P(w) \geq D(\alpha)$ for every pair $(w, \alpha) \in \mathbb{R}^d \times \mathbb{R}^n$ and $P(w^*) = D(\alpha^*)$. Moreover, the primal and dual optimal solutions, $w^*$ and $\alpha^*$, respectively, are unique, and satisfy the relations $w^* = \frac{1}{\lambda n} \mathbf{X} \alpha^*$ and $\alpha_j^* = \phi_j'(\langle \mathbf{X}_{:j}, w^* \rangle)$ for all $j \in [n] ~{:=}~ \{1, \dots, n \}$, which also uniquely characterize them.
\section{Primal and Dual RCD}
In its general ``arbitrary sampling'' form \cite{NSync}, RCD applied to the primal problem \eqref{def:Primal} has the form \begin{equation}\label{eq:RCD-primal-text}w_i^{k+1} \leftarrow w_i^k - \frac{1}{u'_i}\nabla_i P(w^k)\quad \text{for} \quad i\in S_k, \qquad \qquad w_i^{k+1} \leftarrow w_k^k \quad \text{for} \quad i\notin S_k,\end{equation}
where $u'_1,\dots,u'_d>0$ are parameters of the method and $\nabla_i P(w) = \frac{1}{n}\sum_{j=1}^n \phi_j'(\langle \mathbf{X}_{:j}, w\rangle ) \mathbf{X}_{ij} + \lambda w_i$ is the $i$th partial derivative of $P$ at $w$. This update is performed for a random subset of the coordinates $i\in S_k\subseteq [d]$ chosen in an i.i.d.\ fashion according to some sampling $\hat{S}_P$. The parameters $u'_i$ are computed ahead of the iterative process and need to be selected carefully in order for the method to work \cite{NSync, ESO}. Specifically, one can set $u'_i ~{:=}~ \frac{\beta}{n}u_i + \lambda$, where $u=(u_1,\dots,u_d)$ is chosen so as to satisfy the ESO (expected separable overapproximation) inequality \begin{equation}\label{eq:ESO_P}\mathbf{P} \circ \mathbf{X} \mathbf{X}^\top \preceq {\rm Diag}(p\circ u),\end{equation} where $\mathbf{P}$ is the $d\times d$ matrix with entries $\mathbf{P}_{ij} = \mathbb{P}(i\in \hat{S}, j\in \hat{S})$, $p = {\rm Diag}(\mathbf{P})\in \mathbb{R}^d$ and $\circ$ denotes the Hadamard (element-wise) product of matrices. The method is formally described as Algorithm~\ref{alg:nsync}.
\begin{algorithm}[H]
\caption{Primal RCD: NSync \cite{NSync}} \label{alg:nsync}
\begin{algorithmic}
\STATE \textbf{Input:} initial iterate $w^0 \in \mathbb{R}^d$; sampling $\hat{S}_P$; ESO parameters $u_1, \dots, u_d > 0$
\STATE \textbf{Initialize:} $z^0 = \mathbf{X}^\top w^0$
\FOR{$k=0,1,\dots $}
\STATE Sample $S_k \subseteq [d]$ according to $\hat{S}_P$
\FOR{$i \in S_k$}
\STATE Compute $\Delta^{k}_i = -\frac{n}{\beta u_i + \lambda n}\left( \frac{1}{n}\sum_{j=1}^n \phi_j'(z_j^k) \mathbf{X}_{ij} + \lambda w_i^k \right)$
\STATE Update $w_i^{k+1} = w_i^k + \Delta_i^k$
\ENDFOR
\FOR{$i\notin S_k$}
\STATE $w_i^{k+1} = w_i^k$
\ENDFOR
\STATE Update $z^{k+1} = z^k + \sum_{i \in S_k} \Delta_i^k\mathbf{X}_{i:}^\top $
\ENDFOR
\end{algorithmic}
\end{algorithm}
When applying RCD to the dual problem \ref{def:Dual}, we can't proceed as above since the functions $\phi_j^*$ are not necessarily smooth, and hence we can't compute the partial derivatives of the dual objective. The standard approach here is to use a proximal variant of RCD \cite{PCDM}. In particular, Algorithm~\ref{alg:quartz} has been analyzed in \cite{Quartz}. Like Algorithm~\ref{alg:nsync}, Algorithm~\ref{alg:quartz} is also capable to work with an arbitrary sampling, which in this case is a random subset of $[n]$. The ESO parameters $v=(v_1,\dots,v_j)$ must in this case satisfy the ESO inequality \begin{equation}\label{eq:ESO_D}\mathbf{Q}\circ \mathbf{X}^\top \mathbf{X} \preceq {\rm Diag}(q\circ v),\end{equation} where $\mathbf{Q}$ is an $n\times n$ matrix with entries $\mathbf{Q}_{ij} = \mathbb{P}(i\in \hat{S}_D, j\in \hat{S}_D)$ and $q = {\rm Diag}(\mathbf{Q})\in \mathbb{R}^n$.
\begin{algorithm}[H]
\caption{Dual RCD: Quartz \cite{Quartz}} \label{alg:quartz}
\begin{algorithmic}
\STATE \textbf{Input:} initial dual variables $\alpha^0\in \mathbb{R}^n$, sampling $\hat{S}_D$; ESO parameters $v_1, \dots, v_n > 0$
\STATE \textbf{Initialize:} set $w^0 = \frac{1}{\lambda n} \mathbf{X} \alpha^0 $
\FOR{$k=0, 1, \dots$}
\STATE Sample $S_k \subseteq [n]$ according to $\hat{S}_D$
\FOR{$j \in S_k$}
\STATE Compute $\Delta_j^k = \argmax_{h \in \mathbb{R}} \left\{ -\phi_j^*(-(\alpha_j^k + h)) - h\langle \mathbf{X}_{:j}, w^k \rangle - \frac{v_j h^2}{2 \lambda n} \right\}$
\STATE Update $\alpha_j^{k+1} = \alpha_j^k + \Delta_j^k $
\ENDFOR
\FOR{$j\notin S_k$}
\STATE $\alpha_j^{k+1} = \alpha_j^k$
\ENDFOR
\STATE Update $w^{k+1} = w^k + \frac{1}{\lambda n} \sum_{j \in S_k} \Delta_j^k \mathbf{X}_{:j}$
\ENDFOR
\end{algorithmic}
\end{algorithm}
If we assume that $|\hat{S}_P|=1$ (resp.\ $|\hat{S}_D|=1$) with probability 1 (i.e., of the samplings are ``serial''), then it is trivial to observe that \eqref{eq:ESO_P} (resp.\ \eqref{eq:ESO_D}) holds with
\begin{equation}\label{eq:98s98hsd}u={\rm Diag}(\mathbf{X} \mathbf{X}^\top) \qquad \text{(resp. } v={\rm Diag}(\mathbf{X}^\top \mathbf{X}) \text{)}.\end{equation}
Easily computable expressions for $u$ (resp.\ $v$) for more complicated samplings can be found in \cite{ESO}.
\section{Iteration Complexity and Total Arithmetic Complexity}
In this section we give expressions for the total expected arithmetic complexity of the two algorithms.
\subsection{Number of iterations} Iteration complexity of Algorithms~\ref{alg:nsync} and \ref{alg:quartz} is described in the following theorem.
We include a proof sketch in the appendix.
\begin{theorem} \label{thm:primal_and_dual} \textbf{(Complexity: Primal vs Dual RCD)}
Let $\{\phi_j\}$ be convex and $\beta$-smooth.
\\ (i) If $\hat{S}_P$ is proper (i.e., $p_i>0$ for all $i$), and $u$ satisfies \eqref{eq:ESO_P}, then iterates of primal RCD satisfy
\begin{equation}\label{eq:K_P}k \geq K_P = K_P( \hat{S}_P,\epsilon)~{:=}~ \max_{i \in [d]} \left( \frac{\beta u_i + \lambda n }{p_i \lambda n }\right) \log\left( \frac{C_P}{\epsilon} \right) \quad \Rightarrow \quad \mathbb{E}[P(w^k) - P(w^*)] \leq \epsilon,\end{equation}
where $C_P$ is a constant depending on $w^0$ and $w^*$.\\
(ii) If $\hat{S}_D$ is proper (i.e., $q_i>0$ for all $i$), and $v$ satisfies \eqref{eq:ESO_D}, then iterates of dual RCD satisfy
\begin{equation}\label{eq:K_D}k \geq K_D = K_D(\hat{S}_D,\epsilon) ~{:=}~ \max_{j \in [n]} \left( \frac{\beta v_j + \lambda n }{q_j \lambda n }\right) \log\left( \frac{C_D}{\epsilon} \right) \quad \Rightarrow \quad \mathbb{E}[P(w^k) - P(w^*)] \leq \epsilon, \end{equation}
where $C_D$ is a constant depending on $w^0$ and $w^*$.
\end{theorem}
For the dual method a stronger guarantee can be established (see \cite{Quartz}): as soon as $k\geq L_D(\hat{S}_D,\epsilon)$, we have $\mathbb{E}[P(w^k)-D(\alpha^k)]\leq \epsilon$. Clearly, this stronger result implies the claim in part ii) of the above theorem.
\subsection{Average cost of a single iteration}
Let $\|\cdot\|_0$ be the number of nonzeros in a matrix/vector. It is easy to observe that the average cost of a single iteration of Algorithm~\ref{alg:nsync} is \begin{equation}\label{eq:W_P} W_P(\mathbf{X},\hat{S}_P)~{:=}~ \mathcal{O}\left(\mathbb{E}\left[ \sum_{i \in \hat{S}_P} \|\mathbf{X}_{i:}\|_0\right]\right)= \mathcal{O}\left(\sum_{i=1}^d p_i \|\mathbf{X}_{i:}\|_0\right),\end{equation}
and for Algorithm~\ref{alg:quartz} it is \begin{equation}\label{eq:W_D} W_D(\mathbf{X},\hat{S}_D)~{:=}~ \mathcal{O}\left(\mathbb{E}\left[ \sum_{j \in \hat{S}_D} \|\mathbf{X}_{:j}\|_0\right]\right)= \mathcal{O}\left(\sum_{j=1}^n q_j \|\mathbf{X}_{:j}\|_0\right).\end{equation}
We remark that the constant hidden in $\mathcal{O}$ may be larger for Algorithm~\ref{alg:nsync} than for Algorithm~\ref{alg:quartz}. The reason for this is that for Algorithm~\ref{alg:nsync} we compute the one-dimensional derivative $\phi_j'$ for every nonzero term in the sum, while for Algorithm~\ref{alg:quartz} we do this only once. Depending on the loss $\phi_j$, this may lead to slower iterations. For example, if $\phi_j$ isthe logistic loss, experimentation shows that the constant is around 50. On the other hand, if $\phi_j$ is the squared loss, the constant is 1.
\subsection{Total complexity}
By combining the bounds on the number of iterations provided by Theorem~\ref{thm:primal_and_dual} with the formulas \eqref{eq:W_P} and \eqref{eq:W_D} for the cost of a single iteration we obtain the following expressions for the {\em total complexity} of the two algorithms, where we ignore the logarithmic terms and drop the $\tilde{O}$ symbol:\begin{equation}\label{eq:T_P} T_P = T_P(\mathbf{X},\hat{S}_P) ~{:=}~ K_P W_P \overset{\eqref{eq:K_P} + \eqref{eq:W_P}}{=} \left( \max_{i \in [d]} \frac{\beta u_i + \lambda n }{p_i \lambda n }\right) \left(\sum_{i=1}^d p_i \|\mathbf{X}_{i:}\|_0\right), \end{equation} \begin{equation}\label{eq:T_D} T_D = T_D(\mathbf{X},\hat{S}_D) ~{:=}~ K_D W_D \overset{\eqref{eq:K_D} + \eqref{eq:W_D}}{=} \left( \max_{j \in [n]} \frac{\beta v_j + \lambda n }{q_j \lambda n }\right) \left(\sum_{j=1}^n q_j \|\mathbf{X}_{:j}\|_0\right).\end{equation}
\section{Choosing a Sampling that Minimizes the Total Complexity}
In this section we identify the {\em optimal sampling} in terms of the {\em total complexity.} This is different from previous results on {\em importance sampling}, which neglect to take into account the cost of the iterations \cite{NSync,Quartz,IProx-SDCA, ImportanceSrebro}. For simplicity, we shall only consider {\em serial} samplings, i.e., samplings which only pick a single coordinate at a time. The situation is much more complicated with non-serial samplings where first importance sampling results have only been derived recently \cite{ISM}.
\subsection{Uniform Sampling}
The simplest serial sampling is the {\em uniform sampling}: it selects every coordinate with the same probability, i.e. $p_i = 1/d, ~\forall i \in [d]$ and $q_j = 1/n, ~\forall j \in [n]$. In view of \eqref{eq:T_P}, \eqref{eq:T_D} and \eqref{eq:98s98hsd}, we get
\[T_P =\|\mathbf{X}\|_0 \left( 1 + \frac{\beta}{\lambda n}\max_{i \in [d]} \|\mathbf{X}_{i:}\|_2^2 \right)\] and \[T_D = \|\mathbf{X}\|_0\left( 1 + \frac{\beta}{\lambda n}\max_{j \in [n]} \|\mathbf{X}_{:j}\|_2^2 \right).\]
We can now clearly see that whether $T_P\leq T_D$ or $T_P\geq T_D$ depends does not simply depend on $d$ vs $n$, but instead depends on the relative value of the quantities $\max_{i \in [d]} \|\mathbf{X}_{i:}\|_2^2$ and $\max_{j \in [n]} \|\mathbf{X}_{:j}\|_2^2$. Having said that, we shall not study these quantities in this paper. The reason for this is that for the cake of brevity, we shall instead focus on comparing the primal and dual RCD methods for optimal sampling which minimizes the total complexity, in which case we will obtain different quantities.
\subsection{Importance Sampling}
By {\em importance sampling} we mean the serial sampling $\hat{S}_P$ (resp.\ $\hat{S}_D$) which minimizes the bounds $K_P$ in \ref{eq:K_P} (resp.\ $K_D$ in \eqref{eq:K_D}). It can easily be seen (see also \cite{NSync}, \cite{Quartz}, \cite{IProx-SDCA}), that importance sampling probabilities are given by \begin{equation} \label{def:importance_sampling}
p_i^* = \frac{\beta u_i + \lambda n}{\sum_{l} ( \beta u_l + \lambda n)} \qquad \mbox{and} \qquad q_j^* = \frac{\beta v_j + \lambda n}{\sum_{l} (\beta v_l + \lambda n)}.
\end{equation}
On the other hand, one can observe that the average iteration cost of importance sampling may be larger than the average iteration cost of uniform serial sampling. Therefore, it is a natural question to ask, whether it is necessarily better. In view of \eqref{eq:T_P}, \eqref{eq:T_D} and \eqref{def:importance_sampling}, the total complexities for importance sampling are
\begin{equation} \label{eq:MMM} T_P =\|\mathbf{X}\|_0 + \frac{\beta}{\lambda n}\sum_{i=1}^d \|\mathbf{X}_{i:}\|_0 \|\mathbf{X}_{i:}\|_2^2 , \qquad T_D = \|\mathbf{X}\|_0 + \frac{\beta}{\lambda n}\sum_{j=1}^n \|\mathbf{X}_{:j}\|_0\|\mathbf{X}_{:j}\|_2^2. \end{equation}
Since a weighted average is smaller than the maximum, the total complexity of both methods with importance sampling is always better than with uniform sampling. However, this does not mean that importance sampling is the sampling that minimizes total complexity.
\subsection{Optimal Sampling}
The next theorem states that, in fact, importance sampling {\em does} minimize the total complexity.
\begin{theorem}\label{thm:runtime} The optimal serial sampling (i.e., the serial sampling minimizing the total expected complexity $T_P$ (resp, $T_D$)) is the importance sampling \eqref{def:importance_sampling}.
\end{theorem}
\section{The Face-Off}
In this section we investigate the two quantities in \eqref{eq:MMM}, $T_P$ and $T_D$, measuring the total complexity of the two methods as functions of the data $\mathbf{X}$. Clearly, it is enough to focus on the quantities
\begin{equation} \label{eq:two_quantities} C_P(\mathbf{X}) ~{:=}~ \sum_{i=1}^d \|\mathbf{X}_{i:}\|_0 \| \mathbf{X}_{i:}\|^2\qquad \text{and} \qquad C_D(\mathbf{X}) ~{:=}~ \sum_{j=1}^n \|\mathbf{X}_{:j}\|_0 \| \mathbf{X}_{:j}\|^2.\end{equation}
We shall ask questions such as: when is $C_P(\mathbf{X})$ larger/smaller than $C_D(\mathbf{X})$, and by how much. In this regard, it is useful to note that $C_P(\mathbf{X}) = C_D(\mathbf{X}^\top)$. Our first result gives tight lower and upper bounds on their ratio.
\begin{theorem}\label{thm:Bound_Any_Data} For any $\mathbf{X}\in \mathbb{R}^{d \times n}$ with no zero rows or columns, we have the bounds $\|\mathbf{X}\|_F^2 \leq C_P(\mathbf{X}) \leq n\|\mathbf{X}\|_F^2$ and $\|\mathbf{X}\|_F^2\leq C_D(\mathbf{X}) \leq d\|\mathbf{X}\|_F^2$. It follows that $ 1/d\leq C_P(\mathbf{X}) / C_D(\mathbf{X}) \leq n.$ Moreover, all these bounds are tight.
\end{theorem}
Since $C_P(\mathbf{X})$ (resp. $C_D(\mathbf{X})$) can dominate the expression \eqref{eq:T_P} (resp.\ \eqref{eq:T_D}) for total complexity, it follows that, depending on the data matrix $\mathbf{X}$, {\em the primal method can be up to $d$ times faster than the dual method, and up to $n$ times slower than the dual method.}
\subsection{Random Data and Dense Data}
Assume now that the entries of $\mathbf{X}$ are chosen in an i.i.d.\ manner from some distribution with mean $\mu$ and variance $\sigma^2$. While this is not a realistic scenario, it will help us build intuition about what we can expect the quantities $C_P(\mathbf{X})$ and $C_D(\mathbf{X})$ to look like. A simple calculation reveals that $\mathbb{E}[C_P(\mathbf{X})] = dn\sigma^2 + dn^2\mu^2$, and $\mathbb{E}[C_D(\mathbf{X})] = dn\sigma^2 + nd^2\mu^2$. Hence, \[\mathbb{E}[C_P(\mathbf{X})]\leq \mathbb{E}[C_D(\mathbf{X})]\] precisely when $n\leq d$, which means that the primal method is better when $n < d$ and the dual method is better when $n>d$.
If $\mathbf{X}$ is a dense deterministic matrix ($\mathbf{X}_{ij}\neq 0$ for all $i,j$), then $C_P(\mathbf{X}) = n \|\mathbf{X}\|_F^2$ and $C_D(\mathbf{X})=d \|\mathbf{X}\|_F^2$, and we reach the same conclusion as for random data: everything boils down to $d$ vs $n$.
\subsection{Binary Data} \label{sec:binary}
In this part we identify a class of data matrices for which one can have $C_P\leq C_D$ even if $d\ll n$. This class is by no means exhaustive, and serves as an example which we use to illustrate the phenomenon.
Let $\mathbb{B}^{d\times n}$ denote the set of $d\times n$ matrices $\mathbf{X}$ with (signed) binary elements, i.e., with $\mathbf{X}_{ij} \in \{-1,0,1\}$ for all $i,j$. For $\mathbf{X} \in \mathbb{B}^{d\times n}$, the expressions in \eqref{eq:two_quantities} can be also written in the form $C_P(\mathbf{X}) = \sum_{i=1}^d \|\mathbf{X}_{i:}\|_0^2$ and $C_D(\mathbf{X}) = \sum_{j=1}^n \|\mathbf{X}_{:j}\|_0^2$.
By $\mathbb{B}_{\neq 0}^{d\times n}$ we denote the set of all matrices in $\mathbb{B}^{d\times n}$ with nonzero columns and rows.
For positive integers $a,b$ we write $\bar{a}_b ~{:=}~ b\left\lfloor\frac{a}{b}\right\rfloor$ (i.e., $a$ rounded down to the closest multiple of $b$). Further, we write \[R(\alpha,d,n) ~{:=}~ U(\alpha,d,n)/ L(\alpha,n),\] where
\[L(\alpha,n) ~{:=}~ \frac{1}{n} (\bar{\alpha}_n^2 + (\alpha - \bar{\alpha}_n)(2\bar{\alpha}_n + n))\]
and \[U(\alpha,d,n) ~{:=}~ (d + 1)\overline{(\alpha - n)}_{d-1} + n - 1 + (\alpha - n + 1 - \overline{(\alpha - n)}_{d-1})^2.\]
The following is a refinement of Theorem~\ref{thm:Bound_Any_Data} for binary matrices of fixed cardinality $\alpha$.
\begin{theorem} \label{lem:bound_ratio}
For all $\mathbf{X} \in \mathbb{B}^{d\times n}_{\neq 0}$ with $\alpha = \|\mathbf{X}\|_0$ we have the bounds $1/R(\alpha,n,d) \leq C_P(\mathbf{X}) / C_D(\mathbf{X}) \leq R(\alpha, d, n)$. Moreover, these bounds are tight.
\end{theorem}
The above theorem follows from Lemma~\ref{lem:general_binary_maxmin}, which we formulate and prove in the Appendix. This lemma establishes formulas for the minimum and maximum of $C_D$ and $C_P$, subject to the constraint $\|\mathbf{X}\|_0=\alpha$, in terms of the functions $L$ and $U$. Further, as we show in Lemma~\ref{lem:bound_Rabc} in the Appendix, if $d \geq n$ and $\alpha \geq n^2 + 3n$, then $R(\alpha,d,n) \leq 1$. Likewise, if $n \geq d$ and $\alpha \geq d^2 + 3d$, then $R(\alpha,n,d) \leq 1$. Combined with Theorem~\ref{lem:bound_ratio}, this has an interesting consequence, spelled out in the next theorem and its corollary.
\begin{theorem} \label{thm:98y9s8s} Let $\mathbf{X} \in \mathbb{B}^{d\times n}_{\neq 0}$.
If $d \geq n$ and $\| \mathbf{X} \|_0 \geq n^2 + 3n$, then $C_P(\mathbf{X}) \leq C_D(\mathbf{X})$. By symmetry, if $n \geq d$ and $\| \mathbf{X}\|_0 \geq d^2 + 3d$, then $C_D(\mathbf{X}) \leq C_P(\mathbf{X})$.
\end{theorem}
This result says that for binary data, and $d\geq n$, the primal method is better than the dual method even for non-dense data, as long as the the data is ``dense enough''. Observe that as long as $d \geq n^2 + 3n$, all matrices $\mathbf{X} \in \mathbb{B}^{d\times n}_{\neq 0}$ satisfy $\|\mathbf{X}\|_0 \geq d \geq n^2 + 3n \geq n$. This leads to the following corollary.
\begin{corollary}
If $d \geq n^2 + 3n$, then for all $\mathbf{X} \in \mathbb{B}^{d\times n}_{\neq 0}$ we have $C_P(\mathbf{X}) \leq C_D(\mathbf{X})$. By symmetry, if $n \geq d^2 + 3d$, then for all $\mathbf{X} \in \mathbb{B}^{d\times n}_{\neq 0}$ we have $C_D(\mathbf{X}) \leq C_P(\mathbf{X})$.\end{corollary}
In words, the corollary states that for binary data where the number of features ($d$) is large enough in comparison with the number of examples ($n$), the primal method will be always better. On the other hand, if $n$ is large enough, the dual method will be always better. This behavior can be observed in Figure~\ref{fig:alpha}. For large enough $d$, all the values $R(\alpha,d,n)$ are below 1.
\begin{figure}[H]
\centering
\begin{subfigure}{.30\linewidth}
\label{fig:plotR1000}
\centering
\includegraphics[width = \linewidth]{code/plots/plotR1000.pdf}
\end{subfigure}
\caption{The value $R(\alpha,d,n)$ plotted for $n=10^3$, $n\leq d\leq n^2$ and $\max\{d,n\} \leq \alpha \leq nd$.}
\label{fig:alpha}
\end{figure}
\section{Experiments}
We conducted experiments on both real and synthetic data. The problem we were interested in is a standard logistic regression with an L2-regularizer, i.e., \[P(w) = \frac{1}{n}\sum_{j=1}^n \log(1 + \exp(-y_j \langle \mathbf{X}_{:j}, w \rangle)) + \frac{\lambda}{2}\|w\|_2^2.\] In all our experiments we used $\lambda = 1/n$ and we normalized all the entries of $\mathbf{X}$ by the average column norm.
Note that for logistic loss there is no closed form solution for $\Delta_j^k$ in Algorithm~\ref{alg:quartz}. Therefore we use a variant of Algorithm~\ref{alg:quartz} where $\Delta_j^k = \eta (\phi_j'(\langle \mathbf{X}_{:j}, w \rangle) + \alpha_j^k)$ with the step size $\eta$ defined as $\eta = \min_{j \in [n]}(q_j \lambda n)/(\beta v_j + \lambda n)$. This variant has the same convergence rate guarantees as Algorithm~\ref{alg:quartz} and does not require exact minimization. Details can be found in \cite{Quartz}.
We plot the training error against the number of passes through the data. The number of passes is calculated according to the number of visited nonzero entries in the matrix $\mathbf{X}$. One pass means that we look at $\|\mathbf{X}\|_0$ nonzero entries of $\mathbf{X}$, but not necessarily all of them. We look at the problems from the perspective of the primal approach. The same could be done symmetrically for the dual approach.
\subsection{General Data}
We look at the matrices which give worst-case bounds for general matrices (Theorem~\ref{thm:Bound_Any_Data}) and their empirical properties for different choices of $d$ and $n$. The corresponding figures are Figure~\ref{subfig:general1} and \ref{subfig:general2}. For a square dataset, we can observe a large speed-up. For large $n$ we can observe, that the theory holds and the primal method is still faster, but because of numerical issues (we need very small and very large numbers in matrix) and the fact that the optimal value is very close to an "initial guess" of the algorithm, the difference in speed is more difficult to observe.
\begin{figure}
\centering
\begin{subfigure}{0.25 \textwidth}
\includegraphics[width = \textwidth]{code/plots/general_logistic_mean_n__1000_0,1000_0,2_0e9_.pdf}
\caption{$1,000 \times 1,000$}
\label{subfig:general1}
\end{subfigure}
\begin{subfigure}{0.25 \textwidth}
\includegraphics[width = \textwidth]{code/plots/general_logistic_mean_n__100_0,100000_0,2_0e9_.pdf}
\caption{$100 \times 100,000$}
\label{subfig:general2}
\end{subfigure}
\begin{subfigure}{0.25 \textwidth}
\includegraphics[width = \textwidth]{code/plots/news_logistic_mean_n.pdf}
\caption{news dataset}
\label{subfig:news}
\end{subfigure}
\begin{subfigure}{0.25 \textwidth}
\includegraphics[width = \textwidth]{code/plots/leu_logistic_mean_n.pdf}
\caption{leukemia dataset}
\label{subfig:leukemia}
\end{subfigure}
\caption{Testing the worst case for general matrices and real datasets}
\label{fig:general_real}
\end{figure}
\subsection{Synthetic Binary Data}
We looked at matrices with all entries in $\{a, -a, 0\}$ for some $a \neq 0$. We fixed the number of features to be $d = 100$ and we varied the number of examples $n$ and the sparsity level $\alpha=\|\mathbf{X}\|_0$. For each triplet $[d,n,\alpha]$ we produced the worst-case matrix for dual RCD according to the developed theory. The results are in Figure~\ref*{tab:table_experiments}.
\begin{figure}[]
\begin{tabular}{m{0.02\textwidth} >{\centering\arraybackslash} m{0.28\textwidth}
>{\centering\arraybackslash} m{0.28\textwidth} >{\centering\arraybackslash} m{0.28\textwidth} }
& nnz $ \sim 1\% $ & nnz $ \sim 10\% $ & nnz $ 100\% $ \\
\rotatebox[origin=c]{90}{$ n = 100 $} & \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,100_0,199_0_.pdf}& \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,100_0,1179_0_.pdf}& \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,100_0,10000_0_.pdf}\\
\rotatebox[origin=c]{90}{$ n = 1,000 $} & \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,1000_0,1099_0_.pdf}& \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,1000_0,10989_0_.pdf}& \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,1000_0,100000_0_.pdf}\\
\rotatebox[origin=c]{90}{$ n = 10,000 $} & \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,10000_0,10099_0_.pdf}& \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,10000_0,109089_0_.pdf}& \includegraphics[width=0.25\textwidth]{code/plots/alpha_logistic_mean_n__100_0,10000_0,1_0e6_.pdf}
\end{tabular}
\caption{Worst-case experiments with various dimensions and sparsity levels for $d = 100$}
\label{tab:table_experiments}
\end{figure}
\subsection{Real Data}
We used two real datasets to showcase our theory: news and leukemia\footnote{both datasets are available from https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/}. The news dataset in Figure~\ref{subfig:news} is a nice example of our theory in practice. As shown in Table~\ref{tab:alltable} we have $d \gg n$, but the dual method is empirically faster than the primal one. The reason is simple: the news dataset uses a bag of words representation of news articles. If we look at the distribution of features (words), there are many words which appear just very rarely and there are words commonly used in many articles. The features have therefore a very skewed distribution of their nonzero entries. On the other hand, the examples have quite uniform distribution, as the number of distinct words in an article acts nicely. This distribution of nonzero entries highly favors the dual approach, as shown in the theory. The leukemia dataset in Figure~\ref{subfig:leukemia} is a fully dense dataset and $d \gg n$. Therefore, as our theoretical analysis shows, the primal approach should be better. The ratio between the runtimes is not very large, as the constant $\|\mathbf{X}\|_0$ is of similar order as the additional term in the computation of the true runtime. The empirical speedup in Figures~\ref{subfig:news} and Figures~\ref{subfig:leukemia} matches the theoretical predictions from Table~\ref{tab:alltable}.
\begin{table}[]
{
\begin{center}
\begin{tabular}{l r r r r l l cc}
dataset & $d$ & $n$ & density & $\|\mathbf{X}\|_0$ & $C_P$ & $C_D$ & $T_P/T_D$ \\
\hline news & 1,355,191 & 19,996 & 0.03\% & 9,097,916 & $3 \times 10^7$ & $9 \times 10^6$ & 2.0 \\
leukemia & 7,129 & 38 & 100.00\% & 270,902 & $1\times 10^7$ & $2 \times 10^9$ & 0.5
\\\end{tabular}
\end{center}
}
\caption{Details on the datasets used in the experiments}
\label{tab:alltable}
\end{table}
\section{Conclusions and Extensions} We have shown that the question whether RCD should be applied to the primal or the dual problem depends on the structure of the training dataset. For dense data, this simply boils down to whether we have more data or parameters, which is intuitively appealing. We have shown, both theoretically, and through experiments with synthetic and real datasets, that contrary to what seems to be a popular belief, primal RCD can outperform dual RCD even if $n\gg d$.
In order to focus on the main message, we have chosen to present our results for simple (as opposed to ``accelerated'') variants of RCD. However, our results can be naturally extended to accelerated variants of RCD, such as APPROX \cite{APPROX}, ASDCA~\cite{ASDCA}, APCG~\cite{APCG}, ALPHA~\cite{ALPHA} and SPDC~\cite{SPDC}.
Likewise, for simplicity, we focused on serial sampling (i.e., sampling a single coordinate). However, it is possible to use our approach to gain insights into the performance of primal vs dual RCD for arbitrary sampling strategies \cite{NSync, Quartz, ALPHA, ESO}.
{
|
\section{Introduction}
In \cite{Pull}, Marcus ritt proved that two variants of pushpush \cite{Push}, called pull and pullpull, were np-hard. He wondered about the complexity of a variant of pull in which the boxes have either horizontal (left and right) or vertical (up and down) handles.
We will solve that problem in this document, by showing that the problem is NP-hard. We will also show that other variants of pull are NP-hard as well.
In section two we are going to show that pull is NP-hard.
In section three we are going to show that the variant of pull in which the boxes have handles in opposite directions (i.e left and right) is NP-hard.
Finally, in section four, we will show that the variant of pull where the player can only pull boxes in a single direction is NP-hard.
We will show that the previous variants are NP-hard by reducing satisfiability to those variants of pull.
The proofs in this paper are for variants of pull with fixed blocks (Pull-F), but they can be easily adapted so that they work for variants of pull without fixed blocks.
This can be done by filling with blocks all the positions of the board that are inaccesible in the corresponding Pull-F variants, by making the separation between the "gaps" and the boxes inside the crossovers bigger and by making the walls thicker for those gadgets for which it is necessary.
We will represent movable blocks(boxes) by orange/brown squares and fixed blocks by red squares.
\section{Pull is NP-hard}
In this section we are going to prove that Pull is NP-hard, as they did in \cite{Pull}, but we are going to prove the result in a different way. We will adapt the proof in the next sections to show that other pull variants are NP-hard
First we are going to show the gadgets that we need to prove that pull is NP-hard.
\subsection{No return gadget}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{no_return_2.png}
\caption{No return gadget (the green squares are places in which the player has to place the boxes in order to exit from the gadget).}
\label{fig:no_return}
\end{figure}
This is a no return gadget, similar to the ones in \cite{Pull} and \cite{Push}
We can use these gadget for those path variants in which all the boxes can be moved in all directions or those variants in which boxes can only be moved either in horizontal or vertical directions. So we are going to reuse this gadget for the next section.
This gadget allows passage from A to B, forbids passage from B to A, and can only be traversed once
\begin{lema}
\label{one}
The no return gadget can be traversed from A to B , but after it is traversed, te player cannot go back from B to A.
\end{lema}
\begin{proof}
To go to B, the player has to move the boxes 1 and 2 out of the way. In figure 1, the only way to do it is by pulling the boxes to the left (it can be any other direction if the gadget is rotated and/or reflected).
To be able to pull box 2 out of the narrow corridor, the player has to pull box 1 to the position marked by the leftmost green square, blocking the way to A, the player can then pull the other box to the other green square and exit through B
\end{proof}
\subsection{Unidirectional crossover}
This crossover works like the crossover in \cite{Push} , but it doesn´t matter the order in which the player uses the crossover, he can traverse the crossover horizontally and then vertically or viceversa.
If the player enters the gadget from the entry A1, he can only exit through A2. And if the player enters from B1 he can only exit through B2 (see figure 2).
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{crossover_pull_path.png}
\caption{The crossover in its original state}
\label{crossover_pull_path}
\end{figure}
If the player enters the gadget from A1, he can reach A2 by pulling to the left all boxes that are blocking the way to A2.
If the player enters the gadget from B1, he can reach B2 by pulling downwards all boxes that are blocking the way to B2.
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{crossover_pull_path_open.png}
\caption{A crossover that has been crossed from both directions}
\label{crossover_A_2}
\end{figure}
\begin{lema}
\label{two}
In its original state, the crossover gadget can be traversed from A1 to A2 or it can be traversed from B1 to B2. Those are the only possible traversals of the gadget
\end{lema}
\begin{proof}
Let´s suppose that the player enters the gadget from A1, the other direction is simmetrical.
When the player enters, the only thing he can do is pull the two boxes that he finds at the entrance. From there, he cannot go to B2 or B1 because the player has not enough space to pull down the boxes that are blocking the path to B2, and he has not enough space to pull up the boxes that are blocking the way to B1
\end{proof}
Variants of this gadget will form an important component of the different \\crossovers that we will use for the other variants of pull.
\subsection{Bidirectional crossover}
\begin{comment}
An schematic view of the bidirectional crossover
\end{comment}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{crossover_scheme.png}
\caption{An schematic view of the bidirectional crossover}
\label{complete_crossover_1}
\end{figure}
This is a crossover that is formed by four limited unidirectional crossovers, rotated and placed as is necessary to make the gadget work. At the entries and exits of the unidirectional crossovers that make the gadget there are no return gadgets.
This gadget is similar to the bidirectional gadget of \cite{Push} and allows the player to traverse it in the horizontal direction and then the vertical direction or viceversa. This gadget only works for the path versions of pull that let the player move the boxes in all directions.
\subsection{Variable gadget}
\begin{comment}
The variable gadget for standard pull.
\end{comment}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{variable.png}
\caption{An schematic picture of a variable gadget. The arrows represent no return gadgets inside the gadget. In this example, the variable appears four times in the SAT instance. Two times as a positive literal and other two times as a negative literal. If the player takes the left path, that means that the player has decided to set that variable positive. If the player takes the right path, the player has decided to set the variable negative}
\label{variable_1}
\end{figure}
For this variant of the game and the following ones, we replace every literal wire inside the variable gadget with two parallel wires. One is used to go to the clause and the other is used to return from the clause. We have to ensure that the player can only use these wires if he comes from the right place.
If the player comes from the path of a positive literal, we don´t have to do any change, since the one way gadget that is just after the literal wires of the negative literal path prevent him from accessing those literal wires.
Now if the player comes from the negative literal path, we have to ensure that the player cannot access any of the literal wires of the positive literal path.
To do that, we simply place an unidirectional crossover for each crossing of the wires inside the variable gadget
\subsection{Clause gadget}
\begin{comment}
The clause gadget for standard pull.
\end{comment}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{clause_2_B.png}
\caption{A picture of a clause gadget with two literals, the wires A1 and B1 correspond to entrances to the clause. The wires A2 and B2 are exits of the clause gadget. This gadget works for those variants of pull where the player can pull the boxes in all directions or only in opposite directions (either left and right or up and down)}
\label{clause_1}
\end{figure}
The gadget has $n + 1$ entrances and $n$ exits where $n$ is the number of literals in the clause.
Every entrance that comes from a literal wire is adjacent to an exit. The entrances and the exits are connected by a corridor
The gadget has a no return gadget before all entrances and exits, including the C1 entrance and the C2 exit
\begin{lema}
\label{six}
If the player wants to go from C1 to C2 in figure 6, the player had to access the gadget previously from an entrance that belongs to a literal wire
\end{lema}
\begin{proof}
When the player enters the gadget, there is a box (with the label O in figure 6) that it´s blocking the way to B2.
In order to get this box out of the way, the player has to access the gadget from one of the other entrances (A1 or B1) and pull the box one square to the left, so that the box doesn´t block the corridor anymore.
\end{proof}
\begin{lema}
\label{seven}
When the player enters the gadget from an entrance that belongs to a literal wire he has to exit through the adjacent exit
\end{lema}
\begin{proof}
When the player comes from an entrance wire (A1, for instance) he can go to the wire below using the corridor, then he can move the box labelled "1" to the center of the "gap" at his left, because there is a no return gadet at the end of the wire, he has to return to the wire above him and cross a no return gadget in order to get to the clause. Then he has to cross a no return gadget and move the box labelled "3" to the "gap" at the left of the box, then he can enter the gadget and move the box labelled "O" to the left. The no return gadgets that are at the entrance wire (or the boxes at the end of the wires, if the entrances hasn´t been traversed before) doesn´t allow the player to return from these wires.
If he tries to exit from another return wire different from the one that he has unlocked (for instance B2) he will get trapped between a no return gadget and a box (with the label "2" in figure 6) in case the wire had not been traversed before, or he won´t be able to go through the wire, because the no return gadget has already been traversed before. He also won´t be able to exit from the wire he entered because a no return gadget is blocking the way.
\end{proof}
\subsection{General view of the \\board}
We have seen the gadgets that we are going to use. Now we are going to shown how to arrange them to create a board.
In the board, the variable gadgets are at the left of the board and the clause gadgets are at the right of the board. Each variable gadget is put below the next one, the same for the clause gadgets.
\begin{figure}[H]
\centering
\includegraphics[scale=0.8]{clauses_variables.png}
\caption{An example board. The variables at the left are $x_1$ $x_2$ and $x_3$. The clauses at the right are $(\neg x_1 \lor \neg x_3) \wedge (\neg x_1 \lor x_2 \lor x_3) \wedge (x_1 \lor x_2) \wedge (x_1 \lor \neg x_2)$}
\label{fig:board}
\end{figure}
\begin{teorema}
\label{T2}
Pull is NP-hard
\end{teorema}
\section{Pull where boxes have handles in opposite directions is NP-hard}
We are going now to prove that all variants where boxes have handles in opposite directions (i.e. left and right or up and down) are NP-hard
In this section, and in the following ones, we will only describe the changes that we must do to the gadgets we presented in the previous section so that they work correctly for this variant of pull.
\subsection{Half unidirectional\\ crossover}
This is a crossover that works like the unidirectional crossover of previous section, but it only works in a single direction. That is, it is a crossover that can, for instance, prevent a player from accesing the horizontal direction if the player enters the gadget from the vertical direction. But it can´t prevent the player from traversing it vertically if he has entered the gadget through the horizontal wire
When the gadget prevents the player from traversing it vertically from inside the horizontal wires, we will say that the gadget is a "vertical crossover"
When the gadget prevents the player from traversing it horizontally from inside the vertical wire, we will say that the gadget is an "horizontal crossover"
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{half_crossover.png}
\caption{Example picture of an horizontal crossover. If we rotate the gadget 90 degrees, it becomes a vertical crossover}
\label{fig:horizontal_half_crossover}
\end{figure}
We will use this crossover for all those variants in which we cannot create unidirectional crossovers like the ones from the previous section.This crossover will form a component of bigger crossovers with more functionality
\subsection{Variable gadget}
This gadget is the same variable gagdet the one used in the previous section, but we cannot use the same crossovers we used in the previous section. We have to replace the crossovers of the previous section with half crossovers. See figure 9
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{variable_gadget_2.png}
\caption{This picture represents the intersection point between an horizontal and a vertical wire inside a variable gadget. The two horizontal wires are a closeup of the two literal wires that the player has to use to go to a clause (up wire) and return from that clause (down wire)}
\label{fig:var_closeup}
\end{figure}
\subsection{Two way gadget}
This is a gadget that allows the player to go to a position and return back. Once the player returns, the player cant go back to the position he visited previously.
This gadget is formed by two no return gadgets, the arrows in figure 10
\begin{figure}[H]
\centering
\includegraphics[scale=0.8]{two_way.png}
\caption{Two way gadget, the direction in which the arrow is pointing represents the direction we must follow to pass through the wire}
\label{fig:two_way}
\end{figure}
We will only use this gadget inside crossovers
\subsection{Limited Unidirectional Crossover}
This is a unidirectional crossover, similar to the one shown in previous section, but it only works for variants of pull in which the player can only pull boxes in opposite directions. The gadget allows a player to cross from A1 to A2 without allowing the player to exit through B1 or B2 and it allows the player to enter from B1 to B2 without allowing the player to exit through A1 or A2
\begin{figure*}
\begin{center}
\includegraphics[scale=0.8]{LUC_crossover_2.png}
\caption{Complete drawing of the unidirectional crossover. In this example croosover, the player can only pull the boxes left and right.}
\label{fig:unidirectional_crossover}
\end{center}
\end{figure*}
If the player enters from A1, reaching A2 is straightforward. The player has only to follow the path through A2, going through all the half crossovers and the one way gadgets.
To reach B2 from B1, though, is not so simple. The player can´t simply go upwards until he reaches B2, because he will get trapped between a one way gadget and the block 2. The player must first move through the first corridor to the right of the vertical wire, use it to move the block 2 out of the way upwards and then return to the vertical wire. He can now reach B2 without trouble
\begin{lema}
\label{eleven}
The crossover allows the player to traverse from A1 to A2 without allowing him to reach B1 or B2.
\end{lema}
\begin{proof}
We observe by looking at the gadget that there are only two intersections in which the horizontal wire crosses the vertical one. We will proof that attempting to enter into the vertical wires by any of the two intersections is impossible or will only get the player trapped inside the gadget.
In the leftmost intersection, we observe that the player can´t enter inside B1 because he is prevented to do so by a one way gadget that cannot be traversed downwards. If the player tries to move upwards, either an already traversed one way gadget or the block 2 will prevent the player from reaching upwards.
In the rightmost intersection, the player is blocked downwards by the block 1 or by two one way gadgets that have already been traversed. This is inevitable, since to pull the block 1 out of the way the player must cross one of the two one way gadgets and then it must return through the other one, closing it. The player cannot go upwards because there is the block 2 preventing him from reaching B2.
\end{proof}
\begin{lema}
\label{eleven}
The crossover allows the player to traverse from B1 to B2 without leaking to A1 or A2.
\end{lema}
\begin{proof}
We observe by looking at the gadget that there are only two intersections in which the vertical wires cross the horizontal one. All of them are blocked either by a unidirectional crossover or by a one way gadget that it´s blocked. This is inevitable since we showed in the previous lemma that a player entering through A1 can only exit through A2, going through all the one way gadgets in his way.
\end{proof}
\subsection{Bidirectional crossover}
\begin{comment}
An schematic view of the bidirectional crossover
\end{comment}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{crossover_scheme.png}
\caption{An schematic view of the bidirectional crossover}
\label{complete_crossover_2}
\end{figure}
This is a crossover that is formed by four limited unidirectional crossovers, rotated and placed as is necessary to make the gadget work. At the entries and exits of the unidirectional crossovers that make the gadget, there are no return gadgets.
This gadget is similar to the bidirectional gadget of \cite{Push} and allows the player to traverse it in the horizontal direction and then the vertical direction or viceversa.
With these components, we can prove the hardness of pull for those versions in which the boxes can only be pulled from opposite directions
\begin{teorema}
\label{thirteen}
The path version of pull in which all boxes can only be pulled from opposite directions is NP-hard
\end{teorema}
\section{Pull with one handle is NP-hard}
In this section, we are going to prove the main claim of our paper. That is, we are going to show that the version of pull where all boxes can only be pulled from one direction (in our examples, the left direction) is NP-hard.
As we did in previous sections, we are only going to describe the \\changes that must be done to the gadgets so that they work for this variant
\subsection{No return gadget}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{no_return_2_B.png}
\caption{No return gadget for variants in which the player can pull a box in a direction but not the opposite direction.}
\label{fig:no_return1}
\end{figure}
This gadget is a no return gadget like the one in the previous sections, but adapted for this variant of pull. It allows passage from A to B a single time and once we traverse the gadget we cannot return from B to A.
\subsection{Variable gadget}
We are going to describe the changes we need to make to the variable gadget so that it works for this variant.
If the player comes from the path of a positive literal, we don´t have to do any changes to the gadget, since the one way gadget that is just after the literal wires of the negative literal prevent him from accessing those literal wires.
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{variable_gadget_3.png}
\caption{This picture represents the intersection point between an horizontal and a vertical wire inside a variable gadget. The two horizontal wires are a closeup of the two literal wires that the player has to use to go to a clause (up wire) and return from that clause (down wire). In this example figure, the player can only pull the boxes left and down}
\label{fig:var_closeup_2}
\end{figure}
Now, if the player comes from the negative literal path, we have to ensure that the player cannot access any of the literal wires that belong to the other path.
In the example shown in figure 14, this is done by placing an horizontal crossover in the wire that goes to the clause.
For the return wire, a single box prevents the player from accesing the wire. This box can be unlocked from a corridor that comunicates the two wires. This corridor can only be traversed from the interior of the parallel wire above it.
\subsection{Clause gadget}
This clause gadget is similar the one that appears in the previous, but the no return gadgets have been changed so that they work for this variant of pull.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{clause_3.png}
\caption{Clause gadget, it is similar to the one shown in section 2 but changing the no return gadgets}
\label{clause_4}
\end{figure}
\subsection{Crossover gadget}
\begin{figure*}
\begin{center}
\includegraphics[scale=0.8]{crossover_scheme_2.png}
\caption{schematic representation of the crossover.}
\label{fig:crossover_scheme_2}
\end{center}
\end{figure*}
This is a crossover that is valid for all variants of pull where the player can only pull the boxes from a single direction. In our example, the left direction. This crossover is formed by two "halfs". Each of these halfs contains a vertical wire and the intersection of the two horizontal wires with the vertical wire.
The first half contains a vertical crossover to go towards the clause and the second half contains a vertical wire to return. Of course, the two halfs can be "inverted" (that is, the "left" half can be the "right" half and viceversa)
Now, we will show that the first half of the gadget works as it should by a series of lemmas. The other half works the same, so the proof is symmetrical. The only difference is that the vertical wire is "flipped" vertically in the second half of the gadget.
\begin{lema}
\label{fifteen}
The crossover allows the player to go from A1 to the exit of the first half of the crossover without letting him exit from B1 or B2
\end{lema}
\begin{proof}
After the player crosses the first no return gadget, he has to pull a block in the corridor just below him. This is so that he can return through the wire below him without getting trapped. He can´t use that corridor to reach the other horizontal wire since the block prevents him from reaching it. After crossing another no return gadget and opening the first crossover, he finds the first intersection with the vertical wire. If the player has not traversed the vertical wire before and he goes upward, after using a no return gadget, he will find himself trapped because there is a block preventing him from going upwards. If, instead, he goes downwards, another no return gadget will block his way. If the player has traversed the vertical wire before, the player will have two no return gadgets, one below him and other above him that will prevent him from going upwards or downwards.
If, instead, he moves forward, after crossing the second crossover, he will find another intersection with the vertical wire. From here the player can move upwards, but there is an obstacle that blocks his way up. The player can move the obstacle out of the way but he can´t go back to the previous intersection and move upwards from there, since he had to cross a no return gadget to get to the corridor, and that no return gadget prevents him from going to the previous intersection. If the player decides to go down there will be a block preventing him from entering the vertical wire from there. However, if the player traversed the vertical wire before, two no return gadgets will block his way. This is because the player has to use the two way gadget there in order to move the block inside the vertical corridor.
The player is then forced to exit through A2
\end{proof}
\begin{lema}
\label{fifteen}
The crossover allows the player to go from the entrance of the first half of the crossover to A4 without letting the player go to B1 or B2
\end{lema}
\begin{proof}
After crossing the first no return gadget to his left, the player will find himself in the middle of an intersection. If the player has not traversed the vertical wire before, he can go through a no return gadget in his way forwards, then he will find himself trapped between the previous no return gadget and a block above him. If the player has already traversed the vertical wire, a no return gadget will prevent him from going upwards. If the player goes downwards instead, a no return gadget will stand in his way.
If the player moves left instead, after crossing another no return gadget, he will reach another intersection with the vertical wire. He can´t go into the vertical wire by moving upwards because a block stands in his way. He can´t move this block out of the way and then try to go upwards from the previous intersection because a no return gadget that he has already traversed doesn´t allow him to move right towards the intersection. If the player has crossed the vertical wire before, a used two way gadget will prevent him from reaching the vertical wire. This is because the player has to use that gadget in order to clear the path inside the vertical wire.
The player has to keep moving left and, after traversing another no return gadget, he finds a corridor connecting the two horizontal wires, but because there is a block inside the corridor, and this block can only be pulled to the left, he can´t use the corridor to reach the other horizontal wire. He is then forced to exit through A4.
\end{proof}
\begin{figure*}
\includegraphics[scale=0.4]{crossover_pull_one_handle.png}
\caption{Drawing of the first half of the crossover. In this example, the player can only pull the boxes to the left }
\label{fig:horizontal_crossover_one_handle}
\end{figure*}
\begin{figure*}
\includegraphics[scale=0.4]{crossover_pull_one_handle_2.png}
\caption{Drawing of the second half of the crossover.}
\label{fig:horizontal_crossover_one_handle_2}
\end{figure*}
\newpage
\begin{lema}
\label{fifteen}
The crossover allows the player to go from B1 to B2 without letting the player go to A1 or A4
\end{lema}
\begin{proof}
After crossing the first no return gadget above him, the player will find himself in the middle of an intersection. If the player has not traversed the horizontal wire before, he can go through a no return gadget that is to the left of him, then he can cross another no return gadget to his left, but then, he will find a block that prevents him from moving further left. That block can only be moved out of the way if the player traversed the other horizontal wire before. If the player has already traversed the horizontal wire, a used no return gadget will prevent him from going to the left. If the player goes to the right instead, a no return gadget will stand in his way.
If the player moves up there are two points where his path crosses with the horizontal wire. But he is prevented from entering the horizontal wire from those intersections, either because there is a crossover that hasn´t been opened or because the two no return gadgets have already been traversed.
He is then forced to go out through B2.
\end{proof}
\begin{teorema}
\label{sixteen}
The path version of pull in which all boxes can only be pulled from a single direction is NP-hard
\end{teorema}
\section{Open questions and future directions}
Can the results proven here for the path versions of pull be proven for the storage versions of pull?. That is, will it be possible to show that the storage versions of pull where the player can only pull boxes in opposite directions are NP-hard?. Can it be proven that the storage variants of pull where all boxes can only be pulled in only one direction are NP-hard?
We think that the versions where the player can only pull boxes in opposite directions are NP-hard as well. If there exists a variant of planar SAT where the negative clauses are outside the cycle of variables and the positive clauses are inside the cycle of variables, and this variant of SAT is NP-hard, then we can reduce that variant of SAT to the version of storage pull where the player can only pull boxes in opposite directions.
Which variants of Push are NP-hard when the player can only push boxes in certain directions?.
Is the variant of pull where the player cannot revisit a square (called pull-X) still hard if the player can only pull boxes in certain directions?. If the problem is easy this will solve the open question of \cite{Push2} about finding an interesting, but tractable, block-moving puzzle. What about push-X?
|
\section{Introduction}
The analogy between primes and knots is known (\cite{Morishita2012}).
Among other things, there is close analogy between
Iwasawa theory and Alexander--Fox theory, and
topological analogues of Iwasawa's class number formula are studied (\cite{HMM2006}, \cite{KM2008}).
Let $p$ denote a fixed prime number, and let
$\bb{Z}_{p}$ denote the ring of $p$-adic integers through out this paper.
We call a field $K$ simply a \emph{$\bb{Z}_{p}$-field} if it can be obtained as a $\bb{Z}_{p}$-extension of a finite number field $k$.
If the Iwasawa $\mu$-invariants vanish,
extensions of $\bb{Z}_{p}$-fields \blue{resemble}
those of function fields,
and satisfy Kida's formula (\cite{Kida1980}),
which is an arithmetic analogue of the classical Riemann--Hurwitz formula.
In this paper, we formulate their topological analogue,
\blue{in a very parallel manner to }
Iwasawa's second proof in \cite{Iwasawa1981},
\blue{using $p$-adic representation theory of finite groups
and
the Tate cohomologies.}
Here are the contents of this paper.
In Section 2, we review some basic analogies in branched Galois theories for finite degrees.
An analogy between unit groups and 2-cycle groups was pointed out in
our previous paper (\cite{Ueki1}). We calculate the Tate cohomology of 2-cycles
explicitly. We also discuss analogue objects of $S$-ideal groups and others.
In Section 3, we review Iwasawa's class number formula of $\bb{Z}_{p}$-fields.
In Section 4, as an analogue of $\bb{Z}_{p}$-extension, we introduce the notion of
a \emph{branched $\bb{Z}_{p}$-cover} of rational homology 3-spheres ($\bb{Q}$HS$^3$),
namely, an inverse system of branched cyclic $p$-covers of $\bb{Q}$HS$^3$'s, and
generalize the Iwasawa type formula:
\setcounter{section}{4}
\setcounter{thm}{8}
\begin{thm} \blue{Let $\wt{M}=\(M_n\)_n$ be a branched $\bb{Z}_{p}$-cover consisting of $\bb{Q}$HS$^3$'s. Then there are some $\lambda, \mu, \nu\in \bb{Z}$ satisfying $\#H_1(M_n, \bb{Z}_{p})=p^{\lambda n+\mu p^n +\nu}$ for any $n\gg 0$.}
\end{thm}
We also employ the structure theorem of $\Lambda$-modules, and prove a $p$-adic variant of Sakuma exact sequence.
Then we deduce an alternative proof for the Iwasawa type formula, and also a proposition on the direct limit of homology groups.
In Section 5, we further introduce an analogous notion of \blue{a Galois} extension of $\bb{Z}_{p}$-fields,
say \blue{an \emph{equivariant Galois morphism} (or a \emph{branched Galois cover})} of branched $\bb{Z}_{p}$-covers.
It is a compatible system of branched covers on each layer
between two branched $\bb{Z}_{p}$-covers.
In addition, we define chains of branched $\bb{Z}_{p}$-covers by
direct limits with respect to the transfer maps,
and compute the Tate cohomologies for them.
In Section 6, we review
Kida's formula for an extension of $\bb{Z}_{p}$-fields.
In Section 7, we establish an analogue of Kida's formula for branched $\bb{Z}_{p}$-covers:
\setcounter{section}{7}
\setcounter{thm}{0}
\noindent
\begin{thm
\emph{Let $f:\wt{N\,}\!\to \wt{M}$ be \blue{an equivariant Galois morphism of degree \red{$p$-power} of
branched $\bb{Z}_{p}$-covers of $\bb{Q}$HS$^3$,
let $\overline{S}$ be the branch link of $f_0:N\to M$,
\blue{and let $\mca{S}$ denote the inverse limit of the preimages of $\overline{S}$ in $\wt{N\,}\!$.}
If $\overline{S}$ is \blue{infinitely} inert in $\wt{M}$ and
$\mu_{\wt{M}}=0$, then $\mu_{\wt{N\,}\!}=0$.
\bred{If in addition any component of $\overline{S}$ is not inert in $f_0$, then}
\blue{the Iwasawa $\lambda$-invariants and}
the branch indices $e_w$ of components $w$ of \bred{$\mca{S}$} satisfy}
$$\lambda_{\wt{N\,}\!}-1=\deg(f)(\lambda_{\wt{M}}-1)+\sum_{w\subset \blue{\mca{S}}}(e_w-1).$$
\end{thm}
\setcounter{section}{1}
\section{Theories for finite degrees}
To begin with, we briefly review some basic analogies between primes and knots.
In addition,
for finite Galois extensions of number fields and finite branched Galois covers of 3-manifolds,
we study
analogies between unit groups and 2-cycle groups, $S$-ideals and ``$S$-chains'', and compute their Tate cohomologies
in an explicit way.
\subsection{M$^2$KR-dictionary}
The analogy between 3-dimensional topology and number theory was first pointed out by B.~Mazur in the mid 1960's (\cite{Mazur1963}),
and has been studied systematically
by M.~Kapranov (\cite{Kapranov1995}), A.~Reznikov (\cite{Reznikov1997}, \cite{Reznikov2000}) and M.~Morishita (\cite{Morishita2010}, \cite{Morishita2012}).
In their dictionary of analogies, for example, knots and 3-manifolds correspond to primes and number rings respectively.
The study of these analogies is christened ``arithmetic topology'' now.
Here is \blue{a basic dictionary}
we shall use in this paper.
For a number field $k$,
$\mca{O}_{k}$ denotes the ring of integers.
\begin{center}
\begin{tabular}{|c||c|}
\hline
Number theory&3-dimensional topology\\
\hline
number ring $\mathop{\mathrm{Spec}}\nolimits \O_k$ or& (oriented, connected, closed)\\
$\overline{\mathop{\mathrm{Spec}}\nolimits \O_k}=\mathop{\mathrm{Spec}}\nolimits\O_k\cup\({\rm infinite\ primes}\)$& 3-manifold $M$ \\
\hline
prime ideal $\mf{p}: \mathop{\mathrm{Spec}}\nolimits \bb{F}_\mf{p} \hookrightarrow \mathop{\mathrm{Spec}}\nolimits \O_k$ & knot $K:S^1 \hookrightarrow M$ \\
prime ideals $S=\{ \mf{p}_1,...,\mf{p}_r \}$
& link $L=\{K_1,...,K_r\} $ \\
\hline
extension $F/k$ & branched cover $f:N\to M$\\
\hline
\'{e}tale fundamental group $\pi_{1}(\mathop{\mathrm{Spec}}\nolimits \O_k$) & fundamental group $\pi_{1}(M)$\\
$\pi_{1}^{\text{\'et}}(\mathop{\mathrm{Spec}}\nolimits\O_{k}-S$) & link group $\pi_{1}(M-L)$\\
\hline
geometric point $\mathop{\mathrm{Spec}}\nolimits \bb{C} \hookrightarrow \mathop{\mathrm{Spec}}\nolimits \mca{O}_k$ & base point $\(\rm pt\) \hookrightarrow M$\\
\hline
ideal group $I(k)$ & 1-cycle group $Z_{1}(M)$\\
$k^{*} \to I(k);\ a \mapsto (a)$ & $C_{2}(M)\to Z_{1}(M);\ c\mapsto \partial c$ \\
principal ideal group $P(k)$& 1-boundary group $B_{1}(M)$\\
\hline
\end{tabular}
\ \\
\begin{tabular}{|c||c|}
\hline
ideal class group $\mathop{\mathrm{Cl}}\nolimits(k)=I(k)/P(k)$& 1st-homology $H_{1}(M)=Z_{1}(M)/B_{1}(M)$\\
Fact: $\#\mathop{\mathrm{Cl}}\nolimits(k)<\infty$ & \blue{Condition:} $\#H_1(M)<\infty$ (i.e. $M$: $\bb{Q}$HS$^3$),\\
& or consider torsion subgroup $H_1(M)_{\rm tor}$\\
\hline
Artin reciprocity & Hurewicz isomorphism\\
$\pi_1^{\text{\'et}}(\overline{\mathop{\mathrm{Spec}}\nolimits\mca{O}_k})^{\rm ab}\cong \mathop{\mathrm{Gal}}\nolimits(k^{\rm ur}_{\rm ab}/k) \cong \mathop{\mathrm{Cl}}\nolimits(k) $
& $\pi_1(M)^{\rm ab}\cong \mathop{\mathrm{Gal}}\nolimits(M_{\rm ab}/M) \cong H_{1}(M)$\\
$k^{\rm ur}_{\rm ab}/k$ Hilbert class field
& $M_{\rm ab}\to M$ maximal abelian cover\\
\hline
unit group $\O_{k}^{*}$ & 2-cycle group $Z_{2}(M)$,\\
& or 2nd-homology $H_{2}(M)\ (\cong H_1(M)_{\rm free})$\\
\hline
\end{tabular}
\end{center}
We assume that number fields are finite over $\bb{Q}$ and are contained in $\bb{C}$,
3-manifolds are oriented, connected, and closed.
A branched cover means an isomorphism class of branched covers \blue{with base points} branched over links.
We have two attitudes about analogues of ideal class groups $\mathop{\mathrm{Cl}}\nolimits(k)$ and unit groups $\mca{O}_k^*$.
In the conventional attitude, we do not assume $M$ to be $\bb{Q}$HS$^3$, and
we
\blue{regard} the torsion subgroups $H_1(M)_{\rm tor}$ as an analogue of $\mathop{\mathrm{Cl}}\nolimits(k)$.
Then a non trivial term $H_2(M)\cong H_1(M)_{\rm free}$ plays an analogous role to $\mca{O}_k^*$ (\cite{Sikora2003}, \cite{Morin2008}).
In another one, which we pointed out in \cite{Ueki1}, we assume $M$ to be a $\bb{Q}$HS$^3$, fix a CW-structure, and \blue{regard}
$Z_2(M)$ as an analogue of $\mca{O}_k^*$. Then, we have the following parallel exact sequences:
\begin{center}
\begin{tabular}{|c||c|}
\hline
$1\to P(k)\to I(k)\to \mathop{\mathrm{Cl}}\nolimits(k)\to 1$ & $0\to B_{1}(M)\to C_{1}(M)\to H_{1}(M)\to 0$\\
\hline
$1\to \O_{k}^{*}\to k^{*}\to P(k)\to 1$ & $0\to Z_{2}(M)\to C_{2}(M)\to B_{1}(M)\to 0$\\
\hline
\end{tabular}
\end{center}
In addition, we have the following theorem.
\begin{thm}[Ueki \cite{Ueki1}]
Let $f:N\to M$ be a branched Galois cover of 3-manifolds.
Let $G=\mathop{\mathrm{Gal}}\nolimits(h)$, and fix CW-structures compatible with $h$.
Then the Tate cohomology of $\wh{H}^i(G,Z_2(N))$
is independent of the choice of $CW$-structures,
and is a topological invariant of branched covers.
\end{thm}
\blue{We also remark that $\wh{H}^i(G,Z_2(N))$ has more information than $\wh{H}^i(G,H_2(N))$.
These above enable further translation.
In our previous paper \cite{Ueki1}, we gave an analogue of Yokoi's formulation of genus theory (\cite{Yokoi1967}).}
The Tate cohomology of the unit group played an important role in Iwasawa's second proof of Kida's formula\blue{. In this article,}
that of the 2-cycle group
\blue{plays} a similar role in the proof of our main result.
\subsection{Computations of Tate cohomologies}
In this subsection, we \blue{present} some explicit computations of Tate cohomologies \blue{for number fields extensions and branched covers of 3-manifolds.}
For a group $G$ and a $G$-module $A$, $\wh{H}^i(G,A)$ denotes the Tate cohomology \magenta{for each $i\in \bb{Z}$}. We abbreviate $\wh{H}^i(G,A)=\wh{H}^i(A)$ \blue{in the proofs} if there is no ambiguity of $G$.
\blue{The following facts in number theory are well-known.}
\begin{prop} Let $F/k$ be a Galois extension of number fields \blue{with $G=\mathop{\mathrm{Gal}}\nolimits(F/k)$.}
Then,\\
(1) {\rm [Hilbert's Satz 90]} \blue{T}he equality $\wh{H}^{\magenta{1}}(G,F^*)=0$ holds.\\
(2) {\rm (Iwasawa \cite{Iwasawa1956u})} If $F/k$ is unramified and $\mathop{\mathrm{Cl}}\nolimits(F)=1$, then
$\wh{H}^i(G,\mca{O}_F^*)\overset{\cong}{\to} \wh{H}^{i-1}(G,C_F)\overset{\cong}{\to} \wh{H}^{i-3}(G,\bb{Z})$ \magenta{for each $i \in \bb{Z}$},
\blue{where $C_F$ denote the idele class group of $F$.}
Some cases with restricted ramifications, such as cyclotomic extensions, are also computable.
\end{prop}
We have the following on the topology side.
\begin{prop}
Let $f: N\to M$ be a branched Galois cover of 3-manifolds,
Put $G=\mathop{\mathrm{Gal}}\nolimits(f)$, and fix a CW-structures or PL-structures on $M$ and $N$ compatible with $f$. Then \magenta{for each $i\in \bb{Z}$},\\
(1)
The equality $\wh{H}^i(G,C_2(N))=0$ holds. \\
(2) There is \blue{a} long exact sequence
$\cdots \to \wh{H}^{i+1}(G,H_3(N))\to \wh{H}^i(G,Z_2(N))\to \wh{H}^i(G,H_2(N))\to \wh{H}^{i+2}(G,H_3(N))\to \cdots .$
\blue{If in addition} $N$ is a $\bb{Q}$HS$^3$\blue{,}
then $\wh{H}^i(G,Z_2(N))\cong \wh{H}^{i+1}(G,H_3(N))$\blue{, where}
$H_3(N)=\langle[N]\rangle\cong \bb{Z}$.
\end{prop}
\begin{proof}
(1) Since the subset of $G$-fixed points is 1-dimensional, $C_2(N)$ is $\bb{Z}[G]$-free.
(2) \blue{The} exact sequence $0\to B_2(N)\to Z_2(N)\to H_2(N)\to 0$ induces a long exact sequence
$\cdots \to \wh{H}^i(B_2(N))\to \wh{H}^i(Z_2(N))\to \wh{H}^i(H_2(N))\to \wh{H}^{i+1}(B_2(N))\to \cdots .$ We
\blue{prove} $\wh{H}^i(B_2(N))\cong \wh{H}^{i+1}(H_3(N))$.
By the same reason as (1), $C_3(N)$ is a $\bb{Z}[G]$-free module,
and $\wh{H}^i(C_3(N))=0$. Hence
\blue{the} exact sequence $0\to Z_3(N)\to C_3(N)\to B_2(N)\to 0$ yields
an isomorphism $\wh{H}^i(B_2(N))\cong \wh{H}^{i+1}(Z_3(N))$.
In addition, since $N$ is 3-dimensional, we have $B_3(N)=0$.
Hence \blue{the} exact sequence $0\to B_3(N)\to Z_3(N)\to H_3(N)\to 0$
yields isomorphisms $Z_3(N)\cong H_3(N)=\langle[N]\rangle\cong \bb{Z}$.
Thereby, we obtain $\wh{H}^i(B_2(N))\cong \wh{H}^{i+1}(Z_3(N))\cong \wh{H}^{i+1}(H_3(N))$.
Especially, if $N$ is a $\bb{Q}$HS$^3$, the Poincar\`e duality and the universal coefficient theorem ensure $H_2(N)=0$, and hence
$\wh{H}^{i+1}(H_3(N))\overset{\cong}{\to} \wh{H}^i(Z_2(N))$.
\end{proof}
\subsection{$S$-ideals and $S$-chains}
In this subsection, we \blue{recall some properties of $S$-ideals and others of number fields \magenta{and $\bb{Z}_{p}$-fields}.
Then we introduce their analogues for 3-manifolds,}
and obtain further results of Tate cohomologies.
Let $F/k$ be a finite Galois extension of number fields \magenta{or $\bb{Z}_{p}$-fields}.
\magenta{Let} $S$ be a finite set of primes in \violet{$k$, and \magenta{let $S$} also \magenta{denote} the set of primes above $S$ in $F$}.
Then \magenta{the $S$-ideal group $I_{F,S}$, the principal $S$-ideal group $P_{F,S}$, the $S$-ideal class group $\mathop{\mathrm{Cl}}\nolimits_{F,S}$, and the $S$-unit group $\mca{O}^*_{F,S}$}
are defined, and there are exact sequences of $G=\mathop{\mathrm{Gal}}\nolimits(F/k)$-modules (\cite{Brumer1966}, \cite{Iwasawa1981}, \cite{NSW}).
\begin{center}
$0\to P_{F,S}\to I_{F,S}\to \mathop{\mathrm{Cl}}\nolimits_{F,S}\to 0$\\
$0\to \mca{O}^*_{F,S}\to F^* \to P_{F,S}\to 0$
\end{center}
For a prime $\mf{p}$ in $k$, let $I_{F,\mf{p}}$ denote the group of ideals of $F$ over $\mf{p}$, and let $Z_{\mf{p}}$ denote the decomposition group of $\mf{p}$. Then we have isomorphisms
$\wh{H}^i(G,I_{\mf{p}})\cong \wh{H}^i(G,\bb{Z}[G/Z_{\mf{p}}])\cong \wh{H}^i(Z_{\mf{p}},\bb{Z})$.
The direct sum decomposition $I_F=\oplus_{\mf{p}} I_{F,\mf{p}}$ yields
$\wh{H}^i(G,I_F)=\oplus_{\mf{p}} \wh{H}^i(G,I_{F,\mf{p}})\cong \oplus_{\mf{p}} \wh{H}^i(Z_{\mf{p}},\bb{Z})$.
\begin{prop}[Iwasawa \magenta{\cite{Iwasawa1981}}]
Let $F/k$ be an extension of degree $p$ unramified at infinite primes, \blue{put $G=\mathop{\mathrm{Gal}}\nolimits(F/k)$,} and let $S$ \blue{denote}
the set of ramified non-$p$ primes. Then, $S$ is a finite set. \\
(1) The Tate cohomology $\wh{H}^i(G,I_{F,S})=\oplus_{\mf{p} \not\in S}\wh{H}^i(G,I_{\mf{p}})$ \blue{of $S$-ideals} \magenta{($i \in \bb{Z})$} vanishes if and only if no ideals outside $S$ inert. (If $k$ is a number field, then it does not vanish.)\\
\blue{(2) \magenta{Let $I_S$ denote the subgroup of $I(F)$ generated by elements of $S$. Then the}
$S$-unit group satisfies $\mca{O}^*_{F,S}/\mca{O}^*_F\cong \magenta{P(F)}\cap I_S$ for a subgroup $\magenta{P(F)}\cap I_S<I_S\cong \bb{Z}^{\#S}$ of finite index, and
$\wh{H}^1(G, \mca{O}^*_{F,S}/\mca{O}^*_F)=0, \wh{H}^2(G, \mca{O}^*_{F,S}/\mca{O}^*_F) \cong (\bb{Z}/p\bb{Z})^{\#S}$.}
\end{prop}
We consider analogues of these above.
Let $f:N \to M$ be a finite Galois branched cover of 3-manifolds,
put $G=\mathop{\mathrm{Gal}}\nolimits(f)$ and
fix CW-structures or PL-structures compatible with $f$.
Let $\overline{S}\subset M$ be a link \violet{and suppose that its each component is in or} \blue{outside the branch link of $f$}.
Note that analogues of $S$-ideals are \emph{not} the ones obtained from $C_*(N,S):=\mathop{\mathrm{Coker}}\nolimits(C_*(S)\hookrightarrow C_*(N))$.
Instead, we consider the following commutative diagram, whose rows are exact.
$$
\xymatrix{
0 \ar[r] & B_1(S)=0 \ar[r] \ar[d] & Z_1(S) \ar^{\cong}[r] \ar^{\iota_*}[d] & H_1(S) \ar[r] \ar^{\iota_*}[d] & 0\\
0 \ar[r] & B_1(N) \ar[r] & Z_1(N) \ar[r] & H_1(N) \ar[r] & 0
} $$
\magenta{We identify $Z_1(S)$ and $\iota_*(Z_1(S))$.}
We
put $Z_1(N)_S:=Z_1(N)/\magenta{Z_1(S)}$, $H_1(N)_S:=H_1(N)/\iota_*(H_1(S))$,
and $B_1(N)_S:= B_1(N)/(B_1(N)\cap Z_1(S))$.
\red{In addition,} for the boundary map $\partial:C_2(N)\to C_1(N)$,
we put $Z_2(N)_S:=\partial^{-1}(Z_1(S))$. \blue{Then we have $Z_2(N)_S\cong Z_2(N,S)$.}
\blue{Thus} we obtain exact sequences of $G$-modules parallel to the case of number theory.
\begin{center}
$0\to B_1(N)_S \to Z_1(N)_S\to H_1(N)_S\to 0$\\
$0\to Z_2(N)_S \to C_2(N)\to B_1(N)_S\to 0$
\end{center}
\blue{The} analogue of the $S$-ideal group satisfies the following proposition.
\begin{prop} \blue{Let $f:N\to M$ be a finite Galois branched cover with $G=\mathop{\mathrm{Gal}}\nolimits(f)$, let $\overline{S}\subset M$ be a link,
and put $S=f^{-1}(\overline{S})$.} \\
{\rm (i)} If $S$ is not empty, then there is an isomorphism $\wh{H}^i(G,Z_1(N)_S)\cong \wh{H}^{i-1}(\blue{G,}\mathop{\mathrm{Ker}}\nolimits(\iota_*:H_0(S)\to H_0(N)))$ \magenta{for each $i\in \bb{Z}$}.
\blue{If in addition} $f$ is of degree $p$ and \blue{$\overline{S}$ is the branch link with $s$ components,}
then
$\wh{H}^0(G,Z_1(N)_S)=0$ and $\wh{H}^1(G,Z_1(N)_S)\cong (\bb{Z}/p\bb{Z})^{s-1}$.\\
{\rm (ii)} If $S$ is empty so that $f$ is unbranched, then
$\wh{H}^i(G,Z_1(N))\cong \wh{H}^i(\blue{G},H_0(N)))\cong \wh{H}^i(\blue{G},\bb{Z})$ \magenta{for each $i\in \bb{Z}$}.
If \blue{in addition} $\deg(f)=p$,
then $\wh{H}^0(G,Z_1(N))\cong \bb{Z}/p\bb{Z},$ $\wh{H}^1(G,Z_1(N))=0$.
\end{prop}
\begin{proof} If $S$ is not empty, then
\blue{the} exact sequences
$0\to Z_1(S)\to C_1(S)\to B_0(S)\to 0$,
$0\to Z_1(N)\to C_1(N)\to B_0(N)\to 0$
and the snake lemma yield an exact sequence
$0\to Z_1(N)_S\to C_1(N)/C_1(S) \to B_0(N)/B_0(S)\to 0$.
Here, $C_1(N)/C_1(S)$ is a $\bb{Z}[G]$-free module. %
In addition, \blue{the} exact sequences
$0\to B_0(S)\to Z_0(S)\to H_0(S)\to 0$,
$0\to B_0(N)\to Z_0(N)\to H_0(N)\to 0$
and the snake lemma yield an exact sequence
$0\to \mathop{\mathrm{Ker}}\nolimits(\iota_*:H_0(S)\to H_0(N))\to B_0(N)/B_0(S)\to Z_0(N)/Z_0(S) \to 0$.
Here, $Z_0(N)/Z_0(S)$ is a $\bb{Z}[G]$-free module.
As a consequence, isomorphisms
$\wh{H}^i(Z_1(N)_S)\cong \wh{H}^{i-1}(B_0(N)/B_0(S))
\cong \wh{H}^{i-1}(\mathop{\mathrm{Ker}}\nolimits(\iota_*:H_0(S)\to H_0(N))$ are obtained.
Especially, if $\deg(f)=p$ and $\overline{S}$ is properly branched, then
$\mathop{\mathrm{Ker}}\nolimits(\iota_*:H_0(S)\to H_0(N))\cong \bb{Z}^{s-1}$.
If $S$ is empty, we consider \blue{the} exact sequences
$0\to Z_1(N)\to C_1(N)\to B_0(N)\to 0$ and
$0\to B_0(N)\to Z_0(N)\to H_0(N)\to 0$,
in which $C_1(N)$ and $Z_0(N)$ are $\bb{Z}[G]$-free modules, and $H_0(N)\cong \bb{Z}$. Then, isomorphisms
$\wh{H}^i(Z_1(N)_S)=\wh{H}^i(Z_1(N))\cong \wh{H}^{i-1}(B_0(N))\cong \wh{H}^{i-2}(H_0(N))\cong \wh{H}^i(\bb{Z})$ are obtained. \end{proof}
If we fix a $\bb{Z}$-basis of $Z_1(M)$ and \blue{regard}
it as an analogue of the set of primes in the base field, then we obtain a direct sum decomposition of $Z_1(N)$ as $G$-modules.
\blue{The analogue of the $S$-unit group satisfies the following.}
\begin{prop}
\blue{Let $f:N\to M$ be a Galois branched cover of degree $p$ with $G=\mathop{\mathrm{Gal}}\nolimits(f)$,
let $\overline{S}$ be the branch link with $s$-components,
and put $S=f^{-1}(\overline{S})$. Suppose that $S$ consists of $\bb{Q}$-null-homologous components. (This assumption holds if $N$ is a $\bb{Q}$HS$^3$.) Then $S$-2-cycles satisfy $Z_2(N)_S/Z_2(N)\cong B_1(N)\cap Z_1(S)$ for a subgroup $B_1(N)\cap Z_1(S)<Z_1(S)\cong \bb{Z}^s$ of finite index, and
$\wh{H}^1(G,Z_2(N)_S/Z_2(N))=0,$ $\wh{H}^2(G,Z_2(N)_S/Z_2(N))\cong (\bb{Z}/p\bb{Z})^s$.}
\end{prop}
\begin{proof}
\blue{Note that $Z_2(N)_S\cong Z_2(N,S)$.
There is a natural exact sequence $0\to Z_2(N)\to Z_2(N,S)\overset{\partial}{\to} Z_1(S)$.
Since every component of $S$ is $\bb{Q}$-null-homologous, $\partial Z_2(N,S)=B_1(N)\cap Z_1(S)< Z_1(S)$ is a subgroup of finite index. Since $S$ is the preimage of the branch link, $G$ acts on $B_1(S)$ trivially.}
\end{proof}
\blue{These computations above will be extended for $\bb{Z}_{p}$-covers in \magenta{Subsection} 5.4.}
\section{Iwasawa's class number formula of $\bb{Z}_{p}$-extensions and $\Lambda$-modules}
It has been known that
there is close analogy between Iwasawa theory and Alexander--Fox theory (\cite{Mazur1963}, \cite{Morishita2012}).
For an inverse system of branched cyclic $p$-covers of $\bb{Q}$HS$^3$ obtained from a $\bb{Z}$-cover,
an analogue of Iwasawa's class number formula is formulated (\cite{HMM2006}, \cite{KM2008}, \cite{KM2013}).
In this paper,
we generalize this formula for an inverse system
which is not necessarily obtained from a $\bb{Z}$-cover.
\begin{center}
\begin{tabular}{|c||c|}
\hline Iwasawa theory& Alexander--Fox theory\\
\hline
\hline
$\bb{Z}_{p}$-extension of $k$ & $\bb{Z}$-cover over $M - L$, or\\
&branched $\bb{Z}_{p}$-cover over $(M,L)$\\
\hline
Iwasawa module & link module \\
Iwasawa polynomial & Alexander polynomial\\
\hline
\end{tabular}
\end{center}
In this section, we recall algebraic lemmas on $\Lambda$-modules,
a proof of Iwasawa's class number formula,
and an assertion on some direct limit module.
Their analogues will be discussed in the next section.
Let $\Lambda=\bb{Z}_{p}[[T]]$ denote the ring of formal power series.
For $\Lambda$-modules $M$ and $M'$, a \emph{pseudo isomorphism} $M\sim M'$ is a homomorphism with finite kernel and cokernel.
For finitely generated compact $\Lambda$-modules, pseudo isomorphisms give an equivalence relation.
There is a structure theorem on compact $\Lambda$-modules.
\begin{lem}[Whasington \cite{Washington}, Chapter 13
\label{Lambda} Let $E$ be a compact $\Lambda$-module. \\
(1) {\rm (Nakayama's lemma)} $E$ is a finitely generated $\Lambda$-module if and only if $E/(p,T)$ is a finite group.\\
(2) Let $E$ be a finitely generated $\Lambda$-module. Then, there is a pseudo isomorphism
$$E\sim \Lambda^{\oplus r}\oplus (\oplus_i \Lambda/(f_i^{e_i}))\oplus (\oplus_j\Lambda/(p^{m_j}))$$
to a unique normal form.
Here, $r,e_i, m_j \in \bb{N}$, and $f_i\in \bb{Z}_{p}[T]$ is a irreducible Weierstrass polynomial, namely, the coefficient of its highest term is 1 and others are multiples of $p$. \\
(3) Let $\mu=\sum m_j, \lambda=\sum e_i \deg (f_i)$ in the normal form above, and put $\nu_{p^n}:=(t^{p^n}-1)/(t-1)=\sum_{0\leq i<p^n}t^i$.
If $E/\nu_{p^n}E$ is finite for all $n$, then $r=0$, and there is some $\nu,n_0$ such that for any $n>n_0$,
$$E/\nu_{p^n}E=p^{\lambda n+ \mu p^n +\nu}.$$
(4) In this situation, $\mu=0$ if and only if the $p$-rank of $E/\nu_{p^n}E$ is bounded.
\end{lem}
Next, we review number theory.
A \emph{$\bb{Z}_{p}$-field}\footnote{A \emph{cyclotomic $\bb{Z}_{p}$-field} and a cyclotomic $\bb{Z}_{p}$-field which includes $p$-power-th roots of unity in this paper are a \emph{$\bb{Z}_{p}$-field} and a \emph{cyclotomic $\bb{Z}_{p}$-field} in the sense of Iwasawa in \cite{Iwasawa1981} respectively.} is a field obtained as a $\bb{Z}_{p}$-extension of a \magenta{number field}.
\magenta{If $k_\infty/k$ is a $\bb{Z}_{p}$-extension of a number field, then $k_\infty$ is the direct limit (union) of $\bb{Z}/p^n\bb{Z}$-extensions $k_n/k$.}
We assume that $\bb{Z}_{p}$-fields are also contained in $\bb{C}$. A famous formula is stated as follows:
\begin{thm}[Iwasawa's class number formula \cite{Iwasawa1959}] \label{ICF
Let $\blue{k_\infty}/k$ be a $\bb{Z}_{p}$-\blue{extension} of a finite number field.
For each \magenta{$n\in \bb{N}=\bb{N}\cup \{0\}$},
let $k_n/k$ denote the subextension of degree $p^n$, and
let $\mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}$
\blue{denote} the
\blue{$p$-}ideal class group
Then there are some $\lambda,\mu \in \bb{N}$,
$\nu \in \bb{Z}$, and $n_0\blue{\in \bb{N}}$ such that for any $n>n_0$,
$$\#\mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}=p^{\lambda n+\mu p^n +\nu}.$$
\end{thm}
These $\lambda, \mu,$ \magenta{and} $\nu$ are called the \emph{Iwasawa invariants},
\magenta{and denoted as $\lambda_{k_\infty/k}, \mu_{k_\infty/k}$, and $\nu_{k_\infty/k}$.
The value of $\lambda$ and whether $\mu=0$ or not depend only on the $\bb{Z}_{p}$-field $k_\infty$, and are independent of the base field $k$. Hence they are sometimes expressed as $\lambda_{k_\infty}$ and $\mu_{k_\infty}=0$.}
Let $\bb{Q}_\infty$ denote the unique $\bb{Z}_{p}$-extension over $\bb{Q}$.
For a number field $k$, \blue{\emph{the cyclotomic $\bb{Z}_{p}$-extension} of $k$ is defined \magenta{by} $k_\infty^{\rm cyc}:=k\bb{Q}_\infty$, and the Iwasawa invariants of $k$ are defined by those of $k_\infty^{\rm cyc}/k$.}
\blue{
In the following, we review a proof of this formula with use of Lemma 3.1 on $\Lambda$-modules.
Let $k_\infty/k$ and $k_n$ be as above, let $t$ be a topological generator of $\Gamma:=\mathop{\mathrm{Gal}}\nolimits(k_\infty/k)\cong \bb{Z}_{p}$, and fix an identification $\Lambda=\bb{Z}[[T]]\cong \bb{Z}_{p}[[\Gamma]]; 1+T \longleftrightarrow t$. The group $\Gamma$ acts on the inverse system $\(\mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}\)_n$ and hence \emph{the Iwasawa module} $\mca{H}:=\varprojlim \mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}$ continuously, and this action makes $\mca{H}$ a compact $\Lambda$-module. The $p$-class groups $\mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}$ can be expressed as a quotient of $\mca{H}$:}
\begin{prop}[Washington \cite{Washington}, Proposition 13.22 and Chapter 13.3] \label{classnumber}
Let $\blue{k_\infty}/k$ be a $\bb{Z}_{p}$-extension with $n$-th subfield $k_n$, and put $\mca{H}\blue{=}\varprojlim \mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}$.\\
(1) When precisely one prime is ramified in $k_\infty/k$ and it is totally ramified,
for each $n$, (i) there is an isomorphism $$\mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}\cong \mca{H}/(t^{p^n}-1)\mca{H},$$
and (ii) $p|\#\mathop{\mathrm{Cl}}\nolimits(k)$ if and only if $p|\#\mathop{\mathrm{Cl}}\nolimits(k_n)$. \\
(2) More generally, if every ramified prime in $k_\infty/k$ is totally ramified,
then there is a subgroup $\mca{H}'<\mca{H}$ of finite index such that
$\mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]}\cong \mca{H}/\nu_{p^n}\mca{H}'$.
\end{prop}
The number of ramified primes in a $\bb{Z}_{p}$-extension is always finite. General cases reduce to the cases of totally ramified. Indeed,
for any $\bb{Z}_{p}$-extension $k_\infty/k$, there is some $n_0$ such that
$k_\infty/\red{k_{n_0}}$ is totally ramified at every ramified prime.
The following lemma can be obtained from Lemma \ref{Lambda} (1): \begin{lem}[Washington \cite{Washington}, Chapter 13.3]\label{classFT}
The \blue{Iwasawa module} $\mca{H}$ is a finitely generated torsion $\Lambda$-module.
\end{lem}
Now Theorem \ref{ICF} (Iwasawa's class number formula) follows immediately from Lemma \ref{classFT}, Proposition \ref{classnumber}, and Lemma \ref{Lambda} (3) (The structure theorem of compact $\Lambda$-modules). \\
On the other hand, in a $\bb{Z}_{p}$-extension $k_\infty/k$,
ideal class groups of $k_n$ form an inductive system with respect to the maps induced by the natural injections of ideal groups $I_{k_n}\hookrightarrow I_{k_{n+1}}$.
The ideal class group of $k_\infty=\cup k_n$ is defined by $\mathop{\mathrm{Cl}}\nolimits(k_\infty):=I(k_\infty)/P(k_\infty)$,
where $I(k_\infty)=\varinjlim I(k_n), P(k_\infty)=\varinjlim P(k_n)$, and satisifies $\mathop{\mathrm{Cl}}\nolimits(k_\infty)=\varinjlim \mathop{\mathrm{Cl}}\nolimits(k_n)$.
We have the following proposition.
\begin{prop}[Iwasawa \cite{Iwasawa1981}, a remark in Chapter 5]
Let $\blue{k_\infty}/k$
be a cyclotomic $\bb{Z}_{p}$-extension. Then, there is an isomorphism of discrete $\Lambda$-modules
$$\mathop{\mathrm{Cl}}\nolimits(k_\infty)=\varinjlim \mathop{\mathrm{Cl}}\nolimits(k_n)_{[p]} \cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{k_\infty/k}} \oplus A',$$
where $A'$ is a bounded module, namely, there is some $a\in \bb{N}$ such that $p^aA'=0$, and $\mu_{k_\infty/k}=0$ if and only if $A'=0$.
\end{prop}
Such a group is discussed also in Iwasawa \cite{Iwasawa1973Zl}, Chapter 5.
\section{Iwasawa type formula of branched $\bb{Z}_{p}$-covers and $\Lambda$-modules}
In this section, we first introduce the notion of a branched $\bb{Z}_{p}$-cover of 3-manifolds as an analogue of a $\bb{Z}_{p}$-extension, and explain that it essentially generalizes the conventional objects.
Next, we prove the Iwasawa type formula \blue{on the homology growth} for a branched $\bb{Z}_{p}$-covers of $\bb{Q}$HS$^3$\blue{'s}.
Moreover, we state a $p$-adic variant of Sakuma's exact sequence, whose proof will be given afterwards,
and deduce an alternative proof of the Iwasawa type formula in a parallel manner to the case of number theory.
In addition, we deduce a proposition on the direct limit of homology groups which will be used in Section 7.
We also discuss further generalization to the cases of non-$\bb{Q}$HS$^3$ briefly.
\subsection{Branched $\bb{Z}_{p}$-covers
In this subsection, we discuss an analogue of $\bb{Z}_{p}$-extension.
\begin{dfn}
Let $M$ be a 3-manifold and let $L$ be a link in $M$.
A \emph{branched $\bb{Z}_{p}$-cover} over $(M,L)$ is
an inverse system $\wt{M}=\(h_n:M_n\to M\)_n$ of cyclic branched covers of $M$ branched over $L$ with degree $p^n$.
It is a branched $\bb{Z}_{p}$-cover \emph{of $\bb{Q}$HS$^3$} if all $M_n$ are $\bb{Q}$HS$^3$. \end{dfn}
\begin{prop}
Let $L$ be a link in a 3-manifold $M$ and put $X=M-L$.
A branched $\bb{Z}_{p}$-cover over $(M,L)$ corresponds to
a homomorphism $\tau:\pi_1(X)\to \bb{Z}_{p}$ such that $\tau \mathop{\mathrm{mod}}\nolimits p \neq 0$ uniquely up to ${\rm Aut}(\bb{Z}_{p})$.
\end{prop}
\begin{proof}
If such $\tau$ is given, then
the composite
$\tau_n:=(\bb{Z}_{p}\twoheadrightarrow \bb{Z}/p^n\bb{Z})\circ \tau$
with the natural surjection
is a surjective homomorphism for each $n\in \bb{N}$.
A subgroup $\mathop{\mathrm{Ker}}\nolimits(\tau_n)<\pi_1(X)$ corresponds to a cyclic cover $h_n:X_n\to X$ with degree $p^n$, and
hence a cyclic branched cover $h_n:M_n\to M$
by Fox completion.
The family $\(h_n\)_n$ forms an inverse system in a natural way.
Conversely, if such an inverse system $\wt{M}=\(h_n:M_n\to M\)_n$ is given,
a family of surjective homomorphism $\tau_n:\pi_1(X)\twoheadrightarrow \bb{Z}/p^n\bb{Z}$ is obtained.
Since $h_n$ is a subcover of $h_{n+1}$ for each $n$, $\mathop{\mathrm{Ker}}\nolimits \tau_n<\mathop{\mathrm{Ker}}\nolimits\tau_{n+1}$ holds, and there is a surjective homomorphism $q_n:\bb{Z}/p^{n+1}\bb{Z}\twoheadrightarrow \bb{Z}/p^n\bb{Z}$ which satisfies $q_n\circ \tau_{n+1}=\tau_n$.
The inverse limit of an inverse system \blue{$\(\bb{Z}/p^n\bb{Z}, q_n\)_n$} is isomorphic to $\bb{Z}_{p}$, and hence $\tau:\pi_1(X)\to \bb{Z}_{p}$ is obtained up to ${\rm Aut}(\bb{Z}_{p})$.
(In order to obtain $\tau$ explicitly, for each $n$, we replace $\tau_n$ by the composite of $\tau_n$ and an element of ${\rm Aut}(\bb{Z}/p^n\bb{Z})$ such that $q_n$ is the natural surjection $\bb{Z}_{p}\to \bb{Z}/p^n\bb{Z}$.)
Such $\tau$ is unique up to ${\rm Aut}(\bb{Z}_{p})$. \end{proof}
The assumption $\tau \mathop{\mathrm{mod}}\nolimits p \neq 0$ is equivalent to
that $\tau$ sends a generator of $\pi_1(X)$ to a unit of $\bb{Z}_{p}$,
that $\tau$ has a dense image,
and that $\tau$ induce a surjection $\tau: \blue{\wh{\pi}}_1(X) \twoheadrightarrow \bb{Z}_{p}$ from the pro-$p$ completion \blue{defined as} $\blue{\wh{\pi}}_1(X):=\varprojlim_{N} \pi_1(X)/N$ \blue{with} $N$ \blue{running} through the set of normal subgroups with $p$-power indices.
\begin{rem} \blue{Let $\wt{M}$ be a branched $\bb{Z}_{p}$-cover as above.
The Fox completion is defined for more general objects than manifolds called spreads (\cite{Fox1957}), and $M_\infty:=\varprojlim M_n$ is the Fox completion of $X_\infty:=\varprojlim X_n$, which is not necessarily a manifold.
There are some researches on \magenta{such objects} (\cite{Dellomo1986}, \cite{Dellomo1988}, \cite{CS1977}).}
\end{rem}
Since the image of $\tau:\pi_1(X)\to \bb{Z}_{p}$ is abelian, $\tau$ factors through $\pi_1(X)^{\rm ab}\cong H_1(X)$.
In order to see examples, we
\blue{prepare} the following lemma \blue{on $H_1(X)$}.
It was originally given by \cite{KM2008} Lemma 4.2 for $\bb{Q}$HS$^3$, but
it can be prove for a general (oriented, connected, and closed) $M$ in a similar way with use of intersection form.
\begin{lem}
Let $L=\sqcup K_i$ be a $d$-component link
in a 3-manifold $M$, put $X=M-L$, and let $\mu_i\in H_1(X)$ denote the meridian of $K_i$ for each $i$.
Then that $L$ consists of null-homologous components
is equivalent to that \blue{the natural exact sequence}
$$0\to \langle \mu_1,...,\mu_d\rangle\to H_1(X)\to H_1(M)\to 0$$
with $\langle \mu_1,...,\mu_d\rangle\cong \bb{Z}^d$ \blue{splits},
and that $H_1(X)_{\rm free}:=H_1(X)/H_1(X)_{\rm tor}$ has a $\bb{Z}$-basis containing the image of $(\mu_1,\ldots,\mu_d)$.
\end{lem}
\begin{proof}
Fix a tubular neighborhood $V_L=\sqcup V_{K_i}$ of $L$
and put $X^{\circ}=M-{\rm Int}(V_L)$. Then there is a non-degenerate quadratic form $I:H_2(X^\circ,\partial X^\circ)_{\rm free}\times H_1(X^\circ)_{\rm free}\to \bb{Z}$ called \emph{the intersection form}.
Let $\mu_i$ denote the meridians of $K_i$ in \red{all of} $H_1(\partial V_{K_i})$, $H_1(X^\circ)\cong H_1(X)$, and $H_1(X)_{\rm free}$ for each $i$.
Suppose that $[K_i]=0$ in $H_1(M)$ for all $i$. Then for each $i$, there is a surface $S'_i$ in $M$ such that $\partial S'_i=K_i$. We may assume that ${\rm Int}(S'_i)$ intersects both $L$ and $\partial V_L$ transversely. Put $S_i:=S_i'\cap X^{\circ}$.
Since there is a basis of $H_1(\partial X^{\circ})$ containing $(\partial [S_1],\cdots, \partial [S_d])$,
there is a $\bb{Z}$-basis of $H_2(X^{\circ},\partial X^{\circ})_{\rm free}$ containing $([S_1],\cdots, [S_d])$.
Replace elements $T_j$ of this basis other than $S_i$'s so that $I(T_j,\mu_i)=0$.
Then the dual basis of $H_1(X)$ with respect to $I$ contains $(\mu_1,\cdots,\mu_d)$,
and its lift defines an isomorphism $H_1(X)\overset{\cong}{\to} \langle \mu_1,\cdots,\mu_d\rangle_{\bb{Z}} \oplus H_1(M)$.
Conversely, suppose that there is a $\bb{Z}$-basis of $H_1(X)_{\rm free}$ containing $(\mu_1,\ldots,\mu_d)$,
and take the dual basis of $H_2(X^{\circ}, \partial X^{\circ})$ with respect to $I$.
Then for each $i$, there is a surface $S_i$ in $X^{\circ}$ such that
$[S_i]$ is an element of the dual basis satisfying $I([S_i], \mu_i)=1$. We may assume that $S_i$ intersects $\partial X$ transversely.
Since $\partial M=\phi$, $\partial S_i$ is the sum of a longitude of $\partial V_{K_i}$ and some meridians.
Capping off the meridians in $\partial S_i$ by the meridian discs, and extending the longitudes of $\partial V_{K_i}$ to $K_i$, we obtain a Seifert surface $S'_i$ of $K_i$, and hence $[K_i]=0$ in $H_1(M)$.
\end{proof}
Following is an important example obtained from a $\bb{Z}$-cover.
\begin{exa}[TLN-cover, \cite{KM2008}]
Let $L=\cup K_i$ be a link in a 3-manifold $M$, let $\mu_i$ denote the meridian of $K_i$,
and suppose that $L$ consists of null-homologous components.
Then a standard $\bb{Z}$-cover $\wt{X}\to X=M-L$ is defined by $\tau:\pi_1(X)\to \bb{Z}; \forall \mu_i\mapsto 1$ \blue{which induce the zero map on $H_1(M)$
and it is called \emph{the TLN-cover over $(M,L)$}.
We call the inverse system of branched $p$-covers obtained from such a $\bb{Z}$-cover \emph{the TLN-$\bb{Z}_{p}$-cover} $\wt{M}=\(h_n:M_n\to M\)_n$ over $(M,L)$.
\blue{We prove in Subsection 4.4 that}
all the $M_n$ are $\bb{Q}$HS$^3$ if and only if
$M$ is a $\bb{Q}$HS$^3$ and the reduced Alexander polynomial \blue{$\Delta_{L,\tau}(t)$} is not divided by any cyclotomic polynomial of $p$-power-th.
\end{exa}
In order to explain that the notion of a branched $\bb{Z}_{p}$-cover gives an essential generalization, we introduce the notion of an isomorphism of them.
\begin{dfn}
Branched $\bb{Z}_{p}$-covers $\wt{M}=\(h_n:M_n\to M\)_n$ and $\wt{M'}=\{h'_n:M'_n\to M'\}_n$ are \emph{isomorphic} if the following equivalent conditions are satisfied. \\
(1) There is a compatible system $\{f_n:M'_n\overset{\cong}{\to} M_n\}_n$ of
isomorphisms
on each layer.\\% \blue{←ここでeachを使うとき単複?}\\
(2) There is an isomorphic cover $f_{\red 0}:X'\overset{\cong}{\to} X$ of the exteriors of some links in the exterior, and there is an isomorphism $\iota: \bb{Z}_{p}\overset{\cong}{\to} \bb{Z}_{p}$ such that $\iota\circ \tau=\tau' \circ f_{\red 0*}$, where $\tau:\pi_1(X)\to \bb{Z}_{p}$ and $\tau':\pi_1(X')\to \bb{Z}_{p}$ are the the defining homomorphisms and $f_{\red 0*}:\pi_1(X)\overset{\cong}{\to} \pi_1(X')$ is the induced map.
\end{dfn}
\begin{proof} We prove the equivalence of (1) and (2).
Suppose an isomorphism $f_{\red 0}:X'\overset{\cong}{\to} X$ is given.
Let $\tau_n,\tau'_n$ denote the composites of $\tau$, $\tau'$ and the natural surjection $\bb{Z}_{p}\twoheadrightarrow \bb{Z}/p^n\bb{Z}$ respectively,
and consider the following commutative diagram consists of exact rows.
$$
\xymatrix{
0\ar[r] & \pi_1(X'_n) \ar[r] \ar^{f_{n*}}[d] & \pi_1(X') \ar^{\tau'_n}[r] \ar^{f_{0*}}_{\cong}[d]& \bb{Z}/p^n\bb{Z} \ar[r] \ar^{\iota_n}[d] &0\\%
0\ar[r] & \pi_1(X_n) \ar[r] & \pi_1(X) \ar^{\tau_n}[r] & \bb{Z}/p^n\bb{Z} \ar[r] &0}
$$
By diagram chasing, taking $f_n$ is equivalent to taking $\iota_n$, and
$f_{n*}$ is isomorphic if and only if $\iota_n$ is isomorphic.
Then the conclusion follows immediately.
\end{proof}
A branched $\bb{Z}_{p}$-cover $\wt{M}$ over $(M,L)$ is isomorphic to one obtained from a $\bb{Z}$-cover
if and only if the defining homomorphism $\tau$ factors some homomorphism $\bb{Z}\hookrightarrow \bb{Z}_{p}$.
If $L$ is a knot and $M$ is a $\bb{Q}$HS$^3$, then $\wt{M}$ is always isomorphic to one obtained from a $\bb{Z}$-cover.
However, if $L$ has more than one component, \blue{then} $\wt{M}$ may not:
\begin{exa}
Let $p\equiv 1$ (mod 4). Then $\sqrt{-1}\in \bb{Z}_{p}^*$.
Let $L=K_1\cup K_2$ be a 2-component link in \blue{$M=S^3$ and put $X=M-L$.}
Let $\mu_i$ denote the meridians of $K_i$\blue{,} and
let $\wt{M}$ be a branched $\bb{Z}_{p}$-cover over $(M,L)$ defined by
$\tau: \blue{H_1(X)\to \bb{Z}_{p};}\mu_1\mapsto 1, \mu_2\mapsto \sqrt{-1}$.
\blue{Since 1 and $\sqrt{-1}$ cannot move into $\bb{Z}$ at the same time by multiplying any unit of $\bb{Z}_{p}$, }
$\tau$ does not factor through any homomorphism $\bb{Z}\hookrightarrow \bb{Z}_{p}$\blue{. Therefore}
$\wt{M}$ cannot be obtained from a $\bb{Z}$-cover.
More concretely, let $p=5$. Then $\sqrt{-1}=....1212_{(5)}$ (base 5).
Let $\wt{M'}$ be another $\bb{Z}_{p}$-cover defined by $\tau':\mu_1\mapsto 1, \mu_2\mapsto 12_{(5)}$. Then $h_n$ and $h_n$ are isomorphic for $n\leq 2$. Thus, a $\bb{Z}_{p}$-cover can be ``approximate'' by $\bb{Z}$-cover as much as we like.
\end{exa}
\bred{We remark that an analogue of the Hilbert ramification theory for $\bb{Z}_{p}$-extensions holds in a natural way. It is an immediate consequence of
the theory for finite covers given in \cite{Morishita2012} Chapter 5 or \cite{Ueki1}:}
\bred{
Let $\wt{M}=\(h_n:M_n\to M\)_n$ be a branched $\bb{Z}_{p}$-cover over $(M,L)$, let $K\subset M$ be a knot in or outside $L$, and let $\wt{K}=\(K_n\)_n$ denote an inverse system of knots over $K$ in $\wt{N}$.
For each $n$, there is a subgroup of $\mathop{\mathrm{Gal}}\nolimits(h_n)$ called the inertia group $I_{K_n}$ and the decomposition group $D_{K_n}$ of $K_n$ in $h_n$.
They satisfy $I_{K_n}<D_{K_n}<\mathop{\mathrm{Gal}}\nolimits(h_n)$ and control the behavior of $K_n$ as follows: Let $h_n:M_n\to T_n\to Z_n \to M$ denote the corresponding decomposition. Then \violet{(the images of)} $K_n$ is totally branched in $M_n\to T_n$, totally inert in $T_n\to Z_n$, and \red{totally} decomposed in $Z_n\to M$. Moreover, since each subgroup of $\mathop{\mathrm{Gal}}\nolimits(h_n)\cong \bb{Z}/p^n\bb{Z}$ is equal to $\mathop{\mathrm{Gal}}\nolimits(h_{m,n})$ for some $m\leq n$, $T_n=M_{n_1}$ and $Z_n=M_{n_2}$ for some $n_2\leq n_1\leq n$.}
\bred{\violet{Note that $\(I_{K_n}\)_n$ and $\(D_{K_n}\)_n$ form surjective systems.} We define \emph{the inertia group} $I_\wt{K}$ and \emph{the decomposition group} $D_\wt{K}$ of $\wt{K}$ as their inverse limits. \violet{Since $\varprojlim \mathop{\mathrm{Gal}}\nolimits(h_n)\cong \bb{Z}_{p}$,} they are open subgroups with $I_{\wt{K}}<D_{\wt{K}}<\varprojlim \mathop{\mathrm{Gal}}\nolimits(h_n)$, and they control the behavior of $\wt{K}$ in $\wt{M}$ in a similar way to the case of finite covers.
\red{Since each open subgroup $G'$ of $\varprojlim \mathop{\mathrm{Gal}}\nolimits(h_n)\cong \bb{Z}_{p}$ satisfies $\varprojlim \mathop{\mathrm{Gal}}\nolimits(h_n)/G'\cong \mathop{\mathrm{Gal}}\nolimits(h_m)$ for some $m\in \bb{N}$}, we have the following.}
\begin{prop}
\bred{
Let $\wt{M}=\(h_n:M_n\to M\)_n$ be a branched $\bb{Z}_{p}$-cover over $(M,L)$, let $K\subset M$ be a knot in or outside $L$, and let $\wt{K}=\(K_n\)_n$ denote a surjecitve system of knots over $K$ in $\wt{N}$.
Then $\wt{K}$ satisfies one of the following: (i) infinitely branched, finitely inert, and finitely branched, (ii) unbranched, infinitely inert, and finitely decomposed, (iii) unbranched, \violet{non-inert}, and totally decomposed.}
\end{prop}
\violet{We say a branched $\bb{Z}_{p}$-cover $\wt{M}=\(h_n:M_n\to M\)_n$ over $(M,L)$ is \emph{properly branched} if it satisfies the following equivalent conditions: (1) Some $h_n$ is properly branched over some non-empty link, (2) $\wt{M}$ is infinitely branched over some non-empty link, (3) $\tau(\mu_i)\neq 0$ for some meridian $\mu_i$ of $L$. }
\subsection{The Iwasawa type formula and Sakuma's exact sequence}
In this subsection, we prove the Iwasawa type formula for a branched $\bb{Z}_{p}$-cover of $\bb{Q}$HS$^3$. Moreover, we state a $p$-adic variant of Sakuma's exact sequence. Then we deduce an alternative \blue{direct} proof of the Iwasawa type formula, and also an assertion on a direct limit module.
\blue{For any manifold $M$, we denote $H_1(M)_{[p]}:= H_1(M,\bb{Z}_{p})=H_1(M)\otimes \bb{Z}_{p}$. If $M$ is a $\bb{Q}$HS$^3$, then $H_1(M)_{[p]}$ is identified with the $p$-torsion subgroup of $H_1(M)$.}
A generalization of the analogue of Iwasawa's class number formula (\cite{HMM2006}, \cite{KM2008}) describes the
behavior of $p$-torsions in a branched $\bb{Z}_{p}$-cover of $\bb{Q}$HS$^3$:
\begin{thm}[the Iwasawa type formula] \label{ITF}
Let $\wt{M}=\(h_n:M_n\to M\)_n$ be a branched $\bb{Z}_{p}$-cover of $\bb{Q}$HS$^3$ over $(M,L)$.
Then for the $p$-torsion subgroups $H_1(M_n)_{[p]}$ of \blue{the} 1st homology groups,
there are some $\lambda, \mu \in \bb{N}, \nu\in \bb{Z}$ and $n_0 \in \bb{N}$ such that for $n>n_0$,
$$\#H_1(M_n)_{[p]} = p^{\lambda n+\mu p^n+\nu}.$$
\end{thm}
\begin{dfn}
These $\lambda,\mu,$ and $\nu$ are called \emph{the Iwasawa invariants} of $\wt{M}$.
\magenta{They are sometimes denoted as}
\violet{$\lambda_{\wt{M}},\mu_{\wt{M}}$, and $\nu_{\wt{M}}$.}
\end{dfn}
Since a branched $\bb{Z}_{p}$-cover can be approximated by $\bb{Z}$-covers as much as we like, the proof of this formula comes down to the case obtained from a $\bb{Z}$-cover (\cite{KM2008}).
\begin{proof}
For each $n_1\in \bb{N}$, let $\tau_{n_1}:H_1(X)\twoheadrightarrow \bb{Z}/p^{n_1}\bb{Z}$ be the defining homomorphism of $M_{n_1}\to M$.
Since $\tau_{n_1}$ lifts to the homomorphism $\tau$ to $\bb{Z}_{p}$,
$H_1(X)_{\rm tor}<\mathop{\mathrm{Ker}}\nolimits \tau_{n_1}$, and
$\tau_{n_1}$ sends a torsion-free element $b\in H_1(X)$ to a unit of $\bb{Z}_{p}$.
By composing an automorphism of $\bb{Z}/p^{n_1}\bb{Z}$, we may assume that $\tau_{n_1}(b)=1$.
Then $\tau_{n_1}$ lifts to a surjective homomorphism $\wt{\tau_{n_1}}:H_1(X_L)\twoheadrightarrow \bb{Z}$ obtained as follows:
take a $\bb{Z}$-basis of $H_1(X)_{\rm free}:=H_1(X)/H_1(X)_{\rm tor}$, and define $H_1(X)_{\rm free}\twoheadrightarrow \bb{Z}$ by sending basis to the integers with the same presentation of the images by $\tau_{n_1}$, and let $\wt{\tau_{n_1}}$ be the composite with $H_1(X)\twoheadrightarrow H_1(X)_{\rm free}$.
\blue{Note that} there is $n_0$ independent of $n_1$ and
the Iwasawa type formula holds for $n_0<n<n_1$.
Indeed, $n_0$ is determined by $H_1(M)$ and the $p$-adic valuations of the images of $\tau$ (\cite{KM2008}).
Therefore, for sufficiently large $n_1$, the Iwasawa invariants of branched covers $\bb{Z}_{p}$ defined by $\wt{\tau_{n_1}}$ and $\tau$ coincide.
\end{proof}
In the proof of the \blue{Iwasawa type} formula by Kadokami--Mizusawa in \cite{KM2008},
Sakuma's exact sequence (\cite{Sakuma1981} Section 4, \cite{KM2008} Lemma 3.4) played a key role.
A $p$-adic variant of this sequence is stated in the following.
\blue{Let $L=\sqcup K_i$ be a link in a 3-manifold $M$, put $X=M-L$, and let $\mu_i \in H_1(X)$ denote the meridian of $K_i$. Let $\wt{M}$ be a branched $\bb{Z}_{p}$-cover over $(M,L)$ defined by $\tau:H_1(X)\to \bb{Z}_{p}$ \green{and suppose that it is properly branched}.
Then $h_{n,n+1 *}: H_1(M_{n+1})\to H_1(M_n)$ is surjective for any $n\gg 0$ (e.g., \cite{Ueki1} Theorem 6).
We define the Iwasawa module of $\wt{M}$ by $\mca{H}:=\varprojlim H_1(M_n)_{[p]}$.
We fix an identification $\Lambda=\bb{Z}_{p}[[T]]\cong \bb{Z}_{p}[[\wh{\langle t\rangle}]] = \varprojlim \bb{Z}_{p}[t]/(t^{p^n}-1); 1+T\longleftrightarrow t$. Then $\mca{H}$ is a $\Lambda$-module.
For each $n$, $H_1(M_n)_{[p]}$ is a $\bb{Z}_{p}[t]/(t^{p^n}-1)$-module, and hence also is a $\Lambda$-module. Put $\nu_{p^n}=1+t+...+t^{p^n-1}$.
Now we have the following:}
\begin{prop}[Sakuma's exact sequence]\label{Sakuma} Let $\wt{M}$ be a branched $\bb{Z}_{p}$-cover
and let the notation be as above. \green{Suppose that $\wt{M}$ is properly branched.}\\
(1) If $\tau(\mu_i)\neq 0 \mathop{\mathrm{mod}}\nolimits p$ for every $i$, \blue{equivalent to say, if $\wt{M}$ is totally branched over $L$, }then for each $n$ we have an exact sequence
$$H_1(M)_{[p]}\to H_1(M_n)_{[p]}\to \mca{H}/\nu_{p^n}\mca{H} \to 0.$$
\noindent
(2) In any case, there is some $n_0$ such that for any $n>n_0$ we have an exact sequence
$$H_1(M_{n_0})_{[p]}\to H_1(M_n)_{[p]}\to \mca{H}/\nu_{p^{n-n_0}}\mca{H} \to 0.$$
\end{prop}
The proof will be given in the subsequent subsection.
Note that \blue{if $H_1(M)$ or $H_1(M_{n_0})$ is finite, then its image}
by the transfer maps in $H_1(M_n)$'s \blue{is}
constant for $n\gg 0$.
Combining Proposition \ref{Sakuma} and the next lemma, the Iwasawa type formula (Theorem \ref{ITF}) follows immediately from the structure theorem of compact $\Lambda$-modules (Lemma \ref{Lambda} (3)),
similarly to the case of number theory in Section 3.
\begin{lem} \blue{Suppose that $\wt{M}$ consists of $\bb{Q}$HS$^3$'s. Then the}
Iwasawa module
$\mca{H}=\varprojlim H_1(M_n)_{[p]}$ of $\wt{M}$ is a finitely generated torsion $\Lambda$-module.
\end{lem}
\begin{proof}
Since $\mca{H}$ is an inverse limit of $\Lambda$-modules with finite orders, it is a compact $\Lambda$-module.
In the composite $H_{1}(M_n)_{[p]}/h_n^!(H_{1}(M_{1})_{[p]})$ $\cong$ $\mca{H}/\nu_{p^{n}}\mca{H}$ $\twoheadrightarrow$ $\mca{H}/(p,T)$,
the right hand term is finite.
Hence the module $\mca{H}$ is finitely generated by Nakayama's lemma (Lemma \ref{Lambda} (1)).
Moreover, since $\mca{H}/\nu_{p^{n}}\mca{H}$ is a finite group, $\mca{H}$ is a torsion $\Lambda$-module by Lemma \ref{Lambda} (3).
\end{proof}
\blue{We note that $\mca{H}$
can be a finitely generated torsion $\Lambda$-module even if $\wt{M}$ does not consist of $\bb{Q}$HS$^3$'s (See Subsection 4.4).}
Finally, we prove an assertion on a direct limit module by the transfer maps.
\blue{It plays an important role in the proof of Kida's formula in Section 7.}
\begin{prop}\label{injlim} Let $\wt{M}$ be a branched $\bb{Z}_{p}$-cover of $\bb{Q}$HS$^3$ and let $H_1(\wt{M})_{[p]}:=\varinjlim H_1(M_n)_{[p]}$ denote the direct limit by the transfer maps. Then, there is an isomorphism of discrete $\Lambda$-modules $H_1(\wt{M})_{[p]}\cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{M}}}\oplus A'$,
where $A'$ is a bounded $\Lambda$-module, namely, there is some $a\in \bb{N}$ such that $p^aA'=$0, and $A'=0$ holds if and only if $\mu_{\wt{M}}=0$.
\end{prop}
\begin{proof}
\blue{In general, let $E$ be a $\Lambda$-module. If $E\sim \Lambda/(p^m)$, then $\varinjlim E/ E/\nu_{p^n} E$ by $\nu_{p^{n+1}}/\nu_{p^n}$ is a bounded infinite group.
If $E\sim \Lambda/(p^m)$, then $\varinjlim E/ E/\nu_{p^n} E\cong (\bb{Q}_{p}/\bb{Z}_{p})^\lambda$.
Let $E\sim E'$ be a pseudo isomorphism of $\Lambda$-modules. Then their direct limits by $\nu_{p^{n+1}}/\nu_{p^n}$ are isomorphic to each other.}
\blue{Now we} consider the following direct system for the exact sequences of Proposition \ref{Sakuma}.
$$\xymatrix{
H_1(M)_{[p]} \ar^{h_{n+1}^!}[r] &H_1(M_{n+1})_{[p]}\ar[r]
&\mca{H}/\nu_{p^{n+1}}\mca{H} \ar[r] &0\\
H_1(M)_{[p]} \ar^{h_n^!}[r] \ar@{=}[u] &H_1(M_n)_{[p]} \ar[r] \ar^{h_{n+1,n}^!}[u] &\mca{H}/\nu_{p^n}\mca{H} \ar[r] \ar^{\nu_{p^{n+1}}/\nu_{p^n}}[u] &0
}$$
\blue{Since} the direct limit functor is exact, there is an isomorphism
$H_1(\wt{M})_{[p]}/h_\infty^!(H_1(M)_{[p]})$ $\cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{M}}}\oplus A''$,
\blue{where}
$A''$ is a bounded $\Lambda$-module, and $A''=0$ if and only if $\mu_{\wt{M}}=0$.
\blue{Since} $h_\infty^!(H_1(M)_{[p]})$ is a finite group,
there is an isomorphism $H_1(\wt{M})_{[p]}\cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{M}}}\oplus A'$ with a bounded module $A'$.
If $\mu_{\wt{M}}=0$, then by Lemma \ref{Lambda} (4), the $p$-ranks of $H_1(M_n)_{[p]}$ are bounded, and hence the $p$-ranks of $h_n^!(H_1(M)_{[p]})$ are also bounded.
Moreover, since $\tau:H_1(X)\to \bb{Z}_{p}$ sends torsions to zero,
$h_{n,n+1*}:H_1(M_{n+1})\to H_1(M_n)$ is surjective on the torsion subgroup.
Thus, for sufficiently large $n$, a map $h_{n,n+1}^!\circ h_{n,n+1*}$ on $H_1(M_{n+1})$ is multiplication by $p$,
and $h_{n}^!(H_1(M)_{[p]})$ is contained in $pH_1(M_{n+1})_{[p]}$. Therefore $h_{\infty}^!(H_1(M)_{[p]})$ has to be a subgroup of a divisible group, and there is an isomorphism
$H_1(\wt{M})_{[p]}\cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{M}}}$.
Conversely, if this isomorphism holds, then $\mu_{\wt{M}}=0$ clearly holds. \end{proof}
\subsection{Proof of Sakuma's exact sequence (Proposition \ref{Sakuma})}
We prove Proposition \ref{Sakuma} by modifying the argument in \cite{KM2008}, Chapter 3.
The assertion (2) can be reduced to (1). Indeed,
\blue{consider the branched cover $h_{n,n+1}:M_{n+1}\to M_n$ of degree $p$,
let $K$ be a component of $h_n^{-1}(L)$ in $M_n$, let $\mu$ denote its meridian,
and $\wt{\mu}$ \bred{a meridian over $\mu$} in $M_{n+1}$.
If $h_{n,n+1}$ is branched along $K$, then $h_{n,n+1*}(\wt{\mu})=p\mu$. If otherwise, then $h_{n,n+1*}(\wt{\mu})=\mu$.
Therefore, if we put $n_0=\max \(v_p(\tau(\mu_i))\)_i$
\bred{where $v_p(x)$ denote the $p$-adic valuation of a number $x$},
then $\tau_{n_0}:H_1(X_{n_0})\to p^{n_0}\bb{Z}_{p}\overset{\cong}{\to} \bb{Z}_{p}$ sends
every meridian $\mu_i$ of $h_{n_0}^{-1}(L)$ in $H_1(X_{n_0})$
to an element whose image by mod $p$ is non-trivial.
Thus, by replacing the base space by $X_{n_0}$, we can reduce (2) to (1).}
\blue{In order to prove the assertion (1), we prepare several lemmas.}
We put $\Lambda_0=\bb{Z}[\langle t\rangle]$.
Then every $H_1(M_n)$ is a $\Lambda_0=\bb{Z}[\langle t\rangle]$-module.
\begin{lem} \label{lem-mn}
\blue{Suppose $n<m$. If $\wt{M}$ is totally branched over $L$, then there is
\bred{a natural} exact sequence of $\Lambda_0$-modules:}
$$p^{m-n}\bb{Z}/p^m \bb{Z} \to H_1(X_m)/\nu_{p^n}H_1(X_m) \to H_1(M_n)/h_n^!(H_1(M))\to 0.$$
\blue{In addition, for the meridian module $\langle\wt{\mu_i}\rangle :=\mathop{\mathrm{Ker}}\nolimits(H_1(X_m)\to H_1(M_m))$,
there is
\bred{a natural} exact sequence:}
$$0\to \langle\wt{\mu_i}\rangle/\nu_{p^n} \to H_1(X_m)/\nu_{p^n}H_1(X_m
\to H_1(M_m)/\nu_{p^n} H_1(M_m) \to 0.$$
\blue{By taking $\otimes \bb{Z}_{p}$, similar exact sequences of $\Lambda$-modules are obtained.}
\end{lem}
\begin{proof}
Put $G:=\mathop{\mathrm{Gal}}\nolimits(h_m)$ for the subcover $h_m:X_m\to X$ of degree $p^m$. Then there is an exact sequence
$1\to \pi_1(X_m)\to \pi_1(X)\to G \to 1$, and
the Hochschild--Serre spectral sequence (\cite{Brown}) yields
an exact sequence
$H_2(G)\to H_1(\pi_1(X_m))_{G} \to H_1(\pi_1(X)) \to H_1(G)\to 0$.
Since $G=\blue{\langle t \mid t^{p^m}\rangle}\cong \bb{Z}/p^{m}\bb{Z}$ is a finite cyclic group,
by applying the Hurewicz isomorphism, we have
$H_2(G)=0$,
$H_1(\pi_1(X_m))_G=(\pi_1(X_m)^{\rm ab})_G\cong H_1(X_m)_G=H_1(X_m)/(t-1)H_1(X_m)$,
$H_1(\pi_1(X))=\pi_1(X)^{\rm ab}\cong H_1(X)$, and $H_1(G)=G\cong \bb{Z}/p^m\bb{Z}$.
Therefore we obtain an exact sequence $0\to (t-1)H_1(X_m)\to H_1(X_m) \to H_1(X) \to \bb{Z}/p^m\bb{Z} \to 0$.
In a similar way, for a subcover $h_{n,m}:X_m\to X_n$,
an exact sequence
$1\to \pi_1(X_m)\to \pi_1(X_n)\to \mathop{\mathrm{Gal}}\nolimits(h_{n,m})\to 1$
yields an exact sequence
$0\to (t^{p^n}-1)H_1(X_m)\to H_1(X_m) \to H_1(X_n) \to p^n(\bb{Z}/p^m\bb{Z}) \to 0$.
Therefore, a commutative diagram with exact rows
$$\xymatrix{
C_*(X_m) \ar^{t^{p^n} -1}[r] &C_*(X_m) \ar^{h_{n,m*}}[r] &C_*(X_n)\ar[r] & 0\\
C_*(X_m) \ar^{t-1}[r] \ar@{=}[u] &C_*(X_m) \ar^{h_{m*}}[r] \ar^{\nu_{p^n}}[u] &C_*(X)\ar[r] \ar^{h_n^!}[u] &0
}$$
yields the following commutative diagram with exact rows
$$\xymatrix{
H_1(X_m) \ar^{t^{p^n}-1}[r] &H_1(X_m)\ar[r] &H_1(X_n)\ar^{\wt{\tau}_n}[r] &p^n(\bb{Z}/p^m\bb{Z}) \ar[r] &0 \\
H_1(X_m) \ar@{=}[u] \ar^{t-1}[r] &H_1(X_m) \ar[r] \ar^{\nu_{p^n}}[u] &H_1(X)\ar^{\wt{\tau}}[r] \ar^{h_n^!}[u] & \bb{Z}/p^m\bb{Z} \ar[r] \ar^{\times p^n}[u]&0.
}$$
Since $\nu_{p^n}:(t-1)H_1(X_m)\to (t^{p^n}-1)H_1(X_m)$ is surjective,
there is an isomorphism
$\mathop{\mathrm{Coker}}\nolimits(\nu_{p^n}:H_1(X_m)\to H_1(X_m))\cong \mathop{\mathrm{Coker}}\nolimits(\nu_{p^n}: H_1(X_m)/(t-1)H_1(X_m)\to$ $H_1(X_m)/(t^{p^n}-1)H_1(X_m)$.
\blue{By} the Snake lemma, \blue{we obtain}
an exact sequence $p^{m-n}\bb{Z}/p^m \bb{Z} \to H_1(X_m)/\nu_{p^n}H_1(X_m) \to H_1(X_n)/h_n^!(H_1(X))\to 0$.
\blue{Now we consider the following commutative diagram with exact rows.
$$\xymatrix{
H_2(M_n, X_n) \ar^{\partial_n}[r] &H_1(X_n) \ar[r] &H_1(M_n) \ar[r] &0\\
H_2(M, X) \ar[u] \ar^{\partial_1}[r] &H_1(X) \ar^{h_n^!}[u] \ar[r] &H_1(M) \ar^{h_n^!}[u] \ar[r] &0
}$$
By the assumption of (1) that $\wt{M}$ is totally branched over $L$, the transfer map $h_n^!$ is surjective on the meridians. Hence we have an isomorphism $h_n^!: \mathop{\mathrm{Im}}\nolimits \partial_1\to \mathop{\mathrm{Im}}\nolimits \partial_n$.
By the snake lemma, we obtain an isomorphism $H_1(X_n)/h_n^!(H_1(X))\overset{\cong}{\to} H_1(M_n)/h_n^!(H_1(M))$ on the $\mathop{\mathrm{Coker}}\nolimits$ of two vertical morphisms on the right side.}
Thus we have obtained an exact sequence of $\Lambda_0$-modules. \end{proof}
\blue{Next, we}
recall some facts on inverse limits.
We say \violet{that} an inverse system $\(A_n\)_n$ \blue{of abelian groups} satisfies \emph{the Mittag-Leffler condition} (ML-condition) if
``for any $n$ there exists some $n'$ such that for any $n''>n'$, $\mathop{\mathrm{Im}}\nolimits(A_{n''}\to A_n)$ is constant''.
If $\(A_n\)_n$ consists of finite \violet{groups}, or if its morphisms are all surjective,
then it satisfies ML-condition.
An inverse system $\(A_n\)_n$ is said to be \emph{ML-zero} if it satisfies a modified ML-condition in which ``constant'' is replaced by ``zero''.
\violet{Recall $\Lambda=\bb{Z}_{p}[[T]]$. A \emph{profinite $\Lambda$-module} is one obtained as the inverse limit of discrete $\Lambda$-modules. A \emph{homomorphism of profinite $\Lambda$-modules} is a continuous $\Lambda$-module homomorphism. It is always a closed map, because it is a continuous map from a compact space to a Hausdorff space.
Note that profinite $\Lambda$-modules form an abelian category. By \cite{Jannsen1988} \S 1, we have the following lemma.}
\begin{lem}\label{ML} Let $\(A_n\)_n, \(B_n\)_n, \(C_n\)_n$ be inverse systems of \violet{profinite $\Lambda$-modules.}\\
(1) Then, an exact sequence $0\to A_n \to B_n \to C_n \to 0$ of inverse systems yields an exact sequence
$0\to \varprojlim A_n \to \varprojlim B_n \to \varprojlim C_n \to \varprojlim^1 A_n$.
If $\(A_n\)_n$ satisfies \blue{the} ML-condition, then $\varprojlim^1 A_n=0$ holds.\\
(2) If $\(A_n\)_n$ is ML-zero, then an exact sequence $A_n \to B_n \to C_n \to 0$ yields
an isomorphism $\varprojlim B_n \overset{\cong}{\to} \varprojlim C_n$.\\
(3) \violet{Suppose that $\(A_n\)_n$ is a surjective system and let $\(f_n:A_n\to A_n\)_n$ be a family of endomorphisms commutative with the system.}
Put $\mca{A}=\varprojlim A_n$ and let $f:\mca{A}\to \mca{A}$ denote the induced map. Then $\mca{A}/f(\mca{A})\overset{\cong}{\to} \varprojlim A_n/f_n(A_n)$.
\end{lem}
\begin{proof}
\violet{Here we prove (3). Since each $f(\mca{A})\to f_n(A_n)$ is a surjection,
the inclusion map $\iota:f(\mca{A})\hookrightarrow \varprojlim_n f_n(A_n)$ is a continuous homomorphism with dense image. Since $\iota$ is a closed map, $\iota$ is an isomorphism and we have $f(\mca{A})= \varprojlim_n f_n(A_n)$.}
Now consider \violet{exact sequences} $0\to f_n(A_n) \to A_n \to A_n/f_n(A_n)\to 0$ compatible with the inverse system. Since $\(f_n(A_n)\)_n$ is a surjective system, (1) yields an exact sequence $0\to \varprojlim_n f_n(A_n) \to \mca{A} \to \varprojlim_n A_n/f_n(A_n)\to 0$. Thus we obtain $\mca{A}/f(\mca{A})\overset{\cong}{\to} \varprojlim A_n/f_n(A_n)$.
\end{proof}
\blue{Now we give a proof of Proposition 4.11 (1):}
\begin{proof}[Proposition \ref{Sakuma} \blue{(1)}, Sakuma's exact sequence]
We consider the inverse systems with respect to $m$ in Lemma \ref{lem-mn}.
In the first exact sequence, the \red{transition morphisms on} the first terms $L_m:=p^{m-n}\bb{Z}/p^m\bb{Z} \red{\cong \langle t^{p^{m-n}} | t^{p^m}\rangle}$ are
\red{ $\{ \mathop{\mathrm{mod}}\nolimits p^m: L_{m+1}\to L_m \}_m$.
Hence}
if $m'>\red{m+n},$ then $L_{m'}\to L_m$ is the zero map.
Thus the first term is ML-zero,
\red{and Lemma \ref{ML} (2) yields}
an isomorphism $\varprojlim_m (H_1(X_m)_{[p]}/\blue{\nu_{p^n}} H_1(X_m)_{[p]})$ $\red{\overset{\cong}{\to}} H_1(\blue{M_n})_{[p]}/h_n^!(H_1(M)_{[p]})$.
In the second exact sequence in Lemma \ref{lem-mn},
since the
\red{transition maps on} the finite $p$-torsion \red{abelian} groups $\langle\mu_i\rangle/\nu_{p^n}$ are the multiplication by $p$,
\red{these terms} also satisfy ML-zero\blue{. Therefore by}
Lemma \ref{ML} (2), there is an isomorphism
$\varprojlim_m(H_1(X_m)_{[p]}/\nu_{p^n} H_1(X_m)_{[p]})\cong \varprojlim_m(H_1(M_m)_{[p]}/\nu_{p^n} H_1(M_m)_{[p]})$.
\blue{Since $\(H_1(M_m)_{[p]}\)_m$ is a surjective system of pro-$p$ \red{abelian} groups,
Lemma \ref{ML} (3) yields an isomorphism
$\varprojlim_m(H_1(M_m)_{[p]}/\nu_{p^n} H_1(M_m)_{[p]})\cong \mca{H}/\nu_{p^n}\mca{H}$.
}
\end{proof}
\subsection{Remarks on non-$\bb{Q}$HS$^3$ cases}
\blue{In this subsection, we briefly discuss further generalization of Iwasawa type formula for non-$\bb{Q}$HS$^3$ cases.} %
\begin{lem} (1) Let $E\sim \oplus_i \Lambda/(p_i^{m_i}) \oplus \oplus_j \Lambda/(f_j^{e_j})$ be a pseudo isomorphism from a finitely generated torsion $\Lambda$-module to the normal form. Then $E/\nu_{p^n}E$ is a infinite group if and only if $f_i$ are not $p^{n'}$-th cyclotomic polynomial for any $\red{1\leq} n'\leq n$.\\
(2) Let $f$ be \red{a $p$-power-th cyclotomic polynomial in $\bb{Z}[t]$}, which is an irreducible Weierstrass polynomial in $\bb{Z}_{p}[[T]]$, and put $E=\Lambda/(f^e)$. Then the following are equivalent:
that $E/\nu_{p^n}E$ is constant for $n\gg 0$, that $(f^e, \nu_{p^n})$ is constant for $n\gg 0$, and that $e=1$. \end{lem}
\begin{proof} Note that $\nu_{p^n}=\nu_{p^n}(t)=(t^{p^n}-1)/(t-1)=\nu_{p^n}(1+T)$ in $\bb{Z}[t]\subset \bb{Z}_{p}[[T]]; t\mapsto 1+T$ is the product of all the $p^{n'}$-th cyclotomic polynomials for $1\leq n'\leq n$.
(1) A module of the form $(\Lambda/(p^{m_i}))/\nu_{p^n}(\Lambda/(p^{m_i}))$ is finite in any case.
By a general fact that monic polynomials $g,h\in \bb{Z}[t]$ with positive degrees satisfy $\#\bb{Z}[t]/(g,h)<\infty$ if and only if they have no common divisor,
a module of the form $(\Lambda/(f_j^{e^j}))/\nu_{p^n}(\Lambda/(f_j^{e^j}))=\Lambda/(f_j^{e_j}, \nu_{p^n})$ is finite if and only if the monic polynomials $f_j^{e_j}$ and $\nu_{p^n}$ have no common divisor, that is, $f_j$ is not the $p^{n'}$-th cyclotomic polynomial for $1\leq n'\leq n$.
(2) Let $f$ be the $p^n$-th cyclotomic polynomial in $\bb{Z}[t]\subset \bb{Z}_{p}[[T]]$.
Then $(f,\nu_{p^{n'}})=(f)$ holds for any $n'\geq n$.
We suppose both $e>1$ and that $E/\nu_{p^{n'}}E$ is constant for $n'\gg 0$, and lead a contradiction.
By these hypothesises, we have $(f^e, \nu_{p^n})=(f)(f^{e-1}, \nu_{p^n}/f)$,
and there is some $n'\geq n$ with $(f^{e-1}, \nu_{p^{n'}}/f)=(f^{e-1}, \nu_{p^{n'+1}}/f)$.
By taking $\mathop{\mathrm{mod}}\nolimits f$, we have
$(\nu_{p^{n'}}/f) = (\nu_{p^{n'+1}}/f)$ in $\Lambda/(f)$.
Let $\alpha=\alpha_1,\ldots,\alpha_{\blue{\varphi(p^n)}}$ denote the roots of $f$ in an algebraic closure $\overline{\bb{Z}}_p$ of $\bb{Z}_{p}$, where \blue{$\varphi(n)$} is the Euler function.
Then we have a standard injective homomorphism $\Lambda/(f)\hookrightarrow \overline{\bb{Z}}_p^{\blue{\varphi(p^n)}}$,
and $((\nu_{p^{n'}}/f)(\alpha))= ((\nu_{p^{n'+1}}/f)(\alpha))$ in $\overline{\bb{Z}}_p$.
However, the $p^{n'+1}$-th cyclotomic polynomial $q_{p^{n'+1}}$ satisfies $\nu_{p^{n'+1}}/f=q_{p^{n'+1}}\nu_{p^{n'}}/f$ and $(q_{p^{n'+1}}(\alpha))=(p^{n'+1-n})\neq (1)$.
Thus we have $((\nu_{p^{n'}}/f)(\alpha))\neq ((\nu_{p^{n'+1}}/f)(\alpha))$, and hence contradiction.
\end{proof}
\begin{thm}
Let $\wt{M}$ be a branched $\bb{Z}_{p}$-cover of 3-manifolds,
\blue{and suppose that $\mca{H}=\varprojlim H_1(M_n)_{[p]}$ is a finitely generated torsion $\Lambda$-module.} Then\\
\noindent
(1) The $n$-th cover $M_n$ is $\bb{Q}$HS$^3$ if and only if $M$ is a $\bb{Q}$HS$^3$ and the \blue{characteristic} polynomial $char_{\mca{H}}$ of $\mca{H}$ is not divided by $p^{n'}$-th cyclotomic polynomial for $n'\leq n$. \\
(2) If every $p$-power-th cyclotomic polynomial contained in $char_{\mca{H}}$ is of exponent 1, then the $p$-torsion subgroups $H_1(M_n)_{[p]}$ satisfy the Iwasawa type formula. \end{thm}
\begin{proof}
Let $\mca{H}\sim \oplus_i \Lambda/(p_i^{m_i}) \oplus \oplus_j \Lambda/(f_j^{e_j})$ be a pseudo isomorphism to a normal form,
and put $E_j:=\Lambda/(f_j^{e_j})$ for each $j$.
Then $E_j/\nu_{p^n}E_j$ contains a free $\bb{Z}_{p}$-module if and only if
$f_j$ is the $p^{n'}$-th cyclotomic polynomial for some $n'\leq n$.
If $E_j/\nu_{p^n}E_j$ is constant for $n\gg 0$ for every $j$ such that $f_j$ is a $p$-power-th cyclotomic polynomial, then by considering the direct sum of factors which do not correspond to such $j$'s, we have the same augment as before. Therefore the previous lemma yields this theorem.
\end{proof}
\begin{exa}
\blue{Let $M$ be a $\bb{Q}$HS$^3$, let $L=\sqcup K_i$ be a link with null homologous $d$-components in $M$, put $X=M-L$, and let $\mu_i\in H_1(X)$ denote the meridian of $K_i$.
Let $v_1,...,v_d$ be units of $\bb{Z}_{p}$, let $\tau: H_1(X)\to \bb{Z}_{p}$ be a homomorphism defined by $\mu_i \mapsto v_i$, and let $\wt{M}$ be the branched $\bb{Z}_{p}$-cover defined by $\tau$.
Then $\wt{M}$ is totally branched over $L$.
Let $\Delta_{L,\tau}(t)=\Delta_{L}(t^{v_1},\ldots,t^{v_d})$ denote the $p$-adic reduced Alexander polynomial in $\bb{Z}_{p}[[\wh{\langle t\rangle}]]$.
Since we can approximate $\wt{M}$ by $\bb{Z}$-covers as much as we can,
from a well-known fact for $\bb{Z}$-covers and \cite{KM2008} Theorem 3.3 (a variant of the Mayberry--Murasugi formula \cite{MM1982} or Porti's result \cite{Porti2004}),
we can easily deduce isomorphisms $\mca{H}\overset{\cong}{\to} \Lambda/(\Delta_{L,\tau}(t))$ and $H_1(M_n)/h_n^{!}(H_1(M))_{[p]}\overset{\cong}{\to} \mca{H}/\nu_{p^n}\mca{H}$, and
$||H_1(M_n)/h_n^{!}(H_1(M))||_p=\prod_{\zeta^n=1} |\Delta_{L,\tau}(\zeta))|_p$.
(For a group $G$, we put $|G|=\#G$ or zero
according as $G$ is finite or infinite. For a number $x$, $|x|_p$ denotes the $p$-adic norm.)}
Typical examples with cyclotomic Alexander polynomial is torus knots.
\blue{Only for here, the subscription $n$ means the covering degree over $S^3$.
If $L=K$ is the trefoil in $M=S^3$,}
then $\Delta_K(t)=t^2-t+1$ is the 6-th cyclotomic polynomial, and
$H_1(M_6)=\bb{Z}^2$, $H_1(M_3)=(\bb{Z}/2\bb{Z})^2$, $H_1(M_2)=\bb{Z}/3\bb{Z}$ are known.
For 3-fold cover $M_6\to M_2$ and double-cover $M_6\to M_3$,
the maps on $H_1$ are not surjections on 3-torsions and 2-torsions respectively. However, for $n>0$, $M_{2\cdot 3^{n+1}}\to M_{2\cdot 3^n}$ and $M_{3\cdot 2^{n+1}}\to M_{3\cdot 2^n}$ induce surjections on 3-torsions and 2-torsions respectively, and satisfy the Iwasawa type formula with trivial invariants. If we take the connected sum with the trefoil and the figure eight knot, then we obtain a non-trivial example. \end{exa}
\begin{rem}
\blue{For a branched $\bb{Z}$-cover over a link $L$ in $S^3$, we have the balance formula among
\emph{the $p$-adic Mahler measure} of the Alexander polynomial, the
the Iwasawa $\mu$-invariant, and the $p$-adic entropy (\cite{Ueki4}).}
\blue{We expect further study (i) with use of higher Alexander polynomial (\cite{SW2002M}, \cite{Le2014}), (ii) for graph-branched cases (\cite{Porti2004}), (iii) about the asymptotic formula related to hyperbolic volumes (\cite{BV2013}, \cite{Le2014}).}
\end{rem}
\section
\bred{Morphisms} of branched $\bb{Z}_{p}$-covers}
In this section, we introduce an analogue object of an extension of $\bb{Z}_{p}$-fields,
and discuss a condition for $\mu=0$.
We also define \emph{chains} of branched $\bb{Z}_{p}$-covers by the direct limits with respect to the transfer maps, and calculate Tate cohomologies for
\bred{an equivariant Galois morphism} of branched $\bb{Z}_{p}$-covers.
\subsection{Extensions of $\bb{Z}_{p}$-fields}
Let $k_\infty/k$ be a $\bb{Z}_{p}$-extension over a finite number field
and let $F_\infty/k_\infty$ be a $p$-extension.
Then there is a $p$-extension $F/k$ such that $F_\infty/F$ is a $\bb{Z}_{p}$-extension and $F_n=Fk_n$ for each $n$.
In such a situation, behaviors of the Iwasawa invariants are studied (\cite{Iwasawa1973}, \cite{Iwasawa1981}).
Recall that for a $\bb{Z}_{p}$-extension $k_\infty/k$, the value of $\lambda$ and the property whether $\mu=0$ or not are independent of the choice of base field $k$, while the values of $\mu$ and $\nu$ depend.
Let $k_1$ the 1st middle field of $k_\infty/k$. Then a $\bb{Z}_{p}$-extension $k_\infty/k_1$ has $\mu$ and $\nu$ different from those of $k_\infty/k$.
In this case, if we put $F_n:=k_{n+1}$, then we have $F_n=Fk_n$.
However, we do not assume such a case as an essential extension of $\bb{Z}_{p}$-fields.
\subsection
\bred{Morphisms of branched $\bb{Z}_{p}$-covers}}
We define an analogue notion of an extension of $\bb{Z}_{p}$-fields \blue{as follows:}
\begin{dfn}[branched cover of branched $\bb{Z}_{p}$-covers]
\blue{Let $\wt{M}$ and $\wt{N\,}\!$ be branched $\bb{Z}_{p}$-covers over $(M,L)$ and $(N,L')$ respectively. Then \bred{a \emph{morphism} $f:\wt{N\,}\!\to \wt{M}$ of branched $\bb{Z}_{p}$-covers} is a compatible system of branched covers on each layer, that is,
a family $\( f_n:N_n \to M_n\)_n$ of branched covers commutative with the \red{transition maps.}
If
every $f_n$ is Galois,
\red{then} we say this
\bred{morphism} is \emph{Galois}. We put $G_n=\mathop{\mathrm{Gal}}\nolimits(f_n)$. If all the induced maps $G_n\to G_0$ are isomorphisms, then we call this
\bred{morphism \emph{equivariant Galois}, or} a \emph{branched Galois cover of branched $\bb{Z}_{p}$-covers}, and we write $\mathop{\mathrm{Gal}}\nolimits(f)=G_0$.}
\end{dfn}
\blue{
Let $f:\wt{M}\to \wt{N\,}\!$ be a
\bred{morphism of branched $\bb{Z}_{p}$-covers with notation} as above.
We can easily check the following facts by diagram chasing, using Proposition 4.8:
If $\wt{M}$ and $\wt{N\,}\!$ are properly branched over $L$ and $L'$, then $L'=f_0^{-1}(L)$.
Let $\overline{S}_n\subset M_n$ denote the branch link of $f_n$.
If $f$ is \bred{equivariant Galois},
then we have $\overline{S}_n\subset h_n^{-1}(\overline{S}_0)$.}
\violet{If in addition if $\overline{S}_0$ is unbranched in $\wt{M}$, then $\overline{S}_n=h_n^{-1}(\overline{S}_0)$ and $L\cap \overline{S}_0=\phi$.}
\blue{
We denote the restrictions to the exteriors of the preimages of \violet{$L\cup \overline{S}_0$} by the same letters
$h_n:X_n\to X$, $h_n':Y_n\to Y$, and $f_n: X_n\to Y_n$ for each $n$.
}
\begin{comment}
\red{Let $f_0:N\to M$ be a branched cover of 3-manifolds branched over a link $\overline{S}$, let $L$ and $L'$ be links in $M$ and $N$ respectively, and
let $\wt{M}$ and $\wt{N\,}\!$ be branched $\bb{Z}_{p}$-covers over $(M,L)$ and $(N,L')$ respectively. Assume that the union $L\cup f_0(L')\cup \overline{S}$ is a link, and let $f_0:Y\to X$ denote the restriction to the exterior.}
A \emph{compatible system} $\red{f:\wt{N\,}\!\to \wt{M}}$ of branched covers on each layer is a family $\( f_n:N_n \to M_n\)_n$ of branched covers commutative with the \blue{transition maps.}
\red{(If $\wt{M}$ and $\wt{N\,}\!$ are properly branched over $L$ and $L'$, then $L'=f_0^{-1}(L)$.)}
If $f_{\red 0}:N\to M$ is Galois, then every $f_n$ is Galois, and we say this system is \emph{Galois}.
We put
$G_n=\mathop{\mathrm{Gal}}\nolimits(f_n)$. If all the induced maps $G_n\to G_{\red 0}$ are isomorphisms, we call this system a \emph{branched Galois cover of branched $\bb{Z}_{p}$-covers}, \red{and we write $\mathop{\mathrm{Gal}}\nolimits(f)=G_0$}.
\end{comment}
\begin{prop} \blue{Let $\wt{M}$ and $\wt{N\,}\!$ be branched $\bb{Z}_{p}$-covers with notation as above, and let $f_0:N\to M$ be a branched cover. Then}
taking a
\bred{morphism $f:\wt{M}\to \wt{N\,}\!$} is equivalent to taking a homomorphism $\iota:\bb{Z}_{p}\to \bb{Z}_{p}$ which commutes with the defining homomorphism $\tau,\tau'$ and $f_{0*}:\pi_1(Y)\to \pi_1(X)$.
\bred{Suppose $f$ is Galois. Then}
the natural maps $G_n\to G$ are isomorphisms if and only if corresponding $\iota:\bb{Z}_{p}\to \bb{Z}_{p}$ is an isomorphism.
\bred{In other words, $f$ is equivariant Galois if and only if $f$ is $\varprojlim\mathop{\mathrm{Gal}}\nolimits(h_n)=\bb{Z}_{p}$-equivariant.} \end{prop}
\begin{proof}
Consider the following commutative diagram with exact rows. $$
\xymatrix{
0\ar[r] & \pi_1(Y_n) \ar[r] \ar^{f_{n*}}[d] & \pi_1(Y) \ar^{\tau'_n}[r] \ar^{f_{\red 0*}}[d]& \bb{Z}/p^n\bb{Z} \ar[r] \ar^{\iota_n}[d] &0\\%
0\ar[r] & \pi_1(X_n) \ar[r] & \pi_1(X) \ar^{\tau_n}[r] & \bb{Z}/p^n\bb{Z} \ar[r] &0}
$$
Then it can be seen that taking $\(f_n\)_n$ and taking $\iota:\bb{Z}_{p}\to \bb{Z}_{p}$ is equivalent, and that $f_{\red 0}$ is Galois if and only all the $f_n$ are Galois.
If Galois, since
$G_n=\pi_1(X_n)/f_{n*}(\pi_1(Y_n))$, the snake lemma yields an exact sequence $0\to \mathop{\mathrm{Ker}}\nolimits(\iota_n)\to G_n \to G_{\red 0} \to \mathop{\mathrm{Coker}}\nolimits(\iota_n)\to 0$. Thus the natural \blue{maps} $G_n\to G_{\red 0}$ \blue{are isomorphisms} if and only if $\iota$ is an isomorphism. \end{proof}
Even if $\iota$ is not an isomorphism, $G_n$'s can be isomorphic to each other \blue{non-canonically}.
Indeed, $\mathop{\mathrm{Ker}}\nolimits \iota_n$ and $\mathop{\mathrm{Coker}}\nolimits \iota_n$ are isomorphic.
For instance, if $\iota$ is the multiplication by $p$ and $G_{\red 0}=\bb{Z}/p\bb{Z}$, then
the situation resembles to the case of $k_1/k$ in the previous subsection.
We can also define a branched cover of branched $\bb{Z}_{p}$-cover which may not be Galois, however we omit here.
Next, we see some examples. The following lemma is obvious.
\begin{lem} Let $f:N\to M$ be a branched cover of 3-manifolds with degree $p$. Let $K$ be a knot in $M$, let $K'$ be a component of $f^{-1}(K)$,
and let $\mu,\mu'$ denote the meridians of $K,K'$ respectively.
\blue{If} $K$ is inert or decomposed, then $f(\mu')=\mu$.
If $K$ is branched, then $f(\mu')=p\mu$.
\end{lem}
This lemma ensures that the following settings give examples.
\begin{exa} \blue{
Let $\mca{L}=K\sqcup \overline{S}$ be a 2-component link in $M=S^3$
and let $f_{\red 0}:N\to M$ be a branched cover of degree $p$ with the branch link $\overline{S}$.
Put $X=M-\mca{L}$, $Y=N-f_{\red 0}^{-1}(\mca{L})$ and $S=f_{\red 0}^{-1}(\overline{S})$.
Let $\mu_K,\mu_{\overline{S}}\in H_1(X)$ \blue{denote} the meridians of $K$ and $\overline{S}$ respectively.
Here we take $\iota={\rm id}_{\bb{Z}_{p}}$.}
\noindent
(1) \blue{Suppose that $L$ is decomposed in $f_0$ as $f_0^{-1}(K)=K_1\sqcup \cdots \sqcup K_p$} (e.g., let $\mca{L}$ be the trivial 2-component link), and let
(i)
$\tau:H_1(X)\to \bb{Z}_{p}; \mu_K\mapsto1, \mu_{\overline{S}}\mapsto 0,$
$\tau':H_1(Y)\to \bb{Z}_{p}; \mu_{K_i}\mapsto 1, \mu_{S}\mapsto 0.$
(ii) $\tau: \mu_K\mapsto 1, \mu_{\overline{S}}\mapsto \blue{1},$
$\tau': \mu_{K_i}\mapsto 1, \mu_{S}\mapsto \blue{p}.$
\noindent
(2) \blue{Suppose that $L$ is inert in $f_0$ as $f_{\red 0}^{-1}(K)=K'$}
(e.g., let $\mca{L}$ be the Hopf link), and let
(i)
$\tau:\mu_K\mapsto1, \mu_{\overline{S}}\mapsto 0,$
$\tau':\mu_{\blue{K'}}\mapsto 1, \mu_{S}\mapsto 0.$
(ii) $\tau: \mu_K\mapsto 1, \mu_{\overline{S}}\mapsto \blue{1},$
$\tau': \mu_{\blue{K'}}\mapsto 1, \mu_{S}\mapsto \blue{p}.$\\
These above give
\bred{equivariant Galois morphisms} of branched $\bb{Z}_{p}$-covers of degree $p$.
\blue{Next, we change the setting and
see an example of non-Galois case.}\\
\noindent
(3)
Let $K$ be a knot in $M=S^3$, put $X=M-K$.
Let $f_{\red 0}:N\to M$ be a cover of degree $p$ branched over $K$,
and let $f_{\red 0}:Y\to X$ be the restriction to the exterior.
Put $K':=f_{\red 0}^{-1}(K)$, and let $\mu, \mu'$ denote the meridians of $K, K'$ respectively. If we put
$\tau: H_1(X)\to \bb{Z}_{p}; \mu_K \mapsto 1$, $\tau': H_1(Y)\to \bb{Z}_{p}; \mu_{K'}\mapsto 1$, and
$\iota: \bb{Z}_{p}\to \bb{Z}_{p}; 1\mapsto p$,
then we obtain an analogous situation of $k_1/k$ in the previous \blue{sub}section.
\end{exa}
\subsection{On conditions for $\mu=0$}
\blue{In this subsection, we recall analogous natures of cyclotomic and anti-cyclotomic $\bb{Z}_{p}$-extensions for branched $\bb{Z}_{p}$-covers investigated in our another paper \cite{Ueki3}.}
\blue{
In a cyclotomic $\bb{Z}_{p}$-extension over a number field,
any non-$p$ prime is finitely decomposed,
and in an anti-cyclotomic $\bb{Z}_{p}$-extension over a CM-field,
any non-$p$ prime is totally decomposed.
For a branched $\bb{Z}_{p}$-covers, instead of ``any non-$p$ primes'', we focus on a certain link.
\begin{prop}[\cite{Ueki3} Proposition 5.1]
Let $M$ be a 3-manifold, let $L$ be a link in $M$ consisting of null-homologous components, and let $\wt{M}=\(h_n:M_n \to M\)_n$ be the TLN-$\bb{Z}_{p}$-cover over $(M,L)$.
Let $K$ be a knot in $M-L$.
If $\mathop{\mathrm{lk}}\nolimits(K,L) \neq 0$, then $K$ is finitely decomposed into a $p^{v_p({\rm lk}(K,L))}$-component link in $\wt{M}$.
If $\mathop{\mathrm{lk}}\nolimits(K,L)=0$, then $K$ is totally decomposed in $\wt{M}$.
\end{prop}
Let $f:\wt{N\,}\!\to \wt{M}$ be
\bred{an equivariant Galois morphism} of branched $\bb{Z}_{p}$-covers, let $\overline{S}\subset M$ denote the branch link of $f_{\red 0}:N\to M$, and suppose $\overline{S}\cap L=\phi$.
If $\overline{S}$ is finitely decomposed in $\wt{M}$, then $f:\wt{N\,}\!\to \wt{M}$ \emph{resembles} a $p$-extension of \magenta{a} cyclotomic $\bb{Z}_{p}$-field.
If $\overline{S}$ is \red{totally} decomposed in $\wt{M}$, then $f:\wt{N\,}\!\to \wt{M}$ \emph{resembles} a $p$-extension of \magenta{an} anti-cyclotomic $\bb{Z}_{p}$-field.}
\blue{
In \cite{Ueki3}, we established an analogue of relative genus theory and studied the behaviors of Iwasawa $\mu$-invariants by following Iwasawa's argument in \cite{Iwasawa1973}.
Suppose that $\wt{M}$ and $\wt{N\,}\!$ consist of $\bb{Q}$HS$^3$'s and that $f:\wt{N\,}\!\to \wt{M}$ is of degree $p$.
If $\overline{S}$ is finitely decomposed in $\wt{M}$, then by \cite{Ueki3} Theorem 5.2, $\mu_{\wt{M}}=0$ if and only if $\mu_{\wt{N\,}\!}=0$.
If
$\overline{S}$ is infinitely decomposed in $\wt{M}$, then by \cite{Ueki3} Theorem 5.3, we have
$\mu_{\wt{N\,}\!}\geq \#(\text{components of }\overline{S})$.
}
\begin{comment}
＊＊＊
In a cyclotomic $\bb{Z}_{p}$-extension over a number field,
any non-$p$ prime is finitely decomposed,
and in an anti-cyclotomic $\bb{Z}_{p}$-extension over a CM-field,
any non-$p$ prime is totally decomposed.
In this section, we recall an analogue of this nature for 3-manifolds investigated in our another paper \cite{Ueki3}, in which we established an analogue of relative genus theory and proved theorems on the behaviors of $\mu$-invariants by following Iwasawa's argument in \cite{Iwasawa1973}.
For a branched $\bb{Z}_{p}$-cover, an easy observation ensures the following proposition.
\begin{prop} Let $M$ be a 3-manifold, let $L$ be a link in $M$ \blue{consisting} of null-homologous components, and let $\wt{M}$ be a TLN-$\bb{Z}_{p}$-cover over $(M,L)$.
Let $K$ be a knot in $M-L$.
If $\mathop{\mathrm{lk}}\nolimits(K,L)=0$, then $K$ is totally decomposed in $\wt{M}$.
If $\mathop{\mathrm{lk}}\nolimits(K,L) \neq 0$, then $K$ is finitely decomposed into a $p^{v_p({\rm lk}(K,L))}$-component link in $\wt{M}$.
\end{prop}
Instead of putting analogous conditions of
``any non-$p$ primes'', we consider as follows:
Let $f:\wt{N\,}\!\to \wt{M}$ be a branched Galois cover of branched $\bb{Z}_{p}$-covers, let $\overline{S}\subset M$ be the branch link of $f_{\red 0}:N\to M$,
and suppose $\overline{S}\cap L=\phi$.
If $\overline{S}$ is totally decomposed in $\wt{M}$, then $f:\wt{N\,}\!\to \wt{M}$ \emph{resembles} a $p$-extension of \magenta{a} cyclotomic $\bb{Z}_{p}$-field.
If $\overline{S}$ is finitely decomposed in $\wt{M}$, then $f:\wt{N\,}\!\to \wt{M}$ \emph{resembles} a $p$-extension of \magenta{an} anti-cyclotomic $\bb{Z}_{p}$-field.
In order to consider an analogue of an extension of cyclotomic $\bb{Z}_{p}$-fields,
for a Galois $p$-cover of branched $\bb{Z}_{p}$-covers of $\bb{Q}$HS$^3$,
we assume that $\overline{S}$ is finitely decomposed in $\wt{M}$.
Then by \cite{Ueki3} Theorem 5.2, $\mu_{\wt{M}}=0$ if and only if $\mu_{\wt{N\,}\!}=0$.
In addition, by \cite{Ueki3} Theorem 5.3, for a cover of degree $p$,
if $\overline{S}$ is infinitely decomposed in $\wt{M}$, then
$\mu_{\wt{N\,}\!}\geq \#(\text{components of }\overline{S})$.
In other words,
\end{comment}
\blue{As a consequence,
we have the following theorem:}
\begin{thm}\label{mu}
Let $f:\wt{N\,}\!\to \wt{M}$ be
\bred{an equivariant Galois morphism} of degree $p$ of branched $\bb{Z}_{p}$-covers of $\bb{Q}$HS$^3$'s,
let $\overline{S}$
\blue{denote} the branch link of $f_0:N\to M$, and suppose $\mu_{\wt{M}}=0$.
Then $\overline{S}$ is finitely decomposed in $\wt{M}$ if and only if $\mu_{\wt{N\,}\!}=0$.
\end{thm}
\subsection{Tate cohomologies of branched $\bb{Z}_{p}$-covers}
\magenta{
In Iwasawa's second proof for Kida's formula in \cite{Iwasawa1981},
the following assertions were used:
\begin{prop}[\cite{Iwasawa1981}]
Let $F_\infty/k_\infty$ be an extension of \cyan{a cyclotomic $\bb{Z}_{p}$-field} of degree $p$,
and let $S$ denote the set of ramified primes in $F_\infty$. Then for each $i\in \bb{Z}$,\\
(1) {\rm [Lemma 5]} The equality $\wh{H}^i(G, F_\infty^*)=0$ holds.\\
(2)\footnote{\magenta{It seems that this assertion was used implicitly.}} Let $\mca{O}_{F_\infty,S}^{*-}$ denote the minus part of $\mca{O}_{F_\infty,S}^*$ and put $\hbar^-_{i}:={\rm rank}\wh{H}^i(G,\mca{O}_{F_\infty,S}^{*-})$.
If $k_\infty$ contains every $p$-power-th roots of unity, then $\hbar^-_2-\hbar^-_1=-1$.\\
(3) {\rm [Lemma 1]} The equality $\wh{H}^i(G,I_{F_\infty,S})=0$ holds.\\
(4) The equality $q(\mca{O}_{F_\infty,S}^*)=q(\mca{O}_{F_\infty}^*)p^{\#S}$ holds.
\end{prop}}
We consider \magenta{their analogues} in the following.
For a branched $\bb{Z}_{p}$-cover $\wt{M}$, we fix CW-structures or PL-structures compatible with
the inverse system, and
define the modules of \emph{chains, cycles, boundaries,} and \emph{homologies} by the direct limits with respect to the transfer maps.
In addition, we also define \emph{$S$-chains} and others as in Subsection 2.3.
Moreover, for
\bred{an equivariant Galois morphism} of branched $\bb{Z}_{p}$-covers $f:\wt{N\,}\!\to \wt{M}$, we put $G=\mathop{\mathrm{Gal}}\nolimits(f)$, and fix CW-structures
or PL-structures compatible with all the \bred{maps in the system}.
Let $\overline{S}\subset M$ be a link,
put $S=f_{\red 0}^{-1}(\overline{S})$, \blue{$S_n=h_n'^{-1}(S)$, and let $\wt{S}$ denote} \bred{the inverse system $\{S_n\}_n$.}
Since the direct limit functor is exact, every exact sequence in Section 2
holds for $\wt{N\,}\!$ and $(\wt{N\,}\!,\wt{S})$, and
the following proposition is obtained.
\begin{prop}\label{ZpTate} Let the situation be as above. Then \magenta{for each $i\in \bb{Z}$},\\
(1) \blue{The} equation $\wh{H}^i(G,C_2(\wt{N\,}\!))=0$ holds.\\
(2) If $\wt{M}$ and $\wt{N\,}\!$ consist of $\bb{Q}$HS$^3$\blue{'s}, then $\wh{H}^i(G,Z_2(\wt{N\,}\!))\cong \wh{H}^{i+1}(G,\bb{Z})$.
If $G\cong \bb{Z}/p\bb{Z}$ in addition, then $\hbar_i:=\mathop{\mathrm{rank}}\nolimits \wh{H}^i(G,Z_2(\wt{N\,}\!))$ satisf\blue{ies} $\hbar_1=1,\hbar_2=0$.\\
(3) If $G\cong \bb{Z}/p\bb{Z}$, then $\wh{H}^i(G,Z_1(\wt{N\,}\!)_\wt{S})=0$.\\
(4) Suppose that $G\cong \bb{Z}/p\bb{Z}$,
\violet{$S_n$'s consist} of \blue{$\bb{Q}$}-null-homologous components, $\overline{S}$ is properly branched in $f_{\red 0}$, and $\overline{S}$ is \blue{infinitely inert (equivalent to say, unbranched and finitely decomposed)}
in $\wt{M}$.
Then $S$ is also \blue{infinitely inert}
\blue{in} $\wt{N\,}\!$.
Let $s$ denote the number of \blue{connected} component of \bred{$\mca{S}:=\varprojlim S_n$},
which is equal to that of $S_n$ for $n\gg 0$.
Then there is an isomorphism $Z_2(\wt{N\,}\!)_\wt{S}/Z_2(\wt{N\,}\!)\cong \blue{B_1(\wt{N\,}\!)\cap }Z_1(\wt{S})$ \blue{for the subgroup} $\blue{B_1(\wt{N\,}\!)\cap }Z_1(\wt{S})< Z_1(\wt{S})\cong \bb{Z}^s$
of finite index, and satisfies
$\wh{H}^1(G,Z_2(\wt{N\,}\!)_\wt{S}/Z_2(\wt{N\,}\!))=0$, $\wh{H}^2(G,Z_2(\wt{N\,}\!)_\wt{S}/Z_2(\wt{N\,}\!)) \cong (\bb{Z}/p\bb{Z})^s$.
\end{prop}
\begin{proof}
\blue{Let $n\gg0$.}
The transfer map $h'$$_{n,n+1}^!$ is an isomorphism on $H_3(N_n)\cong \bb{Z}$,
and is the multiplication by $p$ on $H_0(N_n)\cong \bb{Z}$ and on $H_0(h_n^{-1}(S))\cong \bb{Z}^s$.
If $S$ is totally inert in $\wt{N\,}\!$, then $h'$$_{n,n+1}^!={\rm id}$ on $Z_1(h'$$_n^{-1}(S))\cong \bb{Z}^s$.
Since ${\rm Nr}=(\text{multiplication by }p)$ is a surjection on a divisible group $\displaystyle \varinjlim_{\times p}\bb{Z}$, \magenta{the assertion} follows from Propositions 2.3, 2.5, and 2.6.
\end{proof}
\section{Kida's formula for extensions of $\bb{Z}_{p}$-fields}
In this section, we review the background of our main result of this paper.
The following is a classical theorem.
\begin{thm}[The Riemann--Hurwitz formula] Let $f:R'\to R$ be an $n$-fold covering of compact, connected Riemann surfaces and let $g$ and $g'$ denote
the genera of $R$ and $R'$ respectively.
the ramification indices $e(P')$ of $P'\in R'$ satisfy
$$2g'-2=(2g-2)n+\sum_{P'\in R'}(e(P')-1).$$
\end{thm}
Y.\ Kida proved a highly interesting analogue of this formula for number fields.
\magenta{For each $k\subset \bb{C}$, we set $k_+=k\cap \bb{R}$.}
\begin{thm}[Kida's formula, \cite{Kida1980}]
Let $p$ be an odd prime,
let $F/k$ be a finite Galois $p$-extension of CM-fields,
and let $F_\infty$ and $k_\infty$ denote the cyclotomic $\bb{Z}_{p}$-extensions of $F$ and $k$ respectively.
If $\mu_{\magenta{k_\infty/k}}=0$, then $\mu_{\magenta{F_\infty/F}}=0$, and
$$\lambda^{-}_{\magenta{F_\infty/F}}-\delta=(F_\infty:k_\infty)(\lambda^{-}_{\magenta{k_\infty/k}}-\delta) +(\sum (e_w-1) -\sum(e_{w_+}-1)),$$
where \magenta{$\lambda^-$ denotes the $\lambda$-invariant of the minus part,}
$\delta$ is 1 or 0 according \magenta{to whether} $k$ contains a primitive $p$-th root of unity or not,
\magenta{$w$ (resp. $w_+$) runs through all the non-$p$ primes of $F_\infty$ (resp. $F_{\infty,+}$), and
$e_w$ (resp. $e_{w_+}$) denotes the ramification index of $w$ (resp. $w_+$) in $F_\infty/k_\infty$ (resp. $F_{\infty,+}/k_{\infty,+}$). }
\end{thm}
Following the method of Chevalley--Weil \cite{CWH1934},
K.\ Iwasawa gave an alternative proof for Kida's formula,
with use of $p$-adic representation theory of finite groups.
He also gave a less explicit formula for a more general situation,
from which his second proof follows:
\begin{thm}[Iwasawa \cite{Iwasawa1981}, a corollary of Theorem 6]
Let $p$ be a prime number,
let \blue{$k_\infty$} be a cyclotomic $\bb{Z}_{p}$-field,
and let \blue{$F_\infty$} be a cyclic extension of degree $p$ over $k_\infty$ unramified at every infinite place of $k_\infty$ \magenta{with $G=\mathop{\mathrm{Gal}}\nolimits(k_\infty/k)$}. Assume that $\mu_{k_\infty}=0$. Then $\mu_{F_\infty}=0$, and
$$\lambda_{F_\infty}=p\lambda_{k_\infty}+\sum_{w}(e_w-1)+(p-1)(\hbar_{2}-\hbar_{1}),$$
where $w$ ranges over all non-$p$ \magenta{primes} on $F_\infty$, $e_w$ denotes the ramification index of $w$ in $F_\infty/k_\infty$, and $\hbar_{i}$ denotes the $p$-rank of the abelian group $H^{i}(G, \O_{F_\infty}^{*})$ for $i=1,2$.
\end{thm}
\magenta{An application of this formula for totally real number fields with some cohomological study of units was given in (\cite{FKOT1997}).}
\section{Kida's formula for branched $\bb{Z}_{p}$-covers
\subsection{Main theorem and example}
\blue{An analogue of the original Kida's formula (Theorem 6.2) is stated as follows:}
\begin{thm}[Kida's formula]\label{Main}
Let $f:\wt{N\,}\!\to \wt{M}$ be
\bred{an equivariant Galois morphism of degree $p$-power} of
branched $\bb{Z}_{p}$-covers of $\bb{Q}$HS$^3$.
Let $\overline{S}\subset M$ denote the branch link of $f_{\red 0}:N\to M$
\blue{and put $S=f_{\red 0}^{-1}(\overline{S})$.}
\blue{If $\overline{S}$ is infinitely inert in $\wt{M}$, then so is $S$ in $\wt{N\,}\!$, and $S_n\blue{:=h_n'^{-1}(S)}$ is the branch link of $f_n$.}
If in addition $\mu_{\wt{M}}=0$, then $\mu_{\wt{N\,}\!}=0$.
\bred{Suppose that any component of $\overline{S}$ is not inert in $f_0$.}
\blue{For each component $w=(w_n)_n$ of \blue{$\mca{S}:=\varprojlim_n S_n$}, let $e_w$ denote the branch index of $w$ defined as that of $w_n$ in $f_n$ for $n\gg 0$.} Then
$$\lambda_{\wt{N\,}\!}-1=\deg(f)(\lambda_{\wt{M}}-1)+\sum_{w\subset \blue{\mca{S}}}(e_w-1).$$
\end{thm}
\blue{By Kida's method in \cite{Kida1980}, this formula is deduced from the case of degree $p$:}
\begin{lem}[The case of degree $p$] \label{deg=p}
Let $f:\wt{N\,}\!\to \wt{M}$ be a branched cover of degree $p$ of branched $\bb{Z}_{p}$-covers of $\bb{Q}$HS$^3$,
let $\overline{S}\subset M$ be a link containing the branch link of $f_{\red 0}:N\to M$,
\blue{put $S=f_{\red 0}^{-1}(\overline{S})$, and let $\mca{S}$ denote the inverse limit of the preimages of $\overline{S}$ in $\wt{N\,}\!$.}
If $\overline{S}$ is \blue{infinitely}
inert in $\wt{M}$, then $S$ is \blue{infinitely}
inert in $\wt{N\,}\!$.
If in addition $\mu_{\wt{M}}=0$, then $\mu_{\wt{N\,}\!}=0$.
\bred{Suppose that any component of $\overline{S}$ is not inert in $f_0$.}
\blue{Let} $e_w$ denote the branch index for each component $w$ of \blue{$\mca{S}$}. Then
$$\lambda_{\wt{N\,}\!}-1=p(\lambda_{\wt{M}}-1)+\sum_{w\subset \blue{\mca{S}}}(e_w-1).$$
\end{lem}
\blue{We give their proofs in the subsequent subsections.}
\begin{rem}
(1) As an intermediate result, we \blue{prove}
the following:
\blue{Let the setting be as in the Lemma above.}
If we take CW-structures or PL-structures compatible with all the covering maps \blue{and} put $\hbar_i\blue{=}\mathop{\mathrm{rank}}\nolimits \wh{H}^i(G,Z_2(\wt{N\,}\!))$ \magenta{for $G=\mathop{\mathrm{Gal}}\nolimits(f)$},
then
$$\lambda_{\wt{N\,}\!}=p\lambda_{\wt{M}}+\sum_{w\subset \blue{\mca{S}}}(e_w-1)+(p-1)(\hbar_2-\hbar_1).$$
This formula is a more direct analogue of Theorem 6.3 (a corollary of Iwasawa \cite{Iwasawa1981}, Theorem 6).
\noindent (2) If we fix a $\bb{Z}$-basis of $Z_1(\wt{N\,}\!)$ containing \bred{the components of $\mca{S}$} and take the sum over it, then $w$\bred{'s} can be seen as analogues of places.
\noindent
\blue{(3)} The original proof by Kida (\cite{Kida1980}) was given by genus theory.
\magenta{We expect an alternative proof of our formula with use of genus theory for 3-manifolds, which was established in \cite{Ueki3}.
There are also analytic proofs with use of $p$-adic $L$-functions by Grass and Sinnott (\cite{Gras1978}, \cite{Sinnott1984}). It seems interesting to consider their analogues, examine analogues of $p$-adic $L$-functions,
and explore an analogue of the Iwasawa main conjecture.}
\end{rem}
Here is an example:
\begin{exa}
Let $p=3$, let $N=M=S^3$, and let $f_{\red 0}:N\to M$ be a branched cover of degree $p$ branched over an unknot $\overline{S}$.
Let $L=K\cup K'$ be a \magenta{Hopf} link as in the figure below, so that $\mathop{\mathrm{lk}}\nolimits(K,\overline{S})=3, \mathop{\mathrm{lk}}\nolimits(K',\overline{S})=\mathop{\mathrm{lk}}\nolimits(K,K')=1$.
Then, $\wt{K}:=\red{f_0}^{-1}(K)=K_{1}\sqcup K_{2}\sqcup K_{3}$ is a Borromean ring, $\wt{K'}:=\red{f_0}^{-1}(K')$ is an unknot, and $\mathop{\mathrm{lk}}\nolimits(K_{i},\wt{K'})=1$ for $i=1,2,3$.
Since $L$ and $L'=f_{\red 0}^{-1}(L)$ consist of null-homologous components,
\magenta{the} TLN-$\bb{Z}_{p}$-cover\magenta{s} $\wt{M}$ and $\wt{N\,}\!$ are defined.
Since $\mathop{\mathrm{lk}}\nolimits(L,\overline{S})=\mathop{\mathrm{lk}}\nolimits(L', S)=4 \not \equiv 0 \mathop{\mathrm{mod}}\nolimits 3$,
$\overline{S}$ and $S:=f_{\red 0}^{-1}(\overline{S})$ are totally inert in $\wt{M}$ and $\wt{N\,}\!$ respectively.
Since the defining homomorphism $\tau,\tau'$ commute with
$f_{\red 0*}:\pi_1(N-L'\cup S)\to \pi_1(M-L\cup \overline{S})$ and $\iota={\rm id}_{\bb{Z}_{p}}$, a branched cover \magenta{$f:\wt{N\,}\!\to \wt{M}$} of branched $\bb{Z}_{p}$-covers of degree $p$ is defined.
\magenta{By Hosokawa's result (\cite{Hosokawa1958}), Theorem 1),
the value of Hosokawa polynomials $H_{L}(t)$ and $H_{L'}(t)$ at $t=1$ are
calculated by using the linking numbers of components of $L$ and $L'$,} and satisfy $H_{L}(1)=H_{L'}(1)=\pm1$.
\magenta{By a result of Kadokami--Mizusawa (\cite{KM2008}, Proposition 4.1), the fact $p \not | \pm1$ implies that ${\wt{M}}$ and $\wt{N\,}\!$ consist of $\bb{Q}$HS$^3$'s, and their Iwasawa invariants satisfy }
$\lambda_{\wt{M}}=2-1=1,$ $\mu_{\wt{M}}=\nu_{\wt{M}}=0$, $\lambda_{\wt{N\,}\!}=4-1=3,$ and $\mu_{\wt{N\,}\!}=\nu_{\wt{N\,}\!}=0$.
\magenta{Thus we have observed that Kida's formula holds for this case:
$$3-1=3\cdot (1-1)+(3-1).$$}
\begin{center} \includegraphics[width=13cm]{kida.eps
\end{center}
\magenta{
Let $p$ be an arbitrary prime number. In the example above, if we replace $p=3$ by $p^r$ for any $r\in \bb{N}$ and take a similar $K$ with ${\rm lk}(K,S)=p^r$, then we obtain a $\bb{Z}_{p}$-cover $\wt{N\,}\!$ with $\lambda_{\wt{N\,}\!}=p^r$ by Kida's formula (Theorem \ref{Main}).
Indeed, we can verify $H_{L'}(1)=\pm1$ in a similar way.}
\end{exa}
\subsection{Proof for degree $p$ (after Iwasawa)}
In this subsection, following Iwasawa's second proof in \cite{Iwasawa1981}, we prove our formula for the case of degree $p$ (Lemma 7.2).
\begin{proof}[Lemma \ref{deg=p}]
The assertion on the $\mu$-invariants is done by Theorem \ref{mu}.
Let $\wt{S}=$ $\(h_n'^{-1}(S)\)_n$ denote the inverse system over $S$ in $\wt{N\,}\!$.
For $\wt{S}$-chains and others, we use the notation in Subsection 5.4.
We have $\wh{H}^n(H_1(\wt{N\,}\!)_{\wt{S}[p]})\cong \wh{H}^n(Z_2(\wt{N\,}\!)_\wt{S})$.
Indeed, consider \blue{the} exact sequences
$0\to B_1(\wt{N\,}\!)_\wt{S}$ $\to Z_1(\wt{N\,}\!)_\wt{S}\to H_1(\wt{N\,}\!)_\wt{S}\to 0$ and
$0\to Z_2(\wt{N\,}\!)_\wt{S}\to C_2(\wt{M})\to B_1(\wt{N\,}\!)_\wt{S}\to0$, and note $G\cong \bb{Z}/p\bb{Z}$. Then \magenta{Proposition} \ref{ZpTate} (1) and (3) show
$\wh{H}^n(H_1(\wt{N\,}\!)_{\wt{S}[p]})\cong \wh{H}^n(H_1(\wt{N\,}\!)_\wt{S})\overset{\cong}{\to} \wh{H}^{n+1}(B_1(\wt{N\,}\!)_\wt{S})
\overset{\cong}{\to} \wh{H}^{n+2}(Z_2(\wt{N\,}\!)_\wt{S}) \cong \wh{H}^{n}(Z_2(\wt{N\,}\!)_\wt{S})$.
Put $A_\wt{S}:=H_1(\wt{N\,}\!)_{\wt{S}[p]}\blue{=H_1(\wt{N\,}\!)/\iota_*(H_1(\wt{\wt{S}}))}$, $A:=H_1(\wt{N\,}\!)_{[p]}$.
Since $\mu_{\wt{N\,}\!}=0$,
Proposition \ref{injlim} gives an isomorphism $A\cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{N\,}\!}}$.
\blue{Since $\wt{N\,}$ consists of $\bb{Q}$HS$^3$, $S_n$ consists of $\bb{Q}$-null-homologous components for each $n$. If $S_n$ is inert in $N_{n+1}\to N_n$, then the restriction of the transfer $H_1(N_n)\to H_1(N_{n+1})$ to the subgroups generated by all the components of $S_n$ and $S_{n+1}$ is surjective, and hence $\iota_*(H_1(\wt{S}))$ is finite.
Hence there is a surjective homomorphism $A\twoheadrightarrow A_\wt{S}$ with finite kernel, and
we have $A_\wt{S}\cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{N\,}\!}}$.}
We consider a linear representation of $G=\bb{Z}/p\bb{Z}$.
Put $X_p:=\bb{Z}_{p}[G], Z_{p-1}:=(1-\sigma)X_p, X_1:=X_p/X_{p-1}$ $=\bb{Z}_{p}, A_i=\mathop{\mathrm{Hom}}\nolimits_{\bb{Z}_{p}}(X_i,\bb{Q}_{p}/\bb{Z}_{p})$ for $i=1,p-1,p$.
Then there are decompositions $A=A_1^{a_1}\oplus A_{p-1}^{a_{p-1}}\oplus A_p^{a_p}$, $a_i\in \bb{N}$.
By using a functor $V:A\mapsto \mathop{\mathrm{Hom}}\nolimits(A,\bb{Q}_{p}/\bb{Z}_{p})$, we put $V_i:=V(A_i), \pi_i:G\to \mathop{\mathrm{GL}}\nolimits(V_i)$, $i=1,p-1,p$.
Then $\pi_1$ is the trivial representation, $\pi_{p-1}$ is the unique faithful irreducible representation, and $\pi_p=\pi_1\oplus \pi_{p-1}=\pi_G$.
Put $V=V(A)$. Then $G\curvearrowright A$ induces a representation $\pi_{f}:G\to \mathop{\mathrm{GL}}\nolimits(V)$,
and there is a decomposition $\pi_{f}=a_1\pi_1\oplus a_{p-1}\pi_{p-1}\oplus a_p\pi_p$.
Next, we calculate $a_i$'s. Since $G\cong \bb{Z}/p\bb{Z}$,
the Tate cohomologies of $G\curvearrowright A_i$ are
abelian groups of exponent $p$, and
their $p$-ranks satisfy $r(\wh{H}^1(G,A_i))=1,0,0, r(\wh{H}^2(G,A_i))=0,1,0$ for $i=1,p-1,p$.
Put
$r_n=r(\wh{H}^n(G,A))=r(\wh{H}^n(G,$ $Z_2(\wt{N\,}\!)_\wt{S}))$.
Then the decomposition of $A_\wt{S}$ yields $r_1=a_1, r_2=a_{p-1}$.
We calculate the Herbrand quotient of $\wt{S}$-2-cycles.
(Recall that for a cyclic group $G$ and $G$-module $B$,
$q(B)=\# \wh{H}^0(B)/\# \wh{H}^1(B)$ is called the Herbrand quotient.
Herbrand's lemma asserts that in a short exact sequence,
if $q(\, )$ is defined for two terms, then so for the rest term, and its Tate cohomologies are finite.)
Now $\wh{H}^n(G,Z_2(\wt{N\,}\!)_\wt{S})$ is finite
and $q(Z_2(\wt{N\,}\!)_\wt{S})$ is defined.
Put $\hbar_n=r(\wh{H}^n(G,Z_2(\wt{N\,}\!)))$
for $Z_2(\wt{N\,}\!)<Z_2(\wt{N\,}\!)_\wt{S}$.
Then $q(Z_2(\wt{N\,}\!)_\wt{S}/Z_2(\wt{N\,}\!))=p^s$ by
\magenta{Proposition} \ref{ZpTate} (4),
and $q(Z_2(\wt{N\,}\!)_\wt{S})$ $=q(Z_2(\wt{N\,}\!))p^s$ by Herbrand's lemma.
Hence $\hbar_1, \hbar_2$ are finite and $a_{p-1}-a_1=\hbar_2-\hbar_1+s$ holds.
Since $\mu_{\wt{M}}=0$, there is an isomorphism $\bar{A}_\wt{S}:=H_1(\wt{M})_{\overline{\wt{S}}} \cong (\bb{Q}_{p}/\bb{Z}_{p})^{\lambda_{\wt{M}}}$ by Proposition \ref{injlim}.
\magenta{Proposition} \ref{ZpTate} (1) asserts $\wh{H}^2(C_2(\wt{N\,}\!))=0$, and hence $\wh{H}^1(Z_2(\wt{N\,}\!)_\wt{S})\cong (B_1(\wt{N\,}\!)_\wt{S})^G/B_1(\wt{M})_\wt{S}$,
$\wh{H}^2(Z_2(\wt{N\,}\!)_\wt{S})\cong \mathop{\mathrm{Coker}}\nolimits((Z_1(\wt{N\,}\!)_\wt{S})^G \to (H_1(\wt{N\,}\!)_\wt{S})^G)$.
Since the first term is finite, maps $H_1(\wt{M})_{\wt{\overline{S}}}$, $\to H_1(\wt{N\,}\!)_\wt{S}$
and $\bar{A_\wt{S}}\to A_\wt{S}^G$ have finite kernels and cokernels, \bred{where we put $\wt{\overline{S}}=\(h_n^{-1}(\overline{S})\)_n$. }
Hence the decomposition of $A_\wt{S}$ tells $\lambda_{\wt{M}}=a_1+a_p$.
Thereby we have obtained $\pi_{f}=a_1\pi_1\oplus a_{p-1}\pi_{p-1}\oplus a_p\pi_p
=(a_1+a_p)\pi_p \oplus (a_{p-1}-a_1)\pi_{p-1}
=\lambda_K \pi_G \oplus \blue{s}\pi_{p-1}\oplus (\hbar_2-\hbar_1)\pi_{p-1}$.
(If $\hbar_2-\hbar_1<0$, the right hand side of the formula should be interpreted as a difference of two representations.)
Comparing the degrees of both side, we obtain
$$\lambda_{\wt{N\,}\!}=p\lambda_{\wt{M}}+\sum_{w\subset \bred{\mca{S}}}(e_w-1)+(p-1)(\hbar_2-\hbar_1).$$
By \magenta{Proposition} \ref{ZpTate} (2), \blue{we have} $\hbar_2-\hbar_1=-1$, \magenta{and obtain the desired formula.
\end{proof}
\subsection{Proof for degree $p$-power (after Kida)}
Following Kida's argument in \cite{Kida1980}, we deduce our formula (Theorem 7.1) from the case of degree $p$ (Lemma 7.2).
\begin{proof}[Theorem \ref{Main}]
Since every nontrivial finite $p$-group has nontrivial center,
a Galois branched cover of $p$-power degree can be realized
as a sequence of branched covers of degree $p$.
Suppose $\deg(f)=p^n$.
We prove the formula by induction on $m$.
The assertion is trivial for $n=0$.
Assume the assertion is true for $n\leq m$. Then for $n=m+1$,
there is a subcover $\wt{N\,}\!\to \wt{H\,}\!$ of $\wt{N\,}\!\to \wt{M}$ of degree $p$.
Let $w$ run through all the components of \blue{$\wt{S}$}, and $u$ run through the image of \blue{$\wt{S}$} in $\wt{H\,}\!$. By the hypothesis of the induction and by Lemma 7.2, they satisfy
$\lambda_{\wt{N\,}\!}-1=p(\lambda_{\wt{H\,}\!} -1)+\sum_w (e(w/u)-1)$ and
$\lambda_{\wt{H\,}\!}-1=p^m(\lambda_{\wt{M}} -1)+\sum_u (e(u/v)-1)$.
Here $u$ and $v$ denote the images of each $w$ in $\wt{H\,}\!$ and $\wt{M}$ respectively, $e(w/u)$ and $e(u/v)$ denote their branch indices in $\wt{N\,}\!\to \wt{H\,}\!$ and $\wt{H\,}\!\to \wt{M}$ respectively, and $e(w/v)=e_w$.
Now we have
$\lambda_{\wt{N\,}\!} -1=p(p^m(\lambda_{\wt{M}} -1)+\sum_u (e(u/v)-1))+\sum_w (e(w/u)-1)
=p^{m+1}(\lambda_{\wt{M}} -1)+\sum_u p(e(u/v)-1) +\sum_w (e(w/u)-1)$.
If $u$ is branched in ${\wt{N\,}\!}\to {\wt{H\,}\!}$,
then the unique $w$ over $u$ satisfies $e(w/v)=pe(u/v)$, $e(w/u)=p$, and hence
$p(e(u/v)-1) + (e(w/u)-1)= e(w/v)-p+p-1=e(w/v)-1$.
If $u$ is decomposed in ${\wt{N\,}\!}\to {\wt{H\,}\!}$, then
the correction $w_1,...,w_p$ of components over $u$ satisfy $e(w_i/v)=e(u/v)$, $e(w_i/u)=1$, and hence
$p(e(u/v)-1) +\sum_i (e(w_i/u)-1)=\sum_i (e(w_i/v)-1)$.
Thus, due to the assumption that $S$ is not inert in $f_{\red 0}:N\to M$, we have generalized the formula to the case of $p$-power degree.
(If $u$ is inert in ${\wt{N\,}\!}\to{\wt{H\,}\!}$, then $w$ over $u$ satisfies
$e(w/v)=e(u/v)$, $e(w/u)=1$ and hence $p(e(u/v)-1)+ (e(w/u)-1)=p e(w/u)-p+1-1=p(e(w/u)-1)$.)
\end{proof}
\section*{Acknowledgments}
The author would like to appreciate Yasushi Mizusawa for informing a paper on Kida's formula \magenta{(\cite{FKOT1997})},
Teruhisa Kadokami for discussion on significance of the notion of a branched $\bb{Z}_{p}$-cover,
Mikio Furuta for instructive advise from topological point of view,
Masanori Morishita for strong \magenta{encouragement} and pointing out errors,
and the anonymous referees for accurate comments.
He is also very grateful to \blue{Ryo Furukawa}, Yu-ich Hirano, Tomoki Mihara, \magenta{Takayuki Morisawa, Hirofumi Niibo, Takayuki Okuda}, and Tatsuro Shimizu for fruitful discussions.
The author is partially supported by Grant-in-Aid for JSPS Fellows (25-2241).
\bibliographystyle{amsalpha
|
\section*{Introduction}
Forecasting changes in volatility is essential for risk management, asset pricing and scenario analysis. Indeed, models for describing and forecasting the evolution of volatility and covariance among financial assets are widely applied in industry \cite{risk_forecasting_book,multivariate_models,bouchaud_book,preis_stress}. Among the most popular approaches are worth mentioning the multivariate extensions of GARCH \cite{multigarch_survey}, the
stochastic covariance models \cite{stochastic_volatility} and realized covariance \cite{realized_volatility}.
However most of these econometrics tools are not able to cope with more than few assets, due to the curse of dimensionality and the increase in the number of parameters \cite{risk_forecasting_book}, limiting their insight into the volatility evolution to baskets of few assets only. This is unfortunate, since gathering insights into systemic risk and the unfolding of financial crises require modelling the evolution of entire markets which are composed by large numbers of assets \cite{risk_forecasting_book}.
We suggest to use network filtering \cite{mantegna1,Tumminello05,dynamic_networks_correlation,asset_graphs,clustering_dyn_asset_graph,buccheri,caldarelli_book} as a valuable tool to overcome this limitation.
Correlation-based filtering networks are tools which have been widely applied to filter and reduce the complexity of covariance matrices made of large numbers of assets (of the order of hundreds), representative of entire markets. This strand of research represents an important part of the Econophysics literature and has given important insights for risk management, portfolio optimization and systemic risk regulation \cite{black_monday,cluster_portfolio,bonanno_mst,invest_periph,musmeci_DBHT,musmeci_jntf}.
The volatility of a portfolio depends on the covariance matrix of the corresponding assets \cite{markowitz}.
Therefore, the latter can provide insights into the former. In this work we elaborate on this connection: we show that correlation matrices can be used to predict variations of volatility, once they are analysed through the lens of network filtering. This is quite an innovative use of correlation-based networks, which have been used mostly for descriptive analyses, with the connections with risk forecasting being mostly overlooked. Some works have shown that is possible to use dimensionality reduction techniques, such as spectral methods \cite{rmt_1}, as early-warning signals for systemic risk \cite{pca_systemic_risk,pca_systemic_risk2}: however these approaches, although promising, do not provide proper forecasting tools, as they are affected by high false positive ratios and are not designed to predict a specific quantity.
The approach we propose exploits network filtering to explicitly predict future volatility of markets made of hundreds of stocks. To this end, we introduce a new dynamical measure that quantifies the rate of change in the structure of the market correlation matrix: the ``correlation structure persistence'' $\langle ES \rangle$. This quantity is derived from the structure of network filtering from past correlations.
Then we show how such measure exhibits significant predicting power on the market volatility, providing a tool to forecast it. We assess the reliability of this forecasting through out-of-sample tests on two different equity datasets.
The rest of this paper is structured as follows: we first describe the two datasets we have analysed and we introduce the correlation structure persistence;
then we show how our analyses point out a strong interdependence between correlation structure persistence and future changes in the market volatility; moreover, we describe how this result can be exploited to provide a forecasting tool useful for risk management, by presenting out-of-sample tests and false positive analysis; then we investigate how the forecasting performance changes in time; finally we discuss our findings and their theoretical implications.
\section*{Results}
\subsection*{A measure of correlation structure persistence}
We have analysed two different datasets of equity data. The first set (NYSE dataset) is composed by daily closing prices of $N=342$ US stocks traded in New York Stock Exchange, covering 15 years from 02/01/1997 to 31/12/2012. The second set (LSE dataset) is composed by daily closing prices of $N=214$ UK stocks traded in the London Stock Exchange, covering 13 years from 05/01/2000 to 21/08/2013. All stocks have been continuously traded throughout these periods of time. These two sets of stocks have been chosen in order to provide a significant sample of the different industrial sectors in the respective markets.
For each asset $i$ ($i=1, ... ,N$) we have calculated the corresponding daily log-return $r_i(t)=log(P_i(t))-log(P_i(t-1))$, where $P_i(t)$ is the asset $i$ price at day $t$. The market return $r_M(t)$ is defined as the average of all stocks returns: $r_M(t)=1/N \sum_i r_i(t)$. In order to calculate the correlation between different assets we have then analysed the observations by using $n$ moving time windows, $T_a$ with $a=1,...,n$. Each time window contains $\theta$ observations of log-returns for each asset, totaling to $N \times n$ observations. The shift between adjacent time windows is fixed to $dT = 5$ trading days. We have calculated the correlation
matrix within each time window, $\{ \rho_{ij}(T_a) \}$, by using an exponential smoothing method \cite{exp_smoothing} that allows to assign more weight on recent observations. The smoothing factor of this scheme has been chosen equal to $\theta/3$ according to previously established criteria \cite{exp_smoothing}.
From each correlation matrix $\{ \rho_{ij}(T_a) \}$ we have then computed the corresponding Planar Maximally Filtered Graph (PMFG) \cite{PMFG2}. The PMFG is a sparse network representation of the correlation matrix that retains only a subset of most significant entries, selected through the topological criterion of being maximally planar \cite{Tumminello05}. Such networks serve as filtering method and have been shown to provide a deep insight into the dependence structure of financial assets \cite{Tumminello05,dynamic_networks_correlation,NJP10}.
Once the $n$ PMFGs, $G(T_a)$ with $a=1,...,n$, have been computed we have calculated two measures, a backward-looking and a forward-looking one. The first is a measure that monitors the
correlation structure persistence, based on a measure of PMFG similarity.
This backward-looking measure, that we call $\langle ES \rangle (T_a)$, relies on past data only and indicates how slowly the correlation structure measured at time window $T_a$ is differentiating from structures associated to previous time windows. The forward-looking measure is the volatility ratio $q(T_a)$ \cite{kondor,kondor2}, that at each time window quantifies how good the market volatility measured at $T_a$ is as a proxy for the next time-window volatility. Unlike $\langle ES \rangle (T_a)$, the value of $q(T_a)$ is not known at the end of $T_a$. Fig. \ref{fig:time_windows} shows a graphical representation of the time window set-up. In the following we define the two measures:
\label{sec:methods}
\begin{figure}[t!]
\includegraphics[trim = 5mm 10mm 20mm 65mm, clip, scale=0.6]{figure/time_windows2.pdf}
\caption{\label{fig:time_windows} {\bf Scheme of time windows setting for $\langle ES \rangle (T_a)$ and $q(T_a)$ calculation}. $T_a$ is a window of length $\theta$. The correlation structure persistence $\langle ES \rangle(T_a)$ (upper axis) is computed by using data in $T_a$ and in the first $L$ time windows before $T_a$. The volatility ratio $q(T_a)$ is computed by using data in $T_a$ and in the future time window $T_a^{forward}$. In the upper axis the time windows are actually overlapping, but they are here represented as disjoint for the sake of simplicity.}
\end{figure}
\begin{itemize}
\item {\bf Correlation structure persistence $\langle ES \rangle (T_a)$}: we define the correlation structure persistence at time $T_a$ as:
\begin{equation}
\langle ES \rangle (T_a) = \sum_{b = a -L }^{a - 1} \omega(T_b) ES(T_a, T_b) ,
\label{eq:es}
\end{equation}
where $\omega(T_b) = \omega_0 \exp(\frac{b-a-1}{L/3})$ is an exponential smoothing factor, $L$ is a parameter and $ES(T_a, T_b)$ is the fraction of edges in common between the two PMFGs $G(T_a)$ and $G(T_b)$, called ``edge survival ratio'' \cite{black_monday}. In formula, $ES(T_a, T_b)$ reads:
\begin{equation}\label{eq:es_ab}
ES(T_a, T_b) = \frac{1}{N_{edges}} \mid E^{T_a} \cap E^{T_b} \mid ,
\end{equation}
where $N_{edges}$ is the number of edges (links) in the two PMFGs (constant and equal to $3N-6$ for a PMFG \cite{PMFG2}), and $E^{T_{a}}$ ($E^{T_{b}}$) represents the edge-sets of PMFG at $T_{a}$ ($T_{b}$). The correlation structure persistence $\langle ES \rangle (T_a)$ is therefore a weighted average of the similarity (as measured by the edge survival ratio) between $G(T_a)$ and the first $L$ previous PMFGs, with an exponential smoothing scheme that gives more weight to those PMFGs that are closer to $T_a$. The parameter $\omega_0$ in Eq. \ref{eq:es} can be calculated by imposing $\sum_{b = a -L }^{a - 1} \omega(T_b) = 1$. Intuitively, $\langle ES \rangle (T_a)$ measures how slowly
the change of correlation structure is occurring in the near past of $T_a$.
\item {\bf Volatility ratio $q(T_a)$} \cite{kondor}:
In order to quantify the agreement between the estimated and the realized risk we here make use of the volatility ratio, a measure which has been used \cite{kondor,non_stationary_corr} for this purpose and defined as follows:
\begin{equation}
\label{eq:q}
q(T_a) = \frac{\sigma(T^{forward}_a)}{\sigma(T_a)} ,
\end{equation}
where $\sigma(T^{forward}_a)$ is the realized volatility of the average market return $r_M(t)$ computed on the time window $T^{forward}_a$; $\sigma(T_a)$ is the estimated volatility of $r_M(t)$ computed on time window $T_a$, by using the same exponential smoothing scheme \cite{exp_smoothing} described for the correlation $\{ \rho_{ij}(T_a) \}$.
Specifically, $T^{forward}_a$ is the time window of length $\theta_{forward}$ that follows immediately $T_a$: if $t_{\theta}$ is the last observation in $T_a$, $T^{forward}_a$ covers observations from $t_{\theta+1}$ to $t_{\theta+1+\theta_{forward}}$ (Fig. \ref{fig:time_windows}). Therefore the ratio in Eq. \ref{eq:q} estimates the agreement between the market volatility estimated with observations in $T_a$ and the actual market volatility observed over an investment in the $N$ assets over $T^{forward}_a$. If $q(T_a)>1$, then the historical data gathered at $T_a$ has underestimated the (future) realized volatilty, whereas $q(T_a)<1$ indicates overestimation.
Let us stress that $q(T_a)$ provides an information on the reliability of the covariance estimation too, given the relation between market return volatility and covariance \cite{markowitz}:
\begin{equation}
\sigma(T_a) = \sqrt{\frac{1}{N^2} \sum_{ij} Cov_{ij}(T_a)} ,
\end{equation}
\begin{equation}
\sigma(T^{forward}_a) = \sqrt{\frac{1}{N^2} \sum_{ij} Cov_{ij}(T^{forward}_a)} ,
\end{equation}
where $Cov_{ij}(T_a)$ and $Cov_{ij}(T^{forward}_a)$ are respectively the estimated and realized covariances.
\end{itemize}
\subsection*{Interplay between correlation structure persistence and volatility ratio}
To investigate the relation between $\langle ES \rangle (T_a)$ and $q(T_a)$ we have calculated the two quantities with different values of $\theta$ and $L$ in Eqs. \ref{eq:es} and \ref{eq:q}, to assess the robustness against these parameters. Specifically, we have used $\theta \in (250, 500, 750, 1000)$ trading days, that correspond to time windows of length 1, 2, 3 and 4 years respectively; $L \in (10, 25, 50, 100)$, that correspond (given $dT = 5$ trading days) to an average in Eq. \ref{eq:es} reaching back to $50$, $125$, $250$ and $500$ trading days respectively. $\theta_{forward}$ has been chosen equal to $250$ trading days (one year) for all the analysis.
In Fig. \ref{fig:es_matrices} we show the $ES(T_a,T_b)$ matrices (Eq. \ref{eq:es_ab})
for the NYSE and LSE dataset, for $\theta=1000$. We can observe a block structure with periods of high structural persistence and other periods whose correlation structure is changing faster. In particular two main blocks of high persistence can be found before and after the 2007-2008 financial crisis; a similar result was found in a previous work \cite{musmeci_jntf} with a different measure of similarity. These results are confirmed for all values of $\theta$ considered.
In Fig. \ref{fig:plot_342} we show $\langle ES \rangle (T_a)$ and $q(T_a)$ as a function of time, for $\theta=1000$ and $L=100$.
As expected, main peaks of $q(T_a)$ occur during the months before the most turbulent periods in the stock market, namely the 2002 market downturn
and the 2007-08 credit crisis. Interestingly, the corresponding $\langle ES \rangle (T_a)$ seems to follow a specular trend. This is confirmed by explicit calculation of Pearson correlation between the two signals, reported in Tabs. \ref{tab:corr_4} - \ref{tab:corr_8}: as one can see, for all combinations of parameters the correlation is negative.
\begin{figure}[t!]
\includegraphics[scale=0.4]{figure/es_matrix_theta1000.pdf}
\hspace{-3em}
\includegraphics[scale=0.4]{figure/LSE/es_matrix_theta1000_LSE.pdf}
\caption{\label{fig:es_matrices} {\bf $ES(T_a, T_b)$ matrices for $\theta = 1000$, for NYSE (left) and LSE dataset (right)}. A block-like structure can be observed in both datasets, with periods of high structural persistence and other periods whose correlation structure is changing faster. The 2007-2008 financial crisis marks a transition between two main blocks of high structural persistence.}
\end{figure}
\begin{figure}[t!]
\includegraphics[scale=0.5]{figure/plot_es_vs_q_NYSE.pdf}
\includegraphics[scale=0.5]{figure/LSE/plot_es_vs_q_LSE.pdf}
\caption{\label{fig:plot_342} {\bf $\langle ES \rangle (T_a)$ and $q(T_a)$ signals represented for $\theta=1000$ and $L=100$}, for both NYSE (left graph) and LSE (right graph) datasets. It is evident the anticorrelation between the two signals. The financial crisis triggers a major drop in the structural persistence and a corresponding peak in $q(T_a)$.
}
\end{figure}
\begin{figure}[t!]
\includegraphics[scale=0.5]{figure/meta_matrix_theta1000.pdf}
\includegraphics[scale=0.5]{figure/LSE/meta_matrix_theta1000_lse.pdf}
\caption{\label{fig:meta_matrices} {\bf $z(T_a, T_b)$ matrices for $\theta = 1000$, for NYSE (left) and LSE dataset (right)}. A block-like structure can be observed in both datasets, with periods of high structural persistence and other periods whose correlation structure is changing faster. The blocks of high similarity show higher compactness than in Fig. \ref{fig:es_matrices}.}
\end{figure}
In order to check the significance of this anticorrelation we cannot rely on standard tests on Pearson coefficient, such as Fisher transform \cite{fisher_transform}, as they assume i.i.d. series \cite{math_statistics}. Our time series are instead strongly autocorrelated, due to the overlapping between adjacent time windows. Therefore we have calculated confidence intervals by performing a block bootstrapping test \cite{block_bootstrapping}. This is a variation of the bootstrapping test \cite{bootstrapping}, conceived to take into account the autocorrelation structure of the bootstrapped series. The only free parameter in this method is the block length, that we have chosen applying the optimal selection criterion proposed in literature \cite{optimal_block_bootstrapping}: such criterion is adaptive on the autocorrelation strength of the series as measured by the correlogram. We have found, depending on the parameters $\theta$ and $L$, optimal block lengths ranging from 29 to 37, with a mean of 34 (corresponding to 170 trading days). By performing block bootstrapping tests we have therefore estimated confidence intervals for the true correlation between $\langle ES \rangle (T_a)$ and $q(T_a)$; in Tabs. \ref{tab:corr_4} - \ref{tab:corr_8} correlations whose $95\%$ and $99\%$ confidence intervals (CI) do not include zero are marked with one and two stars respectively. As we can see, 14 out of 16 correlation coefficients are significantly different from zero within $95\%$ CI in the NYSE dataset, and 12 out of 16 in the LSE dataset. For what concerns the $99\%$ CI, we observe 13 out of 16 for the NYSE and 9 out of 16 for the LSE dataset. Non-significant correlations appear only for $\theta = 250$, suggesting that this length is too small to provide a reliable measure of structural persistence. Very similar results are obtained by using Minimum Spanning Tree (MST) \cite{mst_history} instead of PMFG as network filtering.
Given the interpretation of $\langle ES \rangle (T_a)$ and $q(T_a)$ given above, anticorrelation implies that an increase in the ``speed'' of correlation structure evolution (low $\langle ES \rangle (T_a)$) is likely to correspond to underestimation of future market volatility from historical data (high $q(T_a$)), whereas when the structure evolution ``slows down'' (high $\langle ES \rangle (T_a)$) there is indication that historical data is likely to provide an overestimation of future volatility. This means that we can use
$\langle ES \rangle (T_a)$ as a valuable predictor of current historical data reliability. This result is to some extent surprising as $\langle ES \rangle (T_a)$ is derived from PMFGs topology, which in turns depends only on the ranking of correlations and not on their actual value: yet, this information provides meaningful information about the future market volatility and therefore about the future covariance.
In principle other measures of correlation ranking structure, more straightforward than the correlation persistence $\langle ES \rangle (T_a)$, might capture the same interplay with $q(T_a)$. We have therefore considered also the Metacorrelation $z(T_a,T_b)$, that is the Pearson correlation computed between the coefficients of correlation matrices at $T_a$ and $T_b$ (see Methods for more details). Such measure does not make use of PMFG. Fig. \ref{fig:meta_matrices} displays the similarity matrices obtained with this measure for NYSE and LSE datasets: we can observe again block-like structures, that however carry different information from the $ES(T_a,T_b)$ in Fig. \ref{fig:es_matrices}; in particular, blocks show higher intra-similarity and less structure. Similarly to Eq. \ref{eq:es}, we have then defined $z(T_a)$ as the weighted average over $L$ past time windows (see Methods).
In Tabs. \ref{tab:corr_meta_nyse} and \ref{tab:corr_meta_lse} we show the correlation between $z(T_a)$ and $q(T_a)$. As we can see, although an anticorrelation is present for each combination of parameters $\theta$ and $L$, correlation coefficients are systematically closer to zero than in Tabs. \ref{tab:corr_4} - \ref{tab:corr_8}, where $\langle ES \rangle (T_a)$ was used. Moreover the number of significant Pearson coefficients, according to the block bootstrapping, decreases to 12 out of 16 in NYSE and to 10 out of 16 in LSE dataset. Since $\langle z \rangle (T_a)$ does not make use of PMFG, this result suggests that the filtering procedure associated to correlation-based networks is a necessary step for capturing at best the correlation ranking evolution and its interplay with the volatility ratio.
\begin{table}[h]\centering
\caption{\label{tab:corr_4} {\bf NYSE dataset: correlation between $\langle ES \rangle (T_a)$ and $q(T_a)$}, for different combinations of parameters $\theta$ and $L$. Stars mark those correlation coefficients whose confidence interval excludes zero with a $95\%$ (one star) or a $99\%$ confidence (two stars). The confidence intervals are computed from the block-bootstrapped sample.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & -0.2129 & -0.2224 & $-0.2997^{*}$ & $-0.3498^{**}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & $-0.4276^{**}$ & $-0.4683^{**}$ & $-0.4945^{**}$ & $-0.5354^{**}$ \\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & $-0.4994^{**}$ & $-0.5499^{**}$ & $-0.5837^{**}$ & $-0.6018^{**}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & $-0.5789^{**}$ & $-0.6152^{**}$ & $-0.6480^{**}$ & $-0.6874^{**}$ \\
\hline
\addlinespace[1ex]
\multicolumn{6}{l}{\textsuperscript{**}$p<0.001$, \textsuperscript{*}$p<0.01$,}
\end{tabular}
\end{table}
\begin{table}[h]\centering
\caption{\label{tab:corr_8} {\bf LSE dataset: correlation between $\langle ES \rangle (T_a)$ and $q(T_a)$}, for different combinations of parameters $\theta$ and $L$. Stars mark those correlation coefficients whose confidence interval excludes zero with a $95\%$ (one star) or a $99\%$ confidence (two stars). The confidence intervals are computed from the block-bootstrapped sample.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & $-0.2084^{*}$ & $-0.1887^{*}$ & -0.1872 & $-0.2269^{*}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & $-0.3083^{**}$ & $-0.3343^{**}$ & $-0.3782^{**}$ & $-0.4202^{**}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & $-0.4050^{**}$ & $-0.4409^{**}$ & $-0.4334^{**}$ & $-0.4374^{**}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & $-0.4552^{**}$ & $-0.5285^{**}$ & $-0.5480^{**}$ & $-0.5227^{**}$\\
\hline
\addlinespace[1ex]
\multicolumn{6}{l}{\textsuperscript{**}$p<0.001$, \textsuperscript{*}$p<0.01$,}
\end{tabular}
\end{table}
\begin{table}[h]\centering
\caption{\label{tab:corr_meta_nyse} {\bf NYSE dataset: correlation between $\langle z \rangle (T_a)$ and $q(T_a)$}, for different combinations of parameters $\theta$ and $L$. Stars mark those correlation coefficients whose confidence interval excludes zero with a $95\%$ (one star) or a $99\%$ confidence (two stars). The confidence intervals are computed from the block-bootstrapped sample.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & -0.0992 & -0.0754 & -0.1055 & -0.1157\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & -0.2146 & -0.2232 & -0.2309 & -0.2753\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & -0.2997 & $-0.3706^{*}$ & $-0.4030^{*}$ & $-0.4109^{*}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & $-0.3933^{**}$ & $-0.4290^{**}$ & $-0.4678^{**}$ & $-0.4574^{*}$\\
\hline
\addlinespace[1ex]
\multicolumn{6}{l}{\textsuperscript{**}$p<0.001$, \textsuperscript{*}$p<0.01$,}
\end{tabular}
\end{table}
\begin{table}[h!]\centering
\caption{\label{tab:corr_meta_lse} {\bf LSE dataset: correlation between $\langle z \rangle (T_a)$ and $q(T_a)$}, for different combinations of parameters $\theta$ and $L$. Stars mark those correlation coefficients whose confidence interval excludes zero with a $95\%$ (one star) or a $99\%$ confidence (two stars). The confidence intervals are computed from the block-bootstrapped sample.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & -0.1470 & -0.1095 & -0.1326 & -0.1720\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & $-0.2365^{*}$ & -0.2113 & $-0.2936^{*}$ & $-0.3932^{**}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & $-0.3123^{**}$ & $-0.3379^{*}$ & $-0.3538^{*}$ & $-0.3851^{*}$\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & $-0.2917^{*}$ & -0.2954 & -0.3163 & $-0.4192^{**}$\\
\hline
\addlinespace[1ex]
\multicolumn{6}{l}{\textsuperscript{**}$p<0.001$, \textsuperscript{*}$p<0.01$,}
\end{tabular}
\end{table}
\subsection*{Forecasting volatility: a new approach}
\label{sec:forecasting_outS}
\begin{figure}[ht!]
\includegraphics[scale=0.4]{figure/out_sample_training_set.pdf}
\includegraphics[scale=0.4]{figure/out_sample_test_set.pdf}
\includegraphics[scale=0.4]{figure/LSE/out_sample_training_set_LSE.pdf}
\includegraphics[scale=0.4]{figure/LSE/out_sample_test_set_LSE.pdf}
\caption{ \label{fig:training_test} {\bf Partition of data into training (left graphs) and test (right graphs) set}. Training sets are used to regress $Y(T_a)$ against
$\langle ES \rangle (T_a)$, in order to estimate the coefficents in the logistic regression and therefore identify the regression threshold, shown as a vertical continuous line. The test sets are used to test the forecasting performance of such regression on a subset of data that has not been used for regression; the model predicts $Y(T_a)=1$ ($q(T_a)>1$) if $\langle ES \rangle (T_a)$ is greater than the regression threshold, and $Y(T_a)=0$ ($q(T_a)<1$) otherwise.}
\end{figure}
In this section we evaluate how well the correlation structure persistence $\langle ES \rangle (T_a)$ can forecast the future through its relation with the forward-looking volatility ratio $q(T_a)$. In particular we focus on estimating whether $q(T_a)$ is greater or less than $1$: this information, although less complete than a
precise estimation of $q(T_a)$, gives us an important insight into possible overestimation ($q(T_a)<1$) or underestimation ($q(T_a)>1$) of future volatility.
We have proceeded as follows. Given a choice of parameters $\theta$ and $L$, we have calculated the corresponding set of pairs $\{ \langle ES \rangle (T_a) , q(T_a) \}$, with $a=1,...,n$.
Then we have defined $Y(T_a)$ as the categorical variable that is $0$ if $q(T_a)<1$ and $1$ if $q(T_a)>1$. Finally we have performed a logistic regression of $Y(T_a)$ against $\langle ES \rangle (T_a)$: namely,
we assume that \cite{statistical_learning_book}:
\begin{equation}\label{eq:logistic_regression}
P\big\{Y(T_a)=1| \langle ES \rangle (T_a)=x\big\} = S\big(\beta_0+\beta_1 x \big) ~~,
\end{equation}
where $S(t)$ is the sigmoid function $S(t)=\frac{1}{1+e^{-t}}$ \cite{logistic_regression_book}; we estimate parameters $\beta_0$ and $\beta_1$ from the observations $\{ \langle ES \rangle (T_a) , q(T_a) \}_{a=1,...,n}$ through Maximum Likelihood \cite{machine_learning_book}.
Once the model has been calibrated, given a new observation $\langle ES \rangle (T_{n+1}) = x$ we have predicted $Y(T_{n+1})=1$ if $P\big\{Y(T_{n+1})=1| \langle ES \rangle (T_{n+1}) = x\big\} > 0.5$, and $Y(T_{n+1})=0$ otherwise. This classification criterion, in a case with only one predictor, corresponds to classify $Y(T_{n+1})$ according to whether $\langle ES \rangle (T_{n+1})$ is greater or less than a threshold $r$ which depends on $\beta_0$ and $\beta_1$, as shown in Fig. \ref{fig:training_test} (right graphs) for a particular choice of parameters. Therefore the problem of predicting whether market volatility will increase or decrease boils down to a classification problem \cite{machine_learning_book} with $\langle ES \rangle (T_a)$ as predictor and $Y(T_a)$ as target variable.
We have made use of a logistic regression because it is more suitable than a polynomial model for dealing with classification problems \cite{statistical_learning_book}. Other classification algorithms are available; we have chosen the logistic regression due to its simplicity. We have also implemented the KNN algorithm \cite{machine_learning_book} and we have found that it provides similar outcomes but worse results in terms of the forecasting performance metrics that we discuss in this section.
We have then evaluated the goodness of the logistic regression at estimating $Y(T_{n+1})$ given a new observation $\langle ES \rangle (T_{n+1})$.
To this end, we have computed three standard metrics for assessing the performance of a classification method: the probability of successful forecasting $P^{+}$, the True Positive Rate $TPR$ and the False Positive Rate $FPR$. $P^{+}$ represents the expected fraction of correct predictions, $TPR$ is the method goodness at identifying true positives (in this case, actual increases in volatility) and $FPR$ quantifies the method tendency to false positives (predictions of volatility increase when the volatility will actually decrease): see Methods for more details.
Overall these metrics provide a complete summary of the model goodness at predicting changes in the market volatility \cite{statistical_learning_book}.
In order to avoid overfitting we have estimated the metrics above by means of an out-of-sample procedure \cite{statistical_learning_book,machine_learning_book}. We have divided our dataset into two periods, a training set and a test set. In the training set we have calibrated the logistic equation in Eq. \ref{eq:logistic_regression}, estimating the parameters $\beta_0$ and $\beta_1$; in the test set we have used the calibrated model to measure the goodness of the model predictions by computing the measures of performance in Eq. \ref{eq:prob_forecast}-\ref{eq:fpr}. In Fig. \ref{fig:training_test} this division is shown for a particular choice of $\theta$ and $L$, for both NYSE and LSE dataset. In this example the percentage of data included in the test set (let us call it $f_{test}$) is $30\%$.
Probabilities of successful forecasting $P^{+}$ are reported in Tabs. \ref{tab:nyse_p_outS} and \ref{tab:lse_p_outS}, for $f_{test}=30\%$.
As we can see $P^{+}$ is higher than $50\%$ for all combinations of parameters in NYSE dataset, and in almost all combinations for LSE dataset.
Stars mark those values of $P^{+}$ that are significantly higher than the same probability obtained by using the most recent value of $q$ instead of $\langle ES \rangle (T_a)$ as a predictor for $q(T_a)$ in the logistic regression (let us call $P^{+}_q$ such probability). Specifically, we have defined a null model where variations from such probability $P^{+}_q$ are due to random fluctuations only; given $n$ observations, such fluctuations follow a Binomial distribution $B(P^{+}_q,n)$, with mean $n P^{+}_q$ and variance $n P^{+}_q(1-P^{+}_q)$. Then p-values have been calculated by using this null distribution for each combination of parameters.
This null hypothesis accounts for the predictability of $q(T_a)$ that is due to the autocorrelation of $q(T_a)$ only; therefore $P^{+}$ significantly higher than the value expected under this hypothesis implies a forecasting power of $\langle ES \rangle (T_a)$ that is not explained by the autocorrelation of $q(T_a)$. From the table we can see that $P^{+}$ is significant in 12 out of 16 combinations of parameters for NYSE dataset, and in 13 out of 16 for LSE dataset. This means that correlation persistence is a valuable predictor for future average correlation, able to outperform forecasting method based on past average correlation trends.
These results are robust against changes of $f_{test}$, as long as the training set is large enough to allow an accurate calibration of the logistic regression. We have found this condition is satisfied for $f_{test} < 40\%$.
However $P^{+}$ does not give any information on the method ability to distinguish between true and false positives. To investigate this aspect we need $TPR$ and $FPR$. A traditional way of representing both measures from a binary classifier is the so-called ``Receiver operating characteristic'' (ROC) curve \cite{roc_curve}. In a ROC plot, $TPR$ is plotted against $FPR$ as the discriminant threshold is varied. The discriminant threshold $p_{max}$ is the value of the probability in Eq. \ref{eq:logistic_regression} over which we classify $Y(T_a)=1$: the higher $p_{max}$ is, the less likely the method is to classify $Y(T_a)=1$ (in the analysis on $P^{+}$ we chose $p_{max}=0.5$). Ideally, a perfect classifier would yield $TPR=1$ for all $p_{max}>0$, whereas a random classifier is expected to lie on the line $TPR=FPR$. Therefore a ROC curve which lies above the line $TPR=FPR$ indicates a classifier that is better than chance at distinguishing true from false positives \cite{statistical_learning_book}.
As one can see from Fig. \ref{fig:roc_curve}, the ROC curve's position depends on the choice of parameters $\theta$ and $L$. In this respect our classifier performs better for low values of $L$ and $\theta$. This can be quantified by measuring the area under the ROC curve; such measure, often denoted by AUC \cite{statistical_learning_book}, is shown in Tabs. \ref{tab:auc_nyse_outS}-\ref{tab:auc_lse_outS}. For both datasets the optimal choice of parameters is $\theta=500$ and $L=10$.
\begin{table}[h!]\centering
\caption{\label{tab:nyse_p_outS} NYSE dataset: {\bf Probability of successful forecasting $P^{+}$}, for different combinations of parameters $\theta$ and $L$.
Out-of-sample analysis.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & 0.546 & 0.560* & 0.599** & 0.539**\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & 0.704** & 0.695** & 0.658** & 0.605**\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & 0.634* & 0.585 & 0.539 & 0.708*\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & 0.704* & 0.7638** & 0.839** & 0.860\\
\hline
\addlinespace[1ex]
\multicolumn{6}{l}{\textsuperscript{**}$p<0.001$, \textsuperscript{*}$p<0.01$,}
\end{tabular}
\end{table}
\begin{table}[h!]\centering
\caption{\label{tab:lse_p_outS} LSE dataset: {\bf Probability of successful forecasting $P^{+}$}, for different combinations of parameters $\theta$ and $L$.
Out-of-sample analysis.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & 0.616** & 0.645** & 0.612** & 0.568**\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & 0.652** & 0.635** & 0.598** & 0.393\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & 0.651** & 0.560** & 0.453** & 0.412\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & 0.544** & 0.573** & 0.706** & 0.689\\
\hline
\addlinespace[1ex]
\multicolumn{6}{l}{\textsuperscript{**}$p<0.001$, \textsuperscript{*}$p<0.01$,}
\end{tabular}
\end{table}
\begin{figure}[t!]
\hspace{-4em}
\includegraphics[scale=0.4]{figure/roc_curves.pdf}
\includegraphics[scale=0.4]{figure/LSE/roc_curves_LSE.pdf}
\caption{ \label{fig:roc_curve} {\bf Receiver operating characteristic (ROC) curve}. Upper graph: True positive rate (TPR) against False positive rate (FPR) as the discriminant threshold $p_{max}$ of the classifier is varied, for each combination of parameters $\theta$ and $L$ in the NYSE dataset. The closer the curve is to the upper left corner of each graph, the better is the classifier compared to chance. Bottom graph: True positive rate (TPR) against False positive rate (FPR) as the discriminant threshold $p_{max}$ of the classifier is varied, for each combination of parameters $\theta$ and $L$ in the LSE dataset.}
\end{figure}
\begin{table}[t]\centering
\caption{\label{tab:auc_nyse_outS} NYSE dataset: {\bf Area under the curve (AUC)}, measured from the ROC curve in Fig. \ref{fig:roc_curve}. Values greater than 0.5 indicate that the classifier performs better than chance.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & 0.669 & 0.652 & 0.655 & 0.616 \\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & 0.775 & 0.753 & 0.710 & 0.625\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & 0.663 & 0.6220 & 0.574 & 0.520\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & 0.467 & 0.470 & 0.462 & 0.314\\
\hline
\end{tabular}
\end{table}
\begin{table}[t]\centering
\caption{\label{tab:auc_lse_outS} LSE dataset: {\bf Area under the curve (AUC)}, measured from the ROC curve in Fig. \ref{fig:roc_curve}. Values greater than 0.5 indicate that the classifier performs better than chance.}
\begin{tabular}{ cc | c c c c |}
\cline{3-6}
& & \multicolumn{4}{c|}{L} \\
\cline{3-6}
& & \textbf{10} & \textbf{25} & \textbf{50} & \textbf{100}\\
\hline
\multicolumn{1}{|c}{\multirow{4}{*}{$\theta$}} & \multicolumn{1}{|c|}{\textbf{250}} & 0.673 & 0.658 & 0.618 & 0.524 \\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{500}} & 0.727 & 0.700 & 0.602 & 0.431\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{750}} & 0.324 & 0.274 & 0.234 & 0.148\\
\multicolumn{1}{ |c }{} & \multicolumn{1}{|c|}{\textbf{1000}} & 0.233 & 0.168 & 0.0918 & 0.0160\\
\hline
\end{tabular}
\end{table}
\subsection*{Temporal evolution of forecasting performance}
\begin{figure}[ht!]
\includegraphics[scale=0.4]{figure/forecasting_in_time_nyse.pdf}
\includegraphics[scale=0.4]{figure/LSE/forecasting_in_time_lse.pdf}
\caption{\label{fig:forecast_time} {\bf Fraction of successful predictions as a function of time}. NYSE (left graph) and LSE dataset (right graph). Forecasting is based on logistic regression with predictor $\langle ES(T_a)\rangle$ (top graphs) and most recent value of $q(T_a)$ (bottom graphs). Horizontal lines represent the average over the entire period.}
\end{figure}
In this section we look at how the forecasting performance changes at different time periods. In order to explore this aspect we have counted at each time window $T_a$ the number $N^{+}(T_a)$ of $Y(T_a)$ predictions (out of the 16 predictions corresponding to as many combinations of $\theta$ and $L$) that have turned out to be correct; we have then calculated the fraction of successful predictions $n^{+}(T_a)$ as $n^{+}(T_a)=N^{+}(T_a)/16$. In this way $n^{+}(T_a)$ is a proxy for the goodness of our method at each time window. Logistic regression parameters $\beta_0$ and $\beta_1$ have been calibrated by using the entire time period as training set, therefore this amounts to an in-sample analysis.
In Fig. \ref{fig:forecast_time} we show the fraction of successful predictions for both NYSE and LSE datasets (upper graphs, blue circles). For comparison we also show the same measure obtained by using the most recent value of $q(T_a)$ as predictor (bottom graphs); as in the previous section, it represents a null model that makes prediction by using only the past evolution of $q(T_a)$. As we can see, both predictions based on $\langle ES \rangle (T_a)$ and on past values of $q(T_a)$ display performances changing in time. In particular $n^{+}(T_a)$ drops just ahead of the main financial crises (the market downturn in March 2002, 2007-2008 financial crisis, Euro zone crisis in 2011); this is probably due to the abrupt increase in volatility that occurred during these events and that the models took time to detect. After these drops though performances based on $\langle ES \rangle (T_a)$ recover much more rapidly than those based on past value of $q(T_a)$. For instance in the first months of 2007 our method shows quite high $n^{+}(T_a)$ (more than $60\%$ of successful predictions), being able to predict the sharp increase in volatility to come in 2008 while predictions based on $q(T_a)$ fail systematically until 2009. Overall, predictions based on correlation structure persistence appear to be more reliable (as shown by the average $n^{+}(T_a)$ over all time windows, the horizontal lines in the plot) and faster at detecting changes in market volatility.
\section*{Discussion}
In this paper we have proposed a new tool for forecasting market volatility based on correlation-based information filtering networks and logistic regression, useful for risk and portfolio management. The advantage of our approach over traditional econometrics tools, such as multivariate GARCH and stochastic covariance models, is the ``top-down'' methodology that treats correlation matrices as the fundamental objects, allowing to deal with many assets simultaneously; in this way the curse of dimensionality, that prevents e.g. multivariate GARCH to deal with more than few assets, is overcome. We have proven the forecasting power of this tool by means of out-of-sample analyses on two different stock markets; the forecasting performance has been proven to be statistically significant against a null model, outperforming predictions based on past market correlation trends. Moreover we have measured the ROC curve and identified an optimal region of the parameters in terms of True Positive and False Positive trade-off. The temporal analysis indicates that our method is able to adapt to abrupt changes in the market, such as financial crises, more rapidly than methods based on past volatility.
This forecasting tool relies on an empirical fact that we have reported in this paper for the first time. Specifically, we have shown that there is a deep interplay between market volatility and the rate of change of the correlation structure. The statistical significance of this relation has been assessed by means of a block-bootstrapping technique. An analysis based on metacorrelation has revealed that this interplay is better highlighted when filtering based on Planar Maximally Filtered Graphs is used to estimate the correlation structure persistence.
This finding sheds new light into the dynamic of correlation. The topology of Planar Maximally Filtered Graphs depends on the ranking of the $N(N-1)/2$ pairs of cross-correlations; therefore an increase in the rate of change in PMFGs topology points out a faster change of this ranking. Our result indicates that such increase is typically followed by a rise in the market volatility, whereas decreases are followed by drops.
A possible interpretation of this is related to the dynamics of risk factors in the market. Indeed higher volatility in the market is associated to the emergence of a (possibly new) risk factor that makes the whole system unstable; such transition could be anticipated by a quicker change of the correlation ranking, triggered by the still emerging factor and revealed by the correlation structure persistence. Such persistence can therefore be a powerful tool for monitoring the emergence of new risks, valuable for a wide range of applications, from portfolio management to systemic risk regulation. Moreover this interpretation would open interesting connections with those
approaches to systemic risk that make use of Principal Component Analysis, monitoring the emergence of new risk factors by means of spectral methods \cite{pca_systemic_risk,pca_systemic_risk2}. We plan to investigate all these aspects in a future work.
\section*{Methods}
\subsection*{Metacorrelation as a measure of correlation structure persistence}
Given two correlation matrices $\{\rho_{ij}(T_a)\}$ and $\{\rho_{ij}(T_b)\}$ at two different time windows $T_a$ and $T_b$, their metacorrelation $z(T_a,T_b)$ is defined as follows:
\begin{equation}
z(T_a,T_b) = \frac{\langle \rho_{ij}(T_a) \rho_{ij}(T_b) \rangle_{ij}}{\sqrt{[\langle \rho^2_{ij}(T_a) \rangle_{ij} - \langle \rho_{ij}(T_a) \rangle_{ij}^2][\langle \rho^2_{ij}(T_b) \rangle_{ij} - \langle \rho_{ij}(T_b) \rangle_{ij}^2] }} ,
\label{eq:z}
\end{equation}
where $\langle ... \rangle_{ij}$ is the average over all couples of stocks $i,j$.
Similarly to Eq. \ref{eq:es} we have then defined $z(T_a)$ as the weighted average over $L$ past time windows:
\begin{equation}
\langle z \rangle (T_a) = \sum_{b = a -L }^{a - 1} \omega(T_b) z(T_a, T_b) .
\end{equation}
\subsection*{Measures of classification performance}
With reference to Figs. \ref{fig:training_test} b) and d), let us define the number of observations in each quadrant $Q_i$ ($i=1,2,3,4$) as $|Q_i|$. In the terminology of classification techniques \cite{machine_learning_book}, $|Q_1|$ is the number of True Positive (observations for which the model correctly predicted $Y(T_{a})=1$), $|Q_3|$ is the number of True Negative (observations for which the model correctly predicted $Y(T_{a})=0$), $|Q_2|$ the number of False Negative (observations for which the model incorrectly predicted $Y(T_{a})=0$) and $|Q_4|$ the number of False Positive (observations for which the model incorrectly predicted $Y(T_a)=1$). We have then computed the following measures of quality of classification, that are the standard metrics for assessing the performances of a classification method \cite{machine_learning_book}:
\begin{itemize}
\item {\bf Probability of successful forecasting ($P^{+}$)} \cite{machine_learning_book}: represents the method probability of a correct prediction, expressed as fraction of observed $\langle ES \rangle (T_a)$ values through which the method has successfully identified the correspondent value of $Y(T_{a})$. In classification problems, sometimes, the error rate $I$ is used \cite{statistical_learning_book}, which is simply $I=1-P^{+}$. $P^{+}$ is computed as follows:
\begin{equation}\label{eq:prob_forecast}
P^{+} = \frac{|Q_1|+|Q_3|}{|Q_1|+|Q_2|+|Q_3|+|Q_4|} .
\end{equation}
\item {\bf True Positive Rate ($TPR$)} \cite{machine_learning_book}: it is the probability of predicting $Y(T_a)=1$, conditional to the fact that the real $Y(T_a)$ is indeed $1$ (that is, to predict an increase in volatility when the volatility will indeed increase); it represents the method sensitivity to increase in volatility. It is also called ``recall'' \cite{statistical_learning_book}. In formula:
\begin{equation}\label{eq:tpr}
TPR = \frac{|Q_1|}{|Q_1|+|Q_2|} .
\end{equation}
\item {\bf False Positive Rate ($FPR$)} \cite{machine_learning_book}: it is the probability of predicting $Y(T_a)=1$, conditional to the fact that the real $Y(T_a)$ is instead $0$ (that is, to predict an increase in volatility when the volatility will actually decrease). It is also called ``1-specificity'' \cite{statistical_learning_book}. In formula:
\begin{equation}\label{eq:fpr}
FPR = \frac{|Q_4|}{|Q_3|+|Q_4|} .
\end{equation}
\end{itemize}
|
\section{Introduction}
There is an extensive, more than 100 year old,
literature on the mechanics of
necking in the uniaxial tensile test. The classical criterion of
\cite{C85} holds
for necking of a tensile bar in the limiting case of a infinitely long, thin
bar and states that necking initiates at the maximum
load. For any finite aspect ratio, there is a delay between the
maximum load point and the onset of necking that increases as the bar
becomes more stubby, \cite{N72,HM74,HN77}. The
literature on the analyses of necking in tensile bars
includes one dimensional analyses as well as full three dimensional finite element
solutions, involving both quasi-static and dynamic formulations, and analyses
that account for effects of various mechanical
properties, such as thermal softening, porosity induced softening, bar
geometry, etc. Reviews of tensile bar necking
analyses are provided by \cite{JWH79,MMJ14}.
For rate independent plasticity and quasi-static deformations, the
onset of necking in a uniform circular cylindrical tensile bar is
associated with a
bifurcation from a state of homogeneous uniaxial tension
\cite{CAD71,N72,HM74}. The bifurcation mode is
associated with a sinusoidal variation in the radial dimension of the
bar with the longest possible wavelength consistent with the bar
geometry and the boundary conditions at the bar ends. A geometrical
imperfection leads to the fairly abrupt development
of this mode at an overall strain somewhat less (depending on the
imperfection) then the bifurcation strain. Once the neck develops, the
classic approximate analysis of \cite{B52}, and subsequent
full numerical solutions, e.g. \cite{C71,N72,Im75,norris78,TN84,TN85}, show
that the neck curvature induces stress triaxiality that plays a key
role in the ductile failure process.
For a viscoplastic solid under
quasi-static loading conditions, the
onset of necking is no longer associated
with a bifurcation. However, the onset of necking can be analyzed as
the growth of an initial inhomogeneity, \cite{HN77}. As for a rate
independent plastic solid, a notch serves as an imperfection that
triggers necking and sets the neck location. Material rate
sensitivity leads to a delay in the onset of necking, \cite{HN77}.
In addition, for both rate
independent and rate dependent plastic solids characterized by a
classic plastic constitutive relation, there is no material
length scale in a quasi-static analysis. Hence, the evolution of the
neck with strain (at the same imposed strain rate for viscoplastic solids) is
independent of specimen size.
The necking behavior under
dynamic loading conditions, e.g.
\cite{AN91,KN93,FN94,GF02,MN03,rusinek05,Os13,VR15,ROR15}, can be quite
different than under quasi-static conditions. Material inertia tends to slow neck
development, \cite{AN91,Xue08}; multiple necking can occur,
e.g. \cite{KN93,FN94,GF02}; there are size effects,
e.g. \cite{rusinek05,KN93}, and neck
development can ignore the presence of notches, \cite{ROR15}. Experiments and
modeling carried out in \cite{ROR15} showed that under dynamic loading
conditions, the onset of necking in notched tensile bars could occur
away from the notch location.
Since material inertia implicitly introduces a length scale,
different size specimens deformed at the same strain rate
may respond differently. As a consequence, there can be a
dependence of the failure strain on specimen size, \cite{KN93}.
\cite{KN93} carried out finite deformation
dynamics analyses
aimed at modeling the effect of specimen size at a
fixed imposed strain rate, on ductile failure in
geometrically self-similar tensile bars having various
sizes. The material was
modeled as a viscoplastic progressively cavitating solid. It was found
that the variation of the necking strain with specimen size was not
monotonic; the response of sufficiently small specimens was
essentially quasi-static and size independent, the
failure strain then increased with specimen size before eventually
decreasing for sufficiently large specimens.
In this study, a combination of the
issues addressed in \cite{KN93} and \cite{ROR15} is considered. In
particular, a main
focus in this paper is to continue exploring the issue raised by
\cite{ROR15} concerning the circumstances, for
dynamic loading conditions, under which necking ignores the presence
of a notch.
Calculations are carried out for geometrically
similar, dynamically loaded notched
circular cylindrical tensile bars of various sizes. Attention is restricted to
axisymmetric deformations. The bar material is characterized as an
isotropically hardening viscoplastic Mises solid. A non-dimensional
form of the governing equations is presented and two key
non-dimensional ratios are identified: one relates the bar length
to a characteristic length that depends on material properties and the
imposed velocity, while the other relates the imposed strain
rate (the imposed velocity divided by the bar length) to a material
characteristic strain rate. Both of these non-dimensional ratios involve
the bar length. The focus of the results is on the transition from
necking at the notch cross section to necking away from the notch
cross section as the specimen size is varied.
\section{Problem Formulation}
As in \cite{KN93}, the calculations are based on a convected
coordinate Lagrangian formulation of the field
equations. The independent variables are taken to be the particle
positions in the initial stress free configuration of the axisymmetric
tensile bar and time. In the current configuration the material point
initially at ${\bf X}$ is at ${\bf x}$. The displacement vector ${\bf
u}$ and the deformation gradient ${\bf F}$ are defined by
\begin{equation}
{\bf u}= {\bf x} - {\bf X} \quad , \quad {\bf F}= \frac{\partial {\bf u}}{\partial {\bf X}}
\end{equation}
The principle of virtual work accounting for material inertia is written as
\begin{equation}
\int_V {\bf S}:\delta {\bf F} dV=\int_B ( {\bf S} \cdot {\bf n}) \cdot \delta {\bf u} dB
-\int_V \rho \frac{\partial^2 {\bf u}}{\partial T^2} \cdot \delta
{\bf u} dV
\label{pvw}
\end{equation}
Here, $T$ is time, ${\bf S}$ is the (unsymmetric) nominal stress tensor, $ {\bf S}=
({\rm det} {\bf F}) {\bf F}^{-1} \cdot \boldsymbol{\Sigma}$ with
$\boldsymbol{\Sigma}$ the Cauchy stress, $\rho$ is the mass density,
and $V$ and $B$ are, respectively, the volume and the surface of the
body in the undeformed reference configuration.
Attention is confined to axisymmetric deformations and, for
notational simplicity, we use $r$
and $z$ to denote the convected Lagrangian coordinates.
The initial length of the bar is $2L_0$ and the initial radius, which
varies along the bar is denoted by $R_0(z)$. The bar occupies the
region $-L_0 \le z \le L_0$, $0 \le r \le R_0(z)$.
An axial velocity $V(t)$ is imposed on $z=L_0$ together with
symmetry about $z=0$. This means that the loading is actually applied
at $z=-L_0$ as well. The reason for imposing symmetry about $z=0$ is that
without this symmetry and with shear free conditions on the loading
ends, the preferred quasi-static necking mode would be the long
wavelength mode with the neck forming at one of the ends. Thus, under
quasi-static loading conditions the deformation and stress
concentrations associated with the centrally placed notch would be
competing with those associated with the preferred necking mode. On
the other hand, if shear constraints were imposed at the ends, then
the onset of necking would be affected by
both the notch and the deformation gradient imposed by the
constraints, complicating the interpretation of the notch effect. With
symmetry about $z=0$ imposed, the preferred quasi-static necking mode
is driven by the the presence of the notch and necking occurs
at the notch cross section. The occurrence of
necking away from the notch cross section, when it occurs, is a dynamic
effect.
The boundary conditions on the region analyzed are $u_z(r,L_0,T)=V(T)$ where
\begin{equation}
V(T)=\begin{cases}
V_1 \, T/T_r, &\mbox{for } T \le T_r \\
V_1 ,&\mbox{for } T > T_r
\end{cases}
\end{equation}
Here, $V_1$ is the magnitude of the imposed velocity and $T_r$ is the rise time.
The other displacement boundary conditions imposed are $u_z(r,0,T)=0$
and $u_r(0,z,T)=0$. All other boundary
conditions correspond to zero imposed tractions.
The material is characterized as an elastic-viscoplastic Mises solid.
The total rate of deformation, ${\bf D}$, is written as the sum of an
elastic (actually hypoelastic) part, ${\bf D}^e$, and a viscoplastic part, ${\bf D}^p$, with
\begin{equation}
{\bf D}^e=\frac{1+\nu}{E} {\tilde {\boldsymbol{T}}}-
\frac{\nu}{E} {\rm tr}(\tilde{ {\boldsymbol{T}}}){\bf I}
\label{con1}
\end{equation}
where $E$ is Young's modulus, $\nu$ is Poisson's ratio,
$\boldsymbol{T}=({\rm det} {\bf
F}) \boldsymbol{\Sigma}$, $(\tilde{\
})$ denotes the Jaumann rate based on $T$, ${\rm tr}( \ )$ denotes the trace and
${\bf I}$ is the identity tensor.
The viscoplastic flow rule is
\begin{equation}
{\bf D}^p=\frac{3 \dot {\Lambda}^p}{2 \Sigma_e}
\boldsymbol{T}^\prime
\label{con2}
\end{equation}
where ${\bf I}$ is the identity tensor, $\dot {\Lambda}^p$ is the
effective plastic strain rate, and the Kirchhoff stress deviator
$\boldsymbol{T}^\prime$ and effective stress $ \Sigma_e$ are given by
\begin{equation}
\boldsymbol{T}^\prime=\boldsymbol{T}- \Sigma_h {\bf I} \quad , \quad \Sigma_e=
\sqrt{\frac{3}{2}\boldsymbol{T}^\prime : \boldsymbol{T}^\prime} \quad ,
\quad \Sigma_h=\frac{1}{3} {\rm tr}(\boldsymbol{T}) {\bf I}
\end{equation}
The material response, the specimen geometry and the boundary value
problem are characterized by a collection of non-dimensional
quantities. To put the equations in non-dimensional form, we normalize
all stress quantities by a reference stress $\sigma_0$, all length
quantities by a reference length $L_c$ and all time quantities by a
reference time $t_c$.
The principle of virtual work, Eq.~(\ref{pvw}), can be written as
\begin{equation}
\int_v {\bf s}:\delta {\bf F} dv=\int_b ({\bf s} \cdot {\bf n}) \cdot \delta {\bf w} db
-\int_V \ddot{\bf w} \cdot \delta {\bf w} dv
\label{pvw2}
\end{equation}
provided
\begin{equation}
t_c={L_c}\sqrt{\frac{\rho}{\sigma_0}}
\label{rel1}
\end{equation}
In Eq.~(\ref{pvw2}), ${\bf s} =\sigma_0 {\bf S}$, ${\bf u} = L_c {\bf w}$, $\ t=t_c T$,
$dV=L_c^3 dv$, $dB=L_c^2 db$ and $(\dot \ )$ denotes $\partial (
\ )/\partial t$.
As in \cite{KN93}, we take
\begin{equation}
L_c=L_0 \frac{c_0}{V_1} \ , \ c_0=\sqrt{\frac{E}{\rho}}
\label{rel2}
\end{equation}
Hence, from Eq.~(\ref{rel1})
\begin{equation}
t_c=\frac{L_0}{V_1}\sqrt{\frac{E}{\sigma_0}}
\label{rel3}
\end{equation}
In non-dimensional form, the rate constitutive relation,
Eqs.~(\ref{con1}) and (\ref{con2}) become
\begin{equation}
{\bf d}^e=\epsilon_0\left [ (1+\nu) {\hat {\boldsymbol{\tau}}}-
\nu {\rm tr}(\hat{ {\boldsymbol{\tau}}}){\bf I} \right ]
\quad , \quad \epsilon_0=\frac{\sigma_0}{E}
\label{con3}
\end{equation}
where $(\hat{\ })$ denotes the Jaumann rate based on $t$, and
\begin{equation}
{\bf d}^p=\frac{3 \dot {\epsilon}^p}{2 \sigma_e}
\boldsymbol{\tau}^\prime
\label{con4}
\end{equation}
In Eq.~(\ref{con4}), $\dot {\epsilon}^p=t_c \dot {\Lambda}^p$.
The plastic response of the material is characterized by a power law
rate hardening of the form
\begin{equation}
\dot{\epsilon}_p=t_c \dot{\epsilon}_0 \left( \frac{\sigma_e}{g} \right)^{1/m}
\label{ep-def}
\end{equation}
Here, $\sigma_{e}=\Sigma_e/\sigma_0$, $\dot{\epsilon}_0$ is a
reference strain rate, $m$ is the rate
sensitivity exponent.
The flow strength function
$g$ in Eq.~(\ref{ep-def}) is taken to be a function of $\epsilon_p$
and have the form
\begin{equation}
g\left(\epsilon_p\right)= \left [ 1+\frac{\epsilon_p}{\epsilon_0} \right ]^N
\label{hardening}
\end{equation}
Here, $\epsilon_p=\int \dot{\epsilon}_p dt$ and $N$ is the strain hardening
exponent. Since $g(0)=1$ in Eq.~(\ref{hardening}), $\sigma_0$ is
now identified with the flow strength at zero plastic strain. Note that up
to the identification in Eq.~(\ref{hardening}), $\sigma_0$ could be any
convenient quantity having the dimension of stress.
The constitutive response is characterized by the non-dimensional
material parameters.
$\epsilon_0$, $\nu$, $m$ and $N$. For a given function
$R_0(z)$, the specimen geometry is characterized by the non-dimensional
ratio $R_0(0)/L_0$.
There are two key non-dimensional groups involving both loading and material
parameters. One is
\begin{equation}
\frac{L_0}{L_c}=\frac{V_1}{c_0}
\label{eq:L}
\end{equation}
and the other is
\begin{equation}
\kappa=\frac{V_1/L_0}{\dot{\epsilon}_0}
\label{eq:kap}
\end{equation}
which is the ratio of the imposed strain rate to the
material characteristic strain rate. The ratio $L_0/L_c$ provides a
measure of the effect of material inertia while $\kappa$ provides a measure of
the effect of loading rate, independent of material inertia. In the quasi-static limit
$L_0/L_c \rightarrow 0$ so that, of course, the effect of inertia on
the response vanishes. On the other hand, the role of $\kappa$
persists in the quasi-static limit.
There are other non-dimensional groups relating
material parameters and loading parameters, but they are not
independent of the ones already defined. For example,
$t_c \dot{\epsilon}_0$, which can be regarded as a ratio of
dynamic and constitutive time scales is, from Eqs.~(\ref{rel3}) and
(\ref{eq:kap}), given by
\begin{equation}
t_c \dot{\epsilon}_0=\frac{L_0
\dot{\epsilon}_0}{V_1}\sqrt{\frac{E}{\sigma_0}}
=\frac{1}{\kappa \sqrt{\epsilon_0}}
\label{eq:ratx}
\end{equation}
Eq.~(\ref{ep-def}) then can be rewritten as
\begin{equation}
\dot{\epsilon}_p=\frac{1}{\kappa \sqrt{\epsilon_0}} \left(
\frac{\sigma_e}{g} \right)^{1/m}
\label{ep-def2}
\end{equation}
and used to express the plastic flow rule,
Eq.~(\ref{con4}) in non-dimensional form as
\begin{equation}
{\bf d}^p=\frac{3}{2} \frac{1}{\kappa \sqrt{\epsilon_0}} \left(
\frac{\sigma_e}{g} \right)^{1/m} \left(
\frac{\boldsymbol{\tau}^\prime}{ \sigma_e} \right )
\label{con5}
\end{equation}
Eq.~(\ref{con5}) explicitly shows the constitutive dependence on the
non-dimensional parameter $\kappa$ which gives rise to a coupling
of the constitutive response to parameters involving the bar geometry
and the imposed loading. This coupling occurs because of
material rate sensitivity and does not occur for the corresponding
rate independent plastic flow rule.
Another non-dimensional ratio is the ratio of the stress carried by the
loading wave to the reference stress. A one
dimensional linear elastic wave propagation analysis of a tensile
bar subject to a prescribed end velocity $V_1$ gives the stress
carried by the loading wave as $\rho c_0 V_1$, see e.g.
\cite{Lee67}. However, this is not an independent ratio.
The ratio of this loading wave stress to the reference stress $\sigma_0$ is
\begin{equation}
\frac{\rho c_0 V_1}{\sigma_0}=\frac{1}{\epsilon_0}
\frac{V_1}{c_0}=\frac{1}{\epsilon_0} \frac{L_0}{L_c}
\label{rel4}
\end{equation}
Hence, the stress carried by the loading wave (according to a one
dimensional linear elastic analysis) is proportional to $L_0/L_c$ and is
inversely proportional to the non-dimensional material parameter
$\epsilon_0$. Hence, with material properties fixed, the stress
carried by the loading wave
increases with increasing relative bar size, $L_0/L_c$ with the
quasi-static limit emerging as $L_0/L_c \rightarrow 0$.
Thus, the boundary value problem is characterized by the non-dimensional
material parameters, $\epsilon_0$, $\nu$, $N$ and $m$; the
non-dimensional geometry parameter $R_0/L_0$; the non-dimensional rise
time $t_r$; and the
non-dimensional parameters, $L_0/L_c$ and $\kappa$, that involve
geometry, loading and material parameters.
To illustrate the scaling, let $\rho^*=A \rho$, then with
\begin{equation}
E^*=A^p E \ , \ \sigma_0^*=A^p \sigma_0
\label{eq:sc1}
\end{equation}
the linear elastic wave speed scales as $c_0^*=A^{(p-1)/2}c_0$.
Taking $V_1^*=A^{(p-1)/2}V_1$, $L_0^*=A^{(p-1)/2}L_0$, and
$\dot{\epsilon}_0^*=\dot{\epsilon}_0$, the non-dimensional ratios
$L_0/L_c$, $\epsilon_0$ and $\kappa$, as well as the characteristic
time $t_c$ are unchanged with this scaling. Thus, with
this scaling the solution of the dynamic
initial/boundary value problem (with fixed $R_0/L_0$) coincides for
both these sets of material and loading parameters.
\section{Numerical Method and Results}
\label{result}
The numerical method is basically the same as in \cite{AN91,KN93}.
The discretization is based on linear displacement triangular elements
arranged in quadrilaterals of four ``crossed'' triangles and the time
integration of the discretized governing equations are
integrated numerically by an explicit integration procedure,
\cite{Belytschko76}, with a lumped mass matrix.
The constitutive update is based on the rate
tangent modulus method of \cite{Peirce84}.
\begin{figure}[htb!]
\begin{center}
\resizebox*{35mm}{!}{\includegraphics{L20-grd.eps}}
\end{center}
\caption{The $12 \times 96$ quadrilateral finite element mesh used in
the computations. Each quadrilateral consists of four ``crossed''
linear displacement triangles.}
\label{mesh}
\end{figure}
The fixed (non-dimensional) constitutive parameters are taken to be
$\epsilon_0=0.004$, $\nu=0.3$. The calculations are carried out for strain rate
hardening exponents $m=0.01$ and $m=0.05$. In most calculations the
strain hardening exponent is taken to be
$N=0.01$, which is nearly ideally plastic, but a few calculations employ
$N=0.1$ to assess the effect of strain hardening. The value of the
(non-dimensional) rise time is taken to be $t_r=1.265\times 10^{-3}$ in
all calculations.
To give a perspective on what these non-dimensional parameter values
could correspond to, one possibility is $E=200$GPa, $\Sigma_0=800$MPa,
$c_0=5000$m/s, $V_1/L_0=1000$s$^{-1}$, $\dot{\epsilon}_0=1000$s$^{-1}$ and
$T_r=20 \times 10^{-6}$s. Also, with these values
$L_c=c_0/(V_1/L_0)=5$m (close to the value in \cite{KN93}).
The bar aspect
ratio is fixed at $L_0/R_0=4$ and the bar has a notch as depicted in
Fig.~\ref{mesh}. The notch is a semi-circle of radius $R_0/10$
centered at $z=0$, $r=R_0$. However, as can be seen in Fig.~\ref{mesh}
there are only three nodal points on the notch surface so that the
circular shape is not faithfully represented. The notch mainly serves as an
imperfection to trigger necking at $z=0$.
We define the ratio of current cross section area to initial cross
section area, $A_r(z)$, by
\begin{equation}
A_r(z)=\frac {\pi R_0^2(z)}{\pi [R_0(z)+u_r(R_0(z),z)]^2}
\label{area}
\end{equation}
A calculation is terminated when the maximum of $A_r(z)$ reaches
$2$ (an area reduction of $1/2$). In the calculations here, this
occurs either for $A_r(0)$ or $A_r(L_0)$.
If $A_r(0)$ reaches $2$ first then necking has occurred at the notch,
if $A_r(L_0)$ reaches $2$ first then necking has occurred away from
the notch and necking has ignored the presence of the notch. The value
$A_r=2$ is chosen arbitrarily. This value was chosen because neck
development has clearly occurred when $A_r=2$ yet the strains are not
so large that the finite element grid in Fig.~\ref{mesh} has become
significantly deformed. As will be seen subsequently, the value of
necking strain is not generally very sensitive to the precise cut-off value
chosen.
At the termination of the calculation we define the ratio
\begin{equation}
R_A=\frac{A_r(0)}{A_r(L_0)}
\label{RA}
\end{equation}
so that $R_A>1$ implies necking at the notch, $R_A<1$ implies necking
away from the notch and $R_A=1$ corresponds to simultaneous necking at
the notch and away from the notch. Assuming a non-negative effective
Poisson's ratio, the possible range of values for $R_A$ is $0.5 \le
R_A \le 2$.
\subsection{Fixed $\kappa$, varying $L_0/L_c$}
\label{Lc}
In this section the
value of $\kappa$, Eq.~(\ref{eq:kap}), is fixed at
$\kappa=1$ and $L_0/L_c$, Eq.~(\ref{eq:L}), is varied.
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./L20-80-ld.eps}}}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./LY20-80-ld.eps}}}
\end{center}
\caption{Nominal stress-strain curves. (a) $m=0.01, N=0.01$. (b)
$m=0.05, N=0.01$.}
\label{L-LY-ld}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./arat-20.eps}}}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./arat-60.eps}}}
\end{center}
\caption{Evolution of $A_r(0)$ and $A_r(L_0)$, with $A_r(z)$ given by
Eq.~(\ref{area}). (a) $m=0.01, N=0.01$, $L_0/L_c=0.004$. (b)
$m=0.01, N=0.01$, $L_0/L_c=0.012$.}
\label{ar2060}
\end{figure}
Fig.~\ref{L-LY-ld} shows nominal stress-strain curves for $m=0.01$,
Fig.~\ref{L-LY-ld}a, and for $m=0.05$, Fig.~\ref{L-LY-ld}b. In both
plots $N=0.01$. The quantity $S$ is the nominal stress, the force per unit reference
area, and the end displacement $U$ is $U=\int V dt$. For sufficiently
small $L_/L_c$ the response is essentially quasi-static and this is
nearly the case for $L_0/L_c=0.004$ in Fig.~\ref{L-LY-ld}.
As $L_0/L_c$ increases from $0.004$, inertia plays an increasing role
which delays the onset of necking. The same
overall trend occurs for both $m=0.01$ and $m=0.05$. In
Fig.~\ref{L-LY-ld}a where $m=0.01$ the maximum
necking strain occurs for $L_0/L_c=0.008$. In Fig.~\ref{L-LY-ld}b
where $m=0.05$ the maximum necking strain occurs for
$L_0/L_c=0.012$. Thus, increased strain rate sensitivity results in
the maximum necking strain (as defined here) occurring for a larger
value of $L_0/L_c$. For the larger values of
$L_0/L_c$ wave effects come into play which is what leads to the
reduction in necking strain as also seen by \cite{KN93}.
Fig.~\ref{ar2060} shows the evolution of the area ratios $A_r(0)$ and
$A_r(L_0)$ with strain, $U/L_0$, for two cases. (Recall that $z$
denotes a convected
coordinate so that $z=L_0$ corresponds to the impact end in the
current configuration.) In Fig.~\ref{ar2060}a, $L_0/L_c=0.004$ and
necking occurs at
the notch cross section while in Fig.~\ref{ar2060}b, $L_0/L_c=0.012$
necking occurs at the impact end. In both cases, the area reduction
(recall that $A_r$ in Eq.~(\ref{area}) increases with increasing area
reduction) eventually increases rapidly at one end and remains nearly
constant at
the opposite end. In Fig.~\ref{ar2060}a the response is
quasi-static like and the relative area changes at $z=0$ and $z=L_0$ are
nearly the same initially. On the other hand in
Fig.~\ref{ar2060}b, where material inertia plays an important role,
significant deformation occurs at $z=L_0$ before much plastic
deformation takes place at $z=0$.
It can also be seen in Fig.~\ref{ar2060} from the
rapid increase in $A_r$ when necking occurs, that although the precise
value of the necking strain depends on the cut-off value chosen
($A_r=2$ here), the trends will not be sensitive to this value.
\begin{figure}[htb!]
\begin{center}
\resizebox*{75mm}{!}{\includegraphics{arat2.eps}}
\end{center}
\caption{Ratio of the relative area reductions at $z=0$ and
$z=L_0$, $R_A=A_r(0)/A_r(L_0)$. Values greater than $1$ correspond
to necking at $z=0$ while
values less than $1$ correspond to necking at $z=L_0$.}
\label{arat}
\end{figure}
The values of $R_A$, defined in Eq.~(\ref{RA}), for all values of
$L_0/L_c$ considered
is shown in Fig.~\ref{arat}. The reference line $R_A=1$ corresponds to
simultaneous necking at the two ends of the region analyzed. For
$m=0.01$ necking occurs at the notch for $L_0/L_c$ less than about
$0.0084$ and occurs at the impact end for greater values of
$L_0/L_c$. For $m=0.05$ this transition occurs just about at
$L_0/L_c=0.01$. Increasing the strain hardening exponent from $N=0.01$
to $N=0.1$ leads to the transition
from necking at the notch to necking away from the notch taking place
at a much larger value of $L_0/L_c$. However, although strain hardening
and/or strain rate hardening can strongly affect the size, i.e. the
value of $L_0/L_c$, at which this transition in necking location
occurs, the results in Fig.~\ref{arat} show that it is inertia that drives
the transition. With $L_c=5$m as for the dimensional values given in Section
\ref{result}, the transition value of $L_0$ for $m=0.01$, $N=0.01$ is
$0.044$m while with $m=0.05$ this increases to $\approx 0.05$m and for
$m=0.01$, $N=0.1$, the transition to necking away from the notch takes
place for $L_0 \ge 0.078$m.
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./L20-ep.eps}}}
\hspace{2cm}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./L20-sh.eps}}}
\end{center}
\caption{(a) Distribution of effective plastic strain
$\epsilon_p$. (b) Distribution of hydrostatic stress $\sigma_h$. For
$m=0.01$, $N=0.01$ and $L_0/L_c=0.004$ at $U/L_0=0.120$.}
\label{L20}
\end{figure}
Fig.~\ref{L20} shows contours of effective plastic strain, $\epsilon_p$, and
hydrostatic stress, $\sigma_h=\Sigma_h/\sigma_0$, for a case where
necking occurs at the notch. There
are two concentrations of plastic strain, $\epsilon_p$, in
Fig.~\ref{L20}a: one at the neck
center and one that extends at about $45^\circ$ from the notch
root. The strain concentration that extends at the notch
root does not occur in a naturally necked specimen. The distribution
of $\sigma_h$ has a hydrostatic
tension peak of about $0.9$ at the neck center and
hydrostatic compression away from
the neck along the bar axis that reaches $-0.6$. The
distributions of $\epsilon_p$ and $\sigma_h$ are very similar to
those that would be obtained from a quasi-static
analysis.
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./L60-ep.eps}}}
\hspace{2cm}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./L60-sh.eps}}}
\end{center}
\caption{(a) Distribution of effective plastic strain
$\epsilon_p$. (b) Distribution of hydrostatic stress $\sigma_h$. For
$m=0.01$, $N=0.01$ and $L_0/L_c=0.012$ at $U/L_0=0.161$.}
\label{L60}
\end{figure}
With $L_0/L_c=0.012$ in Fig.~\ref{L60} necking has occurred at the
impact end. As can be
seen in Fig.~\ref{L60}a a strain concentration did initiate at the
notch root and reach $\epsilon_p \approx 0.3$, but the deformation
eventually concentrated at the impact end. The distribution of
$\sigma_h$ in
Fig.~\ref{L60}b shows evidence of the initial neck development at
$z=0$. There are three concentrations of hydrostatic tension; at the
notch root, at the center of
the bar at $z=0$ and at the center of the bar at $z=L_0$. The maximum
positive hydrostatic stress is in the
neck that forms at $z=L_0$, reaching $0.79$. There is also
a region of
negative $\sigma_h$ along the axis as a consequence of the initial
neck formation there. The distributions of plastic strain $\epsilon_p$
and $\sigma_h$ in Fig.~\ref{L60} are very different from what would be
obtained from a quasi-static analysis.
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./LY50-arat.eps}}}\\
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./LY50-ep.eps}}}
\hspace{2cm}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./LY50-sh.eps}}}
\end{center}
\caption{(a) Evolution of the area ratios at $z=0$ and $z=L_0$, $A_r(0)/A_r(L_0)$. (b)
Distribution of effective plastic strain
$\epsilon_p$. (c) Distribution of hydrostatic stress $\sigma_h$. For
$L_0/L_c=0.010$, $m=0.05$, $N=0.01$ at $U/L_0=0.294$.}
\label{LY50}
\end{figure}
The calculation with $m=0.05$, $N=0.01$ and $L_0/L_c=0.010$ is unusual
in that necking
occurred nearly simultaneously at the notch plane ($z=0$) and at the
loaded end ($z=L_0/2$). Fig.~\ref{LY50}a shows the evolution of the
area ratios $A_r(0)$
and $A_r(L_0)$ with strain, $U/L_0$. Although
$A_r(L_0)$ grows slightly faster from the beginning, the two grow at
nearly the same rate and when $A_r(L_0)=2$, the area ratio at the
$z=0$ is $A_r(0)=1.96$. The strain
distribution in the vicinity of the notch in Fig.~\ref{LY50}b is very
similar to that in Fig.~\ref{L20} with the strain concentration from
the notch root emanating at about $45^\circ$. In Fig.~\ref{LY50}b the plastic
strain near the notch root exceeds $1.0$ but this strain concentration
is very localized. The value of $\epsilon_p$ in the neck at $z=L_0$ is
about $0.7$ over a fairly large region. The hydrostatic stress
distribution in Fig.~\ref{LY50}c is nearly symmetrical consistent with
the nearly equal necking at $z=0$ and $z=L_0$.
To give an indication of the length scales involved, with, from
Eq.~(\ref{eq:L}), $L_c=c_0/(V_1/L_0)=5$m
$L_0/L_c=0.01$ corresponds to $5$cm so that the transition from
necking at $z=0$ to $z=L_0$ occurs for $L_0$ between about $4$cm and
$8$cm.
\subsection{Fixed $L_0/L_c$, varying $\kappa$}
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{75mm}{\includegraphics{./LY50-ld-xx.eps}}}
\hspace{2cm}
\subfigure[]
{\resizebox{!}{75mm}{\includegraphics{./LY50-arats-xx.eps}}}
\end{center}
\caption{(a) Stress strain curves. (b) Evolution of $R_A$ defined in
Eq.~(\ref{RA}). For
$L_0/L_c=0.010$, $m=0.05$ and various values of $\kappa$.}
\label{LY50-ld}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./LY50-100-ep.eps}}}
\hspace{2cm}
\subfigure[]
{\resizebox{!}{60mm}{\includegraphics{./LY50-10K-ep.eps}}}
\end{center}
\caption{(a) Distribution of effective plastic strain
$\epsilon_p$ for $L_0/L_c=0.010$, $m=0.05$. (a) $\kappa=10$ at
$U/L_0=0.253$. (b) $\kappa=0.1$ at $U/L_0=0.257$.}
\label{LY50-ep}
\end{figure}
The calculations in this section are carried out for $L_0/L_c=0.01$ and
$m=0.05$. In Fig.~\ref{arat}, where $\kappa=1$, this is the case for
which necking occurs nearly simultaneously at $z=0$ and
$z=L_0$. Fig.~\ref{LY50-ld} shows nominal stress-strain curves and the
evolution of $R_A$ with imposed strain, $U/L_0$, for $\kappa=10$, $1$
and $0.1$. Larger values of $\kappa$, see Eq.~(\ref{eq:kap}),
correspond to larger imposed strain rates
relative to the material strain rate $\dot{\epsilon}_0$.
From Eq.~(\ref{eq:ratx}), $t_c \dot{\epsilon}_0 \rightarrow 0$ as
$\kappa \rightarrow \infty$ and $t_c \dot{\epsilon}_0 \rightarrow
\infty$ as $\kappa \rightarrow 0$. With $t_c \dot{\epsilon}_0 $ the ratio of
inertial and material time scales, this indicates that inertia dominates as
$\kappa \rightarrow 0$ and more quasi-static type behavior occurs as
$\kappa \rightarrow \infty$. With $\dot{\epsilon}_0$ non-zero and finite,
this means that increased the imposed strain rate can correspond to more
quasi-static like behavior, which may seem counter intuitive.
To understand this consider $\dot{\epsilon}_0$ fixed and $V_1/L_0$
varying. Increasing $\kappa$, Eq.~(\ref{eq:kap}), then implies
increasing $V_1/L_0$, but
since $L_0/L_c$ fixed, $V_1/c_0$ is fixed. For $V_1/L_0$ to increase
with $V_1/c_0$ unchanged requires $L_0$ to decrease. Since
with $L_0/L_c$ fixed, the stress carried by the loading
wave is fixed, Eq.~(\ref{rel4}), this scenario
corresponds to a smaller specimen with a fixed loading wave
stress. Note also that varying $V_1/L_0$ with $\dot{\epsilon}_0$
fixed, implies that $t_c$ varies, see Eq.~(\ref{rel3}), so that to
obtain the properly
scaled initial/boundary value problem, $T_r$ must be varied so that
$t_r=T_r/t_c$ remains fixed.
Fig.~\ref{LY50-ld}a shows stress-strain plots for $\kappa=10$, $1$ and
$0.1$. The calculation for $\kappa=1$ is the one discussed in Section
\ref{Lc} where necking occurred nearly simultaneously at $z=0$ and
$z=L_0$. The stress strain curves for all three values of $\kappa$ are
qualitatively similar with the stress levels varying as expected due
to the material strain rate sensitivity. It is worth noting that the
maximum strain to necking (i.e. to $A_r=2$ at some cross section)
occurs for the intermediate value $\kappa=1$.
Curves of the ratio $A_r(0)/A_r(L_0)$ are shown in
Fig.~\ref{LY50-ld}b. For $\kappa=1$, $A_r(0)/A_r(L_0) \approx 1$
consistent with the nearly simultaneous necking at $z=0$ and
$z=L_0$. For the increased relative strain calculation, $\kappa=10$,
$A_r(0)/A_r(L_0)>1$, indicating necking at $z=0$, while for the lower
strain rate, $\kappa=0.10$, $A_r(0)/A_r(L_0)<1$ indicating necking at
$z=L_0$.
Contours of effective plastic strain, $\epsilon_p$, for $\kappa=10$
and $\kappa=0.1$ are shown in Fig.~\ref{LY50-ep} showing neck
development at $z=0$ for the higher relative strain rate and at
$z=L_0$ for the lower relative strain rate.
The value of $\kappa$ can be varied by keeping all material,
geometrical and loading parameters fixed except for
$\dot{\epsilon}_0$. In this situation, the results here show that the
where necking occurs can be sensitive to the value used for
$\dot{\epsilon}_0$. Thus, the predicted response can depend
qualitatively, not only quantitatively, on the value of
$\dot{\epsilon}_0$.
\section{Discussion}
With fixed material properties and bar geometry, the results show that
with a fixed imposed strain rate, $V_1/L_0$, a transition from notch triggered
quasi-static like necking to notch ignoring dynamic necking occurs with
increasing bar size, $L_0/L_c$. Thus with $\kappa$ fixed, the effect of
material inertia on necking location increases with increasing values
of the imposed velocity $V_1$. On the other hand, with $L_0/L_c$
fixed (with fixed bar geometry and with fixed material properties) but
with $\kappa$ varying, a similar transition takes place for a decreasing value of
the imposed strain rate $V_1/L_0$. In both cases, the effect of material inertia
increases with increasing bar size $L_0$.
The idea that there can be a critical imposed
velocity under dynamic loading conditions beyond which the apparent
ductility decreases dates back to the 1930s, see for example
\cite{Mann36,K42,KN93,Kl05,Vaz16}. This critical velocity is associated with
material softening and often attributed to adiabatic heating. In the
circumstances analyzed here there is a
critical value of $L_0/L_c$ beyond which the strain
for necking decreases as seen in Fig.~\ref{L-LY-ld}. Since, for fixed
material properties,
$L_0/L_c \propto V_1$, there is a critical imposed velocity beyond
which the necking strain decreases. This
non-monotonic behavior is a consequence of material inertia; there
is no material softening in the formulation.
As seen in Fig.~\ref{LY50-ld}b the qualitative nature of the response
can depend on value of the parameter
$\kappa$ that couples the loading rate and the constitutive response
through the material parameter $\dot{\epsilon}_0$,
see Eq.~(\ref{con5}). Thus, it is worth noting that, for a given value of the
imposed strain rate, the predicted
response can strongly depend on the specified value of
$\dot{\epsilon}_0$.
The parameter
$\kappa$, the ratio between time scales associated
with the material and the loading, does not enter a rate independent
formulation but does enter the formulation
for any value of the rate hardening exponent $m \ne 0$. Presumably,
the rate independent limit is somehow approached as $m \rightarrow
0$; one possibility is that the applied strain at which the responses
separate in Fig.~\ref{LY50-ld}b approaches infinity as $m \rightarrow
0$. In any case, the dependence of the necking behavior on $\kappa$ as $m
\rightarrow 0$ remains to be investigated. Note
that even in the quasi-static
limit, $\kappa$ can play a significant role, as
seen by \cite{Eran16} in a context different from the one considered
here.
The non-dimensional ratio $L_0/L_c$, Eq.~(\ref{rel2}), does enter in the rate
independent limit as does the characteristic time $t_c$,
Eq.~(\ref{rel3}), which then
only serves as a quantity for normalizing the time,
$t=T/t_c$.
In the calculations here only one notch geometry was considered, a
semi-circular notch of radius $R_0/10$. It is expected that the
transition from necking at the notch cross section to necking at
another cross section will depend on the depth of the notch and,
probably to a lesser extent, on the shape of the notch. The results in
\cite{KN93} show that such a transition occurs (in some case with necking taking
place at an intermediate location) even for naturally
necking bars with very small
geometrical imperfections, but the variation of the transition bar size with
increasing notch depth remains to be investigated. Also, the material
model used in the present calculations excludes any softening
mechanism, such as thermal softening and porosity induced
softening, which were included in the calculations in \cite{KN93}. The
dependence of the necking location transition seen here on such
softening mechanisms, as well as on other aspects of the constitutive
characterization of the material remains to be explored.
Fineberg and co-workers, e.g. \cite{jay15}, have used soft materials
with slow wave speeds to study dynamic fracture processes at low
velocities in order to observe aspects of the crack growth process
that would be difficult or impossible to observe directly in hard
materials. The scaling properties embodied in the non-dimensional
equations here suggest that it may be possible also to do this to
study dynamic plastic instabilities, provided of course, that both the
soft and hard materials can be characterized using the same
elastic-viscoplastic constitutive framework.
\section{Conclusions}
The governing equations for dynamic deformations of an
elastic-viscoplastic
notched bar subject to tensile loading were presented in
non-dimensional form. Axisymmetric calculations were carried out for geometrically
identical bars having various sizes. Two key non-dimensional groups were identified
that contain a parameter characterizing the bar size. The main focus
was on variations of
the values of these two parameters that can be regarded as corresponding to
variations in bar size, although other interpretations are possible
and were discussed. Attention was principally directed at the effect
of material inertia on whether necking developed at the notch cross
section or whether necking ultimately occurred away from the notch.
It was found that:
\begin{description}
\item{1.} With a fixed imposed strain rate, the applied strain to
necking (as defined here) does not depend monotonically on size.
\item{2.} The transition from notch induced necking to notch ignoring
necking depends on size and is driven by material inertia.
\item{3.} One of the key non-dimensional groups that
involves a measure of bar size is the ratio of the imposed
velocity to an elastic wave speed, the other is
the ratio of the imposed strain rate to the material characteristic
strain rate. The second non-dimensional group is absent for a rate
independent solid.
\item{4.} With a fixed imposed strain rate and a fixed elastic
wave speed, necking was notch induced for sufficiently small values
of the imposed velocity (smaller bar sizes).
\item{5.} With a fixed imposed velocity and a fixed material characteristic
strain rate, necking was notch induced for sufficiently large values
of the imposed strain rate (smaller bar sizes).
\end{description}
Thus, smaller may be stronger, see e.g. \cite{fleck94,greer05}, but larger is more
dynamic.
\section*{References}
|
\section{Introduction}
For an arbitrary field $k$, Novakovi\'c stated the following as a conjecture in~\cite{novakovic2015tilting}:
\begin{conjecture}
\label{novikov}
Let $X \neq {\blb P}_k^n$ be a $n$-dimensional Brauer-Severi variety. Then $D^b(X)$ does not admit a full strongly exceptional collection.
\end{conjecture}
He proves the conjecture in dimension $n \leq 3$~\cite{novakovic2016no} by exploiting the transitivity of the braid group action on full exceptional collections for ${\blb P}^n_k$ to reduce to an equivalence $D^b(A) \cong D^b(k)$. If $A \cong M_l(D)$, for $D$ a division algebra over $k$, these are just the categories of ${\blb Z}$-graded vector spaces over $D$, respectively $k$, so there is an equivalence only if $D$ is isomorphic to $k$. Since the transitivity of the braid group action (which is only established for $n \leq 3$) is only used to be able to reduce to a single semi-orthogonal component, this suggests that noncommutative motives might provide the right framework for this conjecture. Using some results from~\cite{tabuada2014noncommutative} on noncommutative motives of separable algebras, we prove a slightly stronger version of Conjecture~\ref{novikov}, showing that non-split Brauer-Severi varieties do not admit full \'etale exceptional collections.
\section{Noncommutative motives of separable algebras}
To any small dg-category $\mathcal{A}$, one can associate (functorially) its noncommutative motive $U(\mathcal{A})$, which takes values in a category ${\tt Hmo}_0(k)$. This category has as objects small dg-categories, and for two such categories $\mathcal{A}$ and $\mathcal{B}$,
$$
\Hom_{{\tt Hmo}_0(k)}(\mathcal{A},\mathcal{B}) \cong K_0 {\tt rep}(\mathcal{A},\mathcal{B}),
$$
where ${\tt rep}(\mathcal{A},\mathcal{B})$ is the full triangulated subcategory of $D(\mathcal{A}^{\operatorname{op}} \otimes^{\mathbb{L}} \mathcal{B})$ consisting of those $\mathcal{A}$-$\mathcal{B}$-bimodules $B$ such that for every $x \in \mathcal{A}$, the right $\mathcal{B}$-module $B(x,-)$ is a compact object in $D(\mathcal{B})$. The composition is induced by the derived tensor product of bimodules.
More details on the construction of $U$ can be found in~\cite{tabuada2005additive}, but for the purposes of this note, we will only need that $U$ is a ``universal additive invariant''. An additive invariant is any functor $E:{\tt dgcat}(k) \to D$ taking values in an additive category $D$ such that:
\begin{enumerate}
\item it sends dg-Morita equivalences to isomorphisms,
\item for any pre-triangulated dg-category $\mathcal{A}$, with full pre-triangulated dg-subcategories $\mathcal{B}$ and $\mathcal{C}$ giving rise to a semi-orthogonal decomposition
$$
{\tt H}^0(\mathcal{A})=\langle {\tt H}^0(\mathcal{B}), {\tt H}^0(\mathcal{C}) \rangle,
$$
the morphism $E(\mathcal{B}) \oplus E(\mathcal{C}) \to E(\mathcal{A})$ induced by the inclusions is an isomorphism.
\end{enumerate}
We now review some results from~\cite{tabuada2014noncommutative}. Remember that the category of noncommutative Chow motives $\text{NChow}(k)$ is defined as the idempotent completion of the full subcategory of ${\tt Hmo}_0(k)$ containing the smooth and proper dg-categories. Now let $\text{Sep}(k)$ (respectively $\text{CSep}(k)$) denote the full subcategory of $\text{NChow}(k)$ consisting of the $U(A)$, for $A$ a separable (respectively commutative separable) $k$-algebra. Also let $\text{CSA}(k)^{\oplus}$ denote the closure under finite direct sums of the full subcategory of $\text{NChow}(k)$ consisting of the $U(A)$, for $A$ a central simple $k$-algebras. Note that the $\oplus$ is there since central simple $k$-algebras are not closed under products, whereas (commutative) separable algebras are. In this way $\text{Sep}(k), \text{CSep}(k)$ and $\text{CSA}(k)^{\oplus}$ are additive symmetric monoidal categories.
\begin{theorem}\cite[Corollary 2.13]{tabuada2014noncommutative}
\label{tabvdb2}
There is an equivalence of categories
$$
\{U(k)^{\oplus n} \vert n \in {\blb N}\}\simeq\text{CSA}(k)^{\oplus} \times_{\text{Sep}(k)}
\text{CSep}(k),
$$
i.e. $\{U(k)^{\oplus n} \vert n \in {\blb N}\}$ is a $2$-pullback of categories with respect to the obvious inclusion morphisms.
\end{theorem}
For a central simple algebra $A$ over $k$, denote by $\operatorname{ind}(A)$ and $\deg(A)$ the index (respectively degree) of $A$. Then by~\cite[Proposition 4.5.16]{MR2266528}, $A$ admits a $p$-primary decomposition
$$
A=\bigotimes_{i=1}^k A^{p_i},
$$
where $A^{p_i}$ is uniquely characterised by the property $\operatorname{ind}(A^{p_i})=p_i^{n_i}$ if
$$
\operatorname{ind}(A)=p_1^{n_1} \cdots p_k^{n_k}
$$
is the primary decomposition.
\begin{theorem}\cite[Theorem 2.19]{tabuada2014noncommutative}
\label{tabvdb}
Given central simple $k$-algebras $A_1, \ldots, A_n$ and $B_1, \ldots, B_m$, the
following two conditions are equivalent:
\begin{enumerate}
\item There is an isomorphism of noncommutative motives:
$$
U(A_1) \oplus \cdots \oplus U(A_n) \simeq U(B_1) \oplus \cdots \oplus U(B_m).
$$
\item The equality $n=m$ holds, and for all $1 \leq j \leq n$ and all $p$
$$
[B_j^p]=[A_{\sigma_p(j)}^p]
$$
holds in $\Br(k)$, for some permutations $\sigma_p$ depending on $p$.
\end{enumerate}
\end{theorem}
\begin{remark}
Though the isomorphism classes of objects in $\text{CSA}(k)^{\oplus}$ are in some sense understood by Theorem~\ref{tabvdb}, this is not true for $\text{CSep}(k)$. In fact, using the (additive) equivalence $\text{CSep}(k) \simeq \text{Perm}(G)$, where $G=\operatorname {Gal}(k_{\text{sep}}/k)$, and $\text{Perm}(G)$ is the category of permutation $G$-modules, interesting examples can be obtained from integral representation theory, see~\cite[Remark 2.5, 2.6]{tabuada2014noncommutative}.
\end{remark}
\section{Brauer-Severi varieties and full \'etale exceptional collections}
Denote by $BS(A)$ the Brauer-Severi variety associated to a central simple $k$-algebra $A$. We will say (see also~\cite{Orlov:2014aa}) that an object $E \in D^b(BS(A))$ satisfying $\Hom(E,E[i])=0$ for all $i \neq 0$ is
\begin{itemize}
\item semi-exceptional if $\Hom(E,E)=S$ is a semisimple $k$-algebra,
\item \'etale exceptional if $\Hom(E,E)=L$ is an \'etale $k$-algebra.
\end{itemize}
It is well known~\cite{bernardara2009semiorthogonal} that $BS(A)$ has a full semi-exceptional collection giving rise to a semi-orthogonal decomposition
\begin{equation}
\label{SOD}
D^b(BS(A)) = \langle D^b(k), D^b(A), \ldots, D^b(A^{\otimes \deg(A)-1}) \rangle.
\end{equation}
The following theorem now provides a positive answer to Conjecture~\ref{novikov}.
\begin{theorem}
\label{theorem:nofullexc}
Non-split Severi-Brauer varieties do not admit full \'etale exceptional collections.
\end{theorem}
\begin{proof}
Suppose $A$ is non-split and $\deg(A)=d$. Then if $BS(A)$ has a full \'etale exceptional collection, we deduce from~\eqref{SOD} and additivity of $U(-)$ with respect to semi-orthogonal decompositions that there is an isomorphism
$$
U(k) \oplus U(A) \oplus \cdots \oplus U(A^{\otimes d-1}) \simeq U(D^b(BS(A))) \cong U(L_1) \oplus \cdots \oplus U(L_d),
$$
where the $L_i$ are \'etale $k$-algebras. Using
Theorem~\ref{tabvdb2} and the universal property of fibre products, this isomorphism
gives rise to an isomorphism
$$
U(k) \oplus U(A) \oplus \cdots \oplus U(A^{\otimes d-1}) \simeq U(k)^{\oplus d}.
$$
Now by Theorem~\ref{tabvdb}, for all $p: [A^p]=[k]$ in $\Br(k)$, so $[A]=[k]$ or in other
words $A$ should split.
\end{proof}
\begin{remark}
This result formalizes (in this case) the intuition that for varieties defined over arbitrary fields, one should consider semi-exceptional collections instead of usual exceptional collections.
\end{remark}
\input{non-split-Brauer-Severi-varieties-do-not-admit-full-exceptional-collections.bbl}
\bibliographystyle{amsplain}
\def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
\end{document}
|
Subsets and Splits